Datasets:

NousResearch

/

dolma-v1_7-arxiv

like 1

\section{Introduction} Networks are powerful sources for describing and modeling complex systems. Mining knowledge from networks has become a popular yet challenging area. Many researchers have started to focus on this area. One of the most important tasks in network analysis is node classification. In a typical node classification task, the aim is to predict the most probable labels for nodes in a given network~\citep{tsoumakas2007multi}. For example, in a protein-protein interaction network, the aim is to predict functional labels of proteins~\citep{radivojac2013large}. A batch of informative features is required and important for a supervised machine learning method~\citep{grover2016node2vec}. Node classification problem is always solved as a supervised machine learning problem. Therefore, a feature vector representation for nodes in the network should be appropriately constructed. Recently, learning the feature representation of a node based on its neighbors and structural information of the network has become one of the most active areas~\citep{grover2016node2vec,Perozzi2014DeepWalk,Zhang2015LINE}. Learning low-dimensional feature representation of nodes is also termed as $Network\ Embedding$, which has recently attracted lots of attentions~\citep{Goodfellow2016}. Besides node classification~\citep{Jian2018Toward}, the outputs of network embedding have also been widely used on many important tasks, such as link prediction~\citep{Li2018Streaming} and community detection~\citep{Yang2013Community}. For many real-world systems, multiple types of relations are naturally represented by multiple networks. For example, in social network, the relations between people include friendship relation, money transferring relation, colleague relation and so on. Since multiple networks can describe the real-world systems better in many cases, multi-networks analysis have attracted lots of attention in network science community recently. Unfortunately, most existing network embedding methods focus on single network, and few approaches concentrate on learning node representation based on multiple networks. Therefore, it is urgent to develop an algorithm for learning the feature representation of a node by integrating multi-networks appropriately. One simple solution for multi-network embedding is to summarize multiple networks into a single network and apply the single-network embedding method on the integrated network. Several multi-networks integration methods have been proposed, such as probabilistic methods~\citep{Franceschini2013STRING}, kernel-based methods~\citep{Yu2015Predicting} or weighted averaging or summing~\citep{Mostafavi2008GeneMANIA}. However, this type of integration methods often result in information loss problem when integrating multiple networks into a single one~\citep{Tsuda2005Fast,Lanckriet2004A}. Some approaches try to train individual classifiers on different networks and combine these predictions to a final result using ensemble learning methods~\citep{Yan2010A}. However, these methods consider different networks as independent ones, ignoring the correlation between different networks. In addition, such methods often suffer from learning time and memory constraints~\citep{Gligorijevic2017deepNF}. In multi-network embedding, multiple networks represent different types of relations among the same set of nodes (representing person, commodity, gene and so on). There may be potential correlations between different networks. For multi-networks embedding, one of the most challenging task is how to consider the correlation between different networks. To address this problem, we try to model the correlation between different networks during the feature learning process. Autoencoder~\citep{Rumelhart1986Learning, Baldi2011Autoencoders} is a typical unsupervised deep learning model, which aims to learning a new encoding representation of input data. It has been proved that autoencoder can solve these non-linear feature learning problems effectively. However, existing autoencoder-based methods are not designed for learning multi-network topological features. To benefit from the feature learning power of autoencoder and consider the correlation between multiple networks, we propose a novel multi-network-based feature learning algorithm, named DeepMNE. Considering correlation between multiple networks, DeepMNE applies stacked semi-autoencoder to map input multi-networks into a low-dimension and non-linear space. Here are the major contributions: \begin{itemize} \item We propose a novel semi-supervised autoencoder model for learning feature representations of nodes based on multiple networks. \item To consider the correlation between different networks, we design a communication mechanism among multiple autoencoders corresponding to multiple networks. \item We empirically evaluate DeepMNE for multi-label classification on two tasks of gene function prediction. The experimental results show that DeepMNE outperforms the existing state-of-the-art methods. \end{itemize} \section{Related Work} \subsection{Multi-network Embedding} As an extension of single network embedding, multi-network embedding aims to represent nodes using low-dimensional topological information from multi-networks. Current network embedding approaches mainly focus on single-network embedding and utilize topological structure information to represent nodes. For instance, DeepWalk~\citep{Perozzi2014DeepWalk} treats nodes as words and generates short random walks as sentences. Then, it uses Skip-gram, a word representation learning model, on these random walks to represent nodes of networks. Similar, node2vec~\citep{grover2016node2vec} utilizes a biased random walking procedure to learn topological information, and it uses negative sampling to optimize the Skip-gram model. DNGR~\citep{Cao2016Deep} is a novel method which uses random surfing to learn topological information and applies stacked denoising autoencoder to generate low-dimensional node representation. In general, all these methods focus on single-network representation learning, and few can apply on multi-networks directly. Besides, some multi-network integration methods have been proposed in biological networks area. Mashup~\citep{Cho2016Compact} is an integrative framework for learning low-dimensional feature representations of genes form multiple networks constructed from heterogeneous data sources. Similarity Network Fusion~\citep{Wang2014Similarity} is a widely used networks integration method, which constructs networks for each available data type and then efficiently fuses these networks into one. In addition, there are some other multi-network integration methods, such as Diffusion State Distance~\citep{Cao2014New} and Collective Matrix Factorization~\citep{Z2015Gene}. However, these methods are linear and shallow approaches which cannot capture complex and highly non-linear structure across all networks. \subsection{Gene function prediction} Accurate annotation of gene function is one of the most important and challenging problems in biological area. Predicting gene function aims to assign an unknown gene to the correct functional categories in the annotation database, such as Gene Ontology. To solve this problem, lots of methods based on different types of biological information have been proposed, such as amino acid sequence-based method~\citep{Clark2011Analysis}, protein structure-based method~\citep{Pal2005Inference} and gene expression-based method~\citep{Huttenhower2006A}. With the improvement of experimental methods, functional associations between genes or proteins are often represented in terms of networks, such as gene co-expression networks and protein-protein interaction networks. Several network-based gene or protein function prediction methods have been proposed~\citep{Lehtinen2015Gene,Sharan2007Network}. Multi-networks-based function predictions have been proved better than those methods based on single data source~\citep{Re2010Integration,Cozzetto2013Protein}, because of the complementary nature of different data sources. Thus, lots of algorithms have been proposed for gene function prediction by integrating multiple biological networks~\citep{Cho2016Compact,Sara2010Fast,Cao2014New}. \section{Our Proposed Approach} Multiple-network embedding can be formulated as a semi-supervised feature learning problem. In this part, we propose a novel semi-supervised autoencoder, termed as DeepMNE, to learn the node representation based on multi-networks. Let $V$ be a set of $n$ nodes $\{v_1,v_2,...,v_n\}$. Let $E$ be a set if edges between pairs of $n$-nodes $\{v_1, v_2,...,v_n\}$. Given $k$ networks that include the same set of nodes $V$ but different connectivity between nodes, labeled as $\{G^{(1)}, G^{(2)}, ..., G^{(k)}\}$, a network $G_i$, each network is represented as $G^{(i)}=(V,E^{(i)})$, where $i \in \{1,2,...,k\}$. Our aim is to learn a low-dimension feature representation for each $v \in V$ based on the topological information contained in $\{G^{(1)}, G^{(2)}, ..., G^{(k)}\}$. DeepMNE contains two main components: obtaining global structure information of each network; learning feature representation of nodes by considering both topology of multiple networks and their correlation. \subsection{Step 1. Obtaining global structure information using RWR} It has been proved that random walk with restart (RWR) could capture global associations between nodes in a network~\citep{Cho2016Compact}. Instead of inputting adjacency matrices into DeepMNE directly, we run RWR on each network to capture single network topological information and convert it into feature representations of nodes. The adjacency matrix only describes the relationships between any directly connected nodes, ignoring the global structure of a network. RWR can overcome this drawback, and represent nodes using these high-dimensional network structural information. Besides, we choose the RWR method instead of other recently proposed network embedding methods, such as node2vec~\citep{grover2016node2vec} and DeepWalk~\citep{Perozzi2014DeepWalk}, to capture the topological information, because these methods are computationally intense and require additional hyper-parameter fitting~\citep{Gligorijevic2017deepNF}. Let $M_k$ denote the adjacency matrix of a network $G^{(k)}=(V,E^{(k)})$. The RWR from node $v_i$ can be described as the following recurrence relation. \begin{equation} s_i^{t+1}=(1-\alpha)Ts_i^t + \alpha e_i\label{eq:01} \end{equation} \noindent where $\alpha$ is the restart probability, which balances the effect of local and global topological information in the network; $e_i$ is a n-dimensional distribution vector with $e_i(i)=1$ and $e_i(j)=0$, $\forall j \neq i$; $s_i^t$ is a n-dimensional distribution (column) vector in which each entry holds the probability of a node being visited after $t$ steps in the random walk, starting from node $v_i$; $T$ is the transition probability matrix, and each entry $T_{ij}$, which stores the probability from node $j$ to node $v_i$, can be calculated as $T_{ij} = \frac{M_{ij}}{\sum_{i}{M_{ij}}}$. Based on RWR, we can obtain a matrix $S$, in which $S_{ij}$ is the relevance score between node $v_i$ and $v_j$ defined by RWR-based steady state probabilities. \begin{figure*}[!tpb] \centerline{\includegraphics[scale=0.061]{framework.eps}} \caption{The structure of DeepMNE algorithm. The whole process mainly contains two parts, obtaining topological information learning and learning multi-network-based features. The output of DeepMNE could feed to the following machine learning model. We firstly run random walk with restart (RWR) to learn global structure of networks. Then, constraints extraction and application with semi-supervised autoencoder are iteratively implemented on DeepMNE algorithm to integrate multi-networks. After obtaining the integrated representations of multi-networks, we can train machine learning model based on the outputs of DeepMNE to classify nodes.}\label{fig:framework} \end{figure*} \subsection{Step 2. Multi-network embedding with semiAE} In this section, we propose a novel multi-network embedding algorithm, termed as DeepMNE. The main framework is a DNN structure with autoencoder (AE) and Semi-Supervised autoencoder (semiAE) as its building block. The whole process includes two parts: constraints extraction and constraints application. We use constraints to capture the correlation between different networks. Given several networks $\{G^{(1)}, G^{(2)},..., G^{(k)} \}$ with same nodes, the input of this step is $\{S^{(1)}, S^{(2)},..., S^{(k)} \}$ calculated based on RWR. The main framework is shown in Figure~1\vphantom{\ref{fig:01}}. The first layer of DeepMNE framework is the original autoencoder, which is used for feature extraction and dimension reduction. Starting from the second layer, a revised autoencoder (semiAE) is used for constraint integration and dimension reduction. The dimension of input networks decreases constantly with the extension of the whole iteration model. \subsubsection{Prior Constraints Extraction} The idea of constraints comes from semi-supervised clustering. The pairwise constraints can be typically formatted as must-link and cannot-link constraints~\citep{Basu2004A}. The pairwise constraints can be described as follows: a must-link constraint indicates that these nodes are highly similar or belong to the same cluster, while a cannot-link constraint indicates that two points in the pair are highly dissimilar or belong to different clusters. Given pairs of nodes, we use two strategies to extract constraints. One is to calculate and sort pairwise pearson correlation coefficient (PCC) of all pairs of nodes based on their feature vectors. The top-$k$ pairs are considered as the must-link constraints and the bottom-$k$ pairs are considered as the cannot-link constraints. The other is to set two thresholds for must-link and cannot-link, labeled as $f_1$ and $f_2$ respectively. In detail, a pairs can be adopted as a must-link constraint if its PCC value is larger than $f_1$, and a pair is considered as cannot-link constraint if the PCC value is smaller than $f_2$. After extracting the constraints from the previous layer ($i$ layer), we can apply the must-link and cannot-link constraints to the next layer ($i$+1 layer) as the prior information. \subsubsection{Novel autoencoder with constraints} The key question of DeepMNE is how to integrate prior constraints into the network representation through autoencoder. We revise the original autoencoder and propose a novel variant of autoencoder, termed as Semi-Supervised AutoEncoder (semiAE). Starting from the second layer, the input includes both low-dimensional representations and constrains from previous layer. It is noted that constraints from previous layers' building blocks are based on different networks. Therefore, constraints may be conflicting. To solve this problem, we would merge these constraints and take the intersection of all the constraints as the input of semiAE. Autoencoder is an unsupervised model which is composed of two parts, i.e. the encoder and decoder. The encoder transform the high-dimensional data into a low-dimensional code, and a similar "decoder" network to recover the data from the low dimensional code. The low-dimensional code is then used as a compressed representation of the original data. Let $x_i$ be the $i$-th input vector or node representation of network, and $f$ and $g$ be the activations of the hidden layer and the output layer respectively. We have $h_i=f(Wx_i+b)$ and $y_i=g(Mh_i+d)$, where $\Theta=\{\theta_1,\theta_2\}=\{W,b,M,d\}$ are the parameters to be learned, $f$ and $g$ are the non-linear operators such as the sigmoid function ($sigmoid(z)=1/(1+exp(-z))$) or tanh function ($tanh(z)=(e^z-e^{-z})/(e^z+e^{-z})$). Then the optimization goal is to minimize the reconstruction error between the original data $x_i$ and the reconstructed data $y_i$ from the new representation $h_i$. \begin{equation} \arg\min_{\theta\in{\Theta}}\sum_{i=1}^{n}\parallel y_i-x_i\parallel^2 \label{eq:04} \end{equation} The original autoencoder cannot model the constraints obtained from previous layers. We propose semiAE to take these constraints into account. Let $M$ be a set of must-link pairwise constraints where $(x_i,x_j)\in M$ implies the strong association between $x_i$ and $x_j$. Let $C$ be a set of cannot-link pairwise constraints where $(x_i, x_j)\in C$ implies $x_i$ and $x_j$ are unrelated. The number of constraints is much less than the size of the network $|M| + |C| \leq |S|$. The hypothesis is that $x_i$ and $x_j$ should also close based on the low-dimensional space if there is a must-link constraint between them in previous layer. Ideally, after encoding, two must-link nodes should be closer, and two cannot-link nodes may be more distant. Mathematically, let $d(h(x_i),h(x_j))$ be the error score (difference) between $x_i$ and $x_j$ in the encoded space. For Must-link, $d(x_i,x_j)$ should be larger than $d(h(x_i),h(x_j))$; for Cannot-link, $d(x_i,x_j)$ should be smaller than $d(h(x_i),h(x_j))$. If the pair $(x_i,x_j)$ is a must-link constraint, we add a penalty on the loss function. Similarity, if the pair $(x_i, x_j)$ is a cannot-link constraint, we add a reward on the loss function. The loss function for modeling constraints is defined as follows: \begin{equation} \begin{split} & L_{mc}= \lambda_1\sum_{(x_i,x_j) \in M}d(h(x_i),h(x_j)) - \lambda_2\sum_{(x_i,x_j) \in C}d(h(x_i),h(x_j))\\ & =\lambda_1\sum_{i,j=1}^nM_{i,j}||h(x_i),h(x_j)||_2^2 - \lambda_2\sum_{i,j=1}^nC_{i,j}||h(x_i),h(x_j)||_2^2 \label{eq:05} \end{split} \end{equation} \noindent where matrix $M$ and $C$ are set of must-link and cannot-link constraints respectively; $h(x_i)$ and $h(x_j)$ are the hidden layer representation of input feature vectors $x_i$ and $x_j$ that are from previous layer; $\lambda_1$ and $\lambda_2$ are the weight coefficient, controlling the influence of penalty and reward respectively. To combine constraints with autoencoder, we propose a novel semi-supervised autoencoder, which integrates Eq.~(\ref{eq:04}) and Eq.~(\ref{eq:05}) and joint minimizes the following objective function: \begin{equation} loss = \arg\min_{\theta\in{\Theta}}\sum_{i=1}^{n}\parallel y_i-x_i\parallel^2 + \lambda L_{mc}\label{eq:06} \end{equation} \noindent The first part of Equation~\ref{eq:06} measures the squared error between input and output node features, and the second part measures error score of constraints in hidden layer. \subsubsection{The DeepMNE multi-networks integration algorithm} To optimize the aforementioned model, the goal is to minimize the loss function Eq.~(\ref{eq:06}). In detail, the key step is to calculate the partial derivative of $\frac{\partial L_{mc}}{\partial W}$. And the loss function of $L_{mc}$ can be rephrased as follows: \begin{equation} \begin{split} & L_{mc}= \lambda_1\sum_{i,j=1}^nM_{i,j}||h(x_i),h(x_j)||_2^2 - \lambda_2\sum_{i,j=1}^nC_{i,j}||h(x_i),h(x_j)||_2^2\\ & = 2\lambda_1tr(H^\mathrm{ T }L_MH)-2\lambda_2tr(H^\mathrm{ T }L_C H) \\ & = tr(H^\mathrm{ T }(L_M-L_C)H) \label{eq:07} \end{split} \end{equation} \noindent where $L_M = D_M - M$, $D_M\in{\mathbb{R}^{n\times n}}$ is a diagonal matrix, $D_{M_{i,j}}$ = $\sum_jM_{i,j}$. And $L_C$ is similar as $L_M$. $H$ is the simplified representation of hidden layer. Thus, $\frac{\partial L_{mc}}{\partial W}$ can be translated as: \begin{equation} \frac{\partial L_{mc}}{\partial W} = \frac{\partial L_{mc}}{\partial H} \centerdot \frac{\partial H}{\partial W}=\frac{\partial tr(H^\mathrm{ T }(L_M-L_C)H)}{\partial H} \centerdot \frac{\partial f(XW+b)}{\partial W} \label{eq:08} \end{equation} \noindent where $f$ is activation function(i.e. sigmoid), and we can obtain the partial derivatives of $L_{MC}$. Thus, with an initialization of the parameters, the novel semiAE can be optimized by using stochastic gradient descent (SGD). The pseudocode for DeepMNE is given in Algorithm 1. \begin{algorithm} \caption{The DeepMNE algorithm} \begin{algorithmic}[1] \renewcommand{\algorithmicrequire}{\textbf{Input:}} \renewcommand{\algorithmicensure}{\textbf{Output:}} \Require Multi-networks $G=\{G^{(1)}, G^{(2)},...,G^{(K)}\}$ with $G^{(i)}=(V,E^{(i)})$, the number of iteration $T$, the percentage of constraints $P$, initialization parameters; \Ensure Feature representation of nodes in $V$; \State Run Random Walk with Restart on multi-networks $G$; \State Train AutoEncoder to obtain novel feature representations of nodes in$G'$ and extract initial must-link, cannot-link constraints $M$, $C$; \ForAll{$i \in T$} \ForAll{$k \in K$} \State $M'$, $C'$ = Merge constraints from other networks $M_{all \neq k}$, $C_{all \neq k}$; \State $G_k'$ = Train semi-AutoEncoder with initial parameters on $G_k$ to optimize Eq.~(\ref{eq:06}); \State $M_k$, $C_k$ = Extract must-link and cannot-link constraints based on $G_k'$; \EndFor \EndFor \\ \Return Feature representation of nodes in $V$ \end{algorithmic} \end{algorithm} In the first phase of DeepMNE algorithm, we run random walk with restart algorithm to learn global structure of single biological network. Then, we train autoencoder to learn low-dimensional feature and extract prior constraints based on the network representation of hidden layer. In the main phase, DeepMNE algorithm use an iterative model to train semi-supervised autoencoder with prior constraints. In each iteration, DeepMNE mainly contains three steps, which are merging constraints, training semi-autoencoder and extracting novel constraints. With the increasing of iterations, the model trend to converge and the constraints trends to be unchanged. Finally, DeepMNE generates several low-dimensional feature representations of nodes. The DeepMNE algorithm is a scalable framework model, its training complexity is linear to the number of vertexes $N$. The part of extracting constraints need to calculate pair-wise PCC value which requires $O(N^2)$ . Therefore, the training complexity of DeepMNE algorithm is $O((N^2 + N)TK)$, where $T$ is the number of iteration and $K$ is the number of multiple-networks. \section{Experiments} In order to evaluate the performance of DeepMNE, we test our method on a task of gene function prediction. Gene function prediction is a multi-label classification problem, which aims to assign unknown genes to the correct functional categories in the annotation database~\citep{Cho2016Compact}. We compare our model with the four state-of-the-art network embedding methods (Mashup~\citep{Cho2016Compact}, SNF~~\citep{Wang2014Similarity}, node2vec~\citep{grover2016node2vec} and DeepWalk~\citep{Perozzi2014DeepWalk}) and apply the integrated outputs on the task of gene function prediction using support vector machine. We adopt accuracy, micro-averaged F1, micro-averaged area under precision-recall curve (micro-AUPRC) and micro-averaged area under receiver operating characteristic curve (micro-AUROC) as evaluation metrics. We adopt 5-fold cross-validation to evaluate the performance. \subsection{Datasets} In the experimental part, we evaluate the performance of our method on datasets of Yeast and Human respectively, which collected from the STRING database v9.1~\citep{Franceschini2013STRING}. \begin{itemize} \item \textbf{Yeast} - consisted of six networks over 6,400 genes. The detailed number of edges are listed on the Table 1, where each edge represents the probability of edge presence and the weight between 0 and 1. The functional labels are obtained from Munich Information Center for Protein Sequences~\citep{Ruepp2004The}. The functional categories in MIPS are organized in a three-layered hierarchy, and the number of functional categories in each layer are listed on the Table 2. \item \textbf{Human} - consisted of six networks over 18,362 genes with the number of edges varying from 3,760 to 1,576,332, and the value of every edge are between 0 and 1 (see Table 1). The functional labels are downloaded from the Gene Ontology database~\citep{Ashburner2000Gene}. We group the GO terms for human to obtain three distinct levels of functional categories for different specificities. The details are listed in Table 2. \end{itemize} \begin{table}[ht] \caption{Statistics of datasets. The number of edges in different networks and the average degree of nodes $\langle k\rangle$.} \centering \begin{tabular}{c c c c c} \hline\hline \multirow{2}{*}{Network type} & \multicolumn{2}{c}{Yeast} & \multicolumn{2}{c}{Human}\\ ~ & \# edges & $\langle k\rangle$ & \# edges & $\langle k\rangle$\\ \hline co-expression & 314,013 & 98.129 & 1,576,332 & 171.695 \\ cooccurence & 2,664 & 0.833 & 36,128 & 3.935 \\ database & 33,486 & 10.464 & 319,004 & 34.746 \\ experimental & 219,995 & 68.748 & 618,574 & 67.375 \\ fusion & 1,361 & 0.425 & 3,760 & 0.410 \\ neighborhood & 45,610 & 14.253 & 104,958 & 11.432 \\ \hline \end{tabular} \label{table:01} \end{table} \begin{table}[ht] \caption{Function statistics. The whole dataset contains twelve mini-datasets with different numbers of functions.} \centering \begin{tabular}{c c c c c} \hline\hline \multicolumn{2}{c}{Dataset} & \multicolumn{3}{c}{\# numbers of labels} \\ \hline \multirow{3}{*}{Yeast} & level-1 & \multicolumn{3}{c}{17} \\ ~ & level-2 & \multicolumn{3}{c}{74} \\ ~ & level-3 & \multicolumn{3}{c}{154} \\ \hline \multirow{4}{*}{Human} & & BP & CC & MF\\ \cline{2-5} ~ & 11-30 & 262 & 82 & 153 \\ ~ & 31-100 & 100 & 46 & 72 \\ ~ & 101-300 & 28 & 20 & 18 \\ \hline \end{tabular} \label{table:02} \end{table} \subsection{Parameter Settings} The parameters vary with different datasets. The dimension of each layer on multi-networks integration framework (DeepMNE) is listed in Table 3. In our model, we used restart probability of 0.5 for RWR, which is same as Mashup. The final dimension of network representation are 500 and 800 respectively. The whole DeepMNE-based multi-networks integration algorithm is optimized by using stochastic gradient descent~\citep{Bottou1991Stochastic}. The batch size is 128, the initial learning rate is 0.1 for yeast and 0.2 for human, and the epochs are 200 and 400 respectively. For SNF, we generate an emsemble network and we run singular value decomposition to learn low-dimensional feature representation, and the dimension is same with our model. For node2vec and DeepWalk, we all use the default parameters. For all compared algorithms, we use SVM as the classifier to predict the function labels of genes. \begin{table}[ht] \caption{Neural Network Structures} \centering \begin{tabular}{c c} \hline\hline Datasets & \#nodes in each layer \\ [0.5ex] \hline Yeast & [6400-5220-4040-2860-1680-500] \\ Human & [18362-9181-4580-2295-1148-800] \\ [1ex] \hline \end{tabular} \label{table:03} \end{table} \subsection{Experimental Results on Yeast} The gene function prediction mainly contains three parts: RWR-based global structure caption of single network, low-dimensional feature learning, and SVM-based gene function prediction. In this section, we apply all compared approaches to predict functions of yeast genes based on six networks. All approaches are tested on three tasks corresponding to function labels at different levels (level 1 with 17 categories, level 2 with 74 categories and level 3 with 154 categories). The functional classification at level 1 is more general than level 2 and level 3. Similarly, The functional classification at level 2 is more general than level 3. \begin{figure}[!tpb] \centerline{\includegraphics[scale=0.225]{yeast.eps}} \caption{Performance comparison of different metrics on the task of predicting functional labels for yeast genes.}\label{fig:yeast_plot} \end{figure} Overall, comparing DeepMNE with four other approaches, DeepMNE can achieve better performance on yeast dataset at all three functional classification levels. At level 1, DeepMNE can achieve the highest ACC score that is 0.8378\ (on average), in constract to 0.8063\ for Mashup and 0.6734\ for SNF. Besides, the micro-F1 score of DeepMNE algorithm is 0.7096, which is also higher than other four methods. The micro-average AUPRC and AUROC achieved by DeepMNE on level 1 of yeast are 0.7405 and 0.9100 respectively, which are significantly higher than the scores of Mashup and SNF (see Figure~\ref{fig:yeast_plot}). The performance ranks of the four compared approaches are different at different levels. For example, Mashup is the second best approach at level 1, but node2vec goes to the second place at level 3. However, DeepMNE consistently achieve the best performance at different levels of labels. \subsection{Experimental Results on Human} For further evaluation, we also apply DeepMNE on human dataset to investigate its performance. Instead of RWR-based global structure information, we use the original adjacent matrix as the input of semi-autoencoder directly in this experiment. We test both types of input. Adjacent matrix can achieve better performance in this dataset, since human gene networks includes 18,362 nodes, which may be too large to capture the global structure. As described on the previous section, nodes in human gene networks have three types of labels, termed biological process (BP) , cellular component (CC) and molecular function (MF), corresponding to three respects of gene description. Each type of labels are grouped to three levels, annotating 11-30, 31-100 and 101-300 genes respectively. Thus, we can obtain nine distinct mini datasets. We test all approaches on these datasets. \begin{figure}[!tpb \centerline{\includegraphics[scale=0.225]{human_cc.eps}} \caption{Performance comparison of different metrics on the task of predicting cellular component labels for human genes. }\label{fig:human_cc} \end{figure} Overall, DeepMNE also achieve the highest performance on human dataset(see Figure ~\ref{fig:human_cc}). The accuracy of DeepMNE on human CC-11-30 is 0.5779, which is higher than Mashup, SNF, node2vec and DeepWalk (0.5644, 0.3167, 0.5480 and 0.5181 respectively). DeepMNE still achieves the highest micro-F1 (0.3392), which is sightly higher than the other four methods (Mashup, SNF, node2vec, DeepWalk are 0.3388, 0.2176, 0.3383 and 0.3309 respectively). The AUPRC values of DeepMNE implemented on three categories of human CC are 0.5430, 0.5418 and 0.5246, which are all higher than other four methods (0.5284, 0.5291 and 0.5202 for Mashup, 0.1852, 0.2588 and 0.3141 for SNF, 0.5272, 0.5242 and 0.5128 for node2vec, 0.4971, 4891 and 0.5043 for DeepWalk). Besides, the AUROC values of DeepMNE (0.8796, 0.8533 and 0.8350 respectively) are all significantly higher than Mashup (0.8489, 0.8304, 0.8148), SNF (0.7602, 0.7623, 0.7633), node2vec (0.8675, 0.8376, 0.8227) and DeepWalk (0.8644, 0.8493, 0.8254). The experimental results of Biological Process and Molecular Function categories are listed on Table 4. It is shown that DeepMNE outperforms other methods on MF category of human dataset and also achieves good performance on BP categories. \subsection{Parameters Analysis} \begin{table}[!t] \small \caption{The Accuracy, AUPRC of DeepMNE on gene function prediction on human dataset.}\label{Tab:effects} \centering \renewcommand\arraystretch{1.1} \setlength{\tabcolsep}{1.8mm}{ \begin{tabular}{cp{1.0cm}ccccc} \hline \hline & & \multicolumn{2}{c}{\textbf{Biological Process}} & \multicolumn{2}{c}{\textbf{Molecular Function}}\\ & & Acc & AUPRC & Acc & AUPRC\\ \hline & Mashup & \textbf{0.3825} & \textbf{0.2318} & 0.4486 & 0.3836 \\ & SNF & 0.1972 & 0.0602 & 0.3006 & 0.1662 \\ 11-30 & node2vec & 0.3647 & 0.2248 & 0.4482 & 0.3742 \\ & DeepWalk & 0.3613 & 0.2224 & 0.4466 & 0.3762 \\ & DeepMNE & 0.3672 & 0.1910 & \textbf{0.4751} & \textbf{0.3897} \\ \hline & Mashup & 0.4113 & \textbf{0.2587} & 0.4717 & 0.3666 \\ & SNF & 0.2376 & 0.0892 & 0.2689 & 0.1546 \\ 31-100 & node2vec & 0.3812 & 0.2454 & 0.4355 & 0.3456 \\ & DeepWalk & 0.3852 & 0.2477 & 0.4488 & 0.3654 \\ & DeepMNE & \textbf{0.4129} & 0.2459 & \textbf{0.4936} & \textbf{0.4002} \\ \hline & Mashup & 0.4809 & 0.3795 & 0.5761 & 0.5236 \\ & SNF & 0.3374 & 0.2023 & 0.4248 & 0.3473 \\ 101-300 & node2vec & 0.4721 & 0.3740 & 0.3782 & 0.4959 \\ & DeepWalk & 0.4802 & \textbf{0.3836} & 0.5365 & 0.5011 \\ & DeepMNE & \textbf{0.4824} & 0.3692 & \textbf{0.5882} & \textbf{0.5406}\\ \hline \end{tabular}} \end{table} To evaluate the effect of restart probability to DeepMNE, we re-run DeepMNE with different numbers of dimensions for function prediction on yeast dataset and fix other parameters. Figure~\ref{fig:robustness}(a) shows that performance of DeepMNE is stable over a wide range of number of dimensions. In addition, we also test the robustness of DeepMNE to the restart probabilities by varying restart probabilities and fixing other parameters. From the results, we can find that the performance of DeepMNE is stable on different restart probabilities. \begin{figure}[!tpb] \centerline{\includegraphics[scale=0.24]{yeast_param.eps}} \caption{The AUPRC and AUROC score of DeepMNE with different restart probabilities and numbers of dimensions for function prediction on yeast dataset.}\label{fig:robustness} \end{figure} The number of integrated layers may impact the performance of DeepMNE.We tested our method on yeast dataset with different number of layers. DeepMNE can achieves the highest performance when the number of layers is 5 (see Table 5). \begin{table}[ht] \caption{Results with different numbers of layers} \centering \begin{tabular}{c c c c c} \hline\hline & Layer-3 & Layer-5 & Layer-7 & Layer-10\\ [0.5ex] \hline Accuracy & 0.8203 & 0.8248 & 0.8086 & 0.8095 \\ micro-F1 & 0.5977 & 0.5994 & 0.5888 & 0.5906 \\ AUPRC & 0.7366 & 0.7452 & 0.7358 & 0.7346 \\ AUROC & 0.9036 & 0.9097 & 0.9056 & 0.9066 \\ [1ex] \hline \end{tabular} \label{table:01} \end{table} \vspace{-5pt} \section{Conclusions} Network Embedding, aiming to learn non-linear and low-dimensional feature representation of nodes in networks, has achieved a huge success on many tasks, such as node classification and link prediction. However, current network embedding methods mainly focus on single-network embedding, and few approaches try to learn multi-networks topological information. In this paper, we propose a novel multi-networks embedding algorithm based on semi-supervised autoencoder, termed as DeepMNE. Our approach captures multi-network topological information and takes the correlation among multi-networks into account. We apply our multi-network embedding method on the task of gene function prediction. The experimental results show that DeepMNE outperforms than other state-of-the-art methods and has strong robustness to the number of dimensions and restart probability.

dae30d3caf0d8577f2d5092fe7060979c583f27a

\section{introduction} Rapid development of ultrashort ultraintense laser techniques \cite{Yanovsky2008,Danson2015,Sung2017,Vulcan,ELI,Exawatt} has significantly stimulated the worldwide research interests not only on novel applications of laser-matter interaction \citep{Leemans2006GeV,Schwoerer2006Laser,Hegelich2006,Sarri2014}, but also on the investigation of fundamental issues \citep{Piazza2012,Esarey2009,Mourou2006,Marklund2006}. An example is radiation reaction (RR), which has been discussed since the early days of classical and quantum electrodynamics \cite{Abraham_1905,Lorentz_1909,Dirac_1938,Heitler_1941}, with the testing of the theory being experimentally realized only recently \cite{Cole_2018,Poder_2018}. In ultrastrong laser fields the radiative processes may reach the quantum regime \cite{Goldman_1964,Nikishov_1964,Ritus_1985,Sokolov2009,Sokolov2010,Piazza2010,Blackburn_2014,Li2014, Li2015,Dinu2016,Vranic2016,Harvey2017,Piazza2013,Piazza2014,Li2017}. One of the significant quantum properties of radiation is the stochastic nature, i.e., the discrete and probabilistic character of photon emission \cite{Piazza2013,Piazza2014,Tamburini_2014,Bashinov2015,Wang2015,Li2017}. One signature of stochasticity effects (SE) of radiation is the so-called electron straggling effect, which results in quantitative increase of the yield of the high-energy photons in strong fields\cite{Blackburn_2014}, and the quantum quenching of radiation losses in subcycle petawatt lasers \cite{Harvey2017}. Theoretically it has also been shown that the SE can broaden the energy spread of the electron beam in a plane laser field \cite{Piazza2013,Piazza2014} and cause electron stochastic heating in a standing wave \cite{Bashinov2015}. In a focused laser pulse the SE modified by the ponderomotive force may produce an additional energy spread, as for instance, has been shown in \cite{Wang2015}. Compared with radiative SE signatures \cite{Blackburn_2014, Harvey2017, Li2017}, the relevant signatures in the electron dynamics may be easier for experimental observation, since the diffraction limitation of an electron is much smaller than that of a photon. In this paper we aim at to identify such SE signature in electron dynamics, which would have a qualitative nature and, consequently, would be straightforwardly distinguishable at current achievable experimental conditions. The invariant parameter that characterizes quantum effects in the strong field processes is $\chi\equiv |e|\hbar\sqrt{(F_{\mu\nu}p^{\nu})^2}/m^3c^4$ \cite{Ritus_1985}, where $F_{\mu\nu}$ is the field tensor, $\hbar$ the reduced Planck constant, $c$ the speed of the light in vacuum, $p^{\nu}=(\varepsilon/c,\textbf{p})$ the incoming electron 4-momentum, and $-e$ and $m$ are the electron charge and mass, respectively. When the electron counterpropagates with the laser beam, one may estimate $\chi\approx 2(\hbar\omega_0/mc^2)\xi\gamma$. Here, $\xi\equiv |e|E_0/(m\omega_0 c)$ is the invariant laser field parameter, $E_0$ and $\omega_0$ are the amplitude and frequency of the laser field, respectively, and $\gamma$ is the electron Lorentz-factor. SE are expected to be large when RR is significant, i.e., in the quantum radiation-dominated regime (QRDR), which requires $R\equiv \alpha\xi\chi\gtrsim 1$ \cite{Piazza2012,Koga2005}, indicating that the radiation losses during a laser period are comparable with the electron initial energy. $\alpha$ is the fine structure constant. With the worldwide construction of petawatt laser facilities, laser pulses with an intensity above $10^{22}$ W/cm$^2$ ($\xi\sim 10^2$) are available nowadays \cite{Danson2015,Yanovsky2008,Sung2017}, and much more intense lasers will be produced in the near future \cite{Vulcan,ELI,Exawatt}. Meanwhile, the energies of electrons accelerated by a laser wakefield can be up to several GeV ($\gamma\sim 10^3$) \cite{Mourou2006,Esarey2009,Leemans2006GeV}. Thus, the conditions for SE measurement, $\chi\sim 1$ and $R\sim 1$, are achievable with current experimental techniques. Recently, innovative experimental evidences of quantum RR effects have been realized through radiation spectra from ultrarelativistic positrons in silicon in a regime where quantum RR effects dominate the positron dynamics \cite{Wistisen2018Experimental}, and through the electron energy loss in all-optical experiments \cite{Cole_2018,Poder_2018}, respectively. However, in those experiments all quantum properties, including SE and photon recoil effect, arise simultaneouly, rendering it challenging to identify SE in combination with an appropriate set of theoretical methods. \begin{figure}[t] \includegraphics[width=1\linewidth]{fig1.pdf} \caption{Scenario of SE observation in electron-beam dynamics in nonlinear Compton scattering. An electron beam with a mean kinetic energy of about GeV generated by the laser wakefield accelerator head-on collides with an ultraintense scattering laser pulse. The electrons aggregate inwards to the propagation axis of the laser pulse due to the stochastic nature of the photon emission, which can be observed by the electron angular distribution at the image plate. A magnet is required to split the electron beam from the $\gamma$-ray radiation. } \label{fig1} \end{figure} In this paper, we investigate the SE of photon emissions on the electron-beam dynamics in QRDR, see the interaction scheme in Fig.~\ref{fig1}. A GeV electron beam generated by the laser wakefield acceleration head-on collides with an ultraintense laser pulse. We consider the parameter conditions $\chi\sim 1$ and $\gamma$ is approximately one order of magnitude larger than $\xi$: the former ensures the SE being significant and dominating the electron-beam dynamics, and the latter facilitates the SE under observation, namely, the electron-beam aggregation effect at the center of the electron angular distribution, which overcomes the electron-beam expansion produced by the ponderomotive force due to the transverse profile of the focused laser fields. The electron-beam aggregation effect produces a peak in the electron angular distribution with a FWHM larger than about $40^{\circ}$, which is robust with respect to the laser and electron parameters for current achievable experimental techniques. This paper is organized as follows. In Sec. II we discuss the applied theoretical models for the calculation of the electron dynamics and radiation. In Sec. III th SE signature in electron angular distribution is represented and analyzed. In Sec. IV we investigate the impacts of the laser and electron parameters on the SE signature. A brief conclusion of our work is given in Sec. V. \section{Applied theoretical models for the calculation of the electron dynamics and radiation} We use a theoretical model to calculate the electron dynamics based on Monte-Carlo simulations employing QED theory for the electron radiation and classical equations of motion for the propagation of electrons between photon emissions \cite{Elkina2011,Ridgers2014,Green2015}, which is indicated as Monte-Carlo model (MCM). In ultraintense laser fields, $\xi \gg1$, the coherence length of the photon emission is much smaller than the laser wavelength and the typical size of the electron trajectory \cite{Ritus1985,Khokonov2010}. As a result, the photon emission probability is determined by the local electron trajectory, consequently, by the local value of the parameter $\chi$ \cite{Baier1994}. In every step (far less than the coherence length of the photon emission), the emission process is implemented as a random process, see below. In MCM, the quantum properties of SE and photon recoil are included. In order to highlight the impacts of SE, we carry out additional calculations excluding SE but including other key quantum effects. The latter is based on the, so-called, modified Landau-Lifshitz equation. Generally, the Landau-Lifshitz equation \cite{Landau1975} describes electron dynamics under the action of RR in the classical regime $\chi\ll 1$. In the case of $\chi\sim 1$, the classical Landau-Lifshitz model (LLM) overestimates the RR force, which is remedied phenomenologically in the modified Landau-Lifshitz model (MLLM) \cite{Poder_2018,Piazza2012}. Note that the latter provides results similar to the semi-classical Sokolov equation \cite{Sokolov2009,Sokolov2010}. The MLLM treats electron dynamics classically taking into account the quantum-recoil in RR, however, neglecting SE in photon emission. For easily comprehending our simulation results, the three models are briefly introduced as follows. \subsection{Landau-Lifshitz Model} In this model, the RR is considered as the effect of the elec- tromagnetic fields emitted by an electron on the motion of itself classically. The dynamics of an electron is discribed by the Landau-Lifshitz (LL) equation \cite{Landau1975} \begin{equation} m \frac{du^{\mu}}{d\tilde{\tau}}=eF^{\mu j}u_j + f^{\mu}, \end{equation} where \begin{equation} f^{\mu}=\frac{2e^3}{3mc^2}(\partial_\alpha F^{\mu \nu}u_\nu u^\alpha)+\frac{2e^4}{3m^2c^4}(F^{\mu \nu}F_{\nu \mu}\mu^\alpha+(F^{\nu \beta}u_\beta F_{\nu \alpha}u^\alpha)u^\mu), \end{equation} $u^\mu=(\gamma,\gamma \textbf{v}/c)$ is four-velocity of the electron, $\tilde{\tau}$ the proper time, \begin{equation} \frac{d}{d\tilde{\tau}}=(k\cdot p)\frac{d}{d\tilde{\eta}}, \quad \tilde{\eta}=(k\cdot \tilde{r}), \end{equation} and $\tilde{r}$ the four-vector of the coordinate. The three-dimention equation is \begin{widetext} \begin{eqnarray}\label{LL} {\bm F_{RR, classical}}&=&\frac{2e^3}{3mc^3} \left(\gamma\left(\left(\frac{\partial}{\partial t}+\frac{\textbf{p}}{\gamma m}\cdot\nabla\right){\textbf{E}}+\frac{{\textbf{p}}}{\gamma m c}\times\left(\frac{\partial}{\partial t}+\frac{{\textbf{p}}}{\gamma m}\cdot\nabla\right){\textbf{B}}\right)\right.\nonumber\\ && +\frac{e}{m c}\left({\textbf{E}}\times{\textbf{B}}+\frac{1}{\gamma m c}{\textbf{B}}\times\left({\textbf{B}}\times{\textbf{p}}\right)+\frac{1}{\gamma m c}{\textbf{E}}\left({\textbf{p}}\cdot{\textbf{E}}\right)\right)\nonumber\\ && \left. -\frac{e\gamma}{m^2 c^2}{\textbf{p}}\left(\left({\textbf{E}}+\frac{{\textbf{p}}}{\gamma m c}\times {\textbf{B}}\right)^2-\frac{1}{\gamma^2m^2c^2}\left({\textbf{E}}\cdot{\textbf{p}}\right)^2\right)\right), \end{eqnarray} \end{widetext} where $\textbf{E}$ and $\textbf{B}$ are the electric and magnetic fields, respectively. \subsection{Modified Landau-Lifshitz Model} In this model we treat the electron dynamics in the external field classically but take into account the quantum-recoil corrections. The equation used to calculate the electron dynamics is the modified-LL equation with the classical RR force in the LL equation replaced by the quantum RR force as \cite{Poder_2018,Piazza2012}: \begin{equation} {\bm F_{RR, quantum}}=\frac{I_{QED}}{I_{C}}{\bm F_{RR, classical}}, \end{equation} where, \begin{eqnarray} I_{QED}&=&mc^2\int c\left(k\cdot k' \right)\frac{d W_{fi}}{d\eta dr_0}dr_0,\\ I_{C}&=&\frac{2e^4E'^2}{3m^2c^3}, \end{eqnarray} $W_{fi}$ is the radiation probability, $r_0=\frac{2\left(k\cdot k'\right)}{3\chi\left(k\cdot p_i\right)}$, and $E'$ is the electric fields in the electron frame. $k$, $k'$ and $p_i$ are the four-vector of the wave vector of the driving laser, the wave vector of the radiated photon, and the momentum of the electron before the radiation, respectively. In the modified-LL equation, the recoil effects are included by renormalizing the RR force by the factor $I_{QED}/I_{C}$, the ratio of the radiation intensities within QED and classical approaches, which will account for the classical overestimation of the RR effects on electron dynamics. Note that the same results as the modified-LL equation are provided by the phenomenologically derived equation of motion for an electron in the $\xi \gg 1$ limit, based on the energy-momentum conservation within the system of the electron and emitted photons at each formation length of radiation \cite{Sokolov2009,Sokolov2010}: \begin{eqnarray} \frac{d p^{\alpha}}{d\tilde{\tau}}=\frac{e}{m c}F^{\alpha\beta}p_\beta-\frac{{\cal I}_{QED}}{m c^2}p^{\alpha}+\tau_c\frac{e^2{\cal I}_{QED}}{m^2c^2{\cal I}_c}F^{\alpha\beta}F_{ \beta\gamma}p^{\gamma}, \end{eqnarray} where $\tau_c\equiv 2e^2/(3mc^3)$. \subsection{Monte-Carlo Model} In this model, the calculation of the electron dynamics is based on the Monte-Carlo simulations employing QED theory for the electron radiation and classical equations of motion for the propagation of electrons between photon emission \cite{Elkina2011,Ridgers2014,Green2015}. In superstrong laser fields $\xi\gg 1$, the photon emission probability $W_{fi}$ is determined by the local electron trajectory, consequently, by the local value of the parameter $\chi$ \cite{Ritus1985}: \begin{eqnarray} \frac{d^2 W_{fi}}{d \tilde{\eta} dr_0}=\frac{\alpha \chi [\int_{r_{\chi}}^{\infty} K_{5/3}(x)dx+r_0 r_{\chi} \chi^2 K_{2/3}(r_{\chi})]}{\sqrt{3}\pi\lambdabar_c(k\cdot p_i)}, \label{W} \end{eqnarray} where the Compton wavelength $\lambdabar_c=\hbar/mc$, and $r_{\chi} =r_0/(1-3\chi r_0/2)$. The photon emission of electrons is considered to be a Monte-Carlo stochastic process \cite{Elkina2011,Ridgers2014,Green2015}. During the electron-laser interaction, for each propagation coherent length $\Delta \tilde{\eta}$, the photon emission will take place if the condition $(dW_{fi}$/d$\tilde{\eta})\Delta\tilde{\eta}\geq N_{r}$ is fulfilled, where $N_r$ is a uniformly distributed random number in $[0, 1]$. Herein, the coherent length $\Delta \tilde{\eta}$ is inversely proportional to the invariant laser field parameter $\xi$, i.e., $\Delta \tilde{\eta}\sim 1/\xi$. However, to keep the total photon emission energy consistent, i.e., to exclude numerical error of the simulation of photon emission, we choose $\Delta \tilde{\eta} \ll 1/\xi$. The photon emission probability \begin{eqnarray} W_{fi} = \Delta \tilde{\eta}\frac{dW_{fi}}{d\tilde{\eta}}=\Delta \tilde{\eta}\int_{ \omega_{min}}^{ \omega_{max}} \frac{d^2 W_{fi}}{d \tilde{\eta} d\omega}d\omega, \nonumber \end{eqnarray} where $\omega_{min}$ and $\omega_{max}$ are assumed to equal the driving laser photon energy and the electron instantaneously kinetic energy, respectively. In addition, the emitted photon energy $\omega_R$ is determined by the relation: \begin{eqnarray} \frac{1}{W_{fi}}\int_{\omega_{min}}^{\omega_R}\frac{d W_{fi}(\omega)}{d\omega}d\omega = \frac{\Delta\tilde{\eta}}{W_{fi}}\int_{\omega_{min}}^{\omega_R}\frac{d^2 W_{fi}(\omega)}{d\tilde{\eta} d\omega}d\omega=\tilde{N}_r,\nonumber \end{eqnarray} where, $\tilde{N}_r$ is another independent uniformly distributed random number in $[0, 1]$. Between the photon emissions, the electron dynamics in the laser field is governed by classical equations of motion: \begin{eqnarray} \frac{d\bf{p}}{dt}=e({\bf E}+\frac{\textbf{v}}{c}\times\textbf{B}). \end{eqnarray} Given the smallness of the emission angle $ \sim 1/\gamma$ for an ultrarelativistic electron, the photon emission is assumed to be along the electron velocity. The photon emission induces the electron momentum change ${\bf p}_f \approx (1-\omega_R/c|{\bf p}|) {\bf p}_i$, where ${\bf p}_{i,f}$ are the electron momentum before and after the emission, respectively. \subsection{Employed electromagnetic fields of the laser pulse} In this work, we employ a circularly-polarized tightly-focused laser pulse with a Gaussian temporal profile, which propagates along +z-direction as a scattering laser beam. The spatial distribution of the electromagnetic fields takes into account up to $\epsilon^3$-order of the nonparaxial solution \cite{Salamin2002,Salamin2002PhysRevSTAB}, where $\epsilon=w_0/z_r$, while $w_0$ is the laser beam waist, $z_r=k_0w_0^2/2$ the Rayleigh length with laser wave vector $k_0=2\pi/\lambda_0$, and $\lambda_0$ the laser wavelength. The expressions of the electromagnetic fields are presented in the following \cite{Salamin2002,Salamin2002PhysRevSTAB}: \begin{eqnarray} E_x &=& -i E\left[1+\epsilon^2\left(f^2\widetilde{x}^2-\frac{f^3\rho^4}{4} \right) \right],\\ E_y &=& -i E \epsilon^2 f^2 \widetilde{x}\widetilde{y},\\ E_z &=& E\left[\epsilon f \widetilde{x} + \epsilon^3 \widetilde{x} \left(-\frac{f^2}{2}+f^3\rho^2-\frac{f^4\rho^4}{4}\right) \right],\\ B_x &=& 0,\\ B_y &=& -i E \left[1+\epsilon^2\left(\frac{f^2\rho^2}{2}-\frac{f^3\rho^4}{4} \right)\right],\\ B_z &=& E\left[\epsilon f \widetilde{y} + \epsilon^3 \widetilde{y} \left(\frac{f^2}{2}+\frac{f^3\rho^2}{2}-\frac{f^4\rho^4}{4}\right) \right], \end{eqnarray} where, \begin{equation} E = E_0 F_n f e^{-f\rho^2} e^{i\left(\eta+\psi_{\rm{CEP}}\right)} e^{-\frac{t^2}{\tau^2}}, \end{equation} $\tau$ is the laser pulse duration, and $E_0$ the amplitude of the laser fields with normalization factor $F_n=i$ to keep $\sqrt{E_x^2+E_y^2+E_z^2}=E_0$ at the focus, yielding the scaled coordinates \begin{equation} \widetilde{x}=\frac{x}{w_0}, \quad \widetilde{y}=\frac{y}{w_0},\quad \widetilde{z}=\frac{z}{z_r}, \quad \rho^2=\widetilde{x}^2+\widetilde{y}^2, \end{equation} where $f=\frac{i}{\widetilde{z}+i}$, $\eta=\omega_0 t-k_0z$, and $\psi_{\rm{CEP}}$ is the carrier-envelope phase. \section{The SE signature in electron angular distribution} \begin{figure}[t] \includegraphics[width=1.0\linewidth]{fig2.pdf} \caption{Electron angular distribution: (a)-(c) log$_{10}$[d$^2N_e$/(d$\Theta_x$d$\Theta_y$)] rad$^{-2}$ vs the transverse deflection angles of the electron momenta $\Theta_x \equiv$ arctan($p_x/p_z$) and $\Theta_y \equiv$ arctan($p_y/p_z$). The color bar in (c) applies for (a) and (b) as well. (d)-(f): The angular distribution integrated over the angle region $-5^{\circ}\leq\Theta_y\leq5^{\circ}$, d$\tilde{N}_e$/d$\Theta_x$ = $\int_{-5^{\circ}}^{5^{\circ}}$d$^2N_e$/[d$\Theta_x$d$\Theta_y$] d$\Theta_y$, vs $\Theta_x$ corresponding to simulations of (a)-(c), respectively. Simulations are calculated via MCM ((a) and (d)), MLLM ((b) and (e)), and LLM ((c) and (f)), respectively. The parameters of the laser and electron beam are given in the text.} \label{fig2} \end{figure} An electron beam with characteristics like via laser wakefield accelerators is employed to interact with a counterpropagating focused laser pulse in QRDR, see Fig.~\ref{fig1}. We consider the interaction regime $\gamma\gg\xi/2$, when even in QRDR the electron forward motion persists and the deflection angle in the transverse plane observed on the image plate is mostly determined by the ponderomotive potential due to the transverse profile of the laser beam. We may estimate the deflection angle as $\theta_{d}\sim |F_{p\perp}|\tau/p_{\parallel}\propto (\xi^2/\gamma^2)(\tau/w_0)$, with relativistic ponderomotive force $\textbf{F}_p = -\nabla \xi^2/(2\gamma)$ \cite{Quesnel_1998}, and laser pulse duration $\tau$. This is in contrast to the, so-called, reflection regime $\gamma\lesssim \xi/2$, when the electron is reflected backwards with respect to its initial motion because of combined action of RR and laser ponderomotive force \citep{Salamin1996,Li2015,Li2018}. We investigate the electron dynamics by employing MCM, MLLM, and LLM, respectively, and the corresponding angle-resolved electron-number distributions with respect to the transverse deflection angles of the electron momenta are illustrated in Fig.~\ref{fig2}. The laser peak intensity $I_0\approx 1.4\times10^{22}$ W/cm$^2$ ($\xi=100$), $\lambda_0=1 \mu$m, $w_0=4 \mu$m, the FWHM of laser pulse duration $\tau=16 T_0$, and $T_0$ is the laser period. The pair production probility in such ultrashort laser pulse is negligible. The initial mean kinetic energy of electrons is $\varepsilon_i=1$ GeV ($\gamma\approx 1956.95$, $R\approx 1$, and $\chi_{max}\approx 1.38$) with an energy spread $\Delta\varepsilon_i/\varepsilon_i=0.02$. A cylindrical electron beam is employed, and the beam parameters are set as: radius $w_e= 2 \lambda_0$, length $L_e = 8 \lambda_0$, angular spread $\Delta \theta_e\approx\pm3.6^\circ$, and electron number $N_e = 1.5\times10^5$ (i.e., density $n_e\approx 10^{15}$ cm$^{-3}$ with a Gaussian density profile on the cross section of the electron beam). Those electron-beam parameters are achievable for current laser-plasma acceleration setups \citep{Leemans2014,Kneip2009,Clayton2010,Pollock2011}. \begin{figure}[t] \includegraphics[width=0.95\linewidth]{fig3.pdf} \caption{(a) The initial transverse coordinate distribution of electrons near the $x-z$ plane at $y=0$, which finally contribute to the angular distribution peaks. The cyan-dotted curve in (a) shows the transverse profile of the laser intensity $I$ scaled by $I_0$. The yellow circle represents the boundary of the electron beam, and different colors show the different sample electrons. (b)-(d) Electron angular distribution after the interaction via MCM, MLLM and LLM, respectively. The dash-dotted curves of different colors represent the electron distributions of different sample electrons indicated in (a). The black-solid curves indicate the total electron angular distribution. Other laser and electron beam parameters are the same as in Fig.~\ref{fig2}. } \label{fig3} \end{figure} The MCM simulation which includes SE in Fig.~\ref{fig2}(a) shows that the electrons move inwards to the propagation axis of the laser pulse, consequently, a broad electron-density peak emerges in the middle of the electron angular distribution, which decays exponentially to the peripheries. The electrons concentrate with an angular radius of about $40^\circ$. When SE are excluded, as in MLLM and LLM simulations, see Figs.~\ref{fig2}(b) and \ref{fig2}(c), respectively, the electron angular distributions in both cases have a ``ring'' structure, and the density decays exponentially inwards to the center and outwards to the peripheries. This is because the ponderomotive force deflects the electrons transversely outwards. The SE overcome the deflection effect of the ponderomotive force and cause electrons aggregation inwards to the laser propagation axis. The angular radius of the density ``ring'' is approximately $20^{\circ}$ for MLLM, but $40^{\circ}$ for LLM, since in the latter the LL equation overestimates the RR force, and in the deflection angle estimation $\gamma$ should be replaced by $(\varepsilon_i-\varepsilon_R)/m $, where $\varepsilon_R$ is the electron energy loss due to the radiation. For a quantitative analysis we integrate the electron differential angular distributions in the angular range of $-5^\circ \lesssim \Theta_y \lesssim 5^\circ$, which are represented in Figs.~\ref{fig2}(d)-\ref{fig2}(f), respectively. For MCM, MLLM, and LLM, the electron-density peaks are at $\Theta_x=0^\circ$, $\pm 12^\circ$, and $\pm 33^\circ$, respectively, and the corresponding FWHMs are about $34^\circ$, $7^\circ$, and $12^\circ$, respectively. The current techniques of electron detectors with an angular resolution less than 0.1 mrad \cite{Wang2013,Leemans2014,Wolter2016,Chatelain2014} will allow to experimentally distinguish the angular distributions of the MCM case with those via MLLM and LLM, and in this way identify the SE role. Since $\varepsilon_i$ is too large at chosen parameters, the observation of the electron-number distribution is more convenient than that of the electron-energy distribution \cite{Wang2013,Leemans2014,Wolter2016,Chatelain2014}. To analyze the role of SE in forming the electron distribution, we follow the tracks of a group of sample electrons near the $x-z$ plane at $y=0$, see Fig.~\ref{fig3}. The initial coordinate distribution of the sample electrons are shown in Fig.~\ref{fig3}(a). Note that the electron density has a transverse Gaussian distribution in the cross section of the electron beam, such that the numbers of electrons marked in blue and red are larger than those in green and magenta. Under the deflection effects of the laser fields, electrons in different groups (marked in different colors) produce different profile curves in the final angular distributions in Figs.~\ref{fig3}(b)-\ref{fig3}(d). Apparently, as SE are excluded, see Figs.~\ref{fig3}(c) and \ref{fig3}(d), the sample electrons mainly move outwards under the transverse ponderomotive force. Since $w_0=2w_e$ and the laser-intensity gradient near the peak is small, see Fig.~\ref{fig3}(a), the electrons experience similar laser fields, and consequently, the deflection angle $\theta_d$ concentrates at $\Theta_x=\pm12^\circ$ and $33^\circ$, respectively, with small angular spreads, which is proportional to the laser intensity gradient. Finally, a ``ring'' structure emerges in the electron angular distributions, see Figs.~\ref{fig2}(b) and \ref{fig2}(c). As the SE are necessarily taken into account, comparing Fig.~\ref{fig3}(b) with Fig.~\ref{fig3}(c), the SE in photon emission induce stochastic electron dynamics, and consequently a large spread of the final electron momenta. All electrons have substantial probabilities of moving inwards to the laser propagation axis, which leads to the overlap of angular distributions from different electron groups and the formation of the electron-density peak at $\Theta_x=0^\circ$. \begin{figure}[t] \includegraphics[width=1.0\linewidth]{fig4.pdf} \caption{(a)-(c): Electron angular distribution after the interaction, log$_{10}$[d$^2N_e$/(d$\Theta_x$d$\Theta_y$)] rad$^{-2}$ vs $\Theta_x$ and $\Theta_y$. The color bar in (c) applies for (a) and (b) as well. (d)-(f): d$\tilde{N}_e$/d$\Theta_x$ with respect to $\Theta_x$ corresponding to (a)-(c), respectively. The electron dynamics are simulated via ((a) and (d)) MCM, ((b) and (e)) MLLM, and ((c) and (f)) LLM, respectively. $\Delta\varepsilon_i/\varepsilon_i=0.1$, and other laser and electron parameters are the same as in Fig.~\ref{fig2}.} \label{fig4} \end{figure} \begin{figure}[t] \includegraphics[width=1.0\linewidth]{fig5.pdf} \caption{(a)-(c): Electron angular distribution after the interaction log$_{10}$[d$^2N_e$/(d$\Theta_x$d$\Theta_y$)] rad$^{-2}$ vs $\Theta_x$ and $\Theta_y$. The color bar in (c) applies for (a) and (b) as well. (d)-(f): d$\tilde{N}_e$/d$\Theta_x$ vs $\Theta_x$ corresponding to (a)-(c), respectively. The electron dynamics are simulated via ((a) and (d)) MCM, ((b) and (e)) MLLM, and ((c) and (f)) LLM, respectively. The collision angle of the electron beam $\theta_e=179^{\circ}$, and other laser and electron parameters are the same as in Fig.~\ref{fig2}.} \label{fig5} \end{figure} \begin{figure}[b] \includegraphics[width=1.0\linewidth]{fig6.pdf} \caption{(a)-(c): Electron angular distribution after the interaction log$_{10}$[d$^2N_e$/(d$\Theta_x$d$\Theta_y$)] rad$^{-2}$ vs $\Theta_x$ and $\Theta_y$. The color bar in (c) applies for (a) and (b) as well. (d)-(f): d$\tilde{N}_e$/d$\Theta_x$ vs $\Theta_x$ corresponding to (a)-(c), respectively. The electron dynamics are simulated via ((a) and (d)) MCM, ((b) and (e)) MLLM, and ((c) and (f)) LLM, respectively. The collision angle of the electron beam $\theta_e=175^{\circ}$, and other laser and electron parameters are the same as in Fig.~\ref{fig2}.} \label{fig6} \end{figure} \section{The impacts of laser and electron parameters on the SE signature} We have further investigated the impacts of the laser pulse and electron beam parameters on the SE signature in the electron angular distribution. For experimental feasibility, we first consider the case of a large energy spread of the electron beam. The results in the case of a large energy spread $\Delta\varepsilon_i/\varepsilon_i=0.1$ show a stable SE signature compared with those in Fig.~\ref{fig2}, see Fig.~\ref{fig4}. And, we also investigate the cases with a collision angle $\theta_e=179^\circ$ in Fig. ~\ref{fig5} and $\theta_e=175^\circ$ in Fig.~\ref{fig6}. Comparing Fig.~\ref{fig5} with Fig.~\ref{fig2}, as $\theta_e$ shifts 1$^\circ$ from 180$^\circ$ to 179$^\circ$, in MCM the electron density peak moves left about $6^\circ$; in MLLM and LLM the electron density peaks in the left rise, and those in the right decline, respectively. However, the electron distribution signature is similar to that in Fig.~\ref{fig2}. Moreover, comparing Fig.~\ref{fig6} with Fig.~\ref{fig2}, as $\theta_e$ shifts 5$^\circ$ from 180$^\circ$ to 175$^\circ$, the electrons deposit mainly in the region of $\Theta_x < 0^\circ$. In MCM, the electrons deposit in a sub-elliptial region in the angle-resolved electron distribution with one peak close to the center, see Fig.~\ref{fig6}(a). In MLLM and LLM, the electrons both deposit in a fan-shaped region with one peak at the edge, see Figs.~\ref{fig6}(b) and \ref{fig6}(c). However, the distinctions between the three models are still obvious, see also Figs.~\ref{fig6}(d)-\ref{fig6}(f). Thus, the expected radiative aggregation dynamics of electrons are clearly distinguishable. \begin{figure}[t] \includegraphics[width=1.0\linewidth]{fig7.pdf} \caption{Impacts of (a)-(c) the pulse duration, (d)-(f) the focal radius, and (g)-(i) the peak intensity of the laser pulse, as well as (j)-(l) the initial kinetic energy of the electron beam on the angle-resolved electron-number distributions. The simulation models are MCM (left column), MLLM (middle column), and LLM (right column), respectively. Other laser and electron parameters are the same as in Fig.~\ref{fig2}. } \label{fig7} \end{figure} Furthermore, the role of the laser pulse duration $\tau$ is analyzed in Figs.~\ref{fig7}(a)-\ref{fig7}(c). As $\tau$ increases from 12$T_0$ to 20$T_0$, the laser-electron interaction time increases gradually, which allows the ponderomotive force to deflect the electrons further outwards. Consequently, the peak strength via MCM declines, and the angular radii of the ``ring'' structures in MLLM and LLM both increase. For clear SE one should choose an intermediate laser pulse duration. In fact, as the laser pulse duration, i.e., the laser-electron interaction time, is too long, in MCM the stochastic-radiation aggregation effect of electrons could not overcome the electron-beam-expansion effect due to the ponderomotive force. On the contrary, if the laser pulse duration is too short, via MLLM and LLM the ponderomotive force cannot deflect the electrons outwards enough to form the ``ring'' structure. The role of the laser focal radius is analysed in Figs.~\ref{fig7}(d)-\ref{fig7}(f). The latter show that the case $w_0\approx 2w_e$ is optimal for the observation of SE. When $w_0 =w_e=2\lambda_0$, electrons near the electron-beam boundaries experience rather weak laser fields, can not be deflected outwards much, and consequently, keep their initial motion directions near $\Theta_x=0^\circ$ for all three models. However, when $w_0$ increases to $3w_e$, the laser-intensity gradient on the cross section of the electron beam becomes much smaller, and the laser ponderomotive force $F_p\propto\bigtriangledown|E^2|$ is rather weak accordingly. Thus, the deflection effects are weakened, and the electron angular distributions in Figs.~\ref{fig7}(e) and \ref{fig7}(f) vary little from the center to the peripheries. The laser peak intensity can remarkably affect the electron dynamics, as shown in Figs.~\ref{fig7}(g)-\ref{fig7}(i). As $\xi$ increases from 80 to 120, the electron density near $\Theta_x=0^\circ$ decreases apparently, since the ponderomotive force increases. However, the distinctions among the three models are obvious. Furthermore, the initial kinetic energy of the electron beam does not evidently affect the electron distribution, see Figs.~\ref{fig7}(j)-\ref{fig7}(l). As $\varepsilon_i$ increases from 0.5 GeV to 2 GeV, $\theta_d $ decreases accordingly. We find that the electron aggregation effect is more obvious when the condition of $\xi/\gamma \sim 1/20$ is fulfilled. Thus, the qualitative SE signature is easily observable at current achievable experimental conditions of the laser and electron beam. \section{Conclusion} In conclusion, we have investigated SE of photon emission on the dynamics of an electron beam colliding head-on with an ultraintense focused circularly-polarized laser pulse in the quantum radiation-dominated regime with the condition of $\gamma\sim 20 \xi$. Due to SE the electrons aggregate inwards to the laser propagation axis, resulting in a peak structure in electron angular distribution near the beam propagation direction, with a FWHM of tens of degrees. This is in contrast to the case without SE, when the ponderomotive force of the laser fields will deflect the electrons outwards and generates a ``ring'' structure in the angular distribution with a spread of about $10^\circ$. The SE signature is very robust with regard to the laser pulse and the electron beam parameters in currently available laser facilities. \section{Acknowledgments} This work is supported by the Science Challenge Project of China (No. TZ2016099), the National Key Research and Development Program of China (Grant No. 2018YFA0404801), and the National Natural Science Foundation of China (Grants Nos. 11874295, 11804269, U1532263).

5ac89ce60fd179b3ee3e8eeb7f6190489145426b

\section{Introduction} The unusually small velocity fields, corresponding to less than $2000\, {\rm km\, s^{-1}}$ in full width at half maximum (FWHM), that characterize the optical spectra of narrow-line Seyfert 1 galaxies (NLS1s, \cite{Osterbrock85}), pose one of the most intriguing challenges to our understanding of active galactic nuclei (AGNs). A simple obscuring structure in front of the central regions of the source, which well explains most of the observational properties of broad-line type 1 AGNs and narrow-line type 2s \cite{Antonucci93}, does not apply to the case of NLS1s. Many of these AGNs show prominent \ion{Fe}{II} multiplets \cite{Laor97,Vestergaard01} and non-stellar continuum in the optical spectra, together with strong soft X-ray emission characterized by rapid variability (down to less than $1\,$hr, \cite{Boller96}). These clues suggest that the central source is not obscured. The low velocity fields of NLS1s are, therefore, interpreted in two ways: they may be either attributed to the gravitational attraction of a relatively low mass ($M \approx 10^7\, {\rm M}_\odot$) central super massive black hole (SMBH, \cite{Mathur00, Mathur01}), or they could be the result of extreme projection effects from a flattened rotating structure, seen nearly face-on \cite{Decarli08}. Combining spectroscopic observations with the host galaxy properties, where it is well established, at least in the low redshift regime, that NLS1s reside in spiral hosts that should correspond to typically low mass black holes \cite{Orban11}, several lines of evidence support the low-mass interpretation over the geometrical one. Although it is demonstrated that geometrical effects play some role in the determination of the optical broad line profiles, NLS1s do not appear to be exceptional compared to other type 1 sources in this point of view \cite{LaMura09}. In addition, radio morphology studies suggest that we observe NLS1s at various orientations \cite{Berton18}. On the contrary, the presence of a low mass SMBH in their nuclei implies that NLS1s work at quite high accretion regimes, close to and sometimes beyond the Eddington limit, in order to account for their observed luminosity. This particular property would identify NLS1s with quickly evolving sources, which might be in the first stages of their SMBH growth history. As a consequence, they may provide fundamental hints about the influence of AGN activity in a cosmological context. With this contribution, we present the results obtained from a set of models designed to reproduce the properties of optical emission lines in large spectroscopic databases, like the Sloan Digital Sky Survey (SDSS, \cite{York00}) or the 6dF Galaxy Redshift Survey (6dFGS, \cite{Jones04, Jones09}). We use a systematic sequence of models in order to extract emission-line parameters from spectra and to distinguish among the various classes of astrophysical sources with optical emission lines. We subsequently focus our analysis on the properties of type 1 AGNs and we use the resulting data sample to compare NLS1s with the broad-line emitting sources. \section{Data selection and analysis} In recent times, several authors have investigated the possibility of identifying NLS1s by modelling the profiles of the emission lines detected in the optical spectra published by extensive spectroscopic surveys \cite{Cracco16, Rakshit17, Chen18}. With the publication of the Fourteenth Data Release of SDSS (DR14, \cite{Abolfathi18}) and plans to extend the sky coverage of spectroscopic observations in the Southern Hemisphere, it has become clear that the possibility of modelling spectra and extracting emission-line parameters will be of fundamental importance in order to take advantage of the wealth of observational material that is being released to the scientific community. As a consequence, we started a project to collect data from multiple frequency archives and to combine them with optical spectra, with the aim of selecting large object samples and investigating their statistical properties. \begin{figure}[t] \includegraphics[width=0.48\textwidth]{Fig01a.png} \includegraphics[width=0.46\textwidth]{Fig01b.png} \caption{{\bf Left panel -} Relationship between the SDSS $E(g -r)$ color excess and the $E(B - V)$ color excess reported for the GALEX observation of 9158 QSOs. The dashed line represents the linear fit of Eq.~(\ref{eqExcess}), with a linear correlation coefficient $r > 0.99$. {\bf Right panel -} BPT diagram of 53951 sources, with succesfully fitted diagnostic lines. Here we represent as red dots objects that were fitted as star forming galaxies, with blue points type 1 AGNs and with cyan points type 2 AGNs. The narrow-line components of type 1 objects are consistent with the type 2 population.} \end{figure} By querying the summary of emission-line measurements of the SDSS database\footnote{For this purpose, we used the spectroscopic pipeline data products of SDSS, stored in files named \texttt{spZline-PPPP-MMMMM.fits} for every exposed plate.}, we selected all the spectra with measurements of the optical emission lines H$\beta$, [\ion{O}{III}]$\lambda\lambda 4959,5007$ and H$\alpha$ detected with a signal-to-noise ratio better than 10 and not classified as \texttt{STAR} by the spectroscopic pipeline. We adopted this strategy, instead of directly selecting spectroscopically identified QSOs, because some type 1 objects with prominent narrow line components may happen to be classified as \texttt{GALAXY} and therefore get lost in samples selected on the basis of the \texttt{QSO} classification. We obtained the spectra of 66423 line-emitting galaxies, for which we performed a series of diagnostic operations. First, we applied a Cardelli-Clayton-Mathis extinction correction (CCM, \cite{Cardelli89}), based on the $E(g-r)$ color excess returned by the SDSS photometric pipeline. We converted this into $E(B-V)$ through the empirical relation: \begin{equation} E(B-V) = 0.984 \cdot E(g-r) \label{eqExcess} \end{equation} determined from 9158 QSOs, selected from a cross match of SDSS-DR14 with GALEX observations \cite{Bianchi11}, and illustrated in Fig.~1 (left). Then we transformed all the collected spectra to the rest frame, using the redshift measurement provided by SDSS. After these preliminary steps, we ran a first simple iterative fit to the H$\beta$\ region to infer whether the object has broad emission lines. If the total H$\beta$\ profile is estimated to be broader than [\ion{O}{III}]$\lambda 5007$, the object is treated as a type 1 AGN candidate. The Balmer lines are fitted with multiple Gaussians (1 narrow component, and 1 or 2 broad components, depending on the signal to noise ratio) and the forbidden lines of [\ion{O}{I}]$\lambda 6300$, [\ion{O}{I}]$\lambda 6364$, [\ion{N}{II}]$\lambda\lambda 6548,6584$ and [\ion{S}{II}]$\lambda\lambda 6717,6731$ are fitted with single Gaussians. The [\ion{O}{III}]$\lambda\lambda 4959,5007$ doublet represents a special case that we modelled with a narrow-core component and, in the high S/N spectra, an additional broad/shifted component, requiring, in this case, to fit both lines with the same profile. If no broad component is detected, the object is modelled as a type 2 AGN or a star-forming galaxy, and the emission lines are fitted with single Gaussians, but we still allow for the possible detection of a broad component in H$\alpha$\ (LINER, Seyfert 1.9). This procedure was designed to extract line flux and profile estimates, taking into account the complexity of the recombination lines in type 1 AGNs, which arises from the combination of several spectral contributions. It is able to extract emission line models for a total of 66097 objects. In order to test the consistency of the fitting procedure, we compared the distribution of the narrow line components in type 1 AGNs with the Baldwin-Phillips-Terlevich diagnostic diagram (BPT, \cite{Baldwin81, Veilleux87, Kewley06}) of star forming galaxies and type 2 AGNs. The right panel of Fig.~1 shows that the samples behave consistently, suggesting that the narrow line contributions are reasonably well identified. \begin{figure}[t] \begin{center} \includegraphics[width=0.60\textwidth]{Fig02.png} \end{center} \caption{FWHM of the broad components from the spectra of type 1 AGNs with detected H$\alpha$\ and H$\beta$ emission lines. Objects with S/N < 20 are shown as red points, while objects with S/N > 20 are marked in black. The black dashed line indicates the 1:1 relationship, the filled blue circle marks the position of 1RXS~J075111.5+174350, while the green circle represents Mrk~1243.} \end{figure} \section{Discussion} The main motivation of our study is to extract emission-line parameters from spectroscopic databases through automated procedures designed to model different types of spectra. Since one of the fundamental parameters used to distinguish NLS1s from other objects involves the width of the H$\beta$\ broad-line component, we extracted the profiles of H$\beta$\ and H$\alpha$\ from a sample of 13283 SDSS spectra, where both emission lines were succesfully modelled. By plotting the relationship between the FWHM of H$\beta$\ and H$\alpha$, as shown in Fig.~2, we find that the two lines typically exhibit similar profiles in the low velocity regime, while H$\beta$\ appears to become increasingly broader than H$\alpha$\ in objects with larger velocity widths. This trend is clearly observable even when reducing the sample to the spectra with a S/N > 20 in the continuum at $5100\,$\AA. Differences in the profiles of these two lines, as well as in their reverberation lags, have already been reported and it has been suggested that they might be connected with changes in the line emissivity across the broad line region (BLR, \cite{Kaspi00, Bentz10}). \begin{figure}[t] \begin{center} \includegraphics[width=\textwidth]{Fig03.png} \end{center} \caption{{\bf Left panel - }{\it EPIC}-pn X-ray spectrum of 1RXS~J075111.5+174350, a broad lined AGN that exhibits a remarkable profile difference in the H$\alpha$\ and H$\beta$\ line profiles, with a corresponding X-ray absorption from an ionized medium exceding the foreground absorption of galactic neutral H. {\bf Right panel - }{\it EPIC}-pn X-ray spectrum of Mrk~1243, a narrow-lined Type 1 object with comparable line profiles and no evidence of absorption from intervening gas.} \end{figure} There is, however, also the possibility that a difference in the optical depth at the wavelength of the two lines may result in the observed effect. In particular, it is possible that a layer of ionized plasma, with recombining H atoms that cascade through the first excited level, may result in a different optical depth to the Balmer line photons, since the absorption coefficient for such atoms is $k_\lambda \propto \lambda^2$. In order to look for possible clues of such structures, we investigated the X-ray emission of the sources with fitted emission lines, matching our sample with XMM sources with an estimated flux larger than $10^{-12}\, {\rm erg\, cm^{-2}\, s^{-1}}$ in the 0.2 -- 12~keV range and having high S/N SDSS spectra. In Fig.~3, we show two examples of X-ray spectra:1RXS~J075111.5+174350, a broad-lined type 1 with remarkable line profile differences, and Mrk~1243, a narrow-lined Seyfert 1 with very similar line profiles. While in the first case, the X-ray spectrum shows the effect of absorption in excess of the Milky Way neutral H, with an estimated column density of $1.19 \times 10^{20}\, {\rm cm}^{-2}$, no excess absorption is detected in Mrk~1243. The intervening medium observed in some AGNs may be the result of material ejected from the nuclear regions and -- lying along the line of sight to the central source, under the influence of its strong ionizing continuum -- it should be probably free of dust. The typically more similar profiles of narrow-line emitting sources, particularly in NLS1s, suggest that this type of effect is not similarly common in these objects. This property is consistent with other evidence which, based on IR observations, suggests that the environment of NLS1s is typically richer in dust than that of other type 1 objects \cite{Chen17}. \section{Conclusions} From analysis of the broad components of the emission lines in type 1 AGNs, we find that a systematic difference between H$\alpha$\ and H$\beta$\ becomes more and more significant with increasing line width. If this effect is related to the presence of a layer of ionized gas in front of the source, as some X-ray spectra appear to support, then narrow lined type 1 AGNs, and in particular NLS1s, might be operating in a phase where the development of this structure has not yet been fully accomplished. This consideration, which favours the evolutionary interpretation of the nature of NLS1s over a mere geometrical effect, is also consistent with the claim that NLS1s are hosted in dust-rich environments that would otherwise be swept by the nuclear activity. Clearly the evidence for such an interpretation is still limited and much more effort must be devoted to the execution of multiple wavelength investigations that may include sufficient optical, IR and X-ray data to better clarify the matter. \section*{Acknowledgements} This conference has been organized with the support of the Department of Physics and Astronomy ``Galileo Galilei'', the University of Padova, the National Institute of Astrophysics INAF, the Padova Planetarium, and the RadioNet consortium. RadioNet has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No~730562. J. H. Fan's work is partially supported by the National Natural Science Foundation of China (NSFC 11733001, NSFC U1531245). We thank the referee for comments and suggestions leading to the improvement of this text. \newcommand{A\&A}{A\&A} \newcommand{AJ}{AJ} \newcommand{ApJ}{ApJ} \newcommand{ApJL}{ApJL} \newcommand{ApJS}{ApJS} \newcommand{Ap\&SS}{Ap\&SS} \newcommand{ARA\&A}{ARA\&A} \newcommand{MemSAIt}{MemSAIt} \newcommand{MNRAS}{MNRAS} \newcommand{NewA}{NewA} \newcommand{PASA}{PASA} \newcommand{PASP}{PASP} \bibliographystyle{JHEP}

d2ed86d74622d8112ae8f94b1b3638b574d0b968

\section{Introduction} \label{sec:introduction} Deep Convolutional Neural Networks~(CNNs) have been widely utilized in many applications and achieved remarkable success in computer vision~\cite{szegedy2015going}, speech recognition~\cite{abdel2012applying}, natural language processing~\cite{collobert2008unified}, etc. Going deeper has been proven as an effective approach to improve the model accuracy in solving high-dimensional problems~\cite{ba2014deep,szegedy2015going}. However, when the network depth increases, the number of parameters of the neural network increases too.\\ \begin{figure}[!t] \centering \includegraphics[width=1.0\columnwidth]{4_exp_method.png} \caption{Basic scheme of our proposed 2PFPCE. This scheme show the network pruning process consisting phase I and II.} \vspace{-10pt} \label{fig:4_exp_method} \end{figure} Model compression techniques aim at reducing the storage and computational costs of deep neural networks~\cite{lecun1990optimal, gong2014compressing, han2015deep, wen2016learning}. Network pruning is one important example of model compression techniques that can reduce the network complexity and suppress the over-fitting issue simultaneously. Han et al.~\cite{han2015deep, han2015learning} proposed to reduce network parameters by pruning the weights with small magnitudes and then retrain the network in an iterative manner to maintain the overall accuracy. Majority of the pruned parameters is actually from fully connected layers. Since fully connected layers contribute to very small portion of the total floating point operations (FLOPS), e.g., less than 1\% in VGG-16~\cite{simonyan2014very}, the overall computational cost reduction achieved by this method is very limited~\cite{wen2016learning}. Moreover, the random distribution of the removed weights in memory hierarchy also incurs a higher cache miss rate, which greatly harms the actual performance acceleration obtained in real systems~\cite{wen2016learning}. In some recent work on CNNs~\cite{szegedy2015going, he2016deep}, the fully connected layers are replaced by average pooling layers in order to build a deep architecture with hundreds of layers. Recently, more and more work focused on pruning convolutional layers to reduce computational cost in inference time~\cite{li2016pruning,molchanov2016pruning,luo2017entropy}. Despite the significant weight sparsity in fully connected layers, the non-structured random connectivity ignores cache and memory access issues as indicated in~\cite{wen2016learning}. In some recent work on CNNs~\cite{szegedy2015going, he2016deep}, the fully connected layers are replaced by average pooling layers in order to build a deep architecture with hundreds of layers. The computational cost of the convolutional layers, hence, dominates when the networks become deeper. We note that CNNs with a large scale usually have significant redundancy of their filters and feature channels, which offer a large compression and pruning space. In this work, we propose a \textbf{T}wo-\textbf{P}hase \textbf{F}ilter \textbf{P}runing framework based on \textbf{C}onditional \textbf{E}ntropy, referred to as \textit{2PFPCE}, to prune the filters of CNNs based on conditional entropy in a two-phase manner, as shown in Figure~\ref{fig:4_exp_method}. The key idea of our proposed approach is to establish a quantitative connection between the filters and the model accuracy. We adopt global pruning as phase-one and layer-wise pruning as phase-two. In Phase I, the filters with the minimum conditional entropy is pruned filter-by-filter, followed by an iterative fine-tuning constrained by an accuracy drop threshold. In Phase II, the filters are pruned layer-by-layer in a greedy manner based on conditional entropy, followed by also a fine-tuning of the neural network constrained by the accuracy loss threshold. Our major contributions can be summarized as: \begin{itemize} \item We calculate the conditional entropy over the filters in a convolutional layer, i.e. the distribution of entropy conditioned on the network loss. We also propose to use conditional entropy as a criteria to select the filters to be pruned in our method: \item Based on the assumption that the information of the neurons in a layer shall be uniformly distributed, we propose a novel fine-tuning approach where the weights in a filter corresponding to the neuron with the maximum entropy is kept constant during the back-propagation to reduce over-fitting. \item Based on our observation on the different pruning efficiencies of global and layer-wise pruning strategies, we propose to combine these two strategies to achieve a higher compression ratio of the neural network compared to applying only one strategy in network pruning. Experimental results show that 2PFPCE can achieve a reduction of 88\% filters on VGG16 with only 2\% accuracy degradation. The data volume is decreased from 310784 bytes to 49165 bytes and the inference time is $\sim 54\%$ of the original model. \end{itemize} \section{Related works} \label{sec:preliminary} \subsection{Model compression} The compression techniques of convolutional layers can be roughly categorized into the following three types according to their approximation level: \textbf{Pruning} reduces the redundancy in parameters which are not sensitive to the performance at a level of weight and filter. Network pruning, which aims at reducing the connectivities of the network, is a classic topic in model compression and has been actively studied in the past years. Pruning has been performed at weight level~\cite{han2015deep, wen2016learning} and filter level~\cite{li2016pruning, molchanov2016pruning}. \textbf{Quantization} compresses the network by reducing the number of bits required to represent the weights~\cite{han2015deep}. Binarization~\cite{rastegari2016xnor} is an extreme case of quantization where each weight is represented using only 1-bit. \textbf{Convolution reconstruction} divides convolution into subproblems based on organization of filters at layer level. Low rank approximation~\cite{denil2013predicting, zhang2015efficient, tai2015convolutional, ioannou2015training} imitate convolutional operations by decomposing the weight matrix as a low rank product of two smaller matrices without changing the original number of filters. Based on the correlation between groups of filters, ~\cite{cohen2016group, zhai2016doubly} build a convolutional layer from a group of base filters. FFT convolution~\cite{vasilache2014fast} designs a set of leaf filters with well-tuned in-register performance and reduce convolution to a combination of these filters by data and loop tiling. \textbf{Knowledge distillation}~\cite{hinton2015distilling} compresses an ensemble of deep networks (teacher) into a student network with similar depth by applying a softened penalty of the teacher’s output to the student.\\ This compression method works at network level. There is no golden rule to measure which one of the three kinds of approach is the best. In this work, we focus on filter pruning. There exist some heuristic criteria to evaluate the importance of each filter in the literature such as APoZ (Average Percentage of Zeros)~\cite{hu2016network}, $\ell_1$-norm~\cite{li2016pruning} and Taylor expansion~\cite{molchanov2016pruning}. \begin{itemize} \item APoZ (Average Percentage of Zeros)~\cite{hu2016network}: calculates the sparsity of each channel in output feature map as its importance score $\frac{\sum_{k}^{N}\sum_{j}^{M}f\left ( O_{c, j}(k)=0\right )}{N\times M}$. \item $\ell_1$-norm~\cite{li2016pruning}: measure the relative importance of a filter in each layer by calculating the sum of its absolute weights $\sum\left | \textit{F}_{i, j} \right |$, i.e., its $\ell_1$-norm $\left \| F_{i, j} \right \|_{1}$. \item Taylor expansion~\cite{molchanov2016pruning}: approximate change in the loss function with accumulation of the product of the activation and the gradient of the cost function w.r.t. to the activation $\left | \frac{1}{M}\sum_{m}\frac{\delta C}{\delta z_{l, m}^{(k)}}z_{l, m}^{(k)} \right |$. \end{itemize} Unlike above mentioned criterion, we directly quantize contribution of each filter to accuracy via conditional entropy, discussed following section. \subsection{Information Plane} There is a growing interest in networking understanding and this motivates our information guided pruning. ~\cite{tishby2015deep} proposed to analyze DNNs in the \textit{Information Plane}. The goal of the network is to optimize the Information Bottleneck (IB) trade-off between compression and prediction, successively, for each layer. Two properties of the IB are very important in the context of network pruning. The first is the necessity of redundancy during model training. According to~\cite{tishby2015deep}, the Stochastic Gradient Decent (SGD) optimization has two different and distinct phases: empirical error minimization (ERM) and representation compression. In ERM, redundancy is necessary since the high non-convex optimization is hard to be solved with current technologies. Considering convergence rate, reducing model size after its training is more time efficient. The second is the conditional distribution of output $y$ on $\widetilde{x}$, i.e. $p\left ( y \mid \widetilde{x} \right )$, where $\widetilde{x}$ is the compact expression of input $x$ follows from the Markov chain condition $Y\leftarrow X\leftarrow \widetilde{X}$. It is important to notice that this not a modeling assumption and the quantization $\widetilde{x}$ is not used as a hidden variable in a model of the data. Hence, a network can be decomposed to a cascade of subnetworks with its compact input feature maps as input and original model's output as output. On information plane, Mutual Information (MI) quantifies the average number of relevant bits that the input variable X contains about the label Y. \begin{equation} \begin{split} I\left ( X, Y \right ) & =\sum_{(x, y)\in A}p(x, y)\log[\frac{p(x,y)}{p(x)p(y)}] \\ & =\sum_{(x, y)\in A}p(x, y)\log[\frac{p(x \mid y)}{p(x)}] \\ & = H(X) - H(X|Y) \end{split} \end{equation} The connection between mutual information and minimal sufficient statistics is based on its invariance to invertible transformations: \begin{equation} I(X, Y) = I(\psi (X), \phi (Y)) \end{equation} for any invertible functions $\psi$ and $\phi$. The invariance of the information measures to invertible transformations comes with a high cost. For deterministic functions, the mutual information is insensitive to the complexity of the function or the class of functions it comes from~\cite{shwartz2017opening}. If we have no information on the structure or topology of X, there is no way to distinguish low complexity classes from highly complex classes by the mutual information alone. In this paper, instead of utilizing noise insensitive MI criteria, we propose to adopt conditional entropy in terms of error probability in guessing a finitely-valued random variable $X$ given another random variable $Y$. \section{Conditional entropy based compression} \label{sec:method} In this section, we first formulate compression as an optimization problem, then propose a conditional entropy based filter selection criteria and compare the statistical result of CIFAR10 on pre-trained VGG-16 model. Furthermore, we discuss the relationship of error probability and conditional entropy. \begin{figure*}[!ht] \centering \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_1}% \caption{Maximum conditional entropy.}% \label{maxv}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_2}% \caption{Activation of max conditional entropy.}% \label{maxp}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_3}% \caption{Number of zero activations.}% \label{act_zero}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_5}% \caption{Average conditional entropy of filter.}% \label{ent}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_6}% \caption{Maximum conditional entropy filter ratio.}% \label{maxv_rate}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_7}% \caption{Zero activation ratio.}% \label{act_zero_rate}% \end{subfigure}% \caption{Statistical result of VGG-16 on CIFAR10} \label{fig:3_stat_vgg} \end{figure*} \subsection{Problem formulation} In~\cite{shwartz2017opening,michael2018on}, each layer is seen as a \emph{single} random variable. And the distribution is calculated by joining all neuron outputs in this layer. However the same method might not be suitable for convolutional network, for losing the structure feature on both the input and the feature map. Therefore we adopt the procedure to estimate the mutual information of each convolutional layer as in \cite{kraskov2004estimating,kolchinsky2017estimating}. The first step is to use the entropy of each layer's output as the measurement of the information flow. The \textit{activation entropy} can be calculated using the function below where $p_i$ denotes the probability of \textit{i}-th filter in the feature map. \begin{equation} H_{C_n} = \sum_{i=0}^{n} {p_i * \log p_i} \end{equation} Considering the general scenario of a neural network whose operation is parametrized by a vector~$\theta \in {\Re}^{W}$ (representing the weights), and whose input/output characteristics are described by a conditional probability distribution $p_{\theta }(x_{out} \mid x_{in})$, in which~$x_{in} \in {\Re}^{N}$ and~$x_{out} \in {\Re}^{M}$ denote input and output vectors, respectively. The performance of this network on a given input is measured some loss function~$\varepsilon(x_{in}, x_{out})$. If the probability of an input~$x_{in}$ to be encountered is defined as~$p(x_{in})$, the global error made by a network with parameter is given by \begin{equation} E_{\theta }={\sum_{}^{x_{in}}}{\sum_{}^{x_{out}}}\epsilon (x_{in}, x_{out})p_{\theta }(x_{out}\mid x_{in})p(x_{in}) \end{equation} We define $p_{\theta}(x) = p_{\theta}(x_{out}\mid x_{in})p(x_{in})$, $x = (x_{in}, x_{out}) \in {\Re}^{N+M}$. It now combines both the parametrized properties of the network and the likelihood of input data. When recognizing neural network as a stack of sub networks, the above definition holds true for each layer and thus error in convolutional layers: \begin{equation} E_{\theta }^{C}={\sum_{}^{C_{in}}}{\sum_{}^{x_{out}}}\epsilon (C_{in}, x_{out})p_{\theta }(x_{out}\mid C_{in})p(C_{in}) \end{equation} Where~$C_{in}$ denotes the input feature maps of the convolutional layer. In this way, we reorganize the compression problem as a optimization problem and minimize the distance~$\left \|E_{\theta}^{C} - E_{{\theta}'}^{{C}'}\right \|$, where~${C}'_{in}$ is a minimum subset of $C_{in}$. We adopt a d-dimensional binary vector $\sigma$: a 1 indicating the filter is selected, a 0 indicating the filter is discarded. Notation $x_{\sigma}$ indicates the vector of selected features, that is, the full vector $x$ projected onto the dimensions specified by $\sigma$. Notation $x_{\widetilde{\sigma}}$ is the complement, that is, the unselected features. The full feature vector can therefore be expressed as $x = \left \{ x_{\sigma}, x_{\widetilde{\sigma}} \right \}$. As mentioned, we assume the process $p$ is defined by a subset of the features, so for some unknown optimal vector $\sigma^{*}$, $p\left ( y \mid x \right )= p\left( y \mid x_{\sigma^{*}}\right)$. We approximate $p$ using an hypothetical predictive model $q$, with two layers of parameters: $\sigma$ representing which filters are selected and $\tau$ representing the parameters used to predict y. Our problem statement is to identify the minimal subset of features such that we maximize the conditional likelihood of the training labels, with respect to these parameters. For i.i.d data $D = \left \{ \left ( x^{i}, y^{i} \right );i = 1..N \right \}$ the conditional likelihood of the labels given parameters $\left \{ \sigma , \tau \right \}$ is \begin{equation} L\left(\sigma,\tau\mid D\right) = \prod_{i=1}^{N}q\left( y^{i}\mid x_{\sigma}^{i},\tau\right) \end{equation} The (scaled) conditional log-likelihood is \begin{equation} l=\frac{1}{N} \sum_{i=1}^{N}\textup{log}q\left ( y^{i}\mid x_{\sigma}^{i}, \tau \right ) \end{equation} This is the error function we wish to optimize with respect to the parameters $\left \{ \sigma , \tau \right \}$; the scaling term has no effect on the optima, but simplifies exposition later. We now introduce the quantity $p\left ( y \mid x_{\sigma} \right )$: this is the true distribution of the loss given the selected filters $x_{\sigma}$. Multiplying and dividing $q$ by $p\left ( y \mid x_{\sigma} \right )$, we can re-write the above as, \begin{equation} l=\frac{1}{N} \sum_{i=1}^{N}\log\frac{q\left ( y^{i}\mid x_{\sigma}^{i}, \tau \right )}{p\left ( y^{i}\mid x_{\sigma}^{i}\right )}+\frac{1}{N} \sum_{i=1}^{N}\log{p\left ( y^{i}\mid x_{\sigma}^{i}\right )} \end{equation} The second term in (8) can be similarly expanded, introducing the probability $p\left ( y \mid x \right )$. These are finite sample approximations, drawing data points i.i.d. with respect to the distribution. We use $E_{xy}\left \{ . \right \}$ to denote statistical expectation, and for convenience we negate the above, turning our maximization problem into a minimization. This gives us \begin{equation} \begin{split} -l & \approx E_{xy}\left \{ \log\frac{p\left ( y\mid x_{\sigma}\right )}{q\left ( y\mid x_{\sigma}, \tau \right )} \right \} + E_{xy}\left \{ \log\frac{p\left ( y\mid x\right )}{p\left ( y\mid x_{\sigma} \right )} \right \} \\ & - E_{xy}\left \{ \log{p\left ( y\mid x\right )} \right \} \end{split} \end{equation} In the experiments below, we use the training loss as the single variable. Then our problem statement is to identify the minimal subset of features such that we~\textit{maximize the conditional likelihood of the training loss, with respect to these parameters}. \subsection{Filter selection algorithm} The algorithm is illustrated below as shown in Algorithm~\ref{algo:3_procedure}. For each sample, we calculate the cross entropy loss and output activation corresponding to each filter. To achieve discrete statistical requirement, each parameter is multiplied by a factor of $1e4$ and quantized as 32-bit integer. For each filter, $c\_val$ denotes 1-D distribution on output activation and $c\_bins$ denote a 2-D statistics on output activation conditioned on loss. $c\_total$ is a number of activations per filter, i.e. samples in a dataset. $act\_ent$ denotes the entropy of feature map, provided the distribution of output activation across the dataset. Notice here zero activations are excluded because it's considered to contain no information with respect to the next layer. Given the probability of a specific output activation, $ent_i$ denotes the entropy of output activation conditioned on the distribution of cross entropy loss. $con\_ent$ denotes the conditional entropy of a filter, which is an accumulation of $ent_i$. Then, the $con\_ent$ is sorted in ascending and the filters corresponding to the top-$r$ $con\_ent$ are selected to be removed. In above single-layer illustration, the layer to prune is predefined. This can be generalized to multiple layers or the whole model. \begin{algorithm} \caption{Filter selection algorithm.} \label{algo:3_procedure} \hspace*{\algorithmicindent} \\\textbf{Input}: a baseline model $\boldsymbol{M}$, convolutional layer to prune $l$, training dataset $\textbf{x}_{train}$, number of filters to prune $r$ \hspace*{\algorithmicindent} \\\textbf{Output}: candidate filter(s) to prune $\sigma^r$ \begin{algorithmic} \Procedure{ConditionalEntropy\_Calculation}{} \State $ eps\_h = 1e4 $ \State $\textit{criteria} = CrossEntropy(reduce=false)$. \For {$batch\_idx$ in $batches$} \State $output \gets \boldsymbol{M}(\textbf{x}_{train}[batch\_idx])$. \State $loss[batch\_size] \gets \textit{criteria}(output, target)$. \State $j=eps\_h * loss[batch\_size]$. \For {$k$ in $l$} \State $i=eps\_h * fmap\_out^{l}[k]$. \State $ans[k].extend(i, j)$. \State $c\_bins[i][j] += 1 $. \State $c\_val[i] += 1 $. \EndFor \EndFor \State $c\_total = \sum\limits_{i}^{c\_val}{c\_val[i]}$. \State $act\_ent = \sum\limits_{j \neq 0}^{c\_val}{-\frac{c\_val[j]}{c\_total}*\log\frac{c\_val[j]}{c\_total}}$ \State $ent_i = \frac{c\_val[i]}{c\_total}\sum\limits_{j}{-\frac{c\_bins[i][j]}{c\_val[i]}*\log\frac{c\_bins[i][j]}{c\_val[i]}}$ \State $con\_ent = \sum\limits_{i}{ent_i}$ \State sort $con\_ent$ of each filter in layer $l$ ascending \State add corresponding filter of top-$r$ $con\_ent \to \sigma^r$ \EndProcedure \end{algorithmic} \end{algorithm} \subsection{Statistical result of CIFAR10 on VGG-16} As one of the core concepts in the convolutional network, feature maps reveals huge amount of information about the information flow within the network. For which we proposed using statistical variable, especially conditional entropy, to measure the connection between feature map and error probability, which may help us understand how the information flows through the network. Figure~\ref{fig:3_stat_vgg} is the statistical result of filters in each convolutional layer of CIFAR10 on VGG-16. The x-axis indicate the index of convolution layer in model. Figure~\ref{maxv}, Figure~\ref{maxp} shows the maximum $ent_i$ and its corresponding output activation, respectively. Figure~\ref{act_zero} depicts the number of zero activations and Figure~\ref{act_zero_rate} shows the negative zero activation ratio. We observe zero activation mostly reside in first and last several convolutional layers. Figure~\ref{ent} demonstrate total conditional entropy of each filter. Figure~\ref{maxv_rate} is the ratio of maximum conditional entropy to total conditional entropy and it's remains relatively uniform in different layers. \begin{figure}[!ht] \centering \begin{subfigure}{1.0\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_1}% \caption{First convolutional layer entropy.}% \label{mnist_1}% \end{subfigure}\hfill% \begin{subfigure}{1.0\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_2}% \caption{Second convolutional layer entropy.}% \label{mnist_2}% \end{subfigure}\hfill% \begin{subfigure}{0.8\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_sample}% \caption{Sample under different noise level.}% \label{mnist_noise}% \end{subfigure} \caption{Two layer CNN on MNIST} \label{mnist} \end{figure} \section{Experiments} \label{sec:evaluation} \subsection{Experiment setup} In the initial stage of our experiment, we assume using entropy $H(T_n)$ to measure information flowed in the neural network. To testify our assumption, we made several experiments with MNIST datasets and a two-layered CNN. We added Gaussian noise to the picture and see how the entropy in neural network would change as the power of noise changes. If the entropy effectively measures the information uncertainty in CNN, $H_{T_1}$ and $H_{T_2}$ should increase as the loss increases. The result is affirmative, $H_{T_1}$ and $H_{T_2}$ do increase as expected. Yet there are still some problems about $H_{T_2}$. As shown in Figure~\ref{mnist}, $H_{T_2}$ does not increase as much as $H_{T_1}$ when var is below 0.1. We assume that CNN reduces the influence of noise in the classification task when the noise is not so significant. So the entropy does not increase so much in the second layer. Here, we trained a VGG-16~\cite{simonyan2014very} network with \emph{tanh} activation function on CIFAR10 dataset~\cite{krizhevsky2009learning} and logged the output of each layers. The network has 13 convolution layers, each layer has 64 filters to 512 filters, with the size of 3. The network converged at around 100 epochs, but we trained it for 400 epochs. The final accuracy is $99.6\%$ on training set and $92.68\%$ on test set. As we can observe from Figure~\ref{ent}, conditional entropy in filters are almost uniformly distributed, except for the first convolutional layer where conditional entropy is slightly higher. Also, number of zero activations shown in Figure~\ref{act_zero} indicates the same fact that the first convolutional layer in VGG-16 is greatly redundant in CIFAR10 classification. However, the philosophy that each neuron in network is interchangeable and informative equivalent~\cite{williamson2015equivalence} indicate a uniform distribution of information across the layers. As a result, filters are supposed to be pruned in a layer-wise manner where the number of filters to be pruned should be proportional in each layer, e.g. prune 16 filters in convolutional layer with 64 filters, 32 filters with 128 etc. In the experiment, we evaluate trade-off between accuracy and pruning ratio in global and layer-wise approach respectively. Inspired from above observation, we propose a two phase filter pruning framework based on conditional entropy, namely \textit{2PFPCE}. The filter selection criteria is described in Algorithm~\ref{algo:3_procedure}. The procedure is shown in Figure~\ref{fig:4_exp_method}. In phase I, filters are pruned and fine-tuned iteratively until the accuracy drop reach the threshold (1\%). The purpose is to remove the redundancy filters with respect to the dataset. Notice that dataset plays a crucial role when it comes to compression. The number of involved features in a 1000-category dataset is probably much larger than that in a 10-category dataset. After a layer is pruned, weights in the filter w.r.t the maximum activation entropy are kept constant and cannot be updated during fine tuning. Similar to dropout~\cite{srivastava2014dropout}, this aims at penalizing any single neuron that may overly fitted to the dataset. In phase II, filters are pruned in a layer-wise manner: in each iteration, a small portion filters (1/32 or 1/16) of each layer are pruned until the accuracy drop reach the threshold $\gamma$. The threshold $\gamma$ is a hyper-parameter and can be adjusted to satisfy application constraint. To retain the information in pruned filters and avoid time consuming fine-tune, we update the bias term as the activation of maximum conditional entropy filter (Figure~\ref{maxp}). \subsection{Global Pruning Approach} \label{sec:exp_glob} In this experiment, we evaluate the trade-offs between computation complexity and classification accuracy in global pruning approach. As shown in Figure~\ref{fig:4_glob_vgg} and Figure~\ref{fig:4_glob_resnet}, the accuracy w.r.t pruning ratio is similar in both networks. Without fine-tuning, the accuracy in VGG-16 and ResNet-18 decrease significantly to 27\% and 16\% within a pruning ratio of 10\% and 21\%, respectively. After the turning point, in VGG-16 the accuracy slowly decrease to 15\% as the pruning ratio increase while in ResNet-18 the accuracy drop is within 1\%. With fine-tuning, each network can remain its accuracy with pruning ratio below 20\% and then slowly degrade with a marginal accuracy drop between 1\%-2\% until the pruning ratio reaches 90\%. The comparison between w/o fine-tune shows the DNNs can recover from small disturb with fine tuning. The contribution of each filter is very similar in a converged convolutional neural network. Though the experiment results on VGG-16 and ResNet-18, we can conclude that our current solution can prune about 40\% filters with 1\% accuracy compromise and 80\% prunes with only 2\% accuracy compromise with global pruning approach. \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth, trim={3cm 1.6cm 3cm 2.5cm},clip]{figure/4_glob_vgg} \caption{Accuracy vs. pruning ratio of VGG16.} \label{fig:4_glob_vgg} \end{figure} \begin{figure}[t] \centering \includegraphics[width=0.98\columnwidth, trim={3cm 1.5cm 2.7cm 2.8cm},clip]{figure/4_glob_resnet} \caption{Accuracy vs. pruning ratio of Preact-ResNet18.} \vspace{-6pt} \label{fig:4_glob_resnet} \end{figure} \subsection{Layer-wise Pruning Approach} \label{sec:exp_layer} Table~\ref{table:res_cifar10_1} compares the obtained accuracy of different filter importance criterion under the same pruning ratios. For a fair comparison, the DNNs are pruned in a layer-wise approach and fine-tuned same number of epochs for accuracy recover after pruning. At each pruning iteration, we remove a percentage of feature maps and then perform 20 minibatch SGD updates with batch-size 32, momentum 0.9, learning rate $10^{-4}$, and weight decay $10^{-4}$. As we can see from the table, our proposed conditional entropy filter selection criteria outperforms counterparts with highest accuracy on both pruning ratio 25\% and 50\%. Table~\ref{table:res_cifar10_2} compares the obtained pruning ratio of filter importance criterion under approximately the same errors, e.g., within 2\% accuracy loss. The baseline is L1-Norm which has a pruning ratio of 60\% on both networks. Our experiments show that with a marginal accuracy loss, conditional entropy based criterion can always achieve largest pruning ratio. \begin{table} \centering \begin{subtable}{0.5\textwidth} \centering \caption{Filter importance criteria vs. accuracy.} \label{table:res_cifar10_1} \vspace{4pt} \begin{tabular}{ |p{1.0cm}|p{1.4cm}|p{1.4cm}|p{1.4cm}|p{1.4cm}| } \hline \multicolumn{5}{|c|}{VGG16/CIFAR10: 92.98\%} \\ \hline Prune & L1-Norm & APoZ & Act. ent & Cond. ent\\ \hline 25\% & 92.86\% & 92.94\% & 92.93\% & 93.50\% \\ 50\% & 92.11\% & 92.02\% & 92.00\% & 92.76\% \\ \hline \multicolumn{5}{|c|}{ResNet50/CIFAR10: 93.16\% } \\ \hline 20\% & 94.42\% & 94.36\% & 94.33\% & 94.84\% \\ 50\% & 94.48\% & 94.25\% & 94.42\% & 94.44\% \\ \hline \end{tabular} \end{subtable} \begin{subtable}{0.5\textwidth} \centering \vspace{6pt} \caption{Filter importance criteria vs. pruning ratio.} \label{table:res_cifar10_2} \vspace{4pt} \begin{tabular}{ |p{1.6cm}|p{1.35cm}|p{1.1cm}|p{1.35cm}|p{1.35cm}| } \hline \multicolumn{5}{|c|}{VGG16/CIFAR10: 92.98\%} \\ \hline Acc. & L1-Norm & APoZ & Act. ent & Cond. ent\\ \hline 91.0($\pm$0.3)\% & $1.0\times$ & $0.88\times$ & $1.27\times$ & $1.32\times$ \\ \hline \multicolumn{5}{|c|}{ResNet50/CIFAR10: 93.16\% } \\ \hline 92.0($\pm$0.2)\% & $1.0\times$ & $0.93\times$ & $0.96\times$ & $1.05\times$ \\ \hline \end{tabular} \end{subtable} \caption{Comparison of filter importance criteria in layer-wise approach VGG-16/ResNet-50 on CIFAR10} \label{4_layer_cmp} \end{table} \subsection{Stage Pruning Approach} \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth]{figure/4_stage_all} \caption{Stage pruning with 98\% accuracy threshold.} \label{fig:4_stage_all} \end{figure} We focus on reducing the number of convolutional feature maps and the total estimated floating point operations (FLOPs). During pruning we were measuring reduction in computations by FLOPs, which is a common practice~\cite{han2015deep}. Improvements in FLOPs result in monotonically decreasing inference time of the networks because of removing entire feature map from the layer. However, time consumed by inference dependents on particular implementation of convolution operator, parallelization algorithm, hardware, scheduling, memory transfer rate etc. Unlike in previous work, where latency is considered via another, often inaccurate proxy (e.g., FLOPS), in our experiments, we directly measure real-world inference time by executing the model on GPU (NVIDIA Titan Xp with CUDNN 8.0). Therefore we measure improvement in the inference time to see real speed up compared to original networks in Figure~\ref{fig:4_stage_all}. Based on the observations mentioned in setup, a two-phase pruning approach is proposed. As shown in Figure~\ref{fig:4_stage_all}, a combination of both approach can achieve a pruning ratio of 88\% within 2\% accuracy drop. The dataset is CIFAR10 with a minibatch size of 32. The inference time on pre-trained VGG-16 is 22.69ms. In Phase I, number of filters decrease from 4224 to 3960 and total bandwidth from 310784 bytes to 273658 bytes. The inference time reduces to 19.24ms. Removed filters are mostly from the first convolutional layer and this verifies our observation on redundancy. In the first convolutional layer, $\sim 60\%$ filters are removed result in a reduction in data volume from 64M to 35M. Because of high parallelization in GPU, inference time decrease in a single layer is not as significant as in data volume. In Phase II, number of filters decrease from 3960 to 532 and total bandwidth from 273658 bytes to 49165 bytes. The inference time reduces to 12.34ms. \section{Conclusion} \label{sec:conclusion} In this work, inspired from the statistical result of information flow in neural network, we propose to use conditional entropy as the filter selection criteria in filter pruning. The performance of conditional based filter selection criteria outperforms approaches based L1-Norm, APoZ and activation entropy. Experimental result shows the proposed criteria can achieve 92.76\% accuracy when pruning ratio is 50\% and $\sim 91\%$ accuracy when pruning ratio is 80\%. In both VGG-16 and ResNet-50, our proposed conditional entropy outperforms the above criterion. To comply with the network information distribution, we adopt a two phase pruning framework which combines global approach with layer-wise approach. In addition, novel model tuning techniques are proposed 1) freeze weights w.r.t maximum activation entropy to avoid over-fitting, 2) update bias w.r.t the activation of the maximum conditional entropy filter. The above framework can achieve a pruning ratio of 88\% within 2\% accuracy drop of pre-trained VGG-16 model on CIFAR10. \bibliographystyle{plain} \section{Introduction} \label{sec:introduction} Deep Convolutional Neural Networks~(CNNs) have been widely utilized in many applications and achieved remarkable success in computer vision~\cite{szegedy2015going}, speech recognition~\cite{abdel2012applying}, natural language processing~\cite{collobert2008unified}, etc. Going deeper has been proven as an effective approach to improve the model accuracy in solving high-dimensional problems~\cite{ba2014deep,szegedy2015going}. However, when the network depth increases, the number of parameters of the neural network increases too.\\ \begin{figure}[!t] \centering \includegraphics[width=1.0\columnwidth]{4_exp_method.png} \caption{Basic scheme of our proposed 2PFPCE. This scheme show the network pruning process consisting phase I and II.} \vspace{-10pt} \label{fig:4_exp_method} \end{figure} Model compression techniques aim at reducing the storage and computational costs of deep neural networks~\cite{lecun1990optimal, gong2014compressing, han2015deep, wen2016learning}. Network pruning is one important example of model compression techniques that can reduce the network complexity and suppress the over-fitting issue simultaneously. Han et al.~\cite{han2015deep, han2015learning} proposed to reduce network parameters by pruning the weights with small magnitudes and then retrain the network in an iterative manner to maintain the overall accuracy. Majority of the pruned parameters is actually from fully connected layers. Since fully connected layers contribute to very small portion of the total floating point operations (FLOPS), e.g., less than 1\% in VGG-16~\cite{simonyan2014very}, the overall computational cost reduction achieved by this method is very limited~\cite{wen2016learning}. Moreover, the random distribution of the removed weights in memory hierarchy also incurs a higher cache miss rate, which greatly harms the actual performance acceleration obtained in real systems~\cite{wen2016learning}. In some recent work on CNNs~\cite{szegedy2015going, he2016deep}, the fully connected layers are replaced by average pooling layers in order to build a deep architecture with hundreds of layers. Recently, more and more work focused on pruning convolutional layers to reduce computational cost in inference time~\cite{li2016pruning,molchanov2016pruning,luo2017entropy}. Despite the significant weight sparsity in fully connected layers, the non-structured random connectivity ignores cache and memory access issues as indicated in~\cite{wen2016learning}. In some recent work on CNNs~\cite{szegedy2015going, he2016deep}, the fully connected layers are replaced by average pooling layers in order to build a deep architecture with hundreds of layers. The computational cost of the convolutional layers, hence, dominates when the networks become deeper. We note that CNNs with a large scale usually have significant redundancy of their filters and feature channels, which offer a large compression and pruning space. In this work, we propose a \textbf{T}wo-\textbf{P}hase \textbf{F}ilter \textbf{P}runing framework based on \textbf{C}onditional \textbf{E}ntropy, referred to as \textit{2PFPCE}, to prune the filters of CNNs based on conditional entropy in a two-phase manner, as shown in Figure~\ref{fig:4_exp_method}. The key idea of our proposed approach is to establish a quantitative connection between the filters and the model accuracy. We adopt global pruning as phase-one and layer-wise pruning as phase-two. In Phase I, the filters with the minimum conditional entropy is pruned filter-by-filter, followed by an iterative fine-tuning constrained by an accuracy drop threshold. In Phase II, the filters are pruned layer-by-layer in a greedy manner based on conditional entropy, followed by also a fine-tuning of the neural network constrained by the accuracy loss threshold. Our major contributions can be summarized as: \begin{itemize} \item We calculate the conditional entropy over the filters in a convolutional layer, i.e. the distribution of entropy conditioned on the network loss. We also propose to use conditional entropy as a criteria to select the filters to be pruned in our method: \item Based on the assumption that the information of the neurons in a layer shall be uniformly distributed, we propose a novel fine-tuning approach where the weights in a filter corresponding to the neuron with the maximum entropy is kept constant during the back-propagation to reduce over-fitting. \item Based on our observation on the different pruning efficiencies of global and layer-wise pruning strategies, we propose to combine these two strategies to achieve a higher compression ratio of the neural network compared to applying only one strategy in network pruning. Experimental results show that 2PFPCE can achieve a reduction of 88\% filters on VGG16 with only 2\% accuracy degradation. The data volume is decreased from 310784 bytes to 49165 bytes and the inference time is $\sim 54\%$ of the original model. \end{itemize} \section{Related works} \label{sec:preliminary} \subsection{Model compression} The compression techniques of convolutional layers can be roughly categorized into the following three types according to their approximation level: \textbf{Pruning} reduces the redundancy in parameters which are not sensitive to the performance at a level of weight and filter. Network pruning, which aims at reducing the connectivities of the network, is a classic topic in model compression and has been actively studied in the past years. Pruning has been performed at weight level~\cite{han2015deep, wen2016learning} and filter level~\cite{li2016pruning, molchanov2016pruning}. \textbf{Quantization} compresses the network by reducing the number of bits required to represent the weights~\cite{han2015deep}. Binarization~\cite{rastegari2016xnor} is an extreme case of quantization where each weight is represented using only 1-bit. \textbf{Convolution reconstruction} divides convolution into subproblems based on organization of filters at layer level. Low rank approximation~\cite{denil2013predicting, zhang2015efficient, tai2015convolutional, ioannou2015training} imitate convolutional operations by decomposing the weight matrix as a low rank product of two smaller matrices without changing the original number of filters. Based on the correlation between groups of filters, ~\cite{cohen2016group, zhai2016doubly} build a convolutional layer from a group of base filters. FFT convolution~\cite{vasilache2014fast} designs a set of leaf filters with well-tuned in-register performance and reduce convolution to a combination of these filters by data and loop tiling. \textbf{Knowledge distillation}~\cite{hinton2015distilling} compresses an ensemble of deep networks (teacher) into a student network with similar depth by applying a softened penalty of the teacher’s output to the student.\\ This compression method works at network level. There is no golden rule to measure which one of the three kinds of approach is the best. In this work, we focus on filter pruning. There exist some heuristic criteria to evaluate the importance of each filter in the literature such as APoZ (Average Percentage of Zeros)~\cite{hu2016network}, $\ell_1$-norm~\cite{li2016pruning} and Taylor expansion~\cite{molchanov2016pruning}. \begin{itemize} \item APoZ (Average Percentage of Zeros)~\cite{hu2016network}: calculates the sparsity of each channel in output feature map as its importance score $\frac{\sum_{k}^{N}\sum_{j}^{M}f\left ( O_{c, j}(k)=0\right )}{N\times M}$. \item $\ell_1$-norm~\cite{li2016pruning}: measure the relative importance of a filter in each layer by calculating the sum of its absolute weights $\sum\left | \textit{F}_{i, j} \right |$, i.e., its $\ell_1$-norm $\left \| F_{i, j} \right \|_{1}$. \item Taylor expansion~\cite{molchanov2016pruning}: approximate change in the loss function with accumulation of the product of the activation and the gradient of the cost function w.r.t. to the activation $\left | \frac{1}{M}\sum_{m}\frac{\delta C}{\delta z_{l, m}^{(k)}}z_{l, m}^{(k)} \right |$. \end{itemize} Unlike above mentioned criterion, we directly quantize contribution of each filter to accuracy via conditional entropy, discussed following section. \subsection{Information Plane} There is a growing interest in networking understanding and this motivates our information guided pruning. ~\cite{tishby2015deep} proposed to analyze DNNs in the \textit{Information Plane}. The goal of the network is to optimize the Information Bottleneck (IB) trade-off between compression and prediction, successively, for each layer. Two properties of the IB are very important in the context of network pruning. The first is the necessity of redundancy during model training. According to~\cite{tishby2015deep}, the Stochastic Gradient Decent (SGD) optimization has two different and distinct phases: empirical error minimization (ERM) and representation compression. In ERM, redundancy is necessary since the high non-convex optimization is hard to be solved with current technologies. Considering convergence rate, reducing model size after its training is more time efficient. The second is the conditional distribution of output $y$ on $\widetilde{x}$, i.e. $p\left ( y \mid \widetilde{x} \right )$, where $\widetilde{x}$ is the compact expression of input $x$ follows from the Markov chain condition $Y\leftarrow X\leftarrow \widetilde{X}$. It is important to notice that this not a modeling assumption and the quantization $\widetilde{x}$ is not used as a hidden variable in a model of the data. Hence, a network can be decomposed to a cascade of subnetworks with its compact input feature maps as input and original model's output as output. On information plane, Mutual Information (MI) quantifies the average number of relevant bits that the input variable X contains about the label Y. \begin{equation} \begin{split} I\left ( X, Y \right ) & =\sum_{(x, y)\in A}p(x, y)\log[\frac{p(x,y)}{p(x)p(y)}] \\ & =\sum_{(x, y)\in A}p(x, y)\log[\frac{p(x \mid y)}{p(x)}] \\ & = H(X) - H(X|Y) \end{split} \end{equation} The connection between mutual information and minimal sufficient statistics is based on its invariance to invertible transformations: \begin{equation} I(X, Y) = I(\psi (X), \phi (Y)) \end{equation} for any invertible functions $\psi$ and $\phi$. The invariance of the information measures to invertible transformations comes with a high cost. For deterministic functions, the mutual information is insensitive to the complexity of the function or the class of functions it comes from~\cite{shwartz2017opening}. If we have no information on the structure or topology of X, there is no way to distinguish low complexity classes from highly complex classes by the mutual information alone. In this paper, instead of utilizing noise insensitive MI criteria, we propose to adopt conditional entropy in terms of error probability in guessing a finitely-valued random variable $X$ given another random variable $Y$. \section{Conditional entropy based compression} \label{sec:method} In this section, we first formulate compression as an optimization problem, then propose a conditional entropy based filter selection criteria and compare the statistical result of CIFAR10 on pre-trained VGG-16 model. Furthermore, we discuss the relationship of error probability and conditional entropy. \begin{figure*}[!ht] \centering \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_1}% \caption{Maximum conditional entropy.}% \label{maxv}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_2}% \caption{Activation of max conditional entropy.}% \label{maxp}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_3}% \caption{Number of zero activations.}% \label{act_zero}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_5}% \caption{Average conditional entropy of filter.}% \label{ent}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_6}% \caption{Maximum conditional entropy filter ratio.}% \label{maxv_rate}% \end{subfigure}\hfill% \begin{subfigure}{.7\columnwidth} \includegraphics[width=\columnwidth]{Figure_7}% \caption{Zero activation ratio.}% \label{act_zero_rate}% \end{subfigure}% \caption{Statistical result of VGG-16 on CIFAR10} \label{fig:3_stat_vgg} \end{figure*} \subsection{Problem formulation} In~\cite{shwartz2017opening,michael2018on}, each layer is seen as a \emph{single} random variable. And the distribution is calculated by joining all neuron outputs in this layer. However the same method might not be suitable for convolutional network, for losing the structure feature on both the input and the feature map. Therefore we adopt the procedure to estimate the mutual information of each convolutional layer as in \cite{kraskov2004estimating,kolchinsky2017estimating}. The first step is to use the entropy of each layer's output as the measurement of the information flow. The \textit{activation entropy} can be calculated using the function below where $p_i$ denotes the probability of \textit{i}-th filter in the feature map. \begin{equation} H_{C_n} = \sum_{i=0}^{n} {p_i * \log p_i} \end{equation} Considering the general scenario of a neural network whose operation is parametrized by a vector~$\theta \in {\Re}^{W}$ (representing the weights), and whose input/output characteristics are described by a conditional probability distribution $p_{\theta }(x_{out} \mid x_{in})$, in which~$x_{in} \in {\Re}^{N}$ and~$x_{out} \in {\Re}^{M}$ denote input and output vectors, respectively. The performance of this network on a given input is measured some loss function~$\varepsilon(x_{in}, x_{out})$. If the probability of an input~$x_{in}$ to be encountered is defined as~$p(x_{in})$, the global error made by a network with parameter is given by \begin{equation} E_{\theta }={\sum_{}^{x_{in}}}{\sum_{}^{x_{out}}}\epsilon (x_{in}, x_{out})p_{\theta }(x_{out}\mid x_{in})p(x_{in}) \end{equation} We define $p_{\theta}(x) = p_{\theta}(x_{out}\mid x_{in})p(x_{in})$, $x = (x_{in}, x_{out}) \in {\Re}^{N+M}$. It now combines both the parametrized properties of the network and the likelihood of input data. When recognizing neural network as a stack of sub networks, the above definition holds true for each layer and thus error in convolutional layers: \begin{equation} E_{\theta }^{C}={\sum_{}^{C_{in}}}{\sum_{}^{x_{out}}}\epsilon (C_{in}, x_{out})p_{\theta }(x_{out}\mid C_{in})p(C_{in}) \end{equation} Where~$C_{in}$ denotes the input feature maps of the convolutional layer. In this way, we reorganize the compression problem as a optimization problem and minimize the distance~$\left \|E_{\theta}^{C} - E_{{\theta}'}^{{C}'}\right \|$, where~${C}'_{in}$ is a minimum subset of $C_{in}$. We adopt a d-dimensional binary vector $\sigma$: a 1 indicating the filter is selected, a 0 indicating the filter is discarded. Notation $x_{\sigma}$ indicates the vector of selected features, that is, the full vector $x$ projected onto the dimensions specified by $\sigma$. Notation $x_{\widetilde{\sigma}}$ is the complement, that is, the unselected features. The full feature vector can therefore be expressed as $x = \left \{ x_{\sigma}, x_{\widetilde{\sigma}} \right \}$. As mentioned, we assume the process $p$ is defined by a subset of the features, so for some unknown optimal vector $\sigma^{*}$, $p\left ( y \mid x \right )= p\left( y \mid x_{\sigma^{*}}\right)$. We approximate $p$ using an hypothetical predictive model $q$, with two layers of parameters: $\sigma$ representing which filters are selected and $\tau$ representing the parameters used to predict y. Our problem statement is to identify the minimal subset of features such that we maximize the conditional likelihood of the training labels, with respect to these parameters. For i.i.d data $D = \left \{ \left ( x^{i}, y^{i} \right );i = 1..N \right \}$ the conditional likelihood of the labels given parameters $\left \{ \sigma , \tau \right \}$ is \begin{equation} L\left(\sigma,\tau\mid D\right) = \prod_{i=1}^{N}q\left( y^{i}\mid x_{\sigma}^{i},\tau\right) \end{equation} The (scaled) conditional log-likelihood is \begin{equation} l=\frac{1}{N} \sum_{i=1}^{N}\textup{log}q\left ( y^{i}\mid x_{\sigma}^{i}, \tau \right ) \end{equation} This is the error function we wish to optimize with respect to the parameters $\left \{ \sigma , \tau \right \}$; the scaling term has no effect on the optima, but simplifies exposition later. We now introduce the quantity $p\left ( y \mid x_{\sigma} \right )$: this is the true distribution of the loss given the selected filters $x_{\sigma}$. Multiplying and dividing $q$ by $p\left ( y \mid x_{\sigma} \right )$, we can re-write the above as, \begin{equation} l=\frac{1}{N} \sum_{i=1}^{N}\log\frac{q\left ( y^{i}\mid x_{\sigma}^{i}, \tau \right )}{p\left ( y^{i}\mid x_{\sigma}^{i}\right )}+\frac{1}{N} \sum_{i=1}^{N}\log{p\left ( y^{i}\mid x_{\sigma}^{i}\right )} \end{equation} The second term in (8) can be similarly expanded, introducing the probability $p\left ( y \mid x \right )$. These are finite sample approximations, drawing data points i.i.d. with respect to the distribution. We use $E_{xy}\left \{ . \right \}$ to denote statistical expectation, and for convenience we negate the above, turning our maximization problem into a minimization. This gives us \begin{equation} \begin{split} -l & \approx E_{xy}\left \{ \log\frac{p\left ( y\mid x_{\sigma}\right )}{q\left ( y\mid x_{\sigma}, \tau \right )} \right \} + E_{xy}\left \{ \log\frac{p\left ( y\mid x\right )}{p\left ( y\mid x_{\sigma} \right )} \right \} \\ & - E_{xy}\left \{ \log{p\left ( y\mid x\right )} \right \} \end{split} \end{equation} In the experiments below, we use the training loss as the single variable. Then our problem statement is to identify the minimal subset of features such that we~\textit{maximize the conditional likelihood of the training loss, with respect to these parameters}. \subsection{Filter selection algorithm} The algorithm is illustrated below as shown in Algorithm~\ref{algo:3_procedure}. For each sample, we calculate the cross entropy loss and output activation corresponding to each filter. To achieve discrete statistical requirement, each parameter is multiplied by a factor of $1e4$ and quantized as 32-bit integer. For each filter, $c\_val$ denotes 1-D distribution on output activation and $c\_bins$ denote a 2-D statistics on output activation conditioned on loss. $c\_total$ is a number of activations per filter, i.e. samples in a dataset. $act\_ent$ denotes the entropy of feature map, provided the distribution of output activation across the dataset. Notice here zero activations are excluded because it's considered to contain no information with respect to the next layer. Given the probability of a specific output activation, $ent_i$ denotes the entropy of output activation conditioned on the distribution of cross entropy loss. $con\_ent$ denotes the conditional entropy of a filter, which is an accumulation of $ent_i$. Then, the $con\_ent$ is sorted in ascending and the filters corresponding to the top-$r$ $con\_ent$ are selected to be removed. In above single-layer illustration, the layer to prune is predefined. This can be generalized to multiple layers or the whole model. \begin{algorithm} \caption{Filter selection algorithm.} \label{algo:3_procedure} \hspace*{\algorithmicindent} \\\textbf{Input}: a baseline model $\boldsymbol{M}$, convolutional layer to prune $l$, training dataset $\textbf{x}_{train}$, number of filters to prune $r$ \hspace*{\algorithmicindent} \\\textbf{Output}: candidate filter(s) to prune $\sigma^r$ \begin{algorithmic} \Procedure{ConditionalEntropy\_Calculation}{} \State $ eps\_h = 1e4 $ \State $\textit{criteria} = CrossEntropy(reduce=false)$. \For {$batch\_idx$ in $batches$} \State $output \gets \boldsymbol{M}(\textbf{x}_{train}[batch\_idx])$. \State $loss[batch\_size] \gets \textit{criteria}(output, target)$. \State $j=eps\_h * loss[batch\_size]$. \For {$k$ in $l$} \State $i=eps\_h * fmap\_out^{l}[k]$. \State $ans[k].extend(i, j)$. \State $c\_bins[i][j] += 1 $. \State $c\_val[i] += 1 $. \EndFor \EndFor \State $c\_total = \sum\limits_{i}^{c\_val}{c\_val[i]}$. \State $act\_ent = \sum\limits_{j \neq 0}^{c\_val}{-\frac{c\_val[j]}{c\_total}*\log\frac{c\_val[j]}{c\_total}}$ \State $ent_i = \frac{c\_val[i]}{c\_total}\sum\limits_{j}{-\frac{c\_bins[i][j]}{c\_val[i]}*\log\frac{c\_bins[i][j]}{c\_val[i]}}$ \State $con\_ent = \sum\limits_{i}{ent_i}$ \State sort $con\_ent$ of each filter in layer $l$ ascending \State add corresponding filter of top-$r$ $con\_ent \to \sigma^r$ \EndProcedure \end{algorithmic} \end{algorithm} \subsection{Statistical result of CIFAR10 on VGG-16} As one of the core concepts in the convolutional network, feature maps reveals huge amount of information about the information flow within the network. For which we proposed using statistical variable, especially conditional entropy, to measure the connection between feature map and error probability, which may help us understand how the information flows through the network. Figure~\ref{fig:3_stat_vgg} is the statistical result of filters in each convolutional layer of CIFAR10 on VGG-16. The x-axis indicate the index of convolution layer in model. Figure~\ref{maxv}, Figure~\ref{maxp} shows the maximum $ent_i$ and its corresponding output activation, respectively. Figure~\ref{act_zero} depicts the number of zero activations and Figure~\ref{act_zero_rate} shows the negative zero activation ratio. We observe zero activation mostly reside in first and last several convolutional layers. Figure~\ref{ent} demonstrate total conditional entropy of each filter. Figure~\ref{maxv_rate} is the ratio of maximum conditional entropy to total conditional entropy and it's remains relatively uniform in different layers. \begin{figure}[!ht] \centering \begin{subfigure}{1.0\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_1}% \caption{First convolutional layer entropy.}% \label{mnist_1}% \end{subfigure}\hfill% \begin{subfigure}{1.0\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_2}% \caption{Second convolutional layer entropy.}% \label{mnist_2}% \end{subfigure}\hfill% \begin{subfigure}{0.8\columnwidth} \includegraphics[width=\columnwidth]{4_mnist_sample}% \caption{Sample under different noise level.}% \label{mnist_noise}% \end{subfigure} \caption{Two layer CNN on MNIST} \label{mnist} \end{figure} \section{Experiments} \label{sec:evaluation} \subsection{Experiment setup} In the initial stage of our experiment, we assume using entropy $H(T_n)$ to measure information flowed in the neural network. To testify our assumption, we made several experiments with MNIST datasets and a two-layered CNN. We added Gaussian noise to the picture and see how the entropy in neural network would change as the power of noise changes. If the entropy effectively measures the information uncertainty in CNN, $H_{T_1}$ and $H_{T_2}$ should increase as the loss increases. The result is affirmative, $H_{T_1}$ and $H_{T_2}$ do increase as expected. Yet there are still some problems about $H_{T_2}$. As shown in Figure~\ref{mnist}, $H_{T_2}$ does not increase as much as $H_{T_1}$ when var is below 0.1. We assume that CNN reduces the influence of noise in the classification task when the noise is not so significant. So the entropy does not increase so much in the second layer. Here, we trained a VGG-16~\cite{simonyan2014very} network with \emph{tanh} activation function on CIFAR10 dataset~\cite{krizhevsky2009learning} and logged the output of each layers. The network has 13 convolution layers, each layer has 64 filters to 512 filters, with the size of 3. The network converged at around 100 epochs, but we trained it for 400 epochs. The final accuracy is $99.6\%$ on training set and $92.68\%$ on test set. As we can observe from Figure~\ref{ent}, conditional entropy in filters are almost uniformly distributed, except for the first convolutional layer where conditional entropy is slightly higher. Also, number of zero activations shown in Figure~\ref{act_zero} indicates the same fact that the first convolutional layer in VGG-16 is greatly redundant in CIFAR10 classification. However, the philosophy that each neuron in network is interchangeable and informative equivalent~\cite{williamson2015equivalence} indicate a uniform distribution of information across the layers. As a result, filters are supposed to be pruned in a layer-wise manner where the number of filters to be pruned should be proportional in each layer, e.g. prune 16 filters in convolutional layer with 64 filters, 32 filters with 128 etc. In the experiment, we evaluate trade-off between accuracy and pruning ratio in global and layer-wise approach respectively. Inspired from above observation, we propose a two phase filter pruning framework based on conditional entropy, namely \textit{2PFPCE}. The filter selection criteria is described in Algorithm~\ref{algo:3_procedure}. The procedure is shown in Figure~\ref{fig:4_exp_method}. In phase I, filters are pruned and fine-tuned iteratively until the accuracy drop reach the threshold (1\%). The purpose is to remove the redundancy filters with respect to the dataset. Notice that dataset plays a crucial role when it comes to compression. The number of involved features in a 1000-category dataset is probably much larger than that in a 10-category dataset. After a layer is pruned, weights in the filter w.r.t the maximum activation entropy are kept constant and cannot be updated during fine tuning. Similar to dropout~\cite{srivastava2014dropout}, this aims at penalizing any single neuron that may overly fitted to the dataset. In phase II, filters are pruned in a layer-wise manner: in each iteration, a small portion filters (1/32 or 1/16) of each layer are pruned until the accuracy drop reach the threshold $\gamma$. The threshold $\gamma$ is a hyper-parameter and can be adjusted to satisfy application constraint. To retain the information in pruned filters and avoid time consuming fine-tune, we update the bias term as the activation of maximum conditional entropy filter (Figure~\ref{maxp}). \subsection{Global Pruning Approach} \label{sec:exp_glob} In this experiment, we evaluate the trade-offs between computation complexity and classification accuracy in global pruning approach. As shown in Figure~\ref{fig:4_glob_vgg} and Figure~\ref{fig:4_glob_resnet}, the accuracy w.r.t pruning ratio is similar in both networks. Without fine-tuning, the accuracy in VGG-16 and ResNet-18 decrease significantly to 27\% and 16\% within a pruning ratio of 10\% and 21\%, respectively. After the turning point, in VGG-16 the accuracy slowly decrease to 15\% as the pruning ratio increase while in ResNet-18 the accuracy drop is within 1\%. With fine-tuning, each network can remain its accuracy with pruning ratio below 20\% and then slowly degrade with a marginal accuracy drop between 1\%-2\% until the pruning ratio reaches 90\%. The comparison between w/o fine-tune shows the DNNs can recover from small disturb with fine tuning. The contribution of each filter is very similar in a converged convolutional neural network. Though the experiment results on VGG-16 and ResNet-18, we can conclude that our current solution can prune about 40\% filters with 1\% accuracy compromise and 80\% prunes with only 2\% accuracy compromise with global pruning approach. \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth, trim={3cm 1.6cm 3cm 2.5cm},clip]{figure/4_glob_vgg} \caption{Accuracy vs. pruning ratio of VGG16.} \label{fig:4_glob_vgg} \end{figure} \begin{figure}[t] \centering \includegraphics[width=0.98\columnwidth, trim={3cm 1.5cm 2.7cm 2.8cm},clip]{figure/4_glob_resnet} \caption{Accuracy vs. pruning ratio of Preact-ResNet18.} \vspace{-6pt} \label{fig:4_glob_resnet} \end{figure} \subsection{Layer-wise Pruning Approach} \label{sec:exp_layer} Table~\ref{table:res_cifar10_1} compares the obtained accuracy of different filter importance criterion under the same pruning ratios. For a fair comparison, the DNNs are pruned in a layer-wise approach and fine-tuned same number of epochs for accuracy recover after pruning. At each pruning iteration, we remove a percentage of feature maps and then perform 20 minibatch SGD updates with batch-size 32, momentum 0.9, learning rate $10^{-4}$, and weight decay $10^{-4}$. As we can see from the table, our proposed conditional entropy filter selection criteria outperforms counterparts with highest accuracy on both pruning ratio 25\% and 50\%. Table~\ref{table:res_cifar10_2} compares the obtained pruning ratio of filter importance criterion under approximately the same errors, e.g., within 2\% accuracy loss. The baseline is L1-Norm which has a pruning ratio of 60\% on both networks. Our experiments show that with a marginal accuracy loss, conditional entropy based criterion can always achieve largest pruning ratio. \begin{table} \centering \begin{subtable}{0.5\textwidth} \centering \caption{Filter importance criteria vs. accuracy.} \label{table:res_cifar10_1} \vspace{4pt} \begin{tabular}{ |p{1.0cm}|p{1.4cm}|p{1.4cm}|p{1.4cm}|p{1.4cm}| } \hline \multicolumn{5}{|c|}{VGG16/CIFAR10: 92.98\%} \\ \hline Prune & L1-Norm & APoZ & Act. ent & Cond. ent\\ \hline 25\% & 92.86\% & 92.94\% & 92.93\% & 93.50\% \\ 50\% & 92.11\% & 92.02\% & 92.00\% & 92.76\% \\ \hline \multicolumn{5}{|c|}{ResNet50/CIFAR10: 93.16\% } \\ \hline 20\% & 94.42\% & 94.36\% & 94.33\% & 94.84\% \\ 50\% & 94.48\% & 94.25\% & 94.42\% & 94.44\% \\ \hline \end{tabular} \end{subtable} \begin{subtable}{0.5\textwidth} \centering \vspace{6pt} \caption{Filter importance criteria vs. pruning ratio.} \label{table:res_cifar10_2} \vspace{4pt} \begin{tabular}{ |p{1.6cm}|p{1.35cm}|p{1.1cm}|p{1.35cm}|p{1.35cm}| } \hline \multicolumn{5}{|c|}{VGG16/CIFAR10: 92.98\%} \\ \hline Acc. & L1-Norm & APoZ & Act. ent & Cond. ent\\ \hline 91.0($\pm$0.3)\% & $1.0\times$ & $0.88\times$ & $1.27\times$ & $1.32\times$ \\ \hline \multicolumn{5}{|c|}{ResNet50/CIFAR10: 93.16\% } \\ \hline 92.0($\pm$0.2)\% & $1.0\times$ & $0.93\times$ & $0.96\times$ & $1.05\times$ \\ \hline \end{tabular} \end{subtable} \caption{Comparison of filter importance criteria in layer-wise approach VGG-16/ResNet-50 on CIFAR10} \label{4_layer_cmp} \end{table} \subsection{Stage Pruning Approach} \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth]{figure/4_stage_all} \caption{Stage pruning with 98\% accuracy threshold.} \label{fig:4_stage_all} \end{figure} We focus on reducing the number of convolutional feature maps and the total estimated floating point operations (FLOPs). During pruning we were measuring reduction in computations by FLOPs, which is a common practice~\cite{han2015deep}. Improvements in FLOPs result in monotonically decreasing inference time of the networks because of removing entire feature map from the layer. However, time consumed by inference dependents on particular implementation of convolution operator, parallelization algorithm, hardware, scheduling, memory transfer rate etc. Unlike in previous work, where latency is considered via another, often inaccurate proxy (e.g., FLOPS), in our experiments, we directly measure real-world inference time by executing the model on GPU (NVIDIA Titan Xp with CUDNN 8.0). Therefore we measure improvement in the inference time to see real speed up compared to original networks in Figure~\ref{fig:4_stage_all}. Based on the observations mentioned in setup, a two-phase pruning approach is proposed. As shown in Figure~\ref{fig:4_stage_all}, a combination of both approach can achieve a pruning ratio of 88\% within 2\% accuracy drop. The dataset is CIFAR10 with a minibatch size of 32. The inference time on pre-trained VGG-16 is 22.69ms. In Phase I, number of filters decrease from 4224 to 3960 and total bandwidth from 310784 bytes to 273658 bytes. The inference time reduces to 19.24ms. Removed filters are mostly from the first convolutional layer and this verifies our observation on redundancy. In the first convolutional layer, $\sim 60\%$ filters are removed result in a reduction in data volume from 64M to 35M. Because of high parallelization in GPU, inference time decrease in a single layer is not as significant as in data volume. In Phase II, number of filters decrease from 3960 to 532 and total bandwidth from 273658 bytes to 49165 bytes. The inference time reduces to 12.34ms. \section{Conclusion} \label{sec:conclusion} In this work, inspired from the statistical result of information flow in neural network, we propose to use conditional entropy as the filter selection criteria in filter pruning. The performance of conditional based filter selection criteria outperforms approaches based L1-Norm, APoZ and activation entropy. Experimental result shows the proposed criteria can achieve 92.76\% accuracy when pruning ratio is 50\% and $\sim 91\%$ accuracy when pruning ratio is 80\%. In both VGG-16 and ResNet-50, our proposed conditional entropy outperforms the above criterion. To comply with the network information distribution, we adopt a two phase pruning framework which combines global approach with layer-wise approach. In addition, novel model tuning techniques are proposed 1) freeze weights w.r.t maximum activation entropy to avoid over-fitting, 2) update bias w.r.t the activation of the maximum conditional entropy filter. The above framework can achieve a pruning ratio of 88\% within 2\% accuracy drop of pre-trained VGG-16 model on CIFAR10. \bibliographystyle{plain}

66b7587d2da4ec6c4af3cf75554997a44f18444b

\section{Introduction} \noindent Let $\bar{\mathbb{R}} = (\mathbb{R},<,+,\cdot,0,1)$ be the real field. For $\lambda \in \mathbb{R}_{>0}$, set $\lambda^{\mathbb{Z}} := \{ \lambda^m : m \in \mathbb{Z}\}$. Throughout this paper $\Gamma$ denotes a discrete subgroup of Gl$_{n}(\mathbb{C})$, and $G$ denotes a subgroup of Gl$_n(\mathbb{C})$. We identify the set M$_n(\mathbb{C})$ of $n$-by-$n$ complex matrices with $\mathbb{C}^{n^2}$ and identify $\mathbb{C}$ with $\mathbb{R}^2$ in the usual way. Our main result is the following classfication of expansions of $\bar{\mathbb{R}}$ by a discrete subgroup of $\Gl_n(\mathbb{C})$. \begin{thmA}\label{thm:main} Let $\Gamma$ be an infinite discrete subgroup of $\Gl_n(\mathbb{C})$. Then either \begin{itemize} \item $(\mathbb{R},\Gamma)$ defines $\mathbb{Z}$ or \item there is $\lambda \in \mathbb{R}_{>0}$ such that $(\bar{\mathbb{R}},\Gamma)$ is interdefinable with $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$. \end{itemize} If $\Gamma$ is not virtually abelian, then $(\bar{\mathbb{R}},\Gamma)$ defines $\mathbb{Z}$. \end{thmA} \noindent By Hieronymi~\cite[Theorem 1.3]{discrete}, the structure $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}},\mu^{\mathbb{Z}})$ defines $\mathbb{Z}$ whenever $\log_{\lambda} \mu \notin \mathbb{Q}$, and is interdefinable with $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ otherwise. Therefore Theorem A extends immediately to expansions of $\bar{\mathbb{R}}$ by multiple discrete subgroups of $\Gl_n(\mathbb{C})$. \begin{corA} Let $\Cal G$ be a collection of infinite discrete subgroups of various $\Gl_n(\mathbb{C})$. Then either \begin{itemize} \item $(\bar{\mathbb{R}},\big(\Gamma\big)_{\Gamma \in \Cal G})$ defines $\mathbb{Z}$ or \item there is $\lambda \in \mathbb{R}_{>0}$ such that $(\bar{\mathbb{R}},\big(\Gamma\big)_{\Gamma \in \Cal G})$ is interdefinable with $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$. \end{itemize} \end{corA} \noindent The dichotomies in Theorem A and Corollary A are arguably as strong as they can be. An expansion of the real field that defines $\mathbb{Z}$, has not only an undecidable theory, but also defines every real projective set in sense of descriptive set theory (see Kechris \cite[37.6]{kechris}). From a model-theoretic/geometric point of view such a structure is a wild as can be. On the other hand, by van den Dries \cite{vdd-Powers2} the structure $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ has a decidable theory whenever $\lambda$ is recursive, and admits quantifier-elimination in a suitably extended language. It satisfies combinatorical model-theoretic tameness conditions such as NIP and distality (see \cite{GH-Dependent, HN-distal}). Furthermore, it follows from these results that every subset of $\mathbb{R}^n$ definable in $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ is a boolean combination of open sets, and thus $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ defines only sets on the lowest level of the Borel hierarchy. See Miller \cite{Miller-tame} for more on tameness in expansions of the real field. \newline \noindent Our proof of Theorem A relies crucially on the following two criteria for the definability of $\mathbb{Z}$ in expansions of the real field. \begin{fact}\label{thm:hier} Suppose $D \subseteq \mathbb{R}^k$ is discrete. \begin{enumerate} \item If $(\bar{\mathbb{R}},D)$ defines a subset of $\mathbb{R}$ that is dense and co-dense in a nonempty open interval, then $(\bar{\mathbb{R}},D)$ defines $\mathbb{Z}$. \item If $D$ has positive Assouad dimension, then $(\bar{\mathbb{R}},D)$ defines $\mathbb{Z}$. \end{enumerate} \end{fact} \noindent The first statement is \cite[Theorem E]{discrete2}, a fundamental theorem on first-order expansions of $\bar{\mathbb{R}}$, and the second claim is proven using the first in Hieronymi and Miller \cite[Theorem A]{HM}. We recall the definition of Assouad dimension in Section \ref{section:exp}. This important metric dimension bounds more familiar metric dimensions (such as Hausdorff and Minkowski dimension) from above. We refer to \cite{HM} for a more detailed discussion of Assouad dimension and its relevance to definability theory.\newline \noindent The outline of our proof of Theorem A is as follows. Let $\Gamma$ be a discrete, infinite subgroup of $\Gl_n(\mathbb{C})$. Using Fact \ref{thm:hier}(1), we first show that $(\bar{\mathbb{R}},\Gamma)$ defines $\mathbb{Z}$ whenever $\Gamma$ contains a non-diagonalizable matrix. It follows from a theorem of Mal\textquotesingle tsev that $(\bar{\mathbb{R}},\Gamma)$ defines $\mathbb{Z}$ when $\Gamma$ is virtually solvable and not virtually abelian. In the case that $\Gamma$ is not virtually solvable, we prove using Tits' alternative that $\Gamma$ has positive Assouad dimension, and hence $(\bar{\mathbb{R}},\Gamma)$ defines $\mathbb{Z}$ by Fact \ref{thm:hier}(2). We conclude the proof of Theorem A by proving that whenever $\Gamma$ is virtually abelian and $(\bar{\mathbb{R}},\Gamma)$ does not define $\mathbb{Z}$, then $(\bar{\mathbb{R}},\Gamma)$ is interdefinable with $(\bar{\mathbb{R}},\lambda^\mathbb{Z})$ for some $\lambda \in \mathbb{R}_{>0}$. Along the way we give (Lemma~\ref{prop:heisen0}) an elementary proof showing that a torsion free non abelian nilpotent subgroup of Gl$_n(\mathbb{C})$ has a non-diagonalizable element. As every finitely generated subgroup of Gl$_n(\mathbb{C})$ is either virtually nilpotent or has exponential growth, this yields a more direct proof of Theorem A in the case when $\Gamma$ is finitely generated. \newline \noindent We want to make an extra comment about the case when $\Gamma$ is a discrete, virtually solvable, and not virtually abelian subgroup of $\Gl_n(\mathbb{C})$. The \textbf{Novosibirsk theorem} \cite{Noskov} of Noskov (following work of Mal\textquotesingle stev, Ershov, and Romanovskii) shows that a finitely generated, virtually solvable and non-virtually abelian group interprets $(\mathbb{Z},+,\cdot)$. It trivially follows that if $G$ is finitely generated, virtually solvable, and non-virtually abelian, then $(\bar{\mathbb{R}},G)$ interprets $(\mathbb{Z},+,\cdot)$. However, it does not directly follow that $(\bar{\mathbb{R}},G)$ defines $\mathbb{Z}$. We use an entirely different method below to show that if $G$ is in addition discrete, then $(\bar{\mathbb{R}},G)$ defines $\mathbb{Z}$. Our method also applies when $G$ is not finitely generated, but relies crucially on the discreteness of $G$.\newline \noindent This paper is by no means the first paper to study expansions of the real field by subgroups of $\Gl_n(\mathbb{C})$. Indeed, there is a large body of work on this subject, often not explicitly mentioning $\Gl_n(\mathbb{C})$. Because we see this paper as part of a larger investigation, we survey some of the earlier results and state a conjecture. It is convenient to consider three disctinct classes of such expansion. By Miller and Speissegger~\cite{MS99} every first-order expansion $\mathscr{R}$ of $\bar{\mathbb{R}}$ satisfies at least one of the following: \begin{enumerate} \item $\mathscr{R}$ is o-minimal, \item $\mathscr{R}$ defines an infinite discrete subset of $\mathbb{R}$, \item $\mathscr{R}$ defines a dense and co-dense subset of $\mathbb{R}$. \end{enumerate} The \textbf{open core} $\mathscr{R}^\circ$ of $\mathscr{R}$ is the expansion of $(\mathbb{R},<)$ generated by all open $\mathscr{R}$-definable subsets of all $\mathbb{R}^k$. By \cite{MS99}, if $\mathscr{R}$ does not satisfy (2), then $\mathscr{R}^\circ$ is o-minimal.\newline The case when $\mathscr{R}$ is o-minimal, is largely understood. Wilkie's famous theorem \cite{wilkie} that $(\bar{\mathbb{R}},\exp)$ is o-minimal is crucial. This shows the expansion of $\bar{\mathbb{R}}$ by the subgroup $$ \left\{ \begin{pmatrix} 1 & 0 & t\\[3pt] 0 & \lambda^t & 0\\[3pt] 0 & 0 & 1 \end{pmatrix} : t \in \mathbb{R} \right\} $$ is o-minimal for $\lambda \in \mathbb{R}_{>0}$, and so is the expansion of $\bar{\mathbb{R}}$ by any subgroup of the form $$ \left\{ \begin{pmatrix} t^s & 0 \\[3pt] 0 & t^r \end{pmatrix} : t \in \mathbb{R}_{>0} \right\} $$ for $s,r \in \mathbb{R}_{>0}$. Indeed, by Peterzil, Pillary, and Starchenko~\cite{PPS-linear}, whenever an expansion $(\bar{\mathbb{R}},G)$ by a subgroup $G$ of $\Gl_n(\mathbb{R})$ is o-minimal, then $G$ is already definable in $(\bar{\mathbb{R}},\exp)$. Futhermore, note that by a classical theorem of Tannaka and Chevalley~\cite{Chevalley} every compact subgroup of Gl$_n(\mathbb{C})$ is the group of real points on an algebraic group defined over $\mathbb{R}$. Thus every compact subgroup of $\Gl_n(\mathbb{C})$ is $\bar{\mathbb{R}}$-definable, and therefore the case of expansions by compact subgroups of $\Gl_n(\mathbb{C})$ is understood as well.\newline We now consider the case when infinite discrete sets are definable. Corollary A for discrete subgroups of $\mathbb{C}^{\times}$ follows easily from the proof of \cite[Theorem 1.6]{discrete}. While Corollary A handles the case of expansions by discrete subgroups of $\Gl_n(\mathbb{C})$, there are examples of subgroups of $\Gl_n(\mathbb{C})$ that define infinite discrete sets, but fail the conclusion of Theorem A. Given $\alpha \in \mathbb{R}^\times$ the logarithmic spiral $$ S_\alpha = \{ (\exp( t) \sin(\alpha t), \exp(t) \cos(\alpha t) ) : t \in \mathbb{R} \} $$ is a subgroup of $\mathbb{C}^\times$. Let $\mathfrak{s}$ and $\mathfrak{e}$ be the restrictions of $\sin$ and $\exp$ to $[0,2\pi]$, respectively. Then $(\bar{\mathbb{R}},S_\alpha)$ is a reduct of $( \bar{\mathbb{R}}, \mathfrak{s}, \mathfrak{e}, \lambda^{\mathbb{Z}})$ when $\lambda = \exp(2\pi\alpha)$, as was first observed by Miller and Speissegger. As $(\bar{\mathbb{R}},\mathfrak{s},\mathfrak{u})$ is o-minimal with field of exponents $\mathbb{Q}$, the structure $(\bar{\mathbb{R}},S_\alpha)$ is d-minimal\footnote{A expansion $\mathscr{R}$ of $\bar{\mathbb{R}}$ is \textbf{d-minimal} if every definable unary set in every model of the theory of $\mathscr{R}$ is a union of an open set and finitely many discrete sets.} by Miller \cite[Theorem 3.4.2]{Miller-tame} and thus does not define $\mathbb{Z}$. It can be checked that $(\bar{\mathbb{R}},S_\alpha)$ defines a analytic function that is not semi-algebraic\footnote{By induction on the complexity of terms it follows easily from [Theorem II, vdD] that the definable functions in $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ are given piecewise by a finite compositions of $x\mapsto \max \Big(\{0\} \cup \big(\lambda^{\mathbb{Z}} \cap [-\infty,x]\big)\Big)$ and functions definable in $\bar{\mathbb{R}}$. From this one can deduce that every definable function in this structure is piecewise semi-algebraic.}, and thus is not interdefinable with $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ for any $\lambda \in \mathbb{R}_{>0}$. \newline Most work in the case of expansions that define dense and co-dense sets, concerns expansions by finite rank subgroups of $\mathbb{C}^\times$ (see introduction of \cite{Erin-thesis} for a thorough discussion of expansions by subgroups of $\mathbb{C}^\times$). In \cite{vdDG-groups} van den Dries and G\"{u}nayd\i n showed that an expansion of $\bar{\mathbb{R}}$ by a finitely generated dense subgroup of $(\mathbb{R}_{>0},\cdot)$ admits quantifier-elimination in a suitably extend language. G\"{u}nayd\i n~\cite{Ayhan-thesis} and Belegradek and Zilber~\cite{Belegradek-Zilber} proved similar results for the expansion of $\bar{\mathbb{R}}$ by a dense finite rank subgroup of the unit circle $\mathbb{U} := \{ a \in \mathbb{C}^\times : |a| = 1\}$. This covers the case when $G$ is the group of roots of unity. In all these cases the open core of the resulting expansion is interdefinable with $\bar{\mathbb{R}}$. This does not always have to be the case. In Caulfield~\cite{Erin-paper} studies expansions by subgroups of $\mathbb{C}^\times$ of the form $$\{ \lambda^{k} \exp(i\alpha l) : k,l \in \mathbb{Z}\} \quad \text{where } \lambda \in \mathbb{R}_{>0} \text{ and } \alpha \in \mathbb{R} \setminus \pi \mathbb{Q}.$$ Such an expansions obviously defines a dense and co-dense subset of $\mathbb{R}$, but by \cite{Erin-paper} its open core is interdefinable with $(\bar{\mathbb{R}},\lambda^\mathbb{Z})$. Futhermore, even if the open core is o-minimal, it does not have to be interdefinable with $\bar{\mathbb{R}}$. By \cite{H-tau} there is a co-countable subset $\Lambda$ of $\mathbb{R}_{>0}$ such that if $r \in \Lambda$ and $H$ is a finitely generated dense subgroup of $(\mathbb{R}_{>0},\cdot)$ contained in the algebraic closure of $\mathbb{Q}(r)$, then the open core of the expansion of $\bar{\mathbb{R}}$ by the subgroup $$ \left\{ \begin{pmatrix} t & 0 \\[3pt] 0 & t^r \end{pmatrix} : t \in H \right\} $$ is interdefinable with the expansion of $\bar{\mathbb{R}}$ by the power function $t \mapsto t^r : \mathbb{R}_{>0} \to \mathbb{R}_{>0}$.\newline All these previous results suggest that the next class of subgroups of $\Gl_n(\mathbb{C})$ for which we can hope to prove a classification comparable to Theorem A, is the class of finitely generated subgroups. Here the following conjecture seems natural, but most likely very hard to prove. Let $\bar{\mathbb{R}}_{\text{Pow}}$ be the expansion of $\bar{\mathbb{R}}$ by all power functions $\mathbb{R}_{>0} \to \mathbb{R}_{>0}$ of the form $t \mapsto t^r$ for $r \in \mathbb{R}^\times$. \begin{conj} Let $G$ be a finitely generated subgroup of $\Gl_n(\mathbb{C})$ such that $(\bar{\mathbb{R}},G)$ does not define $\mathbb{Z}$. Then the open core of $(\bar{\mathbb{R}},G)$ is a reduct of $\bar{\mathbb{R}}_{\textrm{Pow}}$ or of $(\bar{\mathbb{R}},S_\alpha)$ for some $\alpha \in \mathbb{R}_{>0}$. \end{conj} Even when the statement ``$(\bar{\mathbb{R}},G)$ does not define $\mathbb{Z}$'' is replaced by ``$(\bar{\mathbb{R}},G)$ does not interpret $(\mathbb{Z},+,\cdot)$'', the conjecture is open. However, this weaker conjecture might be easier to prove, because the Novosibirsk theorem can be used to rule out the case when $G$ is virtually solvable and non-virtually abelian. It is worth pointing out that Caulfield conjectured that when $G$ is assumed to be a subgroup of $\mathbb{C}^{\times}$, then the open core $(\bar{\mathbb{R}},G)$ is either $\bar{\mathbb{R}}$ or a reduct of $(\bar{\mathbb{R}},S_\alpha)$ for some $\alpha \in \mathbb{R}_{>0}$. See \cite{Erin-paper, Erin-thesis} for progress towards this later conjecture. \section{Notation and Conventions} \noindent Throughout $m,n$ range over $\mathbb{N}$ and $k,l$ range over $\mathbb{Z}$, $G$ is a subgroup of Gl$_n(\mathbb{C})$, and $\Gamma$ is a discrete subgroup of Gl$_n(\mathbb{C})$. Let $\bar{\mathbb{R}}_\Gamma$ be the expansion of $\bar{\mathbb{R}}$ by a $(2n)^2$-ary predicate defining $\Gamma$. We set $\bar{\mathbb{R}}_\lambda := \bar{\mathbb{R}}_{\lambda^{\mathbb{Z}}}$. A subset of $\mathbb{R}^k$ is \textbf{discrete} if every point is isolated. We let UT$_n(\mathbb{C})$ be the group of $n$-by-$n$ upper triangular matrices, D$_n(\mathbb{C})$ be the group of $n$-by-$n$ diagonal matrices, and $\mathbb{U}$ be the multiplicative group of complex numbers with norm one.\newline All structures considered are first-order, ``definable'' means ``definable, possibly with parameters''. Two expansions of $(\mathbb{R},<)$ are \textbf{interdefinable} if they define the same subsets of $\mathbb{R}^k$ for all $k$. If P is a propety of groups then a group $H$ is \textbf{virtually P} if there is finite index subgroup $H'$ of $H$ that is P. \section{Linear Groups} \noindent We gather some general facts on groups. Throughout this section $H$ is a finitely generated group with a symmetric set $S$ of generators. Let $S_m$ be the set of $m$-fold products of elements of $S$ for all $m$. If $S'$ is another symmetric set of generators then there is a constant $k \geq 1$ such that $$ k^{-1} |S_m| \leq |S'_m| \leq k |S_m| \quad \text{for all } m.$$ Thus the growth rate of $m \mapsto | S_m |$ is an invariant of $H$. We say $H$ has \textbf{exponential growth} if there is a $C \geq 1$ such that $|S_m| \geq C^m$ for all $m$ and $H$ has \textbf{polynomial growth} there are $k, t \in \mathbb{R}_{>0}$ such that $|S_m| \leq t m^k$ for all $m$. Note finitely generated non-abelian free groups are of exponential growth. Gromov's theorem~\cite{Gromov-poly} says $H$ has polynomial growth if and only if it is virtually nilpotent. Gromov's theorem for subgroups of Gl$_n(\mathbb{C})$ is less difficult and may be proven using the following two theorems: \begin{fact}\label{thm:tits} If $G$ does not contain a non-abelian free subgroup, then $G$ is virtually solvable. \end{fact} \noindent Fact~\ref{thm:tits} is \textbf{Tits' alternative}~\cite{tits}. Fact~\ref{thm:bmw} is due to Milnor~\cite{Milnor-solv} and Wolf~\cite{Wolf-solv}. \begin{fact}\label{thm:bmw} Suppose $H$ is virtually solvable. Then $H$ either has exponential or polynomial growth. If the latter case holds then $H$ is virtually nilpotent. \end{fact} \noindent Note Fact \ref{thm:tits} and Fact \ref{thm:bmw} imply every finitely generated subgroup of Gl$_n(\mathbb{C})$ is of polynomial or exponential growth. This dichotomy famously does not hold for finitely generated groups in general, see for example~\cite{Grigorchuk}. \medskip \noindent The \textbf{Heisenberg group} $\mathbb{H}$ is presented by generators $a,b,c$ and relations $$ [a,b] = c, \quad ac = ca, \quad bc = cb. $$ The following fact is folklore; we include a proof for the reader. \begin{fact}\label{fact:heisen} Let $E$ be a nilpotent, torsion-free, and non-abelian group. Then there is a subgroup of $E$ isomorphic to $\mathbb{H}$. \end{fact} \begin{proof} Let $e$ be the identity element of $E$. We define the lower central series $(E_k)_{k \in \mathbb{N}}$ of $E$ by declaring $E_0 = E$ and $E_{k} = [E_{k-1},E]$ for $k \geq 1$. Nilpotency means there is an $m$ such that $E_m \neq \{e\}$ and $[E_m,E] = \{e\}$. Moreover $m \geq 1$ as $E$ is not abelian. On one hand, $[E_{m-1},E] = E_m \neq \{e\}$ and so $E_{m-1}$ is not contained in $Z(E)$. Thus, there exists $a \in E_{m-1} \setminus Z(E)$ and $b \in E_m$ that does not commute with $a$. On the other hand, $[E_m,E]=\{e\}$ implies $E_m$ is contained in the center $Z(E)$ of $E$ and is thus abelian. So, $c := [a,b]$ is an element of $Z(E)$ and commutes with both $a$ and $b$. Finally, $a,b,c$ have infinite order because $E$ is torsion-free. So, $a,b,c$ generate a subgroup of $E$ isomorphic to the Heisenberg group. \end{proof} \subsection{Non-diagonalizable elements} We show certain linear groups necessarily contain non-diagonalizable elements. \begin{lem}\label{prop:heisen0} If $G$ is nilpotent, torsion-free, and not abelian, then $G$ contains a non-diagonalizable element. \end{lem} \noindent Lemma~\ref{prop:heisen0} follows from Fact~\ref{fact:heisen} above and Lemma~\ref{lem:heisen1} below. \begin{lem}\label{lem:heisen1} Suppose $a,b,c \in \text{Gl}_n(\mathbb{C})$ satisfy $$ [a,b] = c, \quad ac = ca, \quad bc =cb, $$ and $c$ is not torsion. Then either $a$ or $c$ is not diagonalizable. \end{lem} \begin{proof} Suppose $a,c$ are both diagonalizable. As $a,c$ commute, they are simultaneously diagonalizable and share a basis $\mathfrak{B}$ of eigenvectors. As $c$ is not torsion, there is $\lambda_c \in \mathbb{C}^\times$ which is not a root of unity and $v \in \mathfrak{B}$ such that $cv = \lambda_c v$. Let $\lambda_a \in \mathbb{C}^\times$ be such that $a v = \lambda_a v$. By way of contradiction, we will show $a (b^{k} v) = (\lambda_a \lambda_c^k)(b^{k}v)$ for all $k \geq 1$. As $\lambda_c$ is not a root of unity, this implies $a$ has infinitely many eigenvalues, which is impossible for an $n \times n$ matrix. The base case holds as $$ a(bv) = bac v = (\lambda_a \lambda_c) (bv).$$ Let $k \geq 2$ and suppose $a (b^{k-1} v) = (\lambda_a \lambda_c^{k-1}) (b^{k-1} v)$. As $c$ commutes with $b$, $$ a(b^{k} v) = ab( b^{k-1} v) = bac( b^{k-1} v) = bab^{k-1}cv = (\lambda_c) (bab^{k-1}v). $$ Applying the inductive assumption, \begin{align*} (\lambda_c) (bab^{k-1}v) &= (\lambda_c) b ( \lambda_a \lambda_c^{k-1} b^{k-1} v) = (\lambda_a \lambda_c^{k}) (b^{k} v). \end{align*} \end{proof} \noindent We now prove a slight weakening of Lemma~\ref{prop:heisen0} for solvable groups. Recall $a \in \text{Gl}_n(\mathbb{C})$ is \textbf{unipotent} if some conjugate of $a$ is upper triangular with every diagonal entry equal to one. The only diagonalizable unipotent matrix is the identity. We recall a theorem of Mal\textquotesingle tsev \cite{Mal-sol}. \begin{fact}\label{thm:lkm} Suppose $G$ is solvable. Then there is a finite index subgroup $G'$ of $G$ such that $G'$ is conjugate to a subgroup of UT$_n(\mathbb{C})$. \end{fact} \noindent We now derive an easy corollary from Fact \ref{thm:lkm} \begin{lem}\label{prop:solv-1} Suppose $G$ is solvable and not virtually abelian. Then $G$ contains a non-diagonalizable element. \end{lem} \begin{proof} Suppose every element of $G$ is diagonalizable. After applying Fact \ref{thm:lkm} and making a change of basis if necessary we suppose $G' = G \cap \text{UT}_n(\mathbb{C})$ has finite index in $G$. Let $\rho : \text{UT}_n(\mathbb{C}) \to \text{D}_n(\mathbb{C})$ be the natural quotient map; that is the restriction to the diagonal. Every element of the kernel of $\rho$ is unipotent. Thus the restriction of $\rho$ to $G'$ is injective, and so $G'$ is abelian. \end{proof} \section{Non-diagonalizable matrices} \begin{lem}\label{lem:diag} Suppose $G$ contains a non-diagonalizable matrix. Then there is a rational function $h$ on $\text{Gl}_n(\mathbb{C}) \times \text{Gl}_{n}(\mathbb{C})$ such that $h(G \times G) \subseteq \mathbb{C}$ is dense in $\mathbb{R}_{>0}$. \end{lem} \begin{proof} Suppose $a \in G$ is non-diagonalizable. Let $b \in \text{Gl}_n(\mathbb{C})$ be such that $bab^{-1}$ is in Jordan form, i.e. $$ bab^{-1} = \begin{pmatrix} A_1 & O & \dots & O \\[4pt] O & A_2 & \ldots & O \\[4pt] \vdots & \vdots & \ddots & \vdots \\[4pt] O & O & \ldots & A_l \end{pmatrix} $$ where each $A_i$ is a Jordan block and each $O$ is a zero matrix of the appropriate dimensions. We have $$ b a^k b^{-1} = \begin{pmatrix} A^k_1 & O & \dots & O \\[4pt] O & A^k_2 & \ldots & O \\[4pt] \vdots & \vdots & \ddots & \vdots \\[4pt] O & O & \ldots & A^k_l \end{pmatrix} \quad \text{for all } k. $$ As $a$ is not diagonalizable, $A_k$ has more then one entry for some $k$. We suppose $A_1$ is $m$-by-$m$ with $m \geq 2$. For some $\lambda \in \mathbb{C}^\times$ we have $$ A_1 = \begin{pmatrix} \lambda & 1 & 0 & \ldots & 0 & 0 \\[4pt] 0 & \lambda & 1 & \ldots & 0 & 0 \\[4pt] 0 & 0 & \lambda & \ldots & 0 & 0 \\[4pt] \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\[4pt] 0 & 0 & 0 & \ldots & \lambda & 1 \\ 0 & 0 & 0 & \ldots & 0 & \lambda \\ \end{pmatrix}. $$ It is well-known and easy to show by induction that for every $k \geq 1$: $$ A^k_{1} = \begin{pmatrix} \lambda^k & \binom{k}{1} \lambda^{k-1} & \binom{k}{2} \lambda^{k-2} & \binom{k}{3} \lambda^{k-3} & \binom{k}{4} \lambda^{k-4} & \ldots & \binom{k}{m} \lambda^{k-m+1} \\[4pt] 0 & \lambda^k & \binom{k}{1} \lambda^{k-1} & \binom{k}{2} \lambda^{k-2} & \binom{k}{3} \lambda^{k-3} & \ldots & \binom{k}{m - 1} \lambda^{k-m +2}\\[4pt] 0 & 0 & \lambda^k & \binom{k}{1} \lambda^{k-1} & \binom{k}{2} \lambda^{k-2} & \ldots & \binom{k}{m - 2} \lambda^{k - m +3} \\[4pt] 0 & 0 & 0 & \lambda^{k} & \binom{k}{1} \lambda^{k-1} & \ldots & \binom{k}{m - 3} \lambda^{k - m +4} \\[4pt] 0 & 0 & 0 & 0 & \lambda^k & \ldots & \binom{k}{m - 4} \lambda^{k - m + 5} \\[4pt] \vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & 0 & \vdots & \vdots & \lambda^k & \binom{k}{1} \lambda^{k-1} \\[4pt] 0 & 0 & 0 & \ldots & \ldots & 0 & \lambda^k \\ \end{pmatrix}. $$ Let $g_{ij}$ be the $(i,j)$-entry of $g \in \text{Gl}_n(\mathbb{C})$. Thus, for each $k \geq 1$, $$(ba^{k}b^{-1})_{01} = k\lambda^{k-1} \quad \text{and} \quad (ba^{k}b^{-1})_{11} = \lambda^{k}.$$ We define a rational function $h'$ on $\Gl_n(\mathbb{C}) \times \Gl_n(\mathbb{C})$ by declaring $$ h'(g,g') := \frac{ g_{01} g'_{11} }{ g'_{01} g_{11} } $$ for all $g,g' \in \Gl_n(\mathbb{C})$ such that $g_{11}, g'_{01} \neq 0$. Then define $h$ by declaring $$h(g,g') := h'(bgb^{-1}, bg'b^{-1})$$ We have $$ h(a^i,a^j) = \frac{ (i \lambda^{i-1})(\lambda^j) }{ ( j \lambda^{j-1} ) (\lambda^i) } = \frac{i}{j} \quad \text{for all } i,j \geq 1.$$ Thus $\mathbb{Q}_{>0}$ is a subset of the image of $G \times G$ under $h$. \end{proof} \begin{cor}\label{cor:diag} If $\Gamma$ contains a non-diagonalizable matrix, then $\bar{\mathbb{R}}_\Gamma$ defines $\mathbb{Z}$. In particular, if $\Gamma$ is either \begin{itemize} \item solvable and not virtually abelian, or \item torsion-free, nilpotent and non-abelian, \end{itemize} then $\bar{\mathbb{R}}_\Gamma$ defines $\mathbb{Z}$. \end{cor} \begin{proof} Applying Lemma~\ref{lem:diag}, suppose $h$ is a rational function on Gl$_n(\mathbb{C}) \times \text{Gl}_n(\mathbb{C})$ such that the image of $\Gamma \times \Gamma$ under $h$ is dense in $\mathbb{R}_{>0}$. Note $\Gamma$ is countable as $\Gamma$ is discrete. It follows that the image of $\Gamma \times \Gamma$ under any function is co-dense in $\mathbb{R}_{>0}$. Fact~\ref{thm:hier}(1) implies that $\bar{\mathbb{R}}_\Gamma$ defines $\mathbb{Z}$. The second claim follows from the first by applying Lemma \ref{prop:heisen0} and Lemma \ref{prop:solv-1}. \end{proof} \begin{cor}\label{cor:cyclic} If $a \in \text{Gl}_n(\mathbb{C})$ is non-diagonalizable, then $(\bar{\mathbb{R}}, \{ a^k : k \in \mathbb{Z}\})$ defines $\mathbb{Z}$. \end{cor} \begin{proof} Set $G := \{ a^k : k \in \mathbb{Z}\}$. The proof of Lemma~\ref{lem:diag} shows that in this case $\mathbb{Q}_{>0}$ is the intersection of $h(G \times G)$ and $\mathbb{R}_{>0}$. Thus the corollary follows by Julia Robinson's classical theorem of definability of $\mathbb{Z}$ in $(\mathbb{Q},+,\cdot)$ in \cite{julia}. \end{proof} \section{The case of exponential growth}\label{section:exp} \noindent We recall the \textbf{Assouad dimension} of a metric space $(X,d)$. See Heinonen \cite{Heinonen} for more information. The Assouad dimension of a subset $Y$ of $\mathbb{R}^k$ is the Assouad dimension of $Y$ equipped with the euclidean metric induced from $\mathbb{R}^k$. \medskip \noindent Suppose $A \subseteq X$ has at least two elements. Then $A$ is $\delta$-\textbf{separated} for $\delta \in \mathbb{R}_{>0}$ if $d(a, b) \geq \delta$ for all distinct $a,b \in A$, and $A$ is \textbf{seperated} if $A$ is $\delta$-seperated for some $\delta > 0$. Let $\mathscr{S}(A) \in \mathbb{R}$ be the supremum of all $\delta \geq 0$ for which $A$ is $\delta$-seperated. Let $\mathscr{D}(A)$ be the \textbf{diameter} of $A$; that is the infimum of all $\delta \in \mathbb{R} \cup \{ \infty \}$ such that $d(a,b) < \delta$ for all $a,b \in A$, and $A$ is \textbf{bounded} if $\mathscr{D}(A) < \infty$. Note $\mathscr{S}(A) \leq \mathscr{D}(A)$. The \textbf{Assouad dimension} of $(X,d)$ is the infimum of the set of $\beta \in \mathbb{R}_{>0}$ for which there is a $C > 0$ such that $$ |A| \leq C \left( \frac{ \mathscr{D}(A) }{ \mathscr{S}(A) } \right)^\beta \text{ for all bounded and separated } A \subseteq X. $$ The proof of Fact~\ref{fact:use-assouad} is an elementary computation which we leave to the reader. \begin{fact}\label{fact:use-assouad} Suppose there is a sequence $\{A_m\}_{m \in \mathbb{N}}$ of bounded separated subsets of $X$ with cardinality at least two, and $B,C,t > 1$ are such that $$ |A_m| \geq C^m \quad \text{and} \quad \frac{ \mathscr{D}(A_m) }{ \mathscr{S}(A_m) } \leq t B^m \quad \text{for all } m $$ then $(X,d)$ has positive Assouad dimension. \end{fact} \medskip \noindent Let $|v|$ be the usual euclidean norm of $v \in \mathbb{C}^n$. Given $g \in \text{M}_n(\mathbb{C})$ we let $$ \| g \| =\inf \{ t \in \mathbb{R}_{>0} : | gv | \leq t |v| \text{ for all } v \in \mathbb{C}^n \} $$ be the operator norm of $g$. Then $\|\text{}\|$ is a linear norm on M$_{n}(\mathbb{C})$ and satisfies $\| g h \| \leq \| g \| \| h \|$ for all $g,h \in \text{M}_{n}(\mathbb{C})$. As any two linear norms on M$_{n}(\mathbb{C})$ are bi-Lipschitz equivalent the metric induced by $\| \|$ is bi-Lipschitz equivalent to the usual euclidean metric on $\mathbb{R}^{n^2}$. \begin{prop}\label{prop:exp} Suppose $\Gamma$ contains a finitely generated subgroup $\Gamma'$ of exponential growth. Then $\Gamma$ has positive Assouad dimension. \end{prop} \begin{proof} Because Assouad dimension is a bi-Lipschitz invariant (see \cite{Heinonen}), it suffices to show that $\Gamma$ has positive Assouad dimension with respect to the metric induced by $\| \|$. We let $I$ be the $n$-by-$n$ identity matrix. Let $S$ be a symmetric generating set of $\Gamma'$, and let $S_m$ be the set of $m$-fold products of elements of $S$ for $m \geq 2$. Set $$ B := \max\{ \| g \| : g \in S\} \quad \text{and} \quad D := \min \{ \| g - I \| : g \in \Gamma \}. $$ Note that $D > 0$, as $\Gamma$ is discrete, and that $B > 0$, as $\Gamma \neq \{I\}$. Induction shows that $\| g \| \leq B^m$ when $g \in S_m$. The triangle inequality directly yields $\mathscr{D}(S_m) \leq 2B^m$. Each $S_m$ is symmetric as $S$ is symmetric. Therefore $\| g^{-1} \| \leq B^m$ for all $g \in S_m$. Let $g, h \in \Gamma$. We have $$ \| I - g^{-1} h \| \leq \| g^{-1} \| \| g - h \|. $$ Equivalently, $$ \frac{ \| I - g^{-1} h \| }{ \| g^{-1} \| } \leq \| g - h \|. $$ \medskip\noindent Suppose $g,h \in S_m$ are distinct. Then $g^{-1}h \neq I$, and hence $\| I - g^{-1}h \| \geq D$. So $$ \| g - h \| \geq \frac{ \| I - g^{-1}h \| }{ \| g^{-1} \| } \geq \frac{ D }{ B^m }. $$ Hence $\mathscr{S}(S_m) \geq D/B^m$. Thus $$ \frac{ \mathscr{D}(S_m)}{ \mathscr{S}(S_m) } \leq \frac{ 2B^m }{ D/B^m} = \frac{ 2 }{D } B^{2m} .$$ As $\Gamma'$ has exponential growth, there is a $C > 0$ such that $|S_m| \geq C^m$ for all $m$. An application of Fact~\ref{fact:use-assouad} shows that $\Gamma$ has positive Assouad dimension. \end{proof} \begin{prop}\label{prop:solv-2} Suppose $\Gamma$ is not virtually abelian. Then $\bar{\mathbb{R}}_\Gamma$ defines $\mathbb{Z}$. \end{prop} \begin{proof} By Corollary \ref{cor:diag}, we can assume that $\Gamma$ is solvable. Thus by Fact \ref{thm:tits}, the group $\Gamma$ contains a non-abelian free subgroup. Therefore $\Gamma$ has positive Assouad dimension by Proposition~\ref{prop:exp}. We conclude that $\bar{\mathbb{R}}_{\Gamma}$ defines $\mathbb{Z}$ by Fact~\ref{thm:hier}(2). \end{proof} \section{The virtually abelian case} \noindent We first reduce the virtually abelian case to the abelian case. \begin{lem}\label{lem:ab-1} Suppose $G$ is virtually abelian and every element of $G$ is diagonalizable. Then there is a finite index abelian subgroup $G'$ of $G$ such that $(\bar{\mathbb{R}},G)$ and $(\bar{\mathbb{R}},G')$ are interdefinable. \end{lem} \begin{proof} Let $G''$ be a finite index abelian subgroup of $G$. As every element of $G''$ is diagonalizable, $G''$ is simultaneously diagonalizable. Fix $g \in \text{Gl}_n(\mathbb{C})$ such that $gag^{-1}$ is diagonal for all $a \in G''$. Let $G'$ be the set of $a \in G$ such that $gag^{-1}$ is diagonal, i.e. $G'$ is the intersection of $G$ and $g^{-1} \text{D}_n(\mathbb{C}) g$. Then $G'$ is abelian, $(\bar{\mathbb{R}},G)$-definable, and is of finite index in $G$ as $G'' \subseteq G'$. Because $G'$ has finite index in $G$, we have $$ G = g_1 G' \cup \ldots \cup g_m G' \text{ for some } g_1,\ldots,g_m \in G.$$ So $G$ is $(\bar{\mathbb{R}},G')$-definable. \end{proof} Proposition~\ref{prop:lastab} finishes the proof of Theorem A. \begin{prop}\label{prop:lastab} Suppose $\Gamma$ is abelian and $\bar{\mathbb{R}}_\Gamma$ does not define $\mathbb{Z}$. Then there is $\lambda \in \mathbb{R}_{>0}$ such that $\bar{\mathbb{R}}_\Gamma$ is interdefinable with $\bar{\mathbb{R}}_\lambda$. \end{prop} \noindent Let $\mathfrak{u}: \mathbb{C}^\times \to \mathbb{U}$ be the argument map and $| \text{ } | : \mathbb{C}^\times \to \mathbb{R}_{>0}$ be the absolute value map. Thus $z = \mathfrak{u}(z)|z|$ for all $z \in \mathbb{C}^\times$. Let $\mathbb{U}_m$ be the group of $m$th roots of unity for all $m \geq 1$. In the following proof of Proposition \ref{prop:lastab} we will use the immediate corollary of \cite[Theorem 1.3]{discrete} that the structure $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}},\mu^{\mathbb{Z}})$ defines $\mathbb{Z}$ whenever $\log_{\lambda} \mu \notin \mathbb{Q}$, and is is interdefinable with $(\bar{\mathbb{R}},\lambda^{\mathbb{Z}})$ otherwise. \begin{proof} Fact~\ref{thm:hier}(1) implies every countable $\bar{\mathbb{R}}_\Gamma$-definable subset of $\mathbb{R}$ is nowhere dense. It follows that every $\bar{\mathbb{R}}_\Gamma$-definable countable subgroup of $\mathbb{U}$ is finite and every $\bar{\mathbb{R}}_\Gamma$-definable countable subgroup of $(\mathbb{R}_{>0},\cdot)$ is of the form $\lambda^{\mathbb{Z}}$ for some $\lambda \in \mathbb{R}_{>0}$. Every element of $\Gamma$ is diagonalizable by Corollary~\ref{cor:diag}. Thus $\Gamma$ is simultaneously diagonalizable. After making a change of basis we suppose $\Gamma$ is a subgroup of D$_n(\mathbb{C})$. We identify D$_n(\mathbb{C})$ with $(\mathbb{C}^\times)^n$. Let $\Gamma_i$ be the image of $\Gamma$ under the projection $(\mathbb{C}^\times)^n \to \mathbb{C}^\times$ onto the $i$th cordinate for $1 \leq i \leq n$. Each $\mathfrak{u}(\Gamma_i)$ is finite. Fix an $m$ such that $\mathfrak{u}(\Gamma_i)$ is a subgroup of $\mathbb{U}_m$ for all $1 \leq i \leq n$. For each $1 \leq i \leq n$, $|\Gamma_i|$ is a discrete subgroup of $\mathbb{R}_{>0}$ and is thus equal to $\alpha_i^{\mathbb{Z}}$ for some $\alpha_i \in \mathbb{R}_{>0}$. By \cite[Theorem 1.3]{discrete} each $\alpha_i$ is a rational power of $\alpha_1$. Let $\lambda \in \mathbb{R}_{>0}$ be a rational power of $\alpha_1$ such that each $\alpha_i$ is an integer power of $\lambda$. We show $\bar{\mathbb{R}}_\Gamma$ and $\bar{\mathbb{R}}_\lambda$ are interdefinable. Note that $\lambda^{\mathbb{Z}}$ is $\bar{\mathbb{R}}_\Gamma$-definable; so it suffices to show $\Gamma$ is $\bar{\mathbb{R}}_\lambda$-definable. Every element of $\Gamma_i$ is of the form $\sigma \lambda^k$ for some $\sigma \in \mathbb{U}_m$ and $k \in \mathbb{Z}$. Thus $\Gamma$ is a subgroup of $$ \Gamma' = \left\{ \begin{pmatrix} \sigma_1 \lambda^{k_1} & 0 & \dots & 0 \\[4pt] 0 & \sigma_2 \lambda^{k_2} & \ldots & 0 \\[4pt] \vdots & \vdots & \ddots & \vdots \\[4pt] 0 & 0 & \ldots & \sigma_n \lambda^{k_n} \end{pmatrix} : \sigma_1, \ldots,\sigma_n \in \mathbb{U}_m, k_1,\ldots,k_n \in \mathbb{Z} \right\}. $$ Note $\Gamma'$ is $\bar{\mathbb{R}}_\lambda$-definable. Abusing notation we let $\mathfrak{u} : (\mathbb{C}^\times)^n \to \mathbb{U}^n$ and we let $| \text{ } | : (\mathbb{C}^\times)^n \to (\mathbb{R}_{>0})^n$ be given by $$ \mathfrak{u}(z_1,\ldots,z_n) = (\mathfrak{u}(z_1),\ldots,\mathfrak{u}(z_n)) \quad \text{and} \quad |(z_1,\ldots,z_n)| = (|z_1|,\ldots,|z_n|). $$ Then the map $(\mathbb{C}^\times)^n \to \mathbb{U}^{n} \times (\mathbb{R}_{>0})^n$ given by $\bar{z} \mapsto (\mathfrak{u}(\bar{z}), |\bar{z}|)$ restricts to a $\bar{\mathbb{R}}_\lambda$-definable isomorphism between $\Gamma'$ and $\mathbb{U}^n_m \times (\lambda^{\mathbb{Z}})^n$. Lemma~\ref{lem:group} below implies any subgroup of $\mathbb{U}^n_m \times (\lambda^{\mathbb{Z}})^n$ is $\bar{\mathbb{R}}_\lambda$-definable. \end{proof} \noindent We consider $(\mathbb{Z}/m\mathbb{Z},+)$ to be a group with underlying set $\{0,\ldots,m-1\}$ in the usual way so that $(\mathbb{Z}/m\mathbb{Z},+)$ is a $(\mathbb{Z},+)$-definable group. Lemma~\ref{lem:group} is folklore. We include a proof for the sake of completeness. \begin{lem}\label{lem:group} Every subgroup $H$ of $(\mathbb{Z}/m\mathbb{Z})^l \times \mathbb{Z}^n$ for $l \geq 0$ is $(\mathbb{Z},+)$-definable. \end{lem} \begin{proof} We first reduce to the case $l = 0$. The quotient map $\mathbb{Z} \to \mathbb{Z}/m\mathbb{Z}$ is $(\mathbb{Z},+)$-definable, it follows that the coordinate-wise quotient $\mathbb{Z}^{l} \times \mathbb{Z}^{n} \to (\mathbb{Z}/m\mathbb{Z})^l \times \mathbb{Z}^n$ is $(\mathbb{Z},+)$-definable. It suffices to show the preimage of $H$ in $\mathbb{Z}^{l + n}$ is $(\mathbb{Z},+)$-definable. Suppose $H$ is a subgroup of $\mathbb{Z}^n$. Then $H$ is finitely generated with generators $\beta_1,\ldots,\beta_k$ where $ \beta_i = (b^{i}_1,\ldots,b^{i}_n) \text{ for all } 1 \leq i \leq k. $ Then \begin{align*} H & = \left\{ \sum_{i = 1}^{k} c_i \beta_i : c_1,\ldots,c_k \in \mathbb{Z} \right\} \\[4pt] & = \left\{ \left( \sum_{i = 1}^{k} c_i b^{i}_{1} , \ldots , \sum_{i = 1}^{k} c_i b^{i}_n \right) : c_1,\ldots, c_n \in \mathbb{Z} \right\}. \end{align*} Thus $H$ is $(\mathbb{Z},+)$-definable. \end{proof} \section{countable $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable groups} \noindent Fix $\lambda \in \mathbb{R}_{>0}$ and an o-minimal $\mathscr{R}$ with field of exponents $\mathbb{Q}$. Since $(\mathscr{R},\lambda^{\mathbb{Z}})$ does not define $\mathbb{Z}$ by \cite[Theorem 3.4.2]{Miller-tame}, Theorem~A implies every $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable discrete subgroup of $\Gl_n(\mathbb{C})$ is virtually abelian. We extend this result to all countable interpretable groups. \begin{prop}\label{prop:lam-group} Every countable $(\mathscr{R},\lambda^{\mathbb{Z}})$-interpretable group is virtually abelian. \end{prop} \noindent Proposition~\ref{prop:lam-group} follows directly from several previous results. Every d-minimal expansion of $\bar{\mathbb{R}}$ admits definable selection by Miller~\cite{Miller-selection}. Therefore an $(\mathscr{R},\lambda^{\mathbb{Z}})$-interpretable group is isomorphic to an $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable group. We now recall two results of Tychonievich. The first is a special case of \cite[4.1.10]{Tychon-thesis}. \begin{fact}\label{prop:ty} If $X \subseteq \mathbb{R}^k$ is $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable and countable, then there is an $\bar{\mathbb{R}}_\lambda$-definable surjection $f : (\lambda^\mathbb{Z})^m \to X$ for some $m$. \end{fact} \noindent Fact~\ref{prop:induced} is a minor rewording of \cite[4.1.2]{Tychon-thesis}. \begin{fact}\label{prop:induced} Every $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable subset of $(\lambda^{\mathbb{Z}})^m$ is $(\lambda^{\mathbb{Z}},<,\cdot)$-definable. \end{fact} \noindent Facts \ref{prop:ty} and \ref{prop:induced} together imply that every countable $(\mathscr{R},\lambda^{\mathbb{Z}})$-definable group is isomorphic to a $(\mathbb{Z},<,+)$-definable group. Now apply the following result of Onshuus and Vicaria~\cite{Onshuus-pres} to complete the proof of Proposition~\ref{prop:lam-group}. \begin{fact} Every $(\mathbb{Z},<,+)$-definable group is virtually abelian. \end{fact} \bibliographystyle{abbrv}

e0880c040e43451569042c08df45d864ed94e5e5

\section{Introduction} The common envelope (CE) interaction is a phase of binary stellar evolution invoked to describe a dramatic decrease in the orbital separation of two stars, coupled with the ejection of a large amount of gas. The idea is attributed to \citet[][by whom private communications with Ostriker and Webbink are also credited]{paczynski1976common}, who was attempting to explain the origin of some compact, evolved binaries. As the primary, more massive star in an intermediate separation binary system swells to become a red giant, it transfers mass to its companion in a phase of Roche lobe overflow. If the mass transfer is unstable, the companion is likely swallowed by the expanding outer atmosphere of the primary. What follows is a phase of dramatic reduction in the orbital separation on a dynamical time-scale, after which the envelope is presumably ejected to reveal a compact binary composed of the companion and the core of the giant primary. If the envelope is not ejected then it is expected that the stars will eventually merge. Understanding this interaction is critical to the interpretation of binary phenomena, from intermediate luminosity red transients \citep[e.g.,][]{blagorodnova2017common} to determining the nature of the progenitor of type Ia supernovae \citep[e.g.,][]{toonen2013effect}, and to illuminating the evolutionary path that leads to compact neutron star and black hole systems that can merge by emission of gravitational waves \citep[e.g.,][]{abbott2016observation}. See \citet{de2017dawes} for a review. Several observed objects carry characteristics distinctive of the common envelope interaction and have been interpreted as common envelope mergers. While some have been known for a while (V838~Monocerotis, \citealt{bond2003energetic}), many more are being discovered now, with the intensification of time-resolved observations. V1309~Sco \citep{tylenda2011v1309} was interpreted as the merging of a solar-mass subgiant with a much less massive companion. M31-2015LRN \citep{macleod2017lessons}, and possibly V838~Mon, were likely slightly more massive, in the range of 3-5.5\,{M$_\odot$}. M101-OT \citep{blagorodnova2017common} possibly came from an 18\,{M$_\odot$}\ progenitor, while NGC~4490-OT \citep{smith2016massive} might have derived from an even more massive star. This flurry of observations promises to place much needed constraints on common envelope simulations, which in turn can start to give a physical explanations to these phenomena \citep{galaviz2017common}. Simulations of the common envelope utilising a variety of codes have, with mild success, reproduced observable post-CE system parameters. A non-exhaustive list of recent works includes the AMR simulations \citep{ricker2012amr}, comparisons of SPH and grid techniques \citep{passy2012simulating, iaconi2017effect}, SPH simulations examining the inclusion of recombination energy \citep{nandez2015recombination} and moving-mesh simulations \citep{ohlmann2016hydrodynamic}. Most simulations model lower-mass red giant branch stars and thereby only cover a limited section of the parameter space. In addition, simulations typically start with the companion very close to the surface of the star and stop as soon as the orbital distance has stabilised. This leaves open the question of whether the phases surrounding the dynamic inspiral have an effect on the inspiral itself, on the parameters of the post-CE binaries, including the ejecta or nebula that may derive from it, and on the light properties near the time of the outburst. The effect of starting common envelope simulations at primary Roche lobe contact was investigated by \citet{iaconi2017effect}, who concluded that including the unstable Roche lobe overflow phase leads to a marginally wider final separation. They also showed that starting the simulations at a wider separation does not increase the amount of unbound common envelope mass\footnote{Envelope unbinding remains an issue, although some solutions or partial solutions have started appearing in the literature. Recent discussion has turned towards the ability of recombination energy to act as a source of energy \citep{han1994possible}, allowing further envelope unbinding (see, for example, \citealt{ivanova2018use} and the counter arguments of \citealt{grichener2018limited} and \citealt{soker2018radiating}).}. However, they did not investigate the outflow, nor the details of the early, pre-dynamical, inspiral. In one of their simulations, carried out with the SPH code \textsc{phantom} \citep{price2017phantom}, they noticed that the relatively long phase of Roche lobe overflow, before the fast inspiral, resulted in the ejection of some mass from the second and third Lagrangian points, $L_2$ and $L_3$. This raised the suspicion that the modelled common envelope ejecta would differ from when the simulation starts at smaller orbital separations, something that could impact studies of the nebular shapes resulting from common envelope ejections \citep{frank2018planetary,garcia2018common}. Observations already indicate that phases just before the dynamical common envelope may play a role in the overall outcome of the interaction. For example, V1309~Scorpii was captured in the OGLE-III and OGLE-IV fields between 2001 and 2008, when the system displayed reduction in the orbital period, before undergoing a fast brightening of several magnitudes \citep{tylenda2011v1309}. Before the outburst the object steadily brightened over 3\,yr. These time-scales and the concomitant characteristics of the contact-binary periodic variability are not easily explained with a common envelope dynamical phase, implying a slower interaction leading up to fast dynamical common envelope merger. In order to explain the slower, pre-outburst rise of V1309~Sco, \citet{pejcha2016binary, pejcha2017pre} and \citet{metzger2017shock} took a different approach to simulating common envelope transients. Their study targeted the light curve of a contact binary before inspiral. By simulating the flow through the $L_2$ point, they claimed to have identified the cause of the slow pre-outburst light increase in V1309~Sco, as well as of other characteristics of transient lightcurves. Although they do not comment on the effects of $L_2$ mass-loss on the subsequent phase of dynamical inspiral, the implication is that the two phases are intimately connected. \citet{macleod2018runaway, macleod2018bound} presented additional simulations of the Roche lobe overflow phase leading up to a common envelope inspiral. They simulated a 1\,{M$_\odot$}\ primary and a 0.3\,{M$_\odot$}\ companion, at the start of Roche lobe overflow (20.6\,R$_\odot$, for their system). The simulation was not continued past the beginning of the dynamic inspiral. They showed the decay of the orbit to be driven by exchange of angular momentum from the companion to the mass flow, which is then lost from the system through the $L_2$ point. \citet{macleod2018bound} claim that, contrary to what was concluded by \citet{pejcha2017pre}, mass lost through the $L_2$ point is not unbound until much later in the simulations, when the companion is almost in contact with the primary. This bound outflow forms a large disk of material in the orbital plane which collimates the late time unbinding, forming a bipolar gas distribution. In summary, much is yet to be learned about the common envelope interaction itself and the transient phenomena that it precedes. Including an extended phase of mass transfer and following the distribution of ejecta after the dynamic inspiral moves us towards an integrated understanding of these adjacent phases. In Section~\ref{sec:simulationsetup} we summarise the setup for our simulations. In Section~\ref{sec:numerical}, we give an overview of the simulation behaviour and discuss the effect of increasing resolution on our simulations, along with the conservation of energy and angular momentum. In Section~\ref{sec:orbitalevolution}, we go over the pre-inspiral phase of our simulations, analysing Roche lobe overflow numerically with comparisons to analytical equations. Further, we discuss the different processes contributing to the the reduction in orbital separation both before and during the dynamic inspiral. We also discuss the creation of a decretion disk, and the presence of fallback material. Section~\ref{sec:planetarynebulae} contains a discussion of the formation of planetary nebulae from common envelopes, and the applications of our final gas distributions to further simulations. We conclude and summarise in Section~\ref{sec:conclusions}. \section{Initial conditions} \label{sec:simulationsetup} The simulations presented here follow on from work done by \citet{passy2012simulating} and \citet{iaconi2017effect}. The simulations we analyse include one that was already presented by \citet{iaconi2017effect}, who list further details. In summary, we modelled a binary system made of a 0.88\,{M$_\odot$}, 90\,R$_\odot$\ RGB primary with a 0.392\,{M$_\odot$}\ point mass core in orbit with a 0.6\,{M$_\odot$}\ point mass companion. We used three different resolutions of $8 \times 10^4$, $2.3 \times 10^5$ and $1.1 \times 10^6$ SPH particles. In addition to these simulations, a $2.3 \times 10^5$ particle simulation was also evolved with a primary star in corotation with the orbit. The resolution length in SPH, the smoothing length, is related to the local number density of particles. Initially, SPH particles in our high resolution simulation have smoothing lengths ranging between 0.2 and 9\,R$_\odot$. For more details on the smoothing lengths in the different simulations, refer to Table~\ref{table:resolution}. The core of our giant star and the companion are represented by point masses. These particles do not have associated internal energies or pressures, instead interacting purely gravitationally with other particles in the simulation. Their gravitational potentials have been softened by a cubic spline kernel, with a softening length of 3\,R$_\odot$. This means that within a radius of 3\,R$_\odot$, the potential of the point mass particles is almost flat \citep[for more information on the exact process, see][]{price2017phantom}, and outside a radius of twice the softening length, the potential is exactly the analytical form. Softened potentials are typically used when point mass particles are not allowed to accrete gas, which is the case in our simulations. Before the binary simulations are carried out, the primary star is relaxed in the computational domain, as explained by \citet{iaconi2017effect}. This tends to result in a stellar structure with a slightly larger radius. The relaxed radii differ slightly with resolution, being 87, 91 and 93\,R$_\odot$\ for the low, medium and high resolutions, respectively. We also note that we use the volume-equivalent radius (defined by \citealt{nandez2014v1309} to be the radius of a sphere with the total volume occupied by particles) for gas with a higher density than the initial surface density of the star. The corotating simulation starts with a primary star stabilised in the inertial frame. Once the simulation starts, not only is the star suddenly immersed in the companion's potential, but it is also spinning. The combination of the distortion and expansion resulting from these forces may lead to a larger mass transfer rate. We discuss this further in Section~\ref{sssec:masstransfer}. We were unable to perform all the simulations in the corotating frame, because the length of the Roche lobe overflow phase is approximately doubled, and so the simulations would have become prohibitively long. We chose the initial orbital separation such that the Roche lobe of the giant is approximately equal to its stellar radius. With our parameters, the initial orbital separation is about 218\,R$_\odot$\ \citep[by Eq.~2 of][]{eggleton1983approximations}. With this setup, a phase of mass transfer begins immediately upon starting the simulation. The reduction in separation for systems that are started well outside of Roche lobe contact is, in nature, driven by tidal interactions from distortions of the primary envelope by the potential of the secondary. This is also true in simulations, though the orbit decays more quickly as a result of oscillations artificially set in motion by the initial conditions. This said, starting the simulation with an orbital separation larger than 218\,R$_\odot$\ results in a far more stable orbit and it requires longer simulation times to get to a state of mass transfer. \section{Simulation overview} \label{sec:numerical} \begin{figure*} \includegraphics[width=\linewidth]{fig1} \caption{Surface rendering of our 1.1 million particle simulation. The frames are 5\,AU per side. From left to right, top to bottom, each row is a snapshot taken at 1.25, 5.26, 8.2, 12.8, 13.3 and 13.8\,yr. This is intended to display the overall evolution of the simulation. This image, along with all other renderings of the simulation in this paper, was created with \textsc{splash} \citet{price2007splash}. In surface renderings such as this, a value for the opacity is assigned arbitrarily for the desired visual result. In this case $\kappa = 1.9\times10^{-9}$\,cm$^2$\,g$^{-1}$.} \label{fig:evolution} \end{figure*} We first present an overview of the simulation parameters and their dependence on resolution, as well as conservation properties. We use this as a basis to describe the orbital evolution in Section~\ref{sec:orbitalevolution}. We show a visual overview of the simulation behaviour in Fig.~\ref{fig:evolution}. In the top three panels, the system is undergoing Roche lobe overflow. The bottom three panels show the evolution of the system within a year from beginning the dynamic inspiral. \subsection{System evolution} \label{ssec:overview} The two left-hand columns of Fig.~\ref{fig:slices} shows the density evolution in the orbital plane (left) and along the plane perpendicular to the orbital plane (right) as a function of time (top to bottom) for the 1.1 million particle simulation. The two right-hand columns of Fig.~\ref{fig:slices} shows the corresponding velocity slices. At the start of the simulation, gas moves in a stream from the primary to the companion. This stream does not flow exactly through the inner Lagrange point $L_1$, instead flowing slightly around it. This can be explained by the star not being in corotation with the orbit, which is further evidenced by the fact that the stream does flow through the $L_1$ point in the corotating simulation. Some of this mass collects in a small disk around the companion, while the orbital separation decreases and the mass transfer rate increases. Eventually, gas starts flowing out of the system via the outer Lagrange point $L_2$ and eventually through $L_3$. These outflows are visible in Fig.~\ref{fig:slices} as spiral tails extending from both point masses in the orbital plane. The $L_2$ outflow is present to the right of the companion (right-hand green dot) in the top four rows. The $L_3$ outflow is most prominent in the second and third rows of Fig.~\ref{fig:slices} as a tail extending from the left of the primary. Later, the companion is engulfed in the atmosphere of the donor star, and common envelope inspiral ensues on a dynamical time-scale, leading to a rapid reduction in the orbital separation (bottom row of Fig.~\ref{fig:evolution} and last two rows in Fig.~\ref{fig:slices}). As the companion is engulfed, a shell of material is ejected above escape velocity, becoming unbound. More material is unbound as the inspiral continues, but most of the envelope is only lifted to larger radii, remaining bound. This is consistent with the findings of \citet{iaconi2018effect}, that most of the unbound ejecta is from the surface of the donor star and is lost at the beginning of the dynamic inspiral. The mass that is ejected collides with material that previously flowed out through the $L_2$ and $L_3$ points, shaping the common envelope (we discuss this further in Section~\ref{ssec:gasdistribution}). Due to evacuation of gas within the orbit of the far more compact binary, and a reduced velocity contrast between the cores and the gas due to the envelope being dragged into corotation, the inspiral slows down dramatically. There is still some interaction, hence the cores do not fully halt in their radial migration. Further, during this extended phase of slow inspiral, we begin to witness a partial fallback of some of the gas that was not fully unbound (see Section~\ref{ssec:fallback}). \begin{figure*} \includegraphics[width=\columnwidth]{fig2a} \includegraphics[width=\columnwidth]{fig2b} \caption{Columns 1 and 3: density and velocity slices in the orbital plane respectively. Columns 2 and 4: density and velocity slices in the perpendicular plane respectively. Arrows show the direction and relative magnitude of the velocity. Each of the frames is 12\,AU on each side and was taken from our 1.1 million particle simulation. The two point mass particles are plotted as green points, with the primary to the left and companion to the right. From top to bottom, each row is a snapshot taken at 1.25, 8.2, 12.8, 13.3 and 13.8\,yr.} \label{fig:slices} \end{figure*} \subsection{Resolution Study} \label{ssec:resolution} \begin{table} \begin{tabular}{x{1.1cm} x{1cm} x{1cm} x{1cm} x{1cm} x{1cm}} \hline $n_\text{part}$ & $h_\text{min;0}$ & $h_\text{max;0}$ & $h_\text{min;ins}$ & $h_\text{90;ins}$ & $h_\text{max;ins}$\\ & / R$_\odot$ & / R$_\odot$ & / R$_\odot$ & / R$_\odot$ & / R$_\odot$\\ \hline $7.6 \times 10^{4}$ & 0.52 & 14 & 0.43 & 170 & $1.0 \times 10^4$ \\ $2.3 \times 10^{5}$ & 0.35 & 11 & 0.35 & 75 & $1.2 \times 10^4$ \\ $1.1 \times 10^{6}$ & 0.19 & 8.7 & 0.19 & 34 & $1.9 \times 10^4$ \\ $2.3 \times 10^{5}$\textsuperscript{\textdagger} & 0.35 & 11 & 0.26 & 390 & $2.0 \times 10^4$ \\ \hline \multicolumn{6}{l}{\textsuperscript{\textdagger}\footnotesize{Simulation with corotating primary.}} \end{tabular} \caption{Comparison of SPH particle smoothing lengths, denoted here as $h$, for each of our different resolution simulations. The subscripts min and max refer to the outer limits of the range of $h$, while subscript 90 denotes the $h$ below which 90 per cent of particles lie. Subscript {\it 0} refers to the beginning of the simulation, while subscript {\it ins} refers to the time of fastest descent during the dynamic inspiral (6.4, 8.7, 13.3 and 18.0\,y for the low, medium and high resolution and corotating simulations, respectively).} \label{table:resolution} \end{table} To see whether our results are sensitive to numerical resolution, we performed a resolution study. Because the highest resolution simulation, at 1.1 million SPH particles, required approximately 10 months on a 32 core server, increasing the resolution further is currently unfeasible. Instead, we compared our simulation with a series of lower resolution calculations. Table~\ref{table:resolution} lists the particle numbers and the corresponding minimum and maximum resolution lengths (the SPH smoothing length) in each simulation, both at the beginning of the simulation (second and third columns) and during the dynamical inspiral (fourth, fifth and sixth columns). It should be noted that the smoothing length of a particle scales inversely and smoothly with density. Hence the simulations are most well resolved around the cores, where the gas is most dense. We observe that the 1.1 million particle simulation has smoothing lengths typically smaller than half that of the 80\,000 particle simulation, except for the outer portions of the initial star. \begin{figure} \includegraphics[width=\columnwidth]{fig3} \caption{Comparison of the orbital evolution (solid lines) and bound mass (dashed lines) for simulations with differing resolutions. Resolutions are $8\times10^4$ particles (blue), $2.3\times10^5$ particles (green) and $1.1\times10^6$ particles (red). The $2.3\times10^5$ particle simulation was also run with a corotating primary (purple).} \label{fig:218sepbound} \end{figure} A comparison of the three simulations shows that the orbital separation at the end of the fast inspiral and the average rate of descent during the fast inspiral appear converged with respect to numerical resolution (see Table~\ref{table:convergence}). These data were collected from the simulation outputs by defining the dynamic inspiral to start and finish at times such that $|\frac{\dot{a}}{a}| \geq \frac{1}{15} \max |\frac{\dot{a}}{a}|$ holds true. For more discussion of the time-scales, see Section~\ref{sssec:orbitalevolutiontime-scale}. The full orbital evolution of the simulations is shown as the solid curves in Fig.~\ref{fig:218sepbound}. Table~\ref{table:convergence} also gives the amount of unbound material in our simulations, which is resolution dependent. In this work, we use the sum of kinetic and potential energies to determine unbound particles (if $E_\text{kin} + E_\text{pot} > 0$, the gas is unbound from the system). This gives a lower limit to the unbound mass. In some works the internal (thermal) energy is also considered ($E_\text{kin} + E_\text{pot} + E_\text{int} > 0$), which should lead to a greater amount of unbound gas. We include the amounts of unbound mass calculated with both criteria in Table~\ref{table:convergence}, but it is our experience that even including internal energy in the unbound criterion, the amount of unbound mass does not increase significantly \citep{staff2016agb}. Fig.~\ref{fig:218sepbound} compares the mass of bound material for our three resolutions (dashed lines). Higher resolution simulations unbind less material. Gas is primarily unbound during the fast inspiral, although a small amount is unbound before, during Roche lobe overflow. This pre-inspiral unbinding is smaller at higher resolution, with the high resolution simulation unbinding only 0.05\,{M$_\odot$}\ before the inspiral. This is a consequence of our choice of initial conditions. Comparing the inertial and corotating frame simulations at intermediate resolution we see that the pre-inspiral unbinding is greater for rotating stars. The length of the pre-inspiral phase does not appear to have converged at the resolutions we have tested, meaning that it could be longer in reality. Further, each of the simulations begins to experience renewed unbinding shortly after the time of orbital stabilisation. For the high resolution simulation, this begins to take place shortly before we stop the simulation (hence is not seen in Fig.~\ref{fig:218sepbound}), much longer after stabilisation than for the lower resolution simulation. This unbinding event is somewhat deceptive, in that it does not represent a large shift in the energy of particles that become unbound. In other words, particles that are close to the threshold of being unbound receive relatively small boosts in kinetic energy, from the orbital energy of the two cores, pushing the particles just over the threshold. Although the exact cause is unclear, this resolution-dependent unbinding occurs approximately when the softening radius of the point masses is no longer resolved in the simulation, which happens when the local smoothing length of the gas particles exceeds the softening radius. This occurs at late times because of the decreasing gas density surrounding the remnant cores. \begin{table*} \begin{tabular}{x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm} x{1.2cm}} \hline $n_\text{part}$ & $a_\text{i}$ & $a_\text{f}$ & $t_\text{f}$ - $t_\text{i}$ & $\frac{\Delta a}{\Delta t}$ & $m_\text{u,i}$ & $m_\text{u,p}$ & $\frac{m_\text{u,p}}{m_\text{tot}}$ & $m_\text{u,i}$\textsuperscript{*} & $m_\text{u,p}$\textsuperscript{*} & $\frac{m_\text{u,p}\textsuperscript{*}}{m_\text{tot}}$\\ & / R$_\odot$ & / R$_\odot$ & / d & / R$_\odot$\,d$^{-1}$ & / {M$_\odot$} & / {M$_\odot$} & $\times 100$ & / {M$_\odot$} & / {M$_\odot$} & $\times 100$\\ \hline $7.6 \times 10^{4}$ & 130 & 30 & 276 & 0.36 & 0.045 & 0.157 & 32.2 & 0.053 & 0.160 & 32.9\\ $2.3 \times 10^{5}$ & 126 & 28 & 275 & 0.36 & 0.011 & 0.071 & 14.6 & 0.013 & 0.076 & 15.6\\ $1.1 \times 10^{6}$ & 122 & 28 & 263 & 0.36 & 0.001 & 0.050 & 10.3 & 0.002 & 0.054 & 11.1\\ $2.3 \times 10^{5}$\textsuperscript{\textdagger} & 129 & 31 & 354 & 0.28 & 0.046 & 0.092 & 18.9 & 0.048 & 0.097 & 19.9\\ \hline \multicolumn{7}{l}{\textsuperscript{\textdagger}\footnotesize{Simulation run with corotating primary.}}\\ \multicolumn{7}{l}{\textsuperscript{*}\footnotesize{Unbound criterion includes internal energy.}} \end{tabular} \caption{Comparison of orbital evolution quantities for different resolutions. Here we choose $a$ as the orbital separation, $m_\text{u}$ as the unbound mass, $m_\text{tot} = 0.49$\,{M$_\odot$}\ as the total gas mass in the simulation and $t$ as the time. The subscripts i and f refer to times at the beginning and end of the fast inspiral as defined in the text, respectively. The subscript p refers to the `plateau' value of the unbound mass, after the end of the fast inspiral. $\frac{\Delta a}{\Delta t}$ is the mean rate of descent during the fast inspiral.} \label{table:convergence} \end{table*} \subsection{Energy and Angular Momentum Conservation} \label{ssec:conservation} Our simulations employed a single, global timestep for all particles, which leads to excellent conservation of energy and angular momentum, at the cost of computational speed. Both physical quantities are conserved to better than 0.1 per cent, an improvement on prior SPH simulations \citep[e.g. the 1 per cent conservation observed in the simulations of ][]{passy2012simulating}. Figs~\ref{fig:1e6energy} and \ref{fig:1e6angmom} show the components of energy and angular momentum in our highest resolution simulation. Conservation is approximately 0.1 per cent and 0.3 per cent for our medium and low resolution simulations, respectively. In Figs~\ref{fig:1e6energy} and \ref{fig:1e6angmom}, we plot multiple quantities as a direct comparison to the equivalent figures in \citet{passy2012simulating}. For much of the simulation, none of the components vary significantly. Our primary core and companion particles do not possess internal energy because they interact with the SPH particles only gravitationally. Hence the total internal energy is summed over only the gas particles. All other quantities are found by summing components from gas and point mass particles as needed. Of note is that the envelope energy, calculated as the sum of the envelope potential energy, the total internal energy and the bound kinetic energy, begins negative and remains so for the entire simulation. The fact that by the end of the simulation the average total energy of the envelope is close to zero suggests that the envelope is only marginally bound. It is possible that with the help of part of the recombination energy budget this may lead to complete unbinding of the envelope. Fig.~\ref{fig:1e6angmom} shows only the z-components of the angular momentum (with respect to the centre of mass of the system at the beginning of the simulation, located at the origin, because the majority of rotational motion is in the orbital plane of the system). We plot the angular momentum of the bound and unbound gas, as well as the angular momentum of the two cores. There is a slight peak in the angular momentum curve of the bound mass. This is because, after this point, some mass becomes unbound. The excellent conservation gives an indication that the transfer of angular momentum is being handled correctly within our simulation. The analytical equation for the angular momentum of a binary system is \begin{equation} J = M_1 M_2 \left(\frac{G a}{M}\right)^{\frac{1}{2}}, \label{eq:j} \end{equation} where $M_1$, $M_2$, $M$, $a$ and $G$ are the primary and secondary star masses, the total mass of the system, the orbital separation and the gravitational constant, respectively. Using this, we calculate a total angular momentum budget of $3.87 \times 10^{52}$\,erg\,s. This is consistent with that calculated directly from the simulation. \begin{figure} \includegraphics[width=\columnwidth]{fig4} \caption{Evolution of the energy components in time. Here $K_\text{c}$ (solid blue) is the total kinetic energy of the two cores, $U_\text{tot}$ (dashed green) is the total internal energy of the gas, $K_\text{b}$ (dashed red) is the kinetic energy of bound gas, $K_\text{u}$ (purple) is the unbound kinetic energy, $E_\text{env}$ (dot-dashed black) is the envelope energy, $\phi_\text{env}$ (dashed blue) is the potential energy of the envelope, $E_\text{c}$ (dashed black) is the orbital energy of the cores, E$_\text{tot}$ (solid black) is the total energy of the system, $\phi_\text{c}$ (solid red) is the potential energy between the point mass particles and $\phi_\text{tot}$ (solid green) is the total potential energy.} \label{fig:1e6energy} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig5} \caption{Evolution of the angular momentum components in time. In this plot: $J_\text{b}$ (red line) is the angular momentum of bound gas, $J_\text{orb}$ (blue line) is the angular momentum of the point mass particles, $J_\text{u}$ (green line) is the angular momentum of unbound material, and $J_\text{tot}$ (black line) is the total angular momentum in the simulation.} \label{fig:1e6angmom} \end{figure} \subsection{The Corotating Reference Frame} \label{ssec:corotating} One of our two simulations with an intermediate resolution ($2.3 \times 10^{5}$ particles) was carried out in a frame that was rotating with the same angular velocity as the initial binary system. The primary star was stabilised in isolation in the inertial reference frame. Then it and the companion were placed in the corotating frame, with the gas particles set to be initially stationary in this frame. Ideally, the giant star should be stabilised alongside its companion in the corotating frame, and this step will be included in future simulations. The primary difference between use of a non-rotating primary star and one which is tidally locked with its companion is that the tidally locked star overfills its Roche lobe more. This would suggest that the initial mass transfer should be larger than for simulations with a non-rotating primary. This simulation progressed over a time-scale approximately double that of the equivalent inertial frame simulation. This is because gas particles begin the simulation with greater angular momentum, hence a gas particle requires less additional angular momentum to be ejected. Point mass particles then exchange less angular momentum with the gas as it escapes, resulting in a less rapid decrease in orbital separation. As the orbital separation decreases in the next phase, the relative velocity between the companion and the gas increases, but due to the initial rotation of the primary, the gas is dragged into corotation with the cores more rapidly than for the inertial frame simulation. Hence, the orbit enters the self-regulated inspiral at 31\,R$_\odot$, a slightly larger separation than 28\,R$_\odot$, measured from the equivalent, non-rotating simulation. The mean rate of inspiral is also slower in the corotating simulation (compare 0.28 R$_\odot$\,d$^{-1}$ and 0.36 R$_\odot$\,d$^{-1}$ in Table~\ref{table:convergence}). \section{Orbital evolution} \label{sec:orbitalevolution} Having discussed the numerical aspects of the simulation, we here describe in detail the phases of our simulation that were outlined in Section~\ref{ssec:overview}. \subsection{Roche lobe Overflow} At the onset of Roche lobe overflow, we expect the systems to conserve angular momentum, and for the mass transfer rate to be related to the amount by which the primary overfills its Roche lobe. \subsubsection{Orbital evolution time-scale} \label{sssec:orbitalevolutiontime-scale} The primary star is not stabilised in the potential of the companion, so the giant is not in perfect equilibrium at the start of the simulation. The distortions to the primary envelope that result are small, but likely lead to an orbital evolution that is too fast. In addition, the rate at which the Roche lobe overflow phase proceeds and how quickly the system enters a common envelope, depends on resolution (see Fig.~\ref{fig:218sepbound}). At our lowest resolution of approximately 80\,000 SPH particles, the Roche lobe overflow phase progresses in about half the time required for our high resolution simulation of 1.1 million particles. This by itself demonstrates that we cannot trust the simulated length of the Roche lobe overflow phase. This reaction to the new potential is also likely to be the cause of the initial mild eccentricity, visible in Fig.~\ref{fig:218sepbound}, that develops in the orbit early in the simulation. Mass transfer between the stars and mass loss from the $L_2$ and $L_3$ locations drives orbital shrinkage (see Eq.~\ref{eq:RLOF}), and it is likely affected by resolution, because the mass is discretised differently. Below we compare the inverse of the orbital time-scale, $\frac{\dot{a}}{a}$, calculated purely from the orbital separation of the two cores in the simulation with that derived from Eq.~\ref{eq:j} and Kepler's third law, \begin{figure} \includegraphics[width=\columnwidth]{fig6} \caption{Comparison of the simulated orbital evolution timescale (solid green line) for the 1.1 million SPH particle simulation, with the analytic time-scale computed from Eq.~\ref{eq:RLOF} (solid blue line). Also displayed lines computed with Eq.~\ref{eq:RLOF} where we have set $\dot{J} = 0$ (dashed red line) and $\dot{M_2} = - \dot{M_1}$ (dash-dotted red line).} \label{fig:RLOF} \end{figure} \begin{equation} \frac{\dot{a}}{a} = \frac{2 \dot{J}}{J} - \frac{2 \dot{M_1}}{M_1} - \frac{2 \dot{M_2}}{M_2} + \frac{\dot{M_1} + \dot{M_2}}{M_1 + M_2}, \label{eq:RLOF} \end{equation} \noindent where the symbols have the same meaning as in Eq.~\ref{eq:j}. In Eq.~\ref{eq:RLOF}, we use the mass contained in the primary and companion Roche lobes ($M_1$ and $M_2$, respectively). For the angular momentum $J$, we sum the contributions from the two core particles and the angular momentum of the gas inside the Roche lobes of the binary, all with respect to the centre of mass. During the interaction, some gas is expelled from the Roche lobes of the binary, forming the spiral density arms that are visible in Fig.~\ref{fig:slices}. The angular momentum carried off in the spiral waves drives the decrease in orbital separation. It should be noted that the separation should decrease even if $\dot{J} = 0$, as long as $M_1 > M_2$ and $\dot{M_2} = - \dot{M_1}$. The disk formed by this outflow is discussed in Section~\ref{sssec:edisk}. We measure the mass transfer rate in the equation, $\dot{M_1}$, by counting the SPH particles that have entered the Roche lobe of the companion star for the first time, and then dividing this mass by the time between two simulation outputs. The Roche lobe radius is updated at every code output to reflect the changes in orbital separation and stellar masses. This method of measuring mass transfer is an approximation because, for example, it does not take into account particles that enter the companion's Roche lobe but exit it before the next code output. The simulated and analytically derived time-scales, computed for the highest resolution simulation, compare to within 10 per cent (Fig.~\ref{fig:RLOF}, green and blue lines, respectively), although the numerical time-scale displays periodic behaviour because of the slight orbital eccentricity. This means that the relationship between mass transfer rate, angular momentum loss and orbital separation are in line with the analytical prescription. We further plot in Fig.~\ref{fig:RLOF}, Eq.~\ref{eq:RLOF}, but with $\dot{J} = 0$ (dashed red line), or with $\dot{M_2} = - \dot{M_1}$ (dash-dot red line). In the case where $\dot{J} = 0$, $\dot{a}/a$ is positive throughout, while if $\dot{M_2} = - \dot{M_1}$, $\dot{a}/a$ is mainly negative, and similar to the observed orbital decay. By separating the effect of mass transfer from the effect of angular momentum loss, we can deduce that it is primarily the loss of angular momentum via $L_2$ and $L_3$ that drives the decrease in orbital separation. After the end of the Roche lobe overflow phase, at around 4600\,d in the high resolution simulation, dynamic inspiral begins. At this time Eq.~\ref{eq:RLOF} can no longer be used to predict the time-scale accurately, because it becomes difficult to define what is meant by the masses of the stars and their ejecta when they are orbiting within a common envelope, so the time-scales do not match as closely as before the dynamic inspiral but still display the same basic shape. \subsubsection{The Mass Transfer Rate} \label{sssec:masstransfer} We next investigate the mass transfer during the Roche lobe overflow phase and the ability of simulations to reproduce it. The aim is to determine whether the simulations can indicate the duration of the unstable Roche lobe overflow phase, and ultimately lead to a better determination of the criterion of instability. For the simple case of a binary with a non-spinning donor, which can be modelled as a polytrope, and for which the mass transfer is perfectly conservative, unstable mass transfer occurs for $q \gtrsim \frac{2}{3}$ or 0.67 \citep{tout1991wind}. The mass ratio of our system is almost exactly $q = 0.68$, our primary star starts with no spin, and is fairly well approximated by a polytrope of index $n = 1.5$ (suitable for red giant stars). Our mass transfer is not conservative, though the mass lost is relatively low, only about 7 per cent over the entire Roche lobe overflow phase. It is therefore expected that our system should be unstable, as demonstrated by our simulations. In Fig.~\ref{fig:massflow} we plot the simulated mass transfer rate determined as described in Section~\ref{sssec:orbitalevolutiontime-scale}. The lines are drawn only to show the time span of interest, namely before the system enters the dynamic inspiral. As mentioned in Section~\ref{ssec:resolution}, this time is chosen as the first point when |$\frac{\dot{a}}{a}| \geq \frac{1}{15} \max |\frac{\dot{a}}{a}|$. After that time (about 4600\,d in the high resolution simulation) the Roche geometry is no longer applicable and measuring mass transfer from one star to another no longer makes sense. \begin{figure} \includegraphics[width=\columnwidth]{fig7} \caption{Mass transfer rate through $L_1$ as a function of time. As previously, blue, green and red lines represent the low, medium and high resolution simulations, respectively, while the purple line is used for the medium resolution simulation in a corotating frame. Main panel: solid lines are mass transfer rates measured from the simulations. Dashed lines represent the analytical form of the expected mass transfer rate given inputs from their respective simulations (see Eq.~\ref{eq:mdot}). Insert: same as the main panel, but for the first 300 days of the simulations.} \label{fig:massflow} \end{figure} As expected, the transfer of mass from the primary to the companion starts slowly (about $10^{22}$\,g\,s$^{-1} = 6.6 \times 10^{-6}$\,{M$_\odot$}\,yr$^{-1}$), and increases by two orders of magnitude as the orbital separation decreases. Because of the discretisation of mass in SPH simulations, the minimum (non-zero) rate of mass transfer is when one SPH particle passes $L_1$ between two consecutive code dumps (dumps are appoximately every 80\,000 seconds). This corresponds to mass transfer rates of $10^{23}$\,g\,s$^{-1}$ and $10^{22}$\,g\,s$^{-1}$, for the lowest and highest resolutions we adopt, respectively. The measured mass transfer rate curves in Fig.~\ref{fig:massflow} have been smoothed with a Savitzky-Golay filter, showing the underlying trends in the mass transfer measurements (which can otherwise be noisy when close to the minimum rates), so the plotted rate can be smaller than the minimum simulated rate. It is important to stress that this is a result of smoothing the curves, and no rate below the minimum is actually measured. We can only consider the measured mass transfer rates after a simulation time of 300\,d, when the transfer rates, at all resolutions, are larger than the minimum measurable. At a time of 300\,d mass transfer rates are quite similar for all resolutions, varying from $10^{23}$\,g\,s$^{-1}$ at our lowest resolution to $5\times10^{22}$\,g\,s$^{-1}$ at the highest resolution. A question is whether we can verify these mass transfer rates. An analytical expression for stable and conservative mass transfer rate though the $L_1$ point is that of J\k{e}drzejec \citep{paczynski1972evolution}, \begin{equation} \dot{M}_1 = - S_1 \left(\frac{\mu m_H}{k_\mathrm{B} T}\right)^{1.5}W(M_\text{1,rel})\rho G^2 M_1^2 \left(\frac{\Delta R}{R_1}\right)^3, \label{eq:mdot} \end{equation} \noindent where $S_1 \approx 0.215$ is a constant factor related to the polytropic index, $\mu$ is the mean molecular weight, $m_H$ is the mass of a hydrogen atom, $k_\mathrm{B}$ is Boltzmann's constant, $T$ is the temperature at the photosphere of the donor, $\rho$ is its density at the photosphere, $G$ is the gravitational constant, $M_1$ is the mass of the primary and $\Delta R = R_1 - R_\text{L,1}$, where $R_1$ is the radius of the primary and $R_\text{L,1}$ is its Roche lobe radius. The quantity $W(M_\text{1,rel})$ is a function of $M_\text{1,rel}$ = $M_1$/($M_1$ + $M_2$), defined as \begin{equation} W(M_\text{1,rel}) = \frac{\sqrt{M_\text{1,rel}}\sqrt{1-M_\text{1,rel}}}{(\sqrt{M_\text{1,rel}} + \sqrt{1-M_\text{1,rel}})^4}\left(\frac{a M_\text{1,rel}}{R_\text{L,1}}\right)^{n+1.5}, \end{equation} \noindent where $a$ is the orbital separation and $n$ is the polytropic index, taken to be $1.5$ for a red giant. A similar derivation for $\dot{M}_1$ with an isothermal flow was made by \citet{ritter1988turning}, although we use Eq.~\ref{eq:mdot} because it is valid for adiabatic flows. The mass transfer rate Eq.~\ref{eq:mdot} applies to giants transferring mass conservatively. It assumes a system with a corotating primary star and that, as the mass flows though the $L_1$ point, it transitions from subsonic to supersonic. The first of these conditions is only true for one of our simulations. However, all of the simulations do have a sonic point between the two stars. With this in mind, we apply it to our simulations, but acknowledge that there are some discrepancies resulting from the non-rotating primary. To determine the mass transfer rate from Eq.~\ref{eq:mdot}, we used inputs from the initial stellar profile, specifically $\mu = 1.34$, $T = 3500$\,K and $\rho = 7\times 10^{-9}$\,g\,cm$^{-3}$. These were kept constant over time owing to difficulties in measuring them from the simulation, generating uncertainty. When the star is distorted by the presence of the companion, it is not trivial to determine its radius. This difficulty is compounded by material filling the orbit after having been stripped from the primary during Roche lobe overflow. Hence, in this case we simply determine the average radius of the stellar gas distribution whose surface has the approximate density of the stellar photosphere at the start of the simulation. The analytical form of the mass transfer rate (Eq.~\ref{eq:mdot}) is compared with the measured numerical rate in Fig.~\ref{fig:massflow} for the high and intermediate resolution simulations (red and green lines) and for the simulation carried out in the corotating frame (purple line), but our discussion below is valid for the other simulations as well. The analytical mass transfer rate is slightly lower than the numerical rate at 300\,d, our fiducial start of the simulation as described above and the difference is smaller for higher resolutions. Both the simulation and the analytical approximation indicate that the mass transfer rate should increase in time, although the numerical rate starts larger but ends smaller than the analytical one. Finally, the analytical mass transfer rate is about four times larger than the numerical rate at the end of the higher resolution simulation and three times larger at the end of both the inertial and corotating frame, intermediate resolution simulations. Considering the imprecision with which we define the stellar radius and the third power dependence of the mass transfer rate on that number, the discrepancy here is relatively small. \subsubsection{The Decretion Disk} \label{sssec:edisk} \begin{figure*} \includegraphics[width=\linewidth]{fig8} \caption{Cross sections of density (left) and velocity (right) in the orbital plane of the excretion disk at the beginning of the dynamic inspiral, at 12.8\,yr, for the 1.1 million particle simulation. These frames are taken at the same time as the third row of Fig.~\ref{fig:slices} (12.8\,yr), but are zoomed out such that the frames are 70\,AU per side. The green dots represent the cores.} \label{fig:disk} \end{figure*} As was mentioned in Section~\ref{sssec:orbitalevolutiontime-scale}, gas is ejected during Roche lobe overflow in the form of spiral density waves originating at $L_2$, and later at $L_3$. From our high resolution simulation we estimate that over the course of the Roche lobe overflow phase, these waves carry approximately $9 \times 10^{51}$\,erg\,s of angular momentum away from the binary (23 per cent of the total angular momentum budget of $3.87 \times 10^{52}$\,erg\,s) and about $7.8 \times 10^{-2}$\,{M$_\odot$}, driving the decrease in the orbital separation. The decretion disk is contained within approximately 5000\,R$_\odot$\ from the centre of the binary (see Fig.~\ref{fig:disk}). Within the disk, the range of smoothing lengths $h$ is about 30 to 1000\,R$_\odot$. Approximately $1.5 \times 10^{-3}$\,{M$_\odot$}\ of gas lies outside this radius and contains gas unbound from the system within the first year of the simulation. \citet{tocknell2014constraints} performed a calculation (their Section~5), using data from the simulations of \citet{passy2012simulating}, estimating the final radius of a disk of material that would be formed by infalling bound gas at the end of the common envelope. Their estimate considered gas moving in ballistic trajectories, and used conservation of energy and angular momentum to determine the radius and time-scale of formation. Using the quantities in the previous paragraph, we calculate a specific angular momentum for the disk of $5.8 \times 10^{19}$\,erg\,s\,g$^{-1}$. Hence, following the procedure of \citet{tocknell2014constraints}, we find that this bound disk would fall back to a radius of about 130\,R$_\odot$\ from the centre of mass. This is well outside the final orbital separation of the system. The time taken for this trajectory would be at least 20\,yr, though this assumes that the bound disk gas has no left over outward motion, making the timescale a lower limit. However, in this simulation this structure is unlikely to survive for long enough, because a dense, fast outburst occurs at the beginning of the dynamic inspiral and quickly overtakes the expanding disk. We discuss in Section~\ref{ssec:gasdistribution} the bipolar gas distribution at the end of our simulation. \subsection{The Dynamical Inspiral Phase} \label{ssec:inspiral} At some stage in a common envelope interaction, the orbital separation starts to decrease at a faster rate. This occurs in our high resolution simulation at approximately 4600\,d, as indicated by the time-scale of the system (Fig.~\ref{fig:RLOF}). At this point, the mass transfer is no longer characterised by the mechanism of Roche lobe overflow because the system is no longer described accurately by the Roche geometry. As the companion splashes into the atmosphere of the primary it unbinds a shell of material which travels outwards at speeds of about 40\,km\,s$^{-1}$. This is visible in Fig.~\ref{fig:slices} at 13.3\,yr, and later on as it has travelled outwards at 13.8\,yr, located at approximately 1000\,R$_\odot$\ from the binary at the centre. We can also see in the perpendicular plane (far right column) of Fig.~\ref{fig:slices} at 13.8\,yr how the material that expands above and below the orbital plane is faster, while the material in the orbital plane encounters the torus and decelerates. This effect is similar to what is observed in the simulations of \citet{macleod2018bound}. The binary has entered a common envelope and the inspiral is now driven by gravitational and hydrodynamic drag forces. Gravitational drag operates purely through gravitational interactions between the gas and the companion, while hydrodynamic drag is instead the force felt by a body owing to direct collisions between the body and the particles of the ambient medium through which it is moving. Hydrodynamic drag tends to be insignificant in typical binary common envelope interactions, while for lighter companions, such as planets, it can be relatively more important due to a lower intensity of the gravitational drag \citep{staff2016hydrodynamic}. We can only simulate gravitational drag, because the point mass particles have no physical size. This said, gas tends to collect in the shallow potential wells around point masses and travels with them, something that does effectively give them a radius. However, any collisions between the ambient medium and the material in the potential wells still only affects the cores via gravitational interactions. While one could expect that the strength of the drag is affected by resolution, it has been shown previously (e.g.,~\citealt{passy2012simulating,iaconi2017effect}) that it is only marginally so. This means that resolution does not dramatically affect the final separation or the speed of the inspiral (we can appreciate this by looking at Table~\ref{table:convergence}). \citet{staff2016hydrodynamic} showed that in the case of their interaction with a planetary mass companion, the strength of the gravitational drag was commensurate with that determined by an analytical approximation. Here we carry out a similar comparison for our stellar-mass companion. Below we compare the drag calculated in three distinct ways, two from the simulations directly and one using an analytical prescription together with simulation quantities, as done before. The physical parameters that influence the gravitational drag are the mass of the embedded body, the density of the surrounding fluid and the velocity contrast of the body and the fluid. In a common envelope, however, there may be additional factors playing a role, such as the density gradient \citep{macleod2017common}. \begin{figure} \includegraphics[width=\columnwidth]{fig9} \caption{Gravitational drag force on the companion in the 1.1 million particle simulation, calculated with different approximations: summing interactions with nearby gas particles (blue line), by the observed loss of angular momentum from the orbit (green line; Eq.~\ref{eq:jdotdrag}) and by the analytical form of the drag force (red line; Eq.~\ref{eq:ana_drag}). The orbital separation is plotted in black. The inset is a zoom out over the entire simulation, for reference. Dashed black lines in the inset show the time range of the main plot, which is the time range during which the system is undergoing fast inspiral.} \label{fig:drag} \end{figure} The gravitational drag has been represented by several authors, with some variability of form \citep[e.g. see][]{shima1985hydrodynamic}, as \begin{equation} F_\text{drag} \approx \xi \pi R_\text{a}^2 \rho v_\text{rel}^2, \label{eq:ana_drag} \end{equation} where $\xi$ is a factor dependent on the Mach number, $R_\text{a}=2GM_2/(v_\text{rel}^2+c_s^2)$ is the accretion radius\footnote{We use the form adopted for subsonic motion, typically observed in common envelope inspirals.}, $M_2$ is the mass of the companion, $c_s$ is the local sound speed, $\rho$ is the density around the core and $v_\text{rel}$ is the relative velocity of the core and the surrounding gas. The parameter $\xi$ is greater than 2 for supersonic motion and less than 1 for subsonic motion \citep{shima1985hydrodynamic} but, owing to difficulties in calculating this quantity on the fly, it is typically just held at unity. We use Eq.~\ref{eq:ana_drag} to verify our simulation in the knowledge that the expression itself may have only limited validity in our situation. To use this expression we need to derive from the simulations the relative velocity of the companion and the density of surrounding material. First, we find the mean velocity vector of all gas within a given radius from the companion. In this case, the radius we use is the distance from the point mass particle to the centre of mass of the orbit. This varies from about 100\,R$_\odot$\ to about 10\,R$_\odot$\ as the orbit shrinks. The mean velocity vector of the gas is then projected onto the velocity vector of the companion, and the difference is found. The average density is similarly found within the same radius. An alternative method is to use a constant radius. This cannot be too small or else only the material caught in the potential well is encapsulated and cannot be too large or else in the late inspiral the core of the primary is included. \citet{iaconi2017effect} use a static radius of 20\,R$_\odot$. Both methods have their benefits but, within the timespan of the dynamic inspiral, we find our method is more stable. Any method which averages quantities over a volume around the companion includes substantial stationary gas caught in its potential well. This decreases the average velocity contrast, while it increases the density. Further, the accretion radius used in Eq.~\ref{eq:ana_drag} is also calculated with the the mean relative velocity and sound speed and is quite sensitive to slight changes in these quantities. Finally, Eq.~\ref{eq:ana_drag} is meant for use in a homogeneous medium, which is not the case in our simulations. In a common envelope, there is a relatively strong density gradient perpendicular to the motion of the core, along with an overdensity caught in its potential well, so the output of Eq.~\ref{eq:ana_drag} should be taken to be indicative at best. There are two ways to calculate the drag force directly from the calculations. The first is via the change in angular momentum $\dot{J}$ of the point mass particles, with respect to the centre of the orbit by \begin{equation} F_\text{drag} = \frac{\dot{J}}{r}, \label{eq:jdotdrag} \end{equation} where $r$ is the distance of the companion from the centre of mass. It should be noted that this only takes into account the component of the drag which is perpendicular to the radius from the centre of the orbit. Hence, while it is a reasonable approximation for circular orbits, if the cores are on an elliptical orbit this equation would yield a lower limit. \begin{figure*} \includegraphics[width=\linewidth]{fig10}\\ \caption{Slices in the XY-plane of the 1.1 million particle simulation, rendering the quantity $\Delta\omega/\omega_\text{bin}$, where $\Delta\omega \equiv \omega_\text{bin} - \omega_\text{gas}$, and $\omega_\text{bin}$ and $\omega_\text{gas}$ are the angular velocity of the point masses and the gas, respectively. The two snapshots are taken when the system is experiencing the most rapid reduction in orbital separation (left), and at the end of the dynamic inspiral (right). The frames have a side length of 1.5\,AU. Black dots represent the point masses.} \label{fig:corotation} \end{figure*} The second way to determine the dynamical friction in a simulation is to directly calculate the forces experienced by the cores as a result of other particles in the simulation. It is not useful to sum up the contributions from all particles, because the gravitational forces are dominated by the high density region surrounding the core of the giant. We select out a certain volume around the companion and determine the local gravitational attraction in this way. If the gas around the companion were spherically symmetric then there would be no net force on the particle. However, owing to the motion of the particle through the gas we see that the symmetry is broken by a high density wake that forms behind the companion. To calculate this force, we add up the gravitational forces between the companion and gas particles residing within the radius from the companion to the centre of mass of the orbit. This total force is then projected onto the velocity vector of the companion to give the dynamical friction of the gas on the companion particle. The analytical and two numerical forces are plotted in Fig.~\ref{fig:drag}, where the force is negative as it acts in the opposite direction to the orbital motion. In Fig.~\ref{fig:drag} we can see that the two numerical methods for calculating the drag force are comparable to within a factor of two. We recall that the torque calculation is a lower limit. However, the analytical force displays some differences though the peak magnitude remains within a factor of two of the blue line. Some variation can be obtained by changing the volume within which we measure the averaged quantities. The period of apparent strong analytical gravitational drag before the inspiral, seen in the inset of Fig.~\ref{fig:drag}, is an artefact of averaging over a large volume. The opposite effect occurs when the volume shrinks too much, because the force becomes entirely dominated by contributions of particles stuck in the potential wells of the cores, sending the drag force to zero earlier than it should. Also worth mentioning is the parameter $\xi$, which is dependent on the Mach number, and may vary the expected drag force by a factor of order unity. \subsection{The Orbital Stabilisation Phase} \label{ssec:theorbitalstabilisation} The orbit of the two cores stabilises at around 20\,R$_\odot$, independent of resolution. By the end of the high resolution simulation, the rate of decrease of the orbit has slowed down to about 0.5\,R$_\odot$\,yr$^{-1}$, though this descent is still decelerating. In Fig.~\ref{fig:drag} we see that at some point there is a significant reduction of the drag force. The cause of this lower force is not so much the lower density of the surrounding gas, but that the gas has been dragged into corotation. Fig.~\ref{fig:corotation} shows two snapshots from the high resolution simulation, showing two cross-section slices of the quantity $(\omega_\text{gas} - \omega_\text{bin})/{\omega_\text{bin}}$, where $\omega_\text{gas}$ and $\omega_\text{bin}$ are the angular velocity of the gas and of the point masses around the centre of mass, respectively. The left-hand panel shows the time when $|\frac{\dot{a}}{a}|$ is a maximum, corresponding to the maximum rate of inspiral, while the right-hand panel shows the same quantities about a few months later, at the end of the dynamic inspiral. The gas surrounding the cores is rotating more uniformly in the second panel and is in approximate corotation, preventing the cores from losing further angular momentum to the envelope gas. This feature can also be seen in Fig.~\ref{fig:1e6angmom}, in which the angular momentum carried by the point mass particles remains approximately constant after the end of the dynamic inspiral. What follows is complex and depends on how much of the envelope is unbound. Stellar structure dictates that for a giant star to depart from the giant branch and shrink, all but a very small amount of envelope must be lost. If the envelope is merely lifted by the common envelope inspiral, but not unbound, the orbit may stabilise while it has little gas within it. We could expect that if the envelope is ejected and the star departs the giant branch shrinking to within its Roche lobe, a small part of the envelope that was lifted but not ejected would return to form a low mass circumbinary disk. However, if the majority of the lifted envelope is not ejected, then a substantial amount of mass would return to the binary, resulting in a further interaction. The time-scale of this return could be very small, though the ballistic calculation of \citet{kuruwita2016considerations} is only a lower limit because it does not account for gas pressure. Ultimately this question cannot be answered while we do not have a definitive answer to the energetics of the envelope and when it does or does not become unbound. \subsection{Common Envelope Fallback} \label{ssec:fallback} If most of the envelope is still bound after orbital stabilisation, we expect it to fall back rapidly. Simulations that use recombination energy tend to unbind much more of the envelope than those, such as ours, that do not \citep{nandez2015recombination}. It is likely that, had these simulations included recombination energy, they would be placed amongst the simulations that succeed in fully unbinding the entire envelope \citep{iaconi2018effect}. However, even leaving aside the controversy of how much recombination energy can be used to unbind the envelope \citep{ivanova2018use, soker2018radiating}, there are situations where the envelope would not be fully unbound even with recombination energy, such as those with initial masses $\gtrapprox 2$\,{M$_\odot$}\ \citep{nandez2016common}. In these cases substantial mass would fall back so that it is worth expending a few words on common envelope fallback gas. \citet{kuruwita2016considerations} investigated how the bound part of the common envelope returns and whether it should form a disk thanks to the added angular momentum. They could not follow the entire outward and return journey of the gas, but using their setup could deduce that gas would return in a matter of months to a few years, and that much of this gas would return all the way to the core. However, their calculation did not account for the fact that gas, upon returning, collides with outflowing gas. Since SPH allows us to track envelope material far from the central binary, we can investigate the dynamics of this gas over a longer time-scale. This allows us to take another look at the idea of an envelope fallback event, beyond what was done by \citet{kuruwita2016considerations}. We consider SPH particles to be falling back if they have a negative radial velocity component with respect to the centre of mass and if they are located outside of a spherical volume encompassing the two stars. The total mass (blue line), angular momentum (red line), and the mean radial velocity (green line) of these particles are plotted in Fig.~\ref{fig:fallback}. Before the dynamic inspiral, there is very little material meeting our fallback criterion, as expected. During the inspiral, material is very rapidly evacuated from the region around the binary, catching up and colliding with material that was ejected previously. However, at the end of inspiral this evacuation stops and the material is able to fall back onto the central binary. The distribution of fallback mass is shown in Fig.~\ref{fig:fallbackslice}, and can be seen to be approximately circular at larger radii. None of this fallback material returns to interact directly with the binary. Instead it impacts with more gas, flowing out from the region around the binary. It is possible that the region immediately surrounding the binary would become almost completely evacuated at some point, allowing future fallback material to return to the very inner portions of the system. However, we have not been able to follow the simulation for long enough to see this. \begin{figure} \includegraphics[width=\columnwidth]{fig11} \caption{Fallback mass (blue), the average radial velocity of this mass (green), and the angular momentum tied up in this mass (red) as functions of time.} \label{fig:fallback} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{fig12} \caption{Distribution of gas displaying a negative radial velocity at t = 5188\,d for the 1.1 million particle simulation, when the most fallback mass was present in the simulation. Material is rendered by its \textit{negative} radial velocity value. Material with a positive radial velocity is not rendered. The frame is 7\,AU on each side.} \label{fig:fallbackslice} \end{figure} \section{Planetary Nebulae and the Shape of the Common Envelope} \label{sec:planetarynebulae} At least one in five planetary nebulae (PNe) are ejected common envelopes \citep{miszalski2009binary}, and the symmetry axis of post-CE PNe with well characterised binaries can be shown to align with the orbital axis \citep{hillwig2016observational}. The shapes of the approximately 40 known post-CE PNe are interestingly inhomogeneous, although some common features, such as bipolarity, are observed in many cases \citep{de2013binary}. However, there is ultimately no obvious set of features that identifies the origin of a given PN as from a common envelope interaction. PNe derive their shapes from the interaction of the fast and tenuous post-asymptotic giant branch (AGB) wind with a previous slow and massive AGB wind \citep{balick2002shapes}. The AGB slow wind creates a mould into which the fast wind expands. The common envelope provides gas distributions with an equatorial density enhancement. These are relatively small and dense compared to those generated by a slower mass-loss process. It is useful to determine the key differences between PN generated from post-CE distributions and those generated by single stars or other binary phenomena. \citet{garcia2018common} have recently demonstrated how a common envelope density distribution, generated by the simulation of \citet{ricker2012amr}, can lead to a bipolar PN. They mapped the gas distribution at the end of common envelope simulation of \citet{ricker2012amr} into a 2D plane, where only a quarter of the domain was simulated assuming two axes of symmetry, one along the orbital plane and one perpendicular to it. The input gas distribution from the common envelope simulation was taken after 56.7\,d of common envelope evolution, when the binary separation between the 0.36\,{M$_\odot$}\ degenerate core and 0.6\,{M$_\odot$}\ companion was approximately 9\,R$_\odot$. A further 0.69~{M$_\odot$}\ of gas mass is in the ejected envelope. \citet{garcia2018common} modelled the central binary as a point with an effective temperature of 29\,000\,K, corresponding to a star with a wind velocity of 600\,km\,s$^{-1}$ (though they also performed two simulations with faster and slower winds). By evolving their simulation for $10^4$\,yr, they demonstrated that forming a bipolar nebula is, as predicted, a natural consequence of blowing a spherical wind into a toroidal gas distribution. Their simulations offer compelling evidence that connecting the common envelope with bipolar nebulae is a path certainly worth further study. \citet{frank2018planetary} used instead the simulation described here as a starting point for a PN simulation. Below we give further information as to the shape of the common envelope into which the post-AGB fast wind expands and compare their work further to that of \citet{garcia2018common}. \subsection{The common envelope's gas distribution at the onset of the post-AGB fast wind} \label{ssec:gasdistribution} \begin{figure*} \includegraphics[width=0.75\linewidth]{fig13}\\ \caption{All frames are slices in the XZ-plane, rendered in density (left column) and velocity (right column). Top and middle rows display the our 1.1 million particle simulation from this paper, and differ by the scale: 15\,AU per side for the top row and 150\,AU for the middle row. The bottom row displays a simulation started with a smaller orbital separation of 100\,R$_\odot$, but that is otherwise identical, and has side length of 150\,AU to match the middle row. Note the different scales of the colour bars.} \label{fig:frames} \end{figure*} The binary star considered by our simulation differs from that of \citet{ricker2012amr}. Our primary star is a 0.88\,{M$_\odot$}\ giant with a 93\,R$_\odot$\ radius, while they have a 1.05\,{M$_\odot$}\ giant with a 31\,R$_\odot$\ radius. Therefore their giant, is more bound and is more similar to a lower RGB star than the more extended AGB star that forms a PN. The more bound nature of their giant leads to a faster evolution over all and a more compact, denser common envelope ejection. Additionally, we evolved our simulation for longer, particularly before the inspiral. Despite the fact that both simulations start the companion near the beginning of Roche lobe overflow, our simulation spends a much longer time before inspiral, and more gas is launched out of the $L_2$ and $L_3$ points before the common envelope ejection takes place. This provides us with a more extended common envelope gas distribution into which the post-AGB wind expands. In addition, the shape of the extended common envelope is greatly influenced by the interaction of the Roche lobe overflow with the outflow from the common envelope inspiral. During the Roche lobe overflow phase, gas leaves the system primarily through the second and third Lagrange points (see Fig.~\ref{fig:slices}). There is almost no unbinding of material in this phase (Fig.~\ref{fig:218sepbound}). The following fast inspiral phase interacts with the Roche lobe overflow ejecta. This distribution, shown both in the final row of Fig.~\ref{fig:slices} and in the top and middle frames of Fig.~\ref{fig:frames}, is toroidal. The polar regions in the distribution are approximately one to two orders of magnitude less dense than the material immediately surrounding it on the equatorial plane. The velocities of the ejecta are also considerably greater in a direction perpendicular to the orbital plane. Hence the relative difference in density continues to increase as the distribution expands. The bottom row of Fig.~\ref{fig:frames} shows a slice of a simulation identical to those we have presented here, except that the initial orbital separation was set to 100\,R$_\odot$. For more details on this simulation, see \citet{iaconi2017effect}. All of the slices in Fig.~\ref{fig:frames} are produced at 1000\,d after the beginning of the fast inspiral in their respective simulations. By comparing the middle and bottom row panels, we see that the gas distributions depend heavily on whether the pre-inspiral phase was simulated or not. Specifically, the Roche lobe overflow phase appears to promote symmetry in the final distribution. The bottom row frames show that there is a large lobe to the right of the frame which is rapidly expanding. There is a hint of the funnel regions of low density above and below the orbital plane but it is certainly not as distinct as the middle panels. {\it We suggest that the amount of mass loss before the inspiral makes a difference to the shape of the ensuing planetary nebula.} We therefore predict that differences in the nebular shape are the results of more or less stable pre-common envelope mass transfer, likely induced by a larger or smaller companion mass. It is worth noting that some of the morphological differences between the middle and bottom panels of Fig.~\ref{fig:frames} may be a result of having a non-rotating primary, which could plausibly lead to greater asymmetry in the bottom panels. \subsection{The SPH-AMR PN Simulation} In our high resolution simulation, all particles have an equal mass of $4.54 \times 10^{-7}$\,{M$_\odot$}. Typical post-AGB winds release mass on the order of $10^{-8}$\,{M$_\odot$}\,yr$^{-1}$. As \textsc{Phantom} currently does not support unequal mass particles, we are unable to model both the common envelope and the spherical wind in the same simulation. The 3D post-CE gas distribution has therefore been mapped into a series of nested grids using \textsc{Splash} \citep{price2007splash} and used as input to the 3D AMR grid code AstroBEAR \citep{cunningham2011astrobear}. Preliminary results of this PN simulation have been presented by \citet{frank2018planetary}. In summary, the AstroBEAR domain is 16\,000\,R$_\odot$$^2$\ $\times$ 128\,000\,R$_\odot$\ in size, mapped with seven levels of refinement. Two spherical wind cases are simulated. In one, a spherical wind with a mass-loss rate of $6.4 \times 10^{-4}$\,{M$_\odot$}\,yr$^{-1}$ is injected after a quiescent period of 500\,d, while in the other a $3.2\times 10^{-5}$\,{M$_\odot$}\,yr$^{-1}$ mass loss rate spherical wind is injected after a quiescent period of 6000\,d. Both winds have speeds of 300\,km\,s$^{-1}$. The simulations are each only run for approximately 1000\,d after the injection of the wind, which is much less than for the simulations of \citet{garcia2018common}. The wind is injected into the simulation at a radius of 46.9\,R$_\odot$. A 1\,{M$_\odot$} particle is placed at the core and replaces the binary. This mass is also the only source of gravity in the simulation; no self-gravity is applied. Strong hydrodynamic collimation occurs in both these simulations, as well as in the simulation of \citet{garcia2018common}, due to the narrow, evacuated funnel that can be observed in Fig.~\ref{fig:frames}, upper left frame. We also note that the simulations of \citet{frank2018planetary} with the lower wind momentum display a strong degree of asymmetry between the upper and lower collimated lobes, with one much smaller than the other. Needless to say, these are early days for hybrid common envelope PN simulations such as these, especially because neither of the two sets of simulations have used a CE gas distribution that is unbound. This means that the envelope velocity field is likely not properly reproduced. \section{Summary and Conclusions} \label{sec:conclusions} We have presented a set of hydrodynamic simulations of common envelope evolution aimed at extending the purely dynamical inspiral phase to include the previous, Roche lobe overflow phase as well as part of the post-inspiral phase. We also considered the post-common envelope, nebular phase. These simulations were carried out with the newly released smoothed particle hydrodynamics code, \textsc{phantom}, and are based on an initial model of a 0.88\,{M$_\odot$}, RGB primary star with a radius of about 90\,R$_\odot$, and a 0.6\,{M$_\odot$}\ companion \citep{passy2012simulating}, placed at an initial orbital separation of 218\,R$_\odot$, which is tuned to trigger Roche lobe overflow. Our main results are the following: \begin{enumerate} \item Our {\it unstable Roche lobe overflow} lasts for a short time (between 6.4 and 12.7\,yr, depending on resolution and 18\,yr if the giant is corotating with the orbit). From trends with primary star rotation, stability and resolution, we conclude that, in nature, this phase should last longer. Although we are unable to ascertain the length of the time between the onset of Roche lobe overflow and the fast inspiral with any certainty, we exclude that this phase is longer than a few centuries for binary parameters similar to those we have modelled. \item By examining the ejection of mass and angular momentum, we are able to see that the orbital evolution of the system responds in a way that is expected analytically. The mass transfer rate during the Roche lobe overflow phase is aligned with analytical predictions to within a factor of a few. This gives confidence that improvements in the simulations may successfully and reliably model the Roche lobe overflow phase preceding a common envelope and soon lead to a better understanding of the stability criteria. \item The {\it flow ejected via the $L_2$ and $L_3$ Lagrangian points} remains mostly bound: only 10$^{-3}$\,{M$_\odot$}\ are unbound with a further $7.8 \times 10^{-2}$\,{M$_\odot$}\ remaining bound. If unobstructed, fallback of this material would form a disk at a radius of about 130\,R$_\odot$\ on a time-scale of at least 20\,yr. In our case, the unbound outflow associated with the subsequent dynamic inspiral disrupts this structure, in a manner similarly observed by \citet{macleod2018bound}. However, it is possible that, if the $L_2$/$L_3$ outflow phase were longer (mass transfer more stable), it may be followed by a weaker common envelope ejection, so that the disk structure formed during Roche lobe overflow may remain to become a long-lived circumbinary disk, as has been observed around post-AGB binaries with periods between 100 and 1500\,d \citep{vanwinckel2009post}. \item \citet{pejcha2016binary} simulated an $L_2$ flow in a binary that was not allowed to change its orbital separation. While they concluded that the $L_2$ outflow becomes unbound for mass ratios (accretor and donor) in the range 0.06--0.8, this is at odds with both the results of our simulations and those of \citet{macleod2018bound}. \item The {\it gravitational drag felt by the cores during the fast inspiral} is within a factor of two of that calculated analytically using quantities from the simulation, as was also concluded for planetary mass companions by \citet{staff2016hydrodynamic}. This consistency exists despite the many approximations and assumptions and it increases our confidence in the inspiral timescale being a reasonable approximation of reality. The {\it end of the inspiral} is brought about by gas being brought into co-rotation with the binary orbital motion. \item The low, medium and high resolution {\it simulations unbind 33, 16 and 11 per cent of the gas mass, respectively} (calculated without including internal energy). Hence there is a trend to unbind less mass as resolution increases \citep{iaconi2017effect}. Without having stabilised the star in the corotating frame, we are unable to definitively state what effect a rotating star has on the amount of unbound mass. \item During the early stages of the {\it self-regulated inspiral} following the dynamic inspiral, as much as 30 per cent of the gas displays negative radial velocities. This fallback material is typically within 500\,R$_\odot$\ of the central binary though, if the simulation were to be run for longer, material at larger radii would start to fall back. \item The final gas distribution of our simulation is toroidal, displaying a strong density contrast between the orbital plane and the polar directions. The common envelope ejection following the disk formation develops high velocities in the direction perpendicular to the orbital plane. However, this perpendicular ejection contains relatively little mass. It is not this mass that becomes visible as a bipolar planetary nebula later on. The highly evacuated polar funnels provides strong density contrast and powerful hydrodynamic collimation, which can be seen in the nebular simulations of \citet{frank2018planetary}, carried out with the output of this simulation. \item By comparing our simulation to an identical one, except for the fact that the companion was placed at the giant's surface, and in which the dynamic inspiral develops immediately, suggests that simulating the Roche lobe overflow phase results in a far more symmetric distribution of gas, due to the regulating effect of the disk on the subsequent morphology. This is something that would impact the shape of subsequent PNe. This in turn suggests that the length of the Roche lobe overflow phase or, in other words, {\it the degree of stability of the mass transfer, may explain the range of PN shapes of common envelope PN.} \end{enumerate} \section*{Acknowledgements} TR acknowledges financial support through the Macquarie University Research Excellence scholarship associated with Future Fellowship grant to O. De Marco (FT120100452). OD and DP acknowledge financial support through the Australian Research Council Future Fellowship scheme (FT120100452 and FT130100034, respectively). CAT thanks Churchill College for his fellowship and Macquarie University for supporting a visit. This work was supported by access to the swinSTAR supercomputing facility at Swinburne University of Technology. \bibliographystyle{mnras}

d260ca0e7cadcf119be29a7cb6bb8ec2042941b8

\section{Introduction} Magneto-transport in two-dimensional (2D) electronic ga\-ses at low temperatures presents ubiquitous features that are observed in vastly different classes of systems, from semiconducting heterostructures~\cite{RMPklitzing}, to graphene~\cite{Novo,Zhang,QHEgraphene} and other carbon based materials, oxide interfaces~\cite{ReviewMnZnO}, and topological systems~\cite{QHEtopo}. While details in the transport characteristics will strongly depend on the peculiarities of a given material (for instance the sign of magnetoconductance variation at low field, or the value of Hall conductance quantization plateaus), Landau level formation (or its counterpart at decreasing field, Landau level breakdown) appears as a very generic phenomenon. Indeed, Landau levels start to be witnessed only from an intermediate magnetic field regime, in which mild oscillation of thermodynamic and transport coefficients are observed. Only in a second range~\cite{Fogler,Floser} of even higher magnetic field does full quantization of the Hall conductance fi\-nally emer\-ge, with the Landau level index becoming a good quantum number. In order to explain these observations, the electronic motion in a perpendicular magnetic field and subject to confining or disordered electrostatic potentials has been thoroughly studied by many different theoretical methods \cite{Laughlin,Halperin,Streda,Thouless1982,Thouless1984,MacDo,Prange,Buttiker1988,Janssen,Huckestein1995,Dmitriev}. One clear limitation of current theories is their inability to comprehend all regimes of magnetic field (from low to intermediate and high) in a unified way, so that the question of the sharpness of the transitions between each regime is not easily established on general grounds. Our goal in this paper is to propose an exactly solvable model of Landau quantization breakdown that exhibits clearly two sharp transitions. This simple model is based on a non-rotationally invariant and inverted 2D parabolic potential, that we exactly solve in two dimensions for all values of the perpendicular magnetic field. Two mathematically related but physically distinct quadratic potential models are already well-known from the literature. The first model~\cite{Fock1928,Darwin1931}, often used to describe quantum dots, considers a fully confining 2D parabolic potential, and was solved around the same time as the Landau states \cite{Landau1930} for the free motion problem. This solution led to the Fock-Darwin eigenstates ~\cite{Fock1928,Darwin1931}, showing a discrete energy spectrum for all magnetic field values. Indeed, the effect of finite magnetic field amounts to redefine the quantum states while renormalizing the harmonic spectrum. The second model, relevant for quantum point contacts, was proposed and solved decades later by Fer\-tig and Hal\-pe\-rin~\cite{Fertig1987}, who considered quantum motion in a quadratic saddle point potential. The mathematical solution is here more involved due to the use of scattering states in a potential that is unbounded from below. Again, this model presents the same feature that the physics is weakly dependent on magnetic field, with tunneling being mostly renormalized by cyclotron motion~\cite{Buttiker}. In contrast, we will find that the inverted non-rotationally invariant quadratic po\-ten\-tial, which is re\-le\-vant to des\-cri\-be anti-dots or Coulomb impurities at a local level, displays markedly different electronic states from the low to intermediate and finally high magnetic field regimes. While this 2D inverted parabolic potential model could be solved by wavefunction techniques (adapted to each specific magnetic field range), or using more general path integral approaches~\cite{Jain1988,Entelis1992,Kagalovsky1996,Tochishita1996}, we propose here an analytic and unified phase space solution, that naturally encompasses all field ranges. As a matter of fact, this single solution also accounts for the Fock-Darwin wavefunctions and Fertig-Halperin scattering states in the case where the sign of two or one curvatures of the potential is inversed respectively. We note that the case of an inverted one-dimensional parabolic potential was recently solved using special functions~\cite{Krason}, and displays a similar albeit simpler phenomenology compared to the two-dimensional situation with a finite magnetic field. The approach that we follow here extends previous phase space quantization ideas \cite{Malkin1969,Champel2007} that were used to derive semiclassical approximations in the large magnetic field limit \cite{Champel2008,Champel2009rapid,Champel2009,Champel2010}, allowing good understanding of local density of states measurements \cite{Hashi2008,Hashi2012} in the quantum Hall regime. This formalism, best suitable at high field, relies upon wavefunctions that are eigenstates of the free Hamiltonian (with pure Landau level spectrum), while maintaining a semi-classical behavior through their coherent state character with respect to the guiding center coordinate (displaying hence a 2D phase space). This property allows one to easily perform the projection of any states of the Hilbert space onto a given arbitrary Landau level, a procedure which assumes a full energy decoupling between the orbital and guiding center degrees of freedom of the electron in the plane. The semiclassical-type approximations are then vindicated at high magnetic fields by the slow dynamics of the guiding center. The main technical development made in the present paper is an extension of the coherent state formalism to account efficiently for Landau level mixing. Such a mixing inherently induced by a nonuniform electrostatic potential signals that the Landau level index may not be a good quantum number anymore, and that the cyclotron motion can in general not be treated independently of the guiding center motion. The key physical insight relies on the use of a 4D phase space, which allows one to treat both orbital motion and guiding center drifting on an equal footing. This approach relies on the general dynamics of phase space distribution functions, which has been proved to be an autonomous formulation of quantum mechanics \cite{Bayen1978,Zachos}. After establishing the general formalism describing the dynamics for the corresponding 4D Wigner functions, we obtain a unique compact analytical solution for the case of an arbitrary quadratic potential, accounting for the three different physical situations discussed above (quantum dots, quantum point contacts, quantum antidots). The case of a quantum antidot is found to display rich physics as a function of magnetic field. The strong magnetic field regime where both orbital and guiding center motions are associated with discrete energy levels gets substituted below a critical magnetic field by a regime with broadened Landau and antidot energy levels, before an ultimate breakdown of guiding center and orbital motions at very low magnetic fields. The behavior shown by our toy model of quantum antidot is clearly relevant for the understanding of the quantum Hall effect breakdown occuring in the more complicated case of a random potential. The plan is organised as follows. For completeness, Sec.~2 briefly reviews the high magnetic field coherent state formalism and the general equations determining the electronic motion in the corresponding 2D coherent state representation. Since the Landau level index is not a good quantum number at any finite magnetic fields, it becomes relevant in general to replace this discrete quantum number by an extra continuous degree of freedom with coherent state character, in the same way as for the guiding center degree of freedom. This leads us to work preferentially in a full 4D phase space representation provided by a basis of doubly coherent states \cite{Feldman1970,Varro1984,Manko2012,Zhebrak2013}. Sec.~3 provides the derivation of the general equation describing electronic motion in the plane under a perpendicular magnetic field within a 4D phase space representation (technical details are provided in Appendix A). The full quantum equation is solved in Sec.~4 for the case of arbitrary quadratic electrostatic potentials in terms of two independent {\em effective} cyclotron and guiding center motions. Remarkably, in the full 4D phase-space representation, all types of quadratic potential enjoy a generic (unique) and compact exact quantum solution, which is valid at any magnetic fields. The physics of the seldom considered inverted parabolic potential is investigated in Sec.~5, as a model of Landau quantization breakdown. \section{Review of the 2D-coherent state representation} In this paper, we consider a single electron of charge $e=-|e|$ and effective mass $m^{\ast}$ at position ${\bf r}=(x,y)$ in a two-dimensional plane subject to a perpendicular uniform magnetic field ${\bf B}=B \hat{{\bf z}}$ and an electrostatic potential $V({\bf r})$. The Hamiltonian reads \begin{eqnarray} H=\frac{1}{2 m^{\ast}} \left(-i \hbar {\bm \nabla}_{{\bf r}} -e{\bf A}({\bf r}) \right)^2 +V({\bf r}), \label{Ham} \end{eqnarray} where the vector potential ${\bf A}({\bf r})$ is related to the magnetic field with the equation ${\bm \nabla} \times {\bf A}={\bf B}$. In the absence of potential [i.e., for $V({\bf r})=0$], the corresponding quantum mechanical problem can be readily solved and yields the well-known quantization of the kinetic energy into discrete Landau levels $E_{n}=(n+1/2) \hbar \omega_c$ where $n$ is a positive integer (a.k.a. the Landau level index) and $\omega_c$ is the cyclotron frequency proportional to the magnetic field amplitude $B$ as $\omega_c=|e|B/m ^{\ast}$. Owing to the degeneracy of the kinetic energy levels, it is possible to showcase different bases of eigenstates associated with this Landau level quantization. A physically transparent basis is provided by the vortex set of states \cite{Malkin1969,Champel2007} expressed in the symmetrical gauge ${\bf A}={\bf B} \times {\bf r}/2$ as \begin{eqnarray} \langle{\bf r} | n,{\bf R} \rangle = \frac{1}{l_B \sqrt{2 \pi n!}} \left(\frac{z-Z}{\sqrt{2} l_B} \right)^{n} \, e^{-\frac{|z|^2+|Z|^2-2 Z z^{\ast}}{4 l_B^2}} , \label{vortex} \end{eqnarray} where $l_B=\sqrt{\hbar/|e|B}$ is the magnetic length and $z=x+iy$ refers to the electron position in the complex plane. Within this peculiar set of eigenstates of the Landau level problem, the degeneracy quantum number is provided by the vortex position ${\bf R}=(X,Y)$, associated with the complex coordinate $Z=X+iY$ in the complex plane, which uniquely characterizes for $n \geq 1$ the location of the zeros of the wave function in the two-dimensional plane. In the limit of vanishing $l_B$ the positions ${\bf R}$ reduce to the classical guiding center location. Despite presenting a nonorthogonal overlap with respect to the quantum number ${\bf R}$ typical of coherent states \begin{eqnarray} \langle n_1,{\bf R}_1| n_2,{\bf R}_2 \rangle= \delta_{n_1,n_2}\, e^{-\frac{|Z_1|^2+|Z_2|^2-2Z_1^{\ast} Z_2}{4 l_B^2}} , \end{eqnarray} the states (\ref{vortex}) form a coherent state basis with respect to the guiding center coordinate (within each Landau level), obeying the completeness relation \begin{eqnarray} \int \!\! \frac{d^2 {\bf R}}{2\pi l_B^2} \sum_{n=0}^{+\infty} |n,{\bf R} \rangle \langle n,{\bf R} |=1. \end{eqnarray} By associating the incremental area $d^2{\bf R}$ with the area $2\pi l_B^2$, this relation explicitly points out the degeneracy of the Landau levels to be $(2 \pi l_B^2)^{-1}$ per unit area. This Landau level degeneracy gets lifted when considering a non-uniform potential $V({\bf r})$. At high magnetic fields, i.e., when Landau level mixing can reasonably be neglected, the degeneracy lifting process becomes nonperturbative in nature and is the source of theoretical difficulties. The continuous character of the degeneracy quantum number ${\bf R}$ in the vortex state basis $|n,{\bf R}\rangle$ then offers a differential perspective of this process by an arbitrary potential, which has been thoroughly studied during the last decade in a series of papers \cite{Champel2008,Champel2009,Champel2010}. Due to the coherent state nature of the degree of freedom ${\bf R}$, the electronic Green's function in the time domain $t$ corresponding to Hamiltonian (\ref{Ham}) can be written as the convolution \begin{eqnarray} G({\bf r},{\bf r'};t)=\int \!\! \frac{d^2 {\bf R}}{2 \pi l_B^2} \sum_{n_1,n_2} K_{n_1,n_2}({\bf r},{\bf r}';{\bf R}) \, g_{n_1,n_2}({\bf R};t) \label{Green} \end{eqnarray} where the electronic structure factor defined by \begin{eqnarray} K_{n_1,n_2}({\bf r},{\bf r}'; {\bf R})=e^{-(l_B^2/4) \Delta_{{\bf R}}} \left[ \langle n_2,{\bf R}|{\bf r}' \rangle \, \langle {\bf r} | n_1,{\bf R}\rangle \right] \end{eqnarray} is independent of the electrostatic potential $V({\bf r})$ and embodies the quantum contribution arising from the pure orbital motion of the electron (here $\Delta_{{\bf R}}$ is the Laplacian operator taken with respect to the position ${\bf R}$). The vortex Green's function components $g_{n_1,n_2}({\bf R};t)$, which encode the quantum drift of the guiding center induced by $V({\bf r})$, obey the equations \begin{eqnarray} \hspace*{-2cm} \left(i\hbar \partial_{t}-E_{n_1} \pm i 0^+\right) g_{n_1,n_2}({\bf R};t) - \sum_{n_3} v_{n_1,n_3}({\bf R}) \star_{{\bf R}} g_{n_3,n_2}({\bf R};t)=\delta_{n_1,n_2} \delta(t) \label{Dyson} \end{eqnarray} with the effective potential matrix elements \begin{eqnarray} v_{n_1,n_2}({\bf R})=\int \!\! d^2 {\bf r} \, K_{n_1,n_2}({\bf r},{\bf r};{\bf R}) \, V\left({\bf r} \right) \label{pot} \end{eqnarray} expressing the average of the bare potential $V({\bf r})$ over the quantized orbital motion. Here the infinitesimal quantity $\pm 0^{+}$ relates to the retarded or advanced Green's functions. The symbol $\star_{{\bf R}}$ is a pseudodifferential infinite-order symplectic operator \begin{eqnarray} \star_{{\bf R}}=\exp\left[i \frac{l_B^2}{2}\left(\overleftarrow{\partial}_{X}\overrightarrow{\partial}_{Y} -\overleftarrow{\partial}_{Y}\overrightarrow{\partial}_{X} \right) \right], \label{star} \end{eqnarray} where the arrows above the partial derivatives indicate to which side (left or right) they have to be applied. It is a magnetic version of the Groenewold-Moyal star product \cite{Zachos}, with $l_B^2$ playing the role of an effective Planck's constant and the one-dimensional conjugated variables, position and momentum, being replaced by the components $X$ and $Y$ of the orbit center in the two-dimensional plane. The exact expression (\ref{Green}) translates into the quantum mechanical language the natural decomposition of the electronic motion into orbital and orbit center degrees of freedom. The vortex representation introducing both discrete and continuous quantum numbers turns out to be well-suited to treat quantitatively the resulting electronic dynamics at high magnetic fields, since it structurally encodes that these two elementary motions are characterized by very different time scales: the fast orbital degree of freedom is described in discrete terms, while a continuous classical phase space representation of the Landau level degeneracy is vindicated by the slow dynamics of the orbit center. In the high magnetic field regime, a good (perturbative) approximation is to entirely separate these two time scales by considering that the orbital motion gets decoupled from the guiding center motion. Technically, this implies restricting the electron dynamics to a given Landau level subspace. This state projection is conveniently performed for any Landau levels through the analyticity property of the vortex state basis (\ref{vortex}) in the complex guiding center variable $Z$, which holds irrespective of the Landau level index $n$ (in contrast, the well-known anti-analyticity property of the wave functions in the electronic variable $z$ only holds for the lowest Landau level). In terms of vortex Green's functions, only diagonal elements $g_{n_1,n_1}({\bf R};t)$ contribute to the overall electron dynamics in expression (\ref{Green}) after projection. Then, at the level of the guiding center motion, the star product operator (\ref{star}) generates a hierarchy of local energy scales ordered by powers of $l_B^2$ and successive spatial derivatives of the effective potential (\ref{pot}), which allows one to devise semiclassical nonperturbative approximation schemes for the vortex Green's functions $g_{n_1,n_1}({\bf R};t)$ valid at small times $t$ (and physically justified at finite temperatures). The objective of this paper is to address the situation beyond the Landau level projection, i.e., to eventually relax the high magnetic field constraint. This means to deal in Eq. (\ref{Dyson}) with the entire matrix structure of the vortex Green's functions associated to the Landau levels together with the differential aspects related to the guiding center dependence. In the following, we shall develop an alternative strategy valid at any magnetic fields, which requires a reformulation of the quantum representation of the states. \section{General equation of motion in the 4D-coherent state representation} The idea is to treat the two electronic degrees of freedom associated to the cyclotron motion and the guiding center motion on an equal footing, i.e., within a fully differential 4-dimensional phase-space perspective. For this purpose, we introduce a coherent state representation of the orbital degree of freedom by defining the doubly coherent states $|{\bm \rho},{\bf R} \rangle$, built from the vortex states as \begin{eqnarray} |{\bm \rho},{\bf R} \rangle = e^{- \frac{|\zeta|^2}{4 l_B^2}} \, \sum_{n=0}^{+ \infty} \frac{1}{\sqrt{n!}} \left(\frac{\zeta}{\sqrt{2}l_B }\right)^n | n,{\bf R} \rangle, \label{defrho} \end{eqnarray} where the orbital position ${\bm \rho}=(\rho_x,\rho_y)$ replaces the quantized Landau level index $n$ by a continuous cyclotron motion around the guiding center ${\bm R}$ in the two-dimensional plane, with $\zeta=\rho_x+i \rho_y$ its complex number representation (thus $\zeta^{\ast}=\rho_x-i \rho_y$). It can be easily established that this set of states form a bi-coherent states basis, with the standard non-orthogonal overlap expression: \begin{eqnarray} \hspace*{-1.5cm} \langle {\bm \rho}_1 ,{\bf R}_1| {\bm \rho}_2 ,{\bf R}_2 \rangle = \langle {\bm \rho}_1 | {\bm \rho}_2 \rangle \, \langle {\bf R}_1| {\bf R}_2 \rangle = e^{-\frac{|\zeta_1|^2+|\zeta_2|^2-2\zeta_1^{\ast} \zeta_2 }{4 l_B^2} } \, e^{-\frac{|Z_1|^2+|Z_2|^2-2Z_1^{\ast} Z_2 }{4 l_B^2} } , \nonumber \\ \end{eqnarray} and the completeness relation \begin{eqnarray} \int \!\! \frac{d^2 {\bf R}}{2 \pi l_B^2} \int \!\! \frac{d^2 {\bm \rho}}{2\pi l_B^2} | {\bm \rho} ,{\bf R} \rangle \langle {\bm \rho},{\bf R} |=1 . \end{eqnarray} From Eqs. (\ref{vortex}) and (\ref{defrho}) one easily gets the expression for the fully coherent wave function (which already appeared in the literature several decades ago, see e.g. Refs. \cite{Feldman1970,Varro1984}) \begin{eqnarray} \langle {\bf r} | {\bm \rho},{\bf R} \rangle= \frac{1}{l_B \sqrt{2 \pi}} \, e^{-\frac{|z|^2+|Z|^2+|\zeta|^2-2Z z^{\ast}-2 \zeta (z-Z)}{4 l_B ^2}}. \end{eqnarray} The corresponding Green's functions in this representation of bi-coherent states are obtained from the vortex Green's functions components via a simple change of basis as \begin{eqnarray} \hspace*{-1.5cm} g_{{\bm \rho}_1,{\bm \rho}_2}\left({\bf R};t\right)= e^{- \frac{|\zeta_1|^2+|\zeta_2|^2}{4 l_B^2}} \,\sum_{n_1=0}^{+\infty} \sum_{n_2=0}^{+\infty} \left(\frac{\zeta_1^{\ast}}{\sqrt{2}l_B }\right)^{n_1} \left(\frac{\zeta_2}{\sqrt{2}l_B }\right)^{n_2} \frac{g_{n_1,n_2}({\bf R};t)}{\sqrt{n_1!\, n_2!}}. \label{defgrho} \end{eqnarray} The analytical dependence of these functions on the variables $\zeta_1^{\ast}$ and $\zeta_2$ is put to good use in order to write down a general ``diagonal'' expression for the electronic Green's function (see Appendix A for a detailed derivation) similarly to Eq. (\ref{Green}) \begin{eqnarray} G({\bf r},{\bf r'};t)=\int \!\! \frac{d^2 {\bf R}}{2 \pi l_B^2} \int \!\! \frac{d^2 {\bm \rho}}{2 \pi l_B^2} K({\bf r},{\bf r}';{\bm \rho},{\bf R}) \, g({\bm \rho},{\bf R};t) \label{Green2} \hspace*{0.5cm} \end{eqnarray} with the Kernel function \begin{eqnarray} K({\bf r},{\bf r}';{\bm \rho},{\bf R})=e^{-(l_B^2/4) \left(\Delta_{{\bf R}}+\Delta_{{\bm \rho}} \right)} \left[ \langle {\bm \rho},{\bf R}|{\bf r}' \rangle \, \langle {\bf r} | {\bm \rho},{\bf R}\rangle \right], \hspace*{0.5cm} \end{eqnarray} and where the diagonal component functions $g({\bm \rho},{\bf R};t)$ obey the relatively compact (exact) equation \begin{eqnarray} \left(i \hbar \partial_{t} \pm i 0^+\right) g({\bm \rho},{\bf R};t) - E({\bm \rho},{\bf R}) \star_{{\bf R}} \star_{{\bm \rho}} \, g({\bm \rho},{\bf R};t)=\delta(t). \nonumber \\ \label{Dysonrho} \end{eqnarray} The matrix structure encountered into the previous system of equations (\ref{Dyson}) has been replaced in the present four-dimensional phase space representation by the presence of an additional pseudodifferential infinite-order symplectic operator \begin{eqnarray} \star_{{\bm \rho}}=\exp\left[i \frac{l_B^2}{2}\left(\overleftarrow{\partial}_{\!\! \rho_x}\overrightarrow{\partial}_{\!\! \rho_y} -\overleftarrow{\partial}_{\!\! \rho_y}\overrightarrow{\partial}_{\!\! \rho_x} \right) \right], \label{starrho} \end{eqnarray} whose structure is the same as that of the star product operator $\star_{\bf{R}}$ which governs the quantum motion of the guiding center ${\bf R}$. Here, the quantity $E({\bm \rho},{\bf R})$ expresses the classical total energy \begin{eqnarray} E({\bm \rho},{\bf R})=\frac{1}{2} m^{\ast}\omega_c^2 {\bm \rho}^2 + V\left({\bm \rho}+{\bf R} \right) , \label{potrho} \end{eqnarray} which includes both the classical (rotational) kinetic energy contribution (which was previously associated with the Landau levels) and the potential energy contribution. Expression~(\ref{potrho}) is obvious on semi-classical grounds, and the only difficulty brought by quantum mechanics in its 4D phase space representation is the necessity to deal with the star-product~(\ref{starrho}). The full phase space formulation provided by the use of the bi-coherent state set thus offers a physically transparent perspective, with the explicit implementation of the electron motion decomposition ${\bf r}={\bm \rho}+{\bf R}$ in the quantum realm. Note that the 4D phase space is characterized here by two spatial coordinates ${\bm \rho}$ and ${\bm R}$, in contrast to the more standard phase space representation with electronic coordinate ${\bm r}$ and its zero-field conjugate momentum ${\bm p = -i\hbar \nabla}$. In fact, under a finite magnetic field, the canonical quantization readily shows that $(X,Y)$ and $(\rho_x,\rho_y)$ each constitute a quantum conjugate pair, vindicating our choice of 4D phase space representation. As a consequence, the main difficulty in this deformation quantization formulation is entirely embodied in the infinite-order differential operators $\star_{{\bf R}}$ and $\star_{{\bm \rho}}$ appearing in Eq. (\ref{Dysonrho}). In general, the electronic potential energy term $V({\bf r})$ introduces a coupling between the orbital ${\bm \rho}$ and the guiding center ${\bf R}$ degrees of freedom, which makes this quantum problem generically quite complicated to solve. Nevertheless, as shown in the next section, an exact decoupling can be handled for any quadratic potentials. \section{Generic solution for arbitrary quadratic potentials} So far, we have derived the general quantum equation (\ref{Dysonrho}) obeyed by the Green's functions in the 4D phase space representation, without resorting to any specific form for the potential $V({\bf r})$. The case of a linear potential term does not present peculiar difficulties, since it does not lead to a coupling between the orbital and guiding center degrees of freedom. Consequently, from now on we focus on the case of quadratic potentials which can be written without loss of generality (a translation and a rotation of the coordinates lead immediately to the most generic quadratic form) as \begin{eqnarray} V({\bf r})=a x^2 + b y^2, \label{potV} \end{eqnarray} where $a$ and $b$ are arbitrary real coefficients, which encompass the three possible cases of potentials: i) confining (parabolic case, $a>0$ and $b>0$); ii) saddle point (hyperbolic case, $ab<0$); iii) impurity-like (inverted parabolic case, $a<0$ and $b<0$). Therefore, the total energy (\ref{potrho}) reads \begin{eqnarray} E({\bm \rho},{\bf R})=a (X+\rho_x)^2+b(Y+\rho_y)^2+c {\bm \rho}^2 \end{eqnarray} with $c=\frac{1}{2} m^{\ast} \omega_c^2$. The difficulty obviously comes from the presence of terms mixing the ${\bm \rho}$ and ${\bf R}$ coordinates, a hallmark of quadratic (squared) contributions. The above equation (\ref{Dysonrho}) for the phase space Green's functions can be solved exactly through the introduction of a well-chosen change in variables $({\bm \rho},{\bf R}) \to ({\bf R}_1,{\bf R}_2)$, which allows us to simultaneously decouple the spatial dependences in the total energy and in the differential star-operators. More explicitly, we impose that the total energy reads after the variable transformation as \begin{eqnarray} E({\bm \rho},{\bf R})=V_1({\bf R}_1)+V_2({\bf R}_2), \label{cons} \end{eqnarray} where the new (quadratic) potential functions $V_1$ and $V_2$ will be determined later on. A second condition is that the new star-products $\star_{{\bf R}_1}$ and $\star_{{\bf R}_2}$ defined with respect to the new variables ${\bf R}_1$ and ${\bf R}_2$ remain decoupled (typically, we do not want to generate cross-derivative terms like $\overleftarrow{\partial}_{\!\! X_1} \overrightarrow{\partial}_{\!\! Y_2}$). The solutions to the differential equations in ${\bf R}_1$ and ${\bf R}_2$ are then derived separately (without the $\delta(t)$ source term) and are appointed in the following as the functions $f_1({\bf R}_1;t)$ and $f_2({\bf R}_2;t)$ with the property that $f_1({\bf R}_1;0)=f_2({\bf R}_2;0)=1$. Hence, it can be easily shown that the full solution of Eq. (\ref{Dysonrho}) is given by the product function \begin{eqnarray} \label{g4D} g({\bm \rho},{\bf R};t)= \mp i \theta(\pm t) \,f_1({\bf R}_1;t) f_2({\bf R}_2;t) \label{fullsolution} \end{eqnarray} with $\theta(t)$ the Heaviside step function, and where the functions $f_{j}({\bf R}_j;t)$ with $j=1$ or $2$ obey the equation \begin{eqnarray} \left(i \hbar \partial_{t} \pm i 0^+\right) f_j({\bf R}_j;t) - V_j({\bf R}_j) \star_{{\bf R}_j} \, f_j({\bf R}_j;t)=0. \hspace*{0.5cm} \label{equationf} \end{eqnarray} This latter equation is very similar to the one obtained for the pure (decoupled) guiding center motion at high magnetic fields after Landau level projection. We can thus follow the derivation detailed in Ref. \cite{Champel2009} to directly write down the solution \begin{eqnarray} \label{fRt} f_j({\bf R}_j;t)=\frac{e^{-i[V_j({\bf R}_j)-V_j({\bf R}_{j0})] \tau_j(t)}} {\cos \left[\sqrt{\gamma_j} t/\hbar \right]} \, e^{-\frac{it}{\hbar} \left[ V_j({\bf R}_{j0}) \mp i0^+\right]} \label{solution}, \hspace*{0.5cm} \end{eqnarray} where \begin{eqnarray} \tau_j(t)=\frac{1}{\sqrt{\gamma_j}} \tan \left( \sqrt{\gamma_j} t /\hbar \right). \label{tau} \end{eqnarray} Here the point ${\bf R}_{j0}$ refers to the critical point of the qua\-dra\-tic potential $V_j$, i.e., $\left. {\bm \nabla}_{{\bf R}_j}V_j({\bf R}_j) \right|_{{\bf R}_{j}={\bf R}_{j0}}={\bf 0}$, and the (uniform) quantity $\gamma_j$ is related to the Gaussian curvature of the potential $V_j$ as \begin{eqnarray} \gamma_j= \frac{l_B^4}{4} \left[\partial^2_{X_j}V_j \partial^2_{Y_j}V_j -\left(\partial_{X_j} \partial_{Y_j} V_j \right)^2 \right]. \end{eqnarray} This quantity plays a pivotal role, since its square root crucially determines the relative time dependence of the Green's function, and thus the spectral properties of the electronic motion. For instance, when $\gamma_j$ is real positive, the function $ f_j({\bf R}_j;t)$ contains a periodical dependence in time, which can be restated as a Fourier series expansion to yield the alternative expression \begin{eqnarray} f_j({\bf R}_j;t)=\sum_{p_j=-\infty}^{+ \infty} a_{p_j}({\bf R}_j) \, e^{\frac{it}{\hbar} \left[p_j \sqrt{\gamma_j}- V_j({\bf R}_{j0}) \pm i0^+\right]}. \label{fconfin} \end{eqnarray} It has been shown in Appendix A of Ref. \cite{Champel2010} that the series coefficients read \begin{eqnarray} a_{p_j}({\bf R}_j) &=&2 (-1)^{n_j} e^{-|\rho_j ({\bf R}_j) |} L_n \left(2|\rho_j ({\bf R}_j) | \right), \\ \rho_j ({\bf R}_j) &=&\frac{V_j({\bf R}_j)-V_j({\bf R}_{j0})}{\sqrt{\gamma_j}}, \end{eqnarray} whenever $p_j=\chi_j (2n_j+1)$ with $n_j$ a positive integer and $\chi_j=\pm 1$ whether the potential $V_j$ is convex or concave (here $L_n(x)$ is the Laguerre polynomial of degree $n$), and $a_{p_j}({\bf R}_j)=0$ whenever $p_j \neq \chi_j (2n_j+1)$. From expression (\ref{fconfin}) valid when $\gamma_j \geq 0$ it is thus readily understood that the energy contribution arising from the potential $V_j$ is quantized with energy gaps given by $2 \sqrt{\gamma_j}$. For $\gamma_j<0$, it is understood in Eqs. (\ref{solution}) and (\ref{tau}) that $\sqrt{\gamma_j}=i \sqrt{-\gamma_j}$, so that the cosine and tangent trigonometric functions transform into their hyperbolic counterparts. As a result, the time periodicity of the Green's function is replaced by a decay on the time scale $1/\sqrt{-\gamma_j}$ due to the cutoff function $1/\cosh(\sqrt{-\gamma_j}t)$, which can be seen as a manifestation of quantum tunneling effects. From the above requirements on the variables decoupling, it is clear that a linear transformation of the coordinates will fit our purpose. Let us write the original variables in terms of the new ones as \begin{eqnarray} X &=& \lambda(X_1+\alpha X_2) , \hspace*{0.5cm} \rho_x = \lambda(X_2 + \beta X_1), \nonumber \\ Y &=& \lambda(Y_1+ \eta Y_2), \hspace*{0.75cm} \rho_y = \lambda(Y_2 + \delta Y_1), \end{eqnarray} with $\lambda>0$. The condition for the absence of cross-terms in the star products yields $\beta=\eta$ and $\alpha=\delta$. Furthermore, we get $ \overleftarrow{\partial}_{\!\! X} \overrightarrow{\partial}_{\!\! Y}-\overleftarrow{\partial}_{\!\! \rho_x} \overrightarrow{\partial}_{\!\! \rho_y} = \overleftarrow{\partial}_{\!\! X_1} \overrightarrow{\partial}_{\!\! Y_1}-\overleftarrow{\partial}_{\!\! X_2} \overrightarrow{\partial}_{\!\! Y_2} $ provided that $\lambda^{-2} =1-\alpha \beta$. The other constraint (\ref{cons}) leads to $\beta= \alpha a/b$ with \begin{eqnarray} \alpha=-\frac{1}{2a} \left[a+b+c - \sqrt{(a+b+c)^2-4ab} \right]. \end{eqnarray} Note that only this combination is compatible with the equality $a \alpha =b \beta$ (this comes out by considering, e.g., the limit $b \to 0$ which necessarily implies $\alpha \to 0$). From this, we obtain \begin{eqnarray} \lambda^2=\frac{1}{2} \frac{a+b+c + \sqrt{(a+b+c)^2-4ab}}{\sqrt{(a+b+c)^2-4ab}}. \end{eqnarray} The effective quadratic potentials read $V_j({\bf R}_j)=a_j X_j^2+b_j Y_j^2$ with \begin{eqnarray} \label{a1} a_1 &=& \lambda^2 \left[a(1+\beta)^2+c \beta^2 \right] \nonumber \\ &= & \frac{a}{2b} \left[b-a-c+ \sqrt{(a+b+c)^2-4ab} \right] , \hspace*{1cm} \\ \label{b1} b_1 &=& \lambda^2 \left[b(1+\alpha)^2+c \alpha^2 \right] \nonumber \\ &=& \frac{b}{2a} \left[a-b-c+ \sqrt{(a+b+c)^2-4ab} \right] , \\ \label{a2} a_2 &=& \lambda^2 \left[a(1+\alpha)^2+c \right] \nonumber \\ &=& \frac{1}{2} \left[c+a-b+ \sqrt{(a+b+c)^2-4ab} \right] , \\ \label{b2} b_2 &=& \lambda^2 \left[b(1+\beta)^2+c \right] \nonumber \\ &=& \frac{1}{2} \left[c+b-a+ \sqrt{(a+b+c)^2-4ab} \right] . \end{eqnarray} The new variables are expressed in terms of the original guiding center and orbital coordinates as \begin{eqnarray} \label{X1} X_1 &=& \lambda(X- \alpha \rho_x), \hspace*{0,5cm} \label{X2} X_2 = \lambda(\rho_x- \beta X), \\ \label{Y1} Y_1 &=& \lambda(Y-\beta \rho_y), \hspace*{0,6cm} \label{Y2} Y_2 = \lambda(\rho_y- \alpha Y). \end{eqnarray} By considering the high magnetic field limit $c \gg |a|,|b|$ for which $\lambda=1$ and $\alpha=\beta=0$, it is clear that, in general, the degree of freedom ${\bf R}_1$ plays the role of an effective guiding center, while ${\bf R}_2$ corresponds to an effective orbital degree of freedom. The final explicit solution for arbitrary quadratic potentials can be read off from expression~(\ref{g4D}) for the 4D phase-space Green's function $g({\bm \rho},{\bm R};t)$, expressed from the functions $f_j({\bm R_j};t)$ in Eq.~(\ref{fRt}), with the coordinates ${\bm R_j}$ given in Eqs.~(\ref{X1})-(\ref{Y2}) and the effective potentials $V_j({\bm R_j})=a_jX_j^2+b_jY_j^2$ determined by the four coefficients in Eqs.~(\ref{a1})-(\ref{b2}). One remarkable aspect of this general solution is that it does not require the computation of any special functions. \section{Model of Landau quantization breakdown} Let us finally analyze some physical features of the exact quantum solution (\ref{fullsolution}). As underlined above, the Gaussian curvatures $\gamma_1$ and $\gamma_2$ of the effective potentials $V_1$ and $V_2$ are key quantities determining the nature itself of the energy spectrum: \begin{eqnarray} \label{gamma1} \gamma_1 &=& l_B^4 a_1 b_1 = l_B^4 \frac{c}{2} \left[a+b+c- \sqrt{(a+b+c)^2-4ab} \right] , \nonumber \\ && \\ \label{gamma2} \gamma_2 &=& l_B^4 a_2 b_2 = l_B^4 \frac{c}{2} \left[a+b+c+ \sqrt{(a+b+c)^2-4ab} \right] . \nonumber \\ \end{eqnarray} Note that here $l_B^2 c=\frac{1}{2}\hbar \omega_c \geq 0$, while the coefficients $a$ and $b$ can be chosen positive or negative real numbers depending on the spatial configuration for the original potential energy $V({\bf r})$. A rich variety of regimes, showing structural changes in the energy spectrum, only takes place when both $a$ and $b$ are negative, corresponding to a toy-model of quantum antidot. Indeed, in the other circumstances (i.e., for $ab <0$, or for $a$ and $b$ both positive), one always gets a real positive $\gamma_2$, which signals the discrete quantization of the effective orbital motion with gaps given by the energy scale $ \hbar \Omega_c=2 \sqrt{\gamma_2}$. Landau quantization is thus robust at arbitrary small magnetic field for confining or saddle-point potentials. The renormalization of the cyclotron frequency from $\omega_c$ to $\Omega_c$ due to the Landau level mixing processes can be translated into a renormalization of the magnetic length $l_B$ by introducing the new length $L=l_B \left(l_B^4 c^2/\gamma_2\right)^{1/4} \equiv l_B \left(\omega_c/\Omega_c \right)^{1/2}. $ Moreover, it is instructive to rewrite the curvature of the effective potential $V_1$ as $\gamma_1=l_B^4 ab l_B^4 c^2/\gamma_2 \equiv \gamma \left( \omega_c/\Omega_c\right)^2$, with $\gamma=l_B^4 ab$ the Gaussian curvature of the electronic potential energy $V({\bf r})$. This proportionality relation shows that when $\gamma_2 > 0$ the sign of $\gamma_1$ dictating the effective guiding center motion is in fact entirely determined by the bare potential curvature. Nevertheless, the effective guiding center follows equipotential lines of the effective potential $V_1({\bf R}_1)$, which, in the presence of Landau level mixing, differ from those of $V({\bf R}_1)$ and evolve in magnetic field. \begin{figure}[ht] \begin{center} \resizebox{0.75\textwidth}{!}{ \includegraphics{Fig.eps}} \end{center} \caption{{\bf Dynamical regimes for a non rotationally invariant antidot.} The figure displays the real and imaginary parts of the quantities $\sqrt{\gamma_1}$ and $\sqrt{\gamma_2}$ characterizing the energy spectrum and related to the two effective potentials $V_1$ and $V_2$ (defined in Eqs.~(\ref{a1})-(\ref{b2})), which are associated, respectively, to the effective guiding center coordinate ${\bm R_1}$ and to the effective orbital motion ${\bm R_2}$, as a function of the classical cyclotron energy $c=m^{\ast} \omega_c^2/2$. Here an asymmetric inverted parabolic potential with $a=10 \, b<0$ in Eq. (\ref{potV}) has been considered. Landau quantization at large magnetic field ($c \geq c_{+}$) corresponds to real and positive potential curvatures, while the intermediate magnetic field range ($c_{-} \leq c \leq c_{+}$) shows broadened Landau and antidot levels due to the non-zero imaginary parts of $\sqrt{\gamma_1}$ and $\sqrt{\gamma_2}$. However, the relation $\mathrm{Im}[\sqrt{\gamma_1}]= -\mathrm{Im}[\sqrt{\gamma_2}]$ translates the fact that the effective cyclotron orbits shrink while the effective guiding center follows orbits that increase as the electron spins down the potential landscape. Finally, both curvatures $\gamma_1$ and $\gamma_2$ become negative at low magnetic fields ($c \leq c_{-}$) so that $\sqrt{\gamma_1}$ and $\sqrt{\gamma_2}$ are purely imaginary, corresponding to the electron running down the inverted parabolic potential without performing any cyclotron motion.} \label{Curvatures} \end{figure} Focusing now the analysis on the inverted parabolic case ($a$ and $b$ negative), it is clear that both effective potential curvatures $\gamma_1$ and $\gamma_2$ are real and positive for strong enough magnetic field, as seen by taking the limit of large $c$ in Eqs.~(\ref{gamma1})-(\ref{gamma2}). Having $\gamma_2>0$ signals robust high magnetic field Landau quantization, while $\gamma_1>0$ demonstrates that the effective antidot potential $V_1$ confines the electronic motion due to the strong Lorentz force, despite the bare antidot potential $V({\bf r})$ of Eq. (\ref{potV}) being repulsive and unbounded from below. Decreasing the magnetic field, i.e. reducing the value of $c$, one encounters a first critical value $c_+=(\sqrt{|a|}+\sqrt{|b|})^2$ below which the term under the square root in Eqs.~(\ref{gamma1})-(\ref{gamma2}) becomes negative. In this case, both orbital and guiding center effective motions lock into decaying orbits (in a semiclassical viewpoint), leading to a finite broadening of both the Landau and antidot energy levels associated to the finite imaginary parts of the quantities $\sqrt{\gamma_1}$ and $\sqrt{\gamma_2}$. Landau quantization only survives on short time-scales in this field regime, where cyclotron orbits shrink ($\mathrm{Im}[\sqrt{\gamma_2}]>0$), while the guiding center makes larger and larger loops around the antidot potential ($\mathrm{Im}[\sqrt{\gamma_1}]=-\mathrm{Im}[\sqrt{\gamma_2}]<0$). Finally, when $a\neq b$ so that angular momentum is no more conserved, one finds for lower magnetic fields a second critical value $c_-=(\sqrt{|a|}-\sqrt{|b|})^2$ below which both curvatures $\gamma_1$ and $\gamma_2$ become real and negative, and thus the quantities $\sqrt{\gamma_1}$ and $\sqrt{\gamma_2}$ get purely imaginary. This low field regime for asymmetric antidot potentials corresponds to the rapid runaway of the electron down the inverted potential without any looping motion from the Lorentz force, so that the magnetic orbital effects are totally washed out. This rich scenario of Landau quantization breakdown in an inverted quadratic potential is illustrated on Fig.~\ref{Curvatures} in the case $a=10\, b<0$. Note that the Gaussian curvatures in Eqs. (\ref{gamma1})-(\ref{gamma2}) are purely classical concepts, as the same quantities naturally appear when solving the characteristic Newtonian equation of motion of a charged particule in an arbitrary quadratic potential in presence of a magnetic field. In this sense, the Landau levels breakdown is not intrinsically quantum in nature, although it will affect electronic motion at the quantum level. \section{Conclusion and final remarks} We have developed a 4D phase space representation of the in-plane electronic quantum motion in a perpendicular magnetic field, which is relevant beyond the Landau level projection. While a 2D-coherent state representation considering the discrete Landau level index as a good quantum number appears still efficient at moderately small Landau level mixing, the recourse to a bi-coherent state representation for which both the guiding center and the orbital degrees of freedom are associated with continuous (coherent) quantum numbers turns out to be unavoidable to get phase space solutions describing the electronic quantum motion at any magnetic fields. As an illustration, we have considered the motion in arbitrary quadratic electrostatic potentials, which is known to be exactly solvable by diagonalization of the Schr\"{o}\-din\-ger's equation. The full phase space formulation (unusual in condensed matter when dealing with fully quantum problems) offers an original viewpoint, with a limpid underlying classical physics, on the quantization processes, which is very different from that provided by the conventional (historical) derivations \cite{Fock1928,Darwin1931,Fertig1987} based on the wave function formalism. Especially, thanks to the overcompleteness of the coherent state representation, it yields a generic (unique) solution capable to embrace all types of quadratic potential within a simple compact mathematical expression, without having recourse to the properties of special orthogonal (Hermite, Laguerre, etc...) polynomials or special functions as usually required via the wave function formalism. We have also investigated a simplified model of Landau breakdown in the case of an inverted parabolic potential, showing a surprisingly rich phe\-no\-me\-no\-lo\-gy. Most markedly, this model displays three distinct physical stages when varying the field amplitude, in a very similar way to the situation encountered in disordered two-dimensional electronic gases. A possible application of the full phase space formalism beyond the case of quadratic potentials may be the derivation of approximate functionals for the local density of states valid in a broader magnetic field range than originally devised in Refs. \cite{Champel2009,Champel2010} for a smooth disordered electrostatic potential. In particular, one may expect to get specific signatures of Landau level mixing in the characteristic features of the effective guiding center motion. The extension of the phase space formalism may also be useful for the study of the correlations of the local density of states in a broader regime than in Ref. \cite{Champel2011} which neglects Landau level mixing. However, the present phase space formulation, which naturally allows one to perform semi-classical (local) approximations, is usually not convenient for the study of nonlocal transport properties, which require controlled approximations of the quantum solution on long time scales.

98375bad784db569c3ca4fa771b8bfb1e42a66bb

\section{Theoretical Analysis of Algorithm~\ref{alg:mvp}} \label{sec:theory_app} Now we present the theoretical analysis of Algorithm~\ref{alg:mvp}, with both asymptotic convergence and finite-sample error bound analysis. For the purpose of clarity, in the following analysis, $x$ is defined as $x := [y,\theta]^\top$, where $x^1 :=y,\ x^2 :=\theta$. Similarly, $g_t^i$ (resp. $\beta^i_t$) is used to denote $g_t^y$(resp. $\beta^y_t$) and $g_t^\theta$(resp. $\beta^\theta_t$), where $g_t^1 :=g_t^y,\ g_t^2 :=g_t^\theta$(resp. $\beta^1_t :=\beta^y_t,\ \beta^2_t :=\beta^\theta_t$). Let $\beta_{t}^{\max} = \max\{\beta_{t}^1,\beta_{t}^2\}$, $\beta_{t}^{\min} = \min\{\beta_{t}^1,\beta_{t}^2\}$, and $\|\cdot\|_2$ denotes the Euclidean norm. \subsection{Asymptotic Convergence Proof Algorithm~\ref{alg:mvp}} \label{sec:asym} In this section, we provide the asymptotic convergence analysis for Algorithm~\ref{alg:mvp} with Option I, i.e., the stepsizes are chosen to satisfy the Robbins-Monro condition and the output is the last iteration's result. We first introduce a useful lemma, which follows from Lemma \ref{lem:err_bound}. \begin{lemma} \label{lem:prod} Under Assumption \ref{asm3}, we have \begin{align} \mathbb{E}[\langle \nabla_{y}\hat{f}_\lambda (x_t), \Delta_t^y \rangle] = & 0\\ \mathbb{E}[\langle \nabla_{\theta}\hat{f}_\lambda (x_t), \Delta_t^\theta \rangle] \leq & \beta^{\max}_t A G, \end{align} where $A$ is defined in Eq~\eqref{eq:stocerrbd}. \end{lemma} \begin{proof} Let $\Xi_t$ denote the history of random samples from the first to the $t$--th episode, then $x_t$ is independent of $\Delta_t^i$ conditioned on $\Xi_{t - 1}$ (since $x_t$ is deterministic conditioned on $\Xi_{t - 1}$). Then we have \begin{align} &\mathbb{E}[\langle \nabla_{y}\hat{f}_\lambda (x_t), \Delta_t^y \rangle] = \mathbb{E}_{\Xi_{t - 1}}[\mathbb{E}[\langle \nabla_{y}\hat{f}_\lambda (x_t), \Delta_t^y \rangle|\Xi_{t - 1}]] \\ = & \mathbb{E}_{\Xi_{t - 1}}[\langle \mathbb{E}[\nabla_{y}\hat{f}_\lambda (x_t)|\Xi_{t - 1}], \mathbb{E}[\Delta_t^y|\Xi_{t - 1}] \rangle] = 0, \label{eq:expbound1} \end{align} \begin{align} &\mathbb{E}[\langle \nabla_{\theta}\hat{f}_\lambda (x_t), \Delta_t^\theta \rangle] = \mathbb{E}_{\Xi_{t - 1}}[\mathbb{E}[\langle \nabla_{\theta}\hat{f}_\lambda (x_t), \Delta_t^\theta \rangle|\Xi_{t - 1}]] \\ = & \mathbb{E}_{\Xi_{t - 1}}[\langle \mathbb{E}[\nabla_{\theta}\hat{f}_\lambda (x_t)|\Xi_{t - 1}], \mathbb{E}[\Delta_t^\theta|\Xi_{t - 1}] \rangle] \\ \leq & \mathbb{E}_{\Xi_{t - 1}}[\| \mathbb{E}[\nabla_{\theta}\hat{f}_\lambda (x_t)|\Xi_{t - 1}]\|_2 \cdot \|\mathbb{E}[\Delta_t^\theta|\Xi_{t - 1}] \|_2] \\ \leq & \beta^{\max}_t A \mathbb{E}_{\Xi_{t - 1}}[\| \mathbb{E}[\nabla_{\theta}\hat{f}_\lambda (x_t)|\Xi_{t - 1}]\|_2] \\ \leq & \beta^{\max}_t A \mathbb{E}[\|\nabla_{\theta}\hat{f}_\lambda (x_t)\|_2] \leq \beta^{\max}_t A G, \label{eq:expbound2} \end{align} where the second equality in both Eq.\eqref{eq:expbound1} and Eq.\eqref{eq:expbound2} follows from the conditional independence between $x_t$ and $\Delta_t^i$, the second inequality in Eq.\eqref{eq:expbound2} follows from Assumption \ref{asm:var}, and the last inequality in Eq.\eqref{eq:expbound2} follows from the Jensen's inequality and $G$ is the gradient bound. \end{proof} Then, we introduce Lemma~\ref{lem:1}, which is essential for proving Theorem~\ref{thm:asyn}. \begin{lemma} \label{lem:1} For two non-negative scalar sequences $\{a_t\}$ and $\{b_t\}$, if $\sum_{t = 1}^{\infty}a_t = +\infty$ and $\sum_{t = 1}^{\infty}a_t b_t < +\infty$, we then have \begin{equation} \mathop {\lim }_{t \to \infty } \inf {b_t} = 0. \end{equation} Further, if there exists a constant $K>0$ such that $|b_{t + 1} - b_{t}| \leq a_t K$, then \begin{equation} \lim_{t \rightarrow \infty}b_t = 0. \end{equation} \end{lemma} The detailed proof can be found in Lemma A.5 of \citep{mairal2013stochastic} and Proposition 1.2.4 of \citep{bertsekas:npbook}. Now it is ready to prove Theorem~\ref{thm:asyn}. \begin{proof}[Proof of Theorem \ref{thm:asyn}] We define $\Gamma_t^1 := \hat{f}(x_{t}^{1},x_t^{2}) - \hat{f}(x_{t + 1}^{1},x_t^{2}), \Gamma_t^2 := \hat{f}(x_{t + 1}^{1},x_t^{2}) - \hat{f}(x_{t + 1}^{1},x_{t + 1}^{2})$ to denote the block update. It turns out that for $i=1,2$, and $\Gamma_t^i$ can be bounded following the Lipschitz smoothness as \begin{align} \Gamma_t^i \leq & \langle g_t^i, x_t^i - x_{t+1}^i \rangle + \dfrac{L}{2}\|x_{t}^i - x_{t + 1}^i \|_2^2 \\ = & -\beta_t^i \langle g_t^i, \Tilde{g}_t^i \rangle + \dfrac{L}{2}(\beta_t^i)^2\|\Tilde{g}_t^i\|_2^2 \\ = & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \dfrac{L}{2}(\beta_t^i)^2)\|\Delta_t^i\|_2^2 - (\beta_t^i - L(\beta_t^i)^2)\langle g_t^i,\Delta_t^i \rangle \\ = & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 \\ & - (\beta_t^i - L(\beta_t^i)^2) (\langle g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle + \langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle) \label{eq:lipthmq0} \end{align} where the equalities follow the definition of $\Delta_t^i$ and the update law of Algorithm~\ref{alg:mvp}. We also have the following argument \begin{align} & - (\beta_t^i - L(\beta_t^i)^2)\langle g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \\ \leq & |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \|g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t)\|_2 \\ \leq & L |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \|x_{t + 1} - x_{t}\|_2 \\ \leq & L |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \sqrt{\sum_{j = 1}^2 \|\beta_t^j \tilde{g}_t^j\|_2^2} \\ \leq & L (\beta_t^i + L(\beta_t^i)^2) \beta_t^{\max} \left(\|\Delta_t^i\|_2 + \sum_{j = 1}^2 (\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2) \right), \label{eq:lipargu} \end{align} where the first inequality follows from Cauchy-Schwarz inequality, the second inequality follows from the Lipschitz smoothness of objective function $\hat{f}_\lambda$, the third inequality follows from the update law of Algorithm~\ref{alg:mvp}, and the last inequality follows from the triangle inequality. Combining Eq.~\eqref{eq:lipthmq0} and Eq.~\eqref{eq:lipargu}, we obtain \begin{align} \Gamma_t^i \leq & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 \\ & - (\beta_t^i - L(\beta_t^i)^2)\langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \\ & + L(\beta_t^i + L(\beta_t^i)^2) \cdot \beta_t^{\max} \left(\|\Delta_t^i\|_2^2 + \sum_{j = 1}^2(\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2)\right). \label{eq:lipthmq} \end{align} Summing Eq.~\eqref{eq:lipthmq} over $i$, then we obtain \begin{align} \label{eq:summing} &\hat{f}_\lambda(x_t) - \hat{f}_\lambda(x_{t + 1}) \\ \leq & -\sum_{i = 1}^2(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 - \sum_{i = 1}^2(\beta_t^i - L(\beta_t^i)^2)\langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \\ &+ \sum_{i = 1}^2 \left( \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 + L(\beta_t^i + L(\beta_t^i)^2)\beta_t^{\max}(\|\Delta_t^i\|_2^2 + \sum_{j = 1}^2(\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2)) \right). \end{align} We also have the following fact, \begin{align} \mathbb{E}[\langle \nabla_{y}\hat{f}_\lambda (x_t), \Delta_t^y \rangle] = & 0\\ \mathbb{E}[\langle \nabla_{\theta}\hat{f}_\lambda (x_t), \Delta_t^\theta \rangle] \leq & \beta^{\max}_t A G. \end{align} We prove this fact in Lemma \ref{lem:prod} as a special case of Lemma \ref{lem:err_bound}, and the general analysis can be found in Lemma 1 in \citep{cd:bsg:xu2015}. Taking expectation w.r.t. $t$ on both sides of the inequality Eq.~\eqref{eq:summing}, we have \begin{align} & \mathbb{E}[\hat{f}_\lambda(x_t)] - \mathbb{E}[\hat{f}_\lambda(x_{t + 1})] \\ \leq & -\sum_{i = 1}^2(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\mathbb{E}[\|g_t^i\|_2^2] + (\beta_t^\theta - L(\beta_t^\theta)^2)\beta^{\max}_t A G \\ & + \sum_{i = 1}^2\left( \dfrac{L}{2}(\beta_t^i)^2\mathbb{E}[\|\Delta_t^i\|_2^2]+ L(\beta_t^i + L(\beta_t^i)^2)\beta_t^{\max} (\mathbb{E}[\|\Delta_t^i\|_2^2] + \sum_{j = 1}^2(\mathbb{E}[\|g_t^j\|_2^2] + \mathbb{E}[\|\Delta_t^j\|_2^2])) \right) \\ \leq & -\sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] + (\beta^{\max}_t)^2 A G \\ & + ( L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2)), \label{eq:exp} \end{align} where the first inequality follows from Eq.~\eqref{eq:expbound1} and Eq.~\eqref{eq:expbound2}, and the second inequality follows from the boundedness of $\mathbb{E}[\|\Delta_t^i\|_2]$ and $\mathbb{E}[\|g_t^i\|_2]$. Rearranging Eq.~\eqref{eq:exp}, we obtain \begin{align} &\sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] \\ \leq& \mathbb{E}[\hat{f}_\lambda(x_{t + 1})] - \mathbb{E}[\hat{f}_\lambda(x_t)] + (\beta_t^{\max})^2 A M_\rho + L(\beta_t^{\max})^2\sigma^2 \\ & + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2). \label{eq:rearr} \end{align} By further assuming $0 < \inf_{t}\frac{\{\beta^{\theta}_t\}}{\{\beta^{y}_t\}} \leq \sup_{t}\frac{\{\beta^{\theta}_t\}}{\{\beta^{y}_t\}} < \infty$, it can be verified that $\{\beta^{\max}_t\}$ and $\{\beta^{\min}_t\}$ also satisfy Robbins-Monro condition. Note that $\hat{f}_{\lambda}$ is upper bounded, summing Eq.~\eqref{eq:rearr} over $t$ and using the Robbins-Monro condition of $\{\beta^{\theta}_t\}, \{\beta^{y}_t\}, \{\beta^{\max}_t\}, \{\beta^{\min}_t\}$, we have \begin{equation} \label{eq:lemmf} \sum_{t = 1}^{\infty}\beta_t^{\min}\mathbb{E}[\|g_t^i\|_2^2] < \infty, ~\forall i. \end{equation} Furthermore, let $\xi^1_t = (x_t^1, x_t^2)$ and $\xi^2_t = (x_{t + 1}^1, x_t^2)$, then \begin{align} |\mathbb{E}[\|g_{t + 1}^i\|_2^2] - \mathbb{E}[\|g_t^i\|_2^2]| \leq & \mathbb{E}[\|g_{t + 1}^i + g_{t}^i\|_2 \cdot \|g_{t + 1}^i - g_{t}^i\|_2]\\ \leq & 2LM_{\rho}\mathbb{E}[\|\xi^i_{t + 1} - \xi^i_t\|_2] \\ = & 2LM_{\rho}\mathbb{E}\left[\sqrt{\sum_{j < i}\|\beta_{t + 1}^j\Tilde{g}_{t + 1}^j\|_2^2 + \sum_{j \geq i}\|\beta_{t}^j\Tilde{g}_{t}^j\|_2^2}\right] \\ \leq & 2LM_{\rho}\beta_{t}^{\max}\mathbb{E}\left[\sqrt{\sum_{j < i}\|\Tilde{g}_{t + 1}^j\|_2^2 + \sum_{j \geq i}\|\Tilde{g}_{t}^j\|_2^2} \right] \\ \leq & 2LM_{\rho}\beta_{t}^{\max}\sqrt{\mathbb{E}[\sum_{j < i}\|\Tilde{g}_{t + 1}^j\|_2^2 + \sum_{j \geq i}\|\Tilde{g}_{t}^j\|_2^2]}\\ \leq & 2LM_{\rho}\beta_{t}^{\max}\sqrt{2(G^2 + \sigma^2)}, \label{eq:eq48} \end{align} where the first inequality follows from Jensen's inequality, the second inequality follows from the definition of gradient bound $G$ and the gradient Lipschitz continuity of $\hat{f}_{\lambda}$, the third inequality follows from the Robbins-Monro condition of $\{\beta_t^y\}$ and $\{\beta_t^\theta\}$, and the last two inequalities follow Jensen's inequality in probability theory. Combining Eq.~\eqref{eq:lemmf} and Eq.~\eqref{eq:eq48} and according to Lemma \ref{lem:1}, we have $\lim_{t\rightarrow \infty}\mathbb{E}[\|g_t^i\|_2] = 0$ for $i = 1,2$ by Jensen's inequality. Hence, \begin{align} \mathbb{E}[\|\nabla_{y}\hat{f}_{\lambda}(x_t)\|_2] = & \mathbb{E}[\|g_t^y\|_2] \\ \mathbb{E}[\|\nabla_{\theta}\hat{f}_{\lambda}(x_t)\|_2] \leq & \mathbb{E}[\|\nabla_{\theta}\hat{f}_{\lambda}(x_t) - g_t^\theta\|_2] + \mathbb{E}[\|g_t^\theta\|_2] \\ \leq & L \cdot \mathbb{E}[\|y_{t + 1} - y_t\|_2] + \mathbb{E}[\|g_t^\theta\|_2] \\ \leq & L \beta_t^{\max} \mathbb{E}[\|\tilde{g}_{t}^{y}\|_2] + \mathbb{E}[\|g_t^\theta\|_2] \\ \leq & L\beta_t^{\max}\sqrt{G^2 + \sigma^2} + \mathbb{E}[\|g_t^\theta\|_2] \end{align} where the first inequality follows from the triangle inequality, the second inequality follows from the Lipschitz continuity of $\hat{f}_{\lambda}$, and the last inequality follows from the same argument for Eq.~\eqref{eq:eq48}. Also, note that $\lim_{t\rightarrow \infty}\beta_t^{\max} = 0$, $\lim_{t\rightarrow \infty}L\sqrt{G^2 + \sigma^2} < \infty$, and $\lim_{t\rightarrow \infty}\mathbb{E}[\|g_t^i\|_2] = 0$, so that when $t \to \infty $, $L\beta_t^{\max}\sqrt{G^2 + \sigma^2} + \mathbb{E}[\|g_t^y\|_2] + \mathbb{E}[\|g_t^\theta\|_2]\to 0$. This completes the proof. \end{proof} \begin{remark} Different from the stringent two-time-scale setting where one stepsize needs to be ``quasi-stationary'' compared to the other~\citep{saferl:castro2012}, the stepsizes in Algorithm~\ref{alg:mvp} does not have such requirements, which makes it easy to tune in practice. \end{remark} \subsection{Finite-Sample Analysis of Algorithm~\ref{alg:mvp}} \label{sec:fs-mvp-appendix} The above analysis provides asymptotic convergence guarantee of Algorithm~\ref{alg:mvp}, however, it is desirable to know the sample complexity of the algorithm in real applications. Motivated by offering RL practitioners confidence in applying the algorithm, we then present the sample complexity analysis with Option II described in Algorithm~\ref{alg:mvp}, i.e., the stepsizes are set to be a constant, and the output is randomly selected from $\{x_1,\cdots,x_N\}$ with a discrete uniform distribution. This is a standard strategy for nonconvex stochastic optimization approaches \citep{cd:sbmd:dang2015}. With these algorithmic refinements, we are ready to present the finite-sample analysis as follows. It should be noted that this proof is a special case of the general stochastic nonconvex BSG algorithm analysis provided in Appendix~\ref{sec:proof:thm:cyclicbcd}. \begin{proof}[Proof of Theorem \ref{thm:cyclicbcd2}] The proof of finite-sample analysis starts from the similar idea with asymptotic convergence. The following analysis follows from Eq.~\eqref{eq:rearr} in the proof of Theorem \ref{thm:asyn}, but we are using stepsizes $\{\beta_t^{\theta}\}$, $\{\beta_t^{y}\}$ are constants which satisfy $2\beta_t^{\min} > L(\beta_t^{\max})^2$ for $t = 1,\cdots,N$ in this proof. Summing Eq.~\eqref{eq:rearr} over $t$, we have \begin{align} \label{eq:finalsum} &\sum_{t = 1}^N \sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] \\ \leq& \hat{f}_\lambda^* - \hat{f}_\lambda(x_1) +\sum_{t = 1}^N [ (\beta_t^{\max})^2 A G + L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2)]. \end{align} Next, we bound $\mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2]$ using $\mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2]$ \begin{align} \mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2] = & \mathbb{E}[\|\nabla_\theta\hat{f}_\lambda(x_t) \|_2^2 + \|\nabla_y\hat{f}_\lambda(x_t) \|_2^2] \\ \leq &\mathbb{E}[\|\nabla_\theta\hat{f}_\lambda(\theta_t, y_{t}) - \nabla_\theta\hat{f}_\lambda(\theta_t, y_{t + 1}) + \nabla_\theta\hat{f}_\lambda(\theta_t, y_{t + 1})\|_2^2 + \|g_t^y\|_2^2] \\ \leq &\mathbb{E}[\|\nabla_\theta\hat{f}_\lambda(\theta_t, y_{t}) - g_t^\theta\|_2^2 + \|g_t^\theta\|_2^2 + 2\langle\nabla_\theta\hat{f}_\lambda(\theta_t, y_{t}) - g_t^\theta, g_t^\theta\rangle + \|g_t^y\|_2^2] \\ \leq &\mathbb{E}[\|\nabla_\theta\hat{f}_\lambda(\theta_t, y_{t}) - g_t^\theta\|_2^2 + 2\langle\nabla_\theta\hat{f}_\lambda(\theta_t, y_{t}) - g_t^\theta, g_t^\theta\rangle] + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & \mathbb{E}[L^2 \|y_{t + 1} - y_{t}\|_2^2 + 2L\|y_{t + 1} - y_{t}\|_2\cdot\|g_t^\theta\|_2] + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & \mathbb{E}[L^2 (\beta_{t}^y)^2\|\tilde{g}_t^y\|_2^2 + 2L\beta_{t}^y\|\tilde{g}_t^y\|_2\cdot\|g_t^\theta\|_2] + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & L^2 (\beta_{t}^y)^2(G^2 + \sigma^2) + 2L\beta_{t}^y\mathbb{E}[\|\tilde{g}_t^y\|_2\cdot\|g_t^\theta\|_2] + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & L^2 (\beta_{t}^y)^2(G^2 + \sigma^2) + L\beta_{t}^y\mathbb{E}[\|\tilde{g}_t^y\|_2^2 + \|g_t^\theta\|_2^2] + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & L^2 (\beta_{t}^{\max})^2(G^2 + \sigma^2) + L\beta_{t}^{\max}(2G^2 + \sigma^2) + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2]. \label{eq:newbd} \end{align} Then, combine Eq.~\eqref{eq:newbd} with Eq.~\eqref{eq:finalsum} \begin{align} & \sum_{t = 1}^N (\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2) \mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2] \\ \leq & \sum_{t = 1}^N (\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2) \mathbb{E}[L^2 (\beta_{t}^{\max})^2(G^2 + \sigma^2) + L\beta_{t}^{\max}(2G^2 + \sigma^2) + \|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq & \hat{f}_\lambda^* - \hat{f}_\lambda(x_1) +\sum_{t = 1}^N [(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)(L^2 (\beta_{t}^{\max})^2(G^2 + \sigma^2) \\ & + L\beta_{t}^{\max}(2G^2 + \sigma^2)) + (\beta_t^{\max})^2 A G + L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2)] \\ \leq& \hat{f}_\lambda^* - \hat{f}_\lambda(x_1) + (\beta_t^{\max})^2\sum_{t = 1}^N [(1 - \dfrac{L}{2}\beta_t^{\max})(L^2 \beta_{t}^{\max}(G^2 + \sigma^2) + L(2M_{\rho}^2 + \sigma^2)) + A G + L\sigma^2 \\ & + 2L(1 + L\beta_t^{\max})(3 \sigma^2 + 2 G^2)]. \end{align} Rearrange it, we obtain \begin{align} \mathbb{E}[\|\nabla\hat{f}_\lambda(x_z) \|_2^2] \leq \dfrac{\hat{f}_\lambda^* - \hat{f}_\lambda(x_1) + N (\beta_t^{\max})^2C}{N (\beta_t^{\min} - \frac{L}{2}(\beta_t^{\max})^2)}, \end{align} where \begin{align} C = & (1 - \dfrac{L}{2}\beta_t^{\max})(L^2 \beta_{t}^{\max}(G^2 + \sigma^2) + L(2G^2 + \sigma^2)) + A G + L\sigma^2 + 2L(1 + L\beta_t^{\max})(3 \sigma^2 + 2 G^2). \end{align} \end{proof} \begin{remark} In Theorem \ref{thm:cyclicbcd2}, we have proven the finite-sample analysis of Algorithm~\ref{alg:mvp} with Option II, i.e., constant stepsizes and randomly picked solution. Note that the error bound in Eq.~\eqref{eq:thm2} can be simplified as $\mathcal O(1 / (N \beta_t^{\min})) + \mathcal O(\beta_t^{\max})$. Especially, if $\beta_t^{\max}=\beta_t^{\min}=\beta_t^\theta = \beta_t^y$ are set to be $\Theta(1 / \sqrt{N})$, then the convergence rate of Option II in Algorithm~\ref{alg:mvp} is $\mathcal O(1/\sqrt{N})$. \end{remark} \section{Proofs in Convergence Analysis of Nonconvex BSG} \label{sec:probsg} This section includes proof of Lemma \ref{lem:err_bound} and Algorithm \ref{thm:cyclicbcd}. We first provide the pseudo-code for nonconvex BSG method as Algorithm \ref{alg:c_sbcd}. \begin{algorithm}[thb] \caption{The nonconvex BSG Algorithm} \label{alg:c_sbcd} {\bfseries Input:} Initial point $x_1 \in \mathbb{R}^n$, stepsizes $\{ \beta_t^i: i = 1,\cdots,b \}_{k = 1}^{\infty}$, positive integers $\{m_t\}_{k = 1}^\infty$ that indicate the mini-batch sizes, and iteration limit $N$. \begin{algorithmic}[1] \FOR{$k = 1,2,\dotsc,N$} \STATE Sample mini batch $\Xi_t = \{\xi_{t,1},\xi_{t,2},\dotsc,\xi_{t,m_t}\}$. \STATE Specify update order $\pi_t^i = i$, $i = 1,\cdots,b$, or randomly shuffle $\{i = 1,\cdots,b\}$ to $\{\pi_t^{1},\pi_t^{2},\dotsc,\pi_t^{b}\}$. \FOR{$i = 1,2,\dotsc,b$} \STATE Compute the stochastic partial gradient for the $\pi_t^i$th block as \begin{align} \tilde g_t^i = \frac{1}{m_t} \sum_{j = 1}^{m_t} \nabla_{x^{\pi_t^i}}F(x_{t + 1}^{\pi_t^{< i}}, x_{t}^{\pi_t^{\geq i}}; \xi_{t,j}). \end{align} \STATE Update $\pi_t^i$th block \begin{align} x_{t + 1}^{\pi_t^i} = x_{k}^{\pi_t^i} - \beta_t^i \tilde g_t^i. \end{align} \ENDFOR \ENDFOR \STATE Return $\bar x_{N} = x_{z}$ randomly according to \begin{align} \Pr(z = t) = \frac{\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2}{\sum_{t = 1}^N (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2)}, \ t = 1,\dotsc, N. \label{eq:randompick} \end{align} \end{algorithmic} \end{algorithm} \subsection{Proof of Lemma \ref{lem:err_bound}} \label{sec:prooflem:err_bound} \begin{proof}[Proof of Lemma \ref{lem:err_bound}] We prove Lemma \ref{lem:err_bound} for the case of discrete $\xi$, note that the proof still holds for the case of continuous $\xi$ just by using probability density function to replace probability distribution. Without the loss of generality, we assume a fixed update order in Algorithm \ref{alg:c_sbcd}: $ \pi_{t}^{i} = i, $ for all $i$ and $t$. Let $\Xi_t = \{\xi_{k,1},\xi_{k,2},\dotsc,\xi_{k,m_t}\}$ be any mini-batch samples in the $t$-th iteration. Let $\tilde g_{\Xi_{t},t}^i = \frac{1}{m_t} \sum_{j = 1}^{m_t} \nabla_{x^{i}}F(x_{t + 1}^{< i}, x_{t}^{\geq i}; \xi_{t,j})$ and $g_{\Xi_{t},t}^i = \nabla_{x^{i}}f(x_{t + 1}^{< i}, x_{t}^{\geq i})$, and $x_{\Xi_{t},t + 1}^{i} = x_{t}^{i} - \beta_t^i \tilde g_{\Xi_{t},t}^i$. Then we have \begin{align} \mathbb{E}[\tilde g_{\Xi_{t},t}^i | \mathbf \Xi_{[t - 1]}] = & \mathbb E_{\Xi_{t}} \left[\frac{1}{m_t} \sum_{j = 1}^{m_t} \nabla_{x^{i}}F(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_{t,j}) \right]\\ = & \sum_{\xi_{1},\dotsc,\xi_{m_t}}\Pr(\Xi_t = \{\xi_{1},\xi_{2},\dotsc,\xi_{m_t}\}) \frac{1}{m_t} \sum_{j = 1}^{m_t} \nabla_{x^i}F(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_{j}), \label{eq:exp_tg} \end{align} and \begin{align} \mathbb{E}[g_{\Xi_{t},t}^i | \mathbf \Xi_{[t - 1]}] = & \mathbb E_{\Xi_{t}} \left[\nabla_{x^{i}}f(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}) \right]\\ =& \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \nabla_{x^{i}}f(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i})\\ =& \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \sum_{\xi_l}\Pr(\xi = \xi_l) \nabla_{x^{i}}F(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_l)\\ =& \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \sum_{\xi_{1},\dotsc,\xi_{m_t}} \Pr(\Xi_t = \{\xi_{1},\xi_{2},\dotsc,\xi_{m_t}\}) \\ & \frac{1}{m_t} \sum_{j = 1}^{m_t} \nabla_{x^{i}}F(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j). \label{eq:exp_g} \end{align} Combine Eq.~\eqref{eq:exp_tg} and Eq.~\eqref{eq:exp_g}, we can obtain the expectation of $\Delta_t^i$ as \begin{align} \mathbb{E}[\Delta_t^i | \mathbf \Xi_{[t - 1]}] = &\mathbb{E}[\tilde g_t^i - g_t^i | \mathbf \Xi_{[t - 1]}] \\ =& \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \sum_{\xi_{1},\dotsc,\xi_{m_t}} \Pr(\Xi_t = \{\xi_{1},\xi_{2},\dotsc,\xi_{m_t}\}) \\ & \frac{1}{m_t} \sum_{j = 1}^{m_t} (\nabla_{x^{i}}F(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j) - \nabla_{x^{i}}F(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j)) \\ =& \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \sum_{\xi_{1},\dotsc,\xi_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \Pr(\Xi_t = \{\xi_{1},\xi_{2},\dotsc,\xi_{m_t}\}) \\ & \frac{1}{m_t} \sum_{j = 1}^{m_t} (\nabla_{x^{i}}F(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j) - \nabla_{x^{i}}F(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j)). \label{eq:exp_delta} \end{align} Note that, since the objection function is Lipschitz smoothness, $F(x; \xi)$ is also Lipschitz smoothness if $\Pr(\xi) > 0$, and we use $L$ to denote the maximum Lipschitz constant for all $\Pr(\xi) > 0$. Similarly, we can also obtain the gradient of $F(x; \xi)$ is also bounded using same analysis, and we use $G$ to denote the maximum bound for all $\Pr(\xi) > 0$. Using these two fact, we have \begin{align} & \|\nabla_{x^{i}}F(x_{\Xi_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j) - \nabla_{x^{i}}F(x_{\Xi'_{t},t + 1}^{< i}, x_{t}^{\geq i}; \xi_j)\|_2 \\ \leq & L \|x_{\Xi_{t},t + 1}^{< i} - x_{\Xi'_{t},t + 1}^{< i}\|_2\\ \leq &\sum_{l < i} L \|x_{\Xi_{t},t + 1}^{l} - x_{\Xi'_{t},t + 1}^{l}\|_2\\ \leq & \sum_{l < i} L\beta_t^l \|\tilde g_{\Xi_{t},t}^l - \tilde g_{\Xi'_{t},t}^l\|_2\\ \leq & 2 L b G \beta_t^{\max}. \label{eq:exp_delta_bound} \end{align} Combine Eq.~\eqref{eq:exp_delta} and Eq.~\eqref{eq:exp_delta_bound}, we complete the proof as \begin{align} & \mathbb{E}[\Delta_t^i | \mathbf \Xi_{[t - 1]}] \\ \leq & \sum_{\xi'_{1},\dotsc,\xi'_{m_t}} \sum_{\xi_{1},\dotsc,\xi_{m_t}} \Pr(\Xi'_t = \{\xi'_{1},\xi'_{2},\dotsc,\xi'_{m_t}\}) \Pr(\Xi_t = \{\xi_{1},\xi_{2},\dotsc,\xi_{m_t}\}) 2 L b G \beta_t^{\max}\\ = & 2 L b G \beta_t^{\max}, \end{align} where the last equation follows from the Law of total probability. This completes the proof. \end{proof} \subsection{Proof of Theorem \ref{thm:cyclicbcd}} \label{sec:proof:thm:cyclicbcd} Let $\beta_t^{\max} \coloneqq \max_i \beta_t^{i}$, $\beta_t^{\min} \coloneqq \min_i \beta_t^{i}$. To establish the convergence rate analysis, we start with Lemma~\ref{lem:vecdoterr}. \begin{lemma} \label{lem:vecdoterr} Let $u_k$ be a random vector that only depends on $\mathbf \Xi_{[t - 1]}$. If $u_t$ is independent of $\Delta_t^i$, then \begin{align} \mathbb E [\langle u_t, \Delta_t^i \rangle] \leq A \beta_t^{\max} \mathbb E [\|u_t\|_2], \end{align} where $A$ is defined in Eq~\eqref{eq:stocerrbd}. \end{lemma} Now, it is ready to discuss the main convergence properties of the nonconvex Cyclic SBCD algorithm (Algorithm \ref{alg:c_sbcd}) and provide the rate of convergence for that. \begin{proof}[Proof of Lemma \ref{lem:vecdoterr}] We can obtain the result of Lemma \ref{lem:vecdoterr} by follows \begin{align} \mathbb{E} [\langle u_t, \Delta_t^i \rangle] = & \mathbb{E}_{\mathbf \Xi_{[t - 1]}} \left[\mathbb{E}[ \langle u_t, \Delta_t^i \rangle | \mathbf \Xi_{[t - 1]}]\right]\\ \overset{\text{(a)}} = & \mathbb{E}_{\mathbf \Xi_{[t - 1]}} \left[ \langle \mathbb{E}[ u_t | \mathbf \Xi_{[t - 1]}], \mathbb{E}[ \Delta_t^i | \mathbf \Xi_{[t - 1]}] \rangle \right]\\ \leq & \mathbb{E}_{\mathbf \Xi_{[t - 1]}} \left[ \| \mathbb{E}[ u_t | \mathbf \Xi_{[t - 1]}]\|_2 \cdot \|\mathbb{E}[ \Delta_t^i | \mathbf \Xi_{[t - 1]}] \|_2 \right]\\ \leq & A \beta_t^{\max} \mathbb{E}_{\mathbf \Xi_{[t - 1]}} \left[ \| \mathbb{E}[ u_t | \mathbf \Xi_{[t - 1]}]\|_2 \right]\\ \overset{\text{(b)}} \leq & A \beta_t^{\max} \mathbb{E} \left[ \|u_t\|_2 \right], \end{align} where (a) follows from the conditional independence between $u_t$ and $\Delta_t^i$, and (b) follows from Jensen's inequality. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:cyclicbcd}] From the Lipschitz smoothness, it holds that \begin{align} &f(x_{t + 1}^{\leq i}, x_{t}^{> i}) - f(x_{t + 1}^{< i}, x_{t}^{\geq i}) \\ \leq & \langle g_t^i, x_{t + 1}^i - x_{t}^i \rangle + \frac{L}{2}\|x_{t + 1}^i - x_{t}^i \|_2^2 \\ = & -\beta_t^i \langle g_t^i, \tilde g_t^i \rangle + \frac{L}{2}(\beta_t^i)^2\|\tilde g_t^i\|_2^2 \\ = & -(\beta_t^i - \frac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \frac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 - (\beta_t^i - L(\beta_t^i)^2)\langle g_t^i,\Delta_t^i \rangle \label{eq:lipsm} \end{align} where all the equations follow the definition of $\Delta_t^i$ and the update law of Algorithm~\ref{alg:c_sbcd}. Summing Eq.~\eqref{eq:lipsm} over $i$, then we obtain \begin{align} & f(x_{t + 1}) - f(x_{t}) \leq -\sum_{i = 1}^b(\beta_t^i - \frac{ L }{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \sum_{i = 1}^b \frac{ L }{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 - \sum_{i = 1}^b(\beta_t^i - L (\beta_t^i)^2)\langle g_t^i, \Delta_t^i \rangle. \label{eq:bcdsumming} \end{align} Use Lemma \ref{lem:vecdoterr}, we also have the following fact, \begin{align} \label{eq:expbound} \mathbb{E}[\langle g_t^i, \Delta_t^i \rangle] \leq \beta^{\max}_t A G. \end{align} Taking expectation over Eq.~\eqref{eq:bcdsumming}, we have \begin{align} & \mathbb{E}[ f(x_{t + 1})] - \mathbb{E}[ f(x_{t})] \\ \leq & -\sum_{i = 1}^b(\beta_t^i - \frac{ L }{2}(\beta_t^i)^2)\mathbb{E}[\|g_t^i\|_2^2] + \sum_{i = 1}^b(\beta_t^i - L(\beta_t^i)^2)\beta^{\max}_t A G + \sum_{i = 1}^b\frac{ L }{2}(\beta_t^i)^2\mathbb{E}[\|\Delta_t^i\|_2^2] \\ \leq & -(\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^b \mathbb{E}[\|g_t^i\|_2^2] + \sum_{i = 1}^b\left( (\beta^{\max}_t)^2 A G + \frac{ L }{2}(\beta_t^i)^2\sigma^2 \right), \label{eq:bcdexp} \end{align} where the first inequality follows from Eq.~\eqref{eq:expbound}, and the second inequality follows from the boundedness of $\mathbb{E}[\|\Delta_t^i\|_2]$ and $\mathbb{E}[\|g_t^i\|_2]$. Rearranging Eq.~\eqref{eq:bcdexp}, we obtain \begin{align} &(\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^b \mathbb{E}[\|g_t^i\|_2^2] \leq \mathbb{E}[ f(x_{t})] - \mathbb{E}[ f(x_{t + 1})] + \sum_{i = 1}^b\left( (\beta^{\max}_t)^2 A G + \frac{ L }{2}(\beta_t^i)^2\sigma^2 \right). \label{eq:bcdrearr} \end{align} Also, we have \begin{align} \mathbb{E}[\|\nabla_{x^i} f(x_t)\|_2] \leq & \mathbb{E}[\|\nabla_{x^i} f(x_t) - g_t^i\|_2] + \mathbb{E}[\|g_t^i\|_2] \\ \overset{\text{(a)}} \leq & L \mathbb{E}[\|x_{t + 1}^{<i} - x_t^{<i}\|_2] + \mathbb{E}[\|g_t^i\|_2] \\ \overset{\text{(b)}} \leq & L \mathbb{E}\left[ \sqrt{\sum_{j < i} \| \beta_t^j \tilde g_t^j \|_2^2 } \right] + \mathbb{E}[\|g_t^i\|_2] \\ \leq & L \beta_t^{\max} \mathbb{E}\left[ \sqrt{\sum_{j < i} \|\tilde g_t^j \|_2^2 } \right] + \mathbb{E}[\|g_t^i\|_2] \\ \overset{\text{(c)}} \leq & L \beta_t^{\max} \sqrt{\mathbb{E}[\sum_{j < i} \|\tilde g_t^j \|_2^2 ]} + \mathbb{E}[\|g_t^i\|_2] \\ \overset{\text{(d)}} \leq & L \beta_t^{\max} \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + \mathbb{E}[\|g_t^i\|_2], \label{eq:gradieq} \end{align} where (a) follows from the Lipschitz smoothness of $f$, (b) follows from $x_{t + 1}^j = x_t^j - \beta_t^j \tilde g_t^j$, (c) follows from Jenson's inequality, and (d) follows from the boundedness of gradient and boundedness of variance. Summing Eq.~\eqref{eq:gradieq} over $i$, we can obtain \begin{align} & \mathbb{E}[\|\nabla f(x_t) \|_2^2] = \sum_{i = 1}^{b}\mathbb{E}[\|\nabla_{x^i} f(x_t)\|_2]\leq \sum_{i = 1}^{b} L \beta_t^{\max} \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + \sum_{i = 1}^{b} \mathbb{E}[\|g_t^i\|_2]. \label{eq:gradbd} \end{align} Combine Eq.~\eqref{eq:gradbd} with Eq.~\eqref{eq:bcdrearr}, we can obtain \begin{align} &(\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \mathbb{E}[\|\nabla f(x_t) \|_2^2]\\ \leq & (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^{b} L \beta_t^{\max} \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^{b} \mathbb{E}[\|g_t^i\|_2]\\ \leq & \mathbb{E}[ f(x_{t})] - \mathbb{E}[ f(x_{t + 1})] + (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^{b} L \beta_t^{\max} \sqrt{\sum_{j < i} (G^2 + \sigma^2) }\\ & + \sum_{i = 1}^b\left( (\beta^{\max}_t)^2 A G + \frac{ L }{2}(\beta_t^i)^2\sigma^2 \right), \label{eq:eachite} \end{align} where the first inequality follows from substituting Eq.~\eqref{eq:gradbd} into the left-hand side of Eq.~\eqref{eq:bcdrearr}, and the second inequality follows from substituting Eq.~\eqref{eq:gradbd} into the right-hand side of Eq.~\eqref{eq:bcdrearr}. Summing Eq.~\eqref{eq:eachite} over $t$, we have \begin{align} &\sum_{t = 1}^N (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \mathbb{E}[\|\nabla f(x_t) \|_2^2] \\ \leq & f(x_{1}) - f(x^*) \\ & + \sum_{t = 1}^N \left[(\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2) \sum_{i = 1}^{b} L \beta_t^{\max} \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + \sum_{i = 1}^b\left( (\beta^{\max}_t)^2 A G + \frac{ L }{2}(\beta_t^i)^2\sigma^2 \right)\right] \\ \leq & f(x_{1}) - f(x^*) + \sum_{t = 1}^N(\beta_t^{\max})^2 C_t. \label{eq:bcdfinalsum} \end{align} where $C_t$ is \begin{align} C_t = & (1 - \frac{ L }{2}\beta_t^{\max}) \sum_{i = 1}^{b} L \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + b\left( A G + \frac{ L }{2} \sigma^2\right). \end{align} Using the probability distribution of $R$ given in Eq.~\eqref{eq:randompick}, we completes the proof. \end{proof} \begin{proof}[Proof of Corollary \ref{cor:cyclicrate}] Combine these conditions with Eq.~\eqref{eq:bcdfinalsum}, we have \begin{align} &\sum_{t = 1}^N (\beta^{\min} - \frac{ L }{2}(\beta^{\max})^2) \mathbb{E}[\|\nabla f(x_t) \|_2^2] \leq f(x_{1}) - f(x^*) + N(\beta^{\max})^2 C, \end{align} where $C$ is \begin{align} C = & (1 - \frac{ L }{2}\beta^{\max}) \sum_{i = 1}^{b} L \sqrt{\sum_{j < i} (G_j^2 + \sigma^2) } + b\left( A G + \frac{ L }{2} \sigma^2 \right). \end{align} Using the probability distribution of $z$ given in Eq.~\eqref{eq:randompick}, we can obtain \begin{align} &\mathbb{E}\left[\|\nabla f(x_z)\|_2^2\right] \leq \frac{ f(x_1) - f^* + N (\beta^{\max})^2 C}{N (\beta^{\min} - \frac{ L }{2}(\beta^{\max})^2)}. \end{align} Thus, we can reach the rate of convergence of $\mathcal O(1 / \sqrt{N})$ by setting $\beta^{\min} = \beta^{\max} = \mathcal O(1 / \sqrt{N})$. \end{proof} \section{RCPG and SGA Algorithm} \label{sec:rcpg-sga} \subsection{Randomized Stochastic Block Coordinate Descent Algorithm} \label{sec:rcpg} We propose the randomized stochastic block coordinate descent algorithm as Algorithm~\ref{alg:sbmdpg}. Note that we also use the same notation about gradient from Eq.~\eqref{def:tildeg_y} and Eq.~\eqref{def:tildeg_theta} with very a tiny difference in practical, where $y_{t + 1} = y_{t}$ in Eq.~\eqref{def:tildeg_theta}. \begin{algorithm}[htb!] \caption{Risk-Sensitive Randomized Coordinate Descent Policy Gradient (RCPG)} \label{alg:sbmdpg} \centering \begin{algorithmic}[1] \STATE {\bfseries Input:} Stepsizes $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$, let $\beta_t^{\max} = \max\{\beta_t^\theta, \beta_t^y\}$.\\ \textbf{Option I:} $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$ satisfy the Robbins-Monro condition.\\ \textbf{Option II:} $\beta_t^\theta$ and $\beta_t^y$ are set to be constants. \FOR {episode $t=1,\dotsc,N$} \FOR {time step $k=1,\dotsc,{\tau_t}$} \STATE Compute $a_k \sim \pi_\theta(a|s_k)$, observe $r_k, s_{k+1}$. \ENDFOR \STATE Compute \begin{align} {R_t} &= \sum_{k = 1}^{\tau_t} {{r_k}} \\ {\omega _t}(\theta_t) &= \sum _{k=1}^{{\tau_t}}{\nabla_\theta \ln{\pi_{\theta_t}}({a_{k}}|{s_{k}})}. \end{align} \STATE Randomly select $i_t \in \{1,2\}$ with distribution $[0.5,0.5]$. If $i_t=1$, \begin{align} {y_{t+1}} &= {y_t} + {\beta _t}\left(2R_t + \frac{1}{\lambda} - 2y_t\right),\\ {\theta_{t+1}} &= {\theta _t}. \end{align} else \begin{align} {y_{t+1}} &= {y_t},\\ {\theta_{t+1}} &= {\theta _t} + {\beta _t}\left( {2y_t{R_t} - {{({R_t})}^2}} \right){\omega _t}({\theta _t}). \end{align} \ENDFOR \STATE {\bfseries Output} $\bar{x}_N$: \\ \textbf{Option I:} Set $\bar{x}_N = x_N$. \\ \textbf{Option II:} Set $\bar{x}_N = x_z$, where $z$ is uniformly drawn from $\{ 1,2, \dotsc ,N\}$. \end{algorithmic} \end{algorithm} Note that, the main difference between Cyclic SBCD and Randomized SBCD is that: at each iteration, Cyclic SBCD cyclically updates all blocks of variables, and the later updated blocks depending on the early updated blocks; while Randomized SBCD randomly chooses one block of variables to update. \subsection{Risk-Sensitive Stochastic Gradient Ascent Policy Gradient} \label{sec:sga} We also proposed risk-sensitive stochastic gradient Ascent policy gradient as Algorithm \ref{alg:sgapg}. \begin{algorithm}[htb!] \caption{Risk-Sensitive Stochastic Gradient Ascent Policy Gradient (SGA)} \label{alg:sgapg} \centering \begin{algorithmic}[1] \STATE {\bfseries Input:} Stepsizes $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$, let $\beta_t^{\max} = \max\{\beta_t^\theta, \beta_t^y\}$.\\ \textbf{Option I:} $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$ satisfy the Robbins-Monro condition.\\ \textbf{Option II:} $\beta_t^\theta$ and $\beta_t^y$ are set to be constants. \FOR {episode $t=1,\dotsc,N$} \FOR {time step $k=1,\dotsc,{\tau_t}$} \STATE Compute $a_k \sim \pi_\theta(a|s_k)$, observe $r_k, s_{k+1}$. \ENDFOR \STATE Compute \begin{align} {R_t} &= \sum_{k = 1}^{\tau_t} {{r_k}} \\ {\omega _t}(\theta_t) &= \sum _{k=1}^{{\tau_t}}{\nabla_\theta \ln{\pi_{\theta_t}}({a_{k}}|{s_{k}})}. \end{align} \STATE Update parameters, \begin{align} {y_{t+1}} &= {y_t} + {\beta _t}\left(2R_t + \frac{1}{\lambda} - 2y_t\right),\\ {\theta_{t+1}} &= {\theta _t} + {\beta _t}\left( {2y_t{R_t} - {{({R_t})}^2}} \right){\omega _t}({\theta _t}). \end{align} \ENDFOR \STATE {\bfseries Output} $\bar{x}_N$: \\ \textbf{Option I:} Set $\bar{x}_N = x_N$. \\ \textbf{Option II:} Set $\bar{x}_N = x_z$, where $z$ is uniformly drawn from $\{ 1,2, \dotsc ,N\}$. \end{algorithmic} \end{algorithm} \section{Details of the Experiments} \label{sec:parameters} The parameter settings for portfolio management domain are as follows: $\tau = 50$, $r_l = 1.001$, $r_{\text{nl}}^{\text{high}} = 2$, $r_{\text{nl}}^{\text{low}} = 1.1$, $p_{\text{risk}} = 0.05$, $p_{\text{switch}} = 0.1$, $W = 4$, $\eta = 0.2$, startup cash $\$100,000$. The parameter settings of American-style Option domain are as follows: $K_{\text{put}} = 1$, $K_{\text{call}} = 1.5$, $x_0 = 1.25$, $f_u = 9 / 8$, $f_d = 8 / 9$, $p = 0.45$, $\tau = 20$. The parameter settings of optimal stopping domain are as follows: $x_0 = 1.25$, $f_u = 2$, $f_d = 0.5$, $p = 0.65$, $\tau = 20$. \section{Introduction} Risk management plays a central role in sequential decision-making problems, common in fields such as portfolio management~\citep{lai2011mean}, autonomous driving~\citep{risk:autonomous:maurer2016}, and healthcare~\citep{risk:american2011}. The risk is typically depicted by the variance of the expected sum of returns, and the basic motivation of risk-sensitive approaches is therefore to incorporate the variance-control term into the objective function. The mean-variance trade-off function~\citep{sobel1982variance,saferl:mannor2011} is one of the most widely used objective functions for variance control, but other risk-sensitive functions have also been studied. For example, Borkar ~\citep{saferl:borkar2002q} studied exponential utility functions, Tamar~\citep{saferl:castro2012} studied the Sharpe Ratio measurement, Chow~\citep{saferl:cvar:chow2014} investigated Conditional Value at Risk (CVaR) in a static setting, and Tamar~\citep{saferl:tamar2015coherent} investigated coherent risk for both linear and nonlinear system dynamics. Chow~\citep{saferl:cvar:chow2015} showed the equivalence between a coherent risk-sensitive objective and a robust MDP. Compared with the other widely used risk management performance measurements, such as the Sharpe Ratio and the CVaR measurement, the mean-variance measurement has explicit interpretability and computational advantages~\citep{meanvar:markowitz2000mean,meanvar:li2000optimal}. For example, the CVaR measurement fails to consider the mean return, and the Sharpe Ratio tends to lead to solutions with less mean return~\citep{saferl:castro2012}. Existing mean-variance reinforcement learning algorithms, however, suffer either heavy computational cost or slow convergence without any available finite-sample complexity analysis, which makes the algorithms difficult to be applied to real-world problems. This paper makes the following contributions. The first is a reformulation of the mean-variance function that offers a stochastic (block) coordinate descent (BCD) framework~\citep{cd:overview:wright2015} for risk-sensitive policy search. This framework enables using any off-the-shelf stochastic non-convex block coordinate descent solver~\citep{cd:bsg:xu2015,cd:sbmd:dang2015}, which has the potential for acceleration and variance reduction. To the best of our knowledge, this is the first time block coordinate descent has been introduced into reinforcement learning. The second contribution lies in the algorithmic development. This paper proposes a computationally efficient risk-sensitive algorithm that is both easy to implement in real-world applications, and with known computational complexity. Previous risk-sensitive RL algorithms are either computationally costly, difficult to tune with multiple time-scale stepsizes, or lacking in rigorous sample complexity analysis, which hinders the interest of reinforcement learning practitioners. This newly proposed algorithm will widen the application scenarios of risk-sensitive RL algorithms. Here is a roadmap for the rest of the paper. Section~\ref{sec:backgrounds} offers a brief background on risk-sensitive reinforcement learning and stochastic variance reduction. In Section~\ref{sec:alg}, the problem is reformulated by using the Fenchel duality, and a novel algorithm is proposed based on stochastic block coordinate descent. Section~\ref{sec:theory} details a theoretical analysis for both asymptotic convergence and finite-sample error bound analysis. The experimental study in Section~\ref{sec:experimental} validates the effectiveness of the proposed algorithm. \section{Backgrounds} \label{sec:backgrounds} This section offers a brief overview of risk-sensitive reinforcement learning, including the objective functions and algorithms. We then introduce block coordinate descent methods. Finally, we introduce the Fenchel duality, a key element in formulating the new algorithm. \subsection{Risk-Sensitive Reinforcement Learning} \label{sec:riskrl} Reinforcement Learning (RL)~\citep{ndp:book,sutton-barto:book} is a class of learning problems in which an agent interacts with an unfamiliar, dynamic, and stochastic environment, where the agent's goal is to optimize some measure of its long-term performance. This interaction is conventionally modeled as a Markov decision process (MDP), defined as the tuple $({\mathcal{S},\mathcal{A},P_{ss'}^{a},r,\gamma})$. $\mathcal{S}$ and $\mathcal{A}$ are the sets of states and actions, the transition kernel $P_{ss'}^{a}$ specifying the probability of transition from state $s\in\mathcal{S}$ to the successor state $s'\in\mathcal{S}$ by taking action $a\in\mathcal{A}$, $r(s,a):\mathcal{S}\times\mathcal{A}\to\mathbb{R}$ is the reward function bounded by $R_{\max}$, and $0\leq\gamma<1$ is a discount factor. A \textit{parameterized policy function} $\pi_\theta(a|s):\mathcal{S}\times\mathcal{A}\to\left[{0,1}\right]$ is a probabilistic mapping from states to actions, where $\theta$ is the tunable parameter and $\pi_\theta(a|s)$ is a differentiable function with respect to (w.r.t.) $\theta$. Different probability distributions over $\mathcal{A}$ are associated with each $s\in \mathcal{S}$ for different values of $\theta$. One commonly used performance measure of policies is the finite-horizon cumulative reward from the starting state: $ R = \sum\nolimits_{k = 1}^{{\tau}} {r({s_k},{a_k})} $, where $\tau$ is the first passage time to the recurrent state $s^*$~\citep{puterman,saferl:castro2012}, and therefore $\tau : = \min \{ k > 0|{s_k} = {s^*}\}$. An RL algorithm aims to find a near-optimal policy that maximizes the expected sum of rewards: $ J(\theta) := \mathbb{E}_{\pi_\theta}[R] = \mathbb{E}_{\pi_\theta}[\sum\limits_{k = 1}^{{\tau}} {r({s_k},{a_k})}]. $ We also define $M(\theta ): = {\mathbb{E}_{{\pi _\theta }}}[{R^2}] = {\mathbb{E}_{{\pi _\theta }}}[{(\sum\limits_{k = 1}^{\tau} {r({s_k},{a_k})} )^2}]$. In the following, we drop the subscription $\pi_\theta$ for simplicity. In risk-sensitive reinforcement learning and control, the optimization criterion is transformed into the \textit{mean-variance} trade-off ${J_\lambda(\theta) }$, defined as $ {J_\lambda }(\theta ): = J(\theta ) - \lambda {\rm{Var}}(R), $ where $\lambda >0$ is the risk-control parameter, and ${\rm {Var}}(R) = M(\theta) - J^2(\theta)$ measures the variance of $R$. So ${J_\lambda }(\theta )$ can be formulated as \begin{equation} {J_\lambda }(\theta ): = J(\theta ) - \lambda (M(\theta ) - {J^2}(\theta )),\quad \lambda > 0 \label{eq:regobj} \end{equation} It should be noted that the mean-variance objective function is \textit{NP-hard} in general~\citep{saferl:mannor2011}, and the underlying problem is not necessarily convex ~\citep{sobel1982variance,saferl:mannor2011}. Most current methods to maximize such objective functions are based on the ordinary differential equation (ODE)---which is based on the stochastic approximation method~\citep{borkar:book}---such as \cite{saferl:prashanth2013,saferl:castro2012}. For example, Tamar~\cite{saferl:castro2012} used a two-time-scale ODE approach to maximize the mean-variance objective function. However, these approaches suffer from certain drawbacks. First, ODE-based methods can provide only an asymptotic convergence guarantee, with no sample complexity analysis. It is important to present algorithms with finite-sample analysis, which offers confidence to reinforcement learning practitioners. Second, it is well known that two-time-scale approaches are very sensitive to stepsize tuning, which is a non-trivial burden in real-world practice. Third, the ODE approach does not allow extra penalty functions. Adding penalty functions can strengthen the robustness of the algorithm, encourage sparsity, and incorporate prior information~\citep{esl:ElementsStatisticalLearning}. This is also a benefit of first-order proximal approaches~\citep{proximal:parikh2013}. \subsection{Coordinate Descent Optimization} \label{sec:bcd} Coordinate descent (CD) and the more general block coordinate descent (BCD) algorithms solve minimization problems by iteratively updating variables along coordinate directions or coordinate hyperplanes~\citep{cd:overview:wright2015}. At each iteration of BCD, the objective function is (approximately) minimized w.r.t. one coordinate or a block of coordinates by fixing the remaining ones, and thus an easier lower-dimensional subproblem is solved. A number of comprehensive studies on BCD have already been carried out, such as ~\citep{cd:bcd:luo1992,cd:sbcd:nesterov2012} for convex problems, ~\citep{cd:bcd:tseng2001,xu2013block, razaviyayn2013unified} for non-convex cases, and the review paper ~\citep{cd:overview:wright2015}. For stochastic problems with a block structure, \citep{cd:sbmd:dang2015} proposed stochastic block mirror descent (SBMD) by combining BCD with stochastic mirror descent \citep{MID:2003,nemirovski2009robust}. Another line of research along this topic is block stochastic gradient coordinate descent (BSG)~\citep{cd:bsg:xu2015}. The key difference between SBMD and BSG is that at each iteration, SBMD randomly picks one block of variables to update, while BSG cyclically updates all block variables. Theoretically, SBMD assumes unbiased estimation of the stochastic approximation of a partial gradient, and has both asymptotic convergence and finite-sample analysis results for convex and also non-convex problems. Conversely, BSG makes a biased stochastic gradient approximation. Specifically, the stochastic gradient approximation of the first block of variables is unbiased, and all the rest are biased, which brings extra difficulties for conducting the finite-sample analysis. However, BSG numerically outperforms SBDM because the former has lower per-epoch complexity and the cyclic update leads to Gauss-Seidel type acceleration. Our algorithm inherits advantages from both SBMD and BSG.\footnote{Because the problem we consider is a maximization problem, our algorithm is a block coordinate \emph{ascent} method.} It has the same per-epoch complexity as BSG and also enjoys the same superior finite-sample analysis results as SBMD. \subsection{Fenchel Duality} Next, we introduce the Fenchel duality. A closed and convex function $f(z)$ can be represented by its Fenchel dual formulation~\citep{boyd}: $ f(z) = \mathop {\max }\limits_y ({z^\top}y - {f^*}(y)), $ where $f^*$ is the convex conjugate function~\citep{boyd}. Especially, for $z \in \mathbb{R}$, the convex conjugate function of a power function $f(z)={\frac {1}{p}}|z|^{p},\,1<p<\infty$ is $f^{* }\left(z\right)={\frac {1}{q}}|x|^{q},\,1<q<\infty$, where ${\frac {1}{p}}+{\frac {1}{q}}=1$. It is easy to obtain when $p=2$ that $ f(z) = {f^{*}}(z) = \frac{1}{2}{z^2} $. So for the function $z^2, z \in \mathbb{R}$, we have the Fenchel dual formulation as follows \begin{align} {z^2} = \mathop {\max }\limits_{y\in \mathbb{R}} (2zy - {y^2}). \label{eq:fenchel-square} \end{align} \section{Algorithm Design} \label{sec:alg} In this section, we first introduce the mean-variance function formulation and explain why it is difficult to use the regular stochastic gradient for maximization. Then we offer a reformulation of the objective function, and a novel algorithm based on recent stochastic non-convex block coordinate descent with in-depth theoretical analysis. \subsection{Problem Formulation} In this section, we describe why the vanilla stochastic gradient cannot be applied to maximize $J_\lambda(\theta)$ defined in Eq.~\eqref{eq:regobj}. Taking gradient of $J_\lambda(\theta)$ w.r.t. $\theta$, we have \begin{equation} \nabla_\theta {J_\lambda }({\theta _t}) = {\nabla _\theta }J({\theta _t}) - \lambda {\nabla _\theta }{\rm{Var}}(R). \label{objgradient} \end{equation} The sample-based estimation of $\nabla_{\theta} {\rm{Var}}(R)$ is computed as \begin{align} \nabla_{\theta} {\rm{Var}}(R) = \nabla_{\theta} M(\theta) - 2J({\theta})\nabla_{\theta} J({\theta}). \label{eq:nablavar} \end{align} The computation of Eq.~\eqref{objgradient} therefore involves three items: $\nabla_{\theta} J({\theta}), \nabla_{\theta} M({\theta})$, and $J({\theta})\nabla_{\theta} J({\theta})$. An episode is the trajectory between two visits of the recurrent state $s^*$---for example, the $t$-th episode refers to the trajectory between the ($t$-1)-th and the $t$-th visit, $\tau_t$ denotes the length of the $t$-th episode, and $\mathbb{E}_t[\tau_t]=\tau$. For the $t$-th episode, unbiased estimations of $M(\theta), {\nabla_{\theta}}J(\theta)$ and ${\nabla_{\theta}}M(\theta)$ can be obtained from a single trajectory via the likelihood ratio method~\citep{reinforce:williams1992} via $ {\nabla_{\theta}}J({\theta}) = \mathbb{E}[{R_t}\omega_{t}(\theta)],\ {\nabla_{\theta}}M({\theta}) = \mathbb{E}[({R_t})^{2}\omega_{t}(\theta)] $. $R_t$ is the cumulative reward for the $t$-th episode, i.e., $R_t = \sum\nolimits_{k=1}^{{\tau_t}} r_k$, which is possibly a \textit{non-convex} function, and $\omega_{t}(\theta)$ is the likelihood ratio derivative. The latter is computed as \begin{align} {\omega _t}(\theta ) = \sum\limits _{k=1}^{{\tau_t}}{\nabla_{\theta}\ln{\pi_{\theta}}({a_{k}}|{s_{k}})}. \label{eq:omegat} \end{align} However, the sample-based estimation of $J(\theta)\nabla_{\theta} J(\theta)$ is difficult to obtain in practice. If a generative model is available---for example, for every state-action pair $(s, a)$---we can sample twice (or even several times) from the underlying Markov chain to obtain the successive state $s'$, and thus the sample-based estimation of $J(\theta)\nabla_{\theta} J(\theta)$ is available. However, if a generative model is not available, then the computation of $J(\theta)\nabla_{\theta} J(\theta)$ requires double sampling (sampling twice from two different trajectories) and cannot be exactly estimated from a single trajectory, as also noted in \cite{saferl:castro2012}. To this end, \cite{saferl:castro2012} proposed a two-time-scale ODE-based stochastic approximation algorithm that circumvents the double-sampling problem. However, as discussed in Sec.~\ref{sec:riskrl}, this approach suffers from several weaknesses such as no available finite-sample analysis and difficult-to-tune two-time-scale stepsizes. To overcome these weaknesses, we reformulate the problem and introduce a novel algorithm with in-depth analysis based on this reformulation. \subsection{Block Coordinate Reformulation} Fenchel duality is a powerful tool for avoiding double sampling. Recent research~\citep{liu2015uai,vr:saddle:pe:du2017} has formulated a class of reinforcement learning problems as convex-concave saddle-point problems by introducing the Fenchel-dual to avoid the double sampling problem in off-policy policy evaluation. In this section, we present a new formulation of $J_\lambda$ defined in Eq.~\eqref{eq:regobj} and novel algorithms to avoid the double sampling problem of $J(\theta)\nabla_{\theta} J(\theta)$. Let \begin{align} {F_\lambda }(\theta ) := {\left[ {J(\theta ) + \frac{1}{{2\lambda }}} \right]^2} - M(\theta ). \label{eq:lprimal} \end{align} It is easy to verify ${F_\lambda }(\theta )=\frac{{{J_\lambda }(\theta)}}{\lambda } + \frac{1}{{4{\lambda ^2}}}$. Since $\lambda>0$ and is a constant, maximizing $J_\lambda(\theta)$ is equivalent to maximizing ${F_\lambda }(\theta )$. Using Eq.~\eqref{eq:lprimal} with $z = J(\theta ) + \frac{1}{2\lambda}$, we reformulate ${F_\lambda }(\theta )$ as \begin{align} {F_\lambda }(\theta) = \mathop {\max }\limits_y \left( {2y(J(\theta) + \frac{1}{{2\lambda }}) - {y^2}} \right) - M(\theta). \end{align} This way, the maximization problem $\mathop {\max }\limits_{\theta} {F_\lambda }(\theta )$ is equivalent to $\mathop {\max }\limits_{\theta ,y} {\hat f_\lambda }(\theta ,y)$, where ${\hat f_\lambda }(\theta ,y)$ is defined as \begin{align} {{{\hat f}_\lambda }(\theta ,y) := 2y(J(\theta ) + \frac{1}{{2\lambda }}) - {y^2} - M(\theta )}. \label{eq:max-alt} \end{align} In the following sections, we denote the stepsize of $\theta$ as $\beta_t^\theta$, and the stepsize of $y$ as $\beta_t^y$. The subscript $t$ denotes the episode number, and the subscript $k$ denotes the time step number. \subsection{Mean-Variance Proximal Policy Gradient} In this section, we present a block coordinate update algorithm to maximize Eq.~\eqref{eq:max-alt}. The block gradient can be directly computed and denoted as \begin{align} \label{def:first} g_t^y := {\nabla_{y}}{\hat{f}_{\lambda}}(\theta_t, y_{t}) &= 2J(\theta_t)+\frac{1}{{\lambda}} - 2y_t,\\ \label{def:g_theta} g_t^\theta := \nabla_{\theta}{\hat{f}_{\lambda}}(\theta_t, y_{t + 1}) &= 2{y_{t+1}}{\nabla_{\theta}}J(\theta_t)-{\nabla_{\theta}}M(\theta_t). \end{align} We denote $\tilde{g}_t^\theta$ (resp. $\tilde{g}_t^y$) as the sample-based estimation of $g_t^\theta$ (resp. $g_t^y$), which is \begin{align} \Tilde{g}^y_t = 2 R_t + \frac{1}{\lambda } - 2y_t,\\ \label{def:tildeg_theta} \Tilde{g}^\theta_t = \left( {2y_{t + 1}{R_t} - {{({R_t})}^2}} \right){\omega _t}({\theta _t}). \end{align} The block coordinate update rule is \begin{align} \label{eq:updatelaw} {y_{t+1}} &= {y_{t}}+{\beta^y_{t}}{\tilde{g}_t^y},& {\theta_{t+1}} &= {\theta_{t}}+{\beta^\theta_{t}}{\tilde{g}_t^\theta}. \end{align} The Mean-Variance Proximal Policy Gradient (\textbf{MPPG}) algorithm is therefore formulated in Algorithm \ref{alg:cyclicbcd}. \begin{algorithm}[htb] \caption{Mean-Variance Proximal Policy Gradient (\textbf{MPPG})} \label{alg:cyclicbcd} \centering \begin{algorithmic}[1] \STATE {\bfseries Input:} Stepsizes $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$, let $\beta_t^{\max} = \max\{\beta_t^\theta, \beta_t^y\}$.\\ \textbf{Option I:} $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$ satisfy the Robbins-Monro condition.\\ \textbf{Option II:} $\beta_t^\theta$ and $\beta_t^y$ are set to be constants. \FOR {episode $t=1,\ldots,N$} \FOR {time step $k=1,\ldots,{\tau_t}$} \STATE Compute $a_k \sim \pi_{\theta_t}(a|s_k)$, observe $r_k, s_{k+1}$. \ENDFOR \STATE Update parameters. \begin{align} {R_t} &= \sum\limits_{k = 1}^{\tau_t} {{r_k}} \\ {\omega _t}(\theta_t) &= \sum\limits _{k=1}^{{\tau_t}}{\nabla_\theta \ln{\pi_{\theta_t}}({a_{k}}|{s_{k}})}\\ {y_{t+1}} &= {y_t} + {\beta _t^y}\left(2 R_t + \frac{1}{\lambda } - 2y_t\right)\\ {\theta_{t+1}} &= {\theta _t} + {\beta _t^\theta}\left( {2y_{t+1}{R_t } - {{({R_t })}^2}} \right){\omega _t}({\theta _t}) \end{align} \ENDFOR \STATE {\bfseries Output} $\bar{x}_N$: \\ \textbf{Option I:} Set $\bar{x}_N = x_N = [\theta_N,y_N]^\top$. \\ \textbf{Option II:} Set $\bar{x}_N = x_z= [\theta_z,y_z]^\top$, where $z$ is uniformly drawn from $\{ 1,2, \ldots ,N\}$. \end{algorithmic} \end{algorithm} In practical implementations, we can either choose Option I with the final iteration's result as the output, or pick up one previous iteration's result randomly according to a discrete uniform distribution over $\{1,2,\ldots, N\}$. However, our theoretical analysis depends on both Options I and II. Here, we divide our theoretical analysis into two parts: asymptotic convergence based on Option I, and sample complexity based on Option II. \section{Theoretical Analysis} \label{sec:theory} Now we present the theoretical discussion of Algorithm~\ref{alg:cyclicbcd}, with both asymptotic convergence analysis and finite-sample error bound analysis. For the purpose of clarity, in the following analysis, $x$ is defined as $x := [y,\theta]^\top$, where $x^1 :=y,\ x^2 :=\theta$. Similarly, $g^i$ (resp. $\beta^i_t$) is used to denote $g^y$(resp. $\beta^y_t$) and $g^\theta$(resp. $\beta^\theta_t$), where $g_1 :=g^y,\ g_2 :=g^\theta$(resp. $\beta^1_t :=\beta^y_t,\ \beta^2_t :=\beta^\theta_t$). Let $\beta_{t}^{\max} = \max\{\beta_{t}^1,\beta_{t}^2\}$, $\beta_{t}^{\min} = \min\{\beta_{t}^1,\beta_{t}^2\}$, and $\|\cdot\|_2$ denotes the Euclidean norm. \subsection{Assumptions} First, we introduce the assumptions needed for the proof. \begin{assumption}(\textbf{Boundedness}) \label{asm:bound} There exists a constant $\rho$ such that $\mathbb{E}[\|x_t\|_2^2] \leq \rho^2$ for $\forall t$. The objective function $\hat{f}_{\lambda}$ is upper bounded, i.e., $\hat{f}_{\lambda} < +\infty$. \end{assumption} \begin{assumption}(\textbf{Lipschitz Smoothness}) \label{asm:lips} There is a uniform Lipschitz constant $L > 0$ such that \begin{align} \|\nabla_\theta \hat{f}_{\lambda}(x_1) - \nabla_\theta \hat{f}_{\lambda}(x_2)\|_2 \leq L\|x_1 - x_2\|_2, ~ \forall x_1,x_2; \nonumber \\ \|\nabla_y \hat{f}_{\lambda}(x_1) - \nabla_y \hat{f}_{\lambda}(x_2)\|_2 \leq L\|x_1 - x_2\|_2, ~ \forall x_1,x_2. \nonumber \end{align} \end{assumption} \begin{assumption}(\textbf{Bounded Variance and Bias}) \label{asm3} There exists a constant $A$ and a constant ${\sigma}$ such that for any $t$ we have \begin{align} &\mathbb{E}[\|\Delta^y_t\|_2^2] \leq {\sigma}^2,\quad \mathbb{E}[\|\Delta^\theta_t\|_2^2] \leq {\sigma}^2, \nonumber\\ &\|\mathbb{E}[\Delta^\theta_t |\{ {R_i},{\omega _i}\} _{i = 1}^{t - 1}]\|_2 \leq A\cdot \beta_t^{\max}. \end{align} where ${\Delta}^y_t, {\Delta}^\theta_t$ are defined as \begin{align} {\Delta}^y_t = \Tilde{g}^y_t - {g}^y_t ,\quad{\Delta}^\theta_t = \Tilde{g}^\theta_t - {g}^\theta_t. \label{def:last} \end{align} \label{asm:var} \end{assumption} We also need Assumption~\ref{asm:erg} on the induced Markov chain. \begin{assumption}(\textbf{Ergodicity}) Under all policies, the induced Markov chain is ergodic, i.e., irreducible, aperiodic, and recurrent. \label{asm:erg} \end{assumption} \begin{remark} Assumption \ref{asm:bound} is relatively weaker than the assumptions made in the literature on stochastic gradient methods. For example, \citep{lan2012optimal, nemirovski2009robust} assume $x_t$ is bounded in a fixed set $\mathcal{X}$. Note that Assumption \ref{asm:bound} together with the Lipschitz smoothness in Assumption \ref{asm:lips} implies the boundedness of $g_t^y$ and $g_t^\theta$ by the following arguments \begin{align} & \|\nabla_{x^i}\hat{f}_{\lambda}(x_t)\|_2^2 \nonumber\\ \leq &2\|\nabla_{x^i}\hat{f}_{\lambda}(x_t) - \nabla_{x^i}\hat{f}_{\lambda}(\mathbf{0})\|_2^2 + 2\|\nabla_{x^i}\hat{f}_{\lambda}(\mathbf{0})\|_2^2 \nonumber \\ \leq & 2L^2\|x_t\|_2^2 + 2\|\nabla_{x^i}\hat{f}_{\lambda}(\mathbf{0})\|_2^2, \end{align} where $\mathbf{0}$ is the null vector whose elements are all zeros. We denote $M_\rho$ as \begin{align} \label{def:mrho} M_\rho := \sqrt{2L^2\rho^2 + 2\max_i\|\nabla_{x^i}\hat{f}_{\lambda}(\mathbf{0})\|_2^2}. \end{align} Thus, we have $\mathbb{E}[\|g_t^y\|_2] \leq M_\rho, ~\mathbb{E}[\|g_t^\theta\|_2] \leq M_\rho$, for $\forall t$, by Jensen's inequality. Assumption \ref{asm:lips} is a standard assumption for stochastic gradient methods, especially for the non-convex case. Assumption \ref{asm3} assumes the bounded variance of $\mathbb{E}[\|\Delta^y_t\|_2^2] $ and $\mathbb{E}[\|\Delta^\theta_t\|_2^2] $, and the bias between $\Tilde{g}^\theta_t$ and ${g}^\theta_t$ is also bounded. Note that in the update law as shown in Eq.~\eqref{eq:updatelaw}, we have $\mathbb{E}[\tilde g_t^y] = g_t^y, \mathbb{E}[\tilde g_t^\theta ] \ne g_t^\theta $, as explained in~\cite{cd:bsg:xu2015}. This bias problem can be avoided by using the randomized stochastic block coordinate update\citep{cd:sbmd:dang2015} so that at each iteration either $y_{t+1}$ or $\theta_{t+1}$ (but only one) is updated. However, the cyclic update has better empirical performances as shown in Section~\ref{sec:experimental}, and a tighter finite-sample error bound as well. \end{remark} \subsection{Asymptotic Convergence Analysis} In this section, we provide the asymptotic convergence analysis for Algorithm~\ref{alg:cyclicbcd} with Option I, i.e., the stepsizes are chosen to satisfy the Robbins-Monro condition and the output is the last iteration's result. We first introduce Lemma~\ref{lem:1}, which is essential for proving Theorem~\ref{thm1}. \begin{lemma} \label{lem:1} For two non-negative scalar sequences $\{a_t\}$ and $\{b_t\}$, if $\sum_{t = 1}^{\infty}a_t = +\infty$ and $\sum_{t = 1}^{\infty}a_t b_t < +\infty$, we then have \begin{equation} \mathop {\lim }\limits_{t \to \infty } \inf {b_t} = 0 \end{equation} Further, if there exists a constant $K>0$ such that $|b_{t + 1} - b_{t}| \leq a_t K$, then \begin{equation} \lim_{t \rightarrow \infty}b_t = 0. \end{equation} \end{lemma} The detailed proof can be found in Lemma A.5 of \citep{mairal2013stochastic} and Proposition 1.2.4 of \citep{bertsekas1999nonlinear}. Now it is ready to prove Theorem~\ref{thm1}. \begin{theorem}[\textbf{Asymptotic Convergence}] \label{thm1} Let $\{(\theta_t, y_t)\}$ (i.e. $\{x_t\}$) be generated from Algorithm~\ref{alg:cyclicbcd} with Option I. If $\{\beta_t^\theta\}$, $\{\beta_t^y\}$ are time-diminishing real positive sequences satisfying the Robbins-Monro condition, i.e., if $\sum\limits_{t = 1}^\infty {\beta _t^\theta } = \infty, \sum\limits_{t = 1}^\infty {{{(\beta _t^\theta )}^2}} < \infty $, $\sum\limits_{t = 1}^\infty {\beta _t^y} = \infty, \sum\limits_{t = 1}^\infty {{{(\beta _t^y)}^2}} < \infty$, then Algorithm~\ref{alg:cyclicbcd} will converge to the stationary point: \begin{equation} \lim_{t\rightarrow \infty}\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t) \|_2] = 0. \end{equation} \end{theorem} \textbf{Proof Sketch} The key idea for proving Theorem \ref{thm1} is the utilization of Lemma \ref{lem:1}. First, we use Assumption \ref{asm:bound} and \ref{asm3} to introduce the boundedness of $\{\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t)\|_2]\}$. Then, using the boundedness of $\{\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t)\|_2]\}$, we prove $\sum_{t = 1}^{\infty} \beta_t^{\min} \mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2] < +\infty$, and using the first half of Lemma \ref{lem:1} we obtain $\mathop {\lim }\limits_{t \to \infty } \inf \mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t)\|_2^2] = 0$. We then show that $|\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_{t+1})\|_2^2] - \mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t)\|_2^2]| \leq K \cdot \beta_t^{\max}$ for some fixed constant $K$. Combining this with Lemma \ref{lem:1}, we obtain $\lim_{t \rightarrow \infty}\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t)\|_2^2] = 0$. The detailed proof is provided in the Appendix due to space constraints. \subsection{Finite-Sample Analysis} The above analysis provides asymptotic convergence guarantee of Algorithm~\ref{alg:cyclicbcd}, however, it is desirable to know the sample complexity of the algorithm in real applications. Motivated by offering RL practitioners confidence in applying the algorithm, we then present the sample complexity analysis with Option II described in Algorithm~\ref{alg:cyclicbcd}, i.e., the stepsizes are set to be a constant, and the output is randomly selected from $\{x_1,\cdots,x_N\}$ with a discrete uniform distribution. This is a standard strategy for non-convex stochastic optimization approaches \citep{cd:sbmd:dang2015}. With these algorithmic refinements, we are ready to present the finite-sample analysis as follows. \begin{theorem} \label{thm:cyclicbcd2} Let the output of the Algorithm~\ref{alg:cyclicbcd} be $\bar{x}_N = x_z$ as in Option II. If $\{\beta_t^{\theta}\}$, $\{\beta_t^{y}\}$ are chosen as constants satisfying $\beta_t^\theta,\beta_t^y < 2 / L$ for $t = 1,\cdots,N$, we have \begin{align} &\mathbb{E}\left[\|g_z^\theta\|_2^2 + \|g_z^y\|_2^2\right] \leq \dfrac{\hat{f}_\lambda^* - \hat{f}_\lambda(x_1) + N (\beta_t^{\max})^2 C}{N (\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)} \label{eq:thm2} \end{align} where $\hat{f}_{\lambda}^* = \max_{x} \hat f_\lambda (x)$, and \begin{align} C = A M_\rho + L\sigma^2 + 2L(1 + L\beta_t^{\max})(3 \sigma^2 + 2 M_\rho^2), \label{eq:C} \end{align} A is defined in Assumption \ref{asm:var} and $M_\rho$ is defined in Eq.~\eqref{def:mrho}. \end{theorem} \vspace{-0.64cm} \begin{proof} We define $\Gamma_1 := \hat{f}(x_{t}^{1},x_t^{2}) - \hat{f}(x_{t + 1}^{1},x_t^{2}), \Gamma_2 := \hat{f}(x_{t + 1}^{1},x_t^{2}) - \hat{f}(x_{t + 1}^{1},x_{t + 1}^{2})$ to denote the block update. It turns out that for $i=1,2$, and $\Gamma_i$ can be bounded following the Lipchitz smoothness as \begin{align} \label{eq:lipthmq0} \Gamma_i \leq & \langle g_t^i, x_t^i - x_{t+1}^i \rangle + \dfrac{L}{2}\|x_{t}^i - x_{t + 1}^i \|_2^2 \nonumber \\ = & -\beta_t^i \langle g_t^i, \Tilde{g}_t^i \rangle + \dfrac{L}{2}(\beta_t^i)^2\|\Tilde{g}_t^i\|_2^2 \nonumber\\ = & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 \nonumber \\ & + \dfrac{L}{2}(\beta_t^i)^2)\|\Delta_t^i\|_2^2 - (\beta_t^i - L(\beta_t^i)^2)\langle g_t^i,\Delta_t^i \rangle \nonumber \\ = & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 \nonumber \\ & - (\beta_t^i - L(\beta_t^i)^2) (\langle g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \nonumber \\ & + \langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle) \end{align} where the equalities follow the definition of $\Delta_t^i$ and the update law of Algorithm~\ref{alg:cyclicbcd}. We also have the following argument \begin{align} \label{eq:lipargu} & - (\beta_t^i - L(\beta_t^i)^2)\langle g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \nonumber\\ \leq & |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \|g_t^i - \nabla_{x^i}\hat{f}_\lambda (x_t)\|_2 \nonumber\\ \leq & L |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \|x_{t + 1} - x_{t}\|_2 \nonumber\\ \leq & L |\beta_t^i - L(\beta_t^i)^2| \|\Delta_t^i\|_2 \sqrt{\sum_{j = 1}^2 \|\beta_t^j \tilde{g}_t^j\|_2^2} \nonumber\\ \leq & L (\beta_t^i + L(\beta_t^i)^2) \beta_t^{\max} \left(\|\Delta_t^i\|_2 + \sum_{j = 1}^2 (\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2) \right), \end{align} where the first inequality follows from Cauchy-Schwarz inequality, the second inequality follows from the Lipchitz smoothness of objective function $\hat{f}_\lambda$, the third inequality follows from the update law of Algorithm~\ref{alg:cyclicbcd}, and the last inequality follows from the triangle inequality. Combining Eq.\eqref{eq:lipthmq0} and Eq.\eqref{eq:lipargu}, we obtain \begin{align} \label{eq:lipthmq} \Gamma_i \leq & -(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 + \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 \nonumber \\ & - (\beta_t^i - L(\beta_t^i)^2)\langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \nonumber \\ & + L(\beta_t^i + L(\beta_t^i)^2) \nonumber \\ & \cdot \beta_t^{\max} \left(\|\Delta_t^i\|_2^2 + \sum_{j = 1}^2(\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2)\right). \end{align} Summing Eq.~\eqref{eq:lipthmq} over $i$, then we obtain \begin{align} \label{eq:summing} &\hat{f}_\lambda(x_t) - \hat{f}_\lambda(x_{t + 1}) \nonumber \\ \leq & -\sum_{i = 1}^2(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\|g_t^i\|_2^2 \nonumber\\ & - \sum_{i = 1}^2(\beta_t^i - L(\beta_t^i)^2)\langle \nabla_{x^i}\hat{f}_\lambda (x_t), \Delta_t^i \rangle \nonumber\\ &+ \sum_{i = 1}^2 \Bigg( \dfrac{L}{2}(\beta_t^i)^2\|\Delta_t^i\|_2^2 + L(\beta_t^i + L(\beta_t^i)^2)\beta_t^{\max}(\|\Delta_t^i\|_2^2 \nonumber \\ &+ \sum_{j = 1}^2(\|g_t^j\|_2^2 + \|\Delta_t^j\|_2^2)) \Bigg). \end{align} We also have the following fact, \begin{align} \mathbb{E}[\langle \nabla_{y}\hat{f}_\lambda (x_t), \Delta_t^y \rangle] = & 0\\ \mathbb{E}[\langle \nabla_{\theta}\hat{f}_\lambda (x_t), \Delta_t^\theta \rangle] \leq & \beta^{\max}_t A M_\rho. \end{align} We provide detailed analysis of this fact in Lemma \ref{lem:prod} in Appendix, and the general analysis can be found in Lemma 1 in \citep{cd:bsg:xu2015}. Taking expectation w.r.t. $t$ on both sides of the inequality Eq.~\eqref{eq:summing}, we have \begin{align} \label{eq:exp} & \mathbb{E}[\hat{f}_\lambda(x_t)] - \mathbb{E}[\hat{f}_\lambda(x_{t + 1})] \nonumber\\ \leq & -\sum_{i = 1}^2(\beta_t^i - \dfrac{L}{2}(\beta_t^i)^2)\mathbb{E}[\|g_t^i\|_2^2] + (\beta_t^\theta - L(\beta_t^\theta)^2)\beta^{\max}_t A M_\rho \nonumber \\ & + \sum_{i = 1}^2\Bigg( \dfrac{L}{2}(\beta_t^i)^2\mathbb{E}[\|\Delta_t^i\|_2^2] \nonumber\\ & + L(\beta_t^i + L(\beta_t^i)^2)\beta_t^{\max} (\mathbb{E}[\|\Delta_t^i\|_2^2] \nonumber\\ & + \sum_{j = 1}^2(\mathbb{E}[\|g_t^j\|_2^2] + \mathbb{E}[\|\Delta_t^j\|_2^2])) \Bigg) \nonumber\\ \leq & -\sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] + (\beta^{\max}_t)^2 A M_\rho \nonumber\\ & + ( L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 M_\rho^2)), \end{align} where the first inequality follows from Eq.~\eqref{eq:expbound1} and Eq.~\eqref{eq:expbound2}, and the second inequality follows from the boundedness of $\mathbb{E}[\|\Delta_t^i\|_2]$ and $\mathbb{E}[\|g_t^i\|_2]$. Rearranging Eq.~\eqref{eq:exp}, we obtain \begin{align} \label{eq:rearr} &\sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] \nonumber\\ \leq& \mathbb{E}[\hat{f}_\lambda(x_{t + 1})] - \mathbb{E}[\hat{f}_\lambda(x_t)] + (\beta_t^{\max})^2 A M_\rho \nonumber \\ & + L(\beta_t^{\max})^2\sigma^2 \nonumber\\ & + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 M_\rho^2). \end{align} Summing Eq.~\eqref{eq:rearr} over $t$, we have \begin{align} \label{eq:finalsum} &\sum_{t = 1}^N \sum_{i = 1}^2(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^i\|_2^2] \nonumber\\ \leq& \hat{f}_\lambda^* - \hat{f}_\lambda(x_1) +\sum_{t = 1}^N [ (\beta_t^{\max})^2 A M_\rho + L(\beta_t^{\max})^2\sigma^2 \nonumber\\ & + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 M_\rho^2)]. \end{align} Then, rearranging the terms, using the similar technique with \citet{cd:sbmd:dang2015} (detailed in Section 4), we obtain Eq.~\eqref{eq:thm2} and \eqref{eq:C}. This completes the proof. \end{proof} \begin{remark} In Theorem \ref{thm:cyclicbcd2}, we have proven the finite-sample analysis of Algorithm~\ref{alg:cyclicbcd} with Option II, i.e., constant stepsizes and randomly picked solution. Note that the error bound in Eq.~\eqref{eq:thm2} can be simplified as $O(1 / (N \beta_t^{\min})) + O(\beta_t^{\max})$. Especially, if $\beta_t^{\max}=\beta_t^{\min}=\beta_t^\theta = \beta_t^y$ are set to be $\Theta(1 / \sqrt{N})$, then the convergence rate of Option II in Algorithm~\ref{alg:cyclicbcd} is $O(1/\sqrt{N})$. \end{remark} \section{Experimental Study} \label{sec:experimental} In this section, the proposed algorithms are tested on two risk management domains, i.e., the portfolio management domain~\citep{saferl:castro2012} and the American-style option domain~\citep{tamar2014scaling}. The baseline algorithms are the two-time-scale stochastic approximation based policy gradient method (\textbf{PGSA})~\citep{saferl:castro2012} and the randomized coordinate descent policy gradient (\textbf{RCPG}) algorithm described in the Appendix. \subsection{Portfolio Management} Consider a portfolio domain in \cite{saferl:castro2012} which is composed of two types of assets, i.e., the liquid asset and the non-liquid asset. A liquid asset has a fixed interest rate $r_l$ and can be sold at any time step $k = 1,\cdots, {\tau_t}$. A non-liquid asset, which can be sold only after a fixed period of $W$ time steps, has a time-dependent interest rate $r_{nl}(k)$. At the $k$-th time step, the interest rate of the non-liquid asset $r_{nl}(k)$ can take either $r_{nl}^{low}$ or $r_{nl}^{high}$, and the transitions between these two values are determined by a switching probability $p_{switch}$. Additionally, the non-liquid asset suffers a default risk (i.e. not being paid) with a probability $p_{risk}$. At the $k$-th step, the state is represented by $x(k)\in \mathbb{R}^{W+2}$, where $x_1 \in [0,1]$ is the portion of the investment in liquid assets, $x_2,\cdots,x_{W+1} \in [0,1]$ is the portion in non-liquid assets with time to maturity of $1,\cdots, W$ time steps, respectively, and $x_{W+2}(k) = r_{nl}(k)-\mathbb{E}[r_{nl}(k)]$, as Figure~\ref{fig:dynamic} shows. The investor can choose to invest a fixed portion $\eta$ ($0<\eta<1$) of his total available cash to the non-liquid asset if the investor has at least $\eta$ portion in the liquid asset, otherwise he has to wait for the maturity of the non-liquid asset. The reward at the $k$-th step is defined as the profit from the investment, and we show the logarithm of the total cumulative reward for better visualization. All investments are in liquid assets at the initial time step $k = 1$, and the startup cash is $\$100,000$. \begin{figure}[htb] \centering \includegraphics[width= .45\textwidth, height=.8in]{figure/Dynamic_of_the_investment}\\ \caption{{\small Dynamics of the investment~\citep{saferl:castro2012}}} \label{fig:dynamic} \end{figure} Figure~\ref{fig:portfolio} demonstrates the comparison results of the three algorithms, comparing the final iteration's solution property. The top and the middle figures are the distribution and the histogram distribution of cumulative rewards (in logarithm form), respectively. The top figure shows the distribution of cumulative rewards (in logarithm form) along with their CVaR values,which is another widely used risk control measure~\citep{saferl:cvar:chow2014}, w.r.t. risk level $\alpha =0.05$. The middle figure gives the histogram of cumulative rewards. The bottom figure shows the CVaR value with different risk level $\alpha$. The parameter settings are shown in Table~\ref{tab:portfolio}. The results indicate that MPPG tends to yield a higher cumulative reward, with less variance compared to PGSA and RCPG. \begin{figure}[htb!] \centering \begin{minipage}{1\textwidth} \includegraphics[width= .5\textwidth, height=1.48in]{figure/Tamardist208_3.png}\\ \includegraphics[width= .5\textwidth, height=1.48in]{figure/Tamarhist208_3}\\ \includegraphics[width= .492\textwidth, height=1.48in]{figure/Tamarcvar209_4} \end{minipage} \caption{Portfolio domain results} \label{fig:portfolio} \end{figure} \subsection{American-style Option} We test the algorithms on the American-style Option (put and call) domain introduced in \cite{tamar2014scaling}. The American-style option is a contract that gives the buyer the right, but not the obligation, to buy or sell an underlying asset at a specified strike price $W$ at or before a specific maturity time ${\tau_t}$. The initial price of the option is represented as $x_{0}$, and the price at $k$-th time step is exhibited by $x_{k}$. We consider the scenario that a buyer has bought a put option with strike price $W_{put}<x_{0}$, and a call option with strike price $W_{call}>x_{0}$. The price fluctuation follows a Bernoulli distribution, $x_{k+1} = \begin{cases} f_u x_k, \mathrm{with\ prob.}\ p,\\ f_d x_k, \mathrm{with\ prob.}\ 1-p, \end{cases}$where the up and down factors, $f_u$ and $f_d$, are constants. At the $k$-th time step, the state is represented by $\{x_{k},k\}$, where $x_k$ is the current price of the option. The action is binary, where $a_{k} = 1$ stands for executing the option, and $a_{k}=0$ stands for holding it at the $k$-th time step. The reward is $0$ unless an option has been executed, and the reward for executing an option is $g(x_k) = g_{put}(x_k)+ g_{call} (x_{k})$, where $g_{put} (x_k) = \max (0, W_{put}-x_k)$ and $g_{call} (x_k) = \max (0, x_k-W_{call}$). Once an option is executed, or $k = {\tau_t}$, a transition to terminal state occurs and the reward is given. Figure~\ref{fig:american} shows the comparison results of the three algorithms. The parameter settings are presented in Table~\ref{tab:option}. The top and the middle figures are the distribution ($\alpha = 0.05$) and the histogram distribution of cumulative rewards, respectively. The bottom figure shows the CVaR value with different risk level $\alpha$. The results suggest that MPPG can yield a higher cumulative reward along with less variance compared to PGSA and RCPG. The cumulative reward distribution of MPPG tends to have lower right-tail than PGSA and RCPG. In addition, MPPG appears to have a better chance to avoid getting zero reward, which suggests that MPPG performs a better risk-sensitive property. \begin{figure}[htb!] \centering \begin{minipage}{1\textwidth} \includegraphics[width= .46\textwidth, height=1.44in]{figure/Americanoption208_3}\\ \includegraphics[width= .47\textwidth, height=1.44in]{figure/Americanoptionhist208_3}\\ \includegraphics[width= .47\textwidth, height=1.44in]{figure/AmericanoptionCVaR207} \end{minipage} \caption{American option domain results} \label{fig:american} \end{figure} \begin{table}[htb] \centering \begin{tabular}{|p{0.6in}|p{0.25in}|p{0.25in}|p{0.25in}|p{0.3in}|p{0.32in}|p{0.15in}|} \hline Parameter&$r_{l}$&$r_{nl}^{high}$&$r_{nl}^{low}$&$p_{risk}$&$p_{switch}$&$W$\\ \hline Value&0.007&0.02&0.011&0.0005&0.1&4\\ \hline \end{tabular} \caption{Parameter settings in portfolio domain} \label{tab:portfolio} \end{table} \begin{table}[htb!] \vspace{0.1in} \begin{centering} \begin{tabular}{|p{0.6in}|p{0.2in}|p{0.22in}|p{0.22in}|p{0.22in}|p{0.22in}|p{0.17in}|p{0.13in}|} \hline Parameter&$K_{put}$&$K_{call}$&$x_{0}$&$f_{u}$&$f_{d}$&$p$&${\tau_t}$\\ \hline Value&$1$&$1.5$&$1.25$&$9/8$&$8/9$&$0.45$&$20$\\ \hline \end{tabular} \vspace{0.1in} \caption{Parameter settings in American-style option} \label{tab:option} \end{centering} \end{table} \section{Conclusion} \label{sec:conclusion} \vspace{-3mm} This paper is motivated to provide a risk-sensitive policy search algorithm with provable sample complexity analysis to maximize the mean-variance objective function. To this end, the objective function is reformulated based on the Fenchel duality, and a novel stochastic block coordinate ascent algorithm is proposed with in-depth analysis. There are many interesting future directions on this research topic. Besides stochastic policy gradient, deterministic policy gradient \citep{dpg:silver2014} has shown great potential in large discrete action space. It is interesting to design a risk-sensitive deterministic policy gradient method. Secondly, other reformulations of the mean-variance objective function are also worth exploring, which will lead to new families of algorithms. Thirdly, distributional reinforcement learning~\citep{distributional2016} is strongly related to risk-sensitive policy search. Combining risk-sensitive policy gradient methods with distributional reinforcement learning is also interesting to investigate. \clearpage \bibliographystyle{icml2018} \section{Introduction} \label{sec:intro} Risk management plays a central role in sequential decision-making problems, common in fields such as portfolio management~\citep{lai2011mean}, autonomous driving~\citep{risk:autonomous:maurer2016}, and healthcare~\citep{risk:american2011}. A common risk-measure is the variance of the expected sum of rewards/costs and the mean-variance trade-off function~\citep{sobel1982variance,saferl:mannor2011} is one of the most widely used objective functions in risk-sensitive decision-making. Other risk-sensitive objectives have also been studied, for example,~\citet{saferl:borkar2002q} studied exponential utility functions,~\citet{saferl:castro2012} experimented with the Sharpe Ratio measurement,~\citet{Chow18RC} studied value at risk (VaR) and mean-VaR optimization,~\citet{saferl:cvar:chow2014},~\citet{Tamar15OC}, and~\citet{Chow18RC} investigated conditional value at risk (CVaR) and mean-CVaR optimization in a static setting, and~\citet{saferl:tamar2015coherent} investigated coherent risk for both linear and nonlinear system dynamics. Compared with other widely used performance measurements, such as the Sharpe Ratio and CVaR, the mean-variance measurement has explicit interpretability and computational advantages~\citep{meanvar:markowitz2000mean,meanvar:li2000optimal}. For example, the Sharpe Ratio tends to lead to solutions with less mean return~\citep{saferl:castro2012}. Existing mean-variance reinforcement learning (RL) algorithms~\citep{saferl:castro2012,saferl:prashanth2013,Prashanth16VC} often suffer from heavy computational cost, slow convergence, and difficulties in tuning their learning rate schedules. Moreover, all their analyses are asymptotic and no rigorous finite-sample complexity analysis has been reported. {Recently, \citet{pmlr-v75-dalal18a} provided a general approach to compute finite sample analysis in the case of linear multiple time scales stochastic approximation problems. However, existing multiple time scales algorithms like \citep{saferl:castro2012} consist of nonlinear term in its update, and cannot be analyzed via the method in \citet{pmlr-v75-dalal18a}.} All these make it difficult to use them in real-world problems. The goal of this paper is to propose a mean-variance optimization algorithm that is both computationally efficient and has finite-sample analysis guarantees. This paper makes the following contributions: {\bf 1)} We develop a computationally efficient RL algorithm for mean-variance optimization. By reformulating the mean-variance function with its Legendre-Fenchel dual~\citep{boyd}, we propose a new formulation for mean-variance optimization and use it to derive a computationally efficient algorithm that is based on stochastic cyclic block coordinate descent. {\bf 2)} We provide the sample complexity analysis of our proposed algorithm. This result is novel because although cyclic block coordinate descent algorithms usually have empirically better performance than randomized block coordinate descent algorithms, yet almost all the reported analysis of these algorithms are asymptotic~\citep{cd:bsg:xu2015}. Here is a roadmap for the rest of the paper. Section~\ref{sec:backgrounds} offers a brief background on risk-sensitive RL and stochastic variance reduction. In Section~\ref{sec:alg}, the problem is reformulated using the Legendre-Fenchel duality and a novel algorithm is proposed based on stochastic block coordinate descent. Section~\ref{sec:theory} contains the theoretical analysis of the paper that includes both asymptotic convergence and finite-sample error bound. The experimental results of Section~\ref{sec:experimental} validate the effectiveness of the proposed algorithms. \section{Backgrounds} \label{sec:backgrounds} This section offers a brief overview of risk-sensitive RL, including the objective functions and algorithms. We then introduce block coordinate descent methods. Finally, we introduce the Legendre-Fenchel duality, the key ingredient in formulating our new algorithms. \subsection{Risk-Sensitive Reinforcement Learning} \label{sec:riskrl} Reinforcement Learning (RL)~\citep{sutton-barto:book} is a class of learning problems in which an agent interacts with an unfamiliar, dynamic, and stochastic environment, where the agent's goal is to optimize some measures of its long-term performance. This interaction is conventionally modeled as a Markov decision process (MDP), defined as the tuple $({\mathcal{S},\mathcal{A},P_0,P_{ss'}^{a},r,\gamma})$, where $\mathcal{S}$ and $\mathcal{A}$ are the sets of states and actions, $P_0$ is the initial state distribution, $P_{ss'}^{a}$ is the transition kernel that specifies the probability of transition from state $s\in\mathcal{S}$ to state $s'\in\mathcal{S}$ by taking action $a\in\mathcal{A}$, $r(s,a):\mathcal{S}\times\mathcal{A}\to\mathbb{R}$ is the reward function bounded by $R_{\max}$, and $0\leq\gamma<1$ is a discount factor. A \textit{parameterized stochastic policy} $\pi_\theta(a|s):\mathcal{S}\times\mathcal{A}\to\left[{0,1}\right]$ is a probabilistic mapping from states to actions, where $\theta$ is the tunable parameter and $\pi_\theta(a|s)$ is a differentiable function w.r.t.~$\theta$. One commonly used performance measure for policies in {\em episodic} MDPs is the {\em return} or cumulative sum of rewards from the starting state, i.e.,~$R = \sum\nolimits_{k = 1}^{{\tau}} {r({s_k},{a_k})}$, where $s_1\sim P_0$ and $\tau$ is the first passage time to the recurrent state $s^*$~\citep{puterman,saferl:castro2012}, and thus, $\tau \coloneqq \min \{ k > 0\;|\;{s_k} = {s^*}\}$. In risk-neutral MDPs, the algorithms aim at finding a near-optimal policy that maximizes the expected sum of rewards $J(\theta) \coloneqq \mathbb{E}_{\pi_\theta}[R] = \mathbb{E}_{\pi_\theta}\big[\sum_{k = 1}^{{\tau}}{r({s_k},{a_k})}\big]$. We also define the square-return $M(\theta ) \coloneqq {\mathbb{E}_{{\pi _\theta }}}[{R^2}] = {\mathbb{E}_{{\pi _\theta }}}\Big[{\big(\sum_{k = 1}^{\tau} {r({s_k},{a_k})}\big)^2}\Big]$. In the following, we sometimes drop the subscript $\pi_\theta$ to simplify the notation. In risk-sensitive mean-variance optimization MDPs, the objective is often to maximize $J(\theta)$ with a variance constraint, i.e., \begin{equation} \begin{aligned} \max_{\theta} \quad & J(\theta) = \mathbb{E}_{\pi_\theta}[R] \\ \text{s.t.} \quad & \text{Var}_{\pi_\theta}(R) \leq \zeta, \label{eq:opt-prob1} \end{aligned} \end{equation} where $\text{Var}_{\pi_\theta}(R) = M(\theta) - J^2(\theta)$ measures the variance of the return random variable $R$, and $\zeta > 0$ is a given risk parameter~\citep{saferl:castro2012,saferl:prashanth2013}. Using the Lagrangian relaxation procedure~\citep{bertsekas:npbook}, we can transform the optimization problem~\eqref{eq:opt-prob1} to maximizing the following unconstrained objective function \begin{align} {J_\lambda }(\theta ) \coloneqq & \mathbb{E}_{\pi_\theta}[R] - \lambda\big(\text{Var}_{\pi_\theta}(R) - \zeta) \\ = & J(\theta ) - \lambda\big(M(\theta ) - J(\theta)^2 - \zeta\big). \label{eq:regobj} \end{align} It is important to note that the mean-variance objective function is \textit{NP-hard} in general~\citep{saferl:mannor2011}. The main reason for the hardness of this optimization problem is that although the variance satisfies a Bellman equation~\citep{sobel1982variance}, unfortunately, it lacks the monotonicity property of dynamic programming (DP), and thus, it is not clear how the related risk measures can be optimized by standard DP algorithms~\citep{sobel1982variance}. The existing methods to maximize the objective function~\eqref{eq:regobj} are mostly based on stochastic approximation that often converge to an equilibrium point of an ordinary differential equation (ODE)~\citep{borkar:book}. For example,~\citet{saferl:castro2012} proposed a policy gradient algorithm, a two-time-scale stochastic approximation, to maximize~\eqref{eq:regobj} for a fixed value of $\lambda$ (they optimize over $\lambda$ by selecting its best value in a finite set), while the algorithm in~\citet{saferl:prashanth2013} to maximize~\eqref{eq:regobj} is actor-critic and is a three-time-scale stochastic approximation algorithm (the third time-scale optimizes over $\lambda$). These approaches suffer from certain drawbacks: {\bf 1)} Most of the analyses of ODE-based methods are asymptotic, with no sample complexity analysis. {\bf 2)} It is well-known that multi-time-scale approaches are sensitive to the choice of the stepsize schedules, which is a non-trivial burden in real-world problems. {\bf 3)} The ODE approach does not allow extra penalty functions. Adding penalty functions can often strengthen the robustness of the algorithm, encourages sparsity and incorporates prior knowledge into the problem~\citep{esl:ElementsStatisticalLearning}. \subsection{Coordinate Descent Optimization} \label{sec:bcd} Coordinate descent (CD)\footnote{Note that since our problem is maximization, our proposed algorithms are block coordinate \emph{ascent}.} and the more general block coordinate descent (BCD) algorithms solve a minimization problem by iteratively updating variables along coordinate directions or coordinate hyperplanes~\citep{cd:overview:wright2015}. At each iteration of BCD, the objective function is (approximately) minimized w.r.t.~a coordinate or a block of coordinates by fixing the remaining ones, and thus, an easier lower-dimensional subproblem needs to be solved. A number of comprehensive studies on BCD have already been carried out, such as~\citet{cd:bcd:luo1992} and~\citet{cd:sbcd:nesterov2012} for convex problems, and~\citet{cd:bcd:tseng2001},~\citet{xu2013block}, and~\citet{razaviyayn2013unified} for nonconvex cases (also see~\citealt{cd:overview:wright2015} for a review paper). For stochastic problems with a block structure,~\citet{cd:sbmd:dang2015} proposed stochastic block mirror descent (SBMD) by combining BCD with stochastic mirror descent~\citep{MID:2003,nemirovski2009robust}. Another line of research on this topic is block stochastic gradient coordinate descent (BSG)~\citep{cd:bsg:xu2015}. The key difference between SBMD and BSG is that at each iteration, SBMD randomly picks one block of variables to update, while BSG cyclically updates all block variables. In this paper, we develop mean-variance optimization algorithms based on both nonconvex stochastic BSG and SBMD. Since it has been shown that the BSG-based methods usually have better empirical performance than their SBMD counterparts, the main algorithm we report, analyze, and evaluate in the paper is BSG-based. We report our SBMD-based algorithm in Appendix~\ref{sec:rcpg-sga} and use it as a baseline in the experiments of Section~\ref{sec:experimental}. The finite-sample analysis of our BSG-based algorithm reported in Section~\ref{sec:theory} is novel because although there exists such analysis for convex stochastic BSG methods~\citep{cd:bsg:xu2015}, we are not aware of similar results for their nonconvex version to the best our knowledge. \section{Algorithm Design} \label{sec:alg} In this section, we first discuss the difficulties of using the regular stochastic gradient ascent to maximize the mean-variance objective function. We then propose a new formulation of the mean-variance objective function that is based on its Legendre-Fenchel dual and derive novel algorithms that are based on the recent results in stochastic nonconvex block coordinate descent. We conclude this section with an asymptotic analysis of a version of our proposed algorithm. \subsection{Problem Formulation} \label{subsec:formulation} In this section, we describe why the vanilla stochastic gradient cannot be used to maximize $J_\lambda(\theta)$ defined in Eq.~\eqref{eq:regobj}. Taking the gradient of $J_\lambda(\theta)$ w.r.t.~$\theta$, we have \begin{align} \nabla_\theta {J_\lambda }({\theta _t}) = & {\nabla _\theta }J({\theta _t}) - \lambda {\nabla _\theta }{\rm{Var}}(R) \\ = & {\nabla _\theta }J({\theta _t}) - \lambda\big(\nabla_{\theta} M(\theta) - 2J({\theta})\nabla_{\theta} J({\theta})\big). \label{objgradient} \end{align} Computing $\nabla_\theta {J_\lambda }({\theta _t})$ in~\eqref{objgradient} involves computing three quantities: $\nabla_{\theta} J({\theta}), \nabla_{\theta} M({\theta})$, and $J({\theta})\nabla_{\theta} J({\theta})$. We can obtain unbiased estimates of $\nabla_{\theta}J(\theta)$ and $\nabla_{\theta}M(\theta)$ from a single trajectory generated by the policy $\pi_\theta$ using the likelihood ratio method~\citep{reinforce:williams1992}, as $\nabla_{\theta}J({\theta})=\mathbb{E}[{R_t}\omega_{t}(\theta)]$ and $\nabla_{\theta}M({\theta})=\mathbb{E}[R_t^2\omega_{t}(\theta)]$. Note that $R_t$ is the cumulative reward of the $t$-th episode, i.e.,~$R_t = \sum\nolimits_{k=1}^{{\tau_t}} r_k$, which is possibly a \textit{nonconvex} function, and $\omega _t(\theta ) = \sum _{k=1}^{{\tau_t}}{\nabla_{\theta}\ln{\pi_{\theta}}({a_{k}}|{s_{k}})}$ is the likelihood ratio derivative. In the setting considered in the paper, an episode is the trajectory between two visits to the recurrent state $s^*$. For example, the $t$-th episode refers to the trajectory between the ($t$-1)-th and the $t$-th visits to $s^*$. We denote by $\tau_t$ the length of this episode. However, it is not possible to compute an unbiased estimate of $J(\theta)\nabla_{\theta} J(\theta)$ without having access to a generative model of the environment that allows us to sample at least two next states $s'$ for each state-action pair $(s, a)$. As also noted by~\citet{saferl:castro2012} and~\citet{saferl:prashanth2013}, computing an unbiased estimate of $J(\theta)\nabla_{\theta} J(\theta)$ requires double sampling (sampling from two different trajectories), and thus, cannot be done using a single trajectory. To circumvent the double-sampling problem, these papers proposed multi-time-scale stochastic approximation algorithms, the former a policy gradient algorithm and the latter an actor-critic algorithm that uses simultaneous perturbation methods~\citep{Bhatnagar13SR}. However, as discussed in Section~\ref{sec:riskrl}, multi-time-scale stochastic approximation approach suffers from several weaknesses such as no available finite-sample analysis and difficult-to-tune stepsize schedules. To overcome these weaknesses, we reformulate the mean-variance objective function and use it to present novel algorithms with in-depth analysis in the rest of the paper. \subsection{Block Coordinate Reformulation} In this section, we present a new formulation for $J_\lambda(\theta)$ that is later used to derive our algorithms and do not suffer from the double-sampling problem in estimating $J(\theta)\nabla_{\theta} J(\theta)$. We begin with the following lemma. \begin{lemma} \label{lem:fenchel} For the quadratic function $f(z)=z^2,\;z\in\mathbb{R}$, we define its Legendre-Fenchel dual as $f(z)=z^2=\mathop{\max}_{y\in \mathbb{R}} (2zy - {y^2})$. \end{lemma} This is a special case of the Lengendre-Fenchel duality~\citep{boyd} that has been used in several recent RL papers (e.g.,~\citealt{liu2015uai,vr:saddle:pe:du2017,liubo:jair:2018}). Let ${F_\lambda }(\theta ) \coloneqq {\left({J(\theta ) + \frac{1}{{2\lambda }}} \right)^2} - M(\theta )$, which follows ${F_\lambda }(\theta )=\frac{{{J_\lambda }(\theta)}}{\lambda } + \frac{1}{{4{\lambda ^2}}} - \zeta$. Since $\lambda>0$ is a constant, maximizing $J_\lambda(\theta)$ is equivalent to maximizing ${F_\lambda }(\theta )$. Using Lemma~\ref{lem:fenchel} with $z = J(\theta ) + \frac{1}{2\lambda}$, we may reformulate ${F_\lambda }(\theta )$ as \begin{align} \label{eq:000} {F_\lambda }(\theta) = \mathop {\max }_y \Big( {2y\big(J(\theta) + \frac{1}{{2\lambda }}\big) - {y^2}} \Big) - M(\theta). \end{align} Using~\eqref{eq:000}, the maximization problem $\mathop {\max }_{\theta} {F_\lambda }(\theta )$ is equivalent to \begin{equation} \begin{aligned} \mathop {\max}_{\theta ,y} \qquad & {\hat f_\lambda }(\theta ,y), \\ \text{where} \qquad & {{{\hat f}_\lambda }(\theta ,y) \coloneqq 2y\big(J(\theta ) + \frac{1}{{2\lambda }}\big) - {y^2} - M(\theta )}. \label{eq:max-alt} \end{aligned} \end{equation} Our optimization problem is now formulated as the standard nonconvex coordinate ascent problem~\eqref{eq:max-alt}. We use three stochastic solvers to solve~\eqref{eq:max-alt}: SBMD method~\citep{cd:sbmd:dang2015}, BSG method~\citep{cd:bsg:xu2015}, and the vanilla stochastic gradient ascent (SGA) method~\citep{nemirovski2009robust}. We report our BSG-based algorithm in Section~\ref{subsec:MVPG} and leave the details of the SBMD and SGA based algorithms to Appendix~\ref{sec:rcpg-sga}. In the following sections, we denote by $\beta_t^\theta$ and $\beta_t^y$ the stepsizes of $\theta$ and $y$, respectively, and by the subscripts $t$ and $k$ the episode and time-step numbers. \subsection{Mean-Variance Policy Gradient} \label{subsec:MVPG} We now present our main algorithm that is based on a block coordinate update to maximize~\eqref{eq:max-alt}. Let ${g}_t^\theta$ and ${g}_t^y$ be block gradients and $\tilde{g}_t^\theta$ and $\tilde{g}_t^y$ be their sample-based estimations defined as \begin{align} \label{def:tildeg_y} g^y_t = \mathbb{E}[\tilde{g}^y_t] = 2 J(\theta_t) + \frac{1}{\lambda } - 2y_t \quad , \quad & \tilde{g}^y_t = 2 R_t + \frac{1}{\lambda } - 2y_t,\\ \label{def:tildeg_theta} g^\theta_t = \mathbb{E}[\tilde g^\theta_t] = 2y_{t + 1}\nabla_{\theta}{J(\theta_t)} - \nabla_{\theta} M(\theta_t) \quad , \quad & \tilde{g}^\theta_t = \left( {2y_{t + 1}{R_t} - {{({R_t})}^2}} \right){\omega _t}({\theta _t}). \end{align} The block coordinate updates are \begin{align} \label{eq:updatelaw} {y_{t+1}} = & {y_{t}}+{\beta^y_{t}}{\tilde{g}_t^y}, \\ {\theta_{t+1}} = & {\theta_{t}}+{\beta^\theta_{t}}{\tilde{g}_t^\theta}. \end{align} To obtain unbiased estimates of $g^y_t$ and $g^\theta_t$, we shall update $y$ (to obtain $y_{t+1}$) prior to computing $g^\theta_t$ at each iteration. Now it is ready to introduce the Mean-Variance Policy Gradient (\textbf{MVP}) Algorithm~\ref{alg:mvp}. \begin{algorithm}[htb] \caption{Mean-Variance Policy Gradient (\textbf{MVP})} \label{alg:mvp} \centering \begin{algorithmic}[1] \STATE {\bfseries Input:} Stepsizes $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$, and number of iterations $N$ \\ $\quad$\textbf{Option I:} $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$ satisfy the Robbins-Monro condition \\ $\quad$\textbf{Option II:} $\beta_t^\theta$ and $\beta_t^y$ are set to be constants \FOR {episode $t=1,\dotsc,N$} \STATE Generate the initial state $s_1\sim P_0$ \WHILE{$s_k \neq s^*$} \STATE Take the action $a_k \sim \pi_{\theta_t}(a|s_k)$ and observe the reward $r_k$ and next state $s_{k+1}$ \ENDWHILE \STATE Update the parameters \begin{align} {R_t} = & \sum_{k = 1}^{\tau_t} {{r_k}} \\ {\omega _t}(\theta_t) = & \sum _{k=1}^{{\tau_t}}{\nabla_\theta \ln{\pi_{\theta_t}}({a_{k}}|{s_{k}})} \\ {y_{t+1}} = & {y_t} + {\beta _t^y}\left(2 R_t + \frac{1}{\lambda } - 2y_t\right)\\ {\theta_{t+1}} = & {\theta _t} + {\beta _t^\theta}\left( {2y_{t+1}{R_t } - {{({R_t })}^2}} \right){\omega _t}({\theta _t}) \end{align} \ENDFOR \STATE {\bfseries Output} $\bar{x}_N$: \\ $\quad$\textbf{Option I:} Set $\bar{x}_N = x_N = [\theta_N,y_N]^\top$ \\ $\quad$\textbf{Option II:} Set $\bar{x}_N = x_z= [\theta_z,y_z]^\top$, where $z$ is uniformly drawn from $\{ 1,2, \ldots ,N\}$ \end{algorithmic} \end{algorithm} Before presenting our theoretical analysis, we first introduce the assumptions needed for these results. \begin{assumption}[\textbf{Bounded Gradient and Variance}] \label{asm3} There exist constants $G$ and $\sigma$ such that \begin{align} \|\nabla_{y}\hat f_{\lambda}(x)\|_2 \leq G, ~ & \|\nabla_{\theta} \hat f_{\lambda} (x)\|_2 \leq G, \\ \mathbb{E}[\|\Delta^y_t\|_2^2] \leq {\sigma}^2, ~ & \mathbb{E}[\|\Delta^\theta_t\|_2^2] \leq {\sigma}^2, \end{align} for any $t$ and $x$, where $\|\cdot\|_2$ denotes the Euclidean norm, ${\Delta}^y_t \coloneqq \tilde{g}^y_t - {g}^y_t$ and ${\Delta}^\theta_t \coloneqq \Tilde{g}^\theta_t - {g}^\theta_t$. \label{asm:var} \end{assumption} Assumption~\ref{asm:var} is standard in nonconvex coordinate descent algorithms~\citep{cd:bsg:xu2015,cd:sbmd:dang2015}. We also need the following assumption that is standard in the policy gradient literature. \begin{assumption}[\textbf{Ergodicity}] The Markov chains induced by all the policies generated by the algorithm are ergodic, i.e.,~irreducible, aperiodic, and recurrent. \label{asm:erg} \end{assumption} In practice, we can choose either Option I with the result of the final iteration as output or Option II with the result of a randomly selected iteration as output. In what follows in this section, we report an asymptotic convergence analysis of MVP with Option I, and in Section~\ref{sec:theory}, we derive a finite-sample analysis of MVP with Option II. \begin{theorem}[\textbf{Asymptotic Convergence}] \label{thm:asyn} Let $\big\{x_t=(\theta_t, y_t)\big\}$ be the sequence of the outputs generated by Algorithm~\ref{alg:mvp} with Option I. If $\{\beta_t^\theta\}$ and $\{\beta_t^y\}$ are time-diminishing real positive sequences satisfying the Robbins-Monro condition, i.e.,~$\sum_{t = 1}^\infty {\beta _t^\theta } = \infty$, $\sum_{t = 1}^\infty {{{(\beta _t^\theta )}^2}} < \infty$, $\sum_{t = 1}^\infty {\beta _t^y} = \infty$, and $\sum_{t = 1}^\infty {{{(\beta _t^y)}^2}} < \infty$, then Algorithm~\ref{alg:mvp} will converge such that $\;\lim_{t\rightarrow \infty}\mathbb{E}[\|\nabla\hat{f}_{\lambda}(x_t) \|_2] = 0$. \end{theorem} The proof of Theorem~\ref{thm:asyn} follows from the analysis in~\citet{xu2013block}. Due to space constraint, we report it in Appendix~\ref{sec:theory_app}. Algorithm~\ref{alg:mvp} is a special case of nonconvex block stochastic gradient (BSG) methods. To the best of our knowledge, no finite-sample analysis has been reported for this class of algorithms. Motivated by the recent papers by~\citet{nemirovski2009robust},~\citet{ghadimi2013stochastic},~\citet{cd:bsg:xu2015}, and~\citet{cd:sbmd:dang2015}, in Section~\ref{sec:theory}, we provide a finite-sample analysis for general nonconvex block stochastic gradient methods and apply it to Algorithm~\ref{alg:mvp} with Option II. \section{Finite-Sample Analysis} \label{sec:theory} \vspace{-0.3cm} In this section, we first present a finite-sample analysis for the general class of nonconvex BSG algorithms~\citep{xu2013block}, for which there are no established results, in Section~\ref{sec:fs-bsg}. We then use these results and prove a finite-sample bound for our MVP algorithm with Option II, that belongs to this class, in Section~\ref{sec:fs-mvp}. Due to space constraint, we report the detailed proofs in Appendix~\ref{sec:theory_app}. \subsection{Finite-Sample Analysis of Nonconvex BSG Algorithms} \label{sec:fs-bsg} In this section, we provide a finite-sample analysis of the general nonconvex block stochastic gradient (BSG) method, where the problem formulation is given by \begin{align} \min_{x \in \mathbb{R}^{n}} ~ f(x) = \mathbb{E}_{\xi}[F(x,\xi)]. \label{eq:obj} \end{align} $\xi$ is a random vector, and $F(\cdot,\xi): \mathbb{R}^{n} \rightarrow \mathbb{R}$ is continuously differentiable and possibly nonconvex for every $\xi$. The variable $x \in \mathbb{R}^{n}$ can be partitioned into $b$ disjoint blocks as $x = \{x^1,x^2,\dotsc,x^b\}$, where $x^i \in \mathbb{R}^{n_i}$ denotes the $i$-th block of variables, and $\sum_{i = 1}^{b}n_i = n$. For simplicity, we use $x^{<i}$ for $(x_i,\dots,x_{i-1})$, and $x^{\leq i}$,$x^{>i}$, and $x^{\geq i}$ are defined correspondingly. We also use $\nabla_{x^i}$ to denote $\frac{\partial}{\partial x^i}$ for the partial gradient with respect to $x^i$. $\Xi_t$ is the sample set generated at $t$-th iteration, and $\mathbf \Xi_{[t]} = (\Xi_1,\dots,\Xi_t)$ denotes the history of sample sets from the first through $t$-th iteration. $\{\beta_t^i: i = 1,\cdots,b\}_{t = 1}^{\infty}$ are denoted as the stepsizes. Also, let $\beta_t^{\max} = \max_{i}\beta_t^i$, and $\beta_t^{\min} = \min_{i}\beta_t^i$. Similar to Algorithm \ref{alg:mvp}, the BSG algorithm cyclically updates all blocks of variables in each iteration, and the detailed algorithm for BSG method is presented in Appendix \ref{sec:probsg}. Without loss of generality, we assume a fixed update order in the BSG algorithm. Let $\Xi_t = \{\xi_{t,1},\dotsc,\xi_{t,m_t}\}$ be the samples in the $t$-th iteration with size $m_t \geq 1$. Therefore, the stochastic partial gradient is computed as $ \tilde g_t^i = \frac{1}{m_t} \sum_{l = 1}^{m_t}\nabla_{x^i}F(x_{t+1}^{<i}, x_t^{\geq i}; \xi_{t,l}). $ Similar to Section \ref{sec:alg}, we define $g_t^i = \nabla_{x^i}f(x_{t+1}^{<i}, x_t^{\geq i})$, and the approximation error as $\Delta_t^i = \tilde g_t^i - g_t^i$. We assume that the objective function $f$ is bounded and Lipschitz smooth, i.e., there exists a positive Lipschitz constant $L > 0$ such that $ \|\nabla_{x^i}f(x) - \nabla_{x^i}f(y)\|_2 \leq L \|x - y\|_2 $, $\forall i \in \{1,\dots,b\}$ and $\forall x,y \in \mathbb R^n$. Each block gradient of $f$ is also bounded, i.e., there exist a positive constant $G$ such that $ \|\nabla_{x^i}f(x)\|_2 \leq G $, for any $i \in \{1,\dots,b\}$ and any $x \in \mathbb R^n$. We also need Assumption \ref{asm3} for all block variables, i.e., $\mathbb{E}[\|\Delta_t^i\|_2] \leq \sigma$, for any $i$ and $t$. Then we have the following lemma. \begin{lemma} \label{lem:err_bound} For any $i$ and $t$, there exist a positive constant $A$, such that \begin{align} \|\mathbb{E}[\Delta_t^i | \mathbf \Xi_{[t - 1]}]\|_2 \leq A\beta_t^{\max}. \label{eq:stocerrbd} \end{align} \end{lemma} The proof of Lemma~\ref{lem:err_bound} is in Appendix~\ref{sec:probsg}. It should be noted that in practice, it is natural to take the final iteration's result as the output as in Algorithm~\ref{alg:mvp}. However, a standard strategy for analyzing nonconvex optimization methods is to pick up one previous iteration's result randomly according to a discrete probability distribution over $\{1,2,\dotsc, N\}$~\citep{nemirovski2009robust,ghadimi2013stochastic,cd:sbmd:dang2015}. Similarly, our finite-sample analysis is based on the strategy that randomly pick up $\bar x_{N} = x_{z}$ according to \begin{align} \Pr(z = t) = \frac{\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2}{\sum_{t = 1}^N (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2)}, \ t = 1,\dotsc, N. \label{eq:prob-z} \end{align} Now we provide the finite-sample analysis result for the general nonconvex BSG algorithm as in~\citep{cd:bsg:xu2015}. \begin{theorem} \label{thm:cyclicbcd} Let the output of the nonconvex BSG algorithm be $\bar{x}_N = x_z$ according to Eq.~\eqref{eq:prob-z}. If stepsizes satisfy $2\beta_t^{\min} > L(\beta_t^{\max})^2$ for $t = 1,\cdots,N$, then we have \begin{align} &\mathbb{E}\left[\|\nabla f(\bar{x}_N)\|_2^2\right] \leq \frac{ f(x_1) - f^* + \sum_{t = 1}^N (\beta_t^{\max})^2 C_t}{\sum_{t = 1}^N (\beta_t^{\min} - \frac{ L }{2}(\beta_t^{\max})^2)}, \label{eq:cyclicbcd} \end{align} where $f^* = \max_{x} f(x)$. $ C_t = (1 - \frac{ L }{2}\beta_t^{\max}) \sum_{i = 1}^{b} L \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + b\left( A G + \frac{L}{2} \sigma^2 \right), $ where $G$ is the gradient bound, $L$ is the Lipschitz constants, $\sigma$ is the variance bound, and $A$ is defined in Eq.~\eqref{eq:stocerrbd}. \end{theorem} As a special case, we discuss the convergence rate with constant stepsizes $\mathcal O(1/\sqrt{N})$ in Corollary~\ref{cor:cyclicrate}, which implies that the sample complexity $N = \mathcal O(1 / \varepsilon^2)$ in order to find $\varepsilon$-stationary solution of problem \eqref{eq:obj}. \begin{corollary} \label{cor:cyclicrate} If we take constant stepsize such that $\beta_t^i = \beta^i = \mathcal O(1/\sqrt{N})$ for any $t$, and let $\beta^{\max} \coloneqq \max_i \beta^i$, $\beta^{\min} \coloneqq \min_i \beta^i$, then we have $ \mathbb{E}\left[\|\nabla f(\bar{x}_N)\|_2^2\right] \leq \mathcal O \left( \sqrt{\frac{f(x_1) - f^* + C}{N}} \right), $ where $C_t$ in Eq.~\eqref{eq:cyclicbcd} reduces to a constant $C$ defined as $ C = (1 - \frac{ L }{2}\beta^{\max}) \sum_{i = 1}^{b} L \sqrt{\sum_{j < i} (G^2 + \sigma^2) } + b \left( A G + \frac{L}{2} \sigma^2 \right). $ \end{corollary} \subsection{Finite-Sample Analysis of Algorithm~\ref{alg:mvp}} \label{sec:fs-mvp} We present the major theoretical results of this paper, i.e., the finite-sample analysis of Algorithm~\ref{alg:mvp} with Option II. The proof of Theorem \ref{thm:cyclicbcd2} is in Appendix~\ref{sec:theory_app}. \begin{theorem} \label{thm:cyclicbcd2} Let the output of the Algorithm~\ref{alg:mvp} be $\bar{x}_N$ as in Theorem~\ref{thm:cyclicbcd}. If $\{\beta_t^{\theta}\}$, $\{\beta_t^{y}\}$ are constants as in Option II in Algorithm~\ref{alg:mvp}, and also satisfies $2\beta_t^{\min} > L(\beta_t^{\max})^2$ for $t = 1,\cdots,N$, we have \begin{align} &\mathbb{E}\left[\|\nabla \hat f_{\lambda}(\bar{x}_N)\|_2^2\right] \leq \dfrac{\hat{f}_\lambda^* - \hat{f}_\lambda(x_1) + N (\beta_t^{\max})^2 C}{N (\beta_t^{\min} - \frac{L}{2}(\beta_t^{\max})^2)} \label{eq:thm2} \end{align} where $\hat{f}_{\lambda}^* = \max_{x} \hat f_\lambda (x)$, and \begin{align} C = & (1 - \dfrac{L}{2}\beta_t^{\max})(L^2 \beta_{t}^{\max}(G^2 + \sigma^2) + L(2G^2 + \sigma^2)) + A G + L\sigma^2 + 2L(1 + L\beta_t^{\max})(3 \sigma^2 + 2 G^2). \label{eq:C} \end{align} \end{theorem} \vspace{-0.5cm} \begin{proof}[Proof Sketch] The proof follows the following major steps. \textbf{(\RomanNum{1})}. First, we need to prove the bound of each block coordinate gradient, i.e., $\mathbb{E}[\|g_t^\theta\|_2^2]$ and $\mathbb{E}[ \|g_t^y\|_2^2]$, which is bounded as \begin{align} &(\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq& \mathbb{E}[\hat{f}_\lambda(x_{t + 1})] - \mathbb{E}[\hat{f}_\lambda(x_t)] + (\beta_t^{\max})^2 A M_\rho + L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2). \end{align} Summing up over $t$, we have \begin{align} &\sum_{t = 1}^N (\beta_t^{\min} - \dfrac{L}{2}(\beta_t^{\max})^2)\mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2] \\ \leq& \hat{f}_\lambda^* - \hat{f}_\lambda(x_1) +\sum_{t = 1}^N [ (\beta_t^{\max})^2 A G + L(\beta_t^{\max})^2\sigma^2 + 2L\beta_t^{\max}(\beta_t^{\max} + L(\beta_t^{\max})^2)(3 \sigma^2 + 2 G^2)]. \end{align} \textbf{(\RomanNum{2})}. Next, we need to bound $\mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2]$ using $\mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2]$, which is proven to be \begin{align} & \mathbb{E}[\|\nabla\hat{f}_\lambda(x_t) \|_2^2] \leq L^2 (\beta_{t}^{\max})^2(G^2 + \sigma^2) + L\beta_{t}^{\max}(2G^2 + \sigma^2) + \mathbb{E}[\|g_t^\theta\|_2^2 + \|g_t^y\|_2^2]. \end{align} \textbf{(\RomanNum{3})}. Finally, combining (\RomanNum{1}) and (\RomanNum{2}), and rearranging the terms, Eq.~\eqref{eq:thm2} can be obtained as a special case of Theorem~\ref{thm:cyclicbcd}, which completes the proof. \end{proof} \section{Experimental Study} \label{sec:experimental} In this section, we evaluate our MVP algorithm with Option I in three risk-sensitive domains: the portfolio management~\citep{saferl:castro2012}, the American-style option~\citep{tamar2014scaling}, and the optimal stopping~\citep{saferl:cvar:chow2014,Chow18RC}. The baseline algorithms are the vanilla policy gradient (PG), the mean-variance policy gradient in~\citet{saferl:castro2012}, the stochastic gradient ascent (SGA) applied to our optimization problem~\eqref{eq:max-alt}, and the randomized coordinate ascent policy gradient (RCPG), i.e.,~the SBMD-based version of our algorithm. Details of SGA and RCPG can be found in Appendix~\ref{sec:rcpg-sga}. For each algorithm, we optimize its Lagrangian parameter $\lambda$ by grid search and report the mean and variance of its return random variable as a Gaussian.\footnote{Note that the return random variables are not necessarily Gaussian, we only use Gaussian for presentation purposes.} Since the algorithms presented in the paper (MVP and RCPG) are policy gradient, we only compare them with Monte-Carlo based policy gradient algorithms and do not use any actor-critic algorithms, such as those in~\citet{saferl:prashanth2013} and TRPO~\citep{trpo:schulman:2015}, in the experiments. \begin{figure}[t] \begin{subfigure}{0.32\linewidth} \includegraphics[width=\linewidth]{figure/PF_PDF.pdf} \caption{Portfolio management domain} \label{fig:PF} \end{subfigure} \begin{subfigure}{0.32\linewidth} \includegraphics[width=\linewidth]{figure/AO_PDF.pdf} \caption{American-style option domain} \label{fig:AO} \end{subfigure} \begin{subfigure}{0.32\linewidth} \includegraphics[width=\linewidth]{figure/OS_PDF.pdf} \caption{Optimal stopping domain} \label{fig:OS} \end{subfigure} \caption{Empirical results of the distributions of the return (cumulative rewards) random variable. Note that markers only indicate different methods.} \end{figure} \subsection{Portfolio Management} The portfolio domain~\cite{saferl:castro2012} is composed of the liquid and non-liquid assets. A liquid asset has a fixed interest rate $r_l$ and can be sold at any time-step $k \le {\tau}$. A non-liquid asset can be sold only after a fixed period of $W$ time-steps with a time-dependent interest rate $r_{\text{nl}}(k)$, which can take either $r_{\text{nl}}^{\text{low}}$ or $r_{\text{nl}}^{\text{high}}$, and the transition follows a switching probability $p_{\text{switch}}$. The non-liquid asset also suffers a default risk (i.e.,~not being paid) with a probability $p_{\text{risk}}$. All investments are in liquid assets at the initial time-step $k = 0$. At the $k$-th step, the state is denoted by $x(k)\in \mathbb{R}^{W+2}$, where $x_1 \in [0,1]$ is the portion of the investment in liquid assets, $x_2,\cdots,x_{W+1} \in [0,1]$ is the portion in non-liquid assets with time to maturity of $1,\cdots, W$ time-steps, respectively, and $x_{W+2}(k) = r_{\text{nl}}(k)-\mathbb{E}[r_{\text{nl}}(k)]$. The investor can choose to invest a fixed portion $\eta$ ($0<\eta<1$) of his total available cash in the non-liquid asset or do nothing. More details about this domain can be found in~\citet{saferl:castro2012}. Figure~\ref{fig:PF} shows the results of the algorithms. PG has a large variance and the Tamar's method has the lowest mean return. The results indicate that MVP yields a higher mean return with less variance compared to the competing algorithms. \subsection{American-style Option} An American-style option~\cite{tamar2014scaling} is a contract that gives the buyer the right to buy or sell the asset at a strike price $W$ at or before the maturity time ${\tau}$. The initial price of the option is $x_{0}$, and the buyer has bought a put option with the strike price $W_{\text{put}}<x_{0}$ and a call option with the strike price $W_{\text{call}}>x_{0}$. At the $k$-th step ($k \le \tau$), the state is $\{x_{k},k\}$, where $x_k$ is the current price of the option. The action $a_{k}$ is either executing the option or holding it. $x_{k+1}$ is $f_u x_k$ w.p.~$p$ and $f_d x_k$ w.p.~$1 - p$, where $f_u$ and $f_d$ are constants. The reward is $0$ unless an option is executed and the reward for executing an option is ${r_k} = \max (0,{W_{{\rm{put}}}} - {x_k}) + \max (0,{x_k} - {W_{{\rm{call}}}})$. More details about this domain can be found in~\citet{tamar2014scaling}. Figure~\ref{fig:AO} shows the performance of the algorithms. The results suggest that MVP can yield a higher mean return with less variance compared to the other algorithms. \subsection{Optimal Stopping} The optimal stopping problem~\citep{saferl:cvar:chow2014,Chow18RC} is a continuous state domain. At the $k$-th time-step ($k \le \tau$, $\tau$ is the stopping time), the state is $\{x_k, k\}$, where $x_k$ is the cost. The buyer decide either to accept the present cost or wait. If the buyer accepts or when $k = T$, the system reaches a terminal state and the cost $x_k$ is received, otherwise, the buyer receives the cost $p_h$ and the new state is $\{x_{k+1}, k+1\}$, where $x_{k+1}$ is $f_{u}x_{k}$ w.p.~$p$ and $f_{d}x_{k}$ w.p. $1-p$ ($f_{u} > 1$ and $f_{d}<1$ are constants). More details about this domain can be found in~\citet{saferl:cvar:chow2014}. Figure~\ref{fig:OS} shows the performance of the algorithms. The results indicate that MVP is able to yield much less variance without affecting its mean return. We also summarize the performance of these algorithms in all three risk-sensitive domains as Table \ref{tb:meanvartb}, where Std is short for Standard Deviation. \begin{table}[htb] \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline \multirow{2}{*}{\textbf{}} & \multicolumn{2}{c|}{\textbf{Portfolio Management}} & \multicolumn{2}{c|}{\textbf{American-style Option}} & \multicolumn{2}{c|}{\textbf{Optimal Stopping}} \\ \cline{2-7} & \textbf{Mean} & \textbf{Std} & \textbf{Mean} & \textbf{Std} & \textbf{Mean} & \textbf{Std} \\ \hline \textbf{MVP} & \textbf{29.754} & \textbf{0.325} & \textbf{0.2478} & \textbf{0.00482} & \textbf{-1.4767} & 0.00456 \\ \hline \textbf{PG} & 29.170 & 1.177 & 0.2477 & 0.00922 & -1.4769 & 0.00754 \\ \hline \textbf{Tamar} & 28.575 & 0.857 & 0.2240 & 0.00694 & -2.8553 & \textbf{0.00415} \\ \hline \textbf{SGA} & 29.679 & 0.658 & 0.2470 & 0.00679 & -1.4805 & 0.00583 \\ \hline \textbf{RCPG} & 29.340 & 0.789 & 0.2447 & 0.00819 & -1.4872 & 0.00721 \\ \hline \end{tabular} \vspace{\baselineskip} \caption{Performance Comparison among Algorithms} \label{tb:meanvartb} \end{table} \section{Conclusion} \label{sec:conclusion} \vspace{-3mm} This paper is motivated to provide a risk-sensitive policy search algorithm with provable sample complexity analysis to maximize the mean-variance objective function. To this end, the objective function is reformulated based on the Legendre-Fenchel duality, and a novel stochastic block coordinate ascent algorithm is proposed with in-depth analysis. There are many interesting future directions on this research topic. Besides stochastic policy gradient, deterministic policy gradient \citep{dpg:silver2014} has shown great potential in large discrete action space. It is interesting to design a risk-sensitive deterministic policy gradient method. Secondly, other reformulations of the mean-variance objective function are also worth exploring, which will lead to new families of algorithms. Thirdly, distributional RL~\citep{distributional2016} is strongly related to risk-sensitive policy search, and it is interesting to investigate the connections between risk-sensitive policy gradient methods and distributional RL. Last but not least, it is interesting to test the performance of the proposed algorithms together with other risk-sensitive RL algorithms on highly-complex risk-sensitive tasks, such as autonomous driving problems and other challenging tasks. \section*{Acknowledgments} Bo Liu, Daoming Lyu, and Daesub Yoon were partially supported by a grant (18TLRP-B131486-02) from Transportation and Logistics R\&D Program funded by Ministry of Land, Infrastructure and Transport of Korean government. Yangyang Xu was partially supported by the NSF grant DMS-1719549. \bibliographystyle{apalike}

df946185099cb8f9aa09312387fac53947fff4b2

\section{Introduction} In this article, a linear map from differential two-forms to symmetric two-tensors in two-dimensional Hermitian manifolds introduced in \cite{gi-u1-prl} is studied. The map reveals another aspect of Seiberg-Witten map. The original Seiberg-Witten map is a map from noncommutative gauge fields to commutative gauge fields with a background $B$-field \cite{sw-ncft}. On the other hand, it has been interpreted in \cite{gi-u1-prl,gi-u1-plb,gi-u1-epl} as a map from a noncommutative gauge field to a K\"ahler metric.\\ A purpose of this article is to clarify the map in \cite{gi-u1-prl,gi-u1-plb,gi-u1-epl} which locally maps (anti-)self-dual two-forms on ${\mathbb C}^2$ to Hermitian-Einstein metrics of two-dimensional K\"ahler manifolds. It might be worth noting that it is enough for these two-forms to be defined as a symplectic structure on a commutative manifold, although this map was developed in the context of Seiberg-Witten map in noncommutative gauge theory. But this correspondence between the self-dual two-form and Hermitian-Einstein metric can be lifted to noncommutative spaces after (canonical or deformation) quantization \cite{ly-jkps2018}.\\ The second purpose of this article is to construct explicit examples of Hermitian-Einstein metrics from noncommutative $U(1)$ instantons. $U(1)$ instantons on noncommutative ${\mathbb C}^2$ were found by Nekrasov and Schwarz \cite{Nekrasov:1998ss}. We will construct the two-form from a multi-instanton solution given in \cite{Ishikawa:2001ye} where the noncommutative $U(1)$ instanton solutions are written in an operator form acting on a Fock space. The Fock space is defined by Heisenberg algebra generated by noncommutative complex coordinates. There is a dictionary between the linear operators acting on the Fock space and usual functions \cite{Sako:2016gqb}. The dictionary is applicable for arbitrary noncommutative K\"ahler manifold obtained by deformation quantization with separation of variables \cite{Karabegov1996}. Concrete Hermitian-Einstein metrics are obtained by translating the noncommutative instantons as linear operators into ordinary functions by using the dictionary in \cite{Sako:2016gqb}.\\ The third purpose is to clarify the K\"ahler condition for the metrics derived from noncommutative $U(1)$ instantons. Since a K\"ahler manifold is a symplectic manifold too although the reverse is not necessarily true, one can quantize the K\"ahler manifold by quantizing a Poisson algebra derived from the underlying symplectic structure of the K\"ahler geometry, as recently clarified in \cite{ly-jkps2018}. We will show that the metric derived from noncommutative $U(1)$ instantons becomes a K\"ahler metric if the underlying Poisson algebra of $U(1)$ instantons or its quantization is an associative algebra. \\ Here we mention some studies related with subjects of this article. It has been conjectured in \cite{inov,ideal-sheaf} that NC $U(1)$ gauge theory is the fundamental description of K\"ahler gravity at all scales including the Planck scale and provides a quantum gravity description such as quantum gravitational foams. Recently it was shown in \cite{hsy-jhep09,review4,hsy-jpcs12} that the electromagnetism in noncommutative spacetime can be realized as a theory of gravity and the symplectization of spacetime geometry is the origin of gravity. Such picture is called emergent gravity and it proposes a candidate of the origin of spacetime. See also related works in Refs. \cite{rivelles,review1,review2,review3,yasi-prd10,lee-yang,review6,review7,kawai-ks2016} As a bottom-up approach of the emergent gravity formulated in \cite{our-jhep12}, the Eguchi-Hanson metric \cite{eh-plb,eh-ap} in four-dimensional Euclidean gravity is used to construct anti-self-dual symplectic $U(1)$ gauge fields, and $U(1)$ gauge fields corresponding to the Nekrasov-Schwarz instanton \cite{Nekrasov:1998ss} are reproduced by the reverse process \cite{Lee:2012rb}. As a top-down approach of emergent gravity, the $U(1)$ instanton found by Braden and Nekrasov \cite{bn-inst} derives a corresponding gravitational metric.\\ The organization of this paper is as follows. In section \ref{sect2}, some linear algebraic formulas for self-duality are prepared. In section \ref{Einstein and ASD}, the correspondence between the self-dual two-forms and Hermitian-Einstein metrics is studied. In section \ref{sect4}, Hermitian-Einstein metrics are explicitly constructed from noncommutative $U(1)$ instantons. In section \ref{sect5}, the gauge theory realization of the K\"ahler condition is studied. In section \ref{sect6}, we discuss an outlook of this subject. Some technical details are left for the appendices. \section{Self-duality}\label{sect2} \begin{df}[Hodge star operator] An automorphism $\star$ on the set of $4\times 4$ alternative matrices is defined as \[\star\left[\left( \begin{array}{cccc} 0 &\omega_{12} &\omega_{13} & \omega_{14} \\ -\omega_{12} &0 &\omega_{23} & \omega_{24} \\ -\omega_{13} &-\omega_{23} &0 & \omega_{34} \\ -\omega_{14} & -\omega_{24}&-\omega_{34} & 0 \end{array} \right) \right]:=\left( \begin{array}{cccc} 0 &\omega_{34} &-\omega_{24} & \omega_{23} \\ -\omega_{34} &0 &\omega_{14} & -\omega_{13} \\ \omega_{24} &-\omega_{14} &0 & \omega_{12} \\ -\omega_{23} & \omega_{13}&-\omega_{12} & 0 \end{array} \right),\] \[\left(i.e., ~ \omega_{12}\leftrightarrow \omega_{34}, ~ \omega_{13}\leftrightarrow \omega_{42}, ~ \omega_{14}\leftrightarrow \omega_{23}\right). \] In other words, $\star\omega_{kl}$ is defined as \[\star\omega_{kl} =\frac{1}{2} \sum _{m,n}^4 \varepsilon_{klmn} \omega_{mn},\] where $\varepsilon_{klmn}$ is Levi-Civita symbol. The operator $\star$ is called the Hodge star operation in Euclidean $\mathbb{R}^4$. \end{df} \begin{df}[Anti-self-dual matrix] A $4\times 4$ alternative matrix $\omega^{\pm}$ is an (anti-)self-dual matrix if \begin{align} \star\omega^{\pm}=\pm \omega^{\pm}. \label{asd} \end{align} \end{df} An (anti-)self-dual matrix $\theta^{\pm}$ is defined as \begin{align} \theta^{\pm }:=\left( \begin{array}{cccc} 0 &-\theta & 0& 0 \\ \theta &0 &0 & 0 \\ 0 &0 &0 & \mp \theta \\ 0 &0 & \pm \theta & 0 \end{array} \right) & \label{theta} \end{align} where $\theta$ is a real number. Note that $\omega^{\pm}$ and $\theta^{\mp}$ commute each other: \begin{align} \omega^{\pm}\theta^{\mp}=\theta^{\mp}\omega^{\pm}. \label{omegatheta} \end{align} \begin{df}[Matrix $g^{\pm}$]Let $E_4$ be the $4\times4$ unit matrix and $\omega^{\pm}$ be a $4\times 4$ (anti-)self-dual matrix. Assume that $\det\left[E_4-\omega^{\pm}\theta^{\mp} \right]\neq 0$, then $4\times 4$ matrix $g^{\pm}$ is defined as \[g^{\pm}:=2\left(E_4-\omega^{\pm}\theta^{\mp} \right)^{-1}-E_4. \label{g-metric} \] \end{df} \begin{rem}\label{gsym}$g^{\pm}$ is a symmetric matrix because of (\ref{omegatheta}) and it can be inverted to \[\omega^{\pm} =\left(g^{\pm} - E_4 \right) \left(g^{\pm} + E_4 \right)^{-1} \left(\theta^\mp \right)^{-1}.\] \end{rem} The Remark \ref{gsym} allows us to regard $g^{\pm}$ as a metric tensor since it is symmetric and nondegenerate. \begin{lem}\label{lem2.1} For any $4\times 4$ (anti-)self-dual matrix $\omega^\pm$, \begin{align}\label{det=1} \star\omega^\pm =\pm \omega^\pm \Longrightarrow \det\left[ g^\pm \right]=1. \end{align} \end{lem} This lemma is proved by a direct calculation. \begin{df}\label{iskew}The map $\iota_{skew}:\left\{\omega_{\mathbb{C}}\in M_2[\mathbb{C}] \: |~ \omega_{\mathbb{C}}^\dagger=-\omega_{\mathbb{C}} \right\}\longrightarrow M_4[\mathbb{R}]$ is defined as \[\iota_{skew}\left[\left( \begin{array}{cc} \omega_{\mathbb{C}1\bar{1}} & \omega_{\mathbb{C}1\bar{2}} \\ \omega_{\mathbb{C}2\bar{1}} & \omega_{\mathbb{C}2\bar{2}} \end{array} \right) \right]=\left( \begin{array}{cccc} 0 &2\mathrm{i}\omega_{\mathbb{C}1\bar{1}} & \omega_{\mathbb{C}1\bar{2}}-\omega_{\mathbb{C}2\bar{1}} & \mathrm{i}\left(\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} \right) \\ -2\mathrm{i}\omega_{\mathbb{C}1\bar{1}} &0 &-\mathrm{i}\left(\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} \right) & \omega_{\mathbb{C}1\bar{2}}-\omega_{\mathbb{C}2\bar{1}} \\ -\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} &\mathrm{i}\left(\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} \right) &0 & 2\mathrm{i}\omega_{\mathbb{C}2\bar{2}} \\ -\mathrm{i}\left(\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} \right) &-\omega_{\mathbb{C}1\bar{2}}+\omega_{\mathbb{C}2\bar{1}} &-2\mathrm{i}\omega_{\mathbb{C}2\bar{2}} & 0 \end{array} \right).\] Note that $\omega_{\mathbb{C}1\bar{1}}$ and $\omega_{\mathbb{C}2\bar{2}}$ are pure imaginary. \end{df} If the coordinate transformation on the coordinate neighborhood is $z_1:=x^2+\mathrm{i}x^1,z_2:=x^4+\mathrm{i}x^3$, then the $\iota_{skew}$ is the pull-back of a two-form. This means \[ \sum _{k,l=1}^2\omega_{\mathbb{C}k\bar{l}}\mathrm{d}z_k \wedge \mathrm{d}\bar{z}_l =\frac{1}{2}\sum _{k,l=1}^4\omega_{kl}\mathrm{d}x^k \wedge \mathrm{d}x^l= \frac{1}{2}\sum _{k,l=1}^4\left(\iota_{skew}\left[ \omega_\mathbb{C}\right] \right)_{kl}\mathrm{d}x^k \wedge \mathrm{d}x^l.\] The above $\iota_{skew}$ is defined as satisfying this relation. \begin{rem}$\iota_{skew}$ satisfies the following relation \[\det\left[\iota_{skew}\left[\omega_{\mathbb{C}}\right] \right]=16\left(\det\left[\omega_{\mathbb{C}} \right]\right)^2.\] \end{rem} Using this result, the following lemma can be deduced. \begin{lem}\label{lem2.2} Suppose that the anti-Hermitian matrix $\omega_{\mathbb{C}}$ satisfies $\omega_{\mathbb{C}2\bar{2}}=-\omega_{\mathbb{C}1\bar{1}}$, i.e. ${\rm tr \omega_{\mathbb{C}}} = 0$. Then the two-form $\iota_{skew} [\omega_{\mathbb{C}}]$ is anti-self-dual, i.e., \[\star\left\{ \iota_{skew}\left[\left( \begin{array}{cc} \omega_{\mathbb{C}1\bar{1}} & \omega_{\mathbb{C}1\bar{2}} \\ \omega_{\mathbb{C}2\bar{1}} & \omega_{\mathbb{C}2\bar{2}} \end{array} \right)\right]\right\}=-\iota_{skew}\left[\left( \begin{array}{cc} \omega_{\mathbb{C}1\bar{1}} & \omega_{\mathbb{C}1\bar{2}} \\ \omega_{\mathbb{C}2\bar{1}} & \omega_{\mathbb{C}2\bar{2}} \end{array} \right)\right].\] \end{lem} \section{Hermitian-Einstein metrics and (anti-)self-dual two-forms}\label{Einstein and ASD} In this section, we discuss how to make a Hermitian-Einstein metric from an anti-self-dual two-form. Let us define a $u\left(1 \right)$-valued two-form on $\mathbb{R}^4$ by \[ \sum _{k,l=1}^4\omega_{kl}\mathrm{d}x^k \wedge \mathrm{d}x^l.\] where $\omega$ is an alternative matrix $\left(\omega \right)_{kl}:=\omega_{kl}$. If $\omega$ is an anti-self-dual matrix, then the two-form is called anti-self-dual two-form. \subsection{Ricci flat metrics and Hermitian-Einstein metrics} Let $M$ be a Hermitian manifold and $h$ be its metric. As a well-known fact, Ricci curvature $R_{\bar{j}k}$ for a Hermitian manifold $\left(M,h,\nabla \right)$ with the Levi-Civita connection $\nabla$ takes a simple form \begin{align}\label{ricci} R_{\bar{j}k}=\partial _{\bar{j}}\partial _{k}\log \left(\det\left[ h\right] \right). \end{align} See, for example, \cite{Kobayashi_Nomizu,besse}. Let $\lambda$ be a cosmological constant. When $h$ satisfies the Einstein's equation. \[R_{\bar{k}l}=\lambda h_{\bar{k}l}\] then $M$ is called an Einstein manifold. In this paper we will focus on a Ricci flat manifold (i.e. $R_{\bar{k}l} = 0$ or $\lambda = 0$). We consider $M$ as a real manifold with local coordinates $x^\mu\left(\mu=1,2,3,4 \right)$. \begin{df}\label{isym}The map $\iota_{sym}:\left\{h\in M_2[\mathbb{C}]~|~h^\dagger=h \right\}\longrightarrow M_4[\mathbb{R}]$ is defined as \[\iota_{sym}\left[\left( \begin{array}{cc} h_{1\bar{1}} & h_{1\bar{2}} \\ h_{2\bar{1}} & h_{2\bar{2}} \end{array} \right) \right] =\left( \begin{array}{cccc} h_{1\bar{1}} &0 &\frac{1}{2}\left(h_{1\bar{2}}+h_{2\bar{1}} \right) &\frac{1}{2\mathrm{i}}\left(h_{2\bar{1}}-h_{1\bar{2}} \right) \\ 0 &h_{1\bar{1}} &-\frac{1}{2\mathrm{i}}\left(h_{2\bar{1}}-h_{1\bar{2}} \right) & \frac{1}{2}\left(h_{1\bar{2}}+h_{2\bar{1}} \right) \\ \frac{1}{2}\left(h_{1\bar{2}}+h_{2\bar{1}} \right) &-\frac{1}{2\mathrm{i}}\left(h_{2\bar{1}}-h_{1\bar{2}} \right) &h_{2\bar{2}} &0 \\ \frac{1}{2\mathrm{i}}\left(h_{2\bar{1}}-h_{1\bar{2}} \right) &\frac{1}{2}\left(h_{1\bar{2}}+h_{2\bar{1}} \right) &0 & h_{2\bar{2}} \end{array} \right).\] where $h$ is a matrix and $\left( h\right)_{k\bar{l}}:=h_{k\bar{l}}$. \end{df} \begin{rem}Assume that $h$ is a Hermitian metric. If the coordinate transformation on a coordinate neighborhood is $z^1:=x^2+\mathrm{i}x^1,z^2:=x^4+\mathrm{i}x^3$, the $\iota_{sym}$ is then the pull-back of the Hermitian metric given by \[ \sum _{k,l=1}^2h_{k\bar{l}}\mathrm{d}z_k \mathrm{d}\bar{z}_l= \sum _{k,l=1}^4\left(\iota_{sym}\left[ h\right] \right)_{kl}\mathrm{d}x^k \mathrm{d}x^l.\] Hence $\iota_{sym}$ squares the determinant: \[\det\left[\iota_{sym}\left(h \right) \right]=(\det \left[h \right])^2.\] \end{rem} A Hermitian metric made with $\iota_{sym}^{-1}$ will be used below.\\ \begin{df}If $\tilde{h}\in C^\infty \left(U,M_2[\mathbb{C}] \right)$ and $\tilde{h}^\dagger =\tilde{h}$, then $$\tilde{h}>0~in~U ~~ \Longleftrightarrow ~~ \forall u\in U, ~ \tilde{h}\left(u \right)>0 $$ where $\tilde{h}\left(u \right)>0$ means that $\tilde{h}$ is positive definite as a Hermitian matrix. \end{df} \begin{lem} \label{lem3.1} If $h \in C^\infty \left(U,M_2[\mathbb{C}] \right)$ is a Hermitian matrix with $\det\left[h \right]=1$ and $h$ is positive (negative) at ${}^\exists p\in U$, then $h$ is positive (negative) in $U$. \end{lem} \begin{pf}This follows from \begin{align*} \lefteqn{\left\{h\in M_2[\mathbb{C}]~\big|~h=h^\dagger,~\det\left[h \right]=1 \right\}} \\ &=\left\{\left( \begin{array}{cc} a & b \\ \bar{b} & d \end{array} \right) \in M_2[\mathbb{C}]~\big|~a,d\in\mathbb{R},~a>0,~d>0,~a d\geq 1,~\left|b \right|=\sqrt{a d- 1} \right\}\\ &\coprod \left\{\left( \begin{array}{cc} a & b \\ \bar{b} & d \end{array} \right) \in M_2[\mathbb{C}]~\big|~a,d\in\mathbb{R},~a<0,~d<0,~a d\geq 1,~\left|b \right|=\sqrt{a d- 1} \right\} \end{align*} which means two spaces are disconnected. \qed\end{pf} From the above discussions, the following theorem is obtained. \begin{thm} \label{thm-kal-ma} Let $\omega^\pm$ be an (anti-)self-dual two-form on an open neighborhood $U$, i.e. $\star \omega^\pm =\pm\omega^\pm$, \label{masspro} and \begin{align} \label{sdmetric} h^\pm:=\iota_{sym}^{-1} \left[2 \left(E_4-\omega^\pm \theta^\mp \right)^{-1}-E_4 \right]. \end{align} Then $h^\pm$ gives a Ricci-flat Hermitian metric on $U$. So $(U, h^\pm)$ is a local realization of an Einstein manifold. \end{thm} \begin{pf}Because of Lemma \ref{lem2.1}, if $\star \omega^\pm =\pm\omega^\pm$, then \begin{align}\label{sd} \det\left[h^\pm \right]=1. \end{align} Because of Lemma \ref{lem3.1} and Remark \ref{gsym}, $h^\pm$ is a metric tensor. From equations (\ref{ricci}) and (\ref{sd}), $R_{\bar{j}k}=\partial _{\bar{j}}\partial _{k}\log\left(\det\left[ h^\pm\right] \right)=0$. \qed\end{pf} Local complex coordinates can be arranged in such a way that the Jacobians of the transition functions on overlapping charts are one on all the overlaps. In that case, $\det [h^\pm]$ is a globally defined function and the Ricci-flat condition reduces to the Monge-Amp\'ere equation \cite{ma-eq} \begin{equation}\label{ma-eq} \det [h^\pm] = \kappa, \end{equation} where the constant $\kappa$ is related to the volume of a K\"ahler manifold that depends only on the K\"ahler class. Therefore Theorem \ref{thm-kal-ma} implies that the self-duality for the two-form $\omega^\pm$ is equivalent to the Ricci-flat condition \eq{ma-eq} of K\"ahler manifolds defined by the metric $h^\pm$ \cite{u1-cym}. \section{Hermitian-Einstein metric from noncommutative instanton on $\mathbb{C}^2$}\label{sect4} In the previous section we found the way to construct a Hermitian-Einstein metric from an (anti-)self-dual two-form. To construct the Hermitian-Einstein metric, we will employ the instanton curvature on noncommutative $\mathbb{C}^2$ as the (anti-)self-dual two-form. There are many ways to obtain noncommutative $\mathbb{C}^2$ (see \cite{nc-review} for a review and references therein). We use the Fock representation of noncommutative $\mathbb{C}^2$ given in \cite{Sako:2016gqb}, which is based on the Karabegov's deformation quantization \cite{Karabegov1996}. There is a simple dictionary between the Fock representation and ordinary functions. Using the dictionary, the Hermitian-Einstein metric is expressed in terms of usual functions. \subsection{Noncommutative $\mathbb{C}^2$} Consider a noncommutative algebra $\left(C^\infty \left(\mathbb{C}^2 \right)\left[\left[\hbar \right] \right],* \right)$ led by (\ref{f*g}) in Appendix \ref{dq}. The star product induces a Heisenberg algebra \begin{align} \label{ncp} \left[z^k,\bar{z}^l\right]_*=-\zeta_k\delta_{kl}, \qquad \left[z^k,z^l\right]_*=0,\qquad \left[\bar{z}^k,\bar{z}^l\right]_*=0 , \end{align} where $\left[x,y\right]_*:=x*y-y*x $. We represent it by creation and annihilation operators given by \[a_k:=\frac{\bar{z}^k}{\sqrt{\zeta_k}},\qquad a_k^\dagger:=\frac{z^k}{\sqrt{\zeta_k}},\] then \[\left[a_k,a_l^\dagger \right]_*=\delta_{kl},\qquad \left[a_k^\dagger ,a_l^\dagger \right]_*=0,\qquad \left[a_k,a_l\right]_*=0.\] In the following $\zeta_1=\zeta_2=\zeta>0$ is assumed.\\ Note that the choice of a noncommutative parameter has the freedom associated with a choice of a background two-form \cite{sw-ncft}. Here the $\zeta$ in (\ref{ncp}) is regarded as the only noncommutative parameter. However, in Section \ref{sect5}, we will implicitly assume the identification $\zeta := 2 \theta$ since we will work in the background-independent prescription, i.e. $\theta = B^{-1}$.\\ The algebra $\mathcal{F}$ on $\mathbb{C}$ is defined as follows. The Fock space $\mathcal{H}$ is a linear space spanned by the bases generated by acting $a_l^\dagger$'s on $\Ket{0,0}$ : \begin{align} \frac{1}{\sqrt{m_1!m_2!}}\left( a_1^\dagger\right)^{m_1}_**\left( a_2^\dagger\right)^{m_2}_*\Ket{0,0}= \Ket{m_1,m_2} ,\label{fsp} \end{align} where $m_1$ and $m_2$ are positive integers and $\left( a\right)^m_*$ stands for $\overbrace{a * \cdots * a}^m$. The ground state $\Ket{0,0}$ satisfies $a_l\Ket{0,0}=0, ~ \forall ~ l$. Here, we define the basis of a dual vector space by acting $a_l$'s on $\Bra{0,0}$ as $$\frac{1}{\sqrt{n_1!n_2!}}\Bra{0,0}\left(a_1\right)^{n_1}_**\left(a_2\right)^{n_2}_*=\Bra{n_1,n_2} , $$ where $\Bra{0,0}$ satisfies $\Bra{0,0}a_l^\dagger=0, ~ \forall ~ l$. Then we define a set of linear operators as \begin{align} \mathcal{F}:=span_{\mathbb{C}}\left(\Ket{m_1,m_2}\Bra{n_1,n_2};m_1,m_2,n_1,n_2=0,1,2,\cdots \right) \label{ketbra} \end{align} where $\left( \Ket{m_1,m_2}\Bra{n_1,n_2}\right)\Ket{k_1,k_2}=\delta_{k_1n_1}\delta_{k_2n_2}\Ket{m_1,m_2}$ and $\Bra{k_1,k_2}\left( \Ket{m_1,m_2}\Bra{n_1,n_2}\right)=\delta_{k_1m_1}\delta_{k_2m_2}\Bra{n_1,n_2}$. The product on $\mathcal{F}$ is defined as $$\left(\Ket{j_1,j_2}\Bra{k_1,k_2} \right)\circ\left( \Ket{m_1,m_2}\Bra{n_1,n_2}\right):= \delta_{k_1m_1}\delta_{k_2m_2}\Ket{j_1,j_2}\Bra{n_1,n_2},$$ so, $\mathcal{F}$ is an algebra. There is a one to one correspondence between $\mathcal{F}$ and some subalgebra of $C^\infty \left(\mathbb{C}^2 \right)$. For arbitrary noncommutative K\"ahler manifold obtained by deformation quantization with separation of variables \cite{Karabegov1996}, we can find the similar correspondence \cite{Sako:2016gqb}. The following is the simplest example of the correspondence. \begin{df}\label{iota}(Twisted Fock representation). The linear map $\iota:\mathcal{F}\longrightarrow C^\infty \left(\mathbb{C}^2 \right)$ is defined as \begin{align} \iota\left(\Ket{m_1,m_2}\Bra{n_1,n_2} \right)=e_{\left( m_1,m_2,n_1,n_2\right)}: =\frac{z_1^{m_1}z_2^{m_2}\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}\bar{z}_1^{n_1} \bar{z}_2^{n_2}}{\sqrt{m_1!m_2!n_1!n_2!}\left( \sqrt{\zeta}\right)^{m_1+m_2+n_1+n_2}},\label{Fockrep} \end{align} especially $\iota\left( \Ket{0,0}\Bra{0,0}\right)=e_{\left( 0,0,0,0\right)}=\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}.$ \end{df} \begin{prop}Let $\iota \left(\mathcal{F}\right)$ be defined by \begin{align} \iota \left(\mathcal{F}\right):=span_{\mathbb{C}}\left(e_{\left( m_1,m_2,n_1,n_2\right)};m_1,m_2,n_1,n_2=0,1,2,\cdots \right). \end{align} Then $\left\{\iota \left(\mathcal{F}\right),* \right\}$ is an algebra where $*$ is in (\ref{f*g}). \end{prop} \begin{pf} After a little algebra, one can deduce the following identity \begin{align} e_{\left( k_1,k_2,l_1,l_2\right)}*e_{\left( m_1,m_2,n_1,n_2\right)}=\delta_{l_1m_1}\delta_{l_2m_2}e_{\left( k_1,k_2,n_1,n_2\right)}. \label{e*e} \end{align} Details are given in \cite{Sako:2016gqb}. \qed\end{pf} The identity (\ref{e*e}) derives the following fact. \begin{prop}\label{dictionary}The algebras $\left(\mathcal{F},\circ\right)$ and $\left\{ \iota \left(\mathcal{F}\right),*\right\}$ are isomorphic. \end{prop} This isomorphism $\iota$ is a ``Fock space - function space'' dictionary. From this isomorphism, we do not distinguish these two algebras and we only use $*$ to represent products in the following.\\ \bigskip Here we consider a $U(1)$ gauge theory on noncommutative $\mathbb{C}^2$. $U(1)$ gauge connection in the noncommutative space is defined as follows (see for example \cite{Nekrasov:2000ih}). \begin{df}Rescaled coordinates of $\mathbb{C}^2$ are defined as \[\hat{\partial }_{z_l}:=\frac{\bar{z}_l}{\zeta_l}.\] This acts on $\mathcal{H}$ as a linear operator. \end{df} Using $\hat{\partial }_{z_l},\hat{\partial }_{\bar{z}_m}$, let us introduce covariant derivatives and the gauge curvature as follows. \begin{df}\label{cov}Covariant derivatives for a scalar field in fundamental representation $\phi\in \mathcal{F}$ on noncommutative $\mathbb{C}^2$ are defined as \[\hat{\nabla }_{z_l}\hat{\phi}:=\left[\hat{\partial }_{z_l},\hat{\phi }\right]_*+\hat{A}_{z_l}*\hat{\phi} =-\hat{\phi}*\hat{\partial }_{z_l}+\hat{D}_{z_l}*\hat{\phi}\] where we define a local gauge field $\hat{A}_{z_l}\in \mathcal{F}$ and \[\hat{D}_{z_l}:=\hat{\partial }_{z_l}+\hat{A}_{z_l}.\] The gauge curvature is defined as \begin{align} \hat{F}_{z_l\bar{z}_m}:&=\mathrm{i}\left[\hat{\nabla }_{z_l},\hat{\nabla }_{\bar{z}_m} \right]_* =-\frac{\mathrm{i}\delta_{lm}}{\zeta_l}+\mathrm{i}\left[\hat{D}_{z_l},\hat{D}_{\bar{z}_m} \right]_*, \\ \hat{F}_{z_lz_m}:&=\mathrm{i}\left[\hat{\nabla }_{z_l},\hat{\nabla }_{z_m} \right]_* =\mathrm{i}\left[\hat{D}_{z_l},\hat{D}_{z_m} \right]_*, \nonumber \\ \hat{F}_{\bar{z}_l\bar{z}_m}:&=\mathrm{i}\left[\hat{\nabla }_{\bar{z}_l},\hat{\nabla }_{\bar{z}_m} \right]_* =\mathrm{i}\left[\hat{D}_{\bar{z}_l},\hat{D}_{\bar{z}_m} \right]_*. \nonumber \end{align} \end{df} \subsection{Ricci-flat metrics from noncommutative $k$-instantons} In this section, we make Ricci-flat metrics on a local neighborhood from noncommutative instantons on ${\mathbb C}^2$. As we saw in Section \ref{Einstein and ASD}, (anti)-self-dual two-forms satisfying (\ref{asd}) derive Ricci-flat metrics. Nekrasov and Schwarz found in \cite{Nekrasov:1998ss} how to construct noncommutative instantons on ${\mathbb C}^2$ by using the ADHM method and the general solutions for the $U(1)$ gauge theory are given in \cite{Nekrasov:2000ih}. We introduce the commutation relation of complex coordinates as (\ref{ncp}). As (anti)-self-dual two-forms in Section \ref{Einstein and ASD}, we employ noncommutative instantons given in \cite{Ishikawa:2001ye}.\\ The general instanton solutions (see \cite{Ishikawa:2001ye}) satisfy the (anti)-self-dual relation. An instanton curvature tensor is described by \[\hat{F}^-_{\mathbb{C}}\left[k \right]:=\left( \begin{array}{cc} \hat{F}^-_{z_1\bar{z}_1}\left[k \right] & \hat{F}^-_{z_1\bar{z}_2}\left[k \right] \\ \hat{F}^-_{z_2\bar{z}_1}\left[k \right] & -\hat{F}^-_{z_1\bar{z}_1}\left[k \right] \end{array} \right) , \] and satisfies (\ref{asd}): \begin{align} \star\left(\iota_{skew}\left(\hat{F}^-_{\mathbb{C}}\left[k \right] \right) \right)=-\iota_{skew}\left(\hat{F}^-_{\mathbb{C}}\left[k \right] \right). \end{align} See Lemma \ref{lem2.2} in Section \ref{sect2}. This fact leads to the following result. \begin{prop} If $\hat{F}^-_{\mathbb{C}}$ is a $k$-instanton curvature tensor of $U(1)$ gauge theory on noncommutative ${\mathbb C}^2$, and \begin{align} h\left[k \right]:= & \iota^{-1}_{sym}\left\{ 2\left(E_4-\iota_{skew}\left(\hat{F}^-_{\mathbb{C}}\left[k \right] \right) \theta^+ \right)^{-1}-E_4 \right\} \nonumber \\ =&\frac{1}{4\left|\hat{F}_\mathbb{C}^-\left[k \right] \right|\theta^2-1}\left( \begin{array}{cc} -4\mathrm{i}\hat{F}^-_{z_1\bar{z}_1}\left[k \right]\theta-2 & -4\mathrm{i}\hat{F}^-_{z_1\bar{z}_2}\left[k \right]\theta \\ -4\mathrm{i}\hat{F}^-_{z_2\bar{z}_1}\left[k \right]\theta & 4\mathrm{i}\hat{F}^-_{z_1\bar{z}_1}\left[k \right]\theta-2 \end{array} \right)-\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right), \end{align} then $h\left[k \right]$ is an Einstein (Ricci-flat) metric. \end{prop} A concrete example of $k$-instanton curvature tensors is given in \cite{Ishikawa:2001ye} and the curvature is written by using linear operators on a Fock space. It is known from (\ref{Fockrep}) and Proposition \ref{dictionary} how to translate the operators into functions. (See also Appendix \ref{fock} and \cite{Sako:2016gqb}.) Then the $k$-instanton curvature tensor is expressed by concrete elementary functions as follows: \begin{align*} \lefteqn{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]=\frac{\mathrm{i}}{\zeta}-\frac{\mathrm{i}}{\zeta}\sum _{n_2=0}^\infty \frac{z_2^{n_2}\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}\bar{z}_2^{n_2}}{n_2!\zeta^{n_2}} \left(d_1\left(0,n_2;k \right) \right)^2} \\ &-\frac{\mathrm{i}}{\zeta} \sum _{n_1=1}^\infty \sum _{n_2=0}^\infty \frac{z_1^{n_1}z_2^{n_2}\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}\bar{z}_1^{n_1}\bar{z}_2^{n_2}}{n_1!n_2!\zeta^{n_1+n_2}} \left\{\left(d_1\left(n_1,n_2;k \right) \right)^2-\left(d_1\left(n_1-1,n_2;k \right) \right)^2 \right\}, \end{align*} \begin{align*} \lefteqn{\hat{F}^-_{z_1\bar{z}_2}\left[k \right]=-\frac{\mathrm{i}}{\zeta}\frac{z_1^{k-1}z_2\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}} {\sqrt{\left(k-1 \right)!}\left( \sqrt{\zeta}\right)^{k}}d_1\left(k-1,1;k \right)d_2\left(0,0;k \right)} \\ &-\frac{\mathrm{i}}{\zeta}\sum _{n_1=1}^{k-1} \frac{z_1^{n_1+k-1}z_2\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}\bar{z}_1^{n_1}} {\sqrt{\left(n_1+k-1 \right)!n_1!}\left( \sqrt{\zeta}\right)^{2n_1+k}} \left\{d_1\left(n_1+k-1,1;k \right)d_2\left(n_1,0;k \right)-d_1\left(n_1-1,0;k \right)d_2\left(n_1-1,0;k \right) \right\} \\ &-\frac{\mathrm{i}}{\zeta} \sum _{n_1=1}^\infty \sum _{n_2=1}^\infty \frac{z_1^{n_1-1}z_2^{n_2+1}\mathrm{e}^{-\frac{z^1\bar{z}^1+z^2\bar{z}^2}{\zeta}}\bar{z}_1^{n_1}\bar{z}_2^{n_2}} {\sqrt{\left(n_1-1 \right)!\left(n_2+1 \right)!n_1!n_2!}\left( \sqrt{\zeta}\right)^{2n_1+2n_2}} \\ &\times \left\{d_1\left(n_1-1,n_2+1;k \right)d_2\left(n_1,n_2;k \right)-d_1\left(n_1-1,n_2;k \right)d_2\left(n_1-1,n_2;k \right) \right\}, \end{align*} \[\hat{F}^-_{z_1\bar{z}_2}\left[k \right] =- \hat{F}^-_{z_2\bar{z}_1}\left[k \right]^\dagger,\] where $n_2\neq 0$ and \begin{align} \lefteqn{d_1\left(n_1,0;k \right)=\sqrt{n_1+k+1}\sqrt{\frac{\Lambda\left(n_1+k+1,0 \right)}{\Lambda\left(n_1+k,0 \right)}},}\nonumber \\ &d_1\left(n_1,n_2;k \right)=\sqrt{n_1+1}\sqrt{\frac{\Lambda\left(n_1+1,n_2 \right)}{\Lambda\left(n_1,n_2 \right)}},\label{d1} \\ &d_2\left(n_1,0;k \right)=\sqrt{\frac{\Lambda\left(n_1+k,1 \right)}{\Lambda\left(n_1+k,0 \right)}},\nonumber \\ &d_2\left(n_1,n_2;k \right)=\sqrt{n_2+1}\sqrt{\frac{\Lambda\left(n_1,n_2+1 \right)}{\Lambda\left(n_1,n_2 \right)}} \label{d2} . \end{align} Here $$\Lambda\left[k \right]\left(n_1,n_2 \right)=\frac{w_k\left[k \right]\left(n_1,n_2 \right)} {w_k\left[k \right]\left(n_1,n_2 \right)-2kw_{k-1}\left[k \right]\left(n_1,n_2 \right)},$$ and $$w_n\left[k \right]\left(n_1,n_2 \right)=\sum _{l=0}^n \left\{ \frac{n!}{l!}\frac{\left(n_1-n_2+k+l \right)!}{\left(n_1-n_2-k \right)!} \frac{2^{\left(n-l \right)}}{\left(n-l \right)!}\frac{\left(n_2+\left(n-l \right) \right)!}{n_2!}\right\}.$$ Note that some notations are slightly changed from \cite{Ishikawa:2001ye} and imaginary unit factor causes here. See also Appendix \ref{imaginaryunit}.\\ Using these instanton curvatures, Hermitian-Einstein metrics can be constructed by concrete elementary functions according to the Theorem \ref{masspro}. \subsection{Einstein metric from finite $N$ }\label{finiten} The full noncommutative $U\left(1 \right)$ instanton solution is very complicated. For simplicity, let us consider the $\zeta$-expansion. \\ In the previous subsection, $\hat{F}^-$ is represented by an infinite series \begin{align} \hat{F}^-=\sum _{n=1}^\infty\left(\frac{1}{\zeta} \right)^{\frac{n}{2}} \hat{F}^-_{\left({\frac{n}{2}} \right)}. \label{infiniteseries} \end{align} The anti-self-dual condition $\star \hat{F}^-=-\hat{F}^-$ implies \begin{align} \star \hat{F}^-_{\left({\frac{n}{2}} \right)}=-\hat{F}^-_{\left({\frac{n}{2}} \right)} \end{align} for each $n/2$. Therefore it is possible to employ an arbitrary partial sum of (\ref{infiniteseries}) determined by a subset $\displaystyle S\subset {\frac{1}{2}}\mathbb{Z}_{>0}$ \begin{align} \hat{F}^-_S=\sum _{{\frac{n}{2}} \in S}\left(\frac{1}{\zeta} \right)^{\frac{n}{2}} \hat{F}^-_{\left({\frac{n}{2}} \right)}\label{Fseries} \end{align} for the anti-self-dual two-form to construct a Hermitian-Einstein metric $h$ without losing rigorousness.\footnote{One may choose even more loose condition than (\ref{Fseries}). One can choose a different subset $S$ for each $\hat{F}^-_{z_1\bar{z}_1},\hat{F}^-_{z_1\bar{z}_2}$ to obtain a Hermitian-Einstein metric.} In the following we consider \begin{align} \hat{F}^-_{\left\{{\frac{N}{2}} \right\}}:=\sum _{n=1/2}^{{N}/{2}} \left(\frac{1}{\zeta} \right)^{\frac{n}{2}} \hat{F}^-_{\left({\frac{n}{2}} \right)}. \end{align} \begin{ex} First let us make the Ricci-flat metric $h\left[k \right]_{\left\{1 \right\}}$ from ${\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{1 \right\}}$. The curvature tensor in this case is ${\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{1 \right\}}=\left( \begin{array}{cc} \frac{\mathrm{i}}{\zeta} & 0 \\ 0 & -\frac{\mathrm{i}}{\zeta} \end{array} \right)$, and its determinant is $\det\left[{\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{1 \right\}} \right]=\frac{1}{\zeta^2}.$ So the metric $h\left[k \right]_{\left\{1 \right\}}$ is given by \begin{align} h\left[k \right]_{\left\{1 \right\}} &:=\frac{1}{4~\det\left[{\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{1 \right\}} \right]\theta^2-1}\left( \begin{array}{cc} -4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]}_{\left\{1 \right\}}\theta-2 & -4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_2}\left[k \right]}_{\left\{1 \right\}}\theta \\ -4\mathrm{i}{\hat{F}^-_{z_2\bar{z}_1}\left[k \right]}_{\left\{1 \right\}}\theta & 4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]}_{\left\{1 \right\}}\theta-2 \end{array} \right)-\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right) \nonumber \\ &=\frac{1}{1-4\zeta^{-2}\theta^2}\left( \begin{array}{cc} 1-4\zeta^{-1}\theta+4\zeta^{-2}\theta^2 & 0 \\ 0 & 1+4\zeta^{-1}\theta+4\zeta^{-2}\theta^2 \end{array} \right)=\left( \begin{array}{cc} \frac{1-2\zeta^{-1}\theta}{1+2\zeta^{-1}\theta} & 0 \\ 0 & \frac{1+2\zeta^{-1}\theta}{1-2\zeta^{-1}\theta} \end{array} \right). \nonumber \end{align} This corresponds to the Euclidean metric essentially. \end{ex} \begin{ex} Let us make a Ricci-flat metric $h\left[k \right]_{\left\{2 \right\}}$ from ${\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{2 \right\}}$. From (\ref{f11k}),(\ref{f12k}), \begin{align} {\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{2 \right\}}& =\frac{\mathrm{i}}{\zeta} \left[ 1-\frac{z_2\bar{z}_2}{\zeta} \left(d_1\left(0,1;k \right) \right)^2 -\frac{z_1\bar{z}_1}{\zeta}\left\{\left(d_1\left(1,0;k \right) \right)^2-\left(d_1\left(0,0;k \right) \right)^2 \right\} \right] \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right) \nonumber \\ &-\frac{\mathrm{i}d_1\left(k-1,1;k \right)d_2\left(0,0;k \right)}{\zeta^{1+k/2}\sqrt{\left(k-1 \right)!}}\left( \begin{array}{cc} 0 & z_1^{k-1}z_2 \nonumber\\ \bar{z}_1^{k-1}\bar{z}_2 & 0 \end{array} \right). \nonumber \end{align} Then its determinant is \begin{align} \det\left[{\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{2 \right\}} \right] & =\frac{1}{\zeta^2} \left[ 1-\frac{z_2\bar{z}_2}{\zeta} \left(d_1\left(0,1;k \right) \right)^2 -\frac{z_1\bar{z}_1}{\zeta}\left\{\left(d_1\left(1,0;k \right) \right)^2-\left(d_1\left(0,0;k \right) \right)^2 \right\} \right]^2 \nonumber \\ &+ \frac{\left\{d_1\left(k-1,1;k \right)\right\}^2\left\{d_2\left(0,0;k \right) \right\}^2z_1^{k-1}z_2\bar{z}_1^{k-1}\bar{z}_2 } {\zeta^{2+k}\left(k-1 \right)!}.\nonumber \end{align} So the metric $h\left[k \right]_{\left\{2 \right\}}$ is given by \begin{align} h\left[k \right]_{\left\{2 \right\}} &:=\frac{1}{4\det\left[{\hat{F}^-_\mathbb{C}\left[k \right]}_{\left\{2 \right\}} \right]\theta^2-1}\left( \begin{array}{cc} -4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]}_{\left\{2 \right\}}\theta-2 & -4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_2}\left[k \right]}_{\left\{2 \right\}}\theta \\ -4\mathrm{i}{\hat{F}^-_{z_2\bar{z}_1}\left[k \right]}_{\left\{2 \right\}}\theta & 4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]}_{\left\{2 \right\}}\theta-2 \end{array} \right)-\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right), \nonumber \end{align} which can be calculated concretely though its expression becomes complex. To simplify this we assume $k>3$, then \begin{align} h\left[k \right]_{\left\{2 \right\}}& =\left\{ \frac{2}{1-4\det\left[{\hat{F}_\mathbb{C}^-\left[k \right]}_{\left\{2 \right\}} \right] \theta^2}-1\right\}\left( \begin{array}{cc} 1 & 0 \\ 0& 1 \end{array} \right) +\frac{4\mathrm{i}{\hat{F}^-_{z_1\bar{z}_1}\left[k \right]}_{\left\{2 \right\}}\theta} {1-4\det\left[{\hat{F}_\mathbb{C}^-\left[k \right]}_{\left\{2 \right\}} \right]\theta^2}\left( \begin{array}{cc} 1 & 0 \\ 0& -1 \end{array} \right) \nonumber \\ &=\left\{ \frac{2}{1-4\theta^2\zeta^{-2}\left[ 1-\frac{z_2\bar{z}_2}{\zeta} \left(d_1\left(0,1;k \right) \right)^2 -\frac{z_1\bar{z}_1}{\zeta}\left\{\left(d_1\left(1,0;k \right) \right)^2-\left(d_1\left(0,0;k \right) \right)^2 \right\} \right]^2}-1\right\}\left( \begin{array}{cc} 1 & 0 \\ 0& 1 \end{array} \right) \nonumber \\ &-\frac{\frac{4\theta}{\zeta} \left[ 1-\frac{z_2\bar{z}_2}{\zeta} \left(d_1\left(0,1;k \right) \right)^2 -\frac{z_1\bar{z}_1}{\zeta}\left\{\left(d_1\left(1,0;k \right) \right)^2-\left(d_1\left(0,0;k \right) \right)^2 \right\} \right] } {1-4\theta^2\zeta^{-2}\left[ 1-\frac{z_2\bar{z}_2}{\zeta} \left(d_1\left(0,1;k \right) \right)^2 -\frac{z_1\bar{z}_1}{\zeta}\left\{\left(d_1\left(1,0;k \right) \right)^2-\left(d_1\left(0,0;k \right) \right)^2 \right\} \right]^2}\left( \begin{array}{cc} 1 & 0 \\ 0& -1 \end{array} \right). \nonumber \end{align} \end{ex} In next subsection, we discuss a Hermitian-Einstein metric obtained from $1$-instanton solution. \subsection{Hermitian-Einstein metric from a 1-instanton}\label{1inst} For the simplest example of the Hermitian-Einstein metric given in the previous discussion, we describe a Hermitian-Einstein metric obtained from a single noncommutative $U(1)$ instanton. Now we pay attention to low order terms.\\ For $k=1,\hat{F}^-_{\mathbb{C}}\left[1 \right]$ is \begin{align*} \lefteqn{\hat{F}^-_{z_1\bar{z}_1}\left[1 \right]=\frac{\mathrm{i}}{\zeta}-\frac{\mathrm{i}z_2\bar{z}_2}{\zeta^2} \left(d_1\left(0,1;1 \right) \right)^2 -\frac{\mathrm{i}z_1\bar{z}_1}{\zeta^2}\left\{\left(d_1\left(1,0;1 \right) \right)^2-\left(d_1\left(0,0;1 \right) \right)^2 \right\} +\mathcal{O}\left(\zeta^{-3} \right) } \\ &=\frac{\mathrm{i}}{\zeta}- \frac{2\mathrm{i}}{3}\frac{z_2\bar{z}_2}{\zeta^2}-\frac{\mathrm{i}z_1\bar{z}_1}{\zeta^2} \left\{\frac{5}{2}-\frac{4}{3} \right\}+\mathcal{O}\left(\zeta^{-3} \right) =\frac{\mathrm{i}}{\zeta}- \frac{\mathrm{i}}{6\zeta^2}\left(4z_2\bar{z}_2+7z_1\bar{z}_1\right)+\mathcal{O}\left(\zeta^{-3} \right) \\ &\hat{F}^-_{z_1\bar{z}_2}\left[1 \right]=-\frac{\mathrm{i}z_2}{\zeta^{3/2}}\left(1-\frac{z_1\bar{z}_1}{\zeta} -\frac{z_2\bar{z}_2}{\zeta} \right) d_1\left(0,1;1 \right)d_2\left(0,0;1 \right)+\mathcal{O}\left(\zeta^{-3} \right) \\ &=-\frac{2\mathrm{i}z_2}{3\zeta^{3/2}}\left(1-\frac{z_1\bar{z}_1}{\zeta} -\frac{z_2\bar{z}_2}{\zeta} \right)+\mathcal{O}\left(\zeta^{-3} \right) \\ &\hat{F}^-_{z_2\bar{z}_1}\left[1 \right]=-\frac{\mathrm{i}\bar{z}_2}{\zeta^{3/2}}\left(1-\frac{z_1\bar{z}_1}{\zeta} -\frac{z_2\bar{z}_2}{\zeta} \right) d_1\left(0,1;1 \right)d_2\left(0,0;1 \right)+\mathcal{O}\left(\zeta^{-3} \right) \\ &=-\frac{2\mathrm{i}\bar{z}_2}{3\zeta^{3/2}}\left(1-\frac{z_1\bar{z}_1}{\zeta} -\frac{z_2\bar{z}_2}{\zeta} \right)+\mathcal{O}\left(\zeta^{-3} \right) \end{align*} from (\ref{f11k}),(\ref{f121}). Then \begin{align} det\left[\hat{F}_\mathbb{C}^-\left[1 \right]_{\left\{2 \right\}}\right] &=\frac{4z_2\bar{z}_2}{9\zeta^5}\left(\zeta-z_1\bar{z}_1-z_2\bar{z}_2\right)^2 -\frac{1}{36\zeta^4}\left( 6\zeta- 7z_1\bar{z}_1-4z_2\bar{z}_2\right)^2 \end{align} From this $1$-instaoton curvature, the Hermitian-Einstein metric is given as \begin{align*} &h \left[1 \right]_{\left\{2 \right\}}:=\frac{1}{4~\det\left[\hat{F}_\mathbb{C}^-\left[1 \right]\right]_{\left\{2 \right\}}\theta^2-1}\left( \begin{array}{cc} -4\mathrm{i}\hat{F}^-_{z_1\bar{z}_1}\left[1 \right]_{\left\{2 \right\}}\theta-2 & -4\mathrm{i}\hat{F}^-_{z_1\bar{z}_2}\left[1 \right]_{\left\{2 \right\}}\theta \\ -4\mathrm{i}\hat{F}^-_{z_2\bar{z}_1}\left[1 \right]_{\left\{2 \right\}}\theta & 4\mathrm{i}\hat{F}^-_{z_1\bar{z}_1}\left[1 \right]_{\left\{2 \right\}}\theta-2 \end{array} \right)-\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right) \\ &=\frac{4}{1-4\left\{\frac{4z_2\bar{z}_2}{9\zeta^5}\left(\zeta-z_1\bar{z}_1-z_2\bar{z}_2\right)^2 -\frac{1}{36\zeta^4}\left( 6\zeta- 7z_1\bar{z}_1-4z_2\bar{z}_2\right)^2 \right\}\theta^2} \\ &\times \left\{\frac{1}{2}\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right)+ \frac{\theta}{\zeta}\left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right)+\frac{2\theta}{3\zeta^{3/2}}\left( \begin{array}{cc} 0 & z_2 \\ \bar{z}_2 & 0 \end{array} \right)+\frac{\theta}{6\zeta^2}\left( \begin{array}{cc} -4z_2\bar{z}_2-7z_1\bar{z}_1 & 0 \\ 0 & 4z_2\bar{z}_2+7z_1\bar{z}_1 \end{array} \right) \right. \\ &\left. +\frac{2\theta}{3\zeta^{5/2}}\left( \begin{array}{cc} 0 & -z_2\left( z_1\bar{z}_1+z_2\bar{z}_2 \right) \\ -\bar{z}_2\left(z_1\bar{z}_1+z_2\bar{z}_2 \right) & 0 \end{array} \right) \right\}-\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right). \end{align*} \section{K\"ahler structure and Bianchi identity}\label{sect5} In this section we discuss the K\"ahler condition on the metric derived from (anti-)self-dual two-forms of noncommutative $U(1)$ instantons. We will clarify this issue by illuminating the duality between the K\"ahler geometry and $U(1)$ gauge theory claimed in \cite{inov}. \subsection{K\"ahler geometry and $U(1)$ gauge theory} Let $M$ be a two-dimensional complex manifold with a K\"ahler metric \begin{equation}\label{c-metric} ds^2 = h_{i\bar{j}} (z, \overline{z}) dz^i d \overline{z}^{{j}}, \end{equation} where local complex coordinates are given by $z^i = {x}^{2i} + \mathrm{i}~ x^{2i-1}, \; (i= 1, 2)$. A K\"ahler manifold is described by a single function $K(z, \overline{z})$, so-called K\"ahler potential, defined by \begin{equation}\label{k-metric} h_{i\bar{j}} (z, \overline{z}) = \frac{\partial^2 K(z, \overline{z})}{\partial z^i \partial \overline{z}^{j}}. \end{equation} The K\"ahler potential is not unique but admits a K\"ahler transformation given by \begin{equation}\label{kah-gauge} K(z, \overline{z}) \to K(z, \overline{z}) + f(z) + \overline{f}(\overline{z}) \end{equation} where $f(z)$ and $\overline{f}(\overline{z})$ are arbitrary holomorphic and anti-holomorphic functions. Two K\"ahler potentials related by the K\"ahler gauge transformation \eq{kah-gauge} give rise to the same K\"ahler metric \eq{k-metric}. \begin{df}[K\"ahler form \cite{besse}] Given a K\"ahler metric (\ref{c-metric}), the K\"ahler form is a fundamental closed two-form defined by \begin{align} \Omega = \mathrm{i}~ h_{i\bar{j}} (z, \overline{z}) dz^i \wedge d \overline{z}^{j}. \label{ftwo-form} \end{align} \end{df} Note that the K\"ahler form \eq{ftwo-form} can be written as \begin{equation}\label{k-curvature} \Omega = d \mathcal{A} \qquad \mathrm{and} \qquad \mathcal{A} = \frac{\mathrm{i}}{2} (\overline{\partial} - \partial) K (z, \overline{z}) \end{equation} where the exterior differential operator is given by $d = \partial + \overline{\partial}$ with $\partial = dz^i \frac{\partial}{\partial z^i}$ and $ \overline{\partial} = d \overline{z}^{i} \frac{\partial}{\partial \overline{z}^{i}}$. Then the above K\"ahler transformation \eq{kah-gauge} corresponds to a gauge transformation for the one-form $\mathcal{A}$ given by \begin{equation}\label{k-gauge} \mathcal{A} \to \mathcal{A} + d \lambda \end{equation} where $\lambda = \frac{\mathrm{i}}{2} \big(\overline{f}(\overline{z}) - f(z) \big)$. This implies that the one-form $\mathcal{A}$ corresponds to $U(1)$ gauge fields or a connection of holomorphic line bundle. Note that the K\"ahler form $\Omega$ on a K\"ahler manifold $M$ is a nondegenerate, closed two-form. Therefore the K\"ahler form $\Omega$ is a symplectic two-form. This fact leads to the following proposition: \begin{prop}\label{kahler-symplectic} A K\"ahler manifold $(M, \Omega)$ is a symplectic manifold too although the reverse is not necessarily true. \end{prop} The K\"ahler condition enforces a specific analytic characterization of K\"ahler metrics: \begin{lem}\label{normal-kgh} $ds^2$ is K\"ahler if and only if it osculates to order 2 to the Euclidean metric everywhere. \end{lem} \noindent The proof of this lemma can be found in \cite{griffiths-harris} (Griffiths-Harris, p.~107). It means that the existence of normal holomorphic coordinates around each point of $M$ is equivalent to that of K\"ahler metrics. Let us consider an atlas $\{(U_\alpha, \varphi_\alpha)| \alpha \in I \}$ on the K\"ahler manifold $M$ and denote the K\"ahler form $\Omega$ restricted on a chart $(U_\alpha, \varphi_\alpha)$ as $\omega_\alpha \equiv \Omega|_{U_\alpha}$. According to the Lemma \ref{normal-kgh}, it is possible to write the local K\"ahler form as \begin{equation}\label{kahler-gauge} \omega_\alpha = B + F_\alpha, \end{equation} where $B$ is the K\"ahler form of $\mathbb{C}^2$. Since the two-form $F_\alpha$ must be closed due to the K\"ahler condition, it can be represented by $F_\alpha = d A_\alpha$. Using Eq. \eq{k-curvature} and $F_\alpha = \omega_\alpha - B$, the one-form $A_\alpha$ on $U_\alpha$ can be written as the form \begin{equation}\label{local-one-form} A_\alpha = \frac{\mathrm{i}}{2} (\overline{\partial} - \partial) \phi_\alpha (z, \overline{z}) \end{equation} where $\phi_\alpha (z, \overline{z}) = K_\alpha (z, \overline{z}) - K_0 (z, \overline{z})$ and $K_\alpha (z, \overline{z})$ is the K\"ahler potential on a local chart $U_\alpha$ and $K_0 (z, \overline{z}) = z^i \overline{z}^{\bar{i}}$ is the K\"ahler potential of $\mathbb{C}^2$. On an overlap $U_\alpha \bigcap U_\beta$, two one-forms $A_\alpha$ and $A_\beta$ can be glued using the freedom \eq{k-gauge} such that \begin{equation}\label{g-transf} A_\beta = A_\alpha + d \lambda_{\alpha\beta} \end{equation} where $\lambda_{\alpha\beta}(z, \overline{z})$ is a smooth function on the overlap $U_\alpha \bigcap U_\beta$. The gluing \eq{g-transf} on $U_\alpha \bigcap U_\beta$ is equal to the K\"ahler transformation \begin{equation}\label{holo-gauge} K_\beta (z, \overline{z}) = K_\alpha (z, \overline{z}) + f_{\alpha\beta} (z) + \overline{f}_{\alpha\beta} (\overline{z}) \end{equation} if $\lambda_{\alpha\beta}(z, \overline{z}) = \frac{\mathrm{i}}{2} \big(\overline{f}_{\alpha\beta}(\overline{z}) - f_{\alpha\beta}(z) \big)$. \begin{rem}\label{kahler-section}The K\"ahler transformation (\ref{holo-gauge}) implies the relation \[e^{K_\beta} = | e^{f_{\alpha\beta}}|^2 e^{K_\alpha}.\] So $e^{K(z, \overline{z})}$ is a section of a nontrivial line bundle over $M$. \end{rem} According to the proposition (\ref{kahler-symplectic}), the K\"ahler manifold $(M, h)$ is also a symplectic manifold $(M, \Omega)$. Therefore one can find a coordinate transformation $\varphi_\alpha \in \mathrm{Diff}(U_\alpha)$ on a local coordinate patch $U_\alpha$ such that $\varphi_\alpha^* (B + F) = B$ according to the famous Darboux theorem or Moser lemma in symplectic geometry \cite{symp-book}. In other words, the electromagnetic fields in the local K\"ahler form \eq{kahler-gauge} can always be eliminated by a local coordinate transformation. To be specific, the Darboux theorem ensures the existence of the local coordinate transformation $\varphi_\alpha: y^\mu \mapsto x^a = x^a (y), \; \mu, a = 1, \cdots, 4$, obeying \cite{hliu,sw-darboux} \begin{equation}\label{darboux-tr} \Big (B_{ab} + F_{ab} (x) \Big) \frac{\partial x^a}{\partial y^\mu} \frac{\partial x^b}{\partial y^\nu} = B_{\mu\nu}. \end{equation} Note that $B_{ab}$ and $B_{\mu\nu}$ are constant since they are coming from the K\"ahler form of $\mathbb{C}^2 \cong \mathbb{R}^4$ according to (\ref{kahler-gauge}) (see also the Lemma \ref{normal-kgh}). \begin{rem}\label{gravity-darboux}So far the coordinates $x^\mu$ have been commonly used for both gravity and field theory descriptions since it does not cause any confusion. However it is convenient to distinguish two kinds of coordinates $(x^a, y^\mu)$ appearing in the Darboux transformation (\ref{darboux-tr}). The so-called Darboux coordinates $y^\mu$ will be used for field theory description while the so-called covariant coordinates $x^a$ will be used for gravity description. \end{rem} \begin{df}[Poisson bracket \cite{symp-book}] Let $\theta := B^{-1} = \frac{1}{2} \theta^{\mu\nu} \frac{\partial}{\partial {y}^\mu} \bigwedge \frac{\partial}{\partial y^\nu} \in \Gamma(\Lambda^2 T\mathbb{R}^4)$ be a Poisson bivector. Then the Poisson bracket $\{~\cdot~, ~\cdot~\}: C^\infty (\mathbb{R}^4) \times C^\infty (\mathbb{R}^4) \to C^\infty (\mathbb{R}^4)$ is defined by $\{f, g\} = \theta(df, dg)$ for any smooth functions $f, g \in C^\infty (\mathbb{R}^4)$. \end{df} Since both sides of Eq. \eq{darboux-tr} are invertible, one can take its inverse and derive the following relation \begin{equation}\label{darboux-map1} \Theta^{ab} (x) := \left(\frac{1}{B + F (x) } \right)^{ab} = \theta^{\mu\nu} \frac{\partial x^a}{\partial y^\mu} \frac{\partial x^b}{\partial y^\nu} = \{ x^a (y), x^b (y)\} \end{equation} or \begin{equation}\label{darboux-map2} - \left(\frac{1}{1 + F\theta} B \right)_{ab} (x) = \{\phi_a (y), \phi_b (y)\} \end{equation} where $\phi_a (y) := B_{ab} x^b(y)$. Recall that we have started with a K\"ahler manifold with the metric \eq{c-metric} and applied the Darboux transformation to the local K\"ahler form \eq{kahler-gauge}. Now, in the description \eq{darboux-map1} or \eq{darboux-map2}, the curving of the K\"ahler manifold is described by local fluctuations of $U(1)$ gauge fields on the line bundle $L \to \mathbb{R}^4$. This becomes more manifest by taking the coordinate transformation in Eq. \eq{darboux-tr} as the form \begin{equation}\label{cov-phi} \phi_\mu (y) = p_\mu + a_\mu (y) \end{equation} and by calculating the Poisson bracket \begin{equation}\label{poisson-br} \{\phi_\mu (y), \phi_\nu (y) \} = -B_{\mu\nu} + \partial_\mu a_\nu(y) - \partial_\nu a_\mu (y) + \{a_\mu (y), a_\nu (y) \} \equiv -B_{\mu\nu} + f_{\mu\nu} (y). \end{equation} The functions $a_\mu (y)$ in the Darboux transformation \eq{cov-phi} will be regarded as gauge fields whose field strength is given by $f_{\mu\nu} (y) = \partial_\mu a_\nu(y) - \partial_\nu a_\mu (y) + \{a_\mu (y), a_\nu (y) \}$.\footnote{Here $a_\mu$ is a gauge field of a new $U(1)$ gauge symmetry with the Poisson structure rather than the original $U(1)$ gauge symmetry. From the original $U(1)$ gauge theory point of view, they are local sections of the line bundle $L \to \mathbb{R}^4$.} Since they respect the non-Abelian structure due to the underlying Poisson structure, they are different from ordinary $U(1)$ gauge fields $A_\mu (x)$ in \eq{local-one-form}, so they will be called ``symplectic" $U(1)$ gauge fields. Then Eq. \eq{darboux-map2} leads to the exact Seiberg-Witten map between commutative $U(1)$ gauge fields and symplectic $U(1)$ gauge fields \cite{sw-ncft,hliu,sw-darboux}: \begin{equation}\label{esw-map} f_{\mu\nu} (y) = \left(\frac{1}{1 + F \theta} F \right)_{\mu\nu} (x) \quad \mathrm{or} \quad F_{\mu\nu} (x) = \left(\frac{1}{1 - f \theta} f \right)_{\mu\nu} (y). \end{equation} Thus the following Lemma is conferred \cite{hsy-jhep09,hsy-ijmpa09,hsy-ijmpa15}: \begin{lem}\label{durboux-swmap} The Darboux transformation $\varphi_\alpha \in \mathrm{Diff}(U_\alpha)$ on a local coordinate patch $U_\alpha$ obeying $\varphi_\alpha^* (B + F) = B$ is equivalent to the Seiberg-Witten map between commutative $U(1)$ gauge fields and symplectic $U(1)$ gauge fields. \end{lem} The gauge theory description of K\"ahler gravity is realized by viewing a K\"ahler manifold as a phase space and its K\"ahler form as the symplectic two-form on the phase space \cite{inov}. This viewpoint naturally leads to a Poisson algebra $\mathfrak{P}=\{C^\infty(\mathbb{R}^4), \theta \}$ associated with the K\"ahler geometry we have started with. The underlying Poisson structure is inherited from the symplectic structure, i.e. $\theta = B^{-1} \in \Gamma(\Lambda^2 T\mathbb{R}^4)$, which is a bivector field called the Poisson tensor. \subsection{K\"ahler metric and Bianchi identity} Recall that the Seiberg-Witten map \eq{esw-map} has been derived from the local K\"ahler form \eq{kahler-gauge}. With the identification $\omega^\pm = f^\pm$ and using the map \eq{esw-map}, the metric $g^\pm$ in the Definition \ref{g-metric} can be written as \begin{equation}\label{sw-metg} g^\pm = 2 F^\pm \theta^\mp + E_4 \end{equation} which can be inverted to yield \begin{equation}\label{sw-comf} F^\pm = \frac{1}{2} (g^\pm - E_4) (\theta^\mp)^{-1}. \end{equation} Now we will prove the following proposition \cite{Lee:2012rb}. \begin{prop}\label{kal-bianchi Let $F$ be a two-form in (\ref{sw-comf}). Then the K\"ahler condition for the metric $g$ in (\ref{sw-metg}) is equivalent to the Bianchi identity for the $U(1)$ curvature $f$. \end{prop} \begin{pf} First note that the K\"ahler condition for the metric $g$ in (\ref{sw-metg}) is the closedness of the fundamental two-form $\omega = B + F$, which is equal to $d F = 0$. Consider the Jacobi identity \begin{equation}\label{x-jacobi} \{ x^a, \{ x^b, x^c \} \} + \{ x^b, \{ x^c, x^a \} \} + \{ x^c, \{ x^a, x^b \} \} = 0 \end{equation} that is equivalent to the Bianchi identity of symplectic $U(1)$ gauge fields \begin{equation}\label{g-bianchi} D_a f_{bc} + D_b f_{ca} + D_c f_{ab} = 0, \end{equation} where $D_a f_{bc} =\partial_a f_{bc}+ \left\{a_a,f_{bc} \right\}$. Using Eq. \eq{darboux-map1}, let us rewrite the Jacobi identity \eq{x-jacobi} as \begin{eqnarray}\label{x-bianchi} 0 &=& \{ x^a, \Theta^{bc} (x) \}_\theta + \{ x^b, \Theta^{ca} (x) \}_\theta + \{ x^c, \Theta^{ab} (x) \}_\theta \nonumber\\ &=& \theta^{\mu\nu} \left( \frac{\partial x^a}{\partial y^\mu} \frac{\partial \Theta^{bc} (x)} {\partial y^\nu} + \frac{\partial x^b}{\partial y^\mu} \frac{\partial \Theta^{ca} (x)}{\partial y^\nu} + \frac{\partial x^c}{\partial y^\mu} \frac{\partial \Theta^{ab} (x)}{\partial y^\nu} \right) \nonumber\\ &=& \theta^{\mu\nu} \left( \frac{\partial x^a}{\partial y^\mu} \frac{\partial x^d}{\partial y^\nu} \frac{\partial \Theta^{bc} (x)}{\partial x^d} + \frac{\partial x^b}{\partial y^\mu} \frac{\partial x^d}{\partial y^\nu} \frac{\partial \Theta^{ca} (x)}{\partial x^d} + \frac{\partial x^c}{\partial y^\mu} \frac{\partial x^d}{\partial y^\nu} \frac{\partial \Theta^{ab} (x)}{\partial x^d} \right) \nonumber\\ &=& \{ x^a, x^d \}_\theta \frac{\partial \Theta^{bc} (x)}{\partial x^d} + \{ x^b, x^d \}_\theta \frac{\partial \Theta^{ca} (x)}{\partial x^d} + \{ x^c, x^d \}_\theta \frac{\partial \Theta^{ab} (x)}{\partial x^d} \nonumber\\ &=& \Theta^{ad} (x) \frac{\partial \Theta^{bc} (x)}{\partial x^d} + \Theta^{bd} (x) \frac{\partial \Theta^{ca} (x)}{\partial x^d} + \Theta^{cd} (x) \frac{\partial \Theta^{ab} (x)}{\partial x^d} \nonumber \\ &=& - \Theta^{ad} \Theta^{be} \Theta^{fc} \left( \frac{\partial F_{ef} (x)} {\partial x^d} + \frac{\partial F_{fd} (x)}{\partial x^e} + \frac{\partial F_{de} (x)}{\partial x^f} \right). \end{eqnarray} Since $\Theta^{ab}$ is invertible, we get from \eq{x-bianchi} the Bianchi identity for the $U(1)$ curvature $F$, i.e., \begin{equation}\label{u1-bianchi} \frac{\partial F_{bc} (x)}{\partial x^a} + \frac{\partial F_{ca} (x)}{\partial x^b} + \frac{\partial F_{ab} (x)}{\partial x^c} = 0 \qquad \Longleftrightarrow \qquad dF=0. \end{equation} The same argument shows that the reverse is also true, i.e., if $dF=0$, the Bianchi identity (\ref{g-bianchi}) is deduced. This completes the proof. \qed\end{pf} If one introduces a new bivector $\Theta = \frac{1}{2} \Theta^{ab} (x) \frac{\partial}{\partial x^a} \bigwedge \frac{\partial}{\partial x^b} \in \Gamma(\Lambda^2 T\mathbb{R}^4)$ using the Poisson tensor in \eq{darboux-map1}, Eq. \eq{x-bianchi} shows that the Schouten-Nijenhuis bracket of the bivector $\Theta \in \Gamma(\Lambda^2 TN)$ identically vanishes, i.e., $[\Theta, \Theta]_{SN} = 0$ \cite{vaisman}. This means that the bivector $\Theta$ defines a new Poisson structure on $\mathbb{R}^4 \cong \mathbb{C}^2$. We thus see that the Bianchi identity for symplectic $U(1)$ gauge fields leads to the Bianchi identity of commutative $U(1)$ gauge fields and vice versa. Since the Bianchi identity \eq{u1-bianchi} can be understood as the K\"ahler condition for the local K\"ahler form \eq{kahler-gauge}, the Hermitian-Einstein metrics defined by $g = \omega \cdot J$ must be K\"ahler. Let us quantize the Poisson algebra $\mathfrak{P}$ to get a noncommutative algebra and a corresponding noncommutative $U(1)$ gauge theory. We apply the deformation quantization $\mathcal{Q}$ in Appendix A and define the quantization map for symplectic $U(1)$ gauge fields \cite{ly-jkps2018}: \begin{eqnarray}\label{q-map} && \mathcal{Q} (\phi_\mu) := \widehat{\phi}_\mu (y) = p_\mu + \widehat{A}_\mu (y), \nonumber \\ && \mathcal{Q} (\{\phi_\mu, \phi_\nu \}) := - i [\widehat{\phi}_\mu (y), \widehat{\phi}_\nu (y)] = -i \big(-B_{\mu\nu} + \widehat{F}_{\mu\nu} (y) \big), \end{eqnarray} where $\mathcal{Q} (f_{\mu\nu}) := \widehat{F}_{\mu\nu} (y) = \partial_\mu \widehat{A}_\nu (y) - \partial_\nu \widehat{A}_\mu (y) - i [\widehat{A}_\mu (y), \widehat{A}_\nu (y)]$ is the field strength of noncommutative $U(1)$ gauge fields $\widehat{A}_\mu (y) := \mathcal{Q} (a_\mu)$. After quantization, the symplectic $U(1)$ gauge fields map to noncommutative $U(1)$ gauge fields which contain infinitely many derivative corrections controlled by the noncommutative parameter $\theta^{\mu\nu}$. For example, the Seiberg-Witten map (\ref{esw-map}) receives noncommutative corrections and takes a non-local form whose exact form was conjectured in \cite{hliu}: \begin{equation}\label{esw-liu} F_{\mu\nu} (k) = \int d^4 y L_* \left[ \sqrt{1- \theta \widehat{F}} \left( \frac{1}{1-\widehat{F}\theta} \widehat{F} \right)_{\mu\nu} (y) W(y, C) \right] e^{ik \cdot y}, \end{equation} where $W(x, C)$ is a straight open Wilson line, the determinant and rational function of $\widehat{F}$ should be understood as a power series expansion, and $L_*$ denotes the integrations together with the path ordering procedure. The conjectured form (\ref{esw-liu}) was immediately proved in \cite{Okawa:2001mv,nc-openW}. In a commutative limit where the derivatives of the field strength can be ignored, the map (\ref{esw-liu}) is reduced to the second form in (\ref{esw-map}). An immediate question arises about the status of Proposition \ref{kal-bianchi} after (deformation) quantization. Let us state the result with the following proposition. \begin{prop}\label{cnc-bianchi} Let $F$ be a two-form in (\ref{esw-liu}). Then the closedness condition for the commutative $U(1)$ curvature $F$, $dF=0$, is equivalent to the Bianchi identity for the noncommutative $U(1)$ curvature $\widehat{F}$. \end{prop} This proposition was proved in \cite{Okawa:2001mv} by proving the conjecture by H. Liu. Theorem \ref{masspro} implies that the Hermitian metric $h^\pm$ in (\ref{masspro}) constructed by the identification $\omega^\pm = \widehat{F}^\pm$ still generates a Ricci-flat metric. Therefore Proposition \ref{kal-bianchi} may be lifted to noncommutative spaces although we do not have a rigorous proof yet. \section{Discussion}\label{sect6} We have shown that the K\"ahler geometry can be described by a $U(1)$ gauge theory on a symplectic manifold leading to a natural Poisson algebra associated with the K\"ahler geometry we have started with. Since the Poisson algebra $\mathfrak{P}$ defined by the Poisson bracket $\{f,g\} = \theta(df, dg)$ is mathematically the same as the one in Hamiltonian dynamics of particles, one can quantize the Poisson algebra in the exactly same way as quantum mechanics. Hence we have applied the deformation quantization to the Poisson algebra $\mathfrak{P} = (C^\infty (\mathbb{R}^4), \{-, -\})$. The quantization of the underlying Poisson algebra leads to a noncommutative $U(1)$ gauge theory which arguably describes a quantized K\"ahler geometry, as claimed in \cite{inov} and illuminated in \cite{ly-jkps2018}. Then we get a remarkable duality between K\"ahler gravity and noncommutative $U(1)$ gauge theory depicted by the following flow chart \cite{ly-jkps2018}: \begin{equation} \label{q-diag} \begin{array}[c]{ccc} \mathrm{K\ddot{a}hler~gravity}&\stackrel{\mathfrak{I}^{-1}_\epsilon}{\longrightarrow}& \mathrm{Symplectic~{\it U(1)}~gauge~theory }\\ {\mathcal{Q}}\downarrow\scriptstyle&&\downarrow{\mathcal{Q}}\scriptstyle\\ \mathrm{Quantized~K\ddot{a}hler~gravity} &\stackrel{\mathfrak{I}_\theta}{\longleftarrow}& \mathrm{Noncommutative~{\it U(1)}~gauge~theory } \end{array} \end{equation} Here $\mathcal{Q}: C^\infty (\mathbb{R}^4) \to \mathcal{A}_\theta$ means the quantization and $\mathfrak{I}$ means an isomorphism between two theories. In some sense $\mathfrak{I}$ corresponds to the gauge-gravity duality. It turns out \cite{ly-jkps2018} that it can be interpreted as the large $N$ duality too. Since symplectic $U(1)$ gauge theory is a commutative limit of noncommutative $U(1)$ gauge theory, we understand the classical isomorphism in \eq{q-diag} as $\mathfrak{I}_\epsilon = \mathfrak{I}_\theta|_{\varepsilon = |\theta| \to 0}$. The duality in \eq{q-diag} implies that a quantized K\"ahler gravity is isomorphically described by a noncommutative $U(1)$ gauge theory. Actually this relation was already observed in \cite{inov} in the context of topological strings probing K\"ahler manifolds where several nontrivial evidences have been analyzed to support the picture. In particular, the authors in \cite{inov} argue that noncommutative $U(1)$ gauge theory is the fundamental description of K\"ahler gravity at all scales including the Planck scale and provides a quantum gravity description such as quantum gravitational foams. The duality in \cite{inov} has been further clarified in \cite{neova-kap} by showing that it follows from the S-duality of the type IIB superstring. This duality, if any, suggests an important clue about how to quantize the K\"ahler gravity. Surprisingly, the correct variables for quantization are not metric fields but dynamical coordinates $x^a(y)$ and their quantization is defined in terms of $\alpha'$ rather than $\hbar$. So far, there is no well-established clue to quantize metric fields directly in terms of $\hbar$ in spite of impressive developments in loop quantum gravity. However, the picture in \eq{q-diag} suggests a completely new quantization scheme where quantum gravity is defined by quantizing spacetime itself in terms of $\alpha'$, leading to a dynamical noncommutative spacetime described by a noncommutative $U(1)$ gauge theory \cite{ly-jkps2018}. The duality relation in \eq{q-diag} may be more accessible with the corresponding relation for solutions of the self-duality equation, i.e., $U(1)$ instantons. Indeed it was shown in \cite{gi-u1-prl,gi-u1-plb,gi-u1-epl} that the commutative limit of noncommutative $U(1)$ instantons are equivalent to Calabi-Yau manifolds. {\bf Acknowledgments} \\ A.S. was supported in part by JSPS KAKENHI Grant Number 16K05138. The work of H.S.Y. was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MOE) (No. NRF-2015R1D1A1A01059710) and (No. NRF-2018R1D1A1B07050113).

bf70545735c91d0f916d979c311fac892d468725

\section*{Methods} \subsection*{Sample fabrication} The samples used in this work were deposited on Si substrates coated with $100$ nm of SiO$_{2}$. The deposition was done using DC magnetron sputtering at room temperature. The base pressure in the deposition chamber was $10^{-9}$ mbar. After deposition, Ti/Au contacts were deposited on top of the full sheet sample by a lift-off procedure using UV lithography (ma-N 415 photoresist). Lastly, $5 \ \mu$m wide and $90 \ \mu$m long wires were created using electron beam lithography to create a hard mask (ma-N 2410 photoresist) in combination with argon ion milling. \subsection*{Measurement techniques} The AOS was achieved using linearly polarized laser pulses with a central wavelength of $700$ nm and a pulse duration of $\approx 100$ fs. The pulse energy used throughout the work presented in the Letter was $\approx 12$ nJ, while the laser spot had a radius ($1/\mathrm{e}$ Gaussian pulse) of $\approx 20$ $\mu$m. The out-of-plane magnetization in the Hall crosses was measured using the AHE. To measure the AHE signal a small AC current (150 $\mu$A) is sent through the wire, and the resulting anomalous Hall voltage across the lateral legs is measured using a lock-in amplifier. In case of the Kerr microscope images a differential technique was used. In this technique first an image of the magnetic wire with the magnetization saturated in either the up or down direction was captured. This (background) image was then subtracted from the subsequent images in order to enhance the magnetic contrast. \subsection*{\large Supplementary Note 1: AOS as a function of the overlap between laser spot and Hall cross} The result presented in Fig.\ 1(b) of the main Letter shows full AOS in the Hall cross. This was measured with the centre of the laser spot aligned to the centre of the Hall cross. As was mentioned, the laser spot was larger than the cross, meaning that the area of the Pt/Co/Gd wire that is exposed (and switched) by the laser pulse is larger than the region probed by the Hall cross. It is known that depending on the laser fluence, a multidomain state can form at the centre of the (Gaussian shaped) laser spot, in which case only AOS is observed in an outer rim of the excited area \cite{Hadri2016}. In this section, it is verified that a single homogeneous domain was written in the Pt/Co/Gd wire by the laser pulse. In order to check that a homogeneous domain is written by the laser pulse, the measurement performed in Fig.\ 1 of the main Letter was repeated for different alignments of the laser spot with respect to the Hall cross. At each new alignment, the magnetization in the wire was first saturated using an externally applied field, whereafter the field was turned off and the Hall cross was exposed to a single laser pulse. When there is (partial) overlap between the Hall cross and the center area of the laser spot where the fluence $F\left(x,y\right)$ is above the AOS threshold fluence $F_{0}$, the magnetization in the Hall cross will be switched, which is recorded by a step in the AHE signal (similar as shown in Fig.\ 1(b) of the main Letter). The size of the AHE step is proportional to the area of the Hall cross that is switched by the laser. Figure\ \ref{Fig:AOSLineScan} shows the normalized AHE step size as a function of $x$ and $y$ position, where $\left(x,y\right) = \left(0,0\right)$ corresponds to the center of the (Gaussian) laser spot being aligned to the center of the Hall cross. The $x$ (black) and $y$ (red) scans are performed with $y = 0$ and $x = 0$, respectively, and each data point is an average of 7-8 subsequent measurements. When the laser spot is sufficiently far away from the Hall cross, i.e., for $\left|x\right|,\left|y\right|>15$ $\mu$m, the AHE step size is zero, meaning that there is no overlap between the Hall cross and the laser spot (or at least no sufficient overlap). Moving the laser spot closer to the Hall cross, i.e., decreasing $\left|x\right|$ or $\left|y\right|$, the AHE step size increases towards saturation at $\left|x\right|,\left|y\right|\approx5$ $\mu$m. The increase in AHE step size corresponds to the center part of the laser spot, where $F\left(x,y\right) \geq F_{0}$, moving into the Hall cross area. This is illustrated by the left and right cartoons in the figure in which the laser spot area with $F\left(x,y\right) \geq F_{0}$ (red dotted circle) overlaps $\approx30\%$ of the Hall cross area. For $\left|x\right|,\left|y\right|<5$ $\mu$m the AHE step size is constant and equal to saturation. The presence of these plateaus in both $x$ and $y$ scans demonstrates that for the full area of the laser spot where $F\left(x,y\right) \geq F_{0}$ there is full AOS. In other words, this means that indeed a homogeneous domain was written in the Pt/Co/Gd wire by the laser pulse. As a side note, the full width at half maximum of the curves are equal to the size of the written magnetic domain along the $x$ and $y$ directions, showing a domain size of $\approx 20$ $\mu$m and a slightly elliptically shaped laser spot, which was verified using wide field Kerr microscopy on a full-sheet sample. \begin{figure} \includegraphics[scale=0.35]{AOSLineScan.pdf} \caption{\textbf{AOS as a function of the overlap between laser spot and Hall cross.} Normalized step size of the AHE signal as a function of the $x$ and $y$ position. The step size in the AHE signal measured across the Hall cross is a measure of the area of the Hall cross that is switched by the laser pulse. $\left(x,y\right) = \left(0,0\right)$ corresponds to the center of the laser spot being aligned to the center of the Hall cross.} \label{Fig:AOSLineScan} \end{figure} Lastly, it is known that when the pulse energy is sufficiently increased, the fluence at the center of the laser spot can be increased to a value above a second threshold fluence. At this fluence the lattice temperature is heated above the Curie temperature, resulting in the formation of a multidomain state on cool down. Such a multidomain state at the center of the written domain would show up as a dip in the AHE step size around $\left(x,y\right) = \left(0,0\right)$ in the measurement presented in Fig.\ \ref{Fig:AOSLineScan}. Such a dip was indeed observed when repeating the measurement of Fig.\ \ref{Fig:AOSLineScan} with increasing pulse energy (not shown), verifying that inhomogineties in the written domain can indeed be measured, and thus confirming that for the laser pulses used in Fig.\ \ref{Fig:AOSLineScan} and the main Letter a homogeneous domain was written. \subsection*{\large Supplementary Note 2: DW pinning and Ga$^{+}$ irradiation} In the main Letter it was mentioned that the legs of some of the Hall crosses were irradiated with Ga$^{+}$ ions in order to prevent pinning at the entrance of the cross. The pinning of the DW at the entrance of a non-irradiated Hall cross in a typical on-the-fly AOS measurement as performed in the main Letter is demonstrated in the measurement discussed in Supplementary Note 3 (Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(c)). A visual presentation using a Kerr microscope is presented in the top row of Fig.\ \ref{Fig:GaIrradiation}. In this figure a down domain (dark) in an otherwise up (light) magnetized Pt/Co/Gd wire is located between two Hall crosses. Using three current pulses of alternating direction (see figure) it can be seen that the domain tries to move along the current direction, but gets pinned at the entrance of the cross it is moving towards. Moreover, it can be seen that the end points of the DW get pinned at the start of the legs, while the center of the DW gets pushed into the cross, which will be visible in the AHE signal (as demonstrated in Supplementary Note 3). The pinning of the DW at the entrance of the Hall cross happens due to the fact that the DW length has to increase in order to pass through the cross \cite{Ravelosona2005}. One way to overcome this problem, while still being able to use the legs for the AHE measurement, is to magnetically `cut-off' the legs using a technique called magnetic etching \cite{Franken2012}. In this technique a magnetic sample with perpendicular magnetic anisotropy (PMA) is exposed to Ga$^{+}$ ion irradiation with a relatively high dose ($40$ $\mu$C/cm$^{2}$ in this work), which is enough to destroy the magnetic anisotropy, but not enough to physically remove a significant amount of material. Using this technique the PMA in (part of) the legs is destroyed, causing it to become in-plane magnetized (or even paramagnetic), while the legs keep their conductive properties needed for the AHE measurement. In this way the AHE measurement can still be performed to measure the magnetization in the Hall cross area, while the legs become invisible for the DW. As a result, the DW will not be pinned at the entrance of the Hall cross since it does no longer need to increase in length when passing through the cross. \begin{figure} \includegraphics[scale=0.5]{GaIrradiation.pdf} \caption{\textbf{DW pinning at a Hall cross and the effect of Ga$^{+}$ ion irradiation.} Top row: A Pt/Co/Gd wire with a down domain (dark) in an otherwise up (light) magnetized wire that is located between two Hall crosses. The three figures show the domain after three current pulses of alternating direction (direction indicated in figures). Bottom row: A Hall cross on a $2$ $\mu$m Pt/Co/Gd wire with a DW initially located below the cross. The red squares in the left figure represent the regions that are exposed to the Ga$^{+}$ ions. The figures show three snapshots of the DW moving through the wire by the SHE using a DC current (direction indicated in left figure).} \label{Fig:GaIrradiation} \end{figure} The effect of the Ga$^{+}$ ion irradiation on the DW propagation through the Hall cross is verified in the bottom row of Fig.\ \ref{Fig:GaIrradiation}. In these figures a Hall cross on a $2$ $\mu$m Pt/Co/Gd wire is shown, where the red squares in the left figure represent the regions that are exposed to the Ga$^{+}$ ions. Starting with a DW below the Hall cross in the left figure, three snapshots are shown of the DW moving through the wire by the SHE using a DC current. It is seen that with the legs being magnetically cut-off, the DW indeed is able to move past the Hall cross without getting pinned. \subsection*{\large Supplementary Note 3: On-the-fly AOS with different laser spot to (non-irradiated) Hall cross alignments} A proof-of-concept measurement was presented in Fig.\ 2 of the main Letter, demonstrating on-the-fly single-pulse AOS and simultaneous SHE driven motion of magnetic domains in a single racetrack. In that measurement, the laser spot was aligned to the right side of the first Hall cross to prevent pinning of the DW's (see Fig.\ 2(a) of the main Letter), while a small overlap between laser pulse and the first Hall cross was maintained in order to verify if and when a domain was written. In this section, two similar measurement are presented; (i) with the laser spot centered at the center of the first Hall cross, clearly demonstrating the pinning of the DW at the entrance of the non-irradiated Hall cross, and (ii) with the laser spot aligned completely in between the two Hall crosses. \begin{figure*} \includegraphics[scale=0.33]{ProofOfPrincipleOnCross.pdf} \caption{\textbf{On-the-fly AOS with the laser spot centered on the non-irradiated Hall cross.} (a) Illustration of the on-the-fly AOS measurement performed with the laser spot aligned to the center of the first Hall cross. The AHE signal in the Hall crosses is measured using lock-in amplifiers LI $1$ and LI $2$. The red dotted circle illustrates the region exposed by the laser pulse, and the red blocks indicate the regions exposed to Ga$^{+}$ ion irradiation. Figures I to IIII show different snapshots of the magnetization in the wire during a two-pulse cycle. (b,c) Measurement of the normalized AHE signal as a function of time in the first (top) and second (bottom) Hall cross while at the same time the first Hall cross is exposed to a train of linearly polarized laser pulses ($\approx 100$ fs) at a laser-pulse repetition rate of 0.1 Hz. The measurement presented in (b) was performed without any current sent through the wire, while a current of $+5.5$ mA was sent through the wire during the measurement presented in (c).} \label{Fig:ProofOfPrincipleOnCross} \end{figure*} An illustration of the on-the-fly AOS measurement performed with the laser spot aligned to the center of the first Hall cross is shown in Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(a). The expected magnetic behavior is as follows; (I) With the magnetization in the wire initially saturated, the first (left) cross is exposed to a single laser pulse (red dotted circle), writing a magnetic domain that is larger than the Hall cross, thus creating a DW at either side of the cross. (II) Due to the SHE originating from the DC current that is continuously sent through the wire, both DW's (and thereby the domain) move along the current direction towards the second Hall cross. The DW written in between the two Hall crosses will reach the second cross, where it can pass the cross since its legs are magnetically cut-off using Ga$^{+}$ ion irradiation. The legs of the first Hall cross, however, have not been irradiated. Therefore, the DW written to the left of the first cross will get pinned at the entrance of the cross. The center of the DW is pushed into the cross by the SHE, and will be measurable by the AHE. (III) The next laser pulse will toggle the magnetization in the exposed area, creating two new DW's and reversing the polarity of the DW located in the first Hall cross. (IIII) The newly created DW's are moved along the wire by the SHE and annihilate with the previously written DW's, leaving the wire in the saturated state. This process will then repeat itself at the next laser pulse. First, a test measurement is performed without a DC current being sent through the wire, and using a repetition rate of 0.1 Hz to clearly see the effect of the single laser pulses. The result is presented in Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(b), showing the normalized AHE signal of both crosses as a function of time. The top graph shows the AHE signal of the first cross, demonstrating clear and full single-pulse AOS of the magnetization in this cross, similar as shown in Fig.\ 1(b) of the main Letter. The bottom graph shows the AHE signal of the second Hall cross. As can be seen by the constant AHE signal at the initial saturation value, there is no effect of the laser pulses on the magnetization in the second cross, which is expected since there is no current sent through the wire, and thus no DW motion. Figure \ref{Fig:ProofOfPrincipleOnCross}(c) shows the result of a measurement with a DC current of $+5.5$ mA sent through the wire. Looking at the AHE signal of the first Hall cross (top graph), the signal looks much different than the signal measured without a DC current (Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(b)). Before explaining the observed behavior in more detail, it is noted that the times at which the magnetic domains are written at the first Hall cross are still clearly visible by the sudden steps in the AHE signal (red dotted lines). Looking at the AHE signal in the second Hall cross (bottom graph), it can be seen that shortly after the domains are written, the magnetization in the second cross switches its direction, toggling up and down after each subsequent laser pulse. The toggling behavior corresponds to the DW that is written to the right of the first Hall cross that is transported along the wire by the SHE and passes through the second Hall cross. As shown in Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(a), this DW alternates between an up-down and down-up DW on subsequent laser pulses, leaving the magnetization in the cross in the up and down state after the DW has passed, respectively. Coming back to the AHE signal measured in the first Hall cross (top graph Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(c)), there are two observations that can be made. Firstly, the step in the AHE signal varies in size and has the same sign for every laser pulse. Secondly, the AHE signal after every even number of pulses is not equal to the initial saturation value of $+1$. Both observations will be discussed in the following. \begin{figure*} \includegraphics[scale=0.5]{ProofOfPrincipleNoOverlap.pdf} \caption{\textbf{On-the-fly AOS with the laser spot aligned in between the two Hall crosses.} (a) Illustration of the on-the-fly AOS measurement performed with the laser spot aligned in between the two Hall crosses of the Pt/Co/Gd wire. The AHE signal in the Hall crosses is measured using lock-in amplifiers LI $1$ and LI $2$. The red dotted circle illustrates the region exposed by the laser pulse, and the red blocks indicate the regions exposed to Ga$^{+}$ ion irradiation. (b) Measurement of the (normalized) AHE signal as a function of time in the first (top) and second (bottom) Hall cross while at the same time the wire is exposed to a train of linearly polarized laser pulses ($\approx 100$ fs) at a laser-pulse repetition rate of 0.2 Hz and a DC current of $+5.5$ mA is sent through the wire.} \label{Fig:ProofOfPrincipleNoOverlap} \end{figure*} It can be seen that the very first laser pulse switches the magnetization from the initial saturated up ($+1$) direction to the down ($-1$) direction. In contrast to the case without the DC current, the AHE signal immediately rises again towards $\approx +0.5$. This corresponds to the DW written to the left of the first Hall cross being transported along the wire by the SHE and getting pinned at the entrance of the cross (Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(a).II). The center of the DW is being pushed inside the cross by the SHE, causing about $75\%$ of the Hall cross area being switched back to the initial saturation direction. After this initial (fast) response, the center of the DW keeps getting pushed further inside the cross at a slower rate, causing a further (slow) increase of the AHE signal. When the second pulse hits the sample, most of the Hall cross area ($\approx 85\%$) is switched back to the initial saturation direction (Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(a).III). Therefore, the second step in the AHE signal has the same sign as for the first pulse, however, the size of the step is smaller since it only corresponds to the net switched magnetization. This measurement clearly demonstrates the pinning of the DW at the entrance of the Hall cross when the legs are not magnetically cut-off using Ga$^{+}$ ion irradiation. Taking a closer look at the AHE signal in the first Hall cross just before the third pulse arrives, it can be seen that the signal is not back to the saturation value of $+1$, but is close to $+0.9$ which corresponds to $95\%$ of the first Hall cross having its magnetization up. Looking back at Fig.\ \ref{Fig:ProofOfPrincipleOnCross}(a), it was expected that the DW written on the left side of the cross by the second pulse (or every even amount of pulses) would annihilate with the DW pinned in the cross (written by the previous pulse), leaving the Hall cross in the initial saturated state. This discrepancy is believed to be the result of the center of the DW that is being pushed into the cross not reaching a steady position before the next laser pulse arrives. This could result in the formation of thin rings with alternating magnetization direction in the first Hall cross that slowly move along with the current. The presence of these rings cause the AHE signal not to reach the saturation value of $+1$ after every even amount of pulses. Lastly, the same on-the-fly AOS measurement is performed, but now with the laser spot aligned completely in between the two Hall crosses, as shown in Fig.\ \ref{Fig:ProofOfPrincipleNoOverlap}(a), and a laser-pulse repetition rate of 0.2 Hz. The AHE signal of both Hall crosses as a function of time is shown in Fig.\ \ref{Fig:ProofOfPrincipleNoOverlap}(b). The magnetic domains are written to the right side of the first Hall cross, and they are transported along the current direction. This means that none of the domains or DW's reach the first Hall cross, resulting in the observed constant saturation value of the AHE signal in the first Hall cross (top graph). Looking at the AHE signal of the second Hall cross (bottom graph), it can be seen that all the optically written domains pass the second cross, as was also demonstrated in the measurement presented in Fig.\ 2(b) of the main Letter. Also seen in both measurements is the variation in the width of the domains when they pass the second cross (proportional to the time between down and up switch in the AHE signal). This shows that the variation in domain width measured in Fig.\ 2(b) of the main Letter is not dominated by the slight overlap between laser spot and first Hall cross, but is more likely to be the result of random pinning along the wire, as was already stated in the main Letter. \end{document}

5b6ca1a6eceb9827dc742117edf991a75b48ec2c

\section{Introduction} \label{sec:intro} In the query, or black-box, model of computation, an algorithm aims to evaluate a known Boolean function $f : \{0,1\}^n \to \{0,1\}$ on an unknown input $x \in \{0,1\}^n$ by reading as few bits of $x$ as possible. One of the most basic questions one can ask about query complexity, or indeed any complexity measure of Boolean functions, is how it behaves under \emph{composition}. Namely, given functions $f$ and $g$, and a method of combining these functions to produce a new function $h$, how does the query complexity of $h$ depend on the complexities of the constituent functions $f$ and $g$? The simplest method for combining functions is \emph{block composition}, where the inputs to $f$ are obtained by applying the function $g$ to independent sets of variables. That is, if $f : \{0,1\}^m \to \{0,1\}$ and $g : \{0,1\}^k \to \{0,1\}$, then the block composition $(f \circ g): \{0,1\}^{m \cdot k} \to \{0,1\}$ is defined by $(f \circ g)(x_1, \dots, x_m) = f(g(x_1), \dots, g(x_m))$ where each $x_i$ is a $k$-bit string. In most reasonable models of computation, one can evaluate $f \circ g$ by running an algorithm for $f$, and using an algorithm for $g$ to compute the inputs to $f$ as needed. Thus, the query complexity of $f \circ g$ is at most the product of the complexities of $f$ and $g$.\footnote{In some ``reasonable models,'' such as those with bounded error, one must take care to ensure that errors in computing each copy of $g$ do not propagate, but we elide these issues for this introduction. Addressing this concern typically adds at most a logarithmic overhead.} For many query models, including those capturing deterministic and quantum computation, this is known to be tight. In particular, letting $Q(f)$ denote the bounded-error quantum query complexity of a function $f$, it is known that $Q(f \circ g) = \Theta(Q(f) \cdot Q(g))$ for all Boolean functions $f$ and $g$~\cite{negativeweights, Rei11}. This result has the flavor of a direct sum theorem: When computing many copies of the function $g$ (in this case, as many as are needed to generate the necessary inputs to $f$), one cannot do better than just computing each copy independently. \subsection{Quantum algorithms for shared-input compositions} While we have a complete understanding of the behavior of quantum query complexity under block composition, little is known for more general compositions. What is the quantum query complexity of a composed function where inputs to $f$ are generated by applying $g$ to overlapping sets of variables? We call these more general compositions \emph{shared-input compositions}. Not only does answering this question serve as a natural next step for improving our understanding of quantum query complexity, but it may lead to more unified algorithms and lower bounds for specific functions of interest in quantum computing. Many of the functions that have played an influential role in the study of quantum query complexity can be naturally expressed as compositions of simple functions with shared inputs, including $k$-distinctness, $k$-sum, surjectivity, triangle finding, and graph collision. \begin{figure}[ht \centering \begin{tikzpicture}[x=4em,y=4em,scale=1] \node at (2,2.75) [] (h) {$h(x)$}; \node at (2,2) [shape=circle, draw] (f) {$f$}; \node at (-1.25,0) [] (x1) {\vphantom{I}$x_1$}; \node at (-0.75,0) [] (notx1) {\vphantom{I}$\overline{x_1}$}; \node at (0.25,0) [] (x2) {\vphantom{I}$x_2$}; \node at (0.75,0) [] (notx2) {\vphantom{I}$\overline{x_2}$}; \node at (1.75,0) [] (x3) {\vphantom{I}$x_3$}; \node at (2.25,0) [] (notx3) {\vphantom{I}$\overline{x_3}$}; \node at (3.25,0) [] (x4) {\vphantom{I}$x_4$}; \node at (3.75,0) [] (notx4) {\vphantom{I}$\overline{x_4}$}; \node at (4.75,0) [] (x5) {\vphantom{I}$x_5$}; \node at (5.25,0) [] (notx5) {\vphantom{I}$\overline{x_5}$}; \node at (-0.5,1) [shape=circle, draw] (g1) {$\wedge$}; \node at (0.5,1) [shape=circle, draw] (g2) {$\wedge$}; \node at (1.5,1) [shape=circle, draw] (g3) {$\wedge$}; \node at (2.5,1) [shape=circle, draw] (g4) {$\wedge$}; \node at (3.5,1) [shape=circle, draw] (g5) {$\wedge$}; \node at (4.5,1) [shape=circle, draw] (g6) {$\wedge$}; \draw (x1) -- (g1); \draw (notx1) -- (g2); \draw (x2) -- (g1); \draw (notx2) -- (g2); \draw (notx2) -- (g3); \draw (notx2) -- (g4); \draw (notx3) -- (g3); \draw (x3) -- (g2); \draw (x3) -- (g4); \draw (notx3) -- (g6); \draw (notx4) -- (g6); \draw (x4) -- (g4); \draw (x4) -- (g5); \draw (x5) -- (g4); \draw (notx5) -- (g5); \draw (notx5) -- (g6); \draw (g1) -- (f); \draw (g2) -- (f); \draw (g3) -- (f); \draw (g4) -- (f); \draw (g5) -- (f); \draw (g6) -- (f); \draw (h) -- (f); \end{tikzpicture} \caption{\label{fig:composition}A depth-2 circuit $h:\{0,1\}^5\to\{0,1\}$ where the top gate is a function $f:\{0,1\}^6 \to \{0,1\}$ and the bottom level gates are $\AND$ gates on a subset of the input bits and their negations. More generally, we consider $h:\{0,1\}^n \to \{0,1\}$, with top gate $f:\{0,1\}^m \to \{0,1\}$.} \end{figure} In this work, we study shared-input compositions between an arbitrary function $f$ and the function $g = \AND$. If $f : \{0,1\}^m \to \{0,1\}$, then we let $h : \{0,1\}^n \to \{0,1\}$ be any function obtained by generating each input to $f$ as an $\AND$ over some subset of (possibly negated) variables from $x_1, \dots, x_n$, as depicted in \fig{composition}. Of course, one can compute the function $h$ by ignoring the fact that the $\AND$ gates depend on shared inputs, and instead regard each gate as depending on its own set of copies of the input variables. Using the quantum query upper bound for block compositions, together with the fact that $Q(\AND_n) = \Theta(\sqrt{n})$~\cite{Gro96,BBBV97}, one obtains \begin{equation} Q(h) = O(Q(f) \cdot Q(\AND_n)) = O(Q(f) \cdot \sqrt{n}). \end{equation} Observe that this bound on $Q(h)$ is non-trivial only if $Q(f) \ll \sqrt{n}$. A priori, one may conjecture that this bound is tight in the worst case for shared-input compositions. After all, if the variables overlap in some completely arbitrary way with no structure, it is unclear from the perspective of an algorithm designer how to use the values of already-computed $\AND$ gates to reduce the number of queries needed to compute further $\AND$ gates. It might even be the case that every pair of $\AND$ gates shares very few common input bits, suggesting that evaluating one $\AND$ gate yields almost no information about the output of any other $\AND$ gate. This intuition even suggests a path for proving a matching lower bound: Using a random wiring pattern, combinatorial designs, etc., construct the set of inputs to each $\AND$ gate so that evaluating any particular gate leaks almost no useful information that could be helpful in evaluating the other $\AND$ gates. In this work, we show that this intuition is wrong: the overlapping structure of the $\AND$ gates can \emph{always} be exploited algorithmically (so long as $Q(f) \ll n$). \para{Results.} Our main result shows that a shared-input composition between a function $f$ and the $\AND$ function always has substantially lower quantum query complexity than the block composition $f \circ \AND_n$. Specifically, instead of having quantum query complexity which is the product $Q(f) \cdot \sqrt{n}$, a shared-input composition has quantum query complexity which is, up to logarithmic factors, the geometric mean $\sqrt{Q(f) \cdot n}$ between $Q(f)$ and the number of input variables $n$. This bound is nontrivial whenever $Q(f)$ is significantly smaller than $n$. \begin{restatable}{theorem}{quantum} \label{thm:quantum} Let $h:\{0,1\}^n \to \{0,1\}$ be computed by a depth-2 circuit where the top gate is a function $f:\{0,1\}^m \to \{0,1\}$ and the bottom level gates are $\AND$ gates on a subset of the input bits and their negations (as depicted in \fig{composition}). Then we have \begin{equation} Q(h) = O\Bigl(\sqrt{Q(f) \cdot n} \cdot \log^2(mn)\Bigr). \end{equation} \end{restatable} Note that \thm{quantum} is nearly tight for every possible value of $Q(f) \in [n]$. For a parameter $t \le n$, consider the block composition (i.e., the composition with disjoint inputs) $\PARITY_t \circ \AND_{n / t}$. Since $Q(\PARITY_t) = \lceil t / 2 \rceil$~\cite{beals}, this function has quantum query complexity \begin{equation} Q\(\PARITY_t \circ \AND_{n / t}\)=\Theta\left(t \cdot \sqrt{n/t}\right) = \Theta\left(\sqrt{Q(\PARITY_t) \cdot n}\right), \end{equation} matching the upper bound provided by \thm{quantum} up to log factors. This shows that \thm{quantum} cannot be significantly improved in general. The proof of \thm{quantum} makes use of an optimal quantum algorithm for computing $f$ and Grover's search algorithm for evaluating $\AND$ gates. Surprisingly, it uses no other tools from quantum computing. The core of the argument is entirely classical, relying on a recursive gate and wire-elimination argument for evaluating $\AND$ gates with overlapping inputs. At a high level, the algorithm in \thm{quantum} works as follows. The overall goal is to query enough input bits such that the resulting circuit is simple enough to apply the composition upper bound $Q(f \circ g) = O(Q(f)Q(g))$. To apply this upper bound and obtain the claimed upper bound in \thm{quantum}, we require $Q(g)$ to be $O(\sqrt{n/Q(f)})$. Since $g$ is just an $\AND$ gate on some subset of inputs, this means we want the fan-in of each $\AND$ gate in our circuit to be $O(n/Q(f))$. If we call $\AND$ gates with fan-in $\omega(n/Q(f))$ ``high fan-in'' gates, then the goal is to eliminate all high fan-in gates. Our algorithm achieves this by judiciously querying input bits that would eliminate a large number of high fan-in gates if they were set to 0. Besides the line of work on the quantum query complexity of block compositions, our result is also closely related to work of Childs, Kimmel, and Kothari~\cite{CKK12} on read-many formulas. Childs et al. showed that any formula on $n$ inputs consisting of $G$ gates from the de Morgan basis $\{\AND, \OR, \NOT\}$ can be evaluated using $O(G^{1/4} \cdot \sqrt{n})$ quantum queries. In the special case of DNF formulas, our result coincides with theirs by taking the top function $f$ to be the $\OR$ function. However, even in this special case, the result of Childs et al. makes critical use of the top function being $\OR$. Specifically, their result uses the fact that the quantum query complexity of the $\OR$ function is the square root of its formula size. Our result, on the other hand, applies without making any assumptions on the top function $f$. This level of generality is needed when using \thm{quantum} to understand \emph{circuits} (rather than just formulas) of depth 3 and higher, as discussed in \sec{intro-lc}. \subsection{Approximate degree of shared-input compositions} We also study shared-input compositions under the related notion of approximate degree. For a Boolean function $f : \{0,1\}^n \to \{0,1\}$, an \emph{$\varepsilon$-approximating polynomial} for $f$ is a real polynomial $p : \{0,1\}^n \to {\mathbb{R}}$ such that $|p(x) - f(x)| \le \varepsilon$ for all $x \in \{0,1\}^n$. The \emph{$\varepsilon$-approximate degree} of $f$, denoted $\deg_\varepsilon(f)$, is the least degree among all $\varepsilon$-approximating polynomials for $f$. We use the term \emph{approximate degree} without qualification to refer to choice $\varepsilon = 1/3$, and denote it $\adeg(f) = \deg_{1/3}(f)$. A fundamental observation due to Beals et al.~\cite{beals} is that any $T$-query quantum algorithm for computing a function $f$ implicitly defines a degree-$2T$ approximating polynomial for $f$. Thus, $\adeg(f) \le 2Q(f)$. This relationship has led to a number of successes in proving quantum query complexity lower bounds via approximate degree lower bounds, constituting a technique known as the polynomial method in quantum computing. Conversely, quantum algorithms are powerful tools for establishing the existence of low-degree approximating polynomials that are needed in other applications to theoretical computer science. For example, the deep result that every de Morgan formula of size $s$ has quantum query complexity, and hence approximate degree, $O(\sqrt{s})$~\cite{FGG08,CCJY07,ACRSZ10,Rei11} underlies the fastest known algorithm for agnostically learning formulas~\cite{kkms, Rei11} (See \sec{intro-learning} and \sec{agnostic} for details on this application). It has also played a major role in the proofs of the strongest formula and graph complexity lower bounds for explicit functions~\cite{tal1}. \para{Results.} We complement our result on the quantum query complexity of shared-input compositions with an analogous result for approximate degree. \begin{restatable}{theorem}{composition} \label{thm:composition} Let $h:\{0,1\}^n\to \{0,1\}$ be computed by a depth-2 circuit where the top gate is a function $f:\{0,1\}^m \to \{0,1\}$ and the bottom level gates are $\AND$ gates on a subset of the input bits and their negations (as depicted in \fig{composition}). Then \begin{equation} \deg_\varepsilon(h) =O\(\sqrt{\deg_\varepsilon(f)\cdot n \cdot \log m} + \sqrt{n\log(1/\varepsilon)}\).\label{eq:composition} \end{equation} In particular, $\adeg(h) = O\(\sqrt{\adeg(f) \cdot n\log m}\)$. \end{restatable} Note that our result for approximate degree is incomparable with \thm{quantum}, even for bounded error, since both sides of the equation include the complexity measure under consideration. Like \thm{quantum}, \thm{composition} can be shown to be tight by considering the block composition of $\PARITY$ with $\AND$, since $\adeg(\PARITY_{t}\circ \AND_{n/t})=\Theta\(\sqrt{\adeg(\PARITY_t)\cdot n}\)$ \cite{sherstovrobust, comm5}. Our proof of \thm{composition} abstracts and generalizes a technique introduced by Sherstov~\cite{algorithmicpolys}, who very recently proved an $O(n^{3/4})$ upper bound on the approximate degree of an important depth-3 circuit of nearly quadratic size called Surjectivity~\cite{algorithmicpolys}. Despite the similarity between \thm{composition} and \thm{quantum}, and the close connection between approximating polynomials and quantum algorithms, the proof of \thm{composition} is completely different from \Cref{thm:quantum}, making crucial use of properties of polynomials that do not hold for quantum algorithms.\footnote{Any analysis capable of yielding a sublinear upper bound on the approximate degree of Surjectivity requires moving beyond quantum algorithms, as its quantum query complexity is known to be ${\Omega}(n)$~\cite{beame,sherstov15}.} In our opinion, this feature of the proof of \thm{composition} makes \thm{quantum} for quantum algorithms even more surprising. We remark that a different proof of the $O(n^{3/4})$ upper bound for the approximate degree of Surjectivity was discovered in \cite{BKT18}, who also showed a matching lower bound. It is also possible to prove \thm{composition} by generalizing the techniques developed in that work, but the techniques of \cite{algorithmicpolys} lead to a shorter and cleaner analysis. \subsection{Application: Evaluating and approximating linear-size {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits} \label{sec:intro-lc} The circuit class {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace consists of constant-depth, polynomial-size circuits over the de Morgan basis $\{\AND, \OR, \NOT\}$ with unbounded fan-in gates. The full class {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace is known to contain very hard functions from the standpoint of both quantum query complexity and approximate degree. The aforementioned Surjectivity function is in depth-3 {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace and has quantum query complexity $\Omega(n)$~\cite{beame,sherstov15}, while for every positive constant $\delta > 0$, there exists a depth-$O(\log(1/\delta))$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit with approximate degree $\Omega(n^{1-\delta})$~\cite{BT17}. Nevertheless, {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace contains a number of interesting subclasses for which nontrivial quantum query and approximate degree upper bounds might still hold. Here, we discuss applications of our composition theorem to understanding the subclass {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace, consisting of {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits of linear size. The class {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace is one of the most interesting subclasses of {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace. It has been studied by many authors in various complexity-theoretic contexts, ranging from logical characterizations~\cite{KLPT06} to faster-than-brute-force satisfiability algorithms~\cite{calabroetal,santhanam2012limits}. {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace turns out to be a surprisingly powerful class. For example, the $k$-threshold function that asks if the input has Hamming weight greater than $k$ is clearly in {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace for constant $k$, by computing the $\OR$ of all $\binom{n}{k}$ possible certificates. But this yields a circuit of size $O(n^k)$, which one might conjecture is optimal. However, it turns out that $k$-threshold is in {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace even when $k$ is as large as $\polylog(n)$~\cite{RW91}. Another surprising fact is that every regular language in {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace can be computed by an {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit of almost linear size (e.g., size $O(n\log^* n)$ suffices)~\cite{Kou09}. By recursively applying \thm{quantum}, we obtain the following sublinear upper bound on the quantum query complexity of depth-$d$ {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits, denoted by $\LCd$: \begin{restatable}{theorem}{quantumLC} \label{thm:quantumLC} For all constants $d\geq 0$ and all functions $h:\{0,1\}^n \to \{0,1\}$ in $\LCd$, we have $Q(h) = \widetilde{O}(n^{1-{2^{-d}}})$. \end{restatable} Our upper bound is nearly tight for every depth $d$, as shown in~\cite{CKK12}. \begin{restatable}[Childs, Kimmel, and Kothari]{theorem}{quantumlower} \label{thm:quantumlower} For all constants $d\geq 0$, there exists a function $h:\{0,1\}^n \to \{0,1\}$ in $\LCd$ with $Q(h) \geq n^{1-{2^{-\Omega(d)}}}$. \end{restatable} By recursively applying \thm{composition}, we obtain a similar sublinear upper bound for the $\varepsilon$-approximate degree of $\LCd$, even for subconstant values of $\varepsilon$. \begin{restatable}{theorem}{approxdegLC}\label{thm:approxdegLC} For all constant $d\geq 0$, and any $\varepsilon > 0$, and all functions $h:\{0,1\}^n \to \{0,1\}$ in $\LCd$, we have \begin{equation} \deg_\varepsilon(h) = \widetilde{O}\Bigl(n^{1-2^{-d}}\log(1/\varepsilon)^{2^{-d}}\Bigr). \end{equation} \end{restatable} For constant $\epsilon$, we prove a lower bound of the same form with quadratically worse dependence on the depth $d$. \begin{restatable}{theorem}{lowerbound} \label{thm:lower} For all constants $d\geq 0$, there exists a function $h:\{0,1\}^n\to\{0,1\}$ in \LCd with $\adeg(h) \geq n^{1-2^{-\Omega(\sqrt{d})}}$. \end{restatable} A lower bound of $\adeg(h) = n^{1-2^{-\Omega(d)}}$ was already known for general {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace functions $f$ \cite{BT17, BKT18}, but the {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits constructed in these prior works are not of linear size. Previously, for any $\ell \geq 1$, \cite{BKT18} exhibited a circuit $C \colon \{0,1\}^n \to \{0,1\}$ of depth at most $3\ell$, size at most $n^2$, and approximate degree $\adeg(C) \geq \widetilde{\Omega}(n^{1-2^{-\ell}})$. We show how to transform this quadratic-size circuit $C$ into a linear-size circuit $C$ of depth roughly $\ell^2$, whose approximate degree is close to that of $C$. Our transformation adapts that of~\cite{CKK12}, but requires a more intricate construction and analysis. This is because, unlike quantum query complexity, approximate degree is not known to increase multiplicatively under block composition. \subsection{Application: Agnostically learning linear-size {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits} \label{sec:intro-learning} The challenging agnostic model~\cite{KSS:1994} of computational learning theory captures the task of binary classification in the presence of adversarial noise. In this model, a learning algorithm is given a sequence of labeled examples of the form $(x, b) \in \{0,1\}^n \times \{0,1\}$ drawn from an unknown distribution $\mathcal{D}$. The goal of the algorithm is to learn a hypothesis $h : \{0,1\}^n \to \{0,1\}$ which does ``almost as well'' at predicting the labels of new examples drawn from $\mathcal{D}$ as does the the best classifier from a known concept class ${\mathcal{C}}$. Specifically, let the Boolean loss of a hypothesis $h$ be $\mathsf{err}_{\mathcal{D}}(h) = \Pr_{(x, b) \sim \mathcal{D}}[h(x) \ne b]$. For a given accuracy parameter $\varepsilon$, the goal of the learner is to produce a hypothesis $h$ such that $\mathsf{err}_{\mathcal{D}}(h) \le \min_{c \in {\mathcal{C}}} \mathsf{err}_{\mathcal{D}}(c) + \varepsilon$. Very few concept classes ${\mathcal{C}}$ are known to be agnostically learnable, even in subexponential time. For example, the best known algorithm for agnostically learning disjunctions runs in time $2^{\tilde{O}(\sqrt{n})}$~\cite{kkms}.\footnote{Throughout this manuscript, $\tilde{O}$ and $\tilde{\Omega}$ notation hides factors polylogarithmic in the input size $n$.} Moreover, several hardness results are known. Proper agnostic learning of disjunctions (where the output hypothesis itself must be a disjunction) is NP-hard~\cite{KSS:1994}. Even improper agnostic learning of disjunctions is at least as hard as PAC learning DNF~\cite{lbw}, which is a longstanding open question in learning theory. \label{sec:introst} The best known general result for more expressive classes of circuits is that all de Morgan \emph{formulas} of size $s$ can be learned in time $2^{\tilde{O}(\sqrt{s})}$~\cite{kkms, Rei11} (\sec{detailed} contains a detailed overview of prior work on agnostic and PAC learning). Both of the aforementioned results make use of the well-known \emph{linear regression} framework of~\cite{kkms} for agnostic learning. This algorithm works whenever there is a ``small'' set of ``features'' $\mathcal{F}$ (where each feature is a function mapping $\{0,1\}^n$ to ${\mathbb{R}}$) such that each concept in the concept class ${\mathcal{C}}$ can be approximated to error $\varepsilon$ in the $\ell_\infty$ norm by a linear combination of features in $\mathcal{F}$. (See \sec{agnostic} for details.) If every function in a concept class ${\mathcal{C}}$ has approximate degree at most $d$, then one obtains an agnostic learning algorithm for ${\mathcal{C}}$ with running time $2^{\tilde{O}(d)}$ by taking $\mathcal{F}$ to be the set of all monomials of degree at most $d$. Applying this algorithm using the approximate degree upper bound of \thm{approxdegLC} yields a subexponential time algorithm for agnostically learning $\LCd$. \begin{restatable}{theorem}{learningresult} \label{thm:learning} The concept class of $n$-bit functions computed by {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits of depth $d$ can be learned in the distribution-free agnostic PAC model in time $2^{\widetilde{O}(n^{1-2^{-d}})}$. More generally, size-$s$ \ACd circuits can be learned in time $2^{\widetilde{O}(\sqrt{n} s^{1/2 - 2^{-d}})}$. \end{restatable} Prior to our work, no subexponential time algorithm was known even for agnostically learning $\mathsf{LC}^0_3$\xspace. Moreover, since our upper bound on the approximate degree of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits is nearly tight, new techniques will be needed to significantly surpass our results, and in particular, learn \emph{all} of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace in subexponential time. (Note that standard techniques~\cite{patmat} automatically generalize the lower bound of \thm{lower} from the feature set of low-degree monomials to \emph{arbitrary feature sets}. See \Cref{sec:agnostics} for details.) \subsection{Application: New Circuit Lower Bounds} An important frontier problem in circuit complexity is to show that the well-known Inner Product function cannot be computed by \ACmodtwo circuits of polynomial size. Here, \ACmodtwo\ refers to {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits augmented with a layer of parity gates at the bottom (i.e., closest to the inputs). Servedio and Viola \cite{servedio2012special} identified this open problem as a first step toward proving matrix rigidity lower bounds, itself a notorious open problem in complexity theory, and Akavia et al. \cite{akavia} connected the problem to the goal of constructing highly efficient pseudorandom generators.\footnote{Superpolynomial lower bounds are known for \ACmodtwo\ circuits computing the Majority function \cite{razborov} (in fact, even for AC$^0[2]$ circuits, which are AC$^0$ circuits augmented with parity gates at any layer). However, these techniques do not apply to the Inner Product function, which does have small AC$^0[2]$ circuits.} Unfortunately, the best known lower bounds against \ACmodtwo circuits computing Inner Product are quite weak. The state of the art result \cite{CGJWX} for any constant depth $d>4$ is that Inner Product cannot be computed by any depth-$(d+1)$ \ACmodtwo circuit of size $O(n^{1+4^{-(d+1)}})$. We show that \Cref{thm:approxdegLC} implies an improved (if still unsatisfying) lower bound of $\widetilde{\Omega}(n^{1/(1- 2^{-d})}) = n^{1+2^{-d} + \Omega(1)}$. More significantly, unlike prior work our lower bound holds even against circuits that compute the Inner Product function on slightly more than half of all inputs. Below, when we refer to the depth of an \ACmodtwo\ circuit, we count the layer of parity gates toward the depth. For example, we consider a DNF of parities to have depth 3. \begin{restatable}{theorem}{apptheorem} \label{apptheorem} For any constant integer $d \geq 4$, any depth-$(d+1)$ \ACmodtwo circuit computing the Inner Product function on $n$ bits on greater than a $1/2 + n^{-\log n}$ fraction of inputs has size $\widetilde{\Omega}\bigl(n^{1/(1- 2^{-d})}\bigr) = n^{1+ 2^{-d} + \Omega(1)}$. \end{restatable} \subsection{Discussion and future directions} \label{s:discussion} Summarizing our results, we established shared-input composition theorems for quantum query complexity (\thm{quantum}) and approximate degree (\thm{composition}), roughly showing that for compositions between an arbitrary function $f$ and the function $g = \AND$, it is always possible to leverage sharing of inputs to obtain algorithmic speedups. We applied these results to obtain the first sublinear upper bounds on the quantum query complexity and approximate degree of \LCd. \para{Generalizing our composition theorems.} Although considering the inner function $g=\AND$ is sufficient for our applications to {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace, an important open question is to generalize our results to larger classes of inner functions. The proof of our composition theorem for approximate degree actually applies to any inner function $g$ that can be exactly represented as a low-weight sum of $\AND$s (for example, it applies to any strongly unbalanced function $g$, meaning that $|g^{-1}(1)|=\poly(n)$). Extending this further would be a major step forward in our understanding of how quantum query complexity and approximate degree behave under composition with shared inputs. While our paper considers the composition scenario where the top function is arbitrary and the bottom function is $\AND$, the opposite scenario is also interesting. Here the top function is $\AND_m$ and the bottom functions are $f_1,\ldots,f_m$, each acting on the same set of $n$ input variables. Now the question is whether we can do better than the upper bound obtained using results on block composition that treat all the input variables as being independent. More concretely, for such a function $F$, the upper bound that follows from block composition is $Q(F)=O(\sqrt{m}\max_i Q(f_i))$. However, this upper bound cannot be improved in general, because the Surjectivity function is an example of such a function. Here the bottom functions $f_i$ check if the input contains a particular range element $i$, and the upper bound obtained from this argument is $O(n)$, which matches the lower bound~\cite{beame,sherstov15}. Surprisingly, this lower bound only holds for quantum query complexity, as we know that the approximate degree of Surjectivity is $\tilde{\Theta}(n^{3/4})$. We do not know if the upper bound obtained from block composition can be improved for approximate degree. \para{Quantum query complexity of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace and DNFs.} For quantum query complexity, we obtain the upper bound $Q(\LCd)=\tilde{O}(n^{1-2^{-d}})$, nearly matching the lower bound $Q(\LCd)=n^{1-2^{-\Omega(d)}}$ from \cite{CKK12}. However, the bounds do not match for any fixed value of $d$. The lack of matching lower bounds can be attributed to the fact that the Surjectivity function, which is known to have linear quantum query complexity, is computed by a quadratic-size depth-3 circuit, rather than a quadratic-size depth-2 circuit (i.e., a DNF). If one could prove a linear lower bound on the quantum query complexity of some quadratic-size DNF, the argument of \cite{CKK12} would translate this into a $\tilde{\Omega}(n^{1-2^{-d}})$ lower bound for \LCd, matching our upper bound. Unfortunately, no linear lower bound on the quantum query complexity of \emph{any} polynomial size DNFs is known; we highlight this as an important open problem (the same problem was previously been posed by Troy Lee with different motivations \cite{troylee}). \begin{openproblem} Is there a polynomial-size DNF with $\widetilde{\Omega}(n)$ quantum query complexity? \end{openproblem} The quantum query complexity of depth-2 {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace, or linear-size DNFs also remains open. The best upper bound is $O(n^{3/4})$, but the best lower bound is $\Omega(n^{0.555})$~\cite{CKK12}. Any improvement in the lower bound would also imply, in a black-box way, an improved lower bound for the Boolean matrix product verification problem. Improving the lower bound all the way to $\Omega(n^{3/4})$ would imply optimal lower bounds for all of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace using the argument in \cite{CKK12}. We conjecture that there is a linear-size DNF with quantum query complexity $\Omega(n^{3/4})$, matching the known upper bound. \para{Approximate degree of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace and DNFs.} For approximate degree, we obtain the upper bound $\adeg(\LCd)=\tilde{O}(n^{1-2^{-d}})$, and prove a new lower bound of $\adeg(\LCd)=n^{1-2^{-\Omega(\sqrt{d})}}$. The reason our approximate degree lower bound approaches $n$ more slowly than the quantum query lower bound from \cite{CKK12} is that, while the quantum query complexity of {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace is known to be $\Omega(n)$, such a result is not known for approximate degree. This remains an important open problem. \begin{openproblem} Is there a problem in {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace with approximate degree $\widetilde{\Omega}(n)$? \end{openproblem} Our lower bound argument would translate, in a black-box manner, any linear lower bound on the approximate degree of a general {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit into a nearly tight lower bound for \LCd. Alternatively, it would be very interesting if one could improve our approximate degree upper bound for \LCd. Even seemingly small improvements to our upper bound would have significant implications. Specifically, standard techniques (see, e.g., \cite{chaudhuri1996deterministic}) imply that for any constant $\delta>0$, there are approximate majority functions\footnote{Here, by an approximate majority function, we mean any total function $f$ on $n$ bits for which there exist constants $0 < p < 1/2 < q$ such that $|x|\leq pn \Longrightarrow f(x)=0$ and $|x|\geq qn \Longrightarrow f(x)=1$.} computable by depth-$(2d+3)$ circuits of size $O(n^{1+2^{-d}+\delta})$.\footnote{This precise result has not appeared in the literature; we prove it in \Cref{app:amaj} for completeness.} This means that, for sufficiently large constant $d$, if one could improve our upper bound on the approximate degree of \LCd from $\widetilde{O}(n^{1-2^{-d}})$ to $\widetilde{O}(n^{1-2^{-d/2.001}})$, one would obtain a sublinear upper bound on the approximate degree of some total function computing an approximate majority. This would answer a question of Srinivasan \cite{filmus2014real}, and may be considered a surprising result, as approximate majorities are currently the primary natural candidate {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace functions that may exhibit linear approximate degree \cite{BKT18}. \subsection{Paper organization and notation} This paper is organized so as to be accessible to readers without familiarity with quantum algorithms. \sec{quantum} assumes the reader is somewhat familiar with quantum query complexity and Grover's algorithm~\cite{Gro96}, but only uses Grover's algorithm as a black box. In \sec{quantum} we show our main result on the quantum query complexity of shared-input compositions (\thm{quantum}). \sec{approx} proves our result about the approximate degree of shared-input compositions (\thm{composition}). \sec{LC} uses the results of these sections (in a black-box manner) to upper bound the quantum query complexity and approximate degree of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits, and proves related lower bounds. \sec{agnostic} uses the results of \sec{LC} to obtain algorithms to agnostically PAC learn {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits. \Cref{s:app} derives our average-case lower bounds on the size of \ACmodtwo\ circuits computing the Inner Product function. In this paper we use the $\widetilde{O}(\cdot)$ and $\widetilde{\Omega}(\cdot)$ notation to suppress logarithmic factors. More formally, $f(n)=\widetilde{O}(g(n))$ means there exists a constant $k$ such that $f(n)=O(g(n)\log^k g(n))$, and similarly $f(n)=\widetilde{\Omega}(g(n))$ means there exists a constant $k$ such that $f(n)=\Omega(g(n)/\log^k g(n))$. For a string $x\in\{0,1\}^n$, we use $|x|=\sum_i x_i$ to denote the Hamming weight of $x$, i.e., the number of entries in $x$ equal to $1$. For any positive integer $n$, we use $[n]$ to denote the set $\{1,2,\ldots,n\}$. Given two functions $f_m, g_k$, let $f_m \circ g_k \colon \{0,1\}^{m \cdot k} \to \{0,1\}$ denote their \emph{block composition}, i.e., $(f_m \circ g_k)(x) = f_m(g_k(x_1), \dots, g_k(x_m))$, where for every $i\in[m]$, $x_i$ is a $k$-bit string. \section{Quantum algorithm for composed functions} \label{sec:quantum} \subsection{Preliminaries} As described in the introduction, our quantum algorithm only uses variants of Grover's algorithm~\cite{Gro96} and is otherwise classical. To make this section accessible to those without familiarity with quantum query complexity, we only state the minimum required preliminaries to understand the algorithm. Furthermore, we do not optimize the logarithmic factors in our upper bound to simplify the presentation. For a more comprehensive introduction to quantum query complexity, we refer the reader to the survey by Buhrman and de Wolf~\cite{dtsurvey}. In quantum or classical query complexity, the goal is to compute some known function $f:\{0,1\}^n \to \{0,1\}$ on some unknown input $x\in\{0,1\}^n$ while reading as few bits of $x$ as possible. Reading a bit of $x$ is also referred to as ``querying'' a bit of $x$, and hence the goal is to minimize the number of queries made to the input. For example, the deterministic query complexity of a function $f$ is the minimum number of queries needed by a deterministic algorithm in the worst case. A deterministic algorithm must be correct on all inputs, and can decide which bit to query next based on the input bits it has seen so far. Another example of a query model is the bounded-error randomized query model. The bounded-error randomized query complexity of a function $f$, denoted $R(f)$, is the minimum number of queries made by a randomized algorithm that computes the function correctly with probability greater than or equal to $2/3$ on each input. In contrast to a deterministic algorithm, such an algorithm has access to a source of randomness, which it may use in deciding which bits to query. The bounded-error quantum query complexity of $f$, denoted $Q(f)$, is similar to bounded-error randomized query complexity, except that the algorithm is now quantum. In particular, this means the algorithm may query the inputs in superposition. Since quantum algorithms can also generate randomness, for all functions we have $Q(f) \leq R(f)$. An important example of the difference between the two models is provided by the $\OR_n$ function, which asks if any of the input bits is equal to 1. We have $R(\OR_n)=\Theta(n)$, because intuitively if the algorithm only sees a small fraction of the input bits and they are all $0$, we do not know whether or not the rest of the input contains a $1$. However, Grover's algorithm is a quantum algorithm that solves this problem with only $O(\sqrt{n})$ queries~\cite{Gro96}. The algorithm is also known to be tight, and we have $Q(\OR_n)=\Theta(\sqrt{n})$~\cite{BBBV97}. There are several variants of Grover's algorithm that solve related problems and are sometimes more useful than the basic version of the algorithm. Most of these can be derived from the basic version of Grover's algorithm (and this sometimes adds logarithmic overhead). In this work we need a variant of Grover's algorithm that finds a $1$ in the input faster when there are many $1$s. Let the Hamming weight of the input $x$ be $t = |x|$. If we know $t$, then we can use Grover's algorithm on a randomly selected subset of the input of size $O(n/t)$, and one of the $1$s will be in this set with high probability. Hence the algorithm will have query complexity $O(\sqrt{n/t})$. With some careful bookkeeping, this can be done even when $t$ is unknown, and the algorithm will have expected query complexity $O(\sqrt{n/t})$. More formally, we have the following result of Boyer, Brassard, H{\o}yer, and Tapp~\cite{BBHT98}. \begin{lemma}\label{lem:BBHT} Given query access to a string $x\in\{0,1\}^n$, there is a quantum algorithm that when $t = |x| > 0$, always outputs an index $i$ such that $x_i=1$ and makes $O(\sqrt{{n}/{t}})$ queries in expectation. When $t = 0$, the algorithm does not terminate. \end{lemma} Note that because we do not know $t=|x|$, we only have a guarantee on the expected query complexity of the algorithm, not the worst-case query complexity. Note also that this variant of Grover's algorithm is a zero-error algorithm in the sense that it always outputs a correct index $i$ with $x_i = 1$ when such an index exists. In our algorithm we use an amplified version of the algorithm of \lem{BBHT}, which adds a log factor to the running time and always terminates in $O(\sqrt{n}\log n)$ time. \begin{lemma}\label{lem:ampBBHT} Given query access to a string $x\in\{0,1\}^n$, there is a quantum algorithm that \begin{enumerate} \item when $|x| = 0$, the algorithm always outputs ``$|x|=0$'', \item when $|x|>0$, it outputs an index $i$ with $x_i=1$ with probability $1-\frac{1}{\poly(n)}$, and \item terminates after $O\(\sqrt{\frac{n}{|x|+1}} \log n\)$ queries with probability $1-\frac{1}{\poly(n)}$. \end{enumerate} \end{lemma} \begin{proof} This algorithm is quite straightforward. We simply run $O(\log n)$ instances of the algorithm of \lem{BBHT} in parallel and halt if any one of them halts. If we reach our budget of $O(\sqrt{n}\log n)$ queries, then we halt and output ``$|x|=0$''. Let us argue that the algorithm has the claimed properties. First, since the algorithm of \lem{BBHT} does not terminate when $|x|=0$, our algorithm will correctly output ``$|x|=0$'' at the end for such inputs. When $|x|>0$, we know that the algorithm of \lem{BBHT} will find an index $i$ with $x_i=1$ with high probability after $O(\sqrt{n})$ time. The probability that $O(\log n)$ copies of this algorithm do not find such an $i$ is exponentially small in $O(\log n)$, or polynomially small in $n$. Finally, our algorithm makes only $O(\sqrt{n}\log n)$ queries when $|x|=0$ by construction. When $|x|>0$, we know that the algorithm of \lem{BBHT} terminates after an expected $O(\sqrt{{n}/{|x|}})$ queries, and hence halts with high probability after $O(\sqrt{{n}/{|x|}})$ queries by Markov's inequality. The probability that none of $O(\log n)$ copies of the algorithm halt after making $O(\sqrt{{n}/{|x|}})$ queries each is inverse polynomially small in $n$ again. \end{proof} \subsection{Quantum algorithm} We are now ready to present our main result for quantum query complexity, which we restate below. \quantum* While \thm{quantum} allows the bottom $\AND$ gates to depend on negated variables, it will be without loss of generality in the proof to assume that all input variables are unnegated. This is because we can instead work with the function $h' : \{0,1\}^{2n} \to \{0,1\}$ obtained by treating the positive and negative versions of a variable separately, increasing our final quantum query upper bound by a constant factor. We now define some notation that will aid with the description and analysis of the algorithm. We know that our circuit $h$ has $m$ $\AND$ gates and $n$ input bits $x_i$. We say an $\AND$ gate has \emph{high fan-in} if the number of inputs to that $\AND$ gate is greater than or equal to $n/Q(f)$. Note that if our circuit $h$ has no high fan-in gates, then we are done, because we can simply use the upper bound for block composition, i.e., $Q(f\circ g) = O(Q(f)Q(g))$, to compute $h$, since we will have $Q(h) = O(Q(f)\times \sqrt{n/Q(f)}) = O(\sqrt{Q(f)\cdot n})$. Our goal is to reduce to this simple case. More precisely, we will start with the given circuit $h$, make some queries to the input, and then simplify the given circuit to obtain a new circuit $h'$. The new circuit will have no high fan-in gates, but will still have $h'(x)=h(x)$ on the given input $x$. Note that $h'$ and $h$ have the same output only for the given input $x$, and not necessarily for all inputs. For any such circuit $h$, let $S \subseteq [m]$ be the set of all high fan-in $\AND$ gates, and let $w(S)$ be the total fan-in of $S$, which is the sum of fan-ins of all gates in $S$. In other words, it is the total number of wires incident to the set $S$. Since the set $S$ only has gates with fan-in at least $n/Q(f)$, we have \begin{equation} w(S) \geq n|S|/Q(f). \end{equation} We now present our first algorithm, which is a subroutine in our final algorithm. This algorithm's goal is to take a circuit $h$, with $S$ high fan-in gates and $w(S)$ wires incident on $S$, and reduce the size of $w(S)$ by a factor of $2$. Ultimately we want to have $|S|=w(S)=0$, and hence if we can decrease the size of $w(S)$ by $2$, we can repeat this procedure logarithmically many times to get $|S|=w(S)=0$. \begin{lemma} \label{lem:halving} Let $h:\{0,1\}^n \to \{0,1\}$ be a depth-2 circuit where the top gate is a function $f:\{0,1\}^m \to \{0,1\}$ and the bottom level gates are $\AND$ gates on a subset of the input bits and their negations (as depicted in \fig{composition}). Let $w(S)$ be the total fan-in of all high fan-in gates in $h$ (i.e., gates with fan-in $\geq n/Q(f)$). Then there is a quantum query algorithm that makes $O(\sqrt{Q(f)\cdot n} \log n)$ queries to $x\in\{0,1\}^n$ and outputs a new circuit $h'$ of the same form such that $w(S')\leq w(S)/2$, where $w(S')$ is the total fan-in of all high fan-in gates in $h'$, and such that with probability $1-\frac{1}{\poly(n)}$ we have $h(x)=h'(x)$. \end{lemma} \begin{proof} The overall structure of the claimed algorithm is the following: We query some well-chosen input bits, and on learning the values of these bits, we simplify the circuit accordingly. If an input bit is 0, then we delete all the $\AND$ gates that use that input bit. If an input bit is 1, we delete all outgoing wires from that input bit since a 1-input does not affect the output of an $\AND$ gate. Since the circuit will change during the algorithm, let us define $S_0$ to be the initial set of high fan-in (i.e., gates with fan-in $\geq n/Q(f)$) $\AND$ gates in $h$. We also define the degree of an input $x_i$, denoted $\deg(i)$, to be the number of high fan-in $\AND$ gates that it is an input to. Note that this is not the total number of outgoing wires from $x_i$, but only those that go to high fan-in $\AND$ gates, i.e., gates in the set $S$. With this definition, note that $\sum_{i\in[n]} \deg(i) = w(S)$, for any circuit. We say an input bit $x_i$ is \emph{high degree} if $\deg(i)\geq |S_0|/(2Q(f))$. This value is chosen since it is at least half the average degree of all $x_i$ in the initial circuit $h$. As the algorithm progresses, the circuit will change, and some inputs that were initially high degree may become low degree as the algorithm progresses, but a low degree input will never become high degree. But note that the definition of a high-degree input bit does not change, since it only depends on $S_0$ and $Q(f)$, which are fixed for the duration of the algorithm. Finally, we call an input bit $x_i$ is \emph{marked} if $x_i=0$. We are now ready to describe our algorithm by the following pseudocode (see \alg{basic}). \begin{algorithm} \caption{The algorithm of \lem{halving}.\label{alg:basic}} \begin{algorithmic}[1] \Statex \Let{$S_0$}{Set of high fan-in $\AND$ gates in $h$} \Repeat \Let{$M$}{Set of high-degree marked inputs \algorithmiccomment{$M := \left\{i: x_i =0 \wedge \deg(i)\geq \frac{|S_0|}{(2Q(f))}\right\}$}} \State{Grover Search for an index $i$ in $M$} \If{we find such an $i$} \State{Delete all $\AND$ gates that use $x_i$ as an input} \EndIf \Until{Grover Search fails to find an $i \in M$} \State{Delete all remaining high-degree inputs and all outgoing wires from these inputs} \end{algorithmic} \end{algorithm} In more detail, we repeatedly use the version of Grover's algorithm in \lem{ampBBHT} to find a high-degree marked input, which is an input $x_i$ such that $x_i=0$ and $\deg(i)\geq \frac{|S_0|}{2Q(f)}$. If we find such an input, we delete all the $\AND$ gates that use $x_i$ as an input, and repeat this procedure. Note that when we repeat this procedure, the circuit has changed, and hence the set of high-degree input bits may become smaller. The algorithm halts when Grover's algorithm is unable to find any high-degree marked inputs. At this point, all the high-degree inputs are necessarily unmarked with very high probability, which means they are set to $1$. We can now delete all these input bits and their outgoing wires because $\AND$ gates are unaffected by input bits set to $1$. Let us now argue that this algorithm is correct. Let $S'$ denote the set of high fan-in $\AND$ gates in the new circuit $h'$ obtained at the end of the algorithm, and $w(S')$ be the total fan-in of gates in $S'$. Note that when the algorithm terminates, there are no high-degree inputs (marked or unmarked). Hence every input bit that has not been deleted has $\deg(i)<\frac{|S_0|}{2Q(f)}$. Since there are at most $n$ input bits, we have \begin{equation} w(S')=\sum_{i \in [n]} \deg(i) < \frac{n}{2Q(f)}|S_0|. \end{equation} But we also know that we started with $w(S)\geq n|S_0|/Q(f)$, since each gate in $S_0$ has fan-in at least $n/Q(f)$. Hence $w(S')\leq w(S)/2$, which proves that the algorithm is correct. We now analyze the query complexity of this algorithm. Let the loop in the algorithm execute $r$ times. It is easy to see that $r \leq 2Q(f)$ because each time a high-degree marked input is found, we delete all the $\AND$ gates that use it as an input, which is at least $|S_0|/(2Q(f))$ gates. Since there were at most $S_0$ gates to begin with, this procedure can only repeat $2Q(f)$ times. When we run Grover's algorithm to search for a high-degree marked input bit $x_i$ in the first iteration of the loop, suppose there are $k_1$ high-degree marked inputs. Then the variant of Grover's algorithm in \lem{ampBBHT} finds a marked high-degree input and makes $O(\sqrt{n/k_1}\log n)$ queries with probability $1-\frac{1}{\poly(n)}$. In the second iteration of the loop, the number of high-degree marked inputs, $k_2$, has decreased by at least one. It can also decrease by more than 1 since we deleted several $\AND$ gates, and some high-degree inputs can become low-degree. In this iteration, our variant of Grover's algorithm (\lem{ampBBHT}) makes $O(\sqrt{n/k_2} \log n)$ queries, and we know that $k_1>k_2$. This process repeats and we have $k_1>k_2>\cdots >k_r$. Since there was at least one high-degree marked input in the last iteration, $k_r\geq 1$. Combining these facts we have for all $j\in[r]$, $k_j \geq r-j+1$. Thus the total expected query complexity is \begin{equation} O\(\sum_{j=1}^r \sqrt{\frac{n}{k_j}} \log n\) = O\(\sum_{j=1}^r \sqrt{\frac{n}{r-j+1}} \log n\) = O\(\sqrt{n} \sum_{j=1}^r \frac{1}{\sqrt{j}} \log n\) = O\(\sqrt{nr} \log n\), \end{equation} which is $O\(\sqrt{n \cdot Q(f)} \log n\).$ We now have a quantum query algorithm that satisfies the conditions of the lemma with probability at least $1-\frac{1}{\poly(n)}$. \end{proof} We are now ready to prove \thm{quantum}. \begin{proof}[Proof of \protect{\thm{quantum}}] We start by applying the algorithm in \lem{halving} to our circuit as many times as needed to ensure that set $S$ is empty. Since each run of the algorithm reduces $w(S)$ by a factor of 2, and $w(S)$ can start off being as large as $m\cdot n$, where $m$ is the number of $\AND$ gates and $n$ is the number of inputs, we need to run the algorithm $\log(mn)$ times. Since the algorithm of \lem{halving} is correct with probability $1-\frac{1}{\poly(n)}$, we do not need to boost the success probability of the algorithm. The total number of queries needed to ensure $S$ is empty is $O(\sqrt{Q(f)\cdot n}\log(n)\log(mn))$. Now we are left with a circuit $h'$ with no high fan-in $\AND$ gates. That is, all $\AND$ gates have fan-in at most $n/Q(f)$. We now evaluate $h'$ using the standard composition theorem for disjoint sets of inputs, which has query complexity \begin{equation} O(Q(f) \cdot Q(\AND_{n/Q(f)}))= O(Q(f) \cdot \sqrt{n/Q(f)}) = O\(\sqrt{Q(f)\cdot n}\). \end{equation} The total query complexity is $O(\sqrt{Q(f)\cdot n}\log(n)\log(mn)) = O(\sqrt{Q(f)\cdot n}\log^2(mn))$. \end{proof} Note that we have not attempted to reduce the logarithmic factors in this upper bound. We believe it is possible to make the quantum upper bound match the upper bound for approximate degree with a more careful analysis and slightly different choice of parameters in the algorithm. \section{Approximating polynomials for composed functions} \label{sec:approx} \subsection{Preliminaries} We now define the various measures of Boolean functions and polynomials that we require in this section. Since we only care about polynomials approximating Boolean functions, we focus without loss of generality on multilinear polynomials as any polynomial over the domain $\{0,1\}^n$ can be converted into a multilinear polynomial (since it never helps to raise a Boolean variable to a power greater than $1$). The approximate degree of a Boolean function, commonly denoted $\adeg(f)$, is the minimum degree of a polynomial that entrywise approximates the Boolean function. It is a basic complexity measure and is known to be polynomially related to a host of other complexity measures such as decision tree complexity, certificate complexity, and quantum query complexity~\cite{dtsurvey}. We also use another complexity measure of polynomials, which is the sum of absolute values of all the coefficients of the polynomial. This is the query analogue of the so-called $\mu$-norm used in communication complexity~\cite[Definition 2.7]{LS09}. We now formally define these measures. \begin{definition} Let $p:{\mathbb{R}}^n \to {\mathbb{R}}$ be a multilinear polynomial \begin{equation} p(x_1,\ldots,x_n)=\sum_{s\in\{0,1\}^n} \alpha_s x_1^{s_1}\cdots x_n^{s_n}. \end{equation} We define the following complexity measures of the polynomial $p$: \begin{align} \deg(p) = \max \Bigl\{\sum_{i\in[n]} |s_i|:\alpha_s\neq 0\Bigr\} \qquad \mathrm{and} \qquad \mon(p) = \sum_{s\in\{0,1\}^n} |\alpha_s|. \end{align} For a Boolean function $f:\{0,1\}^n \to \{0,1\}$, we define the following complexity measures: \begin{align} \deg_{\varepsilon}(f) &= \min \{\deg(p):\forall x\in\{0,1\}^n,~|f(x)-p(x)|\leq \varepsilon\}\\ \mu_{\varepsilon}(f) &= \min \{\mon(p):\forall x\in\{0,1\}^n,~|f(x)-p(x)|\leq \varepsilon\} \end{align} Finally, we define $\adeg(f)=\deg_{1/3}(f)$ and $\amu(f)=\mu_{1/3}(f)$. \end{definition} We use the following standard relationship between the two measures in our results. \begin{lemma} \label{lem:degreebound} For any multilinear polynomial $p:{\mathbb{R}}^n \to {\mathbb{R}}$ such that $|p(x)|=O(1)$ for all $x\in\{0,1\}^n$, we have \begin{equation} \log \mu(p) = O(\deg(p)\log n).\label{eq:mudeg} \end{equation} Consequently, for any Boolean function $f:\{0,1\}^n \to \{0,1\}$ and $\varepsilon\in[0,1/3]$, we have \begin{equation} \log \mu_\varepsilon(f) = O(\deg_\varepsilon(f)\log n). \end{equation} \end{lemma} \begin{proof} First let us switch to the $\{-1,1\}$ representation instead of the $\{0,1\}$ representation we have used so far. Let $y_i = (-1)^{x_i}$, and replace every occurrence of $x_i$ in the polynomial $p$ with $\frac{1}{2}(1+y_i)$ to obtain a multilinear polynomial $p(y_1,\ldots,y_n)=\sum_{s\in\{0,1\}^n} \beta_s y_1^{s_1}\cdots y_n^{s_n}$. In this representation, a coefficient $\beta_s$ is simply the expectation over the hypercube of the product of $p$ and a parity function, and hence is at most $O(1)$ in magnitude. Since there are only $\binom{n}{\deg(p)}$ monomials, the sum of absolute values of all coefficients is $O(n^{\deg(p)})$. When we switch from this representation back to the $\{0,1\}$ representation, we replace every $y_i$ with $2x_i-1$. Consider this transformation on a single monomial with coefficient $1$. This converts the monomial of degree $d$ into a polynomial over those $d$ variables, such that the sum of coefficients in this polynomial is at most $3^d$. Thus the sum of absolute values of all coefficients is $\mu(p)=O(3^{\deg(p)} n^{\deg(p)}) = n^{O(\deg(p))}$, which proves \eq{mudeg}. Now consider any Boolean function $f:\{0,1\}^n\to\{0,1\}$, and a multilinear polynomial $p$ that minimizes $\deg_\varepsilon(f)$. We can apply \eq{mudeg} to this polynomial to obtain $\log \mu(p) = O(\deg(p)\log n)$. Since $\deg(p) = \deg_\varepsilon(f)$ by assumption, and $\mu_\varepsilon(f)\leq\mu(p)$, since $\mu_\varepsilon(f)$ minimizes over all $\varepsilon$-approximating polynomials, we get $\log \mu_\varepsilon(f) = O(\deg_\varepsilon(f)\log n)$. \end{proof} This shows that $\log \mu(p)$ is at most $\deg(p)$ (up to log factors). However, $\log \mu(p)$ may be much smaller than $\deg(p)$, as evidenced by the polynomial $p(x)=x_1\cdots x_n$. Similarly, $\log \amu(f)$ may be much smaller than $\adeg(f)$, as evidenced by the $\AND$ function on $n$ bits, which has $\adeg(\AND_n)=\Theta(\sqrt{n})$ \cite{nisanszegedy}, but $\amu(\AND_n)\leq 1$. \subsection{Polynomial upper bound} In this section we prove \thm{composition}, which follows from the following more general composition theorem. \begin{restatable}{theorem}{composition-full} \label{thm:composition-full} Let $h:\{0,1\}^n\to \{0,1\}$ be computed by a depth-2 circuit where the top gate is a function $f:\{0,1\}^m \to \{0,1\}$ and the bottom level gates are $\AND$ gates on a subset of the input bits and their negations (as depicted in \fig{composition}). Then \begin{equation} \deg_\varepsilon(h) = O\(\sqrt{n\log{\mu_\varepsilon(f)}}+\sqrt{n\log(1/\varepsilon)}\)=O\(\sqrt{n\deg_\varepsilon(f)\log m}+\sqrt{n\log(1/\varepsilon)}\).\label{eq:comp} \end{equation} \end{restatable} \begin{proof} Let us first fix some notation. We will use $x\in\{0,1\}^n$ to refer to the input of the full circuit $h:\{0,1\}^n \to \{0,1\}$. Let the inputs to the top $f:\{0,1\}^m \to \{0,1\}$ gate be called $y_1,\ldots, y_m$. Let $p:\{0,1\}^m \to \{0,1\}$ be a polynomial that minimizes $\mu_\varepsilon(f)$. Thus we have for all $y\in\{0,1\}^m$, $|p(y)-f(y)|\leq \varepsilon$. More explicitly, $p(y_1,\ldots,y_m) = \sum_{s\in\{0,1\}^m} \alpha_s y_1^{s_1}\cdots y_n^{s_n}$, where $\mu_\varepsilon(f)=\sum_{s\in\{0,1\}^m} |\alpha_s|$, and each $y_i$ is the $\AND$ of some subset of bits in $x$. Since the product of $\AND$s of variables is just an $\AND$ of all the variables involved in the product, for each $s \in\{0,1\}^m$, there is a subset $T_s \subseteq [n]$ such that $y_1^{s_1}\cdots y_n^{s_n} = \bigwedge_{i\in T_s} x_i$. Using this we can replace all the $y$ variables in the polynomial $p$, to obtain \begin{equation} q(x)=\sum_{s\in\{0,1\}^m} \alpha_s \bigwedge_{i\in T_s} x_i. \end{equation} Since $p$ was an $\varepsilon$ approximation to $f$, $q$ is an $\varepsilon$ approximation to $h$. Now we can replace every occurrence of $\bigwedge_{i\in T_s} x_i$ with a low error approximating polynomial for the $\AND$ of the bits in $T_s$. We know that the approximate degree of the $\AND$ function to error $\delta$ is $O(\sqrt{n\log(1/\delta)})$~\cite{smallerrorquantum}. If we approximate each $\AND$ to error $\delta=\varepsilon/\mu_\varepsilon(f)$, then by the triangle inequality the total error incurred by this approximation is at most $\sum_{s\in\{0,1\}^m} |\alpha_s|\varepsilon/\mu_\varepsilon(f) = \varepsilon$. Choosing $\delta=\varepsilon/\mu_\varepsilon(f)$, each $\AND$ is approximated by a polynomial of degree $O(\sqrt{n\log(1/\delta)}) = O\left(\sqrt{n\log \mu_\varepsilon(f)}+\sqrt{n\log(1/\varepsilon)}\right)$. Hence the resulting polynomial $q(x)$ has this degree and approximates the function $h$ to error $2\varepsilon$. By standard error reduction techniques~\cite{BuhrmanNRW07}, we can make this error smaller than $\varepsilon$ at a constant factor increase in the degree. This establishes the first equality in \eq{comp}, and the second equality follows from \lem{degreebound}. \end{proof} \section{Applications to linear-size \texorpdfstring{{$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace}{AC0} circuits} \label{sec:LC} \subsection{Preliminaries} A Boolean circuit is defined via a directed acyclic graph. Vertices of fan-in 0 represent input bits, vertices of fan-out 0 represent outputs, and all other vertices represent one of the following logical operations: a $\mathsf{NOT}$ operation (of fan-in 1), or an unbounded fan-in $\mathsf{AND}$ or $\mathsf{OR}$ operation. The size of the circuit is the total number of $\mathsf{AND}$ and $\mathsf{OR}$ gates. The depth of the circuit is the length of the longest path from an input bit to an output bit. or any constant integer $d > 0$, \ACd refers to the class of all such circuits of polynomial size and depth $d$. {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace refers to $\cup_{d=1}^{\infty}$\ACd. Similarly, \LCd refers to the class of all such circuits of size $O(n)$ and depth $d$, while {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace refers to $\cup_{d=1}^{\infty}$\LCd. We will associate any circuit $C$ with the function it computes, so for example $\adeg(C)$ denotes the approximate degree of the function computed by $C$. It will be convenient to assume that any \ACd circuit is layered, in the sense that it consists of $d$ levels of gates which alternate between being comprised of all $\AND$ gates or all $\OR$ gates, and all negations appear at the input level of the circuit. Any \ACd circuit of size $s$ can be converted into a layered circuit of size $O(d \cdot s)$, and hence making this assumption does not change any of our upper bounds. \subsection{Quantum query complexity} Applying our composition theorem for quantum algorithms (\thm{quantum}) inductively, we obtain a sublinear upper bound on the quantum query complexity of \LCd circuits. \quantumLC* \begin{proof} We prove this for depth-$d$ {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits by induction on $d$. The base case is $d=1$, where the function is either $\AND$ or $\OR$ on $n$ variables, both of which have quantum query complexity $O(\sqrt{n})$~\cite{Gro96}. Now consider a function $h$, which is a layered depth-$d$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit of size $O(n)$. It can be written as a depth-$2$ circuit (as in \thm{quantum}) where the top function is a {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuit $f$ of depth $d-1$ on at most $O(n)$ inputs, and the bottom layer has only $\AND$ gates. (If the bottom layer has $\OR$ gates we can consider the negation of the function without loss of generality, since the quantum query complexity of a function and its negation is the same.) By the induction hypothesis we know that the quantum query complexity of any depth-$(d-1)$, size-$O(n)$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit with $O(n)$ inputs is $\widetilde{O}(n^{1-2^{-(d-1)}})$. Invoking \thm{quantum}, we have that the quantum query complexity of the depth-$d$ function $h$ is $\widetilde{O}\bigl(n^{1-2^{-d}}\bigr)$. \end{proof} \subsection{Approximate degree upper bound} We can now prove \Cref{thm:approxdegLC}, restated below for convenience: \approxdegLC* This follows from a more general result: \begin{restatable}{theorem}{upperbound} \label{thm:upper} For any function $h:\{0,1\}^n\to\{0,1\}$ computed by an {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit of size $s\geq 1$ and depth $d\geq 1$, we have \begin{equation} \deg_\varepsilon(h)=\begin{cases} O\(\sqrt{n\log(1/\varepsilon)}\) &\mbox{if } \varepsilon \leq 2^{-s} \Leftrightarrow \log(1/\varepsilon)\geq s\\ \widetilde{O}\(\sqrt{n} s^{{1}/{2}-2^{-d}} \({\log(1/\varepsilon)}\)^{2^{-d}}\) & \mbox{if } \varepsilon > 2^{-s} \Leftrightarrow \log(1/\varepsilon) < s \end{cases}. \end{equation} In particular, for any $h\in\LCd$, we have $\adeg(h)=\widetilde{O}\bigl(n^{1-2^{-d}}\bigr)$. \end{restatable} \begin{proof} We prove this for depth-$d$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits by induction on $d$. The base case is $d=1$, where the function is either $\AND$ or $\OR$ on $n$ variables, both of which have $\varepsilon$-approximate degree $O(\sqrt{n\log (1/\varepsilon)})$~\cite{smallerrorquantum}. Now consider a function $h$, which is a general depth-$d$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit of size $s$. It can be written as a depth-$2$ circuit (as in \thm{composition}) where the top function is a size-$s$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit $f$ of depth $d-1$ on at most $s$ inputs, and the bottom layer has only $\AND$ gates. If the bottom layer has $\OR$ gates we can consider the negation of the function without loss of generality, since the $\varepsilon$-approximate degree of a function and its negation is the same. In the first case, if $\varepsilon \leq 2^{-s}$, then for any function $f:\{0,1\}^s\to\{0,1\}$ there is a polynomial of degree $s$ and sum of coefficients at most $2^s$ that exactly equals $f$ on all Boolean inputs. Hence we can apply \thm{composition} to get that $\deg_\varepsilon(h)=O(\sqrt{ns}+\sqrt{n\log(1/\varepsilon)}) = O(\sqrt{n\log(1/\varepsilon)})$. In the second case, if $\varepsilon > 2^{-s}$, by the induction hypothesis we know that the $\varepsilon$-approximate degree of any depth-$(d-1)$, size-$O(s)$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuit with $s$ inputs is $\widetilde{O}(s^{1-2^{-(d-1)}}(\log(1/\varepsilon))^{2^{-(d-1)}})$. Invoking \thm{composition}, we have that the approximate degree of the depth-$d$ function is \begin{equation} \widetilde{O}\(\sqrt{n s^{1-2^{-(d-1)}}(\log(1/\varepsilon))^{2^{-(d-1)}}}+\sqrt{n\log(1/\varepsilon)}\)=\widetilde{O}\(\sqrt{n} s^{1/2-2^{-d}}(\log(1/\varepsilon))^{2^{-d}}\).\qedhere \end{equation} \end{proof} \subsection{Approximate degree lower bound} \label{sec:lower} In this section we prove our lower bound on the approximate degree of \LCd, restated below for convenience. \lowerbound* Before proving the theorem, we will need to introduce several lemmas. The first lemma follows from the techniques of \cite{AB84} (see \cite{Kop13} for an exposition). \begin{lemma}\label{lem:gapmaj} There exists a Boolean circuit $C$ with $n$ inputs, of depth 3, and size $\widetilde{O}(n^2)$ satisfying the following two properties: \begin{itemize} \item $C(x) = 0$ for all $x$ of Hamming weight at most $n/3$. \item $C(x) = 1$ for all $x$ of Hamming weight at least $2n/3$. \end{itemize} \end{lemma} We refer to the function computed by the circuit $C$ of \lem{gapmaj} as $\mathsf{GAPMAJ}$, short for a gapped majority function (such a function is sometimes also called an \emph{approximate majority} function). The following lemma of \cite{bchtv} says that if $f$ has large $\varepsilon$-approximate degree for $\varepsilon=1/3$, then block-composing $f$ with $\mathsf{GAPMAJ}$ on $O(\log n)$ bits yields a function with just as high $\varepsilon'$-approximate degree, with $\varepsilon'$ very close to $1/2$. \begin{lemma}[\cite{bchtv}] \label{lemmaa} Let $f \colon \{0,1\}^n \to \{0,1\}$ be any function. Then for $\varepsilon=1/2-1/n^2$, $\deg_{\varepsilon}(\mathsf{GAPMAJ}_{10 \log n} \circ f) \geq \adeg(f)$. \end{lemma} The following lemma says that if $f$ has large $\varepsilon$-approximate degree for $\varepsilon$ very close to $1/2$, then block-composing any function $g$ with $f$ results in a function of substantially larger approximate degree than $g$ itself. \begin{lemma}[\cite{sherstovhalfspaces1}] \label{lemmab} Let $g \colon \{0,1\}^m \to \{0,1\}$ and $f \colon \{0,1\}^n \to \{0,1\}$ be any functions. Then $\adeg(g \circ f) \geq \adeg(g) \cdot \deg_{1/2-1/m^2}(f)$. \end{lemma} Combining Lemmas \ref{lemmaa} and \ref{lemmab}, we conclude: \begin{corollary} \label{nomorecors} Let $g \colon \{0,1\}^m \to \{0,1\}$ and $f \colon \{0,1\}^n \to \{0,1\}$ be any functions. Then $\adeg(g \circ \mathsf{GAPMAJ}_{10\log n} \circ f) \geq \adeg(g) \cdot \adeg(f)$. \end{corollary} \begin{proof}[Proof of \Cref{thm:lower}] Let $\ell \geq 1$ be any constant integer to be specified later (ultimately, we will set $\ell=\Theta(\sqrt{d})$, where $d$ is as in the statement of the theorem). \cite{BKT18} exhibit a circuit family $C^* \colon \{0,1\}^n \to \{0,1\}$ of depth at most $3\ell$, size at most $n^2$, and approximate degree satisfying $\adeg(C^*) \geq D$ for some $D\geq \widetilde{\Omega}(n^{1-2^{-\ell}})$. We need to transform this quadratic-size circuit into a circuit $C$ of \emph{linear} size, without substantially reducing its approximate degree, or substantially increasing its depth (in particular, the depth of $C$ should be at most $d$). To accomplish this, we apply the following iterative transformation. At each iteration $i$, we produce a new circuit $C^i \colon \{0,1\}^n \to \{0,1\}$ of linear size, such that $\adeg(C^i)$ gets closer and closer to $\adeg(C)$ as $i$ grows. Our final circuit will be $C:=C^{\ell}$. $C^1$ is defined to simply be $\OR_n$, which is clearly in $\mathsf{LC}^0_1$\xspace. The transformation from $C^{i-1}$ into $C^{i}$ works as follows. $C^i$ feeds $\sqrt{n}$ copies of $C^{i-1}_{\sqrt{n}/(10 \log n)}$ into the circuit $C^*_{\sqrt{n}} \circ \textsf{GAPMAJ}_{10\log n}$. Here, $C^{i-1}_{k}$ denotes the function $C^{i-1}$ constructed in the previous iteration, and defined on $k$ inputs; similarly, $C^*_{k} \colon \{0,1\}^{k} \to \{0,1\}^n$ refers to the function $C^*$ constructed by \cite{BKT18}, defined on $k$ inputs. That is: \begin{equation} \label{cidef} C^i = C^*_{\sqrt{n}} \circ \textsf{GAPMAJ}_{10\log n} \circ C^{i-1}_{\sqrt{n}/(10 \log n)}. \end{equation} Observe that $C^i$ is a function on $\sqrt{n} \cdot 10 \log n \cdot (\sqrt{n}/(10\log n)) = n$ bits. We now establish the following two lemmas about $C^i$. \begin{lemma} \label{lem:circuitlemma} $C^i$ is computed by a circuit of depth at most $(3\ell+3) \cdot i$, and size at most $2\cdot i \cdot n$. \end{lemma} \begin{proof} Clearly this is true for $i=1$, since $C^1$ is computed by a circuit of size and depth 1. Assume by induction that it is true for $i-1$. Recalling that $\textsf{GAPMAJ}_{10\log n}$ is computed by a circuit of size $O(\log^2 n)$ and depth 3, and $C^*_{\sqrt{n}}$ is computed by a circuit of size $n$ and depth $3\ell$, it is immediate from Equation \eqref{cidef} that $C^i$ is computed by a circuit satisfying the following properties: \begin{itemize} \item The depth is at most $3 \ell + 3 + (3 \ell + 3)(i-1) = (3\ell + 3)i$. \item The size is at most $n + O(\sqrt{n} \cdot \log^2 n) + \left(\sqrt{n} \cdot 10 \log n\right) \cdot \left(2 \cdot (i-1) \cdot \sqrt{n}/(10 \log n)\right)$. For large enough $n$, this is at most $2n + 2 \cdot (i-1) \cdot n = 2 \cdot i \cdot n$. \end{itemize} \end{proof} \begin{lemma} \label{lem:adeglemma} For $i > 1$, $\adeg(C^i) \geq \Omega\left( \adeg(C^*_{\sqrt{n}}) \cdot \adeg(C^{i-1}_{\sqrt{n}/(10 \log n)})\right).$ \end{lemma} \begin{proof} Immediate from \Cref{nomorecors}. \end{proof} Since $\adeg(C^1) = \Omega(\sqrt{n})$, repeated application of \lem{adeglemma} implies that $\adeg(C^2) = \Omega(\sqrt{D} \cdot n^{1/4})$, $\adeg(C^3) = \Omega\left(\sqrt{D} \cdot (\sqrt{D} \cdot n^{1/4})^{1/2}\right) = \Omega(D^{3/4} \cdot n^{1/8})$, and in general, $\adeg(C^i) = \Omega\left(D^{1-2^{-i}} \cdot n^{2^{-i}}\right)$. Setting $i=\ell$, we obtain a circuit $C^{\ell} \colon \{0,1\}^n \to \{0,1\}$ with the following properties: \begin{itemize} \item By \lem{circuitlemma}, $C^{\ell}$ has size at most $2 \ell n$ and depth at most $d:=2 \ell^2$. \item There is a constant $c_0$ such that $C^{\ell}$ has approximate degree at least $\Omega\left(c_0^{\ell} \cdot D^{1-2^{-\ell+1}} \cdot n^{2^{-\ell}}\right) \geq \Omega(c_0^{\ell} \cdot n^{1-2^{-\ell+1/2}})$. \end{itemize} Hence, for any constant value of $d=2\ell^2$, we have constructed a circuit of depth $d$, size $O(n)$, and approximate degree at least $\Omega(n^{1-2^{-\Omega(\sqrt{d})}})$, as required by the theorem. \end{proof} \subsection{Sublinear-size circuits of arbitrary depth} \label{sec:sublinear} \thm{quantum} and \Cref{thm:composition} also allow us to prove sublinear quantum query complexity and approximate degree upper bounds for arbitrary circuits of sublinear size. \begin{theorem} \label{thm:sublinearadeg} Let $h : \{0,1\}^n \to \{0,1\}$ be computed by a layered circuit of size $s \le n$. Then $h$ has quantum query complexity $Q(h) = \tilde{O}(\sqrt{n s})$ and approximate degree $\adeg(h) = O(\sqrt{ns})$. \end{theorem} \begin{proof} Without loss of generality, a function $h$ computed by a layered circuit of size $s\leq n$ can be written as a depth-2 circuit with a function $f:\{0,1\}^s \to \{0,1\}$ as the top gate and $\AND$ gates at the bottom. (The case where the bottom level consists of $\OR$ gates can be handled by negating the function.) The quantum query upper bound then follows immediately from \thm{quantum}, as $Q(f) \le s$. Moreover, for any function $f$, we have $\log \mu_0(f) = O(s)$, since the trivial polynomial obtained by adding all conjunctions over yes-inputs of $f$ satisfies this. Hence from \thm{composition} we have $\adeg(h) = O(\sqrt{ns})$. \end{proof} \section{Applications to agnostic PAC learning} \label{sec:agnostic} Our new upper bounds on the approximate degree of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits yield new subexponential time learning algorithms in the agnostic model. In this section, we provide background for, and the proof of, our main learning result restated below. \learningresult* \paragraph{PAC and agnostic learning models.} In the classic Probably Approximately Correct (PAC) learning model of Valiant~\cite{Valiant:1984}, we have access to an unknown function $f:\{0,1\}^n \to \{0,1\}$ from a known class of functions ${\mathcal{C}}$, called the concept class, through samples $(x,f(x))$, where $x$ is drawn from an unknown distribution $\mathcal{D}$ over $\{0,1\}^n$. The goal is to learn a hypothesis $h:\{0,1\}^n \to \{0,1\}$, such that with probability $1-\delta$ (over the choice of samples), $h(x)$ has (Boolean) loss at most $\varepsilon$ with respect to $\mathcal{D}$. Here, the Boolean loss $\mathsf{err}_{\mathcal{D}}(h, f)$ of $h$ is defined to be $\Pr_{x \sim \mathcal{D}}[h(x) \neq f(x)] \leq \varepsilon$. Since the learning algorithm does not know $\mathcal{D}$ and is required to work for all $\mathcal{D}$, this model is also called the distribution-independent (or distribution-free) PAC model. Unfortunately, in the distribution-free setting, very few concept classes are known to be PAC learnable in polynomial time or even subexponential time (i.e., time $2^{n^{1-\delta}}$ for some constant $\delta>0$). Kearns, Schapire, and Sellie~\cite{KSS:1994} then proposed the more general (and challenging) agnostic PAC learning model, which removes the assumption that examples are determined by a function at all, let alone a function in the concept class ${\mathcal{C}}$. The learner now knows nothing about how examples are labeled, but is only required to learn a hypothesis $h$ that is at most $\varepsilon$ worse than the best possible classifier from the class ${\mathcal{C}}$. We now describe the agnostic PAC model more formally. Let $\mathcal{D}$ be any distribution on $\{0,1\}^n \times \{0,1\}$, and let ${\mathcal{C}}$ be a concept class, i.e., a set of Boolean functions on $\{0,1\}^n$. Define the error of $h \colon \{0,1\}^n \to \{0,1\}$ to be $\mathsf{err}_{\mathcal{D}}(h) := \Pr_{(x, b) \sim \mathcal{D}}[h(x) \neq y]$, and define $\mathsf{opt} := \min_{c \in {\mathcal{C}}} \mathsf{err}_{\mathcal{D}}(c)$. We say that ${\mathcal{C}}$ is agnostically learnable in time $T(n, \varepsilon, \delta)$ if there exists an algorithm which takes as input $n$ and $\delta$ and has access to an example oracle $\mathsf{EX}(\mathcal{D})$, and satisfies the following properties. It runs in time at most $T(n, \varepsilon, \delta)$, and with probability at least $1-\delta$, it outputs a hypothesis $h$ satisfying $\mathsf{err}_{\mathcal{D}}(h) \leq \mathsf{opt}+\varepsilon$. We say that the learning algorithm runs in \emph{subexponential time} if there is some constant $\eta > 0$ such that for any constants $\varepsilon$ and $\delta$, the running time $T(n, \varepsilon, \delta) \leq 2^{n^{1-\eta}}$ for sufficiently large $n$. The agnostic model is able to capture a range of realistic scenarios that do not fit within the standard PAC model. In many situations it is unreasonable to know exactly that $f$ belongs to some class ${\mathcal{C}}$, since $f$ may be computed by a process outside of our control. For example, the labels of $f$ may be (adversarially) corrupted by noise, resulting in a function that is no longer in ${\mathcal{C}}$. Alternatively, $f$ may be ``well-modeled,'' but not \emph{perfectly} modeled, by some concept in ${\mathcal{C}}$. In fact, the agnostic learning model even allows the input sample to not be described by a function $f$ at all, in the sense that the distribution over the sample may have both $(x,0)$ and $(x,1)$ in its support. This is also realistic when the model being used does not capture all of the variables on which the true function depends. \subsection{Related work} \label{sec:detailed} Since the agnostic PAC model generalizes the standard PAC model, it is (considerably) harder to learn a concept class in this model. Consequently, even fewer concept classes are known to be agnostically learnable, even in subexponential time. For example, as mentioned in \Cref{sec:introst}, the best known algorithm for agnostically learning the simple concept class of disjunctions, which are size-$1$, depth-$1$ Boolean circuits, runs in time\footnote{For simplicity, we suppress runtime dependence on $\varepsilon$ and $\delta$.} $2^{\widetilde{O}(\sqrt{n})}$~\cite{kkms}. In contrast, they can be learned in polynomial time in the PAC model~\cite{Valiant:1984}. Meanwhile, several hardness results are known for agnostically learning disjunctions, including NP-hardness for proper learning~\cite{KSS:1994}, and that even improper learning is as hard as PAC learning DNF~\cite{lbw}. While it is an important and interesting problem to agnostically learn more expressive classes of circuits in subexponential time, relatively few results are known. The best known general result is that all de Morgan formulas (formulas over the gate set of $\AND$, $\OR$, and $\NOT$ gates) of size $s$ can be learned in time $2^{\widetilde{O}(\sqrt{s})}$~\cite{kkms, Rei11}. In particular, linear-size formulas (i.e., $s=\Theta(n)$) can be learned in time $2^{\tilde{O}(\sqrt{n})}$, which is the same as the best known upper bound for disjunctions. Even in the relatively easier PAC model, only a small number of circuit classes are known to be learnable in subexponential time. For the well-studied class of polynomial-size DNFs, or depth-2 {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits, we have an algorithm running in time $2^{\widetilde{O}(n^{1/3})}$~\cite{klivansservediodnfs}, and we know that new techniques will be needed to improve this bound~\cite{razborovsherstov}. Little is known about larger subclasses of {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace, other than a recent paper that studied depth-3 {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits with top fan-in $t$, giving a PAC learning algorithm of runtime $2^{\tilde{O}(t\sqrt{n})}$~\cite{DRG17}, which is only subexponential when $t \ll \sqrt{n}$. Given the current state of affairs, a subexponential-time algorithm to learn all of {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace in the standard PAC model would represent significant progress. Indeed, for $d>2$, the fastest known PAC learning algorithm for depth-$d$ {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits runs in time $2^{n-\Omega(n/\log^{d-1}n)}$ \cite{nontriviallearn}, which is quite close to the trivial runtime of $2^n$. We view our new results for learning {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace and sublinear-size {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits as intermediate steps toward this goal. We clarify that our results are incomparable to the known results about agnostically learning de Morgan formulas. A simple counting argument \cite{nisanonline} shows that there are linear-size DNFs that are not computable by formulas of size $o(n^2/\log n)$, so one cannot learn even depth-2 {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace in subexponential time via the learning algorithm for de Morgan formulas. On the other hand, there are linear-size de Morgan formulas (of superconstant depth) that are not in {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace, or even {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace. \label{s:otherprior} Motivated by the lack of positive results in the distribution-free PAC learning model, \cite{nontriviallearn} study algorithms for learning various circuit classes, with the goal of ``only'' achieving a \emph{non-trivial savings} over trivial $2^n$-time algorithms. By achieving non-trivial savings, \cite{nontriviallearn} mean a runtime of $2^{n-o(n)}$; prior work had already connected non-trivial learning algorithms to circuit lower bounds \cite{klivans2013constructing, oliveira2016conspiracies}. The subexponential runtimes we achieve in our work are significantly faster than the $2^{n-o(n)}$-time algorithms of \cite{nontriviallearn}; in addition, our algorithms work in the challenging agnostic setting, rather than just the PAC setting. On the other hand, the algorithms of \cite{nontriviallearn} apply to more general circuit classes than {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace. As mentioned previously, \cite{klivansservediodnfs} gave a $2^{\widetilde{O}(n^{1/3})}$-time algorithm for PAC learning polynomial size DNF formulas; their algorithm is based on a $\widetilde{O}(n^{1/3})$ upper bound on the \emph{threshold degree} of such formulas. In unpublished work, \cite{talpersonal} has observed that the argument in \cite[Theorem 4]{klivansservediodnfs} can be generalized to show that for constant $d\geq 2$, any depth-$d$ {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuit has threshold degree at most $\widetilde{O}\bigl(n^{1-1/(3 \cdot 2^{d-3})}\bigr)$. This in turn yields a PAC learning algorithm for {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace running in time $\exp\bigl(\widetilde{O}\bigl(n^{1-1/(3 \cdot 2^{d-3})}\bigr)\bigr)$. Note that this is in the standard PAC model, not the agnostic PAC model. As mentioned in \Cref{sec:intro}, prior to our work, no subexponential time algorithm was known for agnostically learning even $\mathsf{LC}^0_3$\xspace in subexponential time. \subsection{Linear regression and the proof of \thm{learning}} \label{s:techniques} Our learning algorithm applies the well-known \emph{linear regression} framework for agnostic learning that was introduced by \cite{kkms}. The algorithm of \cite{kkms} works whenever there is a ``small'' set of ``features'' $\mathcal{F}$ (where each feature is a function mapping $\{0,1\}^n$ to ${\mathbb{R}}$) such that each concept in the concept class $\mathcal{C}$ can be approximated to error $\varepsilon$ in the $\ell_{\infty}$ norm via a linear combination of the features in $\mathcal{F}$. Roughly speaking, given a sufficiently large sample $S$ from an (unknown) distribution over $\{0,1\}^n \times \{0,1\}$, the algorithm finds a linear combination $h$ of the features of $\mathcal{F}$ that minimizes the empirical $\ell_1$ loss, i.e., $h$ minimizes $\sum_{(x_i, b_i)\in S} |h(x_i)-b_i|$ among all linear combinations of features from $\mathcal{F}$. An (approximately) optimal $h$ can be found in time $\poly(\mathcal{F})$ by solving a linear program of size $\poly(|\mathcal{F}|, |S|)$. \begin{lemma}[\cite{kkms}] \label{lem:kkms} Let $\mathcal{F}$ be a set of functions mapping $\{0,1\}^n$ to ${\mathbb{R}}$, and assume that each $\phi_i \in \mathcal{F}$ is efficiently computable, in the sense that for any $x \in \{0,1\}^n$, $\phi_i(x)$ can be computed in time $\poly(n)$. Suppose that for every $c \in \mathcal{C}$, there exist coefficients $\alpha_i \in \mathbb{R}$ such that for all $x \in \{0,1\}^n$, $|c(x) - \sum_{\phi_i \in \mathcal{F}} \alpha_i \cdot \phi_i(x)|\leq \varepsilon$. Then there is an algorithm that takes as input a sample $S$ of size $|S|=\poly(n, |\mathcal{F}|, 1/\varepsilon, \log(1/\delta))$ from an unknown distribution $\mathcal{D}$, and in time $\poly(|S|)$ outputs a hypothesis $h$ such that, with probability at least $1-\delta$ over $S$, $\Pr_{(x, b) \sim \mathcal{D}}[h(x) \neq b] \leq \varepsilon$. \end{lemma} A feature set $\mathcal{F}$ that is commonly used in applications of \lem{kkms} is the set of all monomials whose degree is at most some bound $d$. Indeed, an immediate corollary of \lem{kkms} is the following. \begin{corollary} \label{cor:thecor} \label{prop:agnosticfromadeg} Suppose that for every $c \in \mathcal{C}$, the $\varepsilon$-approximate degree of $c$ is at most $d$. Then for every $\delta > 0$, there is an algorithm running in time $\poly(n^{d}, 1/\varepsilon, \log(1/\delta))$ that agnostically learns $\mathcal{C}$ to error $\varepsilon$ with respect to any (unknown) distribution $\mathcal{D}$ over $\{0,1\}^n \times \{0,1\}$. \end{corollary} The best known algorithms for agnostically learning disjunctions and de Morgan formulas of linear size~\cite{kkms, Rei11} combine \cor{thecor} with known approximate degree upper bounds for disjunctions and de Morgan formulas of bounded size. We use the same strategy: our results for agnostic learning (\thm{learning}) follow from combining \cor{thecor} with our new approximate degree upper bounds. Specifically, \thm{approxdegLC} shows that the $\varepsilon$-approximate degree of any $\LCd$ circuit is at most $\tilde{O}(n^{1-2^{-d}} \log(1/\varepsilon)^{2^{-d}})$, yielding our new result for agnostically learning {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits. \thm{upper} shows that {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace circuits of size $s$ have $\varepsilon$-approximate degree $\tilde{O}(\sqrt{n} s^{1/2 - 2^{-d}}(\log(1/\varepsilon))^{2^{-d}})$, giving our new result for learning sublinear-size {$\mathsf{AC}^0$}\xspace} %\textsuperscript{0}}\xspace. Furthermore, since our upper bound on the approximate degree of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace circuits is nearly tight, new techniques will be needed to significantly surpass our results. In particular, new techniques will be needed to agnostically learn \emph{all} of {$\mathsf{LC}^{0}$}\xspace} %\textsuperscript{0}}\xspace in subexponential time. \Cref{thm:lower} implies that if $\mathcal{F}$ is the set of all monomials of at most a given degree $d$, then one cannot use \cor{thecor} to learn \LCd in time less than $2^{n^{1-2^{-\Omega(\sqrt{d})}}}$. However, standard techniques \cite{patmat} automatically generalize the lower bound of \Cref{thm:lower} from the feature set of low-degree monomials to \emph{arbitrary feature sets}. Specifically, we obtain the following theorem. \label{sec:agnostics} \begin{theorem} \label{thm:genF} Let $\mathcal{C}=\LCd$, and let $\mathcal{F}^*$ denote the minimum size set of features such that each $c \in {\mathcal{C}}$ can be approximated point-wise to error $1/3$ by a linear combination of the features in $\mathcal{F}$. Then $|\mathcal{F}^*| \geq 2^{n^{1-2^{-\Omega(\sqrt{d})}}}$. \end{theorem} For completeness, we provide the proof of \thm{genF} below. \begin{proof} For a matrix $F \in \{0,1\}^{N \times N}$, the $\varepsilon$-approximate rank of $F$, denoted $\text{rank}_{\varepsilon}(F)$, is the least rank of a matrix $A \in {\mathbb{R}}^{N \times N}$ such that $|A_{ij} - F_{ij}| \leq \varepsilon$ for all $(i, j) \in [N] \times [N]$. Sherstov's pattern matrix method \cite{patmat} allows one to translate in a black-box manner an approximate degree lower bound for a function $f$ into an approximate rank lower bound for a related matrix $F$, called the pattern matrix of $f$. Specifically, invoking \Cref{thm:lower}, let $f$ be the function in \LCdmone satisfying $\adeg(f) \geq D$ for some $D=n^{1-2^{-\Omega(\sqrt{d})}}$. Viewing $F$ as a $2^{4n} \times 2^{4n}$ matrix in the natural way, the pattern matrix method \cite[Theorem 8.1]{patmat} implies that the function $F \colon \{0,1\}^{4n}\times \{0,1\}^{4n} \to \{0,1\}$ given by $F(x, y) = f\left(\dots, \vee_{j=1}^4 \left(x_{i, j} \wedge y_{i, j}\right) \dots\right)$ satisfies \begin{equation} \label{eq:nomore} \text{rank}_{1/3}(F) \geq 2^{\Omega(D)},\end{equation} where the expression $\text{rank}_{1/3}(F)$ views $F$ as a $2^{4n} \times 2^{4n}$ matrix. Let $\mathcal{F}^*$ be a feature set satisfying the hypothesis of \Cref{thm:genF}, i.e., for every function $c \colon \{0,1\}^{4n} \to \{0,1\}$ in \LCd, there exist constants $\alpha_1, \dots, \alpha_{|\mathcal{F}|}$ such that \begin{equation} |c(x)-\sum_{\phi_j \in \mathcal{F}} \alpha_j \phi_j(x)| \leq 1/3 \end{equation} for all $x \in \{0,1\}^{4n}$. We claim that this implies that \begin{equation}\label{Really} \text{rank}_{1/3}(F) \leq |\mathcal{F}^*|. \end{equation} \Cref{thm:genF} then follows by combining Equation \eqref{Really} with Equation \eqref{eq:nomore}. Let us view each row $i$ of $F$ as a function $F_i$ mapping $\{0,1\}^{4n} \to \{0,1\}$. Then clearly, if $f$ is in \LCdmone, each row $F_i$ is in \LCd. Hence, there exist constants $\alpha_{i,1}, \dots, \alpha_{i,|\mathcal{F}|}$ such that \begin{equation} \label{anotherone} |F_i(x)-\sum_{\phi_j \in \mathcal{F}} \alpha_{i, j} \cdot \phi_j(x)| \leq 1/3 \text{ for all } x \in \{0,1\}^{4n}. \end{equation} Let $M$ denote the $2^{4n} \times |\mathcal{F}|$ matrix whose $i, j$'th entry is $\alpha_{i, j}$. And let $R$ denote that $|\mathcal{F}| \times 2^{4n}$ matrix whose $(j, x)$'th entry is $\phi_{j}(x)$, where we associate $x$ with an input in $\{0,1\}^{4n}$. Then Equation \eqref{anotherone} implies that $|M \cdot R - F_{ij}| \leq 1/3$ for all $(i, j) \in [2^{4n}] \times [2^{4n}]$. Since $M \cdot R$ is a matrix of rank at most $|\mathcal{F}|$, Equation \eqref{Really} follows. \end{proof} \section{Circuit Lower Bounds (Proof of \Cref{apptheorem})} \label{s:app} As a warmup, we start by establishing a worst-case version of \Cref{apptheorem}. \begin{proposition} \label{worstcaseprop} The Inner Product function on $n$ bits cannot be computed by any depth-$(d+1)$ \ACmodtwo\ circuit of size $\widetilde{\Omega}\bigl(n^{1/(1-2^{-d})}\bigr)$. \end{proposition} \begin{proof} \Cref{thm:approxdegLC} shows that any depth-$d$ AC$^0$ circuit of size $s \geq n$ on $n$ inputs has approximate degree at most $D= \tilde{O}(s^{1-2^{-d}})$. Clearly, the approximating polynomial has at most ${s \choose D}$ many monomials. From this, one can conclude that any depth-$(d+1)$ \ACmodtwo\ circuit ${\mathcal{C}}$ on $n$ inputs of size $s \geq n$ can be approximated by a polynomial $p$ over $\{-1, 1\}^n$ with at most ${s \choose D}$ many monomials. To see why, let $q$ approximate the ``AC$^0$ part'' of ${\mathcal{C}}$. This is an AC$^0$ circuit of depth $d$ and size $s$ on at most $s$ inputs, and hence has approximate degree at most $D$. Now obtain $p$ by replacing each input to $q$ with the corresponding $\oplus$ gate of ${\mathcal{C}}$. Since $q$ is defined over domain $\{-1, 1\}^n$, replacing any input to $q$ with an XOR function preserves the number of monomials of $q$. On the other hand, it is known that well-known that any polynomial $p$ over $\{-1, 1\}^n$ that uniformly approximates the Inner Product function (to any error strictly less than 1) requires $2^{\Omega(n)}$ many monomials \cite{bruck1992polynomial}. Combining the above two facts means that ${s \choose D}$ must be at least $2^{\Omega(n)}$, which means that $s$ must be at least $\tilde{\Omega}(n^{1/(1-2^{-d})})$. \end{proof} We now prove \Cref{apptheorem}, restated here for convenience. \apptheorem* \paragraph{Proof Outline.} The proof follows a similar outline to \Cref{worstcaseprop}, but builds on an observation of Tal \cite[Lemma 4.2]{tal2016bipartite}. Roughly, Lemma 4.2 of \cite{tal2016bipartite} shows that bipartite de Morgan formulas of size $s$ cannot compute the Inner Product function on more than a $1/2 + n^{- \log n}$ fraction of inputs unless they have size at least roughly $n^2$. The only property of de Morgan formulas of size $\ll n^2$ that Tal uses is that they have sublinear approximate degree. Similarly, \Cref{thm:approxdegLC} shows that an AC$^0$ circuit of size $s$ and depth $d$ on $n$ inputs, for which $n \leq s \ll n^{1/(1-2^{-d})}$, has sublinear approximate degree. Any parity function is an example of a bipartite function of size $O(1)$, meaning that the parity function applied to some subset of an input $(x, y) \in \{-1, 1\}^n \times \{-1, 1\}^n$ is computable by a constant-sized circuit with leaves computing a function of only $x$ or $y$. Hence, Tal's argument applies with cosmetic changes not only to sub-quadratic size bipartite de Morgan formulas, but also to \ACmodtwo\ circuits of size $s \ll n^{1/(1-2^{-d})}$. We remark that the entire argument (and hence the lower bound of \Cref{apptheorem} itself) applies not only to \ACmodtwo\ circuits, but more generally to depth-$d$ AC$^0$ circuits augmented with a layer of low-communication gates above the inputs; we omit this extension for brevity. \begin{proof}[Proof of \Cref{apptheorem}, closely following the proof of Lemma 4.2 of \cite{tal2016bipartite}] Let $$\mathsf{IP}(x, y) = \oplus_{i=1}^n (x_i \wedge y_i)$$ denote the Boolean inner product on $2n$ bits. Let ${\mathcal{C}} \colon \{-1, 1\}^{2n} \to \{-1, 1\}$ be an \ACmodtwo\ circuit of depth $(d+1)$ and size $s \geq n$, and let $$q = \Pr_{x, y \in \{-1, 1\}^n}[{\mathcal{C}}(x, y) = \mathsf{IP}(x, y)].$$ Suppose that $q \geq 1/2 + \varepsilon$. Our goal is to show that $s$ must be large, even for negligible values of $\varepsilon$. Let $N \leq s$ denote the number of parity gates in ${\mathcal{C}}$, with the $i$th parity gate denoted by $h_i(x) \colon \{-1, 1\}^n \to \{-1, 1\}$. Let ${\mathcal{C}}'$ denote just the AC$^0$ part of the circuit, i.e., ${\mathcal{C}}(x, y) = {\mathcal{C}}'(h_1(x, y), \dots, h_N(x, y))$. Note that ${\mathcal{C}}'$ is an AC$^0$ circuit on at most $s$ inputs, of depth $d$ and size at most $s$. By \Cref{thm:approxdegLC}, there exists a polynomial $p$ of degree at most $D \leq \tilde{O}\left(s^{1-2^{-d}} \log(1/\varepsilon)^{2^{-d}}\right)$ such that, for all $w \in \{-1, 1\}^N$, $|p(w)-{\mathcal{C}}'(w)| \leq \varepsilon$. Next, we show that under the uniform distribution, the function $\mathsf{IP}(x, y)$ correlates well with $p(h_1(x), \dots, h_N(x))$. We decompose the expectation $\mathbf{E}_{x, y \in \{-1, 1\}^n}[p(x, y) \cdot \mathsf{IP}(x, y)]$ according to whether or not $\mathsf{IP}(x, y) = {\mathcal{C}}(x, y)$: \begin{flalign} \mathbf{E}_{x, y \in \{-1, 1\}^n}[p( h_1(x, y), \dots, h_N(x, y)) \cdot \mathsf{IP}(x, y)] = \notag \\ \mathbf{E}_{x, y \in \{-1, 1\}^n}[p( h_1(x, y), \dots, h_N(x, y)) \cdot \mathsf{IP}(x, y) | \mathsf{IP}(x, y) = {\mathcal{C}}(x, y)] \cdot \Pr[\mathsf{IP}(x, y) = {\mathcal{C}}(x, y)] + \notag \\ \mathbf{E}_{x, y \in \{-1, 1\}^n}[p( h_1(x, y), \dots, h_N(x, y)) \cdot \mathsf{IP}(x, y) | \mathsf{IP}(x, y) \neq {\mathcal{C}}(x, y)] \cdot \Pr[\mathsf{IP}(x, y) \neq {\mathcal{C}}(x, y)] \notag \\ \geq (1 - \varepsilon) \cdot q + (-1 - \varepsilon) \cdot (1 - q) \notag \\ = 2q - 1 - \varepsilon \geq 2 \cdot (1/2 + \varepsilon) - 1 - \varepsilon = \varepsilon. \label{eq:corrbound} \end{flalign} Next, we write $p(z)$ as a multi-linear polynomial: $p(z) = \sum_{S \subseteq [N], |S| \leq D} \hat{p}(S) \cdot \prod_{i \in S} z_i$. Since $\hat{p}(S) = \mathbf{E}_{z \in \{-1,1\}^N} [p(z) \cdot \prod_{i \in S} z_i]$, we have that $|\hat{p}(S)| \leq 1 + \varepsilon$ for every $S$. Note that there are at most ${N \choose D}$ monomials in $p$. Invoking Inequality \eqref{eq:corrbound}, we have: \begin{align*} \varepsilon &\leq \mathbf{E}_{x, y \in \{-1, 1\}^n}\left[p(h_1(x, y), \dots , h_N(x, y)) \cdot \mathsf{IP}(x, y)\right]\\ &=\mathbf{E}_{x, y \in \{-1, 1\}^n}\left[\sum_{S \subseteq [N], |S| \leq D} \hat{p}(S) \prod_{i \in S} h_i(x, y) \cdot \mathsf{IP}(x, y)\right]\\ &=\sum_{S \subseteq [N], |S|\leq D} \hat{p}(S) \cdot \mathbf{E}_{x, y \in \{-1, 1\}^n}\left[\prod_{i \in S} h_i(x, y) \cdot \mathsf{IP}(x, y)\right]\\ &\leq \sum_{S \subseteq [N], |S|\leq D} (1+\varepsilon) \left| \mathbf{E}_{x, y \in \{-1, 1\}^n}\left[\prod_{i \in S} h_i(x, y) \cdot \mathsf{IP}(x, y)\right]\right|. \end{align*} Hence there must exist a set $S \subseteq [N]$ with size at most $D$ such that $$\left| \mathbf{E}_{x, y \in \{-1, 1\}^n}\left[\prod_{i \in S} h_i(x, y) \cdot \mathsf{IP}(x, y)\right]\right| \geq \frac{\varepsilon}{{N \choose D}\left(1+\varepsilon\right)} \geq (\varepsilon/2) \cdot s^{-D} \geq \exp\left({\tilde{O}(-s^{1-2^{-d}} \log^{2^{-d}}(1/\varepsilon))}\right).$$ It is well-known that $\mathsf{IP}$ is $2^{-\Omega(n)}$ correlated with any parity function $h_i$ (indeed, $\mathsf{IP}$ on $2n$ bits is a \emph{bent} function, meaning that all its Fourier coefficients have magnitude $2^{-n}$, and hence its correlation with any parity is at most $2^{-n}$). We conclude that $$s^{1-2^{-d}} \log^{2^{-d}}(1/\varepsilon) \geq \tilde{\Omega}(n).$$ The theorem is an immediate consequence of this inequality. \end{proof} \section*{Acknowledgements} We thank Nikhil Mande, Ronald de Wolf, and Shuchen Zhu for comments on earlier drafts of this paper. R.K. thanks Luke Schaeffer for comments on the proof of \thm{quantum}. \bibliographystyle{alphaurl} \phantomsection\addcontentsline{toc}{section}{References} \renewcommand{\UrlFont}{\ttfamily\small} \let\oldpath\path \renewcommand{\path}[1]{\small\oldpath{#1}}

63888f90e7ccd3d0e8d37a3cbcf5f0a0b481218e

\section{Introduction} The allure of fractals lies not only in their aesthetic beauty and the mathematical beauty of their self-similarity, but also in that such complexity can be achieved by very simple chaotic equations. Nature seemingly utilises this as an engineering tool, with symmetric fractal structures appearing in many diverse forms, from romanesco broccoli to ammonite sutures and ferns, while statistical fractals are seen in salt flats, mountains, coastlines and clouds. Popularised by Benoit Mandelbrot \cite{Mandelbrot-1982}, fractals can be thought of the the mathematical instance of ``plus ca change, plus c'est la meme chose'' (``the more things change, the more they stay the same''). Light too can be fractal. The (dark) vortex lines in random light fields have fractal scaling properties \cite{OHolleran-et-al-2008}, and light's spatial (and spectral \cite{Lehman-Garavaglia-1999}) distribution can be directly made fractal by interaction with a fractal object, for example by emitting it from a fractal antenna \cite{Werner-Ganguly-2003}, by passing it through a fractal aperture \cite{Lehman-2001,Saavedra-et-al-2003}, or by resonating it in a cavity that contains a fractal scatterer \cite{Takeda-et-al-2004}. Perhaps more surprisingly, due to the fractal Talbot effect the light field behind a (non-fractal) Ronchi grating illuminated by a uniform plane wave evolves, on propagation, into a fractal~\cite{Berry-Klein-1996a,Berry-et-al-2001}. A glance at intensity cross-sections through the eigenmodes of unstable canonical resonators (e.g.\ \cite{Siegman-1986-unstable-resonator-eigenmode-distributions}) reveals complex and fractal-looking structure, but the first suggestion that these eigenmodes are fractals came only in 1998 \cite{Karman-Woerdman-1998,Karman-et-al-1999}. This is surprising, as canonical resonators are very simple, consisting of a pair of spherical mirrors and any apertures in the resonator. Initially, the discussion of the mechanism involved mostly the round-trip magnification due to geometrical imaging by the spherical mirrors, which leads to similar patterns to appear at a cascade of different length scales, one of the hallmarks of fractals, but it also hinted at the role of diffraction, which gives rise to the ripples in the pattern in the first place \cite{New-et-al-2001}. Clearly, without diffraction, successive magnifications would simply make any initial pattern increasingly uniform. Detailed theoretical studies of these intensity distributions found them to be statistical fractals \cite{Karman-Woerdman-1998,Karman-et-al-1999,Berry-2001,Berry-et-al-2001,New-et-al-2001,Yates-New-2002,Loaiza-2005,Huang-et-al-2006}. Upon magnification, statistical fractals look like the same \emph{type} of pattern, but not actually the same pattern. Like in all physical fractals, the range of length scales over which this scaling behaviour holds (the scaling range) is limited \cite{Avnir-et-al-1998}, here by diffraction. \begin{figure}[ht] \begin{center} \includegraphics[width=1\linewidth]{resonator.pdf} \end{center} \caption{Imaging inside an unstable canonical resonator. The two spherical mirrors, M$_1$ (focal length $f$) and M$_2$ (focal length $F$), perform geometric imaging. We define two longitudinal coordinates, $z$ and $Z$. Both the $z$ and the $Z$ axes coincide with the optical axis, but the $z=0$ plane coincides with the plane S, the magnified self-conjugate plane, and the $Z=0$ plane coincides with the plane of the mirror M$_2$. We use $z$ in our theoretical analysis, $Z$ in the experimental part. The transverse coordinates are $x$ (not shown) and $y$. Three-dimensional imaging during one round trip is indicated by an object in the shape of the letter ``P'' (shown in black) and its image, which looks like a horizontally elongated letter ``b'' (shown in grey): the ``P'' has turned into a ``b'' because the transverse magnification, $M$, is negative, and so the image of the ``P'' is upside-down; the ``b'' is horizontally elongated because the longitudinal magnification, $M_l$, is positive and its magnitude is greater than that of the transverse magnification. $A$ is an aperture immediately in front of M$_2$. The figure is drawn for a the particularly simple case of a confocal resonator (length $F + f$; the plane S then coincides with the common focal plane) with $M = -2$ and $M_l = +4$.} \label{resonator-figure} \end{figure} Shortly after the original explanation for the fractal character of the eigenmodes of unstable resonators it was found that the role of diffraction was particularly simple in the plane that is geometrically imaged into itself with a magnification $M$ of modulus $|M| >1$, the \emph{magnified self-conjugate plane} \cite{Courtial-Padgett-2000b,Watterson-et-al-2003}. In this plane, shown in Fig.\ \ref{resonator-figure}, the intensity distribution is a diffraction-limited self-similar fractal \cite{Courtial-Padgett-2000b}, with an example shown in Fig.\ \ref{transverse-intensity-distribution-3D-resonator-figure}. The mechanism, called the \emph{monitor-outside-a-monitor effect} after a video-feedback analogy \cite{Courtial-et-al-2001,Leach-et-al-2003}, is that each round trip through the resonator, starting and finishing in the magnified self-conjugate plane, can be approximated as simple scaling by a factor $M$ of the initial beam and addition of the aperture diffraction pattern under spherical-wave illumination. \begin{figure}[ht \begin{center} \includegraphics[width=1\linewidth]{SCrossSection.pdf} \end{center} \caption{\label{transverse-intensity-distribution-3D-resonator-figure}Self-similarity of the simulated lowest-loss eigenmode's intensity distribution in the magnified self-conjugate plane, S. The frames show the intensity after 20 round trips, starting with a uniform plane wave. The self-similarity of the pattern is demonstrated by showing its centre at different magnifications ($\times 2$, $\times 4$, $\times 8$), resulting in a similar pattern (rotated by $180^\circ$ after each magnification by a factor 2 due to the resonator's transverse magnification, $M$, being negative); the dotted white square in the centre of the frame marked $\times 1$ shows the outline of the area shown in the next frame. The horizontal dotted line is the orthographic projection of the lateral self-conjugate plane $S_L$, in which we demonstrate, below (see Fig.\ \ref{lateral-intensity-distribution-3D-resonator-figure}), the 3D self-similarity of the light. The figure is calculated for light of wavelength $\lambda = 632.8$\,nm in a resonator of the type shown in Fig.\ \ref{resonator-figure} with $F=16.5$\,cm, $f=8.25$\,cm, $M = -2$, and a seven-sided regular polygonal aperture of circumradius $r_0 = 2.4$\,mm. The beam's transverse cross-sections were represented by a $1024 \times 1024$ array of complex numbers sampled over a physical area of size 1\,cm$\times$1\,cm. For further details of the way this simulation, and indeed all the other simulations in this paper, was performed see App.~\ref{a:simulations}. } \end{figure} After magnification, a part of a self-similar pattern looks not just to be the same pattern type as a corresponding, unscaled, part the pattern, but \emph{the same}. Note that, with all physical fractals, this is only true over a finite range of sizes, here limited by the smallest size allowed by diffraction and the overall size of the beam. Suitable choice of the resonator parameters has been predicted to lead to intensity distributions closely related to classic fractals such as the Weierstrass-Mandelbrot function, the Sierpinski gasket, and the Koch snowflake~\cite{Watterson-et-al-2003,Courtial-Padgett-2000b}. Despite these early advances, experimental observations have been scarce. A pulse of light was injected into a passive canonical cavity and observed to evolve over a number of round trips into a fractal pattern \cite{Loaiza-et-al-2003}; curiously, that work was never published in a peer-reviewed journal. Very recently, small areas containing fractal structure were found in the eigenmode of a non-canonical resonator comprising an array of microspheres sandwiched between planar mirrors \cite{Rivera-et-al-2018} --- the first observation of fractal structure in the eigenmode of an (active) laser. These are the works that are most relevant to this study, but the relevance is limited as they either worked in a passive cavity and did not study the eigenmode in the magnified self-conjugate plane \cite{Loaiza-et-al-2003}, or in a different laser configuration altogether \cite{Rivera-et-al-2018}. Further, the discussion was entirely limited to the light structure in transverse planes, resulting in what we will refer to as 2D fractals\footnote{Note that the dimension of the plane is, in general, different from the fractal dimension of such intensity distributions~\cite{Berry-et-al-2001}.}. Here we experimentally verify the existence of self-similar fractal light from canonical lasers by observing the 2D intensity structure of laser light at the magnified self-conjugate plane inside the cavity and studying its self-similarity directly, rather than through the fractal dimension. Further, we show that fractals can form in the three-dimensional (3D) intensity distributions of light in unstable canonical resonators. We find that, around the centre of the magnified self-conjugate plane, this intensity distribution in 3D space is a self-similar fractal, albeit with different transverse and longitudinal characteristic scaling factors. While we outline this structure in 3D space theoretically, the experimental verification remains an open task. \section{Theory} \subsection{\label{s:transverseFractals}Transverse Fractals} We start by reviewing the mechanism that leads to self-similar fractal structure at the self-conjugate plane in an unstable canonical resonator. Without loss of generality, we restrict ourselves to confocal resonators, as these are particularly simple but at the same time representative of all canonical unstable resonators (with the same round-trip magnification, $M$, and the same Fresnel number \cite{Siegman-1986}). Consider the example shown in Fig.~\ref{resonator-figure}. In such a resonator, each mirror is spherical and so images like a lens, but in reflection. During one round trip, i.e.\ reflection off both mirrors, the image produced by the first mirror is imaged again by the second mirror. In stable resonators this imaging explains the eigenmodes' structural stability \cite{Forrester-et-al-2002}. In unstable canonical resonators, one round trip images two ``self-conjugate'' planes back to their original positions, one with (transverse) magnification $M$ ($|M| \geq 1$), the other with magnification $1/M$ \cite{Courtial-Padgett-2000b}. The former is the magnified self-conjugate plane, S, the latter the de-magnified self-conjugate plane, s. In a confocal resonator, these planes are a focal distance on either side of the two mirrors (see Fig.\ \ref{resonator-figure}), and so the field in these planes forms a Fourier pair. Geometrical imaging stretches, during every round trip through the resonator, the intensity distributions in the planes S and s by a factor $M$ and $1/M$, respectively. Any apertures in the resonator simply add some diffractive ``decoration'' to this image. After a number of round trips, the pattern is essentially unchanged between successive round trips (the complex amplitude cross-section is unchanged apart from complex factor representing a uniform phase change and loss), which means the field has settled into an eigenmode. In our case, the lowest-loss eigenmode is reached after approx.\ 20 round trips. Once the eigenmode has formed, the decoration pattern is the same during successive round trips. Once added, it gets magnified with the rest of the intensity distribution, which results in the presence of the decoration pattern in a number of sizes: the pattern added during the most recent round trip is at the original size; that added during the previous round trip is magnified by $M$; that added two round trips ago is magnified by $M^2$; and so on. The presence of a pattern on such a cascade of length scales is a hallmark of self-similarity. The mechanism outlined above is called the monitor-outside-a-monitor effect (MOM effect), named so because of analogies with video feedback~\cite{Courtial-et-al-2001,Leach-et-al-2003}. \subsection{3D Fractals} For the same resonator, Fig.~\ref{lateral-intensity-distribution-3D-resonator-figure} shows a lateral intensity distribution around the centre of the self-conjugate plane S. This lateral intensity distribution shows some signs of self-similarity: if the pattern is stretched by $M$ in the direction representing the transverse direction, and by a factor $M^2$ in the longitudinal direction, the pattern's centre (which is the point where the plane S intersects the resonator's optical axis) looks similar to what it was before magnification. \begin{figure}[ht] \begin{center} \includegraphics[width=1\linewidth]{SLongitudinalCrossSection.pdf} \end{center} \caption{Self-similarity of the intensity distribution in a lateral self-conjugate plane, $S_L$, which contains the optical axis and intersects the plane S horizontally in Fig.~\ref{transverse-intensity-distribution-3D-resonator-figure}. The vertical dotted line is the orthographic projection of the plane S. Vertically, the plots are centred on the optical axis, $z$. The beam is the same as that shown in Fig.\ \ref{transverse-intensity-distribution-3D-resonator-figure}. After each magnification horizontally by a factor 4 and vertically by $-2$, the pattern looks similar again, which is shown for different magnifications. The dotted box in the centre of the frame marked $\times 1$ marks the outline of the area shown in the next frame ($\times (-2)$ vertically, $\times 4$ horizontally). The $\times 1$ frame represents a physical area of size 2\,m (horizontally) by 10\,mm (vertically).} \label{lateral-intensity-distribution-3D-resonator-figure} \end{figure} \begin{figure}[ht] \begin{center} \includegraphics{SLongitudinalCrossSectionStripResonator.pdf} \end{center} \caption{Self-similarity of the intensity distribution in the lateral plane of a strip resonator of the type shown in Fig.\ \ref{resonator-figure}. The different frames show the centre of the intensity distribution, successively magnified by a factor $M$ in the vertical direction and by $M^{2}$ in the horizontal direction. The $\times 1$ frame represents a physical area of size 20\,m (horizontally) by 2.82\,cm (vertically), centred on the magnified self-conjugate plane and the optical axis in the horizontal and vertical direction, respectively. The dotted box shown in the top left frame outlines the area shown in the top right frame. The figure was calculated for light of wavelength $\lambda = 632.8$\,nm, resonator parameters $F = 70.7$\,cm and $f = 50$\,cm (corresponding to transverse magnification $M = -\sqrt{2}$), and the aperture $A$ was a slit of width 2.08\,cm. Each beam cross-section was represented on 4096-element array of complex numbers, representing a physical width 4\,cm. } \label{lateral-intensity-distribution-2D-resonator-figure} \end{figure} \begin{figure}[ht] \begin{center} \includegraphics[width=\linewidth]{LinePlotsStripResonator.pdf} \end{center} \caption{Axial~(a) and transverse~(b) intensity cross-section through the field around the self-conjugate point at the centre of the plane S in the strip resonator from Fig.\ \ref{lateral-intensity-distribution-2D-resonator-figure}. Like in Figs \ref{transverse-intensity-distribution-3D-resonator-figure}, \ref{lateral-intensity-distribution-3D-resonator-figure} and \ref{lateral-intensity-distribution-2D-resonator-figure}, the self-similarity is demonstrated by successive magnifications, each of which stretches the part of the curve between the vertical dotted lines to the full width. The width of the curves marked $\times 1$ represents a physical length of 2\,m (a) and 2.08\,cm (b). The intensity range represented by the different curves has been adjusted so that corresponding features in the curves have roughly the same vertical size.} \label{line-intensity-distributions-2D-resonator-figure} \end{figure} This self-similarity can be seen much clearer in Fig.~\ref{lateral-intensity-distribution-2D-resonator-figure}, which was calculated for a strip resonator, i.e.\ a resonator that is invariant in one transverse direction. It can therefore be treated as a 2D resonator with only one transverse direction, which means that, along that transverse direction, the light field can be represented in computer simulations by a much greater number of grid points without increasing memory or complexity requirements. This in turn allows an increase in the Fresnel number by increasing the aperture size, resulting in a lateral intensity cross-section with significantly more detail. For that same strip resonator, Fig.~\ref{line-intensity-distributions-2D-resonator-figure} compares the intensity cross-sections along the transverse direction in the plane S with that along the resonator's optical axis. The intensity cross-section along the optical axis is not symmetrical with respect to the position of the plane S, whereas that in the plane S is symmetric with respect to the position of the optical axis. Irrespective of this complication, both curves are strikingly self-similar. This observation can be explained as follows. Spherical mirrors (and lenses) image not only any transverse plane into a corresponding transverse plane, they image any point into a corresponding point. For light initially travelling to the right in the resonator shown in Fig.~\ref{resonator-figure}, any lateral plane that includes the optical axis is being imaged into itself, as is the magnified self-conjugate plane S; no other planes are being imaged into themselves (but other surfaces are, specifically the paraboloids $z = a r^2$, where $z$ and $r$ are cylindrical coordinates as shown in Fig.\ \ref{resonator-figure} and $a$ is an arbitrary constant). One point is imaged into itself (``self-conjugate point''), namely the intersection of the self-conjugate plane S with the optical axis. The volume around this point is also imaged into itself, but the image is distorted as the longitudinal and transverse magnifications are different (the longitudinal magnification is the square of the transverse magnification) and both change with position. (Similar statements are true for light initially traveling to the left, but we do not consider these here.) Close to the self-conjugate point, the longitudinal magnification is constant. This imaging of the volume around the centre of the plane S is indicated in Fig.~\ref{resonator-figure}. As before, the effect of any apertures in the system is the addition of a diffractive decoration pattern, which is now 3D. In a 3D extension of the MOM effect, this pattern gets added to the field during each round trip and magnified during each subsequent round trip, again resulting in its presence on a cascade of length scales, complicated and enriched by the different characteristic stretch factors in the longitudinal and transverse directions. \subsection{Self-similarity of transverse fractals} The mechanism for the emergence of fractals in the transverse intensity cross-sections, described in section \ref{s:transverseFractals}, suggests that the cross section is most self-similar in the magnified self-conjugate plane, but this has never been tested quantitatively. Here we provide the first quantitative evidence for this argument. For the intensity cross-section in one transverse plane at a time we calculate the normalised squared Euclidean distance, $d^2$, between the centre of this intensity cross-section and the centre of the same intensity cross-section, stretched by a factor $M$ (see App.\ \ref{a:nsed} for details). This is a measure of the difference between the stretched and unstretched centre of the intensity cross-section, and $-d^2$ is therefore a measure of their similarity. We then plot $-d^2$ as a function of the $z$ coordinate of the transverse plane, defined as the Cartesian coordinate aligned with the optical axis such that the plane $z=0$ is the magnified self-conjugate plane (Fig.\ \ref{resonator-figure}). From the above argument we expect a peak at $z = 0$, that is, in the magnified self-conjugate plane~S. \begin{figure}[ht] \begin{center} \includegraphics{selfSimilarity.pdf} \end{center} \caption{\label{f:selfSimilarity}Evolution of the self-similarity of transverse intensity cross-sections upon propagation for the eigenmodes shown in Figs \ref{transverse-intensity-distribution-3D-resonator-figure} and \ref{lateral-intensity-distribution-3D-resonator-figure} (a) and in Figs \ref{lateral-intensity-distribution-2D-resonator-figure} and \ref{line-intensity-distributions-2D-resonator-figure}~(b). The self-similarity is quantified here by $-d^2$, the negative of the normalised squared Euclidean distance between the central rectangle of the intensity cross-section in the transverse plane with the given $z$ coordinate, and the central rectangle of the same size of the intensity cross-section after being stretched in the transverse directions by a factor $M$. The magnification was $M = -3$ in (a) and $M=-2$ in~(b). The width and height of the central rectangle were arbitrarily chosen to be $1/4$ of the width and height of the calculated intensity cross-section. The dotted vertical lines indicate the planes of the mirrors, M$_1$ and M$_2$, and the magnified self-conjugate plane, S.} \end{figure} Fig.\ \ref{f:selfSimilarity} shows such curves, calculated for the two eigenmodes discussed earlier, namely that of a resonator with a heptagonal aperture and transverse magnification $M=-3$ (Figs \ref{transverse-intensity-distribution-3D-resonator-figure} and \ref{lateral-intensity-distribution-3D-resonator-figure}), and that of a strip resonator with transverse magnification $M=-\sqrt{2}$ (Figs \ref{lateral-intensity-distribution-2D-resonator-figure} and \ref{line-intensity-distributions-2D-resonator-figure}). The expected peak at $z = 0$ is clearly visible and, especially in the case of the strip resonator, reaches close to $d^2 = 0$, proving the near-exact --- but diffraction-limited --- self-similarity of the intensity cross-section in the magnified self-conjugate plane. Note that these results can be replicated with other measures of difference between images, and we did this with Euclidean distance and image Euclidean distance \cite{Wang-et-al-2005a} (IMED, which we calculated only for the strip-resonator eigenmode). These gave a curve with a different shape to those shown in Fig.\ \ref{f:selfSimilarity}, but always produced a very prominent peak in the magnified self-conjugate plane. \begin{figure}[ht] \begin{center} \includegraphics[width=\columnwidth]{setup2.pdf} \end{center} \caption{Experimental setup, (a)~schematic and (b)~photo with the laser running. (a)~The cavity comprised two concave mirrors, M$_1$ (radius of curvature $R_1$, corresponding to focal length $f = R_1/2$) and M$_2$ (radius of curvature $R_2$, focal length $F = R_2/2$), and an output coupler (OC) angled at 45$^\circ$ with a 99.8\% reflectivity. The geometrical length of the cavity was $G$, that of the Nd:YAG gain medium was $g$. S is the magnified self-conjugate plane. A polygonal aperture (A) was positioned in front of M$_2$. The intensity cross-section in the self-conjugate plane S was captured outside the cavity at a distance $R_2/2$ from M$_2$ using a CCD camera. (b)~Clearly visible is the gain-medium assembly (bright yellow; centre) and the mounts for the end mirrors and the output coupler. The two tubes (light blue; bottom left) connecting to the gain-medium assembly are cooling tubes for the gain medium. } \label{setup} \end{figure} \section{Experiment} \label{s:experiment} We constructed a laser, sketched in Fig.~\ref{setup}, consisting of a flash-lamp pumped Nd:YAG gain medium (6.35\:mm $\times$ 76\:mm) inside an L-shaped, confocal, unstable cavity comprising two concave, high-reflectivity, spherical end mirrors, M$_1$ and M$_2$, and a 45$^\circ$ output coupler positioned at the apex of the L. The radius of curvature of mirror M$_1$ is $R_1$, that of M$_2$ is $R_2$; their focal lengths are respectively $f = R_1/2$ and $F = R_2/2$. The geometrical length of the resonator is $G$. An aperture in the shape of a regular hexagon was positioned in front of the end mirror with the greater focal length, M$_2$. Like that sketched in Fig.\ \ref{resonator-figure}, our cavity is canonical and unstable, and as such it contains magnified and de-magnified self-conjugate planes. Unlike that sketched in Fig.\ref{resonator-figure}, our cavity was not confocal, and so the positions of the self-conjugate planes did not simply coincide with the focal planes, but their $Z$ coordinate was calculated from the geometric-imaging properties of the cavity (see App.\ \ref{a:S-plane}). The output beam was captured using a CCD camera (Spiricon SPU260 BeamGage), placed in an image plane of S. Note that the field depends on the propagation direction (even in stable canonical resonators \cite{Naidoo-et-al-2012}), and the camera has been placed to record the image in the magnified self-conjugate plane, S, not the de-magnified self-conjugate plane, s, which corresponds to the same plane but the opposite propagation direction. One of the intensity cross-sections observed in the magnified self-conjugate plane is shown as part of Fig.\ \ref{setup}(a). It has the expected complex structure characteristic of a fractal. In almost all images we note an unexpectedly bright central intensity peak. We speculate that it is due to a low divergence mode that is also able to lase in the cavity. \begin{figure*}[ht] \begin{center} \includegraphics[width=\linewidth]{fracZ.pdf} \end{center} \caption{Evolution of the eigenmode intensity cross-section between mirrors M$_2$ ($Z = 0\,\mathrm{mm}$) and M$_1$ ($Z = 343\,\mathrm{mm}$). The magnification was $M$ = 3.32 ($F=250$ \,mm, $f = 75$\,mm) with an aperture of circumradius $r_0 = 2\,$mm. The geometrical length of the cavity was $G = 343$\,mm, corresponding to an effective length of $L = 308.5$\,mm due to the refractive index of the gain medium (see App.~\ref{a:S-plane}).} \label{fracz} \end{figure*} \begin{figure}[ht] \begin{center} \includegraphics[width=\linewidth]{frac3.pdf} \end{center} \caption{A few of the transverse intensity patterns recorded inside a resonator for parameters $F = 250\,\mathrm{mm}$, $f = 75\,\mathrm{mm}$, and $G = 343\,\mathrm{mm}$. The self-conjugate plane S is positioned at $Z = Z_S = 231.8\,\mathrm{mm}$. } \label{frac3} \end{figure} For one set of parameters, namely $F = 250\,\mathrm{mm}$, $f = 75\,\mathrm{mm}$, $G = 343\,\mathrm{mm}$, and a hexagonal aperture A of circumradius $r_0 = 2.5\,\mathrm{mm}$ we recorded intensity cross-sections in a number of transverse planes across the resonator (Fig.\ \ref{fracz}), with examples shown in Fig.~\ref{frac3}. The magnified self-conjugate plane is located at $Z_S = 231.8\,\mathrm{mm}$ (calculated using Eqn [\ref{e:Z_S}]), the round-trip transverse magnification is $M = -3.32$~(calculated from Eqn~[\ref{e:M}]). \begin{figure*}[ht] \begin{center} \includegraphics{selfSimilarityZ23To26.pdf} \end{center} \caption{\label{f:selfSimilarityZ23To26}Self-similarity of the intensity cross-sections in different transverse planes near the predicted position of the self-conjugate plane, $Z = Z_S = 231.8\,\mathrm{mm}$. The images in the top row show the central $\approx 4\,\mathrm{mm} \times 4\,\mathrm{mm}$ of the experimentally recorded intensity cross-sections in the transverse planes $Z = 220\,\mathrm{mm}$~(left), $Z = 230\,\mathrm{mm}$~(center), and $Z = 240\,\mathrm{mm}$~(right). The images in the bottom row show the centers of the corresponding top-row images, stretched by a factor $M = -3.32$, that is, stretched by a factor $3.32$ and rotated by $180^\circ$. To show structure over a wider intensity range, the brightness of each pixel is proportional to the logarithm of the recorded intensity.} \end{figure*} Fig.\ \ref{f:selfSimilarityZ23To26} allows visual assessment of the self-similarity of the intensity cross-sections recorded in three transverse planes near the self-conjugate plane S. When stretched by the transverse magnification $M$, the centers of the intensity cross-sections show some similarity to the unstretched intensity cross-sections; for example, dark, centered, hexagons or circles of approximately the same size are present in the stretched and unstretched intensity cross-sections, especially in the planes $Z = 230\,\mathrm{mm}$ and $Z = 240\,\mathrm{mm}$, which are closest to the self-conjugate plane~S. \begin{figure}[ht] \begin{center} \includegraphics{selfSimilarity_DeltaZ_Experiment.pdf} \end{center} \caption{\label{f:selfSimilarityExperiment}Evolution of the self-similarity upon propagation of measured transverse intensity cross-sections. Like in Fig.\ \ref{f:selfSimilarity}, the plot shows the negative normalised squared Euclidean distance, $-d^2$, as a function of axial coordinate $Z$. The predicted position of the self-conjugate plane S, $Z = Z_S = 231.8\,\mathrm{mm}$, is indicated by a dotted black line surrounded by a grey area representing experimental uncertainties. The planes of the mirrors, M$_1$ and M$_2$, and the self-conjugate plane, S, are indicated by vertical dotted lines. } \end{figure} We analysed the recorded intensity cross-sections quantitatively by evaluating their self-similarity and plotting the evolution of this self-similarity upon propagation --- the experimental analog to the curves shown in Fig.\ \ref{f:selfSimilarity}. The result is shown in Fig.\ \ref{f:selfSimilarityExperiment}. The mirror planes are in no way special, which was also the case in the plot calculated from simulated data (Fig.\ \ref{f:selfSimilarity}). The self-similarity is greatest around the expected position of the plane S, but the sharp peak visible in the curves shown in Fig.\ \ref{f:selfSimilarity} is absent. One possible explanation for the absence of this peak is that, due to the theoretical sharpness of the peak (Fig.\ \ref{f:selfSimilarity}), we did not record the intensity in a plane close enough to S despite sampling a plane only $1.8\,\mathrm{mm}$ from S. Another possible explanation is that the lack of the sharp peak is caused by experimental imperfections. One type of such experimental imperfections, including thermal effects and errors on mirror curvatures and distances, could have resulted in S being located further than expected from the nearest plane that was sampled. Another type, experimental imaging imperfections, most likely due to the effect of the gain medium (which images only in the paraxial limit), could have led to fine detail in the intensity cross-section missing. \begin{figure*}[ht] \begin{center} \includegraphics[width=\linewidth]{CollageLogPlot.pdf} \end{center} \caption{Collage of experimentally obtained intensity cross-sections measured close to the self-conjugate plane S inside a variety of laser cavities. The edge of the aperture in the central column is that of the 3rd iteration of the Koch snowflake (the first iteration being an equilateral triangle), the other apertures are regular hexagons. $r_0$ is the circumradius of the apertures, $M$ is the approximate transverse round-trip magnification, and $N_F$ is the Fresnel number. Like in Fig.\ \ref{f:selfSimilarityZ23To26}, brightness is proportional to the logarithm of intensity. } \label{collage} \end{figure*} Finally, we investigated the generation of these fractal modes in several laser resonator configurations of differing magnification factors (and Fresnel numbers), as detailed in Table~\ref{parameters}, with the results shown in Fig.~\ref{collage}. Most of the lasers had hexagonal apertures, but we also experimented with an aperture in the shape of the 3rd iteration of the Koch-snowflake, a shape approximating a fractal. We do not analyse the intensity cross-sections in the laser with the Koch-snowflake aperture in any detail here, but note that it is not surprising that the resulting diffraction patterns are fractals \cite{Berry-1979} as there are \emph{several} mechanisms\footnote{Here there are two mechanisms, specifically the MOM effect and diffraction behind a fractal aperture. This could easily be extended to \emph{three} by periodically repeating the fractal aperture shape across the aperture plane, thereby adding the fractal Talbot effect \cite{Berry-Klein-1996a} to the mix.} at work that all simultaneously shape the eigenmode intensity cross-sections into fractals. \begin{table} \begin{center} \begin{tabular*}{0.48\textwidth}{|c| @{\extracolsep{\fill}} c|c|c|c|} \hline $R_1$ & 200 mm& 200 mm & 150 mm & 100 mm\\ \hline $R_2$&400 mm& 500 mm & 500 mm &500 mm \\ \hline $M$& 2.0 & 2.5 & 3.3 & 5.0 \\ \hline \end{tabular*} \end{center} \caption{Resonator parameters used to design fractal cavities of various magnifications. } \label{parameters} \end{table} \section{Discussion and Conclusion} When not considered in the context of resonators, the existence of 3D self-similar fractal light fields is surprising: the 3D intensity distribution of any light field is fully determined by any transverse cross-section, and so the lowest-loss eigenmode is fully determined by its cross-section in the magnified self-conjugate plane. The existence of self-similar transverse cross-sections whose corresponding beams --- their 3D diffraction patterns --- are also self-similar is far from obvious. While we have attempted to observe this experimentally, the experimental requirements on imaging are at present prohibitive. In the 2D case we have been able to confirm the emergence of fractal light from carefully constructed lasers. We have shown experimentally that fractals can be created directly from such laser cavities, confirming a theoretical prediction of some decades. While the experimental confirmation of 2D fractals reported here concludes an open question in the community, the extension of the theory to 3D opens new exciting avenues for further exploration.

1ea709243c2729e6957ece516ddf93ed266a32d0

\section{Introduction} Cross-lingual word representation learning has been recognized as a very important research topic in natural language processing (NLP). It aims to represent multilingual word embeddings in a common space, and has been applied to many multilingual tasks, such as machine translation \cite{MT_crossemb} and bilingual named entity recognition \cite{named_entity}. It also enables the transfer of knowledge from one language into another \cite{parse_crossemb,Adams}. A number of methods have been proposed that obtain multilingual word embeddings. The key idea is to learn a linear transformation that maps word embedding spaces of different languages. Most of them utilize parallel data such as parallel corpus and bilingual dictionaries to learn a mapping \cite{sup_map}. However, such data are not readily available for many language pairs, especially for low-resource languages. To tackle this problem, a few unsupervised methods have been proposed that obtain cross-lingual word embeddings without any parallel data \cite{MUSE,Zhang,Zhang2,vecmap,vecmap2}. Their methods have opened up the possibility of performing unsupervised neural machine translation \cite{FB_UNMT,UNMT}. \citeauthor{MUSE} \shortcite{MUSE}, \citeauthor{Zhang} \shortcite{Zhang} propose a model based on adversarial training, and similarly \citeauthor{Zhang2} \shortcite{Zhang2} propose a model that employs Wasserstein GAN \cite{WGAN}. Surprisingly, these models have outperformed some supervised methods in their experiments. Recently, however, \citeauthor{limitation} \shortcite {limitation} have pointed out that the model of \citeauthor{MUSE} \shortcite{MUSE} is effective only when the domain of monolingual corpora is the same across languages and languages to align are linguistically similar. \citeauthor{vecmap2} \shortcite{vecmap2}, on the other hand, have overcome this problem and proposed a more robust method that enables to align word embeddings of distant language pairs such as Finnish and English. However, all of these approaches still have a common significant bottleneck: they require a large amount of monolingual corpora to obtain cross-lingual word embedddings, and such data are not readily available among minor languages. In this work, we propose a new unsupervised method that can obtain cross-lingual embeddings even in a low-resource setting. We define our method as {\it multilingual neural language model}, that obtains cross-lingual embeddings by capturing a common structure among multiple languages. More specifically, our model employs bidirectional LSTM networks \cite{BiRNN,LSTM} that respectively perform as forward and backward language models \cite{LM}, and these parameters are shared among multiple languages. The shared LSTM networks learn a common structure of multiple languages, and the shared network encodes words of different languages into a common space. Our model is significantly different from the existing unsupervised methods in that while they aim to align two pre-trained word embedding spaces, ours jointly learns multilingual word embeddings without any pre-training. Our experiments show that our model is more stable than the existing methods under a low-resource condition, where it is difficult to obtain fine-grained monolingual word embeddings. \section{Our Model} \begin{figure}[] \begin{center} {\includegraphics[]{new_fig_small.pdf}} \end{center} \caption{Illustration of our proposed {\it multilingual neural language model}. The parameters shared among across multiple languages are the ones of forward and backward LSTMs $\overrightarrow{f}$ and $\overleftarrow{f}$, the embedding of $<$BOS$>$, $E^{\rm BOS}$, and the linear projection for $<$EOS$>$, $W^{\rm EOS}$. On the other hand, word embeddings, $E^{\ell}$, and linear projection $W^{\ell}$ are specific to each language $\ell$. The shared LSTMs capture a common structure of multiple languages, and that enables us to map word embeddings $E^{\ell}$ of multiple languages into a common space. \label{model}} \end{figure} \subsection{Overview} We propose a model called multi-lingual neural language model, which produces cross-lingual word embeddings in an unsupervised way. Figure \ref{model} briefly illustrates our proposed model. The model consists of the shared parameters among multiple languages and the specific ones to each language. In what follows, we first summerize which parameters are shared or separate across languages: \begin{itemize} \item Shared Parameters \begin{itemize} \item $\boldsymbol{\overrightarrow{f}}$ and $\boldsymbol{\overleftarrow{f}}$: LSTM networks which perform as forward and backward language models, independently. \item $\boldsymbol{{E^{\rm BOS}}}$: The embedding of $<$BOS$>$, an initial input to the language models. \item $\boldsymbol{W^{\rm EOS}}$: The linear mapping for $<$EOS$>$, which calculates how likely it is that the next word is the end of a sentence. \end{itemize} \item Separate Parameters \begin{itemize} \item $\boldsymbol{E^{\ell}}$: Word embeddings of language $\ell$ \item $\boldsymbol{W^{\ell}}$: Linear projections of language $\ell$, which is used to calculate the probability distribution of the next word. \end{itemize} \end{itemize} The LSTMs $\overrightarrow{f}$ and $\overleftarrow{f}$ are shared among multiple languages and capture a common language structure. On the other hand, the word embedding function $E^{\ell}$ and liner projection $W^{\ell}$ are specific to each language $\ell$. Since different languages are encoded by the same LSTM functions, similar words across different languages should have a similar representation so that the shared LSTMs can encode them effectively. For instance, suppose our model encodes an English sentence ``He drives a car." and its Spanish translation ``El conduce un coche." In these sentences, each English word corresponds to each Spanish one in the same order. Therefore, these equivalent words would have similar representations so that the shared language models can encode the English and Spanish sentences effectively. Although in general each language has its different grammar rule, the shared language models are trained to roughly capture the common structure such as a common basic word order rule (e.g. subject-verb-object) among different languages. Sharing $<$BOS$>$ and $<$EOS$>$ symbols further helps to obtain cross-lingual representations, ensuring that the beginning and end of the hidden states are in the same space regardless of language. In particular, sharing $<$EOS$>$ symbol indicates that the same linear function predicts how likely it is that the next word is the end of a sentence. In order for the forward and backward language models to predict the end of a sentence with high probability, the words that appear near the end or beginning of a sentence such as punctuation marks and conjunctions should have very close representations among different languages. \subsection{Network Structure} Suppose a sentence with $N$ words in language $\ell$, $\langle w^\ell_{1},w^\ell_{2}, ..., w^\ell_{N} \rangle$. The forward language model calculates the probability of upcoming word $w^\ell_{t}$ given the previous words $w^\ell_{1},w^\ell_{2}, ..., w^\ell_{t-1}$. \begin{equation} P( w^\ell_{1},w^\ell_{2},..., w^\ell_{N}) = \prod_{t=1}^N p(w^\ell_{t}|w^\ell_{1},w^\ell_{2},..., w^\ell_{t-1}). \end{equation} The backward language model is computed similarly given the backward context: \begin{equation} P( w^\ell_{1},w^\ell_{2},..., w^\ell_{N}) = \prod_{t=1}^N p(w^\ell_{t}|w^\ell_{t+1},w^\ell_{t+2},..., w^\ell_{N}). \end{equation} The $t$th hidden states ${h}^\ell_{t}$ of the forward and backward LSTMs are calculated based on the previous hidden state and word embedding, \begin{equation} \overrightarrow{h}^\ell_{t} = \overrightarrow{f}(\overrightarrow{h}^\ell_{t-1}, x_{t-1}^\ell), \end{equation} \begin{equation} \overleftarrow{h}^\ell_{t} = \overleftarrow{f}(\overleftarrow{h}^\ell_{t+1}, x_{t+1}^\ell), \end{equation} \begin{equation} x^\ell_{t} = \begin{cases} E^{\rm BOS} & \text{if \textit{t} = 0 or \textit{N}+1},\\ E^\ell(w^\ell_{t}) & \text{otherwise}, \end{cases} \end{equation} where $\overrightarrow{f}(\cdot)$ and $\overleftarrow{f}(\cdot)$ are the standard LSTM functions. $E^{\rm BOS} $ is the embedding of $<$BOS$>$, which is shared among all the languages. Note that the same word embedding function $E^\ell$ is used among the forward and backward language models. The probability distribution of the upcoming word $w^\ell_{t}$ is calculated by the forward and backward models \textbf{independently} based on their current hidden state: \begin{equation} p(w^\ell_{t}|w^\ell_{1},w^\ell_{2},..., w^\ell_{t-1}) = {\rm softmax}(g^\ell(\overrightarrow{h}^\ell_{t}))), \end{equation} \begin{equation} p(w^\ell_{t}|w^\ell_{t+1},w^\ell_{t+2},..., w^\ell_{N}). = {\rm softmax}(g^\ell(\overleftarrow{h}^\ell_{t})), \end{equation} \begin{equation} g^\ell(h^\ell_{t}) =[W^{\rm EOS}(h^\ell_{t}),W^\ell(h^\ell_{t})], \end{equation} where $[x, y]$ means the concatenation of $x$ and $y$. $W^{\rm EOS}$ is a matrix with the size of ($1\times d$), where $d$ is the dimension of the hidden state. This matrix is a mapping function for $<$EOS$>$, and shared among all of the languages. $W^\ell$ is a matrix with the size of ($V^\ell \times d$), where $V^\ell$ is the vocabulary size of language $\ell$ excluding $<$EOS$>$. Therefore, $g$ is a linear transformation with the size of ($(V^\ell+1)\times d$). As with the word embeddings, the same mapping functions are used among the forward and backward language models. The largest difference between our model and a standard language model is that our model shares LSTM networks among different languages, and the shared LSTMs capture a common structure of multiple languages. Our model also shares $<$BOS$>$ and $<$EOS$>$ among languages, which encourages word embeddings of multiple languages to be mapped into a common space. The proposed model is trained by maximizing the log likelihood of the forward and backward directions for each language $\ell$: \begin{equation*} \begin{split} \sum_{l=1}^{L}\sum_{i=1}^{S^\ell}\sum_{t=1}^{N^i} &\log p(w^\ell_{i, t}|w^\ell_{i,1},w^\ell_{i, 2},...w^\ell_{i, t-1};\overrightarrow{\theta})\\ & + \log p(w^\ell_{i, t}|w^\ell_{i, t+1},w^\ell_{i, t+2},...w^\ell_{i, N^i}; \overleftarrow{\theta}), \end{split} \end{equation*} where $L$ and $S^\ell$ denote the number of languages and sentences of language $\ell$. $\overrightarrow{\theta}$ and $\overleftarrow{\theta}$ denote the parameters for the forward and backward LSTMs $\overrightarrow{f}$ and $\overleftarrow{f}$, respectively. \section{Related Work} \subsection{Unsupervised Word Mapping} A few unsupervised methods have been proposed that obtain cross-lingual representations in an unsupervised way. Their goal is to find a linear transformation that aligns pre-trained word embeddings of multiple languages. For instance, \citeauthor{vecmap} \shortcite{vecmap} obtain a linear mapping using a parallel vocabulary of automatically aligned digits (i.e. 1-1, 2-2, 15-15...). In fact, their method is weakly supervised because they rely on the aligned information of Arabic numerals across languages. \citeauthor{Zhang} \shortcite{Zhang} and \citeauthor{MUSE} \shortcite{MUSE}, on the other hand, propose fully unsupervised methods that do not make use of any parallel data. Their methods are based on adversarial training \cite{adversarial}: during the training, a discriminator is trained to distinguish between the mapped source and the target embeddings, while the mapping matrix is trained to fool the discriminator. \citeauthor{MUSE} \shortcite{MUSE} further refine the mapping obtained by the adversarial training. They build a synthetic parallel vocabulary using the mapping, and apply a supervised method given the pseudo parallel data. \citeauthor{Zhang2} \shortcite{Zhang2} employ Wasserstein GAN and obtain cross-lingual representations by minimizing the earth-mover's distance. \citeauthor{vecmap2} \shortcite{vecmap2} propose an unsupervised method using a significantly different approach from them. It first roughly aligns words across language using structural similarity of word embedding spaces, and refines the word alignment by repeating a robust self-learning method until convergence. They have found that their approach is much more effective than \citeauthor{Zhang} \shortcite{Zhang} and \citeauthor{MUSE} \shortcite{MUSE} on realistic scenarios, namely when languages to align are linguistically distant or training data are non-comparable across language. The common objective among all these unsupervised methods is to map word embeddings of multiple languages into a common space. In their experiments. the word embeddings are pre-trained on a large amount of monolingual data such as Wikipedia before their methods are applied. Therefore, they haven't evaluated their method on the condition when only a small amount of data are available. That condition is very realistic for minor languages, and an unsupervised method can be very useful for these languages. In our experiments, it turns out that existing approaches do not perform well without enough data, while our proposed method can align words with as small data as fifty thousand sentences for each language. \subsection{Siamese Neural Network} Our model embeds words of multiple languages into a common space by sharing LSTM parameters among the languages. In general, the model architecture of sharing parameters among different domains is called the ``Siamese Neural Network" \cite{Siamese}. It is known to be very effective at representing data of different domains in a common space, and this technique has been employed in many NLP tasks. For example, \citeauthor{zeroshot_NMT} \shortcite{zeroshot_NMT} built a neural machine translation model whose encoder and decoder parameters are shared among multiple languages. They have observed that sentences of multiple languages are mapped into a common space, and that has made it possible to perform zero-shot translation. \citeauthor{named_entity} \shortcite{named_entity} share LSTM networks of their named entity recognition model across multiple languages, and improve the performance in resource-poor languages. Note that these models are fully supervised and require parallel data to obtain cross-lingual representations. Our model, on the other hand, does not require any parallel or cross-lingual data, and it acquires cross-lingual word embeddings through finding a common language structure in an unsupervised way. \section{Experiments } \subsection{Data sets} We considered two learning scenarios that we deem realistic for low-resource languages: \begin{enumerate} \item Only a small amount of monolingual data are available. \item The domains of monolingual corpora are different across languages. \end{enumerate} For the first case, we used the News Crawl 2012 monolingual corpus for every language except for Finnish, for which we used News Crawl 2014. These data are provided by WMT2013\footnote{\url{http://www.statmt.org/wmt13/translation-task.html}} and 2017\footnote{\url{http://www.statmt.org/wmt17/translation-task.html}}. We randomly extracted 50k sentences in each language, and used them as training data. We also extracted 100k, 150k, 200k, and 250k sentences and analyzed the impact of the data size. For the second scenario, we used the Europarl corpus \cite{Europarl} as an English monolingual corpus, and the News Crawl corpus for the other languages. We randomly extracted one million sentences from each corpus and used them as training data. The full vocabulary sizes of the Europarl and News Crawl corpora in English were 79258 and 265368 respectively, indicating the large difference of the domains. We did not use any validation data during the training. We tokenized and lowercased these corpora using Moses toolkit\footnote{\url{https://github.com/moses-smt/mosesDecoder}}. We evaluated models in the pairs of \{French, German, Spanish, Finnish, Russian, Czech\}-English. \subsection{Evaluation} \begin{table*}[!] \hbox to\hsize{\hfil \makebox[\linewidth]{ \scalebox{0.9}[1.0]{ \centering \begin{tabular}{|l|r|r||r|r||r|r||r|r||r|r||r|r|}\hline & \multicolumn{2}{c||}{fr-en} &\multicolumn{2}{c||}{de-en} &\multicolumn{2}{c||}{es-en} &\multicolumn{2}{c||}{fi-en} &\multicolumn{2}{c||}{ru-en} &\multicolumn{2}{c|}{cs-en} \\\cline{2-13} & p@1 &p@5 &p@1 &p@5 &p@1 &p@5 &p@1 &p@5&p@1 &p@5&p@1 &p@5 \\\hline \textsc{Random} &0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5 \\ \citeauthor{MUSE} \shortcite{MUSE} &2.5&7.7&0.6&3.5&3.0&9.0 &0.0 &0.4 &0.1 &0.7 &0.0 &1.2\\ \citeauthor{MUSE} \shortcite{MUSE} + normalize &0.7&3.0&0.6&3.3 &0.5 &2.6 &0.0&0.4 &0.0&0.5 &0.1&0.3 \\ \citeauthor{vecmap2} \shortcite{vecmap2} &2.4&6.8&1.0&4.5&1.0 &5.0&0.0 &0.1 &0.2 &0.9 &0.4 &1.6 \\ \textsc{Ours}&\textbf{7.3}&\textbf{16.5}&\textbf{4.6}&\textbf{12.0}&\textbf{8.2} &\textbf{18.0} &\textbf{2.7} &\textbf{7.3} &\textbf{2.7} &\textbf{6.9} &\textbf{3.7} &\textbf{10.2} \\\hline \end{tabular}\hfil } } } \caption{ Word alignment average precisions p@1 and 5 when models are trained on 50k sentences of source and target languages.\label{low_resource}} \end{table*} \begin{table*}[!] \hbox to\hsize{\hfil \makebox[\linewidth]{ \scalebox{0.9}[1.0]{ \begin{tabular}{|l|r|r||r|r||r|r||r|r||r|r||r|r|}\hline & \multicolumn{2}{c||}{fr-en} &\multicolumn{2}{c||}{de-en} &\multicolumn{2}{c||}{es-en} &\multicolumn{2}{c||}{fi-en} &\multicolumn{2}{c||}{ru-en} &\multicolumn{2}{c|}{cs-en} \\\cline{2-13} & p@1 &p@5 &p@1 &p@5 &p@1 &p@5 &p@1 &p@5&p@1 &p@5&p@1 &p@5 \\\hline \textsc{Random} &0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5&0.1&0.5 \\ \citeauthor{MUSE} \shortcite{MUSE} &0.8&4.2&0.2&1.3&1.4&4.6 &0.1 &0.6 &0.6 &2.1 &0.5 &1.3\\ \citeauthor{MUSE} \shortcite{MUSE} + normalize &0.2 &1.2 &0.1&0.8&0.2 &1.0 &0.2&1.1&0.3&1.1 &0.3 &1.2\\ \citeauthor{vecmap2} \shortcite{vecmap2} &6.1&14.7&1.1&5.0&\textbf{29.9} &\textbf{45.3}&0.5 &2.2 & 0.1&1.2 &0.5 &2.2 \\ \textsc{Ours}&\textbf{12.7}&\textbf{26.6}&\textbf{3.4}&\textbf{10.0}&14.9 &28.6 &\textbf{3.0} &\textbf{8.5} &\textbf{3.8} &\textbf{11.1} &\textbf{4.0} &\textbf{10.8}\\\hline \end{tabular}\hfil } } } \caption{ Word alignment average precisions p@1 and 5 when models are trained on one million sentences extracted from different domains between source and target languages. \label{diff_domains}} \end{table*} \begin{table*}[!t] \hbox to\hsize{\hfil \centering \begin{tabular}[c]{| l | l | l | l | | l | l | l | }\hline \multirow{ 2}{*}{source word (es)} &\multicolumn{3}{c||}{ \textsc{Ours} } &\multicolumn{3}{c|}{\citeauthor{vecmap2} \shortcite{vecmap2}} \\\cline{2-7} & top 1 & top 2 & top 3 & top 1 & top 2 & top 3 \\\hline acusado &\textbf{accused} &designed &captured& english & dark & drama \\\hline actor& \textbf{actor}& artist& candidate & appointment& \textbf{actor} &charlie\\\hline casi &\textbf{almost}& approximately& about &around &age &capita \\\hline aunque &\textbf{although}& but &drafting &about & are & been\\\hline d\'{i}as &\textbf{days} &decades &decade &bodies & both &along \\\hline actualmente& \textbf{currently} &clearly& essentially &comes &continued &candidates\\\hline contiene &\textbf{contains} &defines& constitutes &barrier &etiquette& commissioned\\\hline cap\'{i}tulo &\textbf{chapter} &episode& cause &arriving &bulls& dawn\\\hline \end{tabular}\hfil } \caption{Some examples when Spanish and English words matched correctly by our model using 50k sentences, but not by \citeauthor{vecmap2} \shortcite{vecmap2}. Each column indicates 1st, 2nd, and 3rd most similar English words to each Spanish word. English words in bold font are translations of each Spanish word.\label{words}} \end{table*} In this work, we evaluate our methods on a word alignment task. Given a list of $M$ words in a source language s $[x_{1},x_{2},..., x_{M}]$ and target language $t$ $[y_{1},y_{2},..., y_{M}]$, the word alignment task is to find one-to-one correspondence between these words. If a model generates accurate cross-lingual word embeddings, it is possible to align words properly by measuring the similarity of the embeddings. In our experiment, we used the bilingual dictionary data published by \citeauthor{MUSE} \shortcite{MUSE}, and extracted 1,000 unique pairs of words that are included in the vocabulary of the News Crawl data of from 50k to 300k sentences. As a measurement of the word embeddings, we used cross-domain similarity local scaling (CSLS), which is also used in \citeauthor{MUSE} \shortcite{MUSE} and \citeauthor{vecmap2} \shortcite{vecmap2} . CSLS can mitigate the hubness problem in high-dimensional spaces, and can generally improve matching accuracy. It takes into account the mean similarity of a source language embedding $x$ to its $K$ nearest neighbors in a target language: \begin{equation} rT(x)=\frac{1}{K} \sum_{y \in {\mathcal N_{T}(x)}} {\rm cos} (x,y), \end{equation} where cos is the cosine similarity and $N_{T}(x)$ denotes the $K$ closest target embeddings to $x$. Following their suggestion, we set $K$ as 10. $rR(y)$ is defined in a similar way for any target language embedding y. CSLS(x, y) is then calculated as follows: \begin{equation} {\rm CSLS}(x, y) = 2 {\rm cos}(x, y) -rT(x)- rS(y). \end{equation} For each source word $x_{i}$, we extracted the $k$ target words that have the highest CSLS scores ($k$ = 1 or 5). However, since the value of $rT(x)$ does not affect the result of this evaluation, we omit the score from CSLS in our experiments. We report the precision p@k: how often the correct translation of a source word $x_{i}$ is included in the $k$ extracted target words. \subsection{Baseline} As baselines, we compared our model to that of \citeauthor{MUSE} \shortcite{MUSE} and \citeauthor{vecmap2} \shortcite{vecmap2}. \citeauthor{MUSE} \shortcite{MUSE} aim to find a mapping matrix $W$ based on adversarial training. The discriminator is trained to distinguish the domains (i.e. language) of the embeddings, while the mapping is trained to fool the discriminator. Then, $W$ is used to match frequent source and target words, and induce a bilingual dictionary. Given the pseudo dictionary, a new mapping matrix $W$ is then trained in the same manner as a supervised method, which solves the Orthogonal Procrustes problem: \begin{equation*} \begin{split} &W^* = \mathop{\rm arg~min}\limits_{W} \|WX-Y\|_{F} = UV^{\mathrm{T}},\\ & s.t. ~~~ U\sum{}V^{\mathrm{T}} = {\rm SVD}(YX^{\mathrm{T}}). \end{split} \end{equation*} This training can be iterated using the new matrix $W$ to induce a new bilingual dictionary. This method assumes that the frequent words can serve as reliable anchors to learn a mapping. Since they suggest normalizing word embeddings in some language pairs, we evaluated their method with and without normalization. \citeauthor{vecmap2} \shortcite{vecmap2} use a different approach and employ a robust self-learning method. First, they roughly align words based on the similarity of word emebeddings. Then, they repeat the self-learning approach, where they alternatively update a mapping function and word alignment. To implement the baseline methods, we used the code published by the authors\footnote{\url{https://github.com/facebookresearch/MUSE}}\footnote{\url{https://github.com/artetxem/vecmap}}. To obtain monolingual word embeddings, we used word2vec \cite{word2vec}. Note that these embeddings were used only for the baselines, but not for ours since our method does not require any pre-trained embeddings. For a fair comparison, we used the same monolingual corpus with the same vocabulary size for the baselines and our model. \subsection{Training Settings } We preprocessed monolingual data and generated mini-batches for each language. For each iteration, our model alternately read mini-batches of each language, and updated its parameters every time it read one mini-batch. We trained our model for 10 epochs with the mini-batch size 64. The size of word embedding was set as 300, and the size of LSTM hidden states was also set as 300 for the forward and backward LSTMs, respectively. Dropout \cite{dropout} is applied to the hidden state with its rate 0.3. We used SGD \cite{SGD} as an optimizer with the learning rate 1.0. Our parameters, which include word embeddings, were uniformly initialized in [-0.1, 0.1], and gradient clipping \cite{clipping} was used with the clipping value 5.0. We included in the vocabulary the words that were used at least a certain number of times. For the News Crawl corpus, we set the threshold as 3, 5, 5, 5 ,5, 10, and 20 for 50k, 100k, 150k, 200k, 250k, 300k and 1m sentences. For the Europarl corpus, we set the value as 10. We fed 10000 frequent words into the discriminator in \citeauthor{MUSE} \shortcite{MUSE}. As a model selection criterion, we employed a similar strategy used in the baseline. More specifically, we considered the 3,000 most frequent source words, and used CSLS excluding $rT(x)$ to generate a translation for each of them in a target language. We then computed the average CSLS scores between these deemed translations, and used them as a validation metric. \section{Results} \begin{figure}[!h] \begin{center} {\includegraphics[]{seeds_small2.pdf}} \end{center} \caption{Comparison of p@1 accuracy of German-English pair between supervised word mapping method and our model on 50k sentences. The x axis indicates the number of pairs of words $n$ (= 0,50,100,150,..., 450, 500) that were used for the supervised method, but not for ours, to map word embedding spaces in two languages. \label{vs_supervised}} \end{figure} \begin{figure}[h!] \begin{center} \includegraphics[]{change_N_sentences_small.pdf} \end{center} \caption{Graphs show the change in p@1 accuracy of each language pair as the size of training data increases. The x-axis denotes the number of sentences (thousand) in the monolingual training data of the source and target languages. \label{change_N}} \end{figure} \subsection{Bilingual Word Embeddings} First, we trained our model and obtained cross-lingual embeddings between two languages for each language pair. We report our results under the two scenarios that we considerted realistic when dealing with minor languages. In the first scenario, we trained our model on a very small amount of data, and in the second scenario the model was trained on a large amount of data extracted from different domains between source and target languages. Table \ref{low_resource} illustrates the results of the word alignment task under the low-resource scenario. \textsc{Random} is the expected accuracy when words are aligned at random. The result shows that our model outperformed the baseline methods significantly in all of the language pairs, indicating that ours is more robust in a low-resource senario. On the other hand, the baseline methods got poor performance, especially in the Finnish and English pair. Even though \citeauthor{vecmap2} \shortcite{vecmap2} report that their method achieves good performance in that language pair, our experiment has demonstrated that it does not perform well without a large amount of data. Table \ref{diff_domains} shows the results when the domains of training data used to obtain source and target embeddings are different. Our method again outperformed the baselines to a large extent except for the Spanish-English pair. The poor performance of \citeauthor{MUSE} \shortcite{MUSE} in such a setting has also been observed in \citeauthor{limitation} \shortcite{limitation}, even though much larger data including Wikipedia were used for training in their experiments. Table \ref{words} shows some examples when Spanish and English words were correctly matched by our model, but not by \citeauthor{vecmap2} \shortcite{vecmap2} under the low-resource scenario. The table lists the three most similar English words to each Spanish source word. Our method successfully matched similar or semantically related words to the source words, indicating that our method obtained good cross-lingual embeddings. For example, to the Spanish source word ``casi", our model aligned its translation ``almost" and also very similar words ``approximately" and ``about". Indeed, many of the aligned target words in our model have the same part of speech tag as that of the source word, suggesting that our model captured a common language structure such as rules of word order and roles of vocabulary by sharing LSTMs. On the other hand, \citeauthor{vecmap2} \shortcite{vecmap2} could not align words properly, and there do not seem to exist consistent relations between the source and extracted words. \begin{table*}[] \centering \hbox to\hsize{\hfil \begin{tabular}{|l|r|r|r||r|r|r||r|r|r|}\hline & \multicolumn{3}{c||}{fr-en} &\multicolumn{3}{c||}{de-en} &\multicolumn{3}{c|}{es-en} \\\cline{2-10} &50k &100k&300k &50k &100k&300k &50k &100k&300k \\\hline \textsc{\textsc{Ours (bilingual)} }&{7.3}&\textbf{12.8}&{23.1}&{4.6}&{5.7}&{13.1}&\textbf{8.2} &\textbf{14.0} & \textbf{27.9} \\\hline \textsc{\textsc{Ours (quadrilingual)} }&\textbf{8.4}&{12.3}&\textbf{25.6}&\textbf{4.7}&\textbf{8.4}&\textbf{16.6}&{7.6} &{13.8} & {25.3} \\\hline \end{tabular}\hfil } \caption{Word alignment average precisions p@1 in each language pair when 50k, 100k, and 300k sentences were used for training. \textsc{Ours (bilingual)} denotes the accuracy of the models that read source and target languages, generating bilingual word embeddings. \textsc{Ours (quadrilingual)} denotes the accuracy of one model that reads all four languages, producing quadrilingual word embeddings. \label{quad}} \end{table*} \subsubsection{Comparison to Supervised Method} To further investigate the effectiveness of our model, we compared our method to a supervised method under the low-resource setting. The method is a slight modification of \citeauthor{MUSE} \shortcite{MUSE}: it is trained using a bilingual dictionary, and learns a mapping from the source to the target space using iterative Procrustes alignment. We used the code provided by \citeauthor{MUSE} \shortcite{MUSE}. Figure \ref{vs_supervised} compares p@1 accuracy between the supervised method and our model in the German-English pair. The x-axis denotes the number of seeds of the bilingual dictionary that were used for the supervised method, but not for ours. The figure illustrates that our method achieved a better result than the supervised method when the number of seeds was less than 400, which is surprising given that our model is fully unsupervised. \subsubsection{Impact of Data Size} We changed the size of the training data by 50k sentences, and analyzed how the performance of the baselines and our model changed. Figure \ref{change_N} illustrates how the performance changed depending on the data size. It shows that our method achieved a comparable or better result than the baseline methods in all of the language pairs when the number of sentences was not more than 100k. In the closely related language pairs such as \{French, German, Spanish\}-English, the baselines performed better when there were enough amount of data. Among the distant languages such as \{Finnish, Czech\}-English, our model achieved better results overall, while the baseline methods, especially \citeauthor{MUSE} \shortcite{MUSE} got very poor results. \begin{table}[] \centering \begin{tabular}{|l|l|l|l|} \hline En (source) & Fr top1 & De top1 & Es top1 \\ \hline declared & d\'{e}clar\'{e} & erkl\"{a}rt & declarado \\ \hline always & toujours & immer & siempre \\ \hline are & sont & sind & est\'{a}n \\ \hline after & apr\`{e}s & nachdem & despu\'{e}s \\ \hline died & d\'{e}c\'{e}d\'{e} & starb & murieron \\ \hline \end{tabular} \caption{Examples of words that were correctly aligned by \textsc{Ours (quadrilingual)} among the four languages. \label{quadalign}} \end{table} \subsection{Quadrilingual Word Embeddings} Our results of the word alignment task have shown that our model can jointly learn bilingual word embeddings by capturing the common structure of two languages. This success has raised another question: ``Is it also possible to learn a common structure of more than two languages?" To examine this intriguing question, we trained our model that encoded four linguistically similar languages, namely English, French, Spanish, and German, and aimed to capture the common structure among them. We expect that word embeddings of the four languages should be mapped into a common space, generating what we call quadrilingual word embeddings. Table \ref{quad} describes the result of the word alignment when using bilingual and quadrilingual word embeddings of our model. While quadrilingual word embeddings performed slightly worse than bilingual ones in the Spanish-English pair, they brought large gains in the German-English alignment task and achieved comparable performance overall. Our model successfully mapped word embeddings of the four languages into a common space, making it possible to measure the similarity of words across the multiple languages. To investigate whether quadrilingual embeddings were actually mapped into a common space, we aligned each English word to French, German and Spanish words in the bilingual dictionary. Table \ref{quadalign} describes the words that were correctly aligned among the four languages. This result indicates that these equivalent words have very similar representations, and that means our model successfully embedded these languages into a common space. Figure \ref{quad} illustrates the scatter plot of the embeddings of the most 1,000 frequent words in each corpus of the four languages. It clearly shows that the word embeddings were clustered based on their meanings rather than their language. For example, the prepositions of the four languages (e.g. de (fr, es), of (en), von (de)) were mapped into the bottom-right area, and determiners (e.g. la (fr, es), the (en), der, die, das (de)) were in the bottom-left area. Near the area where the embedding of `$<$BOS$>$' was mapped, the words from which a new sentence often starts (`,' , et (fr), y(es), und (de), and (en)) were mapped. \begin{figure}[!h] \begin{center} \centering {\includegraphics[]{figure_small.pdf}} \end{center} \caption{Scatter plot of cross-lingual word embeddings of French, English, German and Spanish obtained by our model. The embeddings are reduced to 2D using tSNE \cite{tSNE}. \label{quad}} \end{figure} \section{Conclusion} In this paper, we proposed a new unsupervised method that learns cross-lingual embeddings without any parallel data. Our experiments of a word alignment task in six language pairs have demonstrated that our model significantly outperforms existing unsupervised word translation models in all the language pairs under a low resource situation. Our model also achieved better results in five language pairs when the domains of monolingual data are different across language. We also compared our unsupervised method to a supervised one in the German-English word alignment task, and our model achieved a better result than the supervised method that were trained with 350 pairs of words from a bilingual dictionary. Our model also succeeded in obtaining cross-lingual embeddings across four languages, which we call quadrilingual embeddings. These embeddings enabled us to align equivalent words among four languages in an unsupervised way. The visualization of the quadrilingual embeddings showed that these embeddings were actually mapped into a common space, and words with similar meanings had close representations across different languages. Potential future work includes extending our approach to a semi-supervised method that utilizes a bilingual dictionary. One possible idea is to set an additional loss function in our model that decreases the distance of embeddings of equivalent words across languages.

bd05a604d499e144e3a622e3b12c742c0f83c71e

\section{Introduction}\label{sec:introduction} This paper is dedicated to so-called \emph{mathematical programs with switching constraints}, MPSCs for short. These are optimization problems of the form \begin{equation}\label{eq:MPSC}\tag{MPSC} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&\qquad&j\in\mathcal P,&\\ G_l(x)H_l(x)&\,=\,0,&\qquad&l\in\mathcal Q,& \end{aligned} \end{equation} where $\mathcal{M}:=\{1,\ldots,m\}$, $\mathcal{P}:=\{1,\ldots,p\}$, $\mathcal{Q}:=\{1,\ldots,q\}$ are index sets and the functions $f,g_i,h_j,G_l,H_l\colon\R^n\to\R$ are continuously differentiable for all $i\in\mathcal M$, $j\in\mathcal P$, and $l\in\mathcal Q$. For brevity, $g\colon\R^n\to\R^m$, $h\colon\R^n\to\R^p$, $G\colon\R^n\to\R^q$, and $H\colon\R^n\to\R^q$ are the mappings which possess the component functions $g_i$ ($i\in\mathcal M$), $h_j$ ($j\in\mathcal P$), $G_l$ ($l\in\mathcal Q$), and $H_l$ ($l\in\mathcal Q$), respectively. The last $q$ constraints in \eqref{eq:MPSC} force $G_l(x)$ or $H_l(x)$ to be zero for all $l\in\mathcal Q$, which gives rise to the terminology ``switching constraints''. Switching structures appear frequently in the context of optimal control, see \cite{ClasonRundKunisch2017,Gugat2008,HanteSager2013,Liberzon2003,Seidman2013,WangYan2015,Zuazua2011}, and the references therein, or as a reformulation of so-called \emph{either-or constraints}, see \cite[Section~7]{Mehlitz2018}. Naturally, \eqref{eq:MPSC} is related to other problem classes from disjunctive programming such as \emph{mathematical programs with complementarity constraints}, MPCCs for short, see \cite{LuoPangRalph1996,OutrataKocvaraZowe1998}, or \emph{mathematical programs with vanishing constraints}, MPVCs for short, see \cite{AchtzigerKanzow2008,HoheiselKanzow2008}. Indeed, similarly to MPCCs and MPVCs, standard constraint qualifications are likely to be violated at the feasible points of \eqref{eq:MPSC}. Recently, stationarity conditions and constraint qualifications for \eqref{eq:MPSC} were introduced in \cite{Mehlitz2018}. Here, we focus on the computational treatment of \eqref{eq:MPSC}. Clearly, standard methods from nonlinear programming may run into difficulties when applied to \eqref{eq:MPSC} due to two reasons: first, the feasible set of \eqref{eq:MPSC} is likely to be disconnected or at least \emph{almost} disconnected. Secondly, standard regularity conditions like the Mangasarian--Fromovitz constraint qualification, MFCQ for short, are likely to fail at the feasible points of \eqref{eq:MPSC} under mild assumptions, see \cite[Lemma~4.1]{Mehlitz2018}. Similar issues appear in the context of MPCCs and MPVCs where several different relaxation schemes were introduced to overcome these shortcomings, see \cite{HoheiselKanzowSchwartz2012,HoheiselKanzowSchwartz2013} and the references therein. Basically, the idea is to relax the \emph{irregular} constraints using a relaxation parameter such that the resulting surrogate problems are (regular) standard nonlinear problems which can be tackled by common methods. The relaxation parameter is then iteratively reduced to zero and, in each iteration, a Karush--Kuhn--Tucker (KKT) point of the surrogate problem is computed. Ideally, the resulting sequence possesses a limit point and, under some problem-tailored constraint qualification, this point satisfies a suitable stationarity condition. Furthermore, it is desirable that the relaxed problems satisfy standard constraint qualifications in a neighborhood of the limit point under reasonable assumptions. In this paper, we show that the relaxation scheme from \cite{KanzowSchwartz2013}, which was designed for the numerical investigation of MPCCs, can be adapted for the computational treatment of \eqref{eq:MPSC}. Particularly, it will be shown that the modified method can be used to find M-stationary points of \eqref{eq:MPSC}. By means of examples it is demonstrated that other relaxation methods which are well known from the theory of MPCCs only yield W-stationary points of \eqref{eq:MPSC}. We present the results of some numerical experiments which show the performance of the proposed method. The remaining part of the paper is structured as follows. In Section~\ref{sec:notation_preliminaries}, we describe the general notation used throughout the paper and recall some fundamental theory on nonlinear programming and switching constraints. Section~\ref{sec:relaxation_scheme_KS} is dedicated to the main relaxation approach and contains various properties of the resulting algorithm as well as convergence results. In Section~\ref{sec:other_relaxations}, we describe how some common regularization methods from MPCCs can be carried over to the switching-constrained setting, and discuss the convergence properties of the resulting algorithms. Section~\ref{sec:numerical_results} contains some numerical applications, including either-or constraints, switching-constrained optimal control, and semi-continuous optimization problems arising in portfolio optimization. We conclude the paper with some final remarks in Section~\ref{sec:final_results}. \section{Notation and preliminaries}\label{sec:notation_preliminaries} \subsection{Basic notation} The subsequently introduced tools of variational analysis can be found in \cite{RockafellarWets1998}. For a nonempty set $A\subset\R^n$, we call \begin{equation*} A^\circ:=\{y\in\R^n\,|\,\forall x\in A\colon\,x\cdot y\leq 0\} \end{equation*} the \emph{polar cone} of $A$. Here, $x\cdot y$ denotes the Euclidean inner product of the two vectors $x,y\in\R^n$. It is well known that $A^\circ$ is a nonempty, closed, convex cone. For any two sets $B_1,B_2\subset\R^n$, the polarization rule $(B_1\cup B_2)^\circ=B_1^\circ\cap B_2^\circ$ holds by definition. The polar of a polyhedral cone can be characterized by means of, e.g., Motzkin's theorem of alternatives. Note that we interpret the relations $\leq$ and $\geq$ for vectors componentwise. \begin{lemma}\label{lem:polar_of_polyhedral_cone} For matrices $\mathbf C\in\R^{m\times n}$ and $\mathbf D\in\R^{p\times n}$, let $\mathcal K\subset\R^n$ be the polyhedral cone \begin{equation*} \mathcal K:=\{d\in\R^n\,|\,\mathbf Cd\leq 0,\,\mathbf Dd=0\}. \end{equation*} Then $\mathcal K^\circ=\{\mathbf C^\top\lambda+\mathbf D^\top\rho\,|\, \lambda\in\R^m,\,\lambda\geq 0,\,\rho\in\R^p\}$. \end{lemma} Let $A\subset\R^n$ be a nonempty set and $\bar x\in A$. Then the closed cone \begin{equation*} \mathcal T_A(\bar x):= \left\{ d\in\R^n\,\middle|\, \exists\{x_k\}_{k\in\N}\subset A\,\exists\{\tau_k\}_{k\in\N}\subset\R_+\colon\,x_k\to\bar x,\,\tau_k\to 0,\, (x_k-\bar x)/\tau_k\to d \right\} \end{equation*} is called \emph{tangent} or \emph{Bouligand cone} to $A$ at $\bar x$. Here, $\R_+:=\{r\in\R\,|\,r>0\}$ denotes the set of all positive reals. The union $\{v^1,\ldots,v^r\}\cup\{w^1,\ldots,w^s\}$ of sets $\{v^1,\ldots,v^r\},\{w^1,\ldots,w^s\}\subset\R^n$ is called \emph{positive-linearly dependent} if there exist vectors $\alpha\in\R^r$, $\alpha\geq 0$, and $\beta\in\R^s$ which do not vanish at the same time such that \begin{equation*} 0=\sum\limits_{i=1}^r\alpha_iv^i+\sum\limits_{j=1}^s\beta_jw^j. \end{equation*} Otherwise, $\{v^1,\ldots,v^r\}\cup\{w^1,\ldots,w^s\}$ is called \emph{positive-linearly independent}. Clearly, if the set $\{v^1,\ldots,v^r\}$ is empty, then the above definitions reduce to linear dependence and independence, respectively. The following lemma will be useful in this paper; its proof is similar to that of \cite[Proposition~2.2]{QiWei2000} and therefore omitted. \begin{lemma}\label{lem:local_stability_of_positive_linear_independence} Let $\{v^1,\ldots,v^r\},\{w^1,\ldots,w^s\}\subset\R^n$ be given sets whose union $\{v^1,\ldots,v^r\}\cup\{w^1,\ldots,w^s\}$ is positive-linearly independent. Then there exists $\varepsilon>0$ such that, for all vectors $\tilde v^1,\ldots,\tilde v^r,\tilde w^1,\ldots,\tilde w^s\in\{z\in\R^n\,|\,\norm{z}{2}\leq\varepsilon\}$, the union $\{v^1+\tilde v^1,\ldots,v^r+\tilde v^r\}\cup\{w^1+\tilde w^1,\ldots,w^s+\tilde w^s\}$ is positive-linearly independent. \end{lemma} For some vector $z\in\R^n$ and an index set $I\subset\{1,\ldots,n\}$, $z_I\in\R^{|I|}$ denotes the vector which results from $z$ by deleting all $z_i$ with $i\in\{1,\ldots,n\}\setminus I$. Finally, let us mention that $\operatorname{supp} z:=\{i\in\{1,\ldots,n\}\,|\,z_i\neq 0\}$ is called the \emph{support} of the vector $z\in\R^n$. \subsection{Standard nonlinear programs} Here, we recall some fundamental constraint qualifications from standard nonlinear programming, see, e.g., \cite{BazaraaSheraliShetty1993}. Therefore, we consider the nonlinear program \begin{equation}\label{eq:nonlinear_program}\tag{NLP} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&&j\in\mathcal P,& \end{aligned} \end{equation} i.e., we forget about the switching constraints in \eqref{eq:MPSC} for a moment. Let $\tilde X\subset\R^n$ denote the feasible set of \eqref{eq:nonlinear_program} and fix some point $\bar x\in \tilde X$. Then \[ I^g(\bar x):=\{i\in\mathcal M\,|\,g_i(\bar x)=0\} \] is called the index set of active inequality constraints at $\bar x$. Furthermore, the set \[ \mathcal L_{\tilde X}(\bar x) := \left\{ d\in\R^n\,\middle|\, \begin{aligned} \nabla g_i(\bar x)\cdot d&\,\leq\,0&&i\in I^g(\bar x)\\ \nabla h_j(\bar x)\cdot d&\,=\,0&&j\in\mathcal P \end{aligned} \right\} \] is called the linearization cone to $\tilde X$ at $\bar x$. Obviously, $\mathcal L_{\tilde X}(\bar x)$ is a polyhedral cone, and thus closed and convex. It is well known that $\mathcal T_{\tilde X}(\bar x)\subset \mathcal L_{\tilde X}(\bar x)$ is always satisfied. The converse inclusion generally only holds under some constraint qualification. In the definition below, we recall several standard constraint qualifications which are applicable to \eqref{eq:nonlinear_program}. \begin{definition}\label{def:CQs} Let $\bar x\in \R^n$ be a feasible point of \eqref{eq:nonlinear_program}. Then $\bar x$ is said to satisfy the \begin{enumerate}[(a)] \item \emph{linear independence constraint qualification (LICQ)} if the following vectors are linearly independent: \begin{equation}\label{eq:NLP_vectors_for_CQs} \{\nabla g_i(\bar x)\,|\,i\in I^g(\bar x)\}\cup\{\nabla h_j(\bar x)\,|\,j\in\mathcal P\}. \end{equation} \item \emph{Mangasarian--Fromovitz constraint qualification (MFCQ)} if the vectors in \eqref{eq:NLP_vectors_for_CQs} are positive-linearly independent. \item \emph{constant positive linear dependence condition (CPLD)} if, for any sets $I\subset I^g(\bar x)$ and $J\subset\mathcal P$ such that the gradients \begin{equation*} \{\nabla g_i(\bar x)\,|\,i\in I\}\cup\{\nabla h_j(\bar x)\,|\,j\in J\} \end{equation*} are positive-linearly dependent, there exists a neighborhood $U\subset\R^n$ of $\bar x$ such that the gradients \begin{equation*} \{\nabla g_i(x)\,|\,i\in I\}\cup\{\nabla h_j(x)\,|\,j\in J\} \end{equation*} are linearly dependent for all $x\in U$. \item \emph{Abadie constraint qualification (ACQ)} if $\mathcal T_{\tilde X}(\bar x)=\mathcal L_{\tilde X}(\bar x)$. \item \emph{Guignard constraint qualification (GCQ)} if $\mathcal T_{\tilde X}(\bar x)^\circ=\mathcal L_{\tilde X}(\bar x)^\circ$. \end{enumerate} \end{definition} Note that the following relations hold between the constraint qualifications from Definition \ref{def:CQs}: \[ \text{LICQ}\,\Longrightarrow\,\text{MFCQ}\,\Longrightarrow\,\text{CPLD}\,\Longrightarrow\, \text{ACQ}\,\Longrightarrow\,\text{GCQ}, \] see \cite[Section~2.1]{HoheiselKanzowSchwartz2013} for some additional information. It is well known that the validity of GCQ at some local minimizer $\bar x\in\R^n$ of \eqref{eq:nonlinear_program} implies that the KKT conditions \[ \begin{split} &0=\nabla f(\bar x)+\sum\limits_{i\in I^g(\bar x)}\lambda_i\nabla g_i(\bar x)+\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\bar x),\\ &\forall i\in I^g(\bar x)\colon\,\lambda_i\geq 0 \end{split} \] provide a necessary optimality condition. Thus, the same holds for the stronger constraint qualifications ACQ, CPLD, MFCQ, and LICQ. \subsection{Mathematical programs with switching constraints} The statements of this section are taken from \cite{Mehlitz2018}. Let $X\subset\R^n$ denote the feasible set of \eqref{eq:MPSC} and fix a point $\bar x\in X$. Then the index sets \[ \begin{split} I^G(\bar x)&:=\{l\in\mathcal Q\,|\,G_l(\bar x)=0,\,H_l(\bar x)\neq 0\},\\ I^H(\bar x)&:=\{l\in\mathcal Q\,|\,G_l(\bar x)\neq 0,\,H_l(\bar x)=0\},\\ I^{GH}(\bar x)&:=\{l\in\mathcal Q\,|\,G_l(\bar x)=0,\,H_l(\bar x)= 0\} \end{split} \] form a disjoint partition of $\mathcal Q$. It is easily seen that MFCQ (and thus LICQ) cannot hold for \eqref{eq:MPSC} at $\bar x$ if $I^{GH}(\bar x)\neq\varnothing$. Taking a look at the associated linearization cone \[ \mathcal L_X(\bar x) = \left\{ d\in\R^n\, \middle|\, \begin{aligned} \nabla g_i(\bar x)\cdot d&\,\leq\,0&&i\in I^g(\bar x)\\ \nabla h_j(\bar x)\cdot d&\,=\,0&&j\in \mathcal P\\ \nabla G_k(\bar x)\cdot d&\,=\,0&&k\in I^G(\bar x)\\ \nabla H_k(\bar x)\cdot d&\,=\,0&&k\in I^H(\bar x) \end{aligned} \right\}, \] which is always convex, one can imagine that ACQ is likely to fail as well if $I^{GH}(\bar x)\neq\varnothing$ since, in the latter situation, $\mathcal T_X(\bar x)$ might be nonconvex. Note that GCQ may hold for \eqref{eq:MPSC} even in the aforementioned context. Due to the inherent lack of regularity, stationarity conditions for \eqref{eq:MPSC} which are weaker than the associated KKT conditions were introduced. \begin{definition}\label{def:stationarities} A feasible point $\bar x\in X$ of \eqref{eq:MPSC} is called \begin{enumerate}[(a)] \item \emph{weakly stationary (W-stationary)} if there exist multipliers $\lambda_i$ ($i\in I^{g}(\bar x)$), $\rho_j$ ($j\in\mathcal P$), $\mu_l$ ($l\in\mathcal Q$), and $\nu_l$ ($l\in\mathcal Q$) which solve the following system: \[ \begin{split} &0=\nabla f(\bar x)+\sum\limits_{i\in I^g(\bar x)}\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in \mathcal P}\rho_j\nabla h_j(\bar x)\\ &\qquad\qquad +\sum\limits_{l\in \mathcal Q}\bigl[\mu_l\nabla G_l(\bar x)+\nu_l\nabla H_l(\bar x)\bigr],\\ &\forall i\in I^g(\bar x)\colon\,\lambda_i\geq 0,\\ &\forall l\in I^H(\bar x)\colon\,\mu_l=0,\\ &\forall l\in I^G(\bar x)\colon\,\nu_l=0. \end{split} \] \item \emph{Mordukhovich-stationary (M-stationary)} if it is W-stationary and the associated multipliers additionally satisfy \[ \forall l\in I^{GH}(\bar x)\colon\,\mu_l\nu_l=0. \] \item \emph{strongly stationary (S-stationary)} if it is W-stationary while the associated multipliers additionally satisfy \[ \forall l\in I^{GH}(\bar x)\colon\,\mu_l=0\,\land\,\nu_l=0. \] \end{enumerate} \end{definition} Clearly, the following implications hold: \[ \text{S-stationarity}\;\Longrightarrow\;\text{M-stationarity}\;\Longrightarrow\;\text{W-stationarity.} \] Moreover, the KKT conditions of \eqref{eq:MPSC} are equivalent to the S-stationarity conditions from Definition \ref{def:stationarities}. One may check Figure \ref{fig:stationarities} for a geometric interpretation of the Lagrange multipliers associated with the switching conditions from $I^{GH}(\bar x)$. \begin{figure}[h]\centering \includegraphics[width=4.0cm]{WeakStationarity.pdf} \includegraphics[width=4.0cm]{MordukhovichStationarity.pdf} \includegraphics[width=4.0cm]{StrongStationarity.pdf} \caption{Geometric illustrations of weak, M-, and S-stationarity for an index $l\in I^{GH}(\bar x)$.} \label{fig:stationarities} \end{figure} In order to ensure that one of these stationarity notions plays the role of a necessary optimality condition for \eqref{eq:MPSC}, suitable problem-tailored constraint qualifications need to be valid. For the definition of such conditions, the following so-called \emph{tightened nonlinear problem} is of interest: \begin{equation}\label{eq:TNLP}\tag{TNLP} \begin{aligned} f(x)&\,\rightarrow\,\min\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&\qquad&j\in\mathcal P,&\\ G_l(x)&\,=\,0,&&l\in I^{G}(\bar x)\cup I^{GH}(\bar x),&\\ H_l(x)&\,=\,0,&&l\in I^{H}(\bar x)\cup I^{GH}(\bar x). \end{aligned} \end{equation} Note that \eqref{eq:TNLP} is a standard nonlinear program. \begin{definition}\label{def:MPSC_CQs} Let $\bar x\in X$ be a feasible point of \eqref{eq:MPSC}. Then \emph{MPSC-LICQ (MPSC-MFCQ)} is said to hold for \eqref{eq:MPSC} at $\bar x$ if LICQ (MFCQ) holds for \eqref{eq:TNLP} at $\bar x$, i.e., if the vectors \[ \begin{split} \{\nabla g_i(\bar x)\,|\,i\in I^g(\bar x)\}\cup \Bigl[ &\{\nabla h_j(\bar x)\,|\,j\in\mathcal P\}\\ &\cup\{\nabla G_l(\bar x)\,|\,l\in I^G(\bar x)\cup I^{GH}(\bar x)\}\\ &\cup\{\nabla H_l(\bar x)\,|\,l\in I^H(\bar x)\cup I^{GH}(\bar x)\}\Bigr] \end{split} \] are linearly independent (positive-linearly independent). \end{definition} It is obvious that MPSC-LICQ is stronger than MPSC-MFCQ. Furthermore, MPSC-LICQ implies standard GCQ for \eqref{eq:MPSC} at the reference point. In this paper, we will use another MPSC-tailored constraint qualification called MPSC-NNAMCQ, where NNAMCQ stands for the \emph{No Nonzero Abnormal Multiplier Constraint Qualification} which has been introduced to investigate optimization problems whose feasible sets are preimages of closed but not necessarily convex sets under continuously differentiable mappings, see \cite[Section~6.D]{RockafellarWets1998} and \cite{YeYe1997}. Clearly, \eqref{eq:MPSC} belongs to this problem class as well if one reformulates the switching constraints as \begin{equation*} (G_l(x),H_l(x))\in C,\qquad l\in\mathcal Q, \end{equation*} with $C:=\{(a,b)\in\R^2\,|\,ab=0\}$. \begin{definition}\label{def:MPSC_NNAMCQ} Let $\bar x\in X$ be a feasible point of \eqref{eq:MPSC}. Then \emph{MPSC-NNAMCQ} is said to hold for \eqref{eq:MPSC} at $\bar x$ if the following condition is valid: \[ \left. \begin{aligned} &0=\sum\limits_{i\in I^g(\bar x)}\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\bar x)\\ &\qquad\qquad +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\bar x) +\nu_l\nabla H_l(\bar x) \Bigr]\\ &\forall i\in I^{g}(\bar x)\colon\,\lambda_i\geq 0,\\ &\forall l\in I^H(\bar x)\colon\,\mu_l=0,\\ &\forall l\in I^{G}(\bar x)\colon\,\nu_l=0,\\ &\forall l\in I^{GH}(\bar x)\colon\,\mu_l\nu_l=0 \end{aligned} \right\} \, \Longrightarrow \, \lambda=0,\,\rho=0,\,\mu=0,\,\nu=0. \] \end{definition} It is easy to check that MPSC-NNAMCQ is implied by MPSC-MFCQ. Note that a local minimizer of \eqref{eq:MPSC} where MPSC-LICQ holds is an S-stationary point. Furthermore, one can easily check that the associated multipliers which solve the system of S-stationarity are uniquely determined. Under MPSC-MFCQ (and, thus, MPSC-NNAMCQ), a local minimizer of \eqref{eq:MPSC} is, in general, only M-stationary. Finally, there exist several problem-tailored constraint qualifications for \eqref{eq:MPSC} which are weaker than MPSC-MFCQ but also imply that local solutions are M-stationary, see \cite{Mehlitz2018}. \section{The relaxation scheme and its convergence properties}\label{sec:relaxation_scheme_KS} \subsection{On the relaxation scheme}\label{sec:relaxation} For our relaxation approach, we will make use of the function $\varphi\colon\R^2\to\R$ defined below: \[ \forall (a,b)\in\R^2\colon \quad \varphi(a,b):= \begin{cases} ab &\text{if }a+b\geq 0,\\ -\tfrac12(a^2+b^2) &\text{if }a+b<0. \end{cases} \] The function $\varphi$ was introduced in \cite{KanzowSchwartz2013} to study a relaxation method for the numerical solution of MPCCs. In the following lemma, which parallels \cite[Lemma~3.1]{KanzowSchwartz2013}, some properties of $\varphi$ are summarized. \begin{lemma}\label{lem:properties_of_varphi} \begin{enumerate} \item[(a)] The function $\varphi$ is an NCP-function, i.e.\ it satisfies \[ \forall (a,b)\in\R^2\colon \quad \varphi(a,b)=0\,\Longleftrightarrow\,a\geq 0\,\land\,b\geq 0\,\land ab=0. \] \item[(b)] The function $\varphi$ is continuously differentiable and satisfies \[ \forall (a,b)\in\R^2\colon \quad \nabla\varphi(a,b)= \begin{cases} \begin{pmatrix} b\\a \end{pmatrix} &\text{if }a+b\geq 0,\\ \begin{pmatrix} -a\\-b \end{pmatrix} &\text{if }a+b<0. \end{cases} \] \end{enumerate} \end{lemma} For some parameter $t\geq 0$ as well as indices $s\in\mathcal S:=\{1,2,3,4\}$ and $l\in\mathcal Q$, we define functions $\Phi^s_l(\cdot;t)\colon\R^n\to\R$ via \[ \begin{aligned} \Phi^1_l(x;t)&:=\varphi(G_l(x)-t,H_l(x)-t),&\quad \Phi^2_l(x;t)&:=\varphi(-G_l(x)-t,H_l(x)-t),&\\ \Phi^3_l(x;t)&:=\varphi(-G_l(x)-t,-H_l(x)-t),&\quad \Phi^4_l(x;t)&:=\varphi(G_l(x)-t,-H_l(x)-t)& \end{aligned} \] for any $x\in\R^n$. Now, we are in position to introduce the surrogate problem of our interest: \begin{equation}\label{eq:relaxedMPSC}\tag{P$(t)$} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0&\qquad&i\in\mathcal M&\\ h_j(x)&\,=\,0&&j\in\mathcal P&\\ \Phi^s_l(x;t)&\,\leq\,0&&s\in\mathcal S,\,l\in\mathcal Q. \end{aligned} \end{equation} The feasible set of \eqref{eq:relaxedMPSC} will be denoted by $X(t)$. Figure \ref{fig:relaxed_feasible_set} provides an illustration of $X(t)$. \begin{figure}[h]\centering \includegraphics[width=4.0cm]{RelaxedFeasibleSet.pdf} \caption{Geometric interpretation of the relaxed feasible set $X(t)$.} \label{fig:relaxed_feasible_set} \end{figure} Let $x\in X(t)$ be a feasible point of \eqref{eq:relaxedMPSC} for some fixed parameter $t>0$. Later on, it will be beneficial to work with the index sets defined below: \begin{align*} I^{00}_{t,1}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=t,\,H_l(x)=t\},& I^{00}_{t,2}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=-t,\,H_l(x)=t\},&\\ I^{0+}_{t,1}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=t,\,H_l(x)>t\},& I^{0+}_{t,2}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=-t,\,H_l(x)>t\},&\\ I^{+0}_{t,1}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)>t,\,H_l(x)=t\},& I^{-0}_{t,2}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)<-t,\,H_l(x)=t\},&\\ I^{00}_{t,3}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=-t,\,H_l(x)=-t\},& I^{00}_{t,4}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=t,\,H_l(x)=-t\},&\\ I^{0-}_{t,3}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=-t,\,H_l(x)<-t\},& I^{0-}_{t,4}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)=t,\,H_l(x)<-t\},&\\ I^{-0}_{t,3}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)<-t,\,H_l(x)=-t\},& I^{+0}_{t,4}(x)&:=\{l\in\mathcal Q\,|\,G_l(x)>t,\,H_l(x)=-t\}.& \end{align*} Note that all these sets are pairwise disjoint. The index sets $I_{t,1}^{00}(x)$, $I_{t,1}^{0+}(x)$, and $I_{t,1}^{+0}(x)$ subsume the three possible cases where the constraints $\Phi_l^1(x;t)\le 0$ ($l\in \mathcal{Q}$) are active. Similarly, the other index sets cover those indices where the constraints $\Phi_l^s(x;t)\le 0$ ($l\in\mathcal{Q}$, $s\in\{2,3,4\}$) are active. It follows that an index $l\in\mathcal{Q}$ which does not belong to any of the above sets is inactive for \eqref{eq:relaxedMPSC} and can therefore be disregarded (locally). In order to address any of the four quadrants separately, we will exploit \begin{align*} I^0_{t,1}(x)&:=I^{00}_{t,1}(x)\cup I^{0+}_{t,1}(x)\cup I^{+0}_{t,1}(x),& I^0_{t,2}(x)&:=I^{00}_{t,2}(x)\cup I^{0+}_{t,2}(x)\cup I^{-0}_{t,2}(x),\\ I^0_{t,3}(x)&:=I^{00}_{t,3}(x)\cup I^{0-}_{t,3}(x)\cup I^{-0}_{t,3}(x),& I^0_{t,4}(x)&:=I^{00}_{t,4}(x)\cup I^{0-}_{t,4}(x)\cup I^{+0}_{t,4}(x), \end{align*} i.e., for fixed $s\in\mathcal S$, $I_{t,s}^0(x)$ collects all indices $l\in\mathcal{Q}$ where the constraint $\Phi_l^s(x;t)\leq 0$ is active. For brevity, we set \begin{align*} I^{00}_t(x)&:=\bigcup_{s\in \mathcal S}I^{00}_{t,s}(x),\\ I^{0\pm}_t(x)&:=I^{0+}_{t,1}(x)\cup I^{0+}_{t,2}(x)\cup I^{0-}_{t,3}(x)\cup I^{0-}_{t,4}(x),\\ I^{\pm 0}_t(x)&:=I^{+0}_{t,1}(x)\cup I^{-0}_{t,2}(x)\cup I^{-0}_{t,3}(x)\cup I^{+0}_{t,4}(x). \end{align*} Thus, we collect all indices in $I^{0\pm}_t(x)$ where $G_l(x)\in\{-t,t\}$ holds while $|H_l(x)|>t$ is valid. Similarly, $I^{\pm 0}_t(x)$ comprises all indices where $H_l(x)\in\{-t,t\}$ and $|G_l(x)|>t$ hold true. The set $I^{00}_t(x)$ contains all those indices where $G_l(x),H_l(x)\in\{-t,t\}$ is valid. If $\bar x\in X$ is feasible to \eqref{eq:MPSC}, $x$ lies in a sufficiently small neighborhood of $\bar x$, and $t\geq 0$ is sufficiently small, then the following inclusions follow from the continuity of all appearing functions: \begin{equation}\label{eq:upper_estimate_index_sets} \begin{split} I^{00}_{t}(x)\cup I^{0\pm}_{t}(x) &\subset I^{G}(\bar x)\cup I^{GH}(\bar x),\\ I^{00}_{t}(x)\cup I^{\pm 0}_t(x) &\subset I^{H}(\bar x)\cup I^{GH}(\bar x). \end{split} \end{equation} In the lemma below, we present explicit formulas for the gradients of $\Phi^s_l(\cdot;t)$ with $l\in\mathcal Q$ and $s\in\mathcal S$. They can be derived exploiting Lemma \ref{lem:properties_of_varphi} as well as the chain rule. \begin{lemma}\label{lem:derivative_of_Phi} For $x\in\R^n$, $t>0$, and $l\in\mathcal Q$, the following formulas are valid: \begin{align*} \nabla_x\Phi^1_l(x;t)&= \begin{cases} (H_l(x)-t)\nabla G_l(x)+(G_l(x)-t)\nabla H_l(x) &\text{if }G_l(x)+H_l(x)\geq 2t,\\ -(G_l(x)-t)\nabla G_l(x)-(H_l(x)-t)\nabla H_l(x) &\text{if }G_l(x)+H_l(x)<2t, \end{cases}\\ \nabla_x\Phi^2_l(x;t)&= \begin{cases} (t-H_l(x))\nabla G_l(x)-(G_l(x)+t)\nabla H_l(x) &\text{if }-G_l(x)+H_l(x)\geq 2t,\\ -(G_l(x)+t)\nabla G_l(x)-(H_l(x)-t)\nabla H_l(x) &\text{if }-G_l(x)+H_l(x)<2t, \end{cases}\\ \nabla_x\Phi^3_l(x;t)&= \begin{cases} (H_l(x)+t)\nabla G_l(x)+(G_l(x)+t)\nabla H_l(x) &\text{if }-G_l(x)-H_l(x)\geq 2t,\\ -(G_l(x)+t)\nabla G_l(x)-(H_l(x)+t)\nabla H_l(x) &\text{if }-G_l(x)-H_l(x)<2t, \end{cases}\\ \nabla_x\Phi^4_l(x;t)&= \begin{cases} -(H_l(x)+t)\nabla G_l(x)+(t-G_l(x))\nabla H_l(x) &\text{if }G_l(x)-H_l(x)\geq 2t,\\ -(G_l(x)-t)\nabla G_l(x)-(H_l(x)+t)\nabla H_l(x) &\text{if }G_l(x)-H_l(x)<2t. \end{cases} \end{align*} Particularly, we have \begin{align*} &\forall l\in I^{0}_{t,1}(x)\colon\quad& \nabla_x\Phi^1_l(x;t)&= \begin{cases} (G_l(x)-t)\nabla H_l(x) & \text{if }l\in I^{+0}_{t,1}(x),\\ (H_l(x)-t)\nabla G_l(x) & \text{if }l\in I^{0+}_{t,1}(x),\\ 0 & \text{if }l\in I^{00}_{t,1}(x), \end{cases}&\\ &\forall l\in I^{0}_{t,2}(x)\colon\quad& \nabla_x\Phi^2_l(x;t)&= \begin{cases} -(G_l(x)+t)\nabla H_l(x) & \text{if }l\in I^{-0}_{t,2}(x),\\ (t-H_l(x))\nabla G_l(x) & \text{if }l\in I^{0+}_{t,2}(x),\\ 0 & \text{if }l\in I^{00}_{t,2}(x), \end{cases}&\\ &\forall l\in I^{0}_{t,3}(x)\colon\quad& \nabla_x\Phi^3_l(x;t)&= \begin{cases} (G_l(x)+t)\nabla H_l(x) & \text{if }l\in I^{-0}_{t,3}(x),\\ (H_l(x)+t)\nabla G_l(x) & \text{if }l\in I^{0-}_{t,3}(x),\\ 0 & \text{if }l\in I^{00}_{t,3}(x), \end{cases}&\\ &\forall l\in I^{0}_{t,4}(x)\colon\quad& \nabla_x\Phi^4_l(x;t)&= \begin{cases} (t-G_l(x))\nabla H_l(x) & \text{if }l\in I^{+0}_{t,4}(x),\\ -(H_l(x)+t)\nabla G_l(x) & \text{if }l\in I^{0-}_{t,4}(x),\\ 0 & \text{if }l\in I^{00}_{t,4}(x). \end{cases}& \end{align*} \end{lemma} As a corollary of the above lemma, we obtain an explicit formula for the linearization cone associated with \eqref{eq:relaxedMPSC}. \begin{corollary}\label{cor:linearization_cone_relaxedMPSC} Fix $t>0$ and a feasible point $x\in X(t)$ of \eqref{eq:relaxedMPSC}. Then the following formula is valid: \[ \mathcal L_{X(t)}(x) = \left\{ d\in\R^n\,\middle|\, \begin{aligned} \nabla g_i(x)\cdot d &\,\leq\,0& &i\in I^g(x)\\ \nabla h_j(x)\cdot d &\,\leq\,0& &j\in\mathcal P\\ \nabla G_l(x)\cdot d &\,\leq\,0& &l\in I^{0+}_{t,1}(x)\cup I^{0-}_{t,4}(x)\\ \nabla G_l(x)\cdot d &\,\geq\,0& &l\in I^{0+}_{t,2}(x)\cup I^{0-}_{t,3}(x)\\ \nabla H_l(x)\cdot d &\,\leq\,0& &l\in I^{+0}_{t,1}(x)\cup I^{-0}_{t,2}(x)\\ \nabla H_l(x)\cdot d &\,\geq\,0& &l\in I^{-0}_{t,3}(x)\cup I^{+0}_{t,4}(x) \end{aligned} \right\}. \] \end{corollary} The upcoming lemma justifies that \eqref{eq:relaxedMPSC} is indeed a \emph{relaxation} of \eqref{eq:MPSC}. Its proof parallels the one of \cite[Lemma~3.2]{KanzowSchwartz2013}. \begin{lemma}\label{lem:properties_of_relaxed_feasible_set} The following statements hold: \begin{enumerate} \item[\textup{(P1)}] $X(0)=X$, \item[\textup{(P2)}] $0\leq t_1\leq t_2\,\Longrightarrow\,X(t_1)\subset X(t_2)$, \item[\textup{(P3)}] $\bigcap_{t> 0}X(t)=X$. \end{enumerate} \end{lemma} Now, we are in position to characterize a conceptual method for the numerical solution of \eqref{eq:MPSC}: First, a sequence $\{t_k\}_{k\in\N}\subset\R_+$ of positive relaxation parameters is chosen which converges to zero. Next, one solves the surrogate problem \hyperref[eq:relaxedMPSC]{(P$(t_k)$)} via standard methods. If one computes a (local) minimizer of one of these surrogate problems which is feasible to \eqref{eq:MPSC}, then it is already a (local) minimizer of \eqref{eq:MPSC}. In general, it will be only possible to compute KKT points of the surrogate problem. However, if such a sequence of KKT points converges to some point $\bar x\in\R^n$, then this point must be feasible to \eqref{eq:MPSC} by construction. Furthermore, it will be shown in Section \ref{sec:convergence_properties} that whenever MPSC-NNAMCQ is valid at $\bar x$, then it is an M-stationary point of \eqref{eq:MPSC}. \subsection{Convergence properties}\label{sec:convergence_properties} In this section, we analyze the theoretical properties of our relaxation scheme. In order to do so, we fix some standing assumptions below. \begin{assumption}\label{ass:convergence_analysis} Let $\{t_k\}_{k\in\N}\subset\R_+$ be a sequence of positive relaxation parameters converging to zero. For each $k\in\N$, let $x_k\in X(t_k)$ be a KKT point of \hyperref[eq:relaxedMPSC]{\textup{(P$(t_k)$)}}. We assume that $\{x_k\}_{k\in\N}$ converges to $\bar x\in \R^n$. Note that $\bar x\in X$ holds by Lemma~\ref{lem:properties_of_relaxed_feasible_set}. \end{assumption} First, we will show that whenever MPSC-NNAMCQ is valid at $\bar x$, then it is an M-stationary point of \eqref{eq:MPSC}. Second, it will be demonstrated that MPSC-LICQ at $\bar x$ implies that GCQ holds for the surrogate problem \eqref{eq:relaxedMPSC} at all feasible points from a sufficiently small neighborhood of $\bar x$ and sufficiently small $t>0$. This property ensures that local minima of the surrogate problem \eqref{eq:relaxedMPSC} which are located near $\bar x$ are in fact KKT points. This way, it is shown that Assumption \ref{ass:convergence_analysis} is reasonable. \begin{theorem}\label{thm:convergence_to_M-stationary_point} Let Assumption \ref{ass:convergence_analysis} be valid. Suppose that MPSC-NNAMCQ holds at $\bar x$. Then $\bar x$ is an M-stationary point of \eqref{eq:MPSC}. \end{theorem} \begin{proof} Since $x_k$ is a KKT point of \hyperref[eq:relaxedMPSC]{(P$(t_k)$)}, there exist multipliers $\lambda^k\in\R^m$, $\rho^k\in\R^p$, and $\alpha^k,\beta^k,\gamma^k,\delta^k\in\R^q$ which solve the following system: \[ \begin{split} &0=\nabla f(x_k) +\sum\limits_{i\in I^g(x_k)}\lambda^k_i\nabla g_i(x_k) +\sum\limits_{j\in\mathcal P}\rho^k_j\nabla h_j(x_k)\\ &\qquad\qquad +\sum\limits_{l\in\mathcal Q} \Bigl[ \alpha^k_l\nabla_x\Phi^1_l(x_k;t_k)+\beta^k_l\nabla_x\Phi^2_l(x_k;t_k) +\gamma^k_l\nabla_x\Phi^3_l(x_k;t_k)+\delta^k_l\nabla_x\Phi^4_l(x_k;t_k) \Bigr],\\ &\forall i\in I^g(x_k)\colon\,\lambda^k_i\geq 0;\,\forall i\in\mathcal M\setminus I^g(x_k)\colon\,\lambda^k_l=0,\\ &\forall l\in I^{0}_{t_k,1}(x_k)\colon\,\alpha^k_l\geq 0;\,\forall l\in\mathcal Q\setminus I^{0}_{t_k,1}(x_k)\colon\,\alpha^k_l= 0,\\ &\forall l\in I^{0}_{t_k,2}(x_k)\colon\,\beta^k_l\geq 0;\,\forall l\in\mathcal Q\setminus I^{0}_{t_k,2}(x_k)\colon\,\beta^k_l= 0,\\ &\forall l\in I^{0}_{t_k,3}(x_k)\colon\,\gamma^k_l\geq 0;\,\forall l\in\mathcal Q\setminus I^{0}_{t_k,3}(x_k)\colon\,\gamma^k_l= 0,\\ &\forall l\in I^{0}_{t_k,4}(x_k)\colon\,\delta^k_l\geq 0;\,\forall l\in\mathcal Q\setminus I^{0}_{t_k,4}(x_k)\colon\,\delta^k_l= 0. \end{split} \] Next, let us define new multipliers $\alpha_G^k,\alpha_H^k,\beta_G^k,\beta^k_H,\gamma^k_G,\gamma^k_H,\delta^k_G,\delta^k_H\in\R^q$ as stated below for all $l\in\mathcal Q$: \begin{align*} \alpha^k_{G,l}&:= \begin{cases} \alpha^k_l(H_l(x_k)-t_k) &l\in I^{0+}_{t_k,1}(x_k),\\ 0 &\text{otherwise,} \end{cases} & \alpha^k_{H,l}&:= \begin{cases} \alpha^k_l(G_l(x_k)-t_k) &l\in I^{+0}_{t_k,1}(x_k),\\ 0 &\text{otherwise,} \end{cases} &\\ \beta^k_{G,l}&:= \begin{cases} \beta^k_l(t_k-H_l(x_k)) &l\in I^{0+}_{t_k,2}(x_k),\\ 0 &\text{otherwise,} \end{cases} & \beta^k_{H,l}&:= \begin{cases} \beta^k_l(-G_l(x_k)-t_k) &l\in I^{-0}_{t_k,2}(x_k),\\ 0 &\text{otherwise,} \end{cases} &\\ \gamma^k_{G,l}&:= \begin{cases} \gamma^k_l(H_l(x_k)+t_k) &l\in I^{0-}_{t_k,3}(x_k),\\ 0 &\text{otherwise,} \end{cases} & \gamma^k_{H,l}&:= \begin{cases} \gamma^k_l(G_l(x_k)+t_k) &l\in I^{-0}_{t_k,3}(x_k),\\ 0 &\text{otherwise,} \end{cases} &\\ \delta^k_{G,l}&:= \begin{cases} \delta^k_l(-H_l(x_k)-t_k) &l\in I^{0-}_{t_k,4}(x_k),\\ 0 &\text{otherwise,} \end{cases} & \delta^k_{H,l}&:= \begin{cases} \delta^k_l(t_k-G_l(x_k)) &l\in I^{+0}_{t_k,4}(x_k),\\ 0 &\text{otherwise.} \end{cases} & \end{align*} Furthermore, we set $\mu^k:=\alpha^k_G+\beta^k_G+\gamma^k_G+\delta^k_G$ and $\nu^k:=\alpha^k_H+\beta^k_H+\gamma^k_H+\delta^k_H$. By definition, we have $\operatorname{supp} \mu^k\subset I^{0\pm}_{t_k}(x_k)$ as well as $\operatorname{supp}\nu^k\subset I^{\pm 0}_{t_k}(x_k)$. Thus, for sufficiently large $k\in\N$, \eqref{eq:upper_estimate_index_sets} yields \[ \operatorname{supp}\mu^k\subset I^{G}(\bar x)\cup I^{GH}(\bar x),\qquad \operatorname{supp}\nu^k\subset I^{H}(\bar x)\cup I^{GH}(\bar x). \] Using Lemma \ref{lem:derivative_of_Phi}, \begin{equation}\label{eq:KKT_new_multipliers} 0=\nabla f(x_k)+\sum\limits_{i\in I^g(x_k)}\lambda^k_i\nabla g_i(x_k) +\sum\limits_{j\in\mathcal P}\rho^k_j\nabla h_j(x_k)\\ +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu^k_l\nabla G_l(x_k)+\nu^k_l\nabla H_l(x_k) \Bigr] \end{equation} is obtained. Next, it will be shown that the sequence $\{(\lambda^k,\rho^k,\mu^k,\nu^k)\}_{k\in\N}$ is bounded. Assuming the contrary, we define \[ \forall k\in\N\colon \quad (\tilde\lambda^k,\tilde\rho^k,\tilde\mu^k,\tilde\nu^k) := \frac{(\lambda^k,\rho^k,\mu^k,\nu^k)}{\norm{(\lambda^k,\rho^k,\mu^k,\nu^k)}{2}}. \] Clearly, $\{(\tilde\lambda^k,\tilde\rho^k,\tilde\mu^k,\tilde\nu^k)\}_{k\in\N}$ is bounded and, thus, possesses a converging subsequence (without relabeling) with nonvanishing limit $(\tilde\lambda,\tilde\rho,\tilde\mu,\tilde\nu)$. The continuity of $g$ yields $\operatorname{supp}\tilde\lambda\subset I^g(\bar x)$. The above considerations yield \[ \operatorname{supp}\tilde\mu\subset I^{G}(\bar x)\cup I^{GH}(\bar x), \quad \operatorname{supp} \tilde\nu\subset I^{H}(\bar x)\cup I^{GH}(\bar x). \] Choose $l\in I^{GH}(\bar x)$ arbitrarily. If $\tilde\mu_l\neq 0$ holds true, then $\tilde\mu^k_l\neq 0$ must be valid for sufficiently large $k\in\N$. By definition of $\mu^k_l$, $l\in I^{0\pm}_{t_k}(x_k)$ follows. Since $I^{0\pm}_{t_k}(x_k)$ and $I^{\pm 0}_{t_k}(x_k)$ are disjoint, $\nu^k_l=0$ holds for sufficiently large $k\in\N$. This yields $\tilde\nu^k_l=0$ for sufficiently large $k\in\N$, i.e.\ $\tilde\nu_l=0$ is obtained. This shows \[\forall l\in I^{GH}(\bar x)\colon\quad \tilde\mu_l\tilde\nu_l=0.\] Dividing \eqref{eq:KKT_new_multipliers} by $\norm{(\lambda^k,\rho^k,\mu^k,\nu^k)}{2}$, taking the limit $k\to\infty$, respecting the continuous differentiability of $f$, $g$, $h$, $G$, as well as $H$, and invoking the above arguments, we obtain \[ \begin{split} &0=\sum\limits_{i\in I^g(\bar x)}\tilde\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in\mathcal P}\tilde\rho_j\nabla h_j(\bar x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \tilde\mu_l\nabla G_l(\bar x)+\tilde\nu_l\nabla H_l(\bar x) \Bigr],\\ &\forall i\in I^g(\bar x)\colon\,\tilde\lambda_i\geq 0;\,\forall i\in\mathcal M\setminus I^g(\bar x)\colon\,\tilde\lambda_i=0,\\ &\forall l\in I^H(\bar x)\colon\,\tilde\mu_l=0,\\ &\forall l\in I^G(\bar x)\colon\,\tilde\nu_l=0,\\ &\forall l\in I^{GH}(\bar x)\colon\,\tilde\mu_l\tilde\nu_l=0. \end{split} \] Due to the fact that $(\tilde\lambda,\tilde\rho,\tilde\mu,\tilde\nu)$ does not vanish, this is a contradiction to the validity of MPSC-NNAMCQ. Thus, $\{(\lambda^k,\rho^k,\mu^k,\nu^k)\}_{k\in\N}$ is bounded. We assume w.l.o.g.\ that $\{(\lambda^k,\rho^k,\mu^k,\nu^k)\}_{k\in\N}$ converges to $(\lambda,\rho,\mu,\nu)$ (otherwise, we choose an appropriate subsequence). Reprising the above arguments, we have \[ \begin{split} &\forall i\in I^g(\bar x)\colon\,\lambda_i\geq 0;\,\forall i\in\mathcal M\setminus I^g(\bar x)\colon\,\lambda_i=0,\\ &\forall l\in I^H(\bar x)\colon\,\mu_l=0,\\ &\forall l\in I^G(\bar x)\colon\,\nu_l=0,\\ &\forall l\in I^{GH}(\bar x)\colon\,\mu_l\nu_l=0. \end{split} \] Taking the limit in \eqref{eq:KKT_new_multipliers} and respecting the continuous differentiability of all appearing mappings, we obtain \[ 0=\nabla f(\bar x)+\sum\limits_{i\in I^g(\bar x)}\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\bar x)\\ +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\bar x)+\nu_l\nabla H_l(\bar x) \Bigr]. \] This shows that $\bar x$ is an M-stationary point of \eqref{eq:MPSC}. \end{proof} In order to show that the validity of MPSC-LICQ at the limit point $\bar x$ ensures that GCQ holds at all feasible points of \eqref{eq:relaxedMPSC} sufficiently close to $\bar x$ where $t$ is sufficiently small, we need to study the variational geometry of the sets $X(t)$ in some more detail. Fix $t>0$ and some point $\tilde x\in X(t)$. For an arbitrary index set $I\subset I^{00}_t(\tilde x)$, we consider the subsequent program: \begin{equation}\label{eq:relaxedMPSC_decomposed}\tag{P$(t,\tilde x,I)$} \begin{aligned} f(x)&\,\rightarrow\,\min\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,\\ h_j(x)&\,=\,0,&&j\in\mathcal P,\\ -t\,\leq\,G_l(x)&\,\leq\,t,&&l\in I^{0\pm}_t(\tilde x)\cup I,\\ -t\,\leq\,H_l(x)&\,\leq\,t,&&l\in I^{\pm 0}_t(\tilde x)\cup (I^{00}_t(\tilde x)\setminus I),\\ \Phi^s_l(x;t)&\,\leq\,0,&&l\in\mathcal Q\setminus I^0_{t,s}(\tilde x),\,s\in\mathcal S.& \end{aligned} \end{equation} The feasible set of \eqref{eq:relaxedMPSC_decomposed} will be denoted by $X(t,\tilde x,I)$. Clearly, $\tilde x$ is a feasible point of \eqref{eq:relaxedMPSC_decomposed} for arbitrary $I\subset I^{00}_t(\tilde x)$. Furthermore, $X(t,\tilde x,I)\subset X(t)$ is valid for any choice of $I\subset I^{00}_t(\tilde x)$. \begin{lemma}\label{lem:decomposition_of_tangent_cone} For fixed $t>0$ and $\tilde x\in X(t)$, we have \[ \mathcal T_{X(t)}(\tilde x) = \bigcup\limits_{I\subset I^{00}_{t}(\tilde x)}\mathcal T_{X(t,\tilde x,I)}(\tilde x). \] \end{lemma} \begin{proof} We show both inclusions separately.\\ ``$\subset$'' Fix an arbitrary direction $d\in\mathcal T_{X(t)}(\tilde x)$. Then we find sequences $\{y_k\}_{k\in\N}\subset X(t)$ and $\{\tau_k\}_{k\in\N}\subset\R_+$ such that $y_k\to \tilde{x}$, $\tau_k\downarrow 0$, and $(y_k-\tilde x)/\tau_k\to d$ as $k\to\infty$. It is sufficient to verify the existence of an index set $\bar I\subset I^{00}_t(\tilde x)$ such that $\{y_k\}_{k\in\N}\cap X(t,\tilde x,\bar I)$ possesses infinite cardinality since this already gives us $d\in\mathcal T_{X(t,\tilde x,\bar I)}(\tilde x)$. Fix $k\in\N$ sufficiently large and $l\in I^{0\pm}_{t}(\tilde x)$. Then, due to continuity of $G_l$, $H_l$, as well as $\varphi$ and feasibility of $y_k$ to \eqref{eq:relaxedMPSC}, we either have $l\in I^{0\pm}_{t}(y_k)$ and, thus, $G_l(y_k)\in\{-t,t\}$, or $-t< G_l(y_k)<t$. Similarly, we obtain $-t\leq H_l(y_k)\leq t$ for all $l\in I^{\pm0}_t(\tilde x)$. Due to feasibility of $y_k$ to \eqref{eq:relaxedMPSC}, we have $-t\leq G_l(y_k)\leq t$ or $-t\leq H_l(y_k)\leq t$. Thus, setting \[ I_k:=I^{00}_t(\tilde x)\cap\{l\in\mathcal Q\,|\,-t\leq G_l(y_k)\leq t\}, \] $y_k\in X(t,\tilde x,I_k)$ is valid. Since there are only finitely many subsets of $I^{00}_t(\tilde x)$ while $\{y_k\}_{k\in\N}$ is infinite, there must exist $\bar I\subset I^{00}_t(\tilde x)$ such that $\{y_k\}_{k\in\N}\cap X(t,\tilde x,\bar I)$ is of infinite cardinality.\\ ``$\supset$'' By definition of the tangent cone, we easily obtain $\mathcal T_{X(t,\tilde x,I)}(\tilde x)\subset\mathcal T_{X(t)}(\tilde x)$ for any $I\subset I^{00}_t(\tilde x)$. Taking the union over all subsets of $I^{00}_t(\tilde x)$ yields the desired inclusion. \end{proof} \begin{theorem}\label{thm:MPSC_LICQ_yields_GCQ} Let $\bar x\in X$ be a feasible point of \eqref{eq:MPSC} where MPSC-LICQ is satisfied. Then there exist $\bar t>0$ and a neighborhood $U\subset\R^n$ of $\bar x$ such that GCQ holds for \eqref{eq:relaxedMPSC} at all points from $X(t)\cap U$ for all $t\in (0,\bar{t}]$. \end{theorem} \begin{proof} Due to the validity of MPSC-LICQ at $\bar x$ and the continuous differentiability of $g$, $h$, $G$, and $H$, the gradients \[ \begin{split} \{\nabla g_i(x)\,|\,i\in I^g(\bar x)\}&\cup\{\nabla h_j(x)\,|\,j\in\mathcal P\}\\ &\cup\{\nabla G_l(x)\,|\,l\in I^G(\bar x)\cup I^{GH}(\bar x)\}\\ &\cup\{\nabla H_l(x)\,|\,l\in I^H(\bar x)\cup I^{GH}(\bar x)\} \end{split} \] are linearly independent for all $x$ which are chosen from a sufficiently small neighborhood $\mathcal V$ of $\bar x$, see Lemma \ref{lem:local_stability_of_positive_linear_independence}. Invoking \eqref{eq:upper_estimate_index_sets}, we can choose a neighborhood $U\subset\mathcal V$ of $\bar x$ and $\bar t>0$ such that for any $\tilde x\in X(t)\cap U$, where $t\in (0,\bar{t}]$ holds, we have \begin{equation}\label{eq:ThmMPSCLICQyieldsGCQ1} \begin{split} I^g(\tilde x)&\subset I^g(\bar x),\\ I^{00}_t(\tilde x)\cup I^{0\pm}_t(\tilde x)&\subset I^G(\bar x)\cup I^{GH}(\bar x),\\ I^{00}_t(\tilde x)\cup I^{\pm 0}_t(\tilde x)&\subset I^{H}(\bar x)\cup I^{GH}(\bar x). \end{split} \end{equation} Particularly, for any such $\tilde x\in\R^n$ and $I\subset I^{00}_t(\tilde x)$, the gradients \[ \begin{split} \{\nabla g_i(\tilde x)\,|\,i\in I^g(\tilde x)\}&\cup\{\nabla h_j(\tilde x)\,|\,j\in\mathcal P\}\\ &\cup\{\nabla G_l(\tilde x)\,|\,l\in I^{0\pm}_t(\tilde x)\cup I\}\\ &\cup\{\nabla H_l(\tilde x)\,|\,l\in I^{\pm 0}_t(\tilde x)\cup (I^{00}_t(\tilde x)\setminus I)\} \end{split} \] are linearly independent, i.e., standard LICQ is valid for \eqref{eq:relaxedMPSC_decomposed} at $\tilde x$ for any set $I\subset I^{00}_t(\tilde x)$. This implies $\mathcal T_{X(t,\tilde x,I)}(\tilde x)=\mathcal L_{X(t,\tilde x,I)}(\tilde x)$ for any $I\subset I^{00}_t(\tilde x)$. Exploiting Lemma \ref{lem:decomposition_of_tangent_cone}, we obtain \[ \mathcal T_{X(t)}(\tilde x) = \bigcup\limits_{I\subset I^{00}_t(\tilde x)}\mathcal L_{X(t,\tilde x,I)}(\tilde x). \] Computing the polar cone on both sides yields \begin{equation}\label{eq:tangents_via_intersection_of_linearized_tangents} \mathcal T_{X(t)}(\tilde x)^\circ = \bigcap\limits_{I\subset I^{00}_t(\tilde x)}\mathcal L_{X(t,\tilde x,I)}(\tilde x)^\circ. \end{equation} Define \[ \begin{split} \mathcal I^+_G(t,\tilde x,I)&:=I^{0+}_{t,1}(\tilde x)\cup I^{0-}_{t,4}(\tilde x) \cup\bigl[I\cap \bigl(I^{00}_{t,1}(\tilde x)\cup I^{00}_{t,4}(\tilde x)\bigr)\bigr],\\ \mathcal I^-_G(t,\tilde x,I)&:=I^{0+}_{t,2}(\tilde x)\cup I^{0-}_{t,3}(\tilde x) \cup\bigl[I\cap \bigl(I^{00}_{t,2}(\tilde x)\cup I^{00}_{t,3}(\tilde x)\bigr)\bigr],\\ \mathcal I^+_H(t,\tilde x,I)&:=I^{+0}_{t,1}(\tilde x)\cup I^{-0}_{t,2}(\tilde x) \cup\bigl[\bigl(I^{00}_t(\tilde x)\setminus I\bigr) \cap \bigl(I^{00}_{t,1}(\tilde x)\cup I^{00}_{t,2}(\tilde x)\bigr)\bigr],\\ \mathcal I^-_H(t,\tilde x,I)&:=I^{-0}_{t,3}(\tilde x)\cup I^{+0}_{t,4}(\tilde x) \cup\bigl[\bigl(I^{00}_t(\tilde x)\setminus I\bigr) \cap \bigl(I^{00}_{t,3}(\tilde x)\cup I^{00}_{t,4}(\tilde x)\bigr)\bigr] \end{split} \] and observe that these sets characterize the indices $l\in\mathcal{Q}$ where the constraints $G_l(\tilde{x})\le t$, $G_l(\tilde{x})\ge -t$, $H_l(\tilde{x})\le t$, and $H_l(\tilde{x})\ge -t$, respectively, are active in \eqref{eq:relaxedMPSC_decomposed}. We therefore obtain \[ \mathcal L_{X(t,\tilde x,I)}(\tilde x) = \left\{ d\in\R^n\,\middle|\, \begin{aligned} \nabla g_i(\tilde x)\cdot d&\,\leq\,0&&i\in I^g(\tilde x)\\ \nabla h_j(\tilde x)\cdot d&\,=\,0&&j\in\mathcal P\\ \nabla G_l(\tilde x)\cdot d&\,\leq\,0&&l\in \mathcal I^+_G(t,\tilde x,I)\\ \nabla G_l(\tilde x)\cdot d&\,\geq\,0&&l\in \mathcal I^-_G(t,\tilde x,I)\\ \nabla H_l(\tilde x)\cdot d&\,\leq\,0&&l\in \mathcal I^+_H(t,\tilde x,I)\\ \nabla H_l(\tilde x)\cdot d&\,\geq\,0&&l\in \mathcal I^-_H(t,\tilde x,I) \end{aligned} \right\}. \] Exploiting Lemma \ref{lem:polar_of_polyhedral_cone}, the polar of this cone is easily computed: \begin{equation}\label{eq:polar_of_decomposed_linearization_cone} \mathcal L_{X(t,\tilde x,I)}(\tilde x)^\circ = \left\{ \eta\in\R^n\,\middle|\, \begin{aligned} &\eta=\sum\limits_{i\in I^g(\tilde x)}\lambda_i\nabla g_i(\tilde x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\tilde x)\\ &\qquad\qquad +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\tilde x)+\nu_l\nabla H_l(\tilde x) \Bigr]\\ &\forall i\in I^g(\tilde x)\colon\,\lambda_i\geq 0\\ &\forall l\in \mathcal Q\colon\,\mu_l \begin{cases} \geq 0 &\text{if }l\in \mathcal I^+_G(t,\tilde x,I),\\ \leq 0 &\text{if }l\in \mathcal I^-_G(t,\tilde x,I),\\ =0 &\text{otherwise}, \end{cases}\\ &\forall l\in \mathcal Q\colon\,\nu_l \begin{cases} \geq 0 &\text{if }l\in \mathcal I^+_H(t,\tilde x,I),\\ \leq 0 &\text{if }l\in \mathcal I^-_H(t,\tilde x,I),\\ =0 &\text{otherwise} \end{cases} \end{aligned} \right\}. \end{equation} Observing that \eqref{eq:relaxedMPSC} is a standard nonlinear problem, $\mathcal T_{X(t)}(\tilde x)\subset\mathcal L_{X(t)}(\tilde x)$ and, thus, $\mathcal L_{X(t)}(\tilde x)^\circ\subset \mathcal T_{X(t)}(\tilde x)^\circ$ are inherent. It remains to show $\mathcal T_{X(t)}(\tilde x)^\circ\subset \mathcal L_{X(t)}(\tilde x)^\circ$. Thus, choose $\eta\in\mathcal T_{X(t)}(\tilde x)^\circ$ arbitrarily. Then, in particular, \eqref{eq:tangents_via_intersection_of_linearized_tangents} yields \[ \eta\in \mathcal L_{X(t,\tilde x,\varnothing)}(\tilde x)^\circ \cap \mathcal L_{X(t,\tilde x,I^{00}_t(\tilde x))}(\tilde x)^\circ. \] Exploiting the representation \eqref{eq:polar_of_decomposed_linearization_cone}, we find $\lambda_i,\lambda'_i\geq 0$ ($i\in I^g(\tilde x)$), $\rho,\rho'\in\R^p$, $\mu,\mu'\in\R^q$, as well as $\nu,\nu'\in\R^q$ which satisfy \[ \begin{split} \eta&=\sum\limits_{i\in I^g(\tilde x)}\lambda_i\nabla g_i(\tilde x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\tilde x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\tilde x)+\nu_l\nabla H_l(\tilde x) \Bigr]\\ &=\sum\limits_{i\in I^g(\tilde x)}\lambda'_i\nabla g_i(\tilde x) +\sum\limits_{j\in\mathcal P}\rho'_j\nabla h_j(\tilde x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu'_l\nabla G_l(\tilde x)+\nu'_l\nabla H_l(\tilde x) \Bigr] \end{split} \] and \begin{align*} \mu_l& \begin{cases} \geq 0 &\text{if }l\in I^{0+}_{t,1}(\tilde x)\cup I^{0-}_{t,4}(\tilde x),\\ \leq 0 &\text{if }l\in I^{0+}_{t,2}(\tilde x)\cup I^{0-}_{t,3}(\tilde x),\\ =0 &\text{otherwise,} \end{cases}\\ \mu'_l& \begin{cases} \geq 0 &\text{if }l\in I^{0+}_{t,1}(\tilde x)\cup I^{0-}_{t,4}(\tilde x)\cup I^{00}_{t,1}(\tilde x)\cup I^{00}_{t,4}(\tilde x),\\ \leq 0 &\text{if }l\in I^{0+}_{t,2}(\tilde x)\cup I^{0-}_{t,3}(\tilde x)\cup I^{00}_{t,2}(\tilde x)\cup I^{00}_{t,3}(\tilde x),\\ =0 &\text{otherwise,} \end{cases}\\ \nu_l& \begin{cases} \geq 0 &\text{if }l\in I^{+0}_{t,1}(\tilde x)\cup I^{-0}_{t,2}(\tilde x)\cup I^{00}_{t,1}(\tilde x)\cup I^{00}_{t,2}(\tilde x),\\ \leq 0 &\text{if }l\in I^{-0}_{t,3}(\tilde x)\cup I^{+0}_{t,4}(\tilde x)\cup I^{00}_{t,3}(\tilde x)\cup I^{00}_{t,4}(\tilde x),\\ =0 &\text{otherwise,} \end{cases}\\ \nu'_l& \begin{cases} \geq 0 &\text{if }l\in I^{+0}_{t,1}(\tilde x)\cup I^{-0}_{t,2}(\tilde x),\\ \leq 0 &\text{if }l\in I^{-0}_{t,3}(\tilde x)\cup I^{+0}_{t,4}(\tilde x),\\ =0 &\text{otherwise} \end{cases} \end{align*} for all $l\in\mathcal Q$. Thus, we obtain \[ \begin{split} 0 &=\sum\limits_{i\in I^g(\tilde x)}(\lambda_i-\lambda_i')\nabla g_i(\tilde x) +\sum\limits_{j\in\mathcal P}(\rho_j-\rho_j')\nabla h_j(\tilde x)\\ &\qquad\qquad +\sum\limits_{l\in\mathcal Q} \Bigl[ (\mu_l-\mu_l')\nabla G_l(\tilde x)+(\nu_l-\nu_l')\nabla H_l(\tilde x) \Bigr]. \end{split} \] Observing $\operatorname{supp}(\mu-\mu')\subset I^{00}_t(\tilde x)\cup I^{0\pm}_t(\tilde x)$ as well as $\operatorname{supp}(\nu-\nu')\subset I^{00}_t(\tilde x)\cup I^{\pm 0}_t(\tilde x)$ and using \eqref{eq:ThmMPSCLICQyieldsGCQ1}, we obtain $\lambda_i=\lambda_i'$ ($i\in I^g(\tilde x)$), $\rho=\rho'$, $\mu=\mu'$, as well as $\nu=\nu'$. Particularly, \[ \eta=\sum\limits_{i\in I^g(\tilde x)}\lambda_i\nabla g_i(\tilde x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\tilde x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\tilde x)+\nu_l'\nabla H_l(\tilde x) \Bigr]. \] is obtained where $\operatorname{supp}\mu\subset I^{0\pm}_t(\tilde x)$ and $\operatorname{supp}\nu'\subset I^{\pm 0}_t(\tilde x)$ hold true. Finally, we exploit Corollary \ref{cor:linearization_cone_relaxedMPSC} in order to see $\eta\in\mathcal L_{X(t)}(\tilde x)^\circ$. This shows the validity of the inclusion $\mathcal T_{X(t)}(\tilde x)^\circ\subset \mathcal L_{X(t)}(\tilde x)^\circ$ and, thereby, GCQ holds true for \eqref{eq:relaxedMPSC} at $\tilde x$. Since $t\in(0,\bar t]$ and $\tilde x\in X(t)\cap U$ were arbitrarily chosen, the proof is completed. \end{proof} \section{Remarks on other possible relaxation schemes}\label{sec:other_relaxations} In this section, we discuss three more relaxation approaches for the numerical treatment of \eqref{eq:MPSC} which are inspired by the rich theory on MPCCs. Particularly, the relaxation schemes of \cite{Scholtes2001}, \cite{SteffensenUlbrich2010}, as well as \cite{KadraniDussaultBenchakroun2009}, are adapted to the setting of switching-constrained optimization. \subsection{The relaxation scheme of Scholtes}\label{sec:Scholtes} For some parameter $t\geq 0$, let us consider the surrogate problem \begin{equation}\label{eq:relaxedMPSC_Scholtes}\tag{P$_\text{S}(t)$} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&&j\in\mathcal P,&\\ -t\,\leq\,G_l(x)H_l(x)&\,\leq\,t,&&l\in\mathcal Q.& \end{aligned} \end{equation} This idea is inspired by Scholtes' global relaxation method which was designed for the computational treatment of MPCCs, see \cite{Scholtes2001} and \cite[Section~3.1]{HoheiselKanzowSchwartz2013}. The feasible set of \eqref{eq:relaxedMPSC_Scholtes} is denoted by $X_\text S(t)$ and visualized in Figure \ref{fig:relaxed_feasible_set_Scholtes}. Note that the family $\{X_\text S(t)\}_{t\geq 0}$ possesses the same properties as the family $\{X(t)\}_{t\geq 0}$ described in Lemma \ref{lem:properties_of_relaxed_feasible_set}. Thus, Scholtes' relaxation is reasonable for switching-constrained problems as well. In contrast to \eqref{eq:relaxedMPSC}, where we need four inequality constraints in order to replace one original switching constraint, one only needs two inequality constraints in \eqref{eq:relaxedMPSC_Scholtes} for the same purpose. This is a significant advantage of \eqref{eq:relaxedMPSC_Scholtes} over the surrogate \eqref{eq:relaxedMPSC}. \begin{figure}[h]\centering \includegraphics[width=4.0cm]{RelaxedFeasibleSet_Scholtes.pdf} \caption{Geometric interpretation of the relaxed feasible set $X_\text{S}(t)$.} \label{fig:relaxed_feasible_set_Scholtes} \end{figure} It is well known from \cite[Theorem~3.1]{HoheiselKanzowSchwartz2013} that Scholtes' relaxation approach finds Clarke-stationary points of MPCCs under an MPCC-tailored version of MFCQ. Note that, in the context of MPCCs, Clarke-stationarity is stronger than weak stationarity but weaker than Mordukhovich-stationarity. Below, we want to generalize the result from \cite{HoheiselKanzowSchwartz2013} to the problem \eqref{eq:MPSC}. For the fixed parameter $t>0$ and a feasible point $x\in X_\text S(t)$ of \eqref{eq:relaxedMPSC_Scholtes}, we introduce the index sets \[ \begin{split} I^+_t(x)&:=\{l\in\mathcal Q\,|\,G_l(x)H_l(x)=t\},\\ I^-_t(x)&:=\{l\in\mathcal Q\,|\,G_l(x)H_l(x)=-t\}. \end{split} \] In the upcoming theorem, we provide a convergence result of Scholtes' relaxation scheme for the problem \eqref{eq:MPSC}. \begin{theorem}\label{thm:convergence_Scholtes} Let $\{t_k\}_{k\in\N}\subset\R_+$ be a sequence of positive relaxation parameters converging to zero. For each $k\in\N$, let $x_k\in X_{\textup{S}}(t_k)$ be a KKT point of \hyperref[eq:relaxedMPSC_Scholtes]{\textup{(P$_{\text{S}}(t_k)$)}}. Assume that the sequence $\{x_k\}_{k\in\N}$ converges to a point $\bar x\in X$ where MPSC-MFCQ holds. Then $\bar x$ is a W-stationary point of \eqref{eq:MPSC}. \end{theorem} \begin{proof} Noting that $x_k$ is a KKT point of \hyperref[eq:relaxedMPSC_Scholtes]{(P$_{\text{S}}(t_k)$)}, we find multipliers $\lambda^k\in\R^m$, $\rho^k\in\R^p$, and $\xi^k\in\R^q$ which satisfy the following conditions: \begin{align*} &0=\nabla f(x_k)+\sum\limits_{i\in I^g(x_k)}\lambda^k_i\nabla g_i(x_k) +\sum\limits_{j\in\mathcal P}\rho^k_j\nabla h_j(x_k)\\ &\qquad\qquad +\sum\limits_{l\in\mathcal Q}\xi^k_l \Bigl[ H_l(x_k)\nabla G_l(x_k)+G_l(x_k)\nabla H_l(x_k) \Bigr],\\ &\forall i\in I^g(x_k)\colon\,\lambda^k_i\geq 0,\, \forall i\in\mathcal M\setminus I^g(x_k)\colon\,\lambda^k_i=0,\\ &\forall l\in I^+_t(x_k)\colon\,\xi^k_l\geq 0,\\ &\forall l\in I^-_t(x_k)\colon\,\xi^k_l\leq 0,\\ &\forall l\in \mathcal Q\setminus(I^+_t(x_k)\cup I^-_t(x_k))\colon\,\xi^k_l=0. \end{align*} For any $k\in\N$ and $l\in\mathcal Q$, let us define artificial multipliers $\mu^k_l,\nu^k_l\in\R$ as stated below: \[ \mu^k_l:= \begin{cases} \xi^k_lH_l(x_k) &l\in I^{G}(\bar x)\cup I^{GH}(\bar x),\\ 0 &l\in I^{H}(\bar x), \end{cases} \quad \nu^k_l:= \begin{cases} \xi^k_lG_l(x_k) &l\in I^{H}(\bar x)\cup I^{GH}(\bar x),\\ 0 &l\in I^{G}(\bar x). \end{cases} \] Thus, we obtain \begin{equation}\label{eq:Scholtes_stationarity_transformed} \begin{split} &0=\nabla f(x_k)+\sum\limits_{i\in I^g(x_k)}\lambda^k_i\nabla g_i(x_k) +\sum\limits_{j\in\mathcal P}\rho^k_j\nabla h_j(x_k) +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu^k_l\nabla G_l(x_k)+\nu^k_l\nabla H_l(x_k) \Bigr]\\ &\qquad\qquad +\sum\limits_{l\in I^H(\bar x)}\xi^k_lH_l(x_k)\nabla G_l(x_k) +\sum\limits_{l\in I^G(\bar x)}\xi^k_lG_l(x_k)\nabla H_l(x_k). \end{split} \end{equation} Next, we are going to show that the sequence $\{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)\}_{k\in\N}$ is bounded where we used $I:=I^G(\bar x)\cup I^H(\bar x)$ for brevity. We assume on the contrary that this is not the case and define \[ \forall k\in\N\colon \quad (\tilde\lambda^k,\tilde\rho^k,\tilde\mu^k,\tilde\nu^k,\tilde\xi^k_I) := \frac{ (\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I) } { \norm{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)}{2} }. \] Clearly, $\{(\tilde\lambda^k,\tilde\rho^k,\tilde\mu^k,\tilde\nu^k,\tilde\xi^k_I)\}_{k\in\N}$ is bounded and, thus, converges w.l.o.g.\ to a nonvanishing vector $(\tilde\lambda,\tilde\rho,\tilde\mu,\tilde\nu,\tilde\xi_I)$ (otherwise, a suitable subsequence is chosen). The continuity of $g$ ensures that $I^g(x_k)\subset I^g(\bar x)$ is valid for sufficiently large $k\in\N$. Dividing \eqref{eq:Scholtes_stationarity_transformed} by $\norm{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)}{2}$ and taking the limit $k\to\infty$ while respecting the continuous differentiability of all involved functions, we come up with \[ \begin{split} &0=\sum\limits_{i\in I^g(\bar x)}\tilde\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in\mathcal P}\tilde\rho_j\nabla h_j(\bar x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \tilde\mu_l\nabla G_l(\bar x)+\tilde\nu_l\nabla H_l(\bar x) \Bigr],\\ &\forall i\in I^g(\bar x)\colon\,\tilde\lambda_i\geq 0,\, \forall i\in\mathcal M\setminus I^g(\bar x)\colon\,\tilde\lambda_i=0,\\ &\forall l\in I^H(\bar x)\colon\,\tilde\mu_l=0,\\ &\forall l\in I^G(\bar x)\colon\,\tilde\nu_l=0. \end{split} \] Now, the validity of MPSC-MFCQ yields $\tilde\lambda=0$, $\tilde\rho=0$, $\tilde\mu=0$, and $\tilde\nu=0$. Hence, $\tilde\xi_{l_0}\neq 0$ holds for at least one index $l_0\in I$. Let us assume $l_0\in I^H(\bar x)$. Then we have $\nu^k_{l_0}=\xi^k_{l_0}G_{l_0}(x_k)$, which leads to \[ \tilde\nu_{l_0} =\lim\limits_{k\to\infty}\frac{\nu^k_{l_0}}{\norm{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)}{2}} =\lim\limits_{k\to\infty}\frac{\xi^k_{l_0}G_{l_0}(x_k)}{\norm{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)}{2}} =\tilde\xi_{l_0}G_{l_0}(\bar x)\neq 0. \] This, however, is a contradiction since $\tilde\nu$ vanishes due to the above arguments. Similarly, the case $l_0\in I^G(\bar x)$ leads to a contradiction. As a consequence, the sequence $\{(\lambda^k,\rho^k,\mu^k,\nu^k,\xi^k_I)\}_{k\in\N}$ is bounded. Thus, we may assume w.l.o.g.\ that this sequence converges to $(\lambda,\rho,\mu,\nu,\xi_I)$. Again, we take the limit in \eqref{eq:Scholtes_stationarity_transformed} and obtain \[ \begin{split} &0=\nabla f(\bar x)+\sum\limits_{i\in I^g(\bar x)}\lambda_i\nabla g_i(\bar x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(\bar x) +\sum\limits_{l\in\mathcal Q} \Bigl[ \mu_l\nabla G_l(\bar x)+\nu_l\nabla H_l(\bar x) \Bigr],\\ &\forall i\in I^g(\bar x)\colon\,\lambda_i\geq 0,\, \forall i\in\mathcal M\setminus I^g(\bar x)\colon\,\lambda_i=0,\\ &\forall l\in I^H(\bar x)\colon\,\mu_l=0,\\ &\forall l\in I^G(\bar x)\colon\,\nu_l=0 \end{split} \] which shows that $\bar x$ is a W-stationary point of \eqref{eq:MPSC}. \end{proof} Noting that no suitable definition of Clarke-stationarity (in particular, a \emph{reasonable} stationarity concept which is stronger than W- but weaker than M-stationarity) seems to be available for \eqref{eq:MPSC}, Theorem \ref{thm:convergence_Scholtes} does not seem to be too surprising at all. The following example shows that we cannot expect any stronger results in general. Thus, the qualitative properties of Scholtes' relaxation method are substantially weaker than those of the relaxation scheme proposed in Section \ref{sec:relaxation}. \begin{example}\label{ex:Scholtes_relaxation} Let us consider the switching-constrained optimization problem \[ \begin{split} \tfrac12(x_1-1)^2+\tfrac12(x_2-1)^2&\,\rightarrow\,\min\\ x_1x_2&\,=\,0. \end{split} \] The globally optimal solutions of this program are given by $(1,0)$ as well as $(0,1)$, and these points are S-stationary. Furthermore, there exists a W-stationary point at $\bar x=(0,0)$ which is no local minimizer. One can easily check that the associated problem \eqref{eq:relaxedMPSC_Scholtes} possesses a KKT point at $(\sqrt t,\sqrt t)$ for any $t\in(0,1]$. Taking the limit $t\downarrow 0$, this point tends to $\bar x$ which is, as we already mentioned above, only W-stationary for the switching-constrained problem of interest. Clearly, MPSC-LICQ is valid at $\bar x$. \end{example} Although the theoretical properties of Scholtes' relaxation approach do not seem to be promising in light of \eqref{eq:MPSC}, we check the applicability of the approach. More precisely, we analyze the restrictiveness of the assumption that a sequence of KKT points associated with \eqref{eq:relaxedMPSC_Scholtes} can be chosen. In order to guarantee that locally optimal solutions of the nonlinear relaxed surrogate problems \eqref{eq:relaxedMPSC_Scholtes} which are located closely to the limit point from Theorem \ref{thm:convergence_Scholtes} are KKT points, a constraint qualification needs to be satisfied. Adapting \cite[Theorem~3.2]{HoheiselKanzowSchwartz2013} to the switching-constrained situation, it is possible to show that whenever MPSC-MFCQ is valid at a feasible point $\bar x\in X$ of \eqref{eq:MPSC}, then standard MFCQ is valid for \eqref{eq:relaxedMPSC_Scholtes} in a neighborhood of $\bar x$. \begin{theorem}\label{thm:MPSC-MFCQ_yields_MFCQ_Scholtes} Let $\bar x\in X$ be a feasible point of \eqref{eq:MPSC} where MPSC-MFCQ is satisfied. Then there exists a neighborhood $U\subset\R^n$ of $\bar x$ such that MFCQ holds for \eqref{eq:relaxedMPSC_Scholtes} at all points from $X_{\textup S}(t)\cap U$ for all $t>0$. \end{theorem} \begin{proof} Due to the validity of MPSC-MFCQ at $\bar x$, the union \[ \begin{split} \{\nabla g_i(\bar x)\,|\,i\in I^g(\bar x)\}\cup \Bigl[\{\nabla h_j(\bar x)\,|\,j\in\mathcal P\} &\cup\{\nabla G_l(\bar x)\,|\,l\in I^{G}(\bar x)\cup I^{GH}(\bar x)\}\\ &\cup\{\nabla H_l(\bar x)\,|\,l\in I^{H}(\bar x)\cup I^{GH}(\bar x)\} \Bigr] \end{split} \] is positive-linearly independent. Invoking Lemma \ref{lem:local_stability_of_positive_linear_independence}, there is a neighborhood $U$ of $\bar x$ such that the vectors \[ \begin{split} \{\nabla g_i(x)\,|\,i\in I^g(\bar x)\}\cup \Bigl[\{\nabla h_j(x)\,|\,j\in\mathcal P\} &\cup\{\nabla G_l(x)\,|\,l\in I^{G}(\bar x)\cup I^{GH}(\bar x)\}\\ &\cup\{\nabla H_l(x)\,|\,l\in I^{H}(\bar x)\cup I^{GH}(\bar x)\} \Bigr] \end{split} \] are positive-linearly independent for any choice of $x\in U$. Now, fix $t>0$ as well as $x\in X_\text{S}(t)\cap U$ and set $I^\text{a}_t(x):=I^+_t(x)\cup I^-_t(x)$. Note that $t>0$ guarantees $I^+_t(x)\cap I^-_t(x)=\varnothing$. Clearly, we have \[ \begin{aligned} &\forall l\in I^H(\bar x)\colon&\quad &G_l(x)\neq 0&\quad&H_l(x)\approx 0,&\\ &\forall l\in I^G(\bar x)\colon&&G_l(x)\approx 0&&H_l(x)\neq 0 \end{aligned} \] if $U$ is sufficiently small. Exploiting Lemma \ref{lem:local_stability_of_positive_linear_independence} once more while recalling that $G$ and $H$ are continuously differentiable, we obtain that the vectors \begin{equation}\label{eq:positive_linearly_independent_vectors_MPSC_MFCQ} \begin{split} \{\nabla g_i(x)\,|\,i\in I^g(\bar x)\}\cup \Bigl[&\{\nabla h_j(x)\,|\,j\in\mathcal P\}\\ &\cup\{H_l(x)\nabla G_l(x)+G_l(x)\nabla H_l(x)\,|\,l\in I^{G}(\bar x)\cap I^\text a_t(x)\}\\ &\cup\{H_l(x)\nabla G_l(x)+G_l(x)\nabla H_l(x)\,|\,l\in I^{H}(\bar x)\cap I^\text a_t(x)\}\\ &\cup\{\nabla G_l(x)\,|\,l\in I^{GH}(\bar x)\cap I^\text a_t(x)\}\\ &\cup\{\nabla H_l(x)\,|\,l\in I^{GH}(\bar x)\cap I^\text a_t(x)\} \Bigr] \end{split} \end{equation} are positive-linearly independent if the neighborhood $U$ is chosen small enough. Suppose that there are vectors $\lambda\in\R^m$, $\rho\in\R^p$, and $\xi\in\R^q$ which satisfy \[ \begin{split} &0=\sum\limits_{i\in I^g(x)}\lambda_i\nabla g_i(x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(x) +\sum\limits_{l\in\mathcal Q}\xi_l \Bigl[ H_l(x)\nabla G_l(x)+G_l(x)\nabla H_l(x) \Bigr],\\ &\forall i\in I^g(x)\colon\,\lambda_i\geq 0,\, \forall i\in\mathcal M\setminus I^g(x)\colon\,\lambda_i=0,\\ &\forall l\in I^+_t(x)\colon\,\xi_l\geq 0,\\ &\forall l\in I^-_t(x)\colon\,\xi_l\leq 0,\\ &\forall l\in \mathcal Q\setminus I^\text a_t(x)\colon\,\xi_l=0. \end{split} \] In order to show the validity of MFCQ for \eqref{eq:relaxedMPSC_Scholtes} at $x$, $\lambda=0$, $\rho=0$, and $\xi=0$ have to be deduced. We get \[ \begin{split} &0=\sum\limits_{i\in I^g(x)}\lambda_i\nabla g_i(x) +\sum\limits_{j\in\mathcal P}\rho_j\nabla h_j(x)\\ &\qquad\qquad +\sum\limits_{l\in (I^G(\bar x)\cup I^H(\bar x))\cap I^\text a_t(x)}\xi_l \Bigl[ H_l(x)\nabla G_l(x)+G_l(x)\nabla H_l(x) \Bigr],\\ &\qquad\qquad +\sum\limits_{l\in I^{GH}(\bar x)\cap I^\text a_t(x)}\xi_lG_l(x)\nabla H_l(x) +\sum\limits_{l\in I^{GH}(\bar x)\cap I^\text a_t(x)}\xi_lH_l(x)\nabla G_l(x). \end{split} \] Noting that $\lambda_i\geq 0$ holds for all $i\in I^g(x)$ while $I^g(x)\subset I^g(\bar x)$ holds whenever $U$ is chosen sufficiently small, we obtain $\lambda=0$, $\rho=0$, $\xi_l=0$ ($l\in (I^G(\bar x)\cup I^H(\bar x))\cap I^\text a_t(x)$), $\xi_lG_l(x)=0$ ($l\in I^{GH}(\bar x)\cap I^\text a_t(x)$), and $\xi_l H_l(x)=0$ ($l\in I^{GH}(\bar x)\cap I^\text{a}_t(x)$) from the positive-linear independence of the vectors in \eqref{eq:positive_linearly_independent_vectors_MPSC_MFCQ}. Since we have $G_l(x)\neq 0$ and $H_l(x)\neq 0$ for all $l\in I^\text a_t(x)$ from $t>0$, $\xi_l=0$ follows for all $l\in I^\text{a}_t(x)$ since $I^G(\bar x)\cup I^H(\bar x)\cup I^{GH}(\bar x)=\mathcal Q$ is valid. This yields $\xi=0$. Thus, MFCQ holds for \eqref{eq:relaxedMPSC_Scholtes} at $x$. \end{proof} \subsection{The relaxation scheme of Steffensen and Ulbrich}\label{sec:SteffensenUlbrich} Here, we adapt the relaxation scheme from \cite{SteffensenUlbrich2010} for the numerical treatment of \eqref{eq:MPSC}. For any $t>0$, let us introduce $\phi(\cdot;t)\colon\R\to\R$ by means of \[ \forall z\in\R\colon \quad \phi(z;t):= \begin{cases} |z| & \text{if }|z|\geq t,\\ t\,\theta(z/t) & \text{if }|z|<t, \end{cases} \] where $\theta\colon[-1,1]\to\R$ is a twice continuously differentiable function with the following properties: \begin{align*} &\text{(a)}\quad\theta(1)=\theta(-1)=1,& &\text{(b)}\quad\theta'(-1)=-1\text{ and }\theta'(1)=1,&\\ &\text{(c)}\quad\theta''(-1)=\theta''(1)=0,& &\text{(d)}\quad\theta''(z)>0\text{ for all }z\in(-1,1).& \end{align*} A typical example for a function $\theta$ with the above properties is given by \begin{equation}\label{eq:example_for_theta} \forall z\in[-1,1]\colon \quad \theta(z):=\tfrac{2}{\pi}\sin\left(\tfrac{\pi}{2}z+\tfrac{3\pi}{2}\right)+1, \end{equation} see \cite[Section~3]{SteffensenUlbrich2010}. Noting that the function $\phi(\cdot;t)$ is smooth for any choice of $t>0$, it can be used to regularize the feasible set of \eqref{eq:MPSC}. A suitable surrogate problem is given by \begin{equation}\label{eq:relaxedMPSC_SteffensenUlbrich}\tag{P$_\text{SU}(t)$} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&&j\in\mathcal P,&\\ G_l(x)+H_l(x)-\phi(G_l(x)-H_l(x);t)&\,\leq\,0,&&l\in\mathcal Q,&\\ G_l(x)-H_l(x)-\phi(G_l(x)+H_l(x);t)&\,\leq\,0,&&l\in\mathcal Q,&\\ -G_l(x)+H_l(x)-\phi(-G_l(x)-H_l(x);t)&\,\leq\,0,&&l\in\mathcal Q,&\\ -G_l(x)-H_l(x)-\phi(-G_l(x)+H_l(x);t)&\,\leq\,0,&&l\in\mathcal Q.& \end{aligned} \end{equation} Its feasible set will be denoted by $X_\text{SU}(t)$ and is visualized in Figure \ref{fig:relaxed_feasible_set_Steffensen_Ulbrich}. Adapting the proof of \cite[Lemma~3.3]{SteffensenUlbrich2010}, the family $\{X_\text{SU}(t)\}_{t>0}$ possesses the properties (P2) and (P3) from Lemma \ref{lem:properties_of_relaxed_feasible_set}. This justifies that \eqref{eq:relaxedMPSC_SteffensenUlbrich} is a relaxation of \eqref{eq:MPSC}. Note that we need to introduce four inequality constraints to replace one of the original switching constraints. \begin{figure}[h]\centering \includegraphics[width=4.0cm]{RelaxedFeasibleSet_SteffensenUlbrich.pdf} \caption{Geometric interpretation of the relaxed feasible set $X_\text{SU}(t)$.} \label{fig:relaxed_feasible_set_Steffensen_Ulbrich} \end{figure} It has been mentioned in \cite{HoheiselKanzowSchwartz2013} that the relaxation scheme of Steffensen and Ulbrich computes Clarke-stationary points of MPCCs under an MPCC-tailored version of CPLD, see \cite[Section~3.4]{HoheiselKanzowSchwartz2013} as well. Recalling some arguments from Section \ref{sec:Scholtes}, the adapted method may only find W-stationary points of \eqref{eq:MPSC} in general. The upcoming example confirms this conjecture. \begin{example}\label{ex:SteffensenUlbrich} Let us consider the switching-constrained optimization problem \begin{equation}\label{eq:ExSteffensenUlbrich} \begin{split} x_1x_2-x_1-x_2&\,\rightarrow\,\min\\ x_1^2+x_2^2-1&\,\leq\,0,\\ x_1x_2&\,=\,0. \end{split} \end{equation} Obviously, the globally optimal solutions of this problem are given by $(1,0)$ as well as $(0,1)$, and these points are S-stationary. Furthermore, there is a W-stationary point at $\bar x=(0,0)$ which is no local minimizer. The global maximizers $(-1,0)$ and $(0,-1)$ do not satisfy any of the introduced stationarity concepts. Let us consider the associated family of nonlinear problems \eqref{eq:relaxedMPSC_SteffensenUlbrich} for $t\in(0,1]$ where the function $\theta$ is chosen as in \eqref{eq:example_for_theta}. It can easily be checked that $x(t):=(\tfrac t2(1-\tfrac{2}{\pi}),\tfrac t2(1-\tfrac{2}{\pi}))$ is a KKT point of \eqref{eq:relaxedMPSC_SteffensenUlbrich}. Note that $x(t)\to\bar x$ as $t\downarrow 0$, and that MPSC-LICQ holds in $\bar{x}$. However, $\bar{x}$ is only a W-stationary point of the switching-constrained problem \eqref{eq:ExSteffensenUlbrich}. \end{example} \subsection{The relaxation scheme of Kadrani, Dussault, and Benchakroun}\label{sec:Kadrani} Finally, we want to take a closer look at the relaxation approach which was suggested by \cite{KadraniDussaultBenchakroun2009} for the treatment of MPCCs. For any $t\geq 0$, let us consider the optimization problem \begin{equation}\label{eq:relaxedMPSC_Kadrani}\tag{P$_\text{KDB}(t)$} \begin{aligned} f(x)&\,\rightarrow\,\min&&&\\ g_i(x)&\,\leq\,0,&\qquad&i\in\mathcal M,&\\ h_j(x)&\,=\,0,&&j\in\mathcal P,&\\ (G_l(x)-t)(H_l(x)-t)&\,\leq\,0,&&l\in\mathcal Q,&\\ (-G_l(x)-t)(H_l(x)-t)&\,\leq\,0,&&l\in\mathcal Q,&\\ (G_l(x)+t)(H_l(x)+t)&\,\leq\,0,&&l\in\mathcal Q,&\\ (G_l(x)-t)(-H_l(x)-t)&\,\leq\,0,&&l\in\mathcal Q,& \end{aligned} \end{equation} whose feasible set will be denoted by $X_\text{KDB}(t)$. The family $\{X_\text{KDB}(t)\}_{t\geq 0}$ only satisfies property (P1) from Lemma \ref{lem:properties_of_relaxed_feasible_set} while (P2) and (P3) are violated in general. Thus, the surrogate problem \eqref{eq:relaxedMPSC_Kadrani} does not induce a relaxation technique for \eqref{eq:MPSC} in the narrower sense. Figure \ref{fig:relaxed_feasible_set_Kadrani} depicts that $X_\text{KBD}(t)$ is \emph{almost} disconnected, i.e., it is close to crumbling into four disjoint sets for any $t>0$. This may cause serious problems when standard techniques are used to solve the associated surrogate problem \eqref{eq:relaxedMPSC_Kadrani}. Moreover, four inequality constraints are necessary to replace one switching constraint from \eqref{eq:MPSC} in \eqref{eq:relaxedMPSC_Kadrani}. \begin{figure}[H]\centering \includegraphics[width=4.0cm]{RelaxedFeasibleSet_Kadrani.pdf} \caption{Geometric interpretation of the relaxed feasible set $X_\text{KDB}(t)$.} \label{fig:relaxed_feasible_set_Kadrani} \end{figure} On the other hand, it is clear from \cite[Section~3.3]{HoheiselKanzowSchwartz2013} that the regularization approach of Kadrani, Dussault, and Benchakroun computes M-stationary points of MPCCs under an MPCC-tailored version of CPLD at the limit point. Furthermore, if an MPCC-tailored LICQ holds at the limit point, then standard GCQ holds for the surrogate problems in a neighborhood of the point for sufficiently small relaxation parameters. These results are closely related to those for the relaxation approach from \cite{KanzowSchwartz2013} which we generalized to \eqref{eq:MPSC} in Sections \ref{sec:relaxation} and \ref{sec:convergence_properties}. Although we abstain from a detailed analysis of the regularization method which is induced by the surrogate problem \eqref{eq:relaxedMPSC_Kadrani} due to the aforementioned shortcomings, the above arguments motivate the formulation of the following two conjectures. \begin{conjecture}\label{con:convergence_Kadrani} Let $\{t_k\}_{k\in\N}\subset\R_+$ be a sequence of positive regularization parameters converging to zero. For each $k\in\N$, let $x_k\in X_{\textup{KDB}}(t_k)$ be a KKT point of \hyperref[eq:relaxedMPSC_Kadrani]{\textup{(P$_{\text{KDB}}(t_k)$)}}. Assume that the sequence $\{x_k\}_{k\in\N}$ converges to a point $\bar x\in X$ where MPSC-MFCQ holds. Then $\bar x$ is an M-stationary point of \eqref{eq:MPSC}. \end{conjecture} \begin{conjecture}\label{con:MPSC-LICQ_yields_GCQ_Kadrani} Let $\bar x\in X$ be a feasible point of \eqref{eq:MPSC} where MPSC-LICQ is satisfied. Then there exist $\bar t>0$ and a neighborhood $U\subset\R^n$ of $\bar x$ such that GCQ holds for \eqref{eq:relaxedMPSC_Kadrani} at all points from $X_\textup{KDB}(t)\cap U$ for all $t\in (0,\bar{t}]$. \end{conjecture} \section{Numerical results}\label{sec:numerical_results} This section is dedicated to a detailed analysis and comparison of various numerical methods for \eqref{eq:MPSC}. To obtain a meaningful comparison, we apply the relaxation scheme from Section \ref{sec:relaxation_scheme_KS} as well as a collection of other algorithms (see below) to multiple classes of MPSCs which possess significant practical relevance. The particular examples we analyze are: \begin{itemize} \item an either-or constrained problem with known local and global solutions \item a switching-constrained optimal control problem involving the non-stationary heat equation in two dimensions \item optimization problems involving semi-continuous variables, in particular special instances of portfolio optimization \end{itemize} For each of the examples, we first provide an overview of the corresponding problem structure, and then give some numerical results. To facilitate a quantitative comparison of the used algorithms, we use performance profiles (see \cite{DolanMore2002}) based on the computed function values. \subsection{Implementation} The numerical experiments in this section were all done in MATLAB R2018a. The particular algorithms we use for our computations are the following: \begin{itemize}[leftmargin=6em] \item[\textbf{KS:}] the adapted Kanzow--Schwartz relaxation scheme from this paper \item[\textbf{FMC:}] the \texttt{fmincon} function from the MATLAB Optimization Toolbox \item[\textbf{SNOPT:}] the SNOPT nonlinear programming solver from \cite{GillMurraySaunders2002}, called through the TOMLAB programming environment \item[\textbf{IPOPT:}] the IPOPT interior-point algorithm from \cite{WaechterBiegler2006} \end{itemize} The overall implementation is done in MATLAB, and each algorithm is called with user-supplied gradients of the objective functions and constraints. The stopping tolerance for all algorithms is set to $10^{-4}$ (although it should be noted that the methods use different stopping criteria, i.e., they impose the accuracy in different ways). For the KS algorithm, the relaxation parameters are chosen as $t_k:=0.01^k$, and the method is also terminated as soon as $t_k$ drops below $10^{-8}$. Finally, to solve the relaxed subproblems in the relaxation method, we employ the SNOPT algorithm with an accuracy of $10^{-6}$. Judging by past experience with MPCCs, the SNOPT algorithm can be expected to rival the relaxation scheme in terms of robustness. To accurately measure the performance of the solvers, it is important to note that MPSCs can, in general, admit a substantial amount of local minimizers. Therefore, the robustness is best measured by comparing the obtained function values (using different methods and starting points) with the globally optimal function value---if the latter is known; otherwise, a suitable approximate is used. To avoid placing too much emphasis on the accuracy of the final output (which does not make sense since the algorithms use completely different stopping criteria), we use the quantity \begin{equation}\label{Eq:PerfProfMetric} Q_\delta(x):= \begin{cases} f(x)-f_{\min}+\delta, & \text{if $x$ is feasible within tolerance}, \\ +\infty, & \text{otherwise} \end{cases} \end{equation} as the base metric for the performance profiles, where $x$ is the final iterate of the given algorithmic run, $f_{\min}$ the (approximate) global minimal value of the underlying problem, and $\delta\ge 0$ is an additional parameter which reduces the sensitivity of the values to numerical accuracy. We have found that an appropriate choice of $\delta$ can significantly improve the meaningfulness of the results. \subsection{Numerical examples} The following pages contain three examples of MPSCs. In Section \ref{sec:exp2}, we deal with an \emph{either-or constrained} problem, which can be reformulated as an MPSC, see \cite{Mehlitz2018}. Section~\ref{sec:exp3} is dedicated to a switching-constrained optimal control problem based on the framework from \cite{ClasonRundKunisch2017}. Finally, in Section~\ref{Sec:Portfolio}, we deal with a class of optimization problems with \emph{semi-continuous variables}, which can again be reformulated as MPSCs. This section contains a particular example from portfolio optimization which originates from \cite{FrangioniGentile2007}. \subsubsection{An either-or constrained example}\label{sec:exp2} Let us consider the optimization problem \begin{equation}\label{eq:exp2}\tag{E$2$} \begin{split} (x_1-8)^2+(x_2+3)^2&\,\rightarrow\,\min\\ x_1-2x_2+4\,\leq\,0\;\lor\;x_1-2&\,\leq\,0,\\ x_1^2-4x_2\,\leq\,0\;\lor\;(x_1-3)^2+(x_2-1)^2-10&\,\leq\,0. \end{split} \end{equation} Here, $\lor$ denotes the logical ``or''. The feasible set of this program is visualized in Figure~\ref{fig:exp2}. {It is easily seen that \eqref{eq:exp2} possesses the unique global minimizer $\bar x=(2,-2)$ and another local minimizer $\tilde x=(4,4)$. \begin{figure}[h]\centering \includegraphics[width=0.45\textwidth]{FeasibleSetE2.pdf} \hspace{1em} \includegraphics[width=0.45\textwidth]{PerfProfEx2.pdf} \caption{The either-or constrained problem from Section~\ref{sec:exp2}: feasible set, some level sets, and local minimizers (left), performance profile (right).} \label{fig:exp2} \end{figure} Arguing as in \cite[Section~7]{Mehlitz2018}, we can transform \eqref{eq:exp2} into a switching-constrained optimization problem by introducing additional variables: \begin{equation}\label{eq:exp2_switching} \begin{split} (x_1-8)^2+(x_2+3)^2&\,\rightarrow\,\min\limits_{x,z}\\ z_1,z_2,z_3,z_4&\,\leq\,0,\\ (x_1-2x_2+4-z_1)(x_1-2-z_2)&\,=\,0,\\ (x_1^2-4x_2-z_3)((x_1-3)^2+(x_2-1)^2-10-z_4)&\,=\,0. \end{split} \end{equation} Note that the local minimizers of \eqref{eq:exp2} can be found among the local minimizers of \eqref{eq:exp2_switching} choosing suitable values for the variable $z$, see \cite[Section~7.1]{Mehlitz2018}. The algorithms in question are each tested with the starting points in the set $\{0,1\}^6$, which makes for a total of $64$ starting points. The resulting performance profile can also be found in Figure~\ref{fig:exp2}; it is based on the metric \eqref{Eq:PerfProfMetric} with $\delta:=1$. Clearly, the KS relaxation method is the most robust of the four algorithms, finding the best function values (among the tested algorithms) in more than 80\% of the test runs. \subsubsection{Switching-constrained optimal control}\label{sec:exp3} Here, we intend to solve a switching-constrained optimal control problem with the proposed relaxation method. The underlying example is taken from \cite[Section~5.2]{ClasonRundKunisch2017}. Let $I:=(0,T)$, with $T:=10$ the final time, $\Omega:=(-1,1)^2$, and let $\Gamma$ be the boundary of $\Omega$. Furthermore, we define $\Omega_u:=(-1,0]\times(-1,1)$ as well as $\Omega_v:=(0,1)\times(-1,1)$. Let us consider the optimal control of the non-stationary heat equation with zero initial and Neumann boundary conditions given below: \begin{equation}\label{eq:heat_equation} \begin{aligned} \partial_t y(t,\omega)-\Delta_\omega y(t,\omega)-\tfrac{1}{10}\chi_{\Omega_u}(\omega)u(t)-\tfrac{1}{10}\chi_{\Omega_v}(\omega)v(t)&\,=\,0&\quad&\text{a.e.\ on }I\times\Omega,\\ \vec{\mathbf n}(\omega)\cdot\nabla_\omega y(t,\omega)&\,=\,0&&\text{a.e.\ on }I\times\Gamma,\\ y(0,\omega)&\,=\,0&&\text{a.e.\ on }\Omega. \end{aligned} \end{equation} Here, $\chi_A\colon\Omega\to\R$ denotes the characteristic function of a measurable set $A\subset\Omega$ which equals $1$ on $A$ and vanishes otherwise. Let $y_\text{d}\in L^2(I;H^1(\Omega))$ be the solution of the state equation associated with the desirable control functions $u_\text{d},v_\text{d}\in L^2(I)$ given by \begin{equation*} \forall t\in I\colon\quad u_\text{d}(t):=20\sin^4 (2\pi t/T ),\quad v_\text{d}(t):=10\cos^4(1.4\pi t/T). \end{equation*} All feasible controls $u,v\in H^1(I)$ shall satisfy the switching requirement \begin{equation}\label{eq:switching_controls} u(t)v(t)=0\quad\text{a.e.\ on }I. \end{equation} Note that $u_\text{d}$ and $v_\text{d}$ violate this switching condition. We aim to find the minimum of the objective function defined by \begin{equation}\label{eq:exp3} \begin{aligned} \mathcal{J}(y,u,v):=\tfrac{1}{2}\norm{y-y_\text d}{L^2(I;L^2(\Omega))}^2 & +\tfrac{\alpha}{2}\left(\norm{u}{L^2(I)}^2+\norm{v}{L^2(I)}^2\right) \\ & +\tfrac{\beta}{2}\left(\norm{\partial_t u}{L^2(I)}^2+\norm{\partial_t v}{L^2(I)}^2\right) \end{aligned} \end{equation} with respect to $(y,u,v)\in L^2(I;H^1(\Omega))\times H^1(I)\times H^1(I)$ such that $(u,v)$ satisfy the switching requirement \eqref{eq:switching_controls} while $y$ solves the associated state equation \eqref{eq:heat_equation}. We chose $\alpha:=10^{-6}$ and $\beta:=10^{-5}$ for our experiments. For the numerical solution of the problem, the domain $\Omega$ is tessellated using the function \texttt{generateMesh} from the MATLAB PDE toolbox and a discretization tolerance of $h:=10^{-1}$. The time interval $I$ is subdivided into equidistant intervals of size $\tau:=10^{-1}$. Both the spatial and temporal discretizations use standard piecewise linear (continuous) finite elements, which leads to a conforming approximation of the $H^1$-norm in \eqref{eq:exp3}. After discretization, the problem turns into a finite-dimensional MPSC comprising the variables $\vec u,\vec v\in\R^{101}$, a quadratic objective function, and the switching constraints $\vec u_i\vec v_i=0$ for all $i=1,\ldots,101$. These correspond to the simple constraint mappings $G(\vec u,\vec v):=\vec u$ and $H(\vec u,\vec v):=\vec v$. Note that the feasible set can be seen as the union of $2^{101}$ convex ``branches'' (obtained by setting either $\vec u_i=0$ or $\vec v_i=0$ for each $i=1,\ldots,101$). Hence, the problem can be expected to admit a substantial amount of local minimizers, and it is unrealistic to expect algorithmic implementations to find the global minimizer, even when tested with a large number of initial points. To facilitate a quantitative comparison of our numerical algorithms (as in the previous section), we use the following heuristic to obtain an upper estimate of the optimal value: using a coarser time discretization (with $\tau=0.4$), we compute the \emph{exact} global minimizer of the resulting problem by minimizing the objective over each of the branches induced by the switching constraints. The corresponding global minimizer is then lifted to the finer time grid (with $\tau=0.1$) by linear interpolation, and the resulting point is used as an initial guess for all the used algorithms. The resulting estimate of the optimal value is $0.2997$, and the associated controls are depicted in Figure~\ref{fig:possible_global_minimizer}. For the numerical tests, we generated $1000$ starting points with coordinates chosen randomly in the interval $[0,10]$. The performance profile was constructed by using the metric \eqref{Eq:PerfProfMetric} with $\delta:=0$, and it too can be found in Figure~\ref{fig:possible_global_minimizer}. As in the previous example, the relaxation method turns out to be the most robust of the tested algorithms, finding lower function values in around 70\% of the test runs. When analyzing the results in more detail, it turns out that, as expected, the algorithms found an exorbitant amount of distinct points (possibly local minimizers). Interestingly, however, the associated function values actually lie quite close to each other. This explains the $x$-axis scaling in the performance profile plot. \begin{figure}[h]\centering \includegraphics[width=0.45\textwidth]{PerfProfEx3.pdf} \hspace{1em} \includegraphics[width=0.45\textwidth]{GlobalMinimizerEx3.pdf} \caption{Performance profile and (possible) global minimizer for the switching-constrained optimal control problem from Section~\ref{sec:exp3}.} \label{fig:possible_global_minimizer} \end{figure} \subsubsection{Semi-continuous variables}\label{Sec:Portfolio} In many optimization scenarios, it is desirable that a nonnegative decision variable is either exactly zero or contained in some interval whose lower bound is strictly positive. This is the case, for instance, in production planning, portfolio optimization, compressed sensing in signal processing, and subset selection in regression. More details can be found in \cite{BurdakovKanzowSchwartz2016,Sun2013}, and the references therein. Given a decision variable $x\in\R^n$ and bounds $\ell,u\in\R^n$, $\ell\le u$, a requirement of the above form can be reformulated as the either-or type constraints \begin{equation}\label{Eq:SemicontinuousConstraint} x_i=0 \; \lor \; x_i\in [\ell_i,u_i],\qquad i=1,\ldots,n. \end{equation} In this context, the variable $x$ is often called \emph{semi-continuous} since it is required to lie in some continuous interval, except for the outlier case when it is equal to zero. (One might also be inclined to call $x$ a \emph{semi-discrete} variable, but we have not seen this terminology elsewhere in the literature.) Constraint systems of the form \eqref{Eq:SemicontinuousConstraint} can be reformulated as switching constraints by using slack variables. Indeed, there are two ways of doing so: On the one hand, we could introduce two nonnegative slack variables to transform the box constraints in \eqref{Eq:SemicontinuousConstraint} into equality constraints; this procedure eventually yields an MPSC with $2 n$ switching constraints. On the other hand, assuming that $u_i\ge 0$ holds for all $i=1,\ldots,n$ (which is the case in nearly all relevant applications), we can simply treat the requirement $x_i\le u_i$ as a standard inequality constraint which should be fulfilled at all times. Clearly, if $x_i=0$ is valid, then the inequality $x_i\le u_i$ holds automatically, so that we can rewrite \eqref{Eq:SemicontinuousConstraint} as \begin{equation*} \begin{aligned} x_i&\, \le\,u_i, \quad &&i=1,\ldots,n,&\\ x_i=0\,\lor\,x_i&\,\ge \ell_i, \quad &&i=1,\ldots,n.& \end{aligned} \end{equation*} Using a single slack vector $y\in\R^n$, we can now rewrite this system as \begin{equation*} \begin{aligned} x&\,\le\,u,&&&\\ y&\,\ge\,0,&&&\\ x_i(x_i-\ell_i-y_i)&\,=\,0,\quad&&i=1,\ldots,n.& \end{aligned} \end{equation*} In the notation of our general framework \eqref{eq:MPSC}, this corresponds to the switching mappings \begin{equation*} G(x,y):=x \quad\text{and}\quad H(x,y):=x-\ell-y. \end{equation*} The inequality constraints $x\le u$ and $y\ge 0$ can be implemented as components of the mapping $g$ (which may contain other constraints depending on the particular problem). Note that the above reformulation approach only results in $n$ switching constraints. \begin{figure}[h]\centering \includegraphics[scale=0.5]{PerfProfEx4.pdf} \caption{Performance profile for the portfolio optimization problem from Section~\ref{Sec:Portfolio}.} \label{fig:portfolio} \end{figure} We now present a concrete example of portfolio optimization based on the test examples in \cite{FrangioniGentile2007}. The problems in this reference have the form \begin{equation}\label{Eq:PortfolioExample} \begin{aligned} x^{\top}Q x & \,\to\,\min \\ e^{\top}x &\, =\,1, \\ \mu^{\top}x &\, \ge \,\rho, \\ \quad x_i\,=\,0\;\lor\,x_i&\,\in\,[\ell_i,u_i],\quad i=1,\ldots,n, \end{aligned} \end{equation} with randomly generated $Q\in\R^{n\times n}$, $\mu,\ell,u\in\R^n$, and $\rho\in\R$. Here, $e\in\R^n$ represents the all-ones vector. More details can be found in \cite{FrangioniGentile2007} and on their webpage \url{http://www.di.unipi.it/optimize/Data/MV.html}. The particular examples we chose are the $30$ instances with size $200$. The corresponding problems \eqref{Eq:PortfolioExample} are reformulated as MPSCs by means of the aforementioned procedure, and the resulting problems are then attacked by the four test algorithms in question. Figure~\ref{fig:portfolio} depicts the resulting performance profile based on the metric \eqref{Eq:PerfProfMetric} with $\delta:=0$. For this particular problem class, it turns out that the performance advantage of the relaxation method is particularly large when compared to its non-relaxed counterparts. In 28 out of 30 examples, the KS algorithm finds the best function value among the tested methods; in addition, it also seems to find feasible points much more reliably than the other algorithms. \section{Final remarks}\label{sec:final_results} In this paper, we have presented a relaxation method for the solution of mathematical programs with switching constraints (MPSCs). Our theoretical analysis yields strong convergence properties for the method; in particular, the algorithm computes M-stationary points of MPSCs under a problem-tailored constraint qualification (MPSC-NNAMCQ) which is weaker than MPSC-MFCQ. The numerical experiments include a wide array of practically relevant problems and demonstrate the computational efficiency of the proposed algorithm. In addition, we have conducted a dedicated analysis for other relaxation schemes which are known in the MPCC literature and can be carried over to the setting of switching constraints. In particular, adapted versions of the relaxation schemes of Scholtes as well as Steffensen and Ulbrich are shown to converge to weakly stationary points only, even if fairly strong regularity properties such as MPSC-LICQ are satisfied. \bibliographystyle{plainnat}

98c2e950ed9b5269974fa210d00997516a573045

\section{Introduction} Cluster tilting theory gives a way to construct abelian categories from some triangulated categories. Let $\mathcal T$ be a triangulated category and $\mathcal C$ a cluster tilting subcategory of $\mathcal T$. Then the quotient category $\mathcal T/\mathcal C$ is abelian. This is due to Koenig and Zhu \cite[Theorem 3.3]{KZ}. Cluster tilting theory is also permitted to construct abelian categories from some exact categories. Demonet and Liu \cite[Theorem 3.2]{DL} provided a general framework for passing from exact categories to abelian categories by factoring out cluster tilting subcategories. We recall the definition of cluster tilting subcategories, which was introduced by Iyama \cite[Definition 2.2]{I}. Let $\mathcal T$ be a triangulated category or exact category and $\mathcal C$ a subcategory of $\mathcal T $. $\mathcal C$ is called cluster tilting if it satisfies: \begin{itemize} \item $\mathcal C$ is contravariantly finite and covariantly finite; \item $\mathcal C=\mathcal C^{\bot_1}={^{\bot_1}}\mathcal C$, where $\mathcal C^{\bot_1}=\{M\in\mathcal B~|~\operatorname{Ext}\nolimits^1(\mathcal C,M)=0\}$ and ${^{\bot_1}}\mathcal C=\{M\in\mathcal B~|~\operatorname{Ext}\nolimits^1(M,\mathcal C)=0\}$. \end{itemize} Now we consider the opposite direction: if we have an ideal quotient $\mathcal T/\mathcal D$ which is abelian, can we get any information of $\mathcal D$? When does $\mathcal D$ become a cluster tilting subcategory? Beligiannis proved the following characterization of cluster tilting subcategories which complements, and was inspired by Koenig and Zhu. \begin{thm}\cite[Theorem 7.3]{B} Let $\mathcal T$ be connected triangulated category with a Serre functor $\mathbb{S}$ and $\mathcal C$ be a non-zero functorially finite rigid subcategory of $\mathcal B$. Then the following statements are equivalent: \begin{itemize} \item[(a)] $\mathcal C$ is cluster tilting; \item[(b)] $\mathcal C$ is a maximal extension closed subcategory of $\mathcal T$ such that $\mathbb{S}\mathcal C=\mathcal C[2]$; \item[(c)] $\mathcal T/\mathcal C$ is abelian and $\mathbb{S}\mathcal C=\mathcal C[2]$. \end{itemize} \end{thm} It is very natural to ask if the similar theory holds on exact category, which also plays an important role in representation theory. For example the extension closed subcategories of module categories of $k$-algebras (where $k$ is a field) are exact categories. Since we usually do not have Serre functors on exact categories, we are also interested in the case for triangulated categories which do not have Serre functors. Hence in this article, we will study a similar case as \cite{B} on a more generalized setting: a category called extriangulated category. The notion of an extriangulated category was introduced in \cite{NP} (please see section 2 for detailed definition of extriangulated category), which is a simultaneous generalization of exact category and triangulated category. For examples of extriangulated categories which are neither exact categories nor triangulated categories, please see \cite{NP,ZZ}. We can also define cluster tilting subcategory on extriangulated categories. Liu and Nakaoka \cite[Theorem 3.2]{LN} showed that any quotient of a extriangulated category modulo a cluster tilting subcategory carried an induced abelian structures, which generalizes both \cite[Theorem 3.3]{KZ} and \cite[Theorem 3.2]{DL}. In this article, let $(\mathcal B,\mathbb E,\mathfrak{s})$ be a Krull-Schmidt extriangulated category over a field $k$. Any subcategory we discuss in this article will be full and closed under isomorphisms. \begin{rem} If $\mathcal B$ is a $k$-linear, Hom-finite extriangulated category with split idempotents, then it is a Krull-Schmidt category. \end{rem} Now we introduce the notion of pre-cluster tilting subcategory. \begin{defn} A subcategory $\mathcal B'$ of $\mathcal B$ is called \emph{contravariantly finite} if any object in $\mathcal B$ admits a right $\mathcal B'$-approximation. Moreover, it is called \emph{strongly contravariantly finite} if any object in $\mathcal B$ admits a right $\mathcal B'$-approximation which is also a deflation. Dually we can define \emph{strongly covariantly finite}. \end{defn} \begin{defn} We call $\mathcal C$ a pre-cluster tilting subcategory of $\mathcal B$ if it satisfies the following conditions: \begin{itemize} \item $\mathcal C$ is closed under direct sums and summands; \item $\mathcal C$ is rigid, that is to say, $\mathbb E(\mathcal C,\mathcal C)=0$; \item $\mathcal C$ is strongly contravariantly finite and strongly covariantly finite. \item $\mathcal C^{\bot_1}={^{\bot_1}}\mathcal C$. where $\mathcal C^{\bot_1}$ is the subcategory of objects $X\in\mathcal B$ satisfying $\mathbb E(\mathcal C,X)=0$ and ${^{\bot_1}}\mathcal C$ is the subcategory of objects $X\in\mathcal B$ satisfying $\mathbb E(X,\mathcal C)=0$. \end{itemize} A subcategory $\mathcal C$ of $\mathcal B$ is called cluster tilting if $\mathcal C$ is a pre-cluster tilting and $\mathcal C=\mathcal C^{\bot_1}={^{\bot_1}}\mathcal C$. \end{defn} We give an example of pre-cluster tilting subcategory. \begin{exm} Let $\Lambda$ be the the $k$-algebra given by the quiver $$\xymatrix@[email protected]{ &&\centerdot \ar[dll]_x \\ \centerdot \ar[dr]_x &&&&\centerdot \ar[ull]_x\\ &\centerdot \ar[rr]_x &&\centerdot \ar[ur]_x } $$ with relation $x^3=0$. Then the AR-quiver of $\mathcal B:=\mod\Lambda$ is given by $$\xymatrix@[email protected]{ \centerdot \ar[dr] &&\centerdot \ar[dr] &&\centerdot \ar[dr] &&\centerdot \ar[dr] &&\centerdot \ar[dr] &&\centerdot \\ \ar@{.}[r] &\centerdot \ar[dr] \ar@{.}[rr] \ar[ur] &&\centerdot \ar@{.}[rr] \ar[dr] \ar[ur] &&\centerdot \ar[dr] \ar[ur] \ar@{.}[rr] &&\centerdot \ar@{.}[rr] \ar[dr] \ar[ur] &&\centerdot \ar[dr] \ar[ur] \ar@{.}[r] &\\ \centerdot \ar@{.}[rr] \ar[ur] &&\centerdot \ar@{.}[rr] \ar[ur] &&\centerdot \ar@{.}[rr] \ar[ur] &&\centerdot \ar@{.}[rr] \ar[ur] &&\centerdot \ar@{.}[rr] \ar[ur] &&\centerdot } $$ where the first and the last column are identical. We denote by ``$\circ$" in the AR-quiver the indecomposable objects belong to a subcategory and by ``$\cdot$'' the indecomposable objects do not belong to it. Through direct calculation, we know that $$\xymatrix@[email protected]{ &\circ \ar[dr] &&\circ \ar[dr] &&\circ \ar[dr] &&\circ \ar[dr] &&\circ \ar[dr] &&\circ \\ {\mathcal C=} &&\cdot \ar[dr] \ar@{.}[l] \ar@{.}[rr] \ar[ur] &&\cdot \ar@{.}[rr] \ar[dr] \ar[ur] &&\cdot \ar[dr] \ar[ur] \ar@{.}[rr] &&\circ \ar@{.}[rr] \ar[dr] \ar[ur] &&\cdot \ar[dr] \ar[ur] \ar@{.}[r] &\\ &\cdot \ar@{.}[rr] \ar[ur] &&\circ \ar@{.}[rr] \ar[ur] &&\cdot \ar@{.}[rr] \ar[ur] &&\cdot \ar@{.}[rr] \ar[ur] &&\cdot \ar@{.}[rr] \ar[ur] &&\cdot } $$ is pre-cluster tilting. \end{exm} Cotorsion pair is a generalization structure of cluster tilting subcategory on both triangulated and exact categories \cite{N1,DL}, now it is also defined on the extriangulated categories \cite{NP}. We recall its definition, which will be used frequently. \begin{defn}\cite[Definition 2.1]{NP} Let $\mathcal U$ and $\mathcal V$ be two subcategories of $\mathcal B$ which are closed under direct summands. We call $(\mathcal U,\mathcal V)$ a \emph{cotorsion pair} if it satisfies the following conditions: \begin{itemize} \item[(a)] $\mathbb E(\mathcal U,\mathcal V)=0$. \item[(b)] For any object $B\in \mathcal B$, there exist two $\mathbb E$-triangles \begin{align*} V_B\rightarrow U_B\rightarrow B\overset{\delta}{\dashrightarrow},\quad B\rightarrow V^B\rightarrow U^B\overset{\sigma}{\dashrightarrow} \end{align*} satisfying $U_B,U^B\in \mathcal U$ and $V_B,V^B\in \mathcal V$. \end{itemize} \end{defn} By definition of a cotorsion pair, we can immediately conclude: \begin{prop}\label{prop0} Let $\mathcal U$ be a subcategory of $\mathcal B$. Then $(\mathcal U,\mathcal U)$ is a cotorsion pair if and only if $\mathcal U$ is a cluster tilting subcategory. \end{prop} For any subcategory $\mathcal U$, we call $\mathcal U\cap {^{\bot_1}}\mathcal U=$ the coheart of $\mathcal U$. We say $\mathcal U$ is maximal if $\mathcal U$ is maximal among those with the same coheart. \vspace{1mm} Now we assume $\mathcal B$ has enough projectives and enough injectives. We denote by $\mathcal P$ the subcategory of projective objects and by $\mathcal I$ the subcategory of injective objects. Under this assumption, if we have a pre-cluster tilting subcategory $\mathcal C$, then we can get two cotorsion pairs $(\mathcal C,\mathcal C^{\bot_1})$ and $(\mathcal C^{\bot_1},\mathcal C)$ (see Lemma \ref{CP}). Our first main result is the following. \begin{thm}\label{main1} The maps $$\mathcal U \mapsto \mathcal C:=\mathcal U\cap {^{\bot_1}}\mathcal U \text{ and } \mathcal C \mapsto \mathcal U:={^{\bot_1}}\mathcal C $$ give mutually inverse bijections between: \begin{itemize} \item Maximal subcategories $\mathcal U$ which admits two cotorsion pairs $(\mathcal U,\mathcal V),(\mathcal V,\mathcal U)$. \item Pre-cluster tilting subcategories $\mathcal C$. \end{itemize} \end{thm} In order to introduce the second main result, we need the following definition. \begin{defn} Let $\mathcal B_j,j\in \mathcal J$ be extriangulated subcategories of $\mathcal B$. We call that $\mathcal B$ is a direct sum of extriangulated subcategories $\mathcal B_j,j\in\mathcal J$ if it satisfies the following conditions: \begin{itemize} \item Any object $M\in\mathcal B$ is a direct sum of finitely many objects $M_j\in\mathcal B_j$; \item $\operatorname{Hom}\nolimits(\mathcal B_i,\mathcal B_j)=0$, for any $i\neq j$. \end{itemize} In this case, we write $\mathcal B=\bigoplus_{j\in\mathcal J}\mathcal B_j$. An extriangulated category is called connected if it can not be written as direct sum of two non-zero extriangulated subcategories. \end{defn} By \cite[Proposition 3.30]{NP}, $\mathcal B/(\mathcal P\cap \mathcal I)$ is still an extriangulated category. We will show the second main result of this article. \begin{thm}\label{main} Let $\mathcal B/(\mathcal P\cap \mathcal I)$ be connected and $(\mathcal U,\mathcal V),(\mathcal V,\mathcal U)$ be cotorsion pairs on $\mathcal B$. Let $\mathcal C=\mathcal U\cap \mathcal V$, if $\mathcal C\supset \mathcal P\cap \mathcal I$, then the following statements are equivalent. \begin{itemize} \item[(a)] $\mathcal C$ is cluster tilting; \item[(b)] If $(\mathcal S,\mathcal R),(\mathcal R,\mathcal S)$ are cotorsion pairs such that $\mathcal S\cap \mathcal R=\mathcal C$, then $\mathcal S=\mathcal C$; \item[(c)] $\mathcal B/\mathcal C$ is abelian. \item[(d)] $\mathcal B/\mathcal U$ and $\mathcal B/\mathcal V$ are abelian. \end{itemize} \end{thm} Theorem \ref{main1} and Theorem \ref{main} is a generalization of related results of Koenig-Zhu \cite[Theorem 3.3]{KZ}, Demonet-Liu \cite[Theorem 3.2]{DL} and Beligiannis \cite[Theorem 7.3]{B}. This article is organized as follows. In Section 2, we review some elementary definitions and facts of extriangulated category that we need. In Section 3, we prove our first and second main result. \section{Preliminaries} Let us briefly recall the definition and basic properties of extriangulated categories from \cite{NP}. Throughout this paper, we assume that $\mathcal B$ is an additive category. \begin{defn} Suppose that $\mathcal B$ is equipped with an additive bifunctor $\mathbb{E}\colon\mathcal B^\mathrm{op}\times\mathcal B\to\mathsf{Ab}$, where $\mathsf{Ab}$ is the category of abelian groups. For any pair of objects $A,C\in\mathcal B$, an element $\delta\in\mathbb{E}(C,A)$ is called an {\it $\mathbb{E}$-extension}. Thus formally, an $\mathbb E$-extension is a triplet $(A,\delta,C)$. For any $A,C\in\mathcal C$, the zero element $0\in\mathbb E(C,A)$ is called the \emph{spilt $\mathbb E$-extension}. Let $\delta\in\mathbb{E}(C,A)$ be any $\mathbb{E}$-extension. By the functoriality, for any $a\in\mathcal B(A,A^{\prime})$ and $c\in\mathcal B(C^{\prime},C)$, we have $\mathbb{E}$-extensions \[ \mathbb{E}(C,a)(\delta)\in\mathbb{E}(C,A^{\prime})\ \ \text{and}\ \ \mathbb{E}(c,A)(\delta)\in\mathbb{E}(C^{\prime},A). \] We abbreviately denote them by $a_{\ast}\delta$ and $c^{\ast}\delta$. In this terminology, we have \[ \mathbb{E}(c,a)(\delta)=c^{\ast} a_{\ast}\delta=a_{\ast} c^{\ast}\delta \] in $\mathbb{E}(C^{\prime},A^{\prime})$. \end{defn} \begin{defn} Let $\delta\in\mathbb{E}(C,A)$ and $\delta^{\prime}\in\mathbb{E}(C^{\prime},A^{\prime})$ be two pair of $\mathbb{E}$-extensions. A {\it morphism} $(a,c)\colon\delta\to\delta^{\prime}$ of $\mathbb{E}$-extensions is a pair of morphisms $a\in\mathcal B(A,A^{\prime})$ and $c\in\mathcal B(C,C^{\prime})$ in $\mathcal B$, satisfying the equality \[ a_{\ast}\delta=c^{\ast}\delta^{\prime}. \] We simply denote it as $(a,c)\colon\delta\to\delta^{\prime}$. \end{defn} \begin{defn} Let $\delta=(A,\delta,C)$ and $\delta^{\prime}=(A^{\prime},\delta^{\prime},C^{\prime})$ be any pair of $\mathbb{E}$-extensions. Let \[ C\xrightarrow{~\iota_C~}C\oplus C^{\prime}\xleftarrow{~\iota_{C^{\prime}}~}C^{\prime} \] and \[ A\xrightarrow{~p_A~}A\oplus A^{\prime}\xleftarrow{~p_{A^{\prime}}~}A^{\prime} \] be coproduct and product in $\mathcal B$, respectively. Remark that, by the additivity of $\mathbb{E}$, we have a natural isomorphism \[ \mathbb{E}(C\oplus C^{\prime},A\oplus A^{\prime})\simeq \mathbb{E}(C,A)\oplus\mathbb{E}(C,A^{\prime})\oplus\mathbb{E}(C^{\prime},A)\oplus\mathbb{E}(C^{\prime},A^{\prime}). \] Let $\delta\oplus\delta^{\prime}\in\mathbb{E}(C\oplus C^{\prime},A\oplus A^{\prime})$ be the element corresponding to $(\delta,0,0,\delta^{\prime})$ through this isomorphism. This is the unique element which satisfies $$ \mathbb{E}(\iota_C,p_A)(\delta\oplus\delta^{\prime})=\delta,\ \mathbb{E}(\iota_C,p_{A^{\prime}})(\delta\oplus\delta^{\prime})=0,\ \mathbb{E}(\iota_{C^{\prime}},p_A)(\delta\oplus\delta^{\prime})=0,\ \mathbb{E}(\iota_{C^{\prime}},p_{A^{\prime}})(\delta\oplus\delta^{\prime})=\delta^{\prime}. $$ \end{defn} \begin{defn} Let $A,C\in\mathcal B$ be any pair of objects. Two sequences of morphisms in $\mathcal B$ \[ A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C\ \ \text{and}\ \ A\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C \] are said to be {\it equivalent} if there exists an isomorphism $b\in\mathcal B(B,B^{\prime})$ which makes the following diagram commutative. \[ \xy (-16,0)*+{A}="0"; (3,0)*+{}="1"; (0,8)*+{B}="2"; (0,-8)*+{B^{\prime}}="4"; (-3,0)*+{}="5"; (16,0)*+{C}="6"; {\ar^{x} "0";"2"}; {\ar^{y} "2";"6"}; {\ar_{x^{\prime}} "0";"4"}; {\ar_{y^{\prime}} "4";"6"}; {\ar^{b}_{\simeq} "2";"4"}; {\ar@{}|{} "0";"1"}; {\ar@{}|{} "5";"6"}; \endxy \] We denote the equivalence class of $A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C$ by $[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]$. \end{defn} \begin{defn} $\ \ $ \begin{enumerate} \item[(1)] For any $A,C\in\mathcal B$, we denote as \[ 0=[A\overset{\Big[\raise1ex\hbox{\leavevmode\vtop{\baselineskip-8ex \lineskip1ex \ialign{#\crcr{$\scriptstyle{1}$}\crcr{$\scriptstyle{0}$}\crcr}}}\Big]}{\longrightarrow}A\oplus C\overset{[0\ 1]}{\longrightarrow}C]. \] \item[(2)] For any $[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]$ and $[A^{\prime}\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C^{\prime}]$, we denote as \[ [A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]\oplus [A^{\prime}\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C^{\prime}]=[A\oplus A^{\prime}\overset{x\oplus x^{\prime}}{\longrightarrow}B\oplus B^{\prime}\overset{y\oplus y^{\prime}}{\longrightarrow}C\oplus C^{\prime}]. \] \end{enumerate} \end{defn} \begin{defn} Let $\mathfrak{s}$ be a correspondence which associates an equivalence class $\mathfrak{s}(\delta)=[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]$ to any $\mathbb{E}$-extension $\delta\in\mathbb{E}(C,A)$. This $\mathfrak{s}$ is called a {\it realization} of $\mathbb{E}$, if it satisfies the following condition $(\star)$. In this case, we say that the sequence $A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C$ {\it realizes} $\delta$, whenever it satisfies $\mathfrak{s}(\delta)=[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]$. \begin{itemize} \item[$(\star)$] Let $\delta\in\mathbb{E}(C,A)$ and $\delta^{\prime}\in\mathbb{E}(C^{\prime},A^{\prime})$ be any pair of $\mathbb{E}$-extensions, with \[\mathfrak{s}(\delta)=[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C]\text{ and } \mathfrak{s}(\delta^{\prime})=[A^{\prime}\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C^{\prime}].\] Then, for any morphism $(a,c)\colon\delta\to\delta^{\prime}$, there exists $b\in\mathcal B(B,B^{\prime})$ which makes the following diagram commutative. $$ \xy (-12,6)*+{A}="0"; (0,6)*+{B}="2"; (12,6)*+{C}="4"; (-12,-6)*+{A^{\prime}}="10"; (0,-6)*+{B^{\prime}}="12"; (12,-6)*+{C^{\prime}}="14"; {\ar^{x} "0";"2"}; {\ar^{y} "2";"4"}; {\ar_{a} "0";"10"}; {\ar^{b} "2";"12"}; {\ar^{c} "4";"14"}; {\ar^{x^{\prime}} "10";"12"}; {\ar^{y^{\prime}} "12";"14"}; {\ar@{}|{} "0";"12"}; {\ar@{}|{} "2";"14"}; \endxy $$ \end{itemize} In the above situation, we say that the triplet $(a,b,c)$ {\it realizes} $(a,c)$. \end{defn} \begin{defn} Let $\mathcal B,\mathbb{E}$ be as above. A realization of $\mathbb{E}$ is said to be {\it additive}, if it satisfies the following conditions. \begin{itemize} \item[{\rm (i)}] For any $A,C\in\mathcal B$, the split $\mathbb{E}$-extension $0\in\mathbb{E}(C,A)$ satisfies \[ \mathfrak{s}(0)=0. \] \item[{\rm (ii)}] For any pair of $\mathbb{E}$-extensions $\delta\in\mathbb{E}(C,A)$ and $\delta^{\prime}\in\mathbb{E}(C^{\prime},A^{\prime})$, we have \[ \mathfrak{s}(\delta\oplus\delta^{\prime})=\mathfrak{s}(\delta)\oplus\mathfrak{s}(\delta^{\prime}). \] \end{itemize} \end{defn} \begin{defn}\cite[Definition 2.12]{NP} A triplet $(\mathcal B,\mathbb{E},\mathfrak{s})$ is called an {\it extriangulated category} if it satisfies the following conditions. \begin{itemize} \item[{\rm (ET1)}] $\mathbb{E}\colon\mathcal B^{\mathrm{op}}\times\mathcal B\to\mathsf{Ab}$ is an additive bifunctor. \item[{\rm (ET2)}] $\mathfrak{s}$ is an additive realization of $\mathbb{E}$. \item[{\rm (ET3)}] Let $\delta\in\mathbb{E}(C,A)$ and $\delta^{\prime}\in\mathbb{E}(C^{\prime},A^{\prime})$ be any pair of $\mathbb{E}$-extensions, realized as \[ \mathfrak{s}(\delta)=[A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C],\ \ \mathfrak{s}(\delta^{\prime})=[A^{\prime}\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C^{\prime}]. \] For any commutative square $$ \xy (-12,6)*+{A}="0"; (0,6)*+{B}="2"; (12,6)*+{C}="4"; (-12,-6)*+{A^{\prime}}="10"; (0,-6)*+{B^{\prime}}="12"; (12,-6)*+{C^{\prime}}="14"; {\ar^{x} "0";"2"}; {\ar^{y} "2";"4"}; {\ar_{a} "0";"10"}; {\ar^{b} "2";"12"}; {\ar^{x^{\prime}} "10";"12"}; {\ar^{y^{\prime}} "12";"14"}; {\ar@{}|{} "0";"12"}; \endxy $$ in $\mathcal B$, there exists a morphism $(a,c)\colon\delta\to\delta^{\prime}$ satisfying $cy=y^{\prime}b$. \item[{\rm (ET3)$^{\mathrm{op}}$}] Dual of {\rm (ET3)}. \item[{\rm (ET4)}] Let $\delta\in\mathbb{E}(D,A)$ and $\delta^{\prime}\in\mathbb{E}(F,B)$ be $\mathbb{E}$-extensions realized by \[ A\overset{f}{\longrightarrow}B\overset{f^{\prime}}{\longrightarrow}D\ \ \text{and}\ \ B\overset{g}{\longrightarrow}C\overset{g^{\prime}}{\longrightarrow}F \] respectively. Then there exist an object $E\in\mathcal B$, a commutative diagram $$ \xy (-21,7)*+{A}="0"; (-7,7)*+{B}="2"; (7,7)*+{D}="4"; (-21,-7)*+{A}="10"; (-7,-7)*+{C}="12"; (7,-7)*+{E}="14"; (-7,-21)*+{F}="22"; (7,-21)*+{F}="24"; {\ar^{f} "0";"2"}; {\ar^{f^{\prime}} "2";"4"}; {\ar@{=} "0";"10"}; {\ar_{g} "2";"12"}; {\ar^{d} "4";"14"}; {\ar^{h} "10";"12"}; {\ar^{h^{\prime}} "12";"14"}; {\ar_{g^{\prime}} "12";"22"}; {\ar^{e} "14";"24"}; {\ar@{=} "22";"24"}; {\ar@{}|{} "0";"12"}; {\ar@{}|{} "2";"14"}; {\ar@{}|{} "12";"24"}; \endxy $$ in $\mathcal B$, and an $\mathbb{E}$-extension $\delta^{\prime\prime}\in\mathbb{E}(E,A)$ realized by $A\overset{h}{\longrightarrow}C\overset{h^{\prime}}{\longrightarrow}E$, which satisfy the following compatibilities. \begin{itemize} \item[{\rm (i)}] $D\overset{d}{\longrightarrow}E\overset{e}{\longrightarrow}F$ realizes $f^{\prime}_{\ast}\delta^{\prime}$, \item[{\rm (ii)}] $d^{\ast}\delta^{\prime\prime}=\delta$, \item[{\rm (iii)}] $f_{\ast}\delta^{\prime\prime}=e^{\ast}\delta^{\prime}$. \end{itemize} \item[{\rm (ET4)$^{\mathrm{op}}$}] Dual of {\rm (ET4)}. \end{itemize} \end{defn} \begin{rem} Note that both exact categories and triangulated categories are extriangulated categories, see \cite[Example 2.13]{NP} and extension-closed subcategories of extriangulated categories are again extriangulated, see \cite[Remark 2.18]{NP} . Moreover, there exist extriangulated categories which are neither exact categories nor triangulated categories, see \cite[Proposition 3.30]{NP} and \cite[Example 4.14]{ZZ}. \end{rem} We will use the following terminology. \begin{defn}{\cite{NP}} Let $(\mathcal B,\mathbb E,\mathfrak{s})$ be an extriangulated category. \begin{itemize} \item[(1)] A sequence $A\xrightarrow{~x~}B\xrightarrow{~y~}C$ is called a {\it conflation} if it realizes some $\mathbb E$-extension $\delta\in\mathbb E(C,A)$. In this case, $x$ is called an {\it inflation} and $y$ is called a {\it deflation}. \item[(2)] If a conflation $A\xrightarrow{~x~}B\xrightarrow{~y~}C$ realizes $\delta\in\mathbb{E}(C,A)$, we call the pair $( A\xrightarrow{~x~}B\xrightarrow{~y~}C,\delta)$ an {\it $\mathbb E$-triangle}, and write it in the following way. $$A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C\overset{\delta}{\dashrightarrow}$$ We usually do not write this $``\delta"$ if it is not used in the argument. \item[(3)] Let $A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C\overset{\delta}{\dashrightarrow}$ and $A^{\prime}\overset{x^{\prime}}{\longrightarrow}B^{\prime}\overset{y^{\prime}}{\longrightarrow}C^{\prime}\overset{\delta^{\prime}}{\dashrightarrow}$ be any pair of $\mathbb E$-triangles. If a triplet $(a,b,c)$ realizes $(a,c)\colon\delta\to\delta^{\prime}$, then we write it as $$\xymatrix{ A \ar[r]^x \ar[d]^a & B\ar[r]^y \ar[d]^{b} & C\ar@{-->}[r]^{\delta}\ar[d]^c&\\ A'\ar[r]^{x'} & B' \ar[r]^{y'} & C'\ar@{-->}[r]^{\delta'} &}$$ and call $(a,b,c)$ a {\it morphism of $\mathbb E$-triangles}. \item[(4)] An object $P\in\mathcal B$ is called {\it projective} if for any $\mathbb E$-triangle $A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C\overset{\delta}{\dashrightarrow}$ and any morphism $c\in\mathcal B(P,C)$, there exists $b\in\mathcal B(P,B)$ satisfying $yb=c$. We denote the subcategory of projective objects by $\mathcal P\subseteq\mathcal B$. Dually, the subcategory of injective objects is denoted by $\mathcal I\subseteq\mathcal B$. \item[(5)] We say that $\mathcal B$ {\it has enough projective objects} if for any object $C\in\mathcal B$, there exists an $\mathbb E$-triangle $A\overset{x}{\longrightarrow}P\overset{y}{\longrightarrow}C\overset{\delta}{\dashrightarrow}$ satisfying $P\in\mathcal P$. Dually we can define $\mathcal B$ {\it has enough injective objects}. \item[(6)] Let $\mathcal{X}$ be a subcategory of $\mathcal B$. We say $\mathcal{X}$ is {\it extension-closed} if a conflation $A\rightarrowtail B\twoheadrightarrow C$ satisfies $A,C\in\mathcal{X}$, then $B\in\mathcal{X}$. \end{itemize} \end{defn} In this article, we always assume $\mathcal B$ has enough projectives and enough injectives. \medskip By \cite{NP}, we give the following useful remark, which will be used later in the proofs. \begin{rem}\label{useful} Let $\xymatrix{A\ar[r]^a &B \ar[r]^b &C \ar@{-->}[r] &}$ and $\xymatrix{X\ar[r]^x &Y \ar[r]^y &Z \ar@{-->}[r] &}$ be two $\mathbb E$-triangles. Then \begin{itemize} \item[(a)] In this following commutative diagram $$\xymatrix{ X\ar[r]^x \ar[d]_f &Y \ar[d]^g \ar[r]^y &Z \ar[d]^h \ar@{-->}[r] &\\ A\ar[r]^a &B \ar[r]^b &C \ar@{-->}[r] &} $$ $f$ factors through $x$ if and only if $h$ factors through $b$. \item[(b)] In the following commutative diagram $$\xymatrix{ A\ar[r]^a \ar[d]_s &B \ar[d]^r \ar[r]^b &C \ar[d]^t \ar@{-->}[r] &\\ X\ar[r]^x \ar[d]_f &Y \ar[d]^g \ar[r]^y &Z \ar[d]^h \ar@{-->}[r] &\\ A\ar[r]^a &B \ar[r]^b &C \ar@{-->}[r] &} $$ $fs=1_A$ implies $B$ is a direct summand of $C\oplus Y$ and $C$ is a direct summand of $Z\oplus B$; $ht=1_C$ implies $B$ is a direct summand of $A\oplus Y$ and $A$ is a direct summand of $X\oplus B$. \item[(c)] If we have $b:B\xrightarrow{d_1} D \xrightarrow{d_2} C$ and $d_2: D\xrightarrow{d_3} B\xrightarrow{b} C$, then $B$ is a direct summand of $A\oplus D$.\\ If we have $a: A\xrightarrow{e_1} E\xrightarrow{e_2} B$ and $e_1:A \xrightarrow{a} B\xrightarrow{e_3} E$, then $B$ is a direct summand of $C\oplus E$. \end{itemize} \end{rem} We first recall the following proposition (\cite[Proposition 1.20]{LN}), which (also the dual of it) will be used many times in the article. \begin{prop} Let $A\overset{x}{\longrightarrow}B\overset{y}{\longrightarrow}C\overset{\delta}{\dashrightarrow}$ be any $\mathbb E$-triangle, let $f\colon A\rightarrow D$ be any morphism, and let $D\overset{d}{\longrightarrow}E\overset{e}{\longrightarrow}C\overset{f_{\ast}\delta}{\dashrightarrow}$ be any $\mathbb E$-triangle realizing $f_{\ast}\delta$. Then there is a morphism $g$ which gives a morphism of $\mathbb E$-triangles $$\xymatrix{ A \ar[r]^{x} \ar[d]_f &B \ar[r]^{y} \ar[d]^g &C \ar@{=}[d]\ar@{-->}[r]^{\delta}&\\ D \ar[r]_{d} &E \ar[r]_{e} &C\ar@{-->}[r]_{f_{\ast}\delta}& } $$ and moreover, the sequence $A\overset{\svecv{f}{x}}{\longrightarrow}D\oplus B\overset{\svech{d}{-g}}{\longrightarrow}E\overset{e^{\ast}\delta}{\dashrightarrow}$ becomes an $\mathbb E$-triangle. \end{prop} We prove the following lemma related to cotorsion pairs. \begin{lem}\label{CP} If $\mathcal C$ is rigid and strongly contravariantly finite, then $(\mathcal C,\mathcal C^{\bot_1})$ is a cotorsion pair. \end{lem} \begin{proof} Since $\mathcal B$ has enough injectives, any object $A\in\mathcal B$ admits an $\mathbb E$-triangle $\xymatrix{A \ar[r] &I \ar[r] &B \ar@{-->}[r] &}$ where $I$ is injective. Since $\mathcal C$ is strongly contravariantly finite, the object $B$ admits a conflation $$\xymatrix{B_1 \ar[r] &C_0 \ar[r]^{f_0} &B \ar@{-->}[r] &}$$ where $f_0$ is a right $\mathcal C$-approximation of $B$. The rigidity of $\mathcal C$ implies $B_1\in \mathcal C^{\bot_1}$. We have the following commutative diagram $$\xymatrix{ &A \ar[d] \ar@{=}[r] &A \ar[d]\\ B_1 \ar@{=}[d] \ar[r] &X \ar[r] \ar[d] &I \ar@{-->}[r]\ar[d]&\\ B_1 \ar[r] &C_0\ar@{-->}[d] \ar[r]^{f_0} &B\ar@{-->}[d]\ar@{-->}[r]&\\ &&} $$ where $X\in \mathcal C^{\bot_1}$. Hence by definition, the pair $(\mathcal C,\mathcal C^{\bot_1})$ is a cotorsion pair. \end{proof} \section{Main results} Let $\mathcal A$ be an additive category and $\mathcal X$ be a subcategory of $\mathcal A$. We denote by $\mathcal A/\mathcal X$ the category whose objects are objects of $\mathcal A$ and whose morphisms are elements of $\operatorname{Hom}\nolimits_{\mathcal A}(A,B)/\mathcal X(A,B)$ for $A,B\in\mathcal A$, where $\mathcal X(A,B)$ the subgroup of $\operatorname{Hom}\nolimits_{\mathcal A}(A,B)$ consisting of morphisms which factor through an object in $\mathcal X$. Such category is called the quotient category of $\mathcal A$ by $\mathcal X$. For any morphism $f\colon A\to B$ in $\mathcal A$, we denote by $\overline{f}$ the image of $f$ under the natural quotient functor $\mathcal A\to\mathcal A/\mathcal X$. \vspace{1mm} We first introduce some notions. Let $\mathcal B'$ and $\mathcal B''$ be two subcategories of $\mathcal B$, denote by $\operatorname{CoCone}\nolimits(\mathcal B',\mathcal B'')$ the subcategory of objects $X$ admitting an $\mathbb E$-triangle $\xymatrix{X \ar[r] &B'\ar[r] &B''\ar@{-->}[r] &}$ where $B'\in \mathcal B'$ and $B''\in \mathcal B''$. We denote by $\operatorname{Cone}\nolimits(\mathcal B',\mathcal B'')$ the subcategory of objects $Y$ admitting an $\mathbb E$-triangle $\xymatrix{B' \ar[r] &B''\ar[r] &Y\ar@{-->}[r] &}$ where $B'\in \mathcal B'$ and $B''\in \mathcal B''$. Let $\Omega \mathcal B'=\operatorname{CoCone}\nolimits(\mathcal P,\mathcal B')$ and $\Sigma \mathcal B'=\operatorname{Cone}\nolimits(\mathcal B',\mathcal I)$. We write an object $D$ in the form $\Omega B$ if it admits an $\mathbb E$-triangle $\xymatrix{D \ar[r] &P \ar[r] &B \ar@{-->}[r] &}$ where $P\in \mathcal P$. We write an object $D'$ in the form $\Sigma B'$ if it admits an $\mathbb E$-triangle $\xymatrix{B' \ar[r] &I \ar[r] &D' \ar@{-->}[r] &}$ where $I\in \mathcal I$. In the rest of this article, let $(\mathcal U,\mathcal V),(\mathcal V,\mathcal U)$ be cotorsion pairs, we denote $\mathcal U\cap \mathcal V$ by $\mathcal C$. We denote subcategory $\{ \text{direct sums of objects in }\mathcal U \text{ and objects in }\mathcal V\}$ by $\mathcal K$, and we say $\mathcal K=\mathcal U+\mathcal V$. Let $\mathcal H=\operatorname{CoCone}\nolimits(\mathcal C,\mathcal U)\cap \operatorname{Cone}\nolimits(\mathcal V,\mathcal C)$, $\mathcal H/\mathcal C$ is called the heart of $(\mathcal U,\mathcal V)$, it is abelian by \cite{LN}. Let $H$ be the cohomological functor defined in \cite{LN}, it sends an $\mathbb E$-triangle $\xymatrix{A \ar[r]^f &B \ar[r]^g &C \ar@{-->}[r] &}$ to an exact sequence $H(A)\xrightarrow{H(f)} H(B)\xrightarrow{H(g)} H(C)$ in $\mathcal H/\mathcal C$, moreover, $H(B)=0$ if and only if $B\in \mathcal K$. Since $\mathcal U,\mathcal V$ are extension closed subcategories of $\mathcal B$, they are extriangulated subcategories. Moreover, $\mathcal C$ is the subcategory of enough projective-injective objects in $\mathcal U$ and $\mathcal V$, according to \cite{NP}, $\mathcal U/\mathcal C$ and $\mathcal V/\mathcal C$ are triangulated categories. \begin{lem}\label{direct summand} $\mathcal H$ is closed under direct summands. \end{lem} \begin{proof} Since $\mathcal H=\operatorname{CoCone}\nolimits(\mathcal C,\mathcal U)\cap \operatorname{Cone}\nolimits(\mathcal V,\mathcal C)$, we will show that $\operatorname{CoCone}\nolimits(\mathcal C,\mathcal U)$ is closed under direct summands. By dual, we can show that $\operatorname{Cone}\nolimits(\mathcal V,\mathcal C)$ is closed under direct summands. Assume we have an $\mathbb E$-triangle $$\xymatrix{X\oplus Y \ar[r]^-{\svech{x}{y}} &C_1 \ar[r] &U_2 \ar@{-->}[r] &}$$ where $C_1\in \mathcal C$ and $\mathcal U_2\in \mathcal U$, then $x$ is an inflation and it admits an $\mathbb E$-triangle $$\xymatrix{X \ar[r]^-{x} &C_1 \ar[r]^c &U \ar@{-->}[r] &}.$$ Since $\mathbb E(U_2,C_1)=0$, there is a morphism $f\colon C_1\to C_1$ such that $f(x\ \ y)=(x\ \ 0)$. In particular, we have $fx=x$ and $fy=0$. Hence we have the following commutative diagram $$\xymatrix{ X \ar[r]^-{x} \ar[d]_-{\svecv{1}{0}} &C_1 \ar@{=}[d] \ar[r]^c &U \ar[d]^a \ar@{-->}[r] &\\ X\oplus Y \ar[r]^-{\svech{x}{y}} \ar[d]_-{\svech{1}{0}} &C_1 \ar[d]^f \ar[r] &U_2 \ar[d]^b \ar@{-->}[r] &\\ X \ar[r]^-{x} &C_1 \ar[r]^c &U \ar@{-->}[r] &. } $$ It follows that there exists a morphism $d:U\rightarrow C_1$ such that $1_{U}-ba=cd$. Hence $U$ is a direct summand of $C_1\oplus U_2\in \mathcal U$ and then $U\in \mathcal U$. This implies $X\in \operatorname{CoCone}\nolimits(\mathcal C,\mathcal U)$. \end{proof} \textbf{Proof of Theorem \ref{main1}:} \begin{proof} We show $(\mathcal C,\mathcal K)$ is a cotorsion pair, then dually $(\mathcal K,\mathcal C)$ is also a cotorsion pair.\\ Let $B$ be an object in $\mathcal B$, since $(\mathcal V,\mathcal U)$ is a cotorsion pair, $B$ admits a commutative diagram of $\mathbb E$-triangles $$\xymatrix{ \Omega U_B \ar[r] \ar[d]_f &P \ar[r] \ar[d] &U_B \ar@{=}[d] \ar@{-->}[r] &\\ B \ar[r] &V_B \ar[r] &U_B \ar@{-->}[r] & }$$ where $U_B\in \mathcal U$ and $V_B$ and $P\in \mathcal P$. We get $H(f)$ is an epimorphism. $\Omega U_B$ admits the following commutative diagram $$\xymatrix{ \Omega U_B \ar[r] \ar[d] &P \ar[r] \ar[d] &U_B \ar@{=}[d] \ar@{-->}[r] &\\ U \ar[r] \ar[d] &C \ar[r] \ar[d] &U_B \ar@{-->}[r] &\\ V \ar@{=}[r] \ar@{-->}[d]^{\sigma} &V \ar@{-->}[d]\\ &&&& } $$ where $U\in \mathcal U$ and $V\in \mathcal V$. From the second column we get $C= P\oplus V\in \mathcal V$. From the second row we get $C\in \mathcal U$, hence $C\in \mathcal C$ and $V$ is a direct summand of $C$, then $V\in \mathcal C$. Now $B$ admits the following commutative diagram: $$\xymatrix{ \Omega U_B \ar[r] \ar[d]_f &U \ar[r] \ar[d] &V \ar@{=}[d] \ar@{-->}[r]^{\sigma} &\\ B \ar[r]_g &K \ar[r]_h &V \ar@{-->}[r]^{f_*\sigma} & } $$ By applying $H$ we get the following exact sequence $H(\Omega B)\xrightarrow{H(f)} H(B)\xrightarrow{H(g)} H(K) \xrightarrow{H(h)} H(V)=0$. Since $H(f)$ is an epimorphism, we have $H(g)=0$, hence $H(K)=0$, which implies $K\in \mathcal K$. $B$ admits an $\mathbb E$-triangle $\xymatrix{\Omega B \ar[r] &P_B \ar[r] &B \ar@{-->}[r] &}$ where $P_B\in \mathcal P$, by the previous argument, $\Omega B$ admits an $\mathbb E$-triangle $\xymatrix{\Omega B \ar[r] &K' \ar[r] &C' \ar@{-->}[r] &}$ where $K'\in \mathcal K$ and $C'\in \mathcal C$, hence we get the following commutative diagram\\ $$\xymatrix{ \Omega B \ar[d] \ar[r] &K' \ar[d] \ar[r] &C' \ar@{=}[d] \ar@{-->}[r] &\\ P_B \ar[r] \ar[d] &P_B\oplus C' \ar[r] \ar[d] & C' \ar@{-->}[r] &\\ B \ar@{=}[r] \ar@{-->}[d] &B \ar@{-->}[d] \\ && && } $$ Since $P_B\oplus C'\in \mathcal C$ and $\mathbb E(\mathcal C,\mathcal K)=0$, by definition $(\mathcal C,\mathcal K)$ is a cotorsion pair.\\ Hence $\mathcal C$ is pre-cluster tilting. Now if $\mathcal U$ is maximal, $\mathcal U$ has to be $\mathcal K$.\\ On the other hand, if $\mathcal C$ is pre-cluster tilting, by Lemma \ref{CP} and its dual, we have cotorsion pairs $(\mathcal K,\mathcal C),(\mathcal C,\mathcal K)$ where $\mathcal K$ is maximal by the previous argument. \end{proof} \begin{lem}\label{serre} Let $A\xrightarrow{ f} B$ be a morphism in $\mathcal B$ such that $A,B$ do not have direct summand in $\mathcal V$. Then \begin{itemize} \item[(a)] If $\overline f$ is a monomorphism in $\mathcal B/\mathcal V$ and $B\in \mathcal U$, then $A\in \mathcal U$. \item[(b)] If $\overline f$ is an epimorphism in $\mathcal B/\mathcal V$ and $A\in \mathcal U$, then $B\in \mathcal U$. \end{itemize} \end{lem} \begin{proof} Since $A$ admits an $\mathbb E$-triangle $\xymatrix{A\ar[r] &V^A \ar[r] &U^A \ar@{-->}[r] &}$ where $V^A\in \mathcal V$ and $U^A\in \mathcal U$, and $B$ admits an $\mathbb E$-triangle $\xymatrix{ U_B \ar[r] &V_B \ar[r] &B \ar@{-->}[r] &}$ where $U_B\in \mathcal U$ and $V_B\in \mathcal V$, we have the following commutative diagram $$\xymatrix{ U_B \ar@{=}[d] \ar[r] &K' \ar[r]^{g'} \ar[d] &A\ar[r] \ar[d]_f &V^A \ar[d] \ar[r] &U^A \ar@{=}[d]\\ U_B \ar[r] &V_B \ar[r] &B \ar[r]_g &K \ar[r] &U^A. } $$ We only prove (a), since (b) is similar.\\ (a) If $\overline f$ is a monomorphism in $\mathcal B/\mathcal V$, then $\overline {g'}=0$, which implies $g'$ factors through $\mathcal V$. If $B\in \mathcal U$, we have an exact sequence $H(K') \xrightarrow{H(g')=0} H(A) \xrightarrow{H(f)} H(B)=0$, which implies $H(A)=0$, hence $A\in \mathcal K$. Since $\mathcal K=\mathcal U+\mathcal V$ and $A$ has no direct summand in $\mathcal V$, we get $A\in \mathcal U$. \end{proof} \begin{prop} If $\mathcal B/\mathcal V$ is abelian, then $\mathcal U/\mathcal C$ is abelian. \end{prop} \begin{proof} Note that a morphism in $\mathcal U$ factors through $\mathcal V$ if and only if it factors through $\mathcal C$. Since $\mathcal B/\mathcal V$ is abelian, then by Lemma \ref{serre}, $\mathcal U/\mathcal C$ has kernels and cokernals. Now it is enough to show any monomorphism in $\mathcal U/\mathcal C$ is also a monomorphism in $\mathcal B/\mathcal V$, the case for epimorphism is by dual. Let $\overline k:U_1\to U_2$ be a monomorphism in $\mathcal U/\mathcal C$, it has a kernel $\overline f:X\to U_1$ in $\mathcal B/\mathcal V$. By Lemma \ref{serre}, $X\in \mathcal U$, then $\overline {kf}=0$ implies $\overline f=0$, which means $\overline k$ is a kernel in $\mathcal U/\mathcal C$. Then $\overline k$ is the kernel of its cokernel $\overline l:U_2\to Y$ in $\mathcal B/\mathcal V$. By Lemma \ref{serre}, $Y\in \mathcal U$, hence $\overline k$ is the kernel some morphism in $\mathcal U/\mathcal C$. This shows that $\mathcal U/\mathcal C$ is abelian. \end{proof} \begin{rem} Since $\mathcal U/\mathcal C$ is also triangulated, it is semi-simple, which means any monomorphism is a section and any epimorphism is a retraction. \end{rem} \begin{cor}\label{zero-iso} If $\mathcal B/\mathcal V$ is abelian, assume that $f: U_1\to U_2$ is a morphism where $U_1,U_2$ are indecomposable objects in $\mathcal U$, then $f$ is an isomorphism in $\mathcal B$ or factors through $\mathcal C$. \end{cor} \begin{proof} If $\overline f\neq 0$, then $\overline f=\overline {hg}$ where $0 \neq \overline g\colon U_1\to B$ is an epimorphism and $0\neq \overline h\colon B\to U_2$ is a monomorphism. By Lemma \ref{serre}, $B$ lies in $\mathcal U$. Since $\mathcal U/\mathcal C$ is semi-simple, $\overline g$ splits, then $B\simeq U_1$ in $\mathcal B/\mathcal V$. By the same method we can get $B\simeq U_2$. Hence $\overline f:U_1\xrightarrow{\simeq} U_2$ in $\mathcal B/\mathcal V$. Since $U_1,U_2$ are indecomposable, then $f$ is an isomorphism in $\mathcal B$. \end{proof} Let $\widetilde{\mathcal K}$ (resp. $\widetilde{\mathcal U}$, $\widetilde{\mathcal V}$) be subcategory of objects which do not have direct summand in $\mathcal C$. \begin{lem}\label{app} Let $\mathcal D$ be a rigid subcategory, $\xymatrix{D \ar[r]^s &C \ar[r]^t &K \ar@{-->}[r] &}$ be an $\mathbb E$-triangle where $K\in \widetilde{\mathcal K}$ and $t$ be a right $\mathcal C$-approximation (resp. $\mathcal P$ or $\mathcal I$-approximation). Then \begin{itemize} \item[(a)] If $K$ is indecomposable, we have $D=C_0\oplus X$ where $C_0\in \mathcal C$ (resp. $\mathcal P$ or $\mathcal I$) and $X$ is indecomposable and $X$ does not belong to $\mathcal C$. Moreover, if $t$ is right minimal, then $D=X$. \item[(b)] $K$ is indecomposable if $D$ is indecomposable. \end{itemize} \end{lem} \begin{proof} We only show the case when $C\in \mathcal C$, the others are similar. Since $K\in \widetilde{\mathcal K}$, $D$ does not belong to $\mathcal C$. (a) Let $X$ be an indecomposable direct summand of $D$ that does not belong to $\mathcal C$, we have the following commutative diagram $$\xymatrix{ X \ar[r]^-{x} \ar[d]_-{\alpha} &C \ar@{=}[d] \ar[r]^c &K' \ar[d]^a \ar@{-->}[r] &\\ D \ar[r]^s \ar[d]_-{\beta} &C \ar[d]^f \ar[r]^t &K \ar[d]^b \ar@{-->}[r] &\\ X \ar[r]^-{x} &C \ar[r]^c &K' \ar@{-->}[r] &. } $$ where $\beta\alpha=1_{X}$. This diagram implies that $K'$ is a direct summand of $K\oplus C$. If $K'$ is also a direct summand of $C$, the first row splits, hence $X\in \mathcal C$, a contradiction. This means $K'=K\oplus C'$ where $C'\in \mathcal C$. Hence we have the following commutative diagram $$\xymatrix{ D \ar[r]^s \ar[d]_{\beta'} &C \ar[r]^t \ar[d]^c &K \ar[d]^-{\svecv{1_K}{0}} \ar@{-->}[r] &\\ X \ar[r]^-{x} \ar[d]_-{\alpha'} &C \ar@{=}[d] \ar[r]^c &K\oplus C' \ar[d]^-{\svech{1_K}{0}} \ar@{-->}[r] &\\ D \ar[r]^s \ &C \ar[r]^t &K \ar@{-->}[r] & } $$ This diagram implies that $D$ is a direct summand of $C\oplus X$. Since $D$ does not belong to $\mathcal C$, we have $D=C_0\oplus X$. If $t$ is right minimal, from the diagram above get $c$ is an isomorphism, hence $\alpha'\beta'$ is also an isomorphism. Since $X$ is indecomposable, $D$ is also indecomposable and $D=X$. \medskip (b) If $t=0$, we have that $C$ is a direct summand of $D$, hence $D\simeq C\in\mathcal C$, a contradiction. We can assume $$C\xrightarrow{t=\svecv{t_1}{t_2}} K_1\oplus K_2=K$$ where $K_1$ is indecomposable and $t_1\neq 0$, then we have the following commutative diagram $$\xymatrix{ D \ar[r]^-{s} \ar[d]_-{\alpha_1} &C \ar@{=}[d] \ar[r]^-{\svecv{t_1}{t_2}} &K_1\oplus K_2 \ar[d]^-{\svech{1_{K_1}}{0}} \ar@{-->}[r] &\\ D_1 \ar[r] \ar[d]_{\beta_1} \ &C \ar[r]^{t_1} \ar[d] &K_1 \ar[d]^-{\svecv{1_{K_1}}{0}} \ar@{-->}[r] &\\ D \ar[r]^-{s} &C \ar[r]^-{\svecv{t_1}{t_2}} &K_1\oplus K_2 \ar@{-->}[r] & } $$ If $\beta_1\alpha_1$ is invertible, then $D_1\simeq D\oplus D_0$. By (a) we have $D_1\simeq X_1\oplus C_0$ where $X$ is indecomposable and $C_0\in \mathcal C$, hence $D$ is a direct summand of $X_1\oplus C_0$, which implies $X_1\simeq D$. Then $D_1\simeq D\oplus C_0$. Now we have the following commutative diagram. $$\xymatrix{ D \oplus C_0 \ar[r]^-{\left(\begin{smallmatrix} s&0\\ 0&1_{C_0} \end{smallmatrix}\right)} \ar[d]_{\gamma_1}^{\simeq} &C\oplus C_0 \ar[r]^-{\left(\begin{smallmatrix} t_1&0\\ t_2&0 \end{smallmatrix}\right)} \ar[d] &K_1\oplus K_2 \ar[d]^-{\eta_1} \ar@{-->}[r] &\\ D_1 \ar[r] \ar[d]_{\gamma_2}^{\simeq} \ &C \ar[r]^{t_1} \ar[d] &K_1 \ar[d]^-{\eta_2} \ar@{-->}[r] &\\ D \oplus C_0 \ar[r]_-{\left(\begin{smallmatrix} s&0\\ 0&1_{C_0} \end{smallmatrix}\right)} &C\oplus C_0 \ar[r]_-{\left(\begin{smallmatrix} t_1&0\\ t_2&0 \end{smallmatrix}\right)} &K_1\oplus K_2 \ar@{-->}[r] & } $$ where $\gamma_2\gamma_1=1_{D\oplus C_0}$, then $1_K-\eta_2\eta_1$ factors through $C\oplus C_0$, hence $K$ is a direct summand of $K_1\oplus C\oplus C_0$. Since $K\in \widetilde{\mathcal K}$, we have $K=K_1$, hence it is indecomposable. If $\beta_1\alpha_1$ is not invertible, since $\operatorname{End}\nolimits_{\mathcal B}(D)$ is local, there is a natural number $n$ such that $(\beta_1\alpha_1)^n=0$, this implies $${0=\left(\begin{smallmatrix} 1_{K_1}&0\\ 0&0 \end{smallmatrix}\right)}^n\svecv{t_1}{t_2}={\left(\begin{smallmatrix} 1_{K_1}&0\\ 0&0 \end{smallmatrix}\right)}\svecv{t_1}{t_2}=\svecv{t_1}{0},$$ hence $t_1=0$, a contradiction. \end{proof} The following lemma plays an important role in the rest of the paper. \begin{lem}\label{mainlem} If $\mathcal B/\mathcal V$ is abelian and $\widetilde{\mathcal U}\neq 0$, then any indecomposable object $K\in \widetilde{\mathcal U}$ admits the following two $\mathbb E$-triangles: \begin{itemize} \item[(a)] $\xymatrix{K' \ar[r] &P \ar[r] &K \ar@{-->}[r] &}$, \item[(b)] $\xymatrix{K \ar[r] &I \ar[r] &K'' \ar@{-->}[r] &}$ \end{itemize} where $P,I \in \mathcal P \cap \mathcal I$ and $K'$ and $K''\in \widetilde{\mathcal U}$ are indecomposables. \end{lem} \begin{proof} Let $A\in \widetilde{\mathcal U}$ be an indecomposable object, it admits an $\mathbb E$-triangle $\xymatrix{K_A \ar[r] &C_A \ar[r] &A \ar@{-->}[r] &}$ where $C_A\in \mathcal C$ and $K_A\in \mathcal U$, by assumption we get $K_A\notin \mathcal C$. Let $K\in \widetilde{\mathcal U}$ be an indecomposable direct summand of $K_A$, let $K_A=K\oplus K_1$. Then we have the following commutative diagram. $$\xymatrix{ K_A \ar@{=}[d] \ar[r] &C_A \ar[d] \ar[r] &A \ar[d]^{a_1} \ar@{-->}[r] &\\ K\oplus K_1 \ar[d]_{\svech{1}{0}} \ar[r] &I_{K_A} \ar[d] \ar[r] &\Sigma K_A \ar[d]^{a_2} \ar@{-->}[r] &\\ K \ar[r] & I_K \ar[r] &\Sigma K \ar@{-->}[r] & } $$ where $\beta$ is split epimorphism and $I_{K_A},I_K\in \mathcal I$. By Lemma \ref{app}, $K$ admits an $\mathbb E$-triangle $$\xymatrix{ K'\oplus C_1' \ar[r] &C'\ar[r] &K \ar@{-->}[r] &} $$ where $C',C_1'\in \mathcal C$ and $K'\in \widetilde{\mathcal U}$ is indecomposable. By the dual of Lemma \ref{app}, $K'$ admits an $\mathbb E$-triangle $\xymatrix{K' \ar[r] &I_{K'} \ar[r] &X\oplus I' \ar@{-->}[r] &}$ where $X$ is indecomposable and $I_{K'},I'\in \mathcal I$. Then X admits an $\mathbb E$-triangle $\xymatrix{K_X \ar[r] &I_{K'} \ar[r] &X \ar@{-->}[r] &}$. Hence we have the following commutative diagram $$\xymatrix{ K_X \ar[r] \ar[d] &I_{K'} \ar[r] \ar[d] &X \ar[d]^{\svecv{1}{0}} \ar@{-->}[r] &\\ K' \ar[r] \ar[d] &I_{K'} \ar[r] \ar@{=}[d] &X\oplus I' \ar[d]^{\svech{1}{0}} \ar@{-->}[r] &\\ K_X \ar[r] &I_{K'} \ar[r]^{j_X} &X \ar@{-->}[r] & } $$ Now $K_X$ is a direct summand of $K'\oplus I_{C'}$, hence $K_X\in \mathcal U$. Since $C_1'$ also admits an $\mathbb E$-triangle $\xymatrix{C_1' \ar[r] &I_1'\ar[r] &\Sigma C_1'\ar@{-->}[r] &}$ where $I_1'\in \mathcal I$, let $I'\oplus \Sigma C_1'=D$,then we get the following commutative diagram $$\xymatrix{ K'\oplus C_1' \ar@{=}[d] \ar[r] &C'\ar[r] \ar[d] &K \ar[d]^{\svecv{k_X}{k_2}} \ar@{-->}[r] &\\ K'\oplus C_1' \ar[r] &I_{K'}\oplus I_1' \ar[r] &X\oplus D\ar@{-->}[r] & } $$ where morphism $k_X$ does not factors through $\mathcal I$, otherwise $\svecv{k_1}{k_2}$ factors through $\mathcal I$ and then $K'$ is a direct summand of $C'$, a contradiction. Then we have the following commutative diagram $$\xymatrix{ K_A \ar@{=}[d] \ar[r] &C_A \ar[d] \ar[r] &A \ar[d]^{a_1} \ar@{-->}[r]^-{\delta} &\\ K\oplus K_1 \ar[d]_{\svech{1}{0}} \ar[r] &I_{K_A} \ar[d] \ar[r] &\Sigma K_A \ar[d]^{a_2} \ar@{-->}[r] &\\ K \ar[r] \ar[d]_{k_X} & I_K \ar[r] \ar[d] &\Sigma K \ar[d]^{a_3} \ar@{-->}[r] &\\ X \ar[r] &I_X \ar[r]^{j} &\Sigma X \ar@{-->}[r]^-{\sigma} & } $$ where $I_X\in \mathcal I$. Let $a=a_3a_2a_1$, we claim that $a$ does not factors through $j$. Otherwise, $a$ factors through $j$, then $(k_X,0)$ factors through $C_A$. Hence it factors through $j_X$, this means $k_1$ factors through $\mathcal I$, a contradiction.\\ Since $a^*\sigma={k_x}_*\delta$, we have the following commutative diagram $$\xymatrix{ K_A \ar[d]_{k_X} \ar[r] &C_A \ar[r]^{c_A} \ar[d]_c &A \ar@{=}[d] \ar@{-->}[r]^-{\delta} &\\ X \ar[r]^g \ar@{=}[d] &B \ar[d] \ar[r]^f &A \ar[d]^a \ar@{-->}[r]^-{a^*\sigma} & \\ X \ar[r] &I_X \ar[r]_{j} &\Sigma X \ar@{-->}[r]^-{\sigma} & } $$ (I) We show $B\in \mathcal C$.\\ Let $\overline f: B\xrightarrow{\overline l} D\xrightarrow{\overline k} A$ be an epic-monic factorization of $\overline f$ in $\mathcal B/\mathcal V$, then we have $f-kl:B\xrightarrow{v_1} V_A\xrightarrow{v_2} A$ where $V_A\in \mathcal V$. Since $\mathcal E(V_A,K_A)=0$, we have $v_2:V_A\xrightarrow{v_3} C_A\xrightarrow{c_A}$. Hence $f-lk=v_2v_1=c_Av_3v_1$. By Lemma \ref{serre}, $D\in \mathcal U$. Then by Corollary \ref{zero-iso}, $D\in \mathcal C$ or $D\simeq A$. If $D\simeq A$, the $\overline f$ is an epimorphism. If $\overline f$ is also a monomorphism, it is an isomorphism since $\mathcal B/\mathcal V$ is abelian. Since $A\in \widetilde{\mathcal U}$ is indecomposable, we get $f$ is a split epimorphism, which implies $a$ factors through $j$, a contradiction. Let $\overline r:R\to B$ be the kernel of $\overline f$, since $\overline {fg}=0$, there is a morphism $s:X\to R$ such that $\overline {rs}=\overline g$. On the other hand, we have the following commutative diagram $$\xymatrix{ &R \ar[r]^{c_1} \ar[d]_r &V \ar[d]^{c_2}\\ X \ar[r]_g &B \ar[r]_f &A \ar@{-->}[r] & } $$ where $V\in \mathcal V$. Since $\mathbb E(V,K_A)=0$, we have $c_2:V\xrightarrow{c_3} C_A \xrightarrow{c_A} A$. Hence $f(r-cc_3c_1)=0$, then there is a morphism $t:R\to X$ such that $\overline {gt}=\overline r$. Hence $\overline {rst}=\overline {gt}=\overline {r}$, which implies $\overline {st}=\overline {1_{R}}$. Since $X$ is indecomposable, $\overline s$ is an isomorphism, hence $\overline g$ is a monomorphism. This implies $\overline {k_X}=0$, $k_X$ factors through $\mathcal V$, then it factors through $C_A$. Hence $A$ is a direct summand of $B$, which implies $a$ factors through $j$, a contradiction. If $D\in \mathcal C$, then we have $k:D\xrightarrow{d} C_A\xrightarrow{c_A} A$, hence we have a morphism $cd:D\to B$ such that $fcd=k$. This implies $f(1_B-cdl-cv_3v_1)=0$, hence there is a morphism $m:B\to X$ such that $gm=1_B-cdl-cv_3v_1$. Hence $\overline {1_B}=\overline {gm}$ and $\overline g$ is a split epimorphism. Since $X$ is indecomposable, $\overline g:X\xrightarrow{\simeq} B$ in $\mathcal B/\mathcal V$ if $B$ does not belong to $\mathcal V$. This means $g$ is a split monomorphism. Then we still have $A$ is direct summand of $B$ and then $a$ factors through $j$, a contradiction. Hence $B\in \mathcal V$. From the following commutative diagram $$\xymatrix{ K_A \ar[d]_{k_X} \ar[r] &C_A \ar[r]^{c_A} \ar[d] &A \ar@{=}[d] \ar@{-->}[r] &\\ X \ar[r]^g \ar[d] &B \ar[d] \ar[r]^f &A \ar@{=}[d] \ar@{-->}[r] & \\ K_A \ar[r] &C_A \ar[r]_{c_A} &A \ar@{-->}[r] & } $$ we get $K_A$ is a direct summand of $X\oplus C_A$, since $K_A\notin \mathcal C$, $X$ is direct summand of $K_A$, hence $X\in \mathcal U$ and $B\in \mathcal U\cap \mathcal V=\mathcal C$.\\[2mm] (II) We show $B\in \mathcal P$. If $f$ factors through an object $P\in \mathcal P$, then by Remark \ref{useful}, $B$ is a direct summand of $X\oplus P$, since $X\notin \mathcal C$, $B$ is a direct summand of $P$, hence $B\in \mathcal P$. We will show $f$ factors through $\mathcal P$. For an object $B$, there exists an $\mathbb E$-triangle $\xymatrix{\Omega B \ar[r] &P_B \ar[r] &B \ar@{-->}[r]^-{\delta_B} &}$ where $P_B\in \mathcal P$, then we have the following commutative diagram $$\xymatrix{ &\Omega B \ar@{=}[r] \ar[d]_{y_1} &\Omega B \ar@{=}[r] \ar[d]^{f'} &\Omega B \ar[d]\\ K_X \ar[r]^y \ar@{=}[d] &Y \ar[r]^{y_2} \ar[d]_{y_3} &\Omega A \ar[r]^p \ar[d]^{h'} &P_B \ar[d] \ar[r] &A \ar@{=}[d] \ar@{-->}[r]^-{\delta_A} &\\ K_X \ar[r]^{i_X} &I_{C'} \ar[r] &X \ar[r] &B \ar@{-->}[d]^-{\delta_B} \ar[r]^f &A\\ &&&&& } $$ where $f'_*\delta_B=f^*\delta_A$. Hence we have the following diagram $$\xymatrix{ \Omega B \ar[r]^{pf'} \ar[d]_{f'} &P_B \ar[d] \ar[r] &B \ar[d]^f \ar@{-->}[r]^-{\delta_B} &\\ \Omega A \ar[r] &P_B \ar[r] &A \ar@{-->}[r]^-{\delta_A} & } $$ Moreover, $\overline {y_1}$ is an epimorphism and by Lemma \ref{app}, $\Omega A$ is indecomposable in $\mathcal B/\mathcal V$. If $\overline {f'}=0$, then $f'$ factors through $pf'$, hence $f$ factors though $P_B$. If $\overline {f'}\neq 0$, then $\overline {y_2}\neq 0$. Consider its epic-monic factorization $Y \xrightarrow{\overline e} E\xrightarrow{\overline m} \Omega A$. We have $\overline {ey}=0$, then there is an object $V_Y\in \mathcal V$ such that $ey:K_X\xrightarrow{s_1} V_Y \xrightarrow{s_2} E$. Since $X\in \mathcal U$, $\mathbb E(X,V_Y)=0$, then there is a morphism $i:I_{C'}\to V_Y$ such that $s_1=ii_X$, we can replace $e$ by $e-s_2iy_3=e'$. then $e'y=0$, hence there is a morphism $m':\Omega A \to E$ such that $m'y_2=e'$. Since $\overline {m'me}=\overline {e}$, we have $\overline {m'm}=\overline {1_E}$. Since the endomorphism (in $\mathcal B/\mathcal V$) algebra of $\Omega A$ is local, then either $\overline {mm'}$ or $\overline{1_{\Omega A}}-\overline {mm'}$ is invertible. In the first case $\overline m$ is invertible, then $\overline {f'}$ is epic, hence $\overline {h'}=0$. This implies $h'$ factor through $P_B$, then A is a direct summand of $B$, a contradiction. In the second case, we have $(\overline {1_{\Omega A}}-\overline {mm'})\overline{f'}=0$, hence $\overline{f'}=0$, a contradiction. This implies $f$ factors through $\mathcal P$. Now we prove that an indecomposable object $K\in \widetilde{\mathcal U}$ admits an $\mathbb E$-triangles $\xymatrix{K' \ar[r] &P \ar[r] &K \ar@{-->}[r] &}$ where $P\in \mathcal P$ and $K'\in \widetilde{\mathcal U}$. Dually we can show it admits an $\mathbb E$-triangle $\xymatrix{K \ar[r] &I \ar[r] &K'' \ar@{-->}[r] &}$ where $I\in \mathcal I$ and $K'' \in \widetilde{\mathcal U}$. Now by this argument $K'$ also admits an $\mathbb E$-triangles $\xymatrix{K' \ar[r] &I' \ar[r] &K''' \ar@{-->}[r] &}$ where $I'\in \mathcal I$ and $K''' \in \widetilde{\mathcal U}$. Then we have the following commutative diagram $$\xymatrix{ K' \ar[r] \ar@{=}[d] &P \ar[d] \ar[r] &K \ar[d] \ar@{-->}[r] &\\ K' \ar[r] \ar@{=}[d] &I' \ar[r] \ar[d] &K''' \ar[d] \ar@{-->}[r] &\\ K' \ar[r] &P \ar[r] &K \ar@{-->}[r] & } $$ Then $P$ is a direct summand of $K'\oplus I'$, since $K\in \widetilde{\mathcal U}$, $P$ is a direct summand of $I'$. Hence $P\in \mathcal P\cap \mathcal I$. Dually we have $I\in \mathcal P\cap \mathcal I$. \end{proof} We can prove the same results for $\mathcal V$, hence we have the following corollary. \begin{cor}\label{maincor} If $\mathcal B/\mathcal V$ and $\mathcal B/\mathcal U$ are abelian and $\widetilde{\mathcal K}\neq 0$, then any indecomposable object $K\in \widetilde{\mathcal K}$ admits the following two $\mathbb E$-triangles: \begin{itemize} \item[(a)] $\xymatrix{K' \ar[r] &P \ar[r] &K \ar@{-->}[r] &}$, \item[(b)] $\xymatrix{K \ar[r] &I \ar[r] &K'' \ar@{-->}[r] &}$ \end{itemize} where $P,I \in \mathcal P \cap \mathcal I$ and $K',K''\in \widetilde{\mathcal K}$ are indecomposables. \end{cor} Let $\widehat{\mathcal K}$ be the subcategory of objects which are direct sums of objects in $\widetilde{\mathcal K}$ and $\mathcal P\cap \mathcal I$. Let $\mathcal H_{\mathcal C}=\operatorname{CoCone}\nolimits(\mathcal C,\mathcal C)$, note that $\mathcal H_{\mathcal C}/\mathcal C$ is the heart of cotorsion pair $(\mathcal C,\mathcal K)$. Then we have the following corollary. \begin{lem}\label{maincor1} If any indecomposable object $K\in \widetilde{\mathcal K}$ admits the following two $\mathbb E$-triangles: \begin{itemize} \item[(a)] $\xymatrix{K' \ar[r] &P \ar[r] &K \ar@{-->}[r] &}$, \item[(b)] $\xymatrix{K \ar[r] &I \ar[r] &K'' \ar@{-->}[r] &}$ \end{itemize} where $P,I \in \mathcal P \cap \mathcal I$ and $K',K''\in \widetilde{\mathcal K}$ are indecomposables, then $\mathcal B/(\mathcal P\cap \mathcal I)=\widehat{\mathcal K}/(\mathcal P\cap \mathcal I)\oplus \mathcal H_{\mathcal C}/(\mathcal P\cap \mathcal I)$. \end{lem} \begin{proof} Let $A,K$ be indecomposable objects where $A\in \mathcal H$ and $K\in \widetilde{\mathcal K}$, then $A$ admits an $\mathbb E$-triangle $\xymatrix{A \ar[r]^c &C^1 \ar[r] &C^2 \ar@{-->}[r] &}$ where $C^1,C^2\in \mathcal C$. $K$ admits $\mathbb E$-triangles $\xymatrix{K' \ar[r] &P \ar[r]^p &K \ar@{-->}[r] &}$ and $\xymatrix{K \ar[r]^i &I \ar[r] &K'' \ar@{-->}[r] &}$ where $P,I\in \mathcal P\cap \mathcal I$, $K,K''\in \widetilde{\mathcal K}$. Let $f:A\to K$ be a morphism, then there is a morphism $r:C^1\to K$ such that $f=rc$, $c$ factors through $p$, hence $\operatorname{Hom}\nolimits_{\mathcal B/(\mathcal P\cap \mathcal I)}(\mathcal H_{\mathcal C},\widehat{\mathcal K})=0$. Let $g:K\to A$ be a morphism, since we have the following commutative diagram $$\xymatrix{ \Omega C^1 \ar@{=}[r] \ar[d] &\Omega C^1 \ar[d]\\ \Omega C^2 \ar[d] \ar[r] &P \ar[r]^p \ar[d] &C^2 \ar@{=}[d] \ar@{-->}[r]&\\ A \ar[r]^c \ar@{-->}[d] &C^1 \ar[r] \ar@{-->}[d] &C^2 \ar@{-->}[r] &\\ && && } $$ Then we have the following commutative diagram $$\xymatrix{ K \ar[ddr]_g \ar@{.>}[dr]^3 \ar[rrr]^-i &&& I \ar@{.>}[ddl]^(.25)1 \ar@{.>}[dl]_2 \ar[r] &K'' \ar@{.>}[d]_{4=c'}\\ &\Omega C^2 \ar[d] \ar[r] &P \ar[d] \ar[rr]_p &&C^2\\ &A \ar[r]^c &C^1 \\ } $$ The numbers for the morphisms denote the order that we get them. Since $K''$ also admits an $\mathbb E$-triangle $$\xymatrix{K'' \ar[r]^{i''} &I'' \ar[r] &K''' \ar@{-->}[r] &}$$ where $I''\in \mathcal P\cap \mathcal I$ and $K'''\in \widetilde{\mathcal K}$, we get $c'$ factors through $I''$, then $c'$ factors through $p$, hence morphism $3$ factors through $I$, then $g$ factors through $I$. This means $\operatorname{Hom}\nolimits_{\mathcal B/(\mathcal P\cap \mathcal I)}(\widehat{\mathcal K},\mathcal H_{\mathcal C})=0$. Assume that $X$ is an indecomposable object, it admits an $\mathbb E$-triangle $$\xymatrix{X \ar[r] &K^X\oplus C^0 \ar[r] &C^X \ar@{-->}[r] &}$$ where $K^X\in \widetilde{\mathcal K}$ and $C^0,C^X\in \mathcal C$. By Lemma \ref{mainlem}, $K^X$ admits an $\mathbb E$-triangle $$\xymatrix{K_0 \ar[r] &P_0 \ar[r] &K^X \ar@{-->}[r] &}$$ where $K_0\in \widetilde{\mathcal K}$ and $P_0\in \mathcal P$. Hence we have the following commutative diagram $$\xymatrix{ K_0 \ar[d]_{k_0} \ar@{=}[r] &K_0 \ar[d]\\ Y \ar[d] \ar[r] &P_0\oplus C^0 \ar[r] \ar[d] &C^X \ar@{=}[d] \ar@{-->}[r] &\\ X \ar[r] \ar@{-->}[d] &K^X\oplus C^0 \ar[r] \ar@{-->}[d] &C^X \ar@{-->}[r] &\\ && && } $$ where $Y\in \mathcal H_\mathcal C$, we have $k_0: K_0\xrightarrow{j_1} I \xrightarrow{j_2} Y$ where $I \in \mathcal P\cap \mathcal I$, then $j_1$ factors through $k_0$, hence $Y$ is a direct summand of $X\oplus I$. If $Y$ is a direct summand of $I$, we have $X\in \mathcal K$. Otherwise $Y=X\oplus I'$. which implies $X\in \mathcal H_{\mathcal C}$ by Lemma \ref{direct summand}. \end{proof} \begin{lem}\label{maincor2} If $\mathcal B/(\mathcal P\cap \mathcal I)=\widehat{\mathcal K}/(\mathcal P\cap \mathcal I)\oplus \mathcal H_\mathcal C/(\mathcal P\cap \mathcal I)$, then $\widehat{\mathcal K}/(\mathcal P\cap \mathcal I),\mathcal H_{\mathcal C}/(\mathcal P\cap \mathcal I)$ are extriangulated subcategories of $\mathcal B/(\mathcal P\cap \mathcal I)$. \end{lem} \begin{proof} It is enough to show that $\widehat{\mathcal K}$ and $\mathcal H_\mathcal C$ are extension closed. Let $\xymatrix{X\ar[r]^{x} &Y\ar[r]^{y} &Z \ar@{-->}[r] &}$ be an $\mathbb E$-triangle where $X,Z\in \widehat{\mathcal K}$. We already have $Y\in \mathcal K$. If $Y$ has a indecomposable direct summand $C \in \mathcal C$. Since $\operatorname{Hom}\nolimits_{\mathcal B/(\mathcal P\cap \mathcal I)}(C,Z)=0$, then $C$ is a direct summand of some $X\oplus I$ where $I\in \mathcal P\cap \mathcal I$. Hence $C\in \mathcal P\cap \mathcal I$. Now let $\xymatrix{X\ar[r]^{x} &Y\ar[r]^{y} &Z \ar@{-->}[r] &}$ be an $\mathbb E$-triangle where $X,Z\in \mathcal H_\mathcal C$. We have $Y=Y_1\oplus Y_2$ where $Y_1\in \mathcal H_\mathcal C$ and $Y_2\in \widetilde{\mathcal K}$. Since $\operatorname{Hom}\nolimits_{\mathcal B/(\mathcal P\cap \mathcal I)}(Y_2,Z)=0$, $Y_2$ is a direct summand of some $X\oplus I'$ where $I'\in \mathcal P\cap \mathcal I$. Hence $Y_2=0$ and $Y\in \mathcal H_\mathcal C$. \end{proof} \begin{prop}\label{mainprop} Let $\mathcal C$ be a pre-cluster tilting subcategory of $\mathcal B$, then the following statements are equivalent. \begin{itemize} \item[(a)] $\mathcal C$ is the category of projective-injective objects of $\mathcal B$. \item[(b)] $\mathcal K/C=\mathcal B/\mathcal C$. \item[(c)] $\mathcal H_{\mathcal C}=\mathcal C$. \end{itemize} \end{prop} \begin{proof} (b) $\Longrightarrow$ (c): $\mathcal H_{\mathcal C}\subseteq \mathcal K$, and $\mathcal H_{\mathcal C}\cap \mathcal K=\mathcal C$, hence $\mathcal H_{\mathcal C}=\mathcal C$. (c) $\Longrightarrow$ (a): For any object $B$ we have $H(B)\in \mathcal H_{\mathcal C}/\mathcal C=0$, hence $B\in \mathcal K$. We know that $\mathcal C$ is the category of projective-injective objects of $\mathcal K=\mathcal B$. (a) $\Longrightarrow$ (b): By definition $(\mathcal C,\mathcal B)$ is a cotorsion pair, hence $\mathcal B=\mathcal K$ and then $\mathcal K/\mathcal C=\mathcal B/\mathcal C$. \end{proof} \textbf{Now we can prove Theorem \ref{main}.} \begin{proof} By Theorem \ref{main1}, $(\mathcal C,\mathcal K),(\mathcal K,\mathcal C)$ are cotorsion pairs. (a) $\Longrightarrow$ (b): when $\mathcal C$ is cluster tilting, we have $\mathcal C\subseteq \mathcal U \subseteq \mathcal K=\mathcal C$, hence $\mathcal C=\mathcal U$. (b) $\Longrightarrow$ (a): we have cotorsion pairs $(\mathcal K,\mathcal C),(\mathcal C,\mathcal K)$, hence $\mathcal K=\mathcal C$, which implies $\mathcal C$ is a cluster tilting subcategory of $\mathcal B$. (a) $\Longrightarrow$ (d): Since $\mathcal C$ is cluster tilting, then $\mathcal K=\mathcal C$ and $\mathcal U=\mathcal C=\mathcal V$, by \cite[Theorem 3.2]{LN}, we know that $\mathcal B/\mathcal U=\mathcal B/\mathcal V=\mathcal B/\mathcal C$ are abelian. (d) $\Longrightarrow$ (a): Assume $\widetilde{\mathcal K}\neq 0$, then by Lemma \ref{maincor2}, $\mathcal H/(\mathcal P\cap \mathcal I)$ is an extriangulated subcategory of $\mathcal B/(\mathcal P\cap \mathcal I)$. Since $\mathcal C/(\mathcal P\cap \mathcal I)\neq 0$, by Proposition \ref{mainprop}, $\mathcal H/(\mathcal P\cap \mathcal I)$ is non-zero. Now by Lemma \ref{mainlem} and Lemma \ref{maincor2}, $\mathcal B/(\mathcal P\cap \mathcal I)$ is a direct sum of two non-zero extriangulated subcategories, a contradiction. (a) $\Longrightarrow$ (c) is by \cite[Theorem 3.2]{LN}. (c) $\Longrightarrow$ (a): let $\mathcal V=\mathcal C$, the other arguments are the same as ``(d) $\Longrightarrow$ (a)". \end{proof} This theorem immediately yields the following important conclusion. \begin{cor} Let $\mathcal B/(\mathcal P\cap \mathcal I)$ be connected and $\mathcal C$ a pre-cluster tilting subcategory of $\mathcal B$. Then $\mathcal C$ is cluster tilting if and only if $\mathcal B/\mathcal C$ is abelian. \end{cor} \begin{proof} This follows that Proposition \ref{prop0} and Theorem \ref{main}. \end{proof} Theorem \ref{main} generalizes \cite[Theorem 7.3]{B} for the following reason: If $\mathcal B$ is a triangulated category with shift functor $[1]$, then $\mathcal P=0=\mathcal I$. The category $\mathcal B$ has a Serre functor $\mathbb{S}$, then $\mathcal C^{\bot_1}={^{\bot_1}}(\mathbb{S}[-2]\mathcal C)$. The subcategory $\mathcal C$ of $\mathcal B$ is functorially finite and rigid, by Lemma \ref{CP} we have two cotorsion pairs $(\mathcal C,\mathcal C^{\bot_1})$ and $({^{\bot_1}}\mathcal C,\mathcal C)$. Note that if $\mathbb{S}\mathcal C=\mathcal C[2]$, then $\mathcal C^{\bot_1}={^{\bot_1}}\mathcal C$, which implies $\mathcal C$ is pre-cluster tilting.

9293030b7152699549a804edc0fc612f163faf84

\section{Introduction} The U.S. Census Bureau is the largest agency in the Federal Statistical System. According to the Census Bureau's mission statement, ``The Census Bureau's mission is to serve as the leading source of quality data about the nation's people and economy. We honor privacy, protect confidentiality, share our expertise globally, and conduct our work openly.''\cite{mission} As the 2020 Census approaches, focus turns to the Census Bureau as it deploys differential privacy to protect privacy in the upcoming decennial census. Invented by Dwork et.\ al in 2006, differential privacy provides a mathematical definition for the privacy loss to individuals associated with the publishing of statistics based on their confidential data. Today the differential privacy literature provides numerous mechanisms for privacy preserving data publishing and privacy preserving data mining while limiting the resulting privacy loss to mathematically provable bounds\cite{dpbook}. The 2020 Census data processing system begins by attempting to collect data from all people living in the United States through a variety of means, including an online instrument, a telephone voice-response system, a form that can be mailed in, and ``enumerators'' who travel from house-to-house for non-response follow-up (NRFU)\cite{csac-admin-modeling}. These confidential data will collected and processed to create the Census Unedited File (CUF), which will contain a block-by-block list of every person in the United States. These data must be completed in time to meet the statutory deadline for reapportioning the House of Representatives (December 31, 2020). Subject matter experts working with Census-developed software review the CUF and make corrections based on their expertise and other data sources. The result is the Census Edited File (CEF). The Disclosure Avoidance System (DAS), currently under development, will use a novel differential privacy mechanism to add noise to the CEF, producing the Microdata Detail File (MDF) that the Census Bureau's tabulation system will use to create the traditional data products. In 2008, the Census Bureau deployed OnTheMap, the first production system to use differential privacy\cite{onthemap}. Six years later, Google deployed RAPPOR\cite{Erlingsson:2014:RRA:2660267.2660348}, the second major production system to use differential privacy, in its Chrome web browser. Today, differential privacy is also being used by Apple\cite{apple-dp} and Microsoft\cite{microsoft-dp}. Although these examples all use differential privacy to protect data supplied by individuals, they use it in different ways, for different purposes. The Census Bureau operates as a \emph{trusted curator}, which collects sensitive data from individuals, performs statistical tabulations, and publishes them. Trusted curators use differential privacy to prevent matching between a respondent's identity, their data, and a specific data release, which is the Census Bureau's legal requirement under Section 9 of the Census Act, U.S. Code Title 13. Google, Apple and Microsoft use the \emph{local model} of differential privacy: randomization is performed by software running on the individual's computer. These companies use differential privacy so that they cannot make reliable inferences about specific users. These companies use differential privacy to increase public acceptance of their data collection methods. In 2017, the Census Bureau announced that it would be using differential privacy as the privacy protection mechanism for the 2020 Census of Population of Housing\cite{census-csac-2018}. There is no off-the-shelf mechanism for applying differential privacy to a national census. Although in principle, the Census Bureau could apply Google's RAPPOR mechanism to the raw census returns, any resulting tabulations would contain far too much noise for any sensible value of $\epsilon$ to be of much statistical value. To use the Census Bureau's terminology, the resulting statistics would likely not meet ``fitness for use'' standards, which are also part of the mandate in the Census Act. The same result would ensue if the Census Bureau employed the original Laplace Mechanism\cite{Dwork:2006:CNS:2180286.2180305} to protect its publication tables. An added complication of the Laplace Mechanism is that the tables would not be internally consistent, which might create concerns for data users. Instead, the Census Bureau revealed that it was developing, implementing, testing, and deploying a new differential privacy mechanism. It committed to publishing the mechanism in the peer-reviewed academic literature and making the implementation available to the public, along with suitable test data. Surprisingly, the Census Bureau's experience with OnTheMap did not significantly prepare the organization for the difficulty of deploying differential privacy for the 2020 Census. OnTheMap was a new product that was designed to incorporate modern privacy protection. In comparison, the decennial Census of Population and Housing, first performed under the direction of Thomas Jefferson in 1790, is the oldest and most expensive statistical undertaking of the U.S. government. Transitioning existing data products to differential privacy has revealed both today's limits in the field of formal privacy, and demonstrated the difficulty of retrofitting legacy statistical products to conform with modern privacy practice. \section{Prior Work} Statistical agencies of the U.S. government have traditionally used statistical disclosure limitation techniques\cite{wp22} to protect confidentiality; \citeauthor{lauger2014} details how those techniques were applied to many data products released by the U.S. Census Bureau\cite{lauger2014}. \citeauthor{abowd2016-why} identifies the challenges faced by statistical agencies in reconciling their traditional disclosure limitation practices with the modern realities of database reconstruction\cite{abowd2016-why,abowd2017-how}, which is made possible because of the large number of statistics published by official statistical agencies, the availability of large scale computational resources, and third-party data that can improve the accuracy of the reconstructed database when used in a re-identification attack. \citeauthor{abowd-new} proposes an approach that statistical agencies can use to set $\epsilon$ using economic theory\cite{abowd-new}. \section{Specific Challenges} Here we present some of the challenges that the Census Bureau has encountered during the deployment of differential privacy. We group the challenges into those that arise from current limitations in the mathematics of differential privacy, those resulting from operational complexities within the Census Bureau, and issues faced by the agency's data users. \subsection{Scientific Issues} Differential privacy is less than 15 years old, and most existing mechanisms were created for computer science applications, not the needs of official statistical agencies. \emph{Hierarchical Mechanisms.} For the 2020 Census, the agency desired a mechanism that controlled the error as statistics were reported from smaller geographies (e.g. blocks and block groups) to larger geographies (e.g. census tracts and counties) such that the error would decrease as the population in the relevant geography increased. This required the Census Bureau to develop a set of novel hierarchical mechanisms designed to optimize the accuracy of multiple queries simultaneously. \emph{Invariants.} For the 2018 End-to-End test, the Census Bureau is reporting exact counts for some statistics (e.g. the number of people living on each block) but privatized counts for other statistics (e.g. the number of Hispanics living on each block). The agency has adopted the term \emph{invariants} to describe statistics that are not changed by the application of differential privacy, and views them as restrictions on the universe of neighboring databases. Nevertheless, there is no well-developed theory for how differential privacy operates in the presence of such invariants. In addition, the historical reasons for having invariants may no longer be consistent with the Census Bureau's confidentiality mandate. \emph{Stratified sampling.} Between 1940 and 2000, the Census Bureau used two census forms: a \emph{short-form} sent to the majority of households, and a \emph{long-form} with more questions that was sent to a subset. In 2005 the Bureau replaced the long-form with the American Community Survey, a project that continuously measures the U.S. population using a stratified probability sample. Currently, there is no accepted mechanism for applying differential privacy to the results of such sampling. This has delayed the introduction of formal privacy mechanisms for the American Community Survey. \emph{Quality Metrics.} While the trade-off between between statistical accuracy and privacy loss is at the heart of differential privacy, there are many metrics for assessing the quality of a published dataset. One approach is to calculate the $L_{1}$ error between the true data (i.e. without disclosure limitation) and the privatized data. This is a coarse measure: a disclosure limited product with a high $L_{1}$ compared to the same product without disclosure limitation may still be very accurate for its intended use. Ideally, if intended uses are known, they can be incorporated into the privacy mechanism so that the usefulness is higher for the same privacy-loss budget, allowing the overall privacy-loss budget to be better deployed. \emph{Presenting and Resolving Equity Issues.} Because the Census Bureau intends to publish many tables drawn from the same confidential database and controlled by an overall privacy-loss budget, there is an opportunity to make some tables more accurate at the cost of making other tables less accurate. These can be thought of as issues of fairness between different consumers of the Census data, which can be described as an \emph{equity issue}. In principle, these issues are no different from the decisions that statistical agencies routinely make about allocating a fixed dollar sampling budget among sub-populations in order to obtain estimates that are fit for use on those sub-populations. Differential privacy lacks a well-developed theory for measuring the relative impact of added noise on the utility of different data products, tuning equity trade-offs, and presenting the impact of such decisions. \emph{Establishing a Value of Epsilon.} Before the arrival of differential privacy at the Census Bureau, disclosure avoidance had aspects of the black arts. Knowledge of the actual disclosure avoidance techniques and parameters was restricted to a small group of specialists, and the remainder of the agency treated disclosure avoidance as a black box that input dangerous data and output clean, safe data. The proponents of differential privacy, in contrast, have always maintained that the setting of $\epsilon$ is a policy question, not a technical one. When the Census Bureau announced that it was adopting differential privacy, it also stated that the value of $\epsilon$ would be set by policy makers, not technologists. But how should policy makers do that? Here, the literature of differential privacy is very sparse. To date, the Census Bureau's Data Stewardship Executive Policy committee (DSEP) has set the values of $\epsilon$ for one data product. The value was set by having the practitioner prepare a set of graphs that showed the trade-off between privacy loss ($\epsilon$) and accuracy. The group then picked a value of $\epsilon$ that allowed for sufficient accuracy, then tripled it, so that the the researchers would be able to make several additional releases with the same data set without having to return to DSEP to get additional privacy-loss budget. The value of $\epsilon$ that was given out was far higher than those envisioned by the creators of differential privacy. (In their contemporaneous writings, differential privacy's creators clearly imply that they expected values of $\epsilon$ that were ``much less than one.''\cite{privacy-integrated-queries}). \emph{Mechanism Development.} More efficient mechanisms and proofs with tighter bounds are needed to lower amounts of noise for the same level of privacy loss, and to make efficient use of the privacy-loss budget for iterative releases of edited and corrected statistics. \subsection{Operational Issues} \emph{Obtaining Qualified Personnel and Tools.} An early problem faced by the Census Bureau was not technical, but operational: it lacked subject matter experts skilled in the theory and practice of differential privacy. In part, this is because only a smattering of universities cover the topic of differential privacy in an instructional setting, and then typically only in a single upper-level computer science course. The Census Bureau, in contrast, typically hires graduates with degrees in mathematics, statistics or economics for its ``mathematical statistician'' career tract. And while there is a now a textbook that covers the theory of differential privacy\cite{dpbook}, reading a textbook does not provide the necessary expertise to develop correct differential privacy algorithms and implementations. The sparsity of expertise was noted by the Bipartisan Commission on Evidence-based Policymaking, which strongly recommended the adoption of privacy-enhancing data analysis tools while recognizing that there was a dearth of existing tools \cite{CEP2017}. Likewise, there is a profound lack of toolkits for performing differential privacy calculations and for verifying the correctness of specific implementations. It is now 12 years since the invention of differential privacy: the situation is analogous to the state of Public Key Cryptography in 1989. This has impacted both high-profile projects such as the 2020 Census, as well as the day-to-day work involved in producing more than 100 regular data products and supporting hundreds of researchers in the Federal Statistical Research Data Centers. \emph{Recasting high-sensitivity queries.} The 2010 Census publications included statistics about individuals, statistics about households, and statistics reflecting the interaction of the two. The sensitivity of most counting queries is 1---for example, a statistic that reports the number of males and females on a block, or the number of households. Some queries that combine these kinds of statistics also have a sensitivity of 1, such as the number of households headed by a female. But some queries have a much higher sensitivity. For example, a query asking the number of children in households headed by a female has a sensitivity equal to the largest permissible household size. An added complication is that this value needs to be specified in advance, as part of the overall design of the survey, rather than derived by looking at the data, lest information about the presence of a specific large family in the survey data be revealed. Currently, the DAS team is working with data users to redesign the publication tables, with the hope of lowering their sensitivity. For example, instead of reporting the number of children that are in a household headed by a person who is Hispanic, the Census could report the number of Hispanic children. It could also protect the original query, but at more aggregated levels of geography. \emph{Structural Zeros.} Bringing differential privacy to the 2020 Census required in-depth discussions of the difference between \emph{structural zeros} and \emph{sampling zeros}\cite{bishop-fienberg-holland-1974}. Structural zeros are those enforced by the Census Bureau's edit rules (``there can be no six-year old mothers with 30-year-old children''), while sampling zeros emerge from the data collection effort (``no women over 65 were found living in this facility''). Injected statistical noise can make sampling zeros positive (2 women over 65 are reported living in the facility), but cannot be allowed to undo the edit rules. In practice, the distinction between structural and sampling zeros in an operational context is far less clear. For example, is the number of females in a male prison zero because there are none living there (a sampling zero), or because they are prohibited from living there (a structural zero)? For that matter, how should the Census determine that a facility is single-sex? Previously, whether or not a group quarters was a single-sex might have been determined by looking at the data; this is not permissible in a system that implements differential privacy. \emph{Obtaining a Suitable Computing Environment.} The algorithms being developed for the 2020 Census require significant post-processing following the application of noise. In order to characterize their behavior, Census Bureau researchers will perform many runs on the algorithms with historical data, requiring at least three orders of magnitude greater computing resources than were needed for the 2010 Census. Although the Bureau is migrating from on-premise computing to a cloud-based environment, this migration was delayed because of security concerns, resulting in substantial delays in the development of the 2020 DAS. \emph{Accounting for All Uses of Confidential Data.} A key feature of the previous disclosure avoidance mechanism was that it did not change the values of many tabulations at high levels of geography. Thus, many reports from the 2000 and 2010 Censuses could be produced using the confidential data and without applying further disclosure avoidance. A fundamental requirement of differential privacy is that all calculations involving private data must have noise added before they can be made public. As a result, the Census Bureau has had to identify every use of confidential data in the execution and processing of the 2020 Census. New and unanticipated requirements have emerged during the design of the system after the team thought that the design was locked down. \emph{Lack of Final Specifications.} Beyond those issues arising from the application of differential privacy, the team building the DAS has also faced by the fluid nature of the decennial census. Many of the Census Bureau data products have been traditionally developed near the end of the decade in consultation with the data users. This collaborative process helps ensure the utility of the census data, but it is at odds with the design and development of a differential privacy system, which requires that all computations be known in advance, or that some amount of privacy-loss budget be reserved for future use. \subsection{Issues Faced by Data Users} \emph{Access to Micro-data.} Many Census Bureau data users are accustomed to using micro-data, like those originally released for the 1960 Census, that are either raw or that have undergone only limited confidentiality edits as part of their disclosure avoidance. Unfortunately, record-level data are exceedingly difficult to protect in a way that offers real privacy protection while leaving the data useful for unspecified analytical purposes. At present, the Census Bureau advises research users who require such data to consider restricted-access modalities\cite{restricted-use-microdata}. \emph{Difficulties Arising from Increased Transparency.} Most users of the 2000 and 2010 Censuses were not aware of the details of the disclosure avoidance mechanism nor its impact on their results. With 2020 Census data, users will be aware that noise has been added, and they will be able to calculate the margin of error that the noise introduces. Some data users are confused about this \emph{margin of error}, a term that they traditionally associate with sample surveys. While coverage error has long been openly discussed and analyzed,\cite{g01} discussion of the error caused by disclosure avoidance procedures, historically called ``confidentiality edits,'' has been terse and limited to qualitative statements\cite{sf1.pdf}. \emph{Misunderstandings about Randomness and Noise Infusion.} A key mechanism of differential privacy is adding random noise to tabulated data before releasing. By design, the noise-injection mechanisms used by the Census Bureau will result in increased accuracy as population sizes increase. Explaining this to data users, community leaders and the general public will be critical to the acceptance of this new disclosure avoidance methodology. For example, some statistical programmers want to use repeatable random number generators for regression testing and production, and have the ability to re-run the privacy mechanism if the first set of coins produces results that are deemed unacceptable. Differential privacy is clearly incompatible with this notion. Although there are many technical papers explaining differential privacy, including the Harvard University Privacy Tools Project\cite{tutorial} and the Duke University tutorials\cite{Machanavajjhala:2017:DPW:3035918.3054779}, their academic language is not accessible to many of the Census Bureau's data users. The lack of simplified materials to promote a general understanding of differential privacy increases the likelihood of misunderstanding. \section{Recommendations} Despite the numerous challenges differential privacy adoption faced, it has taken root in the Census Bureau. Here, we present recommendations for furthering its integration into the Census Bureau and overcoming some of the hurdles outlined above. \emph{Repeated Discussions with Decision Makers.} The deployment of differential privacy within the Census Bureau marks a sea change for the way that official statistics are produced and published. But despite the problems encountered, the Census Bureau has not reconsidered its decision to adopt modern disclosure avoidance mechanisms. We believe that this is a result of the Census Bureau's longstanding commitment to confidentiality protections and the adoption of modern methodological techniques. Repeated discussions with both the Census Bureau's governing boards and with data users are vital in assembling and maintaining institutional support for this transformative effort. \emph{Controlled Vocabulary.} The Census Bureau has found it helpful to establish a controlled vocabulary of terms for discussions of matters involving differential privacy. In computing and mathematics, it is common for practitioners to adopt many different words to mean the same thing (and, conversely, to use the same words to mean different things in different contexts). Internal comprehension as well as the ease of communicating with others has been helped by having a controlled vocabulary, enforced from the highest levels of technical management. \emph{Integrated Communications.} The Census Bureau has created a communications team staffed with senior members of several directorates for the purpose of working with data users and the public on promoting understanding of the new privacy initiative. This team plays a pillar role in the acceptance of differential privacy, both internally and externally to the Census Bureau. With a public-facing educational tutorial forthcoming, and a suite of informative media in the works, they are making user-level understanding of differential privacy rapidly more available to non-experts. Finally, the Census Bureau is expanding its educational efforts on the topic of differential privacy. \section{Conclusion} The Census Bureau is now two years into the process of modernizing its disclosure avoidance systems to incorporate formal privacy protection techniques. Although this process has proven to be challenging across disciplines, it promises to reward the efforts of the Census Bureau. In order to attempt privacy protection on the same scale without differential privacy, the Census Bureau could publish dramatically fewer tables and simply hope that they haven't leaked enough information to allow an attacker to perform database reconstruction. By implementing differential privacy, the Census Bureau can mathematically limit the privacy loss associated with each publication. Beyond the 2020 Census, the Census Bureau intends to use differential privacy or related formal privacy systems to protect all of its statistical publications. It is noteworthy that this institution is not only implementing differential privacy in its statistical analyses, but truly integrating it into its organizational structure. With staff in communications, research, statistics, and computer science familiar with and supportive of differential privacy, a set of diverse employees equipped with privacy tools will be available in the Census Bureau beyond the 2020 Census. The methods put in place for the 2018 and 2020 implementations will act as templates, greatly easing its adoption in future statistical projects. With skilled staff and effective methodology in place, differential privacy can make lasting improvements to privacy protection at the federal government's largest statistical agency. \textbf{\small DISCLAIMER: This paper is presented with the hope that its content may be of interest to the general statistical community. The views in these papers are those of the authors, and do not necessarily represent those of the Census Bureau.} \clearpage \bibliographystyle{ACM-Reference-Format} \balance

2df2c322c25ada0e48a4f5d6ebb5e2c24b59f8e0

\section{Introduction} Stochastic processes are used in option pricing models for multiple purposes. A very common purpose is smile interpolation and extrapolation. Given several co-terminal market quotes, the objective here is to produce implied volatilities at a continuum of strike prices or delta levels. A second purpose is to value path-dependent contingent claims such as quantoed forward contracts or barrier options. For both purposes, it is well known that arbitrage is avoided so long as all relative price processes are specified as martingales under the appropriate probability measure. In general, arbitrages can either be model-based or model-free. An example of a model-free arbitrage is a violation of put call parity. An example of a model-based arbitrage is when two European-style futures options have different implied volatilities in the Black model. A martingale specification produces prices that are free of both types of arbitrage. For example, using driftless geometric Brownian motion to describe a futures price under the futures measure $\mathbb{Q}$ leads to both put call parity holding and to equal implied volatilities across strikes and maturity. Suppose that a market maker uses one martingale specification on an initial date and then uses a different martingale specification on a second date. For example suppose that a market maker uses a geometric Brownian martingale with 10\% volatility on the first date and then uses a geometric Brownian martingale with 20\% volatility on the second date. The prices produced on both dates are devoid of model-free arbitrages. For example put call parity will hold on both dates. The prices produced on both dates do produce an arbitrage based on the Black model being correct. For example, if the actual volatility in the Black model is constant at 10\%, then the prices produced on the second day allow model-based arbitrage. If the actual volatility is instead constant at 20\%, then the prices produced on the first day allow model-based arbitrage. If the actual volatility is instead constant at some other value e.g. 15\%, then the prices produced on both days allow arbitrage based on the Black model being correct. However, if the Black model is not describing the risk-neutral dynamics of the underlying, then the market maker's use of time-inconsistent martingale specifications need not produce any model-based arbitrages. Nonetheless, the use of time-inconsistent martingale specifications does produce a set of values that are devoid of model-free arbitrages. When the only goal is to produce values that are devoid of model-free arbitrages, the only challenge to be met is to be consistent with all of the liquid and transparent quotes. For this purpose, time-inconsistent martingale specifications offer greater flexibility than a time-consistent specification. A market maker using the Black model with the same volatility on both dates is unlikely to be able to match the ATM quote on both dates. In contrast, a market maker using the Black model with the ability to change the volatility on the second date is guaranteed to be able to match the ATM quotes on both dates. In contrast this time-inconsistent Black model does not guarantee the ability to match more than one option price on any given date. When two or more simultaneous quotes differ in maturity, and are devoid of model-free arbitrage, one can match them by moving from the constant volatility model to the deterministic volatility Black model. However, when two co-terminal quotes differ in strike and are devoid of model-free arbitrage, one cannot necessarily match them with the deterministic volatility Black model. A different type of martingale specification is required to guarantee a match. In choosing an alternative martingale specification, it is wise to understand the reasons behind the success of the Geometric Brownian Martingale as the benchmark process. Once these reasons are understood, it becomes clearer as to which properties of GBM should be kept and which properties should be jettisoned. For example, at first glance, driftless arithmetic Brownian motion (ABM) appears to be an attractive alternative to driftless GBM due to its simplicity and tractability. However, it is widely agreed that the failure of ABM to preserve the positivity property of GBM makes it unviable as an alternative. It is widely argued that this positivity property of GBM makes it a good first approximation in describing market prices of assets whose owners enjoy limited liability. However, GBM has state space $(0,\infty)$ while prices of limited liability assets occupy $[0,\infty)$. To capture the possibility that the price of a limited liability asset can vanish, one can add a jump to default to a GBM, as done in \cite{m76}. The GBM remains appropriate as a toy model for a stock index, where it is widely agreed that zero is inaccessible. The inaccessibility of the origin for GBM also makes it a good toy model for an exchange rate, since if $X$ is an exchange rate, $\frac{1}{X}$ needs to be well defined. For a driftless GBM, its state space and dynamics are preserved upon inversion of the coordinate and a change of probability measure. In foreign exchange (FX) markets, inverting an FX rate is a natural operation and the change in probability measure corresponds to a change of numeraire. It is highly likely that these invariance properties of GBM explain why this stochastic process plays such a large role in the FX options market. If one wants to address deficiencies of GBM while retaining applicability to FX options pricing, it stands to reason that preserving at least some notion of invariance under inversion is crucial. The purpose of this paper is to propose a process that generalizes GBM while respecting invariance under inversion. Not surprisingly, hyperbolic functions play a large role in our analysis. It is helpful to begin by reviewing some well-known properties of GBM. Consider an arbitrage-free market and let $\mathbb{Q}$ be an equivalent martingale measure. Let $Z$ denote standard Brownian motion on the real line under $\mathbb{Q}$. Consider the process $g_t = e^{\beta Z_t}, t \geq 0$, where $\beta >0$. Clearly, the process $g$ starts at one and stays positive forever. From It\^o's formula: \begin{equation} \frac{dg_t}{g_t} = \frac{\beta^2}{2} dt + \beta dZ_t, \qquad t \geq 0. \label{gbm} \end{equation} We say the process $g$ has constant proportional drift at rate $\frac{\beta^2}{2}$ and constant proportional variance at rate $\beta^2$. The parameter $\beta$ is called the volatility. The process $g$ is called Geometric Brownian motion. To obtain a non-negative martingale from $g$, there are at least three approaches. First, one can change the probability measure from $\mathbb{Q}$ to $\tilde {\mathbb{Q}}$ by setting $\frac{d\tilde {\mathbb{Q}}}{d \mathbb{Q}} = e^{ - \frac{\beta}{2} Z_T - \frac{\beta^2}{4}T}$. Second, one can alternatively change the coordinate by setting $F_t = g_t e^{- \beta^2 t/2}$. Both of these approaches to creating a martingale preserve the strict positivity of $g$. If only non-negativity of the martingale is required, one can alternatively add a jump to default to the $g$ process with arrival rate $\beta^2/2$. In this paper, we propose a positive process which generalizes GBM $g_t = e^{\beta Z_t}, t \geq 0$ by adding an asymmetry parameter $\alpha \geq 0$. For our new process, $\alpha$ describes the instantaneous volatility whenever a new low is reached. while $\beta$ is the instantaneous volatility whenever the process becomes arbitrarily high. Our generalization preserves the positivity, constant proportional drift, and tractability of GBM, while expressing the instantaneous variance rate at any time as a convex combination of $\alpha^2$ and $\beta^2$. The model actually allows a third parameter $\gamma$ which is the initial instantaneous volatility, and hence is required to lie between $\alpha$ and $\beta$. For many options markets, three parameter models are widely used to interpolate and extrapolate implied volatilities across strikes. Intuitively, market participants agree that options markets display nonzero skewness and kurtosis, but there is little discussion about moments higher than the fourth power. Put another way, market participants agree that it is necessary to match some measure of level, slope, and convexity of implied volatility at the money, but there is little discussion about the third or higher derivative of implied volatility. Unfortunately, our particular three parameter model is not as flexible as some other three parameter models e.g. SABR with fixed $\beta$ or $\rho$. As a result, our three parameter model is only suitable for options markets where the implied volatility slice appears to be monotone across strike e.g. SPX or VIX. For non-monotone slices such as when implied volatilities smile, one must alter the model by adding e.g. stochastic volatility. So long as the implied volatility slice appears to be monotone across strike price, our three parameters, $\alpha \geq 0$, $\beta>0$ and $\gamma \geq 0$ have distinct and well-defined roles. The parameter $\alpha$ controls the asymptotic implied volatility at low strikes, while the parameter $\beta$ controls the asymptotic implied volatility at high strikes. The parameter $\gamma$ is used to meet an at-the-money implied volatility. An overview of this paper is as follows. The next section develops a new special function called the two parameter exponential function. The following section first uses this special function to construct a positive contibuous sub-martingale that has a constant drift. Then we introduce a non-negative martingale by adding a jump to default process to the sub-martingale. This martingale has three parameters $\alpha \geq 0$, $\beta > 0$, and $\gamma$ between $\alpha$ and $\beta$. This is followed by derivations of the transition PDF's for the new martingale. The penultimate section presents closed form valuation formulas for contingent claims written on these martingales. In particular, we examine vanilla options, lookback options and barrier options. The final section provides both a summary of the paper and some suggestions for future research. \section{Two Parameter Exponential Function} In this section, we construct a new special function which we call a two parameter exponential function. In the next section, we will use this special function to construct our three parameter martingale. For $\beta >0$, let $y = e^{\beta x}$ be the standard one parameter exponential function. While the function is defined for $\beta \in \mathbb{C}$ and $x \in \mathbb{C}$, we consider it only for $\beta \in \mathbb{R}^{+}$ and $x \in \mathbb{R}^+$. The defining characteristics of $e^{\beta x}$ are that the ratio of the function's slope to its height is constant at $\beta >0$ for all $x \geq 0$ and that the function has unit height at $x=0$ for all $\beta>0$. Accordingly, our two parameter exponential function will have unit height at $x=0$ for all values of its two parameters $\alpha \geq 0$ and $\beta>0$. We will show that the ratio of the function's slope to its height is $\alpha \geq 0$ at $x=0$ and approaches $\beta >0$ as $x \uparrow \infty$. Since infinitely many functions meet just these criteria, we further require that the ratio of the function's curvature to its height be constant at $\beta^2 >0$ for all $x \geq 0$. This property also belongs to the one parameter exponential function and serves to uniquely\footnote{Our special function $f(x)$ solves the ordinary differential equation $f''(x) = \beta^2 f(x)$ on $x \geq 0$ subject to the Dirichlet boundary condition $f(0) = 1$ and the Neumann boundary condition $ f'(0) = \alpha$.} determine our two parameter exponential function. For $x \geq 0$, $\beta>0$, and $\alpha \geq 0$, we define\footnote{Our function can also be expressed as $\cosh(\beta x) + \frac{\alpha}{\beta} \sinh(\beta x), x \geq 0, \alpha \geq 0, \beta>0$ and so its properties will arise as a consequence of such a representation.} the two parameter exponential function by: \begin{equation} e^{\beta x}_{\beta - \alpha} \equiv \frac{\beta + \alpha}{2 \beta} e^{\beta x} + \frac{\beta - \alpha}{2 \beta} e^{-\beta x}. \label{gex} \end{equation} Thus, the subscripted exponential is a linear combination of the ordinary exponential $e^{\beta x}$ and its reciprocal. The $\beta - \alpha$ subscript in $e^{\beta x}_{\beta - \alpha} $ describes the numerator of the fraction multiplying the reciprocal $e^{-\beta x}$. The numerator of the fraction multiplying $e^{\beta x}$ is always the sum of the asymmetry parameter $\alpha$ and the scaling factor $\beta$ in the ordinary exponential $e^{\beta x}$. The common denominator of both fractions is twice this scaling factor $\beta$. These rules uniquely expand the LHS of (\ref{gex}) into the RHS. On our function's domain $x \geq 0$, the ordinary exponential $e^{\beta x}$ in the linear combination is larger than its reciprocal i.e. $e^{\beta x} \geq e^{-\beta x}$. If $\alpha = 0$, the two fractions simplify to one half and the function is increasing and convex. Increasing $\alpha$ increases the fraction multiplying the larger exponential $e^{\beta x}$ and decreases the fraction multiplying the smaller exponential $e^{-\beta x}$, while keeping the value of the function at $x=0$ fixed at one. As a result, increasing $\alpha$ causes our special function to slope up faster at every $x \geq 0$. If $\alpha = \beta$, then the two parameter exponential $e^{\beta x}_0$ reduces to the one parameter exponential $e^{\beta x}$. Thus the subscript $\beta - \alpha$ on $e^{\beta x}_{\beta - \alpha}$ is also a measure of the deviation of our two parameter exponential function from the one parameter exponential function. Like the one parameter exponential function $e^{\beta x}$, the two parameter exponential function $e^{\beta x}_{\beta - \alpha}$ defined by (\ref{gex}) is positive, increasing, and convex in $x$ for all $x \geq 0$ and for all $\beta > 0$. The derivative w.r.t. $x$ of our two parameter exponential function is: \begin{equation} \frac{d}{dx} e^{\beta x}_{\beta - \alpha} = \beta e^{\beta x}_{\alpha - \beta}, \qquad \alpha \geq 0, \beta > 0, x \geq 0, \label{d} \end{equation} where: \begin{equation} e^{\beta x}_{\alpha - \beta} \equiv \frac{\beta + \alpha}{2 \beta} e^{\beta x} + \frac{\alpha - \beta}{2 \beta} e^{-\beta x}, \qquad \alpha \geq 0, \beta > 0, x \geq 0. \label{gexm} \end{equation} At $\alpha = 0$, $e^{\beta x}_{\alpha - \beta}$ is the right arm of the hyperbolic sine and hence positive. Increasing $\alpha$ increases the weight on both exponentials and hence $e^{\beta x}_{\alpha - \beta} >0$ for all $\alpha \geq 0, \beta > 0, x \geq 0$. Since $\beta >0$ as well, (\ref{d}) implies that the derivative $\frac{d}{dx} e^{\beta x}_{\beta - \alpha}$ is positive. Thus, the $x$-derivative of our two parameter exponential function behaves the same way as the $x$-derivative of the ordinary exponential function w.r.t to its scaling factor $\beta$. Differentiating our two parameter exponential function w.r.t. $x$ also switches the sign on the subscript. To convert $e^{\beta x}_{\alpha - \beta}$ on the RHS of (\ref{d}) back into an expression involving its cohort $e^{\beta x}_{\beta - \alpha}$, one can again differentiate w.r.t. $x$. In particular: \begin{equation} \frac{d^2}{dx^2} e^{\beta x}_{\beta - \alpha} = \beta^2 e^{\beta x}_{\beta - \alpha}, \qquad \alpha \geq 0, \beta > 0, x \geq 0. \label{dd} \end{equation} Thus, the ratio of the function's curvature to its height is constant at $\beta^2>0$ for all $x \geq 0$, as previously indicated. There is an alternative way to convert $e^{\beta x}_{\alpha - \beta}$ back into an expression involving its cohort $e^{\beta x}_{\beta - \alpha}$. The appendix shows that: \begin{equation} e^{\beta x}_{\alpha - \beta} = \sqrt{ \left( e^{\beta x}_{\beta - \alpha} \right)^2 + \frac{\alpha^2 - \beta^2}{\beta^2}}. \label{swap} \end{equation} We now use this alternative conversion mechanism to show that our two parameter exponential function sets the ratio of its slope to its height at $\alpha$ at $x=0$. We will also show in contrast that the ratio of its slope to its height approaches $\beta$ as $x \uparrow \infty$. These behaviors define the role of each parameter in our two parameter exponential function. Consider the ratio of the slope of our two parameter exponential function to its height: \begin{equation} \frac{\frac{d}{dx} e^{\beta x}_{\beta - \alpha} }{e^{\beta x}_{\beta - \alpha} } = \beta \frac{ e^{\beta x}_{\alpha - \beta} }{e^{\beta x}_{\beta - \alpha} }, \label{ratio} \end{equation} from (\ref{d}). Using (\ref{swap}) on the RHS of (\ref{ratio}), this ratio can also be represented as: \begin{equation} \frac{\frac{d}{dx} e^{\beta x}_{\beta - \alpha} }{e^{\beta x}_{\beta - \alpha} } = \beta \frac{ \sqrt{ \left( e^{ \beta x}_{\beta - \alpha} \right)^2 + \frac{\alpha^2 - \beta^2}{\beta^2}}}{e^{\beta x}_{\beta - \alpha} } = \beta \sqrt{1 + \frac{\alpha^2 - \beta^2}{\beta^2 \left(e^{ \beta x}_{\beta - \alpha} \right)^2}}. \label{null} \end{equation} Bringing $\beta$ under the square root: \begin{equation} \frac{\frac{d}{dx} e^{\beta x}_{\beta - \alpha} }{e^{\beta x}_{\beta - \alpha} } = \sqrt{\alpha^2 \frac{1}{ \left(e^{ \beta x}_{\beta - \alpha}\right)^2} + \beta^2 \left[ 1 - \frac{1}{ \left(e^{\beta x}_{\beta - \alpha} \right)^2} \right]}. \label{ratio1} \end{equation} Since $1/\left( e^{ \beta x}_{\beta - \alpha} \right)^2 \in (0,1]$, the radicand is a convex combination of $\alpha^2$ and $\beta^2$. At $x =0$, $e^{ \beta x}_{\beta - \alpha} =1$, so $\frac{1}{ \left( e^{ \beta x}_{\beta - \alpha} \right)^2}$ also $= 1$ and the ratio $\frac{\frac{d}{dx} e^{\beta x}_{\beta - \alpha} }{e^{\beta x}_{\beta - \alpha} } =\alpha$. As $x \uparrow \infty$, $e^{ \beta x}_{\beta - \alpha} \uparrow \infty$, so $\frac{1}{ \left(e^{ \beta x}_{\beta - \alpha} \right)^2} \downarrow 0$ and the ratio $\frac{\frac{d}{dx} e^{\beta x}_{\beta - \alpha} }{e^{\beta x}_{\beta - \alpha} }$ converges to $\beta$. Like the one parameter exponential function, our two parameter exponential function has an explicit inverse. To derive it, let: \begin{equation} y = e^{\beta x}_{\beta- \alpha} = \frac{\beta + \alpha}{2 \beta} e^{\beta x} + \frac{\beta - \alpha}{2 \beta} e^{-\beta x}, \qquad x \geq 0, \alpha \geq 0, \beta>0. \label{y} \end{equation} We need to solve for $x$ as a function of $y$. Multiplying (\ref{y}) by $\beta e^{\beta x}$ leads to a quadratic function of $e^{\beta x}$: \begin{equation} \frac{\beta + \alpha}{2} e^{2 \beta x} - \beta y e^{\beta x} + \frac{\beta - \alpha}{2} = 0, \qquad x \geq 0, \alpha \geq 0, \beta>0. \label{q} \end{equation} By the quadratic root formula: \begin{equation} e^{\beta x} = \frac{ \beta y + \sqrt{ \beta^2 y^2 - (\beta^2 - \alpha^2)} }{\beta + \alpha}, \qquad x \geq 0, \alpha \geq 0, \beta>0, \label{qrf} \end{equation} where we have chosen $+$ in $\pm$ since $e^{\beta x}>0$. Solving for $x$: \begin{equation} x = \frac{1}{\beta} \ln \frac{ \beta y + \sqrt{\alpha^2 + \beta^2 (y^2 - 1)} }{\beta + \alpha}, \qquad x \geq 0, \alpha \geq 0, \beta>0. \label{inv0} \end{equation} Hence, for $y \geq 1$, the function on the RHS of (\ref{inv0}) is the explicit inverse of our two parameter exponential function. Notice that from (\ref{qrf}): \begin{equation} \frac{\beta + \alpha}{2 \beta} e^{\beta x} = \frac{y}{2} + \sqrt{ \frac{y^2}{4} - \frac{\beta^2 - \alpha^2} {4 \beta^2} }, \label{qrf1} \end{equation} where we observe from (\ref{y}) that $\frac{\beta^2 - \alpha^2} {4 \beta^2}$ is just the product of the two terms which sum to $y$. Equation (\ref{qrf1}) is an explicit formula that maps $y$ to the first term in the sum (\ref{y}) defining it. When $x=0$ and $\alpha = 0$, this first term has the same size of $\frac{1}{2}$ as the second term, but otherwise, the first term is larger. To obtain an explicit formula that maps $y$ to the smaller term in the sum defining it, notice that multiplying (\ref{y}) by $\beta e^{-\beta x}$ leads to a quadratic function of $e^{-\beta x}$: \begin{equation} \frac{\beta - \alpha}{2} e^{-2 \beta x} - \beta y e^{\beta x} + \frac{\beta + \alpha}{2} = 0. \label{q3} \end{equation} By the quadratic root formula: \begin{equation} e^{- \beta x} = \frac{ \beta y - \sqrt{ \beta^2 y^2 - (\beta^2 - \alpha^2)} }{\beta - \alpha}, \label{qrf3} \end{equation} where now we have chosen $-$ in $\pm$ since $e^{-\beta x} <1$. Hence: \begin{equation} \frac{\beta - \alpha}{2 \beta} e^{- \beta x} = \frac{y}{2} - \sqrt{ \frac{y^2}{4} - \frac{\beta^2 - \alpha^2} {4 \beta^2} }. \label{qrf4} \end{equation} This equation is an explicit formula that maps $y$ to the last smaller term in the sum (\ref{y}) defining it. For the one parameter exponential function $y = e^{\beta x}, x \geq 0, \beta > 0$, adding one to the input variable $x$ causes the output variable $y$ to grow by the factor $e^{\beta}>1$. We say the exponential function turns addition into multiplication. For our two parameter exponential function defined by (\ref{y}), adding one to the input variable $x$ causes the output variable $y$ to grow as follows. First, split $y$ into its larger term involving $e^{\beta x}$ given explicitly by (\ref{qrf1}) and its smaller term involving $e^{-\beta x}$, given explicitly by (\ref{qrf4}). Next, grow the larger term by a factor $e^{\beta} >1$ and shrink the smaller term by a factor $e^{-\beta} \in (0,1)$. Finally, add the two altered terms together to obtain the new value of $y$. We say the two parameter exponential function turns addition into a blend of multiplication and division. \section{ Constructing a 3 Parameter Non-Negative Continuous Martingale} In this section, we use the two parameter exponential function constructed in the last section to define a new three parameter non-negative continuous martingale denoted by $F_t$. Recall that to create a driftless GBM $F^b$ , one first creates an auxiliary positive continuous process $g_t = e^{\beta Z_t}$ with constant positive drift of $\beta^2/2$ and then one corrects for this constant drift by setting $\frac{F^b_t}{F^b_0} = g_t e^{-\beta^2 t/2}$. We will mimic this construction in the next subsection by first constructing an auxiliary positive continuous process $G$ with positive constant drift of $\beta^2/2$. The following subsection then corrects for this constant drift by adding a jump to default process. \subsection{ Constructing a Positive Continuous Process with Constant Drift} Let $0$ be the valuation time and let $Z$ be a standard Brownian motion $Z$ under $\mathbb{Q}$ whose value at $t=0$ is $Z_0 = 0$ as usual. We allow $Z$ to exist prior to time 0. Let $t_0 \leq 0$ and we suppose that $Z$ exists for all $t \geq t_0$. For $t \geq t_0$, let $\underline{Z}_t \equiv \stackrel{\inf}{\scriptstyle s \in [t_0,t]} Z_s$ denote the running minimum of the standard Brownian motion $Z$ under $\mathbb{Q}$. Notice that $Z$'s path monitoring begins at time $t_0 \leq 0$, so $\underline{Z}_0 \leq 0$. For $t \geq t_0$, let $\check{Z}_t \equiv Z_t - \underline{Z}_t$ denote $Z$'s running drawup process. Let: \begin{equation} \check{G}_t = e^{\beta \check{Z}_t}_{\beta - \alpha}, \qquad t \geq t_0, \beta >0, \label{gdu1} \end{equation} be a new stochastic process with state space $[1,\infty)$. Recall that setting $\alpha $ to zero reduces the two parameter exponential $e^{\beta x}_{\beta - \alpha}, x \geq 0, \beta >0$ to the ordinary exponential $e^{\beta x}, x \geq 0, \beta >0$. The GBM $e^{\beta Z_t}$ and the processes $\cosh(\beta Z_t)$, $\cosh(\beta |Z_t|)$, and $\cosh(\beta \check{Z}_t)$ all grow in expectation at the rate $\beta^2/2$. The hyperbolic cosine is a simple average of the increasing exponential $e^{\beta x}, x \geq 0, \beta >0$, and its reciprocal. When the asymmetry parameter $\alpha$ is made positive, this simple average is replaced with an asymmetric average putting more weight on the increasing exponential. The effect on the mean of this skewing is the same as the effect on the mean of the GBM $e^{\beta Z_t}$ if $Z$ behaved asymmetrically just when visiting its minimum $\bar{Z}$. In particular, if $Z$ is interpreted as a the limit of a scaled random walk, then putting greater probability on rising above the minimum raises the mean growth rate of $e^{\beta Z_t}$ above $\beta^2/2$. Let $\hat{Z}$ denote this skewed Brownian motion. The effect on the mean of $e^{\beta \hat{Z}}_t$ of this rarely imposed asymmetry can be removed by multiplying by $e^{\alpha \bar{Z}_t}$. We will similarly remove the effect on the mean of $\check{G}_t $ of replacing $\cosh(\beta \check{Z}_t)$ with $e^{\beta \check{Z}_t}_{\beta - \alpha}, \beta >0, \alpha \geq 0$ by multiplying $\check{G}_t $ by $e^{\alpha \bar{Z}_t}$. We introduce a new parameter $\gamma$ which will be used to determine the value of $\check{G}_t$ at $t=0$. We require that $\gamma$ be between $\alpha$ and $\beta$. For technical reasons, we allow $\gamma = \alpha$, but we do not allow $\gamma = \beta$. This allows us to set: \begin{equation} \check{G}_0 = \sqrt{\frac{\alpha^2 - \beta^2}{\gamma^2 - \beta^2} }. \label{id} \end{equation} The radicand is $\geq 1$ and hence so is $\check{G}_0$. We next use (\ref{inv0}) to set $\check{Z}_0$: \begin{equation} \check{Z}_0 = \frac{1}{\beta} \ln \frac{ \beta \check{G}_0 + \sqrt{\alpha^2 + \beta^2 [\check{G}_0^2 - 1]} }{\beta + \alpha}, \qquad x \geq 0, \alpha \geq 0, \beta>0. \label{inv1} \end{equation} Since $\check{G}_0 \geq 1$, $\check{Z}_0 \geq 0$. At each $t \geq 0$, $\check{G}_t \geq 1$ defined in (\ref{gdu1}) is increasing in its driver $\check{Z}_t \geq 0$. Equation (\ref{inv}) implies that (\ref{gdu1}) can be explicitly inverted: \begin{equation} \check{Z}_t = \frac{1}{\beta} \ln \left( \frac{ \beta \check{G}_t + \sqrt{ \alpha^2 + \beta^2 [\check{G}_t^2-1]} }{\alpha + \beta} \right), \qquad t \geq t_0, \alpha \geq 0, \beta > 0. \label{inv} \end{equation} We next set $\underline{Z}_0 = - \check{Z}_0$ so that $Z_0 \equiv \underline{Z}_0 + \check{Z}_0 = 0$. With $\underline{Z}_0$ determined at some non-positive value, let: \begin{equation} \underline{G}_t = e^{ \alpha \underline{Z}_t} \qquad t \geq t_0, \alpha \geq 0, \label{Gmindef} \end{equation} be a super-martingale with state space $(0,1]$. The process $\underline{G}_t \in (0,1]$ defined in (\ref{Gmindef}) is increasing in its driver $\underline{Z}_t \leq 0$, For $\alpha > 0$, (\ref{Gmin}) can be explicitly inverted: \begin{equation} \underline{Z}_t = \frac{1}{\alpha} \ln \underline{G}_t,\qquad t \geq t_0. \label{Wmin} \end{equation} For $\alpha \geq 0, \beta > 0$ and $\gamma$ between them, let: \begin{equation} G_t = \underline{G}_t \check{G}_t, \qquad t \geq 0, \label{Gdef} \end{equation} be our auxiliary continuous process with state space $(0,\infty)$. We claim that $\underline{G}_t = \stackrel{\inf}{\scriptstyle s \in [t_0,t]} G_s$. In words, we claim that the super-martingale $\underline{G}_t \in (0,1]$ defined in (\ref{Gmindef}) is just the running minimum of the $G$ process defined in (\ref{Gdef}). To see why, note that substituting (\ref{gdu1}) and (\ref{Gmindef}) in (\ref{Gdef}) implies that for $\alpha \geq 0, \beta > 0$ and $\gamma$ between them: \begin{equation} G_t = e^{ \alpha \underline{Z}_t} e^{\beta \check{Z}_t}_{\beta - \alpha}, \qquad t \geq t_0. \label{Gdef1} \end{equation} Since $\underline{Z}$ only declines when $\check{Z}=0$: \begin{equation} \stackrel{\inf}{\scriptstyle s \in [t_0,t]} G_s = e^{ \alpha \underline{Z}_t}, \qquad t \geq t_0,\alpha \geq 0, \beta > 0, \label{Gmin} \end{equation} which matches the defining equation (\ref{Gmindef}) for $\underline{G}_t$. Hence $\underline{G}_t $ is the running minimum of the $G$ process defined in (\ref{Gdef}). Since $\underline{G}_t$ has state space $(0,1]$, $G$ is positive forever. From (\ref{Gdef}): \begin{equation} \check{G}_t = \frac{G_t}{\underline{G}_t}, \qquad t \geq 0, \label{gdu} \end{equation} so $\check{G}$ is the relative drawup process of $G$. Applying It\^{o}'s formula to (\ref{gdu1}), (\ref{d}) implies that: \begin{equation} d\check{G}_t = \beta e^{\beta \check{Z}_t}_{\alpha - \beta} d\check{Z}_t + \frac{\beta^2}{2} e^{\beta \check{Z}_t}_{\beta - \alpha} d \langle \check{Z} \rangle_t, \qquad t \geq t_0. \label{gdu1a} \end{equation} Thus the increments of $\check{G}_t$ depend on the increments of $\check{Z}_t$ and the squared increments of $\check{Z}_t$. Since $\underline{Z}$ is a process of bounded variation, it has zero quadratic variation and hence: \begin{equation} \langle \check{Z} \rangle_t = \langle Z - \underline{Z} \rangle_t = \langle Z \rangle_t = t, \qquad t \geq t_0. \label{qv} \end{equation} Substituting (\ref{swap}) and (\ref{qv}) in (\ref{gdu1a}) implies that the coefficients just depend on $e^{\beta \check{Z}_t}_{\beta - \alpha}$: \begin{equation} d\check{G}_t = \frac{\beta^2}{2} e^{\beta \check{Z}_t}_{\beta - \alpha} dt + \beta \sqrt{ \left( e^{\beta \check{Z}_t}_{\beta - \alpha} \right)^2 + \frac{\alpha^2 - \beta^2}{\beta^2}} d\check{Z}_t, \qquad t \geq t_0. \label{gdu1b} \end{equation} Substituting (\ref{gdu1}) in (\ref{gdu1b}) implies that $\check{G}$ solves the following stochastic differential equation (SDE): \begin{equation} d\check{G}_t = \frac{\beta^2}{2} \check{G}_t dt + \sqrt{\alpha^2 + \beta^2 \left[ \left(\check{G}_t\right)^2 - 1 \right]} d\check{Z}_t, \qquad t \geq t_0. \label{gdu1c} \end{equation} This SDE is univariate since the coefficients for $\check{G}_t$ just depend on $\check{G}_t$. Dividing by $\check{G}_t$ implies: \begin{equation} \frac{d\check{G}_t}{\check{G}_t} = \frac{\beta^2}{2} dt + \sqrt{ \alpha^2 \frac{1}{\check{G}^2_t} + \beta^2 \left[1 - \frac{1}{\check{G}^2_t} \right]} d\check{Z}_t, \qquad t \geq t_0. \label{gdu1d} \end{equation} Hence, $\check{G}$ solves the above simple SDE when the two drivers are $t$ and $\check{Z}$. To determine the coefficients of $Z_t$ and $\underline{Z}$, note that substituting $d\check{Z}_t = dZ_t - d \underline{Z}_t$ in (\ref{gdu1d}) implies: \begin{equation} \frac{d\check{G}_t}{\check{G}_t} = \frac{\beta^2}{2} dt + \sqrt{ \alpha^2 \left( \frac{1}{\check{G}_t} \right)^2 + \beta^2 \left[1 - \frac{1}{\check{G}^2_t} \right]} \left(dZ_t - d \underline{Z}_t\right), \qquad t \geq t_0. \label{it1g} \end{equation} Since $\underline{Z}$ only decreases when $\check{G} = 1$, the net coefficient of $d\underline{Z}$ in (\ref{it1g}) is zero. As a result, $\check{G}$ also solves the following SDE: \begin{equation} \frac{d\check{G}_t}{\check{G}_t} = - \alpha d \underline{Z}_t + \frac{\beta^2}{2} dt + \sqrt{ \alpha^2 \frac{1}{\check{G}^2_t} + \beta^2 \left[ 1 - \frac{1}{\check{G}^2_t} \right] }dZ_t, \qquad t \geq t_0. \label{Gvsde} \end{equation} The coefficient of $dZ_t$ in (\ref{Gvsde}) is the instantaneous lognormal volatility of $\check{G}$, which is a randomly weighted $L^2$ mean of $\alpha$ and $\beta$. This form is clearly just a consequence of (\ref{ratio1}). Since $\left( \frac{1}{\check{G}_t} \right)^2 \in (0,1]$, the instantaneous lognormal variance rate of $\check{G}_t$ is just a convex combination of $\alpha^2$ and $\beta^2$. When $Z$ is at its minimum $\underline{Z}$, $\check{Z} = 0$, and hence $\check{G}=1$. At such times, (\ref{Gvsde}) implies that the instantaneous volatility of $\check{G}$ is $\alpha$. In contrast, as the difference between $Z$ and its minimum $\underline{Z}$ approaches infinity, $\check{G}$ also approaches infinity, and (\ref{Gvsde}) implies that the instantaneous volatility of $\check{G}$ approaches $\beta$. These results clearly follow from the behavior of our two parameter exponential function $e^{\beta x}_{\beta - \alpha}$ at $x=0$ and at $x=\infty$. The dynamics in (\ref{Gvsde}) clearly depend on our first two parameters $\alpha$ and $\beta$, which are the respective instantaneous volatilities of $\check G$ at $\check G$'s extremes of one and infinity. To interpret our third parameter $\gamma$, note that squaring both sides of (\ref{id}) implies that: \begin{equation} \check{G}_0^2 = \frac{\alpha^2 - \beta^2}{\gamma^2 - \beta^2}. \label{id2} \end{equation} Cross multiplying and re-arranging: \begin{equation} \gamma^2 \check{G}_0^2 = \alpha^2 - \beta^2 + \beta^2 (\check{G}_0)^2. \label{id3} \end{equation} Dividing by $\check{G}_0^2$ and taking the square root implies: \begin{equation} \gamma = \sqrt{ \alpha^2 \frac{1}{\check{G}^2_0} + \beta^2 \left[1 - \frac{1}{\check{G}^2_0} \right]}. \label{invol} \end{equation} Comparing (\ref{invol}) to the volatility in (\ref{Gvsde}) evaluated at $t=0$ implies that our third parameter $\gamma$ is just the initial volatility of $\check{G}$. We next determine the dynamics of the $G$ process, which (\ref{Gdef}) defined as the product: \begin{equation} G_t = \underline{G}_t \check{G}_t \qquad t \geq t_0, \label{gdut} \end{equation} for $\alpha \geq 0, \beta > 0$ and $\gamma$ between them. It\^{o}'s formula implies that: \begin{equation} \frac{dG_t}{G_t} = \frac{d\underline{G}_t}{\underline{G}_t} + \frac{d\check{G}_t}{\check{G}_t} = \alpha d \underline{Z}_t + \frac{d\check{G}_t}{\check{G}_t}, \qquad t \geq t_0, \label{it} \end{equation} since $\underline{G}_t = e^{ \alpha \underline{Z}_t}$. Substituting in (\ref{gdu1d}) implies that $G$ solves the following SDE: \begin{equation} \frac{dG_t}{G_t} = \frac{\beta^2}{2} dt + \sqrt{ \alpha^2 \left( \frac{\underline{G}_t}{G_t} \right)^2 + \beta^2 \left[ 1 - \left(\frac{\underline{G}_t}{G_t} \right)^2 \right] }dZ_t, \qquad t \geq t_0, \label{Ssde} \end{equation} since $\frac{1}{\check{G}_t} = \frac{\underline{G}_t}{G_t}$. Like the $\check{G}$ process, the $G$ process has constant proportional drift at rate $\frac{\beta^2}{2}$. Unlike the SDE (\ref{Gvsde}) for $\check{G}$, the SDE (\ref{Ssde}) for $G$ has coefficients that depend on the auxiliary process $\underline{G}$. Since $\left( \frac{\underline{G}_t}{G_t} \right)^2 \in (0,1]$, the lognormal variance rate of $G$ is also a convex combination of $\alpha^2$ and $\beta^2$. When $G_t = \underline{G}_t$, the $G$ process behaves locally like a GBM with constant proportional drift rate $\frac{\beta^2}{2}$ and constant proportional variance rate $\alpha^2$. As $G$ rises above $\underline{G}_t$, the lognormal variance rate moves towards $\beta^2$ and asymptotes to this value in the limit as $G \uparrow \infty$. Substituting $\frac{1}{\check{G}_0} = \frac{\underline{G}_0}{G_0}$ in (\ref{invol}) implies that: \begin{equation} \gamma = \sqrt{ \alpha^2 \left( \frac{\underline{G}_0}{G_0} \right)^2 + \beta^2 \left[1 - \left( \frac{\underline{G}_0}{G_0} \right)^2 \right]}. \label{invol1} \end{equation} Evaluating the coefficient of $dZ_t$ in (\ref{Ssde}) at $t=0$, (\ref{invol1}) implies that the instantaneous lognormal volatility of $G$ is $\gamma$. Hence, the three parameters $\alpha, \gamma$, and $\beta$ can be respectively interpreted as the instantaneous lognormal volatility of $G$ at each new low, at the initial time, and at infinitely high values of $G$. The bivariate transition PDF of the pair $(\underline{Z},\check{Z})$ is known in closed form and is given in \cite{c14}. Since $\underline{G}$ and $\check{G}$ are each just univariate, increasing, explicitly invertible transformations of $\underline{Z}$ and $\check{Z}$ respectively, it follows that the bivariate transition PDF of the pair $(\underline{G},\check{G})$ can easily be obtained in closed form. Recall from (\ref{Gdef1}) that: \begin{equation} G_t = e^{ \alpha \underline{Z}_t} e^{\beta \check{Z}_t}_{\beta - \alpha}, \qquad t \geq t_0. \label{Gdef1r} \end{equation} As $\beta \downarrow 0$, the $G$ process becomes driftless and two parameter exponential function $e^{\beta x}_{\beta - \alpha}$ in (\ref{Gdef1r}) converges to the linear function $1 + \alpha x$. As a result, the process $G$ converges to the martingale $F$ in \cite{c14} in this limit when $F_0 =1$. Setting $\alpha = \beta$ in (\ref{Gdef1r}), the two parameter exponential reduces to the one parameter exponential and hence: \begin{equation} G_t = e^{ \beta \underline{Z}_t} e^{\beta \check{Z}_t} = e^{\beta (\underline{Z}_t + \check{Z}_t)} = e ^{\beta Z_t}, \qquad t \geq t_0. \label{Gdef1r1} \end{equation} Thus, the $G$ process generalizes the exponential of standard Brownian motion, by adding parameters $\alpha$ and $\gamma$. Being a sub-martingale, the $G$ process can be used directly to model spot price (e.g. spot FX rates) and price derivatives written on $G$ in risk neutral measure. For this purpose, we introduce a new sub-martingale process \begin{eqnarray} F_t=F_0G_t, \quad t\geq t_0,\label{submgl} \end{eqnarray} where $F_0>0$ is the initial value of the process. Like $G$, $F$ is positive and has a positive drift. Note that the positivity of the drift of $G$ is not a binding restriction due to the international put-call equivalence \cite{g83}. For instance, if a positive process $S_t$ has a negative drift, one can use it to model the inverse of a process that has a positive drift via $F_t=\frac 1{S_t}$. For derivatives on future price, the underlying security is required to be driven by a martingale in the risk neutral measure for derivative pricing. In the next subsection we introduce a new martingale process from $G$ by adding a jump to default process which has a negative drift. However, one should interpret the sub-martingale Eqn (\ref{submgl}) and the new martingale as dynamics of two different securities, instead of spot and future prices of one security. \subsection{Constructing a Non-Negative Martingale via Jump to Default} For $\alpha \geq 0, \beta >0$, and for $\gamma$ between them, the $G$ process constructed in the last subsection starts at one and has constant positive drift $\frac{\beta^2}{2}$. In this section, we change the starting point to $F_0 >0$ and interpret this positive drift as compensation for a possible jump to zero. This allows us to construct a tractable non-negative martingale $F$ which starts at $F_0$. Let $N_t$ be a standard Poisson process with arrival rate $\frac{\beta^2}{2}$ under $\mathbb{Q}$. For $F_0 >0$, let: \begin{equation} F_t = F_0 G_t \mathbbm{1}_{N_t = 0}, \qquad t \geq t_0. \label{fjtd} \end{equation} be a non-negative process started at $F_0>0$. Then $F$ is a $\mathbb{Q}$ martingale which drifts up at the constant rate $\frac{\beta^2}{2}$ in order to compensate for a possible jump to zero. Once $F$ hits zero, it is absorbed there. Let: \begin{equation} \underline{F}_t = \stackrel{\inf}{\scriptstyle s \in [t_0,t]} F_s, \qquad t \geq t_0 \label{Fbar} \end{equation} be the running minimum of $F$. Let $\tau$ be the exponentially distributed random time at which $F$ jumps to zero. For $t \in [t_0, \tau)$, (\ref{fjtd}) implies: \begin{equation} \underline{F}_t = F_0 \underline{G}_t. \label{unfjtd} \end{equation} Dividing (\ref{unfjtd}) by (\ref{Fbar}) implies that for $t \in [t_0, \tau)$: \begin{equation} \frac{\underline{F}_t}{F_t}= \frac{\underline{G}_t}{G_t}. \label{rf} \end{equation} As a result, the SDE for $F$ is: \begin{equation} dF_t = F_{t-} \left[ \sqrt{ \alpha^2 \left( \frac{\underline{F}_{t-}}{F_{t-}} \right)^2 + \beta^2 \left[ 1 - \left(\frac{\underline{F}_{t-}}{F_{t-}} \right)^2 \right] }dZ_t - \left(d N_t - \frac{\beta^2}{2} dt \right) \right], \qquad t \geq t_0. \label{Fsdejtd} \end{equation} Substituting (\ref{Gdef}) in (\ref{fjtd}) implies that $F_t$ can be related to the contemporaneous values of the pair $(\underline{Z},\check{Z})$ and $N_t$: \begin{equation} F_t = F_0 e^{ \alpha \underline{Z}_t } e^{\beta \check{Z}_t}_{\beta - \alpha} \mathbbm{1}_{N_t = 0}, \qquad t \geq t_0. \label{new1f} \end{equation} The price relative $\frac{F_t}{F_0}$ is a non-negative martingale started at one. From (\ref{new1f}), this price relative decomposes into the product of a positive strict supermartingale started at one, $e^{ \alpha \underline{Z}_t } \mathbbm{1}_{N_t = 0}$ and a positive strict submartingale started at one, namely $\check{G}_t = e^{\beta \check{Z}_t}_{\beta - \alpha}$. If $\alpha = \beta$, then the two parameter exponential function $e^{\beta x}_{\beta - \alpha}$ in (\ref{new1f}) reduces to the one parameter exponential function $e^{\beta x}$, and hence (\ref{new1f}) simplifies to: \begin{equation} F_t = F_0 e^{ \beta \underline{Z}_t} e^{\beta \check{Z}_t}\mathbbm{1}_{N_t = 0} = F_0 e^{ \beta (\underline{Z}_t + \check{Z}_t)}\mathbbm{1}_{N_t = 0} = e^{ \beta Z_t}\mathbbm{1}_{N_t = 0}, \qquad t \geq t_0, \label{new2f} \end{equation} which is GBM with jump to default. When $\beta \rightarrow 0$, then (\ref{new1f}) asymptotes to: \begin{equation} F_t \rightarrow F_0 e^{ \alpha \underline{Z}_t} (1 + \alpha \check{Z}_t), \qquad t \geq t_0, \label{new3f} \end{equation} which is a two parameter positive continuous martingale. Setting $\gamma = \alpha$ further reduces $F$ to the one parameter positive continuous martingale in \cite{c14}. From \cite{c14}, the bivariate transition PDF of the Brownian Minimum and Brownian Drawup: \begin{eqnarray} \mathbb{Q}_t \{\underline{Z}_T \in dj,\check{Z}_T \in d\check{k}; \underline{Z}_t = \underline{Z}, \check{Z}_t = \check{Z} \} = b(j,\check{k};w,T-t) dj d\check{k} \nonumber\\ b(j,\check{k};w,T-t) \equiv \sqrt{ \frac{2}{\pi (T - t)^3}} (\check{k}-j + w)e^{- \frac{(\check{k}-j+w)^2}{2(T-t)}},\qquad j < \underline{w}, \quad \check{k} \geq 0, \label{jden} \end{eqnarray} where $w = \underline{Z} + \check{Z}$ and $\underline{w}=\underline{Z}$. Note that in a special case when $\underline{Z}_T=\underline{Z}_t$, the bivariate transition PDF becomes a univariate one: \begin{eqnarray} \tilde{\mathbb{Q}}_t \{\underline{Z}_T = \underline{Z}_t,\check{Z}_T \in d\check{k}; \underline{Z}_t = \underline{Z}, \check{Z}_t = \check{Z} \} = \tilde{b}(\check{k};w,T-t) d\check{k} \nonumber\\ \tilde{b}(\check{k};w,T-t)\equiv \sqrt{ \frac{2}{\pi (T - t)}} \left( e^{-\frac{\check{k}^2}{2(T-t)}} - e^{-\frac{(\check{k}+w-\underline{w})^2}{2(T-t)}} \right), \qquad \check{k} \geq 0\,.\label{jden1} \end{eqnarray} Next we construct the bivariate transition PDF for the double-exponential process (\ref{new1f}). Let $\underline{F}^s_T$ be the minimum of $F$ at $T$ conditional on surviving to $T$. Similarly, let $\check{F}^s_T$ be the drawup of $F$ at $T$, conditional on surviving to $T$. The bivariate transition PDF of the Brownian Minimum and Brownian Drawup can be used to derive the bivariate PDF of the pair $(\underline{F}^s_T, \check{F}^s_T)$, conditional both on surviving to $T$ and on $(\underline{F}^s_t,\check{F}_t) = (\underline{F}, \check{F})$. For $J \in (0,F_0]$, and $\check{K} \geq 1$, we seek: $$\mathbb{Q} \{\underline{F}^s_T \in dJ, \check{F}^s_T \in d\check{K} |N_T = 0, \underline{F}^s_t=\underline{F}, \check{F}_t = \check{F} \}.$$ In other words, we wish to know the bivariate conditional PDF when we change variables from $(j,\check{k})$ to: $$ (J,\check{K}) = (F_0 e^{\alpha j}, e^{\beta \check{k}}_{\beta - \alpha}).$$ Let $j(J)$ be the inverse of $J = F_0 e^{\alpha j}$: \begin{equation} j(J) = \frac{1}{\alpha} \ln \left( \frac{J}{F_0} \right), \qquad J \in (0,F_0]. \label{jdef0} \end{equation} Similarly, let $\check{k}(\check{K})$ be the inverse of $\check{K} = e^{\beta \check{k}}_{\beta - \alpha}$: \begin{equation} \check{k}(\check{K}) = \frac{1}{\beta} \ln \left[ \frac{\beta \check{K} + \sqrt{\alpha^2 + \beta^2(\check{K}^2-1)} }{\alpha + \beta} \right], \qquad \check{K} \geq 1. \label{jdef} \end{equation} The determinant of the Jacobian for this change of variables is: \begin{equation} \left(\alpha J \sqrt{\alpha^2 + \beta^2 (\check{K}^2-1)}\right)^{-1}\,. \label{J} \end{equation} Using the standard change of variables formula, it follows that for $J \in (0,F_0], \check{K} \geq 1$, the conditional bivariate PDF of the pair $(\underline{F}^s_T, \check{F}^s_T)$ is given by: \begin{eqnarray} \mathbb{Q} \{\underline{F}^s_T \in dJ, \check{F}^s_T \in d\check{K}|N_T = 0, \underline{F}^s_t=\underline{F}, \check{F}_t = \check{F}\} = f(J,\check{K};w,T-t) dJ d\check{K} \nonumber\\ f(J,\check{K};w,T-t) \equiv \sqrt{ \frac{2}{\pi (T - t)^3}} \frac{\left(\check{k}(\check{K}) - j(J) + w\right)e^{- \frac{\left(\check{k}(\check{K})- j(J) + w \right)^2}{2(T-t)}}}{\alpha J \sqrt{ \alpha^2 + \beta^2 (\check{K}^2-1)}} \,, \end{eqnarray} and \begin{equation} w = j(\underline{F}) + \check{k}(\check{F}). \label{Wt} \end{equation} Note that $w=Z_t$, and the reason we use $w$ is that it is written on market observables $\check{F}$ and $\underline{F}$ while $Z_t$ is not. Let $F^s_T = \underline{F}^s_T \check{F}^s_T$ be the forward price at $T$ conditional on survival to $T$. The bivariate PDF of the pair $(\underline{F}^s_T, \check{F}^s_T)$ can be used calculate the conditional transition PDF of $F^s_T$: \begin{eqnarray} \mathbb{Q} \{F^s_T \in dF|N_T = 0, \underline{F}^s_t=\underline{F}, \check{F}_t = \check{F}\} = g(F;w,T-t) dF \,, \end{eqnarray} where \begin{eqnarray} g(F;w,T-t) & = & \int_0^{F_0} f \left(J, \frac{F}{J};w,T-t \right) dJ \label{Fpdf} \\ & = & \int_0^{F_0} \sqrt{ \frac{2}{\pi (T - t)^3}} \frac{\left(k \left(\frac{F}{J} \right) - j(J) + w\right) e^{- \frac{\left(k \left( \frac{F}{J} \right)- j(J) + w \right)^2}{2(T-t)}}}{\alpha J \sqrt{ \alpha^2 + \beta^2 \left[\left( \frac{F}{J}\right)^2-1 \right]}} dJ, \nonumber \end{eqnarray} and $w$ is given in (\ref{Wt}). When $F$ is only conditioned on surviving to $t$ rather than to $T$, the transition PDF's of both $(\underline{F}_T,\check{F}_T)$ and $F_T$ are just given by the product of their corresponding transition PDF conditioned on survival to $T$ and the probability of further surviving to $T$, which is $e^{-\frac{\beta^2}{2} (T-t)}$: \begin{eqnarray} \mathbb{Q} \{\underline{F}_T \in dJ, \check{F}_T \in d\check{K}|N_t = 0, \underline{F}^s_t=\underline{F}, \check{F}_t = \check{F}\} &=& f(J,\check{K};w,T-t) e^{-\frac{\beta^2}{2} (T-t)} dJ d\check{K}\,, \nonumber\\ \mathbb{Q} \{F_T \in dF|N_t=0, \underline{F}^s_t=\underline{F}, \check{F}_t = \check{F}\} &=& g(F;w,T-t)e^{-\frac{\beta^2}{2} (T-t)}dF\,. \end{eqnarray} The PDF of $F_T$ is an integral over a bounded domain and it cannot be simplified further. We will find that when common payoffs are integrated against this PDF, additional quadratures are not introduced. It is for this reason that we consider the process $F$ to be tractable. There are two similar constructions of a non-negative martingale which also use jump to default. The cumulative hazard process of $N$ is $\Lambda_t = e^{\frac{\beta^2}{2} t}$ which is deterministic. Suppose instead that the cumulative hazard process is $\hat{\Lambda}_t = e^{-\alpha \underline{Z}_t}$, which is random. Let $\hat{N}$ denote the corresponding counting process and let $\hat{F}$ denote the desired non-negative martingale: \begin{equation} \hat{F}_t = F_0 e^{- \frac{\beta^2}{2}t} e^{\beta \check{Z}_t}_{\beta - \alpha} \mathbbm{1}_{N_t = 0}\,, \qquad t \geq t_0 \label{new10f} \end{equation} is a non-negative martingale started at $F_0 > 0$. Since $\underline{Z}_0 = 0$, this process start off with no chance of jumping to zero but soon endures the possibility of such a default. More generally, one can start the process $\underline{Z}$ at some non-positive number $m_0 \leq 0$ and rename the process $\underline{Z}$ to say $m$ since $Z_0$ is still zero. Since $\check{Z}_t = Z_t - m_t$ starts at $-m_0 >0$, one must then also adjust its origin: \begin{equation} F_t = F_0 e^{- \frac{\beta^2}{2}t} e^{\beta (\check{Z}_t + m_0)}_{\beta - \alpha} \mathbbm{1}_{\hat{N}_t = 0}, \qquad t \geq t_0 \label{new11f} \end{equation} There is yet another construction of a non-negative martingale possibly jumping to zero. Now suppose that the cumulative hazard process of $N$ is $\tilde{\Lambda}_t = e^{-\alpha \underline{Z}_t + \frac{\beta^2}{2} t}$, where we return to $\check{Z}_t = Z_t - \underline{Z}_t$ with $\underline{Z}_ 0 = 0$. Let $\tilde{N}$ denote the corresponding counting process and let $\tilde{F}$ denote the desired non-negative martingale: \begin{equation} \tilde{F}_t = F_0 e^{\beta \check{Z}_t}_{\beta - \alpha} \mathbbm{1}_{\tilde{N}_t = 0}, \qquad t \geq t_0 \label{new12f} \end{equation} is a non-negative martingale started at $F_0 > 0$. This process is convenient if an event happens at the first passage time $\tau$ of $F$ to a constant upper barrier $H = e^{\beta h}_{\beta - \alpha}$ where $h > 0$. In this case, $\tau$ is also the first passage time of $\check{Z}$ to $h$. Since $\frac{F}{F_0}$ is a martingale started at one, the bivariate Laplace transform of $ \underline{Z}_{\tau}$ and $\tau$ becomes known: \begin{equation} = E e^{\alpha \underline{Z}_{\tau} - \frac{\beta^2}{2} \tau} \frac{1}{e^{\beta h}_{\beta - \alpha}}. \label{blt} \end{equation} One can develop yet other tractable constructions of non-negative martingales by altering the cumulative hazard process yet again and compensating by coordinate change as was done above. \section{Application in Option Pricing} In risk neutral measure, non-arbitrage insures that the expected payoff of a security is equal to its current price. In this section we show how our model can be applied in derivative pricing assuming the underlying asset follows the dynamics of either the sub-martingale Eqn (\ref{submgl}) or the martingale Eqn (\ref{new1f}) in risk neutral measure. The former is used for derivatives written on spot price of a security while the latter is for future price of a security. Since the two processes only differ by the inclusion of a jump to default process, the pricing formulas for them are quite close. For this reason, we only present the derivation of pricing for the martingale dynamics. The results for the sub-martingale dynamics are labelled by subscripts for clarification. Note since our model tracks the asset's running minimum and drawup rate, it is especially useful in pricing barrier type of path-dependent options. \subsection{One-Touch with a lower barrier}\label{OT} We first price a One-Touch with a lower barrier. A One-Touch option pays one dollar if the underlying asset's price touches the lower barrier price before maturity, and otherwise expires worthless. Assuming that the present time is $t$ and the underlying asset has not defaulted ($N_t = 0$). The price of a One-Touch with a lower barrier $L$ and maturity $T$ is \begin{eqnarray} OT_t(L,T) &=& \mathbbm{1} _{\underline{F}_t \le L} + \mathbbm{1}_{\underline{F}_t >L}\cdot \left(\mathbbm{1}_{N_T = 0} E_t\left[ \mathbbm{1}_{\underline{F}_T\le L} \right] +\mathbbm{1}_{N_T\ne 0}\right)\nonumber\\ &=& \mathbbm{1} _{\underline{F}_t \le L} + \mathbbm{1}_{\underline{F}_t >L} \cdot \left(e^{-\frac{\beta^2(T-t)}{2}}E_t\left[ \mathbbm{1}_{\underline{Z}_T\le \frac{\ln L-\ln F_0}{\alpha}} \right] +1-e^{-\frac{\beta^2(T-t)}{2}} \right)\,, \end{eqnarray} to get the second line, $\underline F _T = F_0 e^{\alpha \underline{Z} _T}$ has been used. After substituting the transition PDF on $\underline{Z}_T$ one obtains \begin{eqnarray} OT_t(L, T) = \mathbbm{1}_{\underline F_t \le L} + \mathbbm{1}_{\underline F_t >L} \cdot \left(e^{-\frac{\beta^2(T-t)}{2}} \left[2N\left(\frac{\frac{\ln L-\ln F_0}{\alpha} - w}{\sqrt{T-t}}\right) -1\right]+1\right)\,, \end{eqnarray} where $w$ is given in (\ref{Wt}) and $N$ is the standard normal distribution function. Taking $\alpha =1$ the price of the One-Touch reduces to that in \cite{c14}. This is because essentially the payoff of a One-Touch option is only determined by the minimum of the underlying, which is driven by the running minimum of a Brownian motion in both cases. A One-Touch written on spot price can be priced similarly with Eqn (\ref{submgl}), which is equivalent to dropping the probability induced by the jump to default process in Eqn (\ref{new1f}). The price is then given by \begin{eqnarray} OT^{\rm{Spot}}_t(L, T) &=&\mathbbm{1} _{\underline{F}_t \le L} + \mathbbm{1}_{\underline{F}_t >L}\cdot E_t\left[ \mathbbm{1}_{\underline{F}_T\le L} \right] \nonumber\\ &=& \mathbbm{1}_{\underline F_t \le L} + \mathbbm{1}_{\underline F_t >L} \cdot 2N\left(\frac{\frac{\ln L-\ln F_0}{\alpha} - w}{\sqrt{T-t}}\right) \,. \end{eqnarray} \subsection{Lookback option} A lookback call option matures at $T$ with a floating strike price pays off the difference between the terminal value of the asset and its minimum, namely the terminal drawup. If default happens $(N_T\ne 0)$, the option expires worthless $(F_T=\underline{F}_T)$. So under the martingale (\ref{new1f}) the value of this option at maturity is then \begin{eqnarray} LC_{float,t}&=& \mathbbm{1}_{N_T=0} E_t\left[ F_T -\underline{F}_T \right] = \mathbbm{1}_{N_T=0} E_t\left[ \underline{F}_T\left(\check{F}_T-1\right)\right] \nonumber\\ &=& \mathbbm{1}_{N_T=0} E_t\left[ F_0e^{\alpha \underline{Z}_T}\left(e^{\beta \check{Z}_T}_{\beta-\alpha}-1\right)\right]\,.\label{lc1} \end{eqnarray} The expectation value in Eqn (\ref{lc1}) can be evaluated using the bivariate transition PDF of $(\underline{Z}_T,\check{Z}_T)$ the in Eqn (\ref{jden}) if the security runs into a new minimum after $t$, or otherwise the univariate transition PDF of $\check{Z}_T$ in Eqn(\ref{jden1}) if $\underline{Z}_T=\underline{Z}_t$: \begin{eqnarray} E_t\left[ F_0 e^{\alpha \underline{Z}_T}\left(e^{\beta \check{Z}_T}_{\beta-\alpha}-1\right)\right] &=& F_0 \int^{\underline{Z}_t}_{-\infty}dj\int_0^{\infty}d\check{k} \sqrt{ \frac{2}{\pi (T - t)^3}} (\check{k}-j + w)e^{- \frac{(\check{k}-j+w)^2}{2(T-t)}} e^{\alpha j}\left(e^{\beta \check{k}}_{\beta-\alpha}-1\right)\nonumber\\ &+&F_0\int_0^{\infty} d\check{k} \sqrt{ \frac{2}{\pi (T - t)}} \left( e^{-\frac{\check{k}^2}{2(T-t)}} - e^{-\frac{(\check{k}+\check{w})^2}{2(T-t)}} \right) e^{\alpha \underline{w}}\left(e^{\beta \check{k}}_{\beta-\alpha}-1\right)\,, \end{eqnarray} where $\check{w}=w-\underline{w}$. By working out the integral we obtain the price of this option evaluated at $t$ \begin{eqnarray} LC_{float, t}&=&F_0 e^{\alpha \underline{w}}\bigg[ \frac{\alpha}{\beta}e^{\beta \check{w}}N\left(\frac{-\check{w}-\beta(T-t)}{\sqrt{T-t}}\right) -\frac{\alpha}{\beta}e^{-\beta \check{w}}N\left(\frac{-\check{w}+\beta(T-t)}{\sqrt{T-t}}\right) \nonumber\\ &+& \frac{\beta +\alpha}{\beta}N\left(\beta \sqrt{T-t}\right)+\frac{\beta -\alpha}{\beta}N\left(-\beta \sqrt{T-t}\right) +e^{-\frac{\beta^2 (T-t)}2} \left(2N\left( \frac{-\check{w}}{\sqrt{T-t}}\right) -1\right) \nonumber\\ &-&2e^{\alpha \check{w}+\frac{(\alpha^2-\beta^2)(T-t)}2}N\left(\frac{-\check{w}-\alpha(T-t)}{\sqrt{T-t}}\right)\bigg]\,. \label{lc} \end{eqnarray} A lookback call option on spot price can be priced the same way: \begin{eqnarray} LC^{\rm{Spot}}_{float,t} &=& E_t\left[ F_T -\underline{F}_T \right] = E_t\left[ F_0e^{\alpha \underline{Z}_T}\left(e^{\beta \check{Z}_T}_{\beta-\alpha}-1\right)\right]\nonumber\\ &=& e^{\frac{\beta^2 (T-t)}2}LC_{float,t}\,.\label{fxlc} \end{eqnarray} If we instead consider a lookback option with a fixed strike price, then the payoff is determined by the minimum/maximum for a put/call lookback option at maturity. Since Eqn (\ref{new1f}) tracks minimum and drawup, it can also be used to evaluate a lookback put option with fixed price. The price is given by \begin{eqnarray} LP_{fixed, t}(K,T)&=&\mathbbm{1}_{N_T=0}E_t\left[(K-\underline{F}_T)^+\right]+\mathbbm{1}_{N_T\ne 0}\cdot K\,,\nonumber\\ LP^{\rm{Spot}}_{fixed, t}(K,T)&=&E_t\left[(K-\underline{F}_T)^+\right]\,, \end{eqnarray} where $K$ is the strike price. This can be evaluated by integrating the price of a one-touch barrier with respect to the barrier, so we will not carry out the derivation for simplicity. We can also engineer another derivative analogous to a lookback call option with a floating strike price, which pays off the ratio between the terminal price and the minimum price before maturity. Since the underlying asset can default ($F_T=\underline{F}_T=0$), we assume the payoff is zero in that case. The price of this option is given by \begin{eqnarray} LC^*_{float,t}&=&\mathbbm{1}_{N_T=0} E_t\left[\frac{F_T-\underline{F}_T}{\underline{F}_T}\right]=\mathbbm{1}_{N_T=0} \left( E_t\left[ \frac{F_T}{\underline{F}_T}\right] -1 \right) \nonumber\\ &=&\mathbbm{1}_{N_T=0} \left( E_t\left[ e^{\beta \check{Z}_T}_{\beta-\alpha}\right] -1 \right)\,, \end{eqnarray} and the expectation can be evaluated with the bivariate PDF: \begin{eqnarray} E_t\left[ e^{\beta \check{Z}_T}_{\beta-\alpha}\right]&=& \int^{\underline{Z}_t}_{-\infty}dj\int_0^{\infty}d\check{k} \sqrt{ \frac{2}{\pi (T - t)^3}} (\check{k}-j + w)e^{- \frac{(\check{k}-j+w)^2}{2(T-t)}} e^{\beta \check{k}}_{\beta-\alpha} \nonumber\\ &+& \int_0^{\infty} d\check{k} \sqrt{ \frac{2}{\pi (T - t)}} \left( e^{-\frac{\check{k}^2}{2(T-t)}} - e^{-\frac{(\check{k}+\check{w})^2}{2(T-t)}} \right) \left(e^{\beta \check{k}}_{\beta-\alpha}-1\right) \end{eqnarray} which can be evaluated similar to Eqn (\ref{lc}), \begin{eqnarray} LC_{float,t}^*&=&\frac{\beta +\alpha}{\beta}N\left(\beta \sqrt{T-t}\right)+\frac{\beta -\alpha}{\beta}N\left(-\beta \sqrt{T-t}\right)-1\,. \end{eqnarray} Note that the value of $LC_{float,t}^*$ is unitless, since the option is written on the drawup ratio. If there is a size associated to the underlying security, it can be multiplied to $LC_{float,t}^*$ which gives it a dollar amount. As in Eqn (\ref{fxlc}), the price for this derivative on spot price is \begin{eqnarray} LC^{*\,\rm{Spot}}_{float,t}=e^{\frac{\beta^2 (T-t)}2}LC^*_{float,t}\,. \end{eqnarray} \subsection{Vanilla and Down-and-In Call} Now we price a Down-and-In Call (DIC) option which becomes from worthless to a vanilla call if the lower barrier is hit before maturity. A vanilla call can be viewed as a special case of a Down-and-In barrier call (DIC) with the lower barrier has been hit prior to presence. The value of a DIC option written on $F_t$ is given by \begin{equation} {\rm{DIC}}_t(L, K, T) =\mathbbm{1}_{\underline{F}_t \le L}\cdot \mathbbm{1}_{N_T = 0}\cdot C_t(K,T)+ \mathbbm{1}_{\underline{F}_t > L}\cdot \mathbbm{1}_{N_T = 0}\cdot E_t \left[ \mathbbm{1}_{\underline{F}_T \le L} (F_T - K)^+\right]\,,\label{dic} \end{equation} where $L$ is the barrier, $K$ is the strike price, $T$ is maturity and $C_t$ is a vanilla call priced at $t$. Note setting $L ={F}_0$ reduces the DIC to a vanilla call. As implied by Eqn (\ref{dic}) if default happens ($N_T\ne 0$), the option becomes worthless. To evaluate the expectation value of the second term in (\ref{dic}), we once again apply the bivariate transition PDF: \begin{eqnarray} E_t \left[\mathbbm{1}_{\underline{F}_T \le L} (F_T - K)^+\right] &=& E_t\left[\mathbbm{1}_{\underline{Z}_T \le \frac 1{\alpha}\ln{\frac L{F_0}}} (F_0 e^{\alpha\underline{Z}_T}e^{\beta \vw_T}_{\beta - \alpha} - K)^+\right]\nonumber\\ &=& \int_{-\infty}^{\frac 1{\alpha}\ln{\frac L{F_0}}} dj \int_{k^*}^{\infty}d\check{k} \sqrt{ \frac{2}{\pi (T - t)^3}} (\check{k}-j + w)e^{- \frac{(\check{k}-j+w)^2}{2(T-t)}} \left(F_0 e^{\alpha j}e^{\beta \check{k}}_{\beta-\alpha}-K\right)\,, \end{eqnarray} where $k^*(j)$ is determined by \begin{eqnarray} k^*={\rm{max}}\left(f^{-1}\left(\frac S {F_0 e^{\alpha j}}\right),0\right)\,, \qquad f(x)=e^{\beta x}_{\beta-\alpha}\,. \end{eqnarray} For the dependence of $k^*$ on $j$, the integral above cannot be obtained in closed form, a similar situation as in \cite{c14}. Nonetheless, the result can be further simplified as \begin{eqnarray} E_t \left[\mathbbm{1}_{\underline{F}_T \le L} (F_T - K)^+\right] &=& F_0 \int_{-\infty}^{\frac 1{\alpha}\ln{\frac L{F_0}}} dj e^{\alpha j+\frac{\beta^2(T-t)}2}\bigg[(\beta+\alpha)e^{\beta(j-w)}N\left(\frac{j-w-k^*+\beta (T-t)}{\sqrt{T-t}}\right) \nonumber\\ &&\qquad -(\beta-\alpha)e^{-\beta(j-w)}N\left(\frac{j-w-k^*-\beta (T-t)}{\sqrt{T-t}}\right) \bigg]\,, \end{eqnarray} which gives rise to the value of the DIC option After replacing $Z_t$ with the market observable $w$, we now have the price for the DIC option: \begin{eqnarray} {\rm{DIC_t}}(L, K, T) &=&\mathbbm{1}_{\underline{F}_t \le L}C_t(K,T) +\mathbbm{1}_{\underline{F}_t > L} F_0 \int_{-\infty}^{\frac 1{\alpha}\ln{\frac L{F_0}}} dj e^{\alpha j}\bigg[(\beta+\alpha)e^{\beta(j-w)}N\left(\frac{j-w-k^*+\beta (T-t)}{\sqrt{T-t}}\right) \nonumber\\ &&\qquad \qquad -(\beta-\alpha)e^{-\beta(j-w)}N\left(\frac{j-w-k^*-\beta (T-t)}{\sqrt{T-t}}\right) \bigg]\,. \label{dic2} \end{eqnarray} In the special case when $L = F_0$, the DIC option reduces to a vanilla call with a price of \begin{eqnarray} C_t( K, T) &=& F_0 \int_{-\infty}^{0} dj e^{\alpha j}\bigg[(\beta+\alpha)e^{\beta(j-w)}N\left(\frac{j-w-k^*+\beta (T-t)}{\sqrt{T-t}}\right)\nonumber\\ &&\quad\quad -(\beta-\alpha)e^{-\beta(j-w)}N\left(\frac{j-w-k^*-\beta (T-t)}{\sqrt{T-t}}\right) \bigg]\,, \label{call} \end{eqnarray} which completes the pricing of a DIC option on Eqn (\ref{new1f}). For a DIC option on spot price, Eqn (\ref{dic}) becomes \begin{equation} {\rm{DIC}}^{\rm{Spot}}_t(L, K, T) =\mathbbm{1}_{\underline{F}_t \le L}\cdot C^{\rm{FX\,\, Spot}}_t(K,T)+ \mathbbm{1}_{\underline{F}_t > L}\cdot E_t \left[ \mathbbm{1}_{\underline{F}_T \le L} (F_T - K)^+\right]\,, \end{equation} which leads to slight modification on both Eqn (\ref{dic2}) and Eqn (\ref{call}), and the results are \begin{eqnarray} {\rm{DIC}}^{\rm{Spot}}_t(L, K, T) &=&\mathbbm{1}_{\underline{F}_t \le L}C^{\rm{Spot}}_t(K,T) \nonumber\\ &+&\mathbbm{1}_{\underline{F}_t > L} F_0 e^{\frac{\beta^2 (T-t)}2} \int_{-\infty}^{\frac 1{\alpha}\ln{\frac L{F_0}}} dj e^{\alpha j}\bigg[(\beta+\alpha)e^{\beta(j-w)}N\left(\frac{j-w-k^*+\beta (T-t)}{\sqrt{T-t}}\right) \nonumber\\ &&\qquad \qquad -(\beta-\alpha)e^{-\beta(j-w)}N\left(\frac{j-w-k^*-\beta (T-t)}{\sqrt{T-t}}\right) \bigg]\,,\nonumber\\ C^{\rm{ Spot}}_t( K, T) &=& F_0 e^{\frac{\beta^2 (T-t)}2} \int_{-\infty}^{0} dj e^{\alpha j}\bigg[(\beta+\alpha)e^{\beta(j-w)}N\left(\frac{j-w-k^*+\beta (T-t)}{\sqrt{T-t}}\right)\nonumber\\ &&\quad\quad -(\beta-\alpha)e^{-\beta(j-w)}N\left(\frac{j-w-k^*-\beta (T-t)}{\sqrt{T-t}}\right) \bigg]\,. \end{eqnarray} Before closing this section, we would like to point out that Eqn (\ref{dic2}) is related to several options. For instance, when $\alpha=1$ and $\beta = 0$ the result reduces to that in \cite{c14}. In the special case of a zero strike DIC option ($K=0$), Eqn (\ref{dic2}) has closed form expressions: \begin{eqnarray} &&DIC_t(L,0,T)=F_0\bigg[\left(\frac L{F_0}\right)^{\alpha+\beta}e^{-\beta w}N\left(\frac{\frac 1{\alpha}\ln \frac L{F_0}-w+\beta(T-t)}{\sqrt{T-t}}\right)\nonumber\\ &&\quad\quad -2e^{\alpha w +\frac{(\alpha^2-\beta^2)(T-t)}2}N\left( \frac{\frac 1{\alpha}\ln \frac L{F_0}-w-\alpha(T-t)}{\sqrt{T-t}}\right)+\left(\frac L{F_0}\right)^{\alpha-\beta}e^{\beta w}N\left(\frac{\frac 1{\alpha}\ln \frac L{F_0}-w-\beta(T-t)}{\sqrt{T-t}}\right)\bigg]. \end{eqnarray} \section{Summary and Extensions} We proposed a three parameter continuous martingale with state space $[0,\infty)$. This is done by first generating a process with a positive drift driven by the running minimum and drawup of a Brownian motion in the Az{\' e}ma-Yor setting, and adding a jump to default process. The process generalizes driftless Geometric Brownian motion by adding two more parameters while preserving its tractability. In particular, its running minimum and drawup rate (the ratio between level and running minimum) are both analytically tractable. The three model parameters $\alpha, \gamma$, and $\beta$ can be respectively interpreted as the instantaneous volatility of the underlying at each new low, at the initial time, and at infinitely high prices of the underlying. The parameter $\alpha$ controls the implied volatility at low strikes, while the parameter $\beta$ controls the implied volatility at high strikes. So long as implied volatility is monotonic in strike price, the parameter $\gamma$ can be used to meet an at-the-money implied volatility. It is shown that in certain limits, this new process can reduce to Geometric Brownian motion and the positive martingale given in \cite{c14}. We also presented the bivariate transition PDF of the process' running minimum and drawup rate. By utilizing the transition PDF, we priced several options assuming the dynamics are driven by the three parameter martingale in risk neutral measure. The options include a one-touch option with a lower barrier, lookback options with floating and fixed strike prices, vanilla call and a down-and-in call option. Since not all implied volatility slices are monotonic, future research should be directed towards extending the model by introducing either stochastic volatility or jumps. One can also use the process without jump to default to model dynamics that involve a positive drift, for instance, the cumulative return of an investment strategy. Moreover, Girsanov's theorem can be used to remove the drift of $G$, at which point a reflection principle becomes available. In the interests of brevity, these extensions are best left for future research. \section*{Acknowledgement} We are grateful to Matthew Lorig, Vasily Strela, Jane Yu, and especially Travis Fisher, for their comments. They are not responsible for any errors. \section*{Appendix} \subsection*{1. More about $e^{\beta x}_{\beta - \alpha}$} This technical appendix proves the result (\ref{swap}). For $x \geq 0$, $\alpha \geq 0$, and $\beta>0$, our two parameter exponential function is defined as: \begin{equation} e^{\beta x}_{\alpha - \beta} \equiv \frac{\alpha + \beta}{2 \beta} e^{\beta x} + \frac{\alpha - \beta}{2 \beta} e^{-\beta x}. \label{agex2} \end{equation} Squaring this result implies that: \begin{equation} \left( e^{ \beta x}_{\alpha - \beta} \right)^2 = \left( \frac{\alpha + \beta}{2 \beta} \right)^2 e^{2\beta x} + \frac{\alpha^2 - \beta^2}{2 \beta^2} + \left( \frac{\alpha - \beta}{2 \beta} \right)^2e^{-2\beta x}. \label{agex3} \end{equation} Consider the cohort of (\ref{agex2}): \begin{equation} e^{\beta x}_{\beta - \alpha} \equiv \frac{\alpha + \beta}{2 \beta} e^{\beta x} + \frac{\beta - \alpha}{2 \beta} e^{-\beta x}. \label{agex} \end{equation} Squaring this cohort implies that: \begin{equation} \left( e^{ \beta x}_{\beta - \alpha} \right)^2 = \left( \frac{\alpha + \beta}{2 \beta} \right)^2 e^{2\beta x} - \frac{\alpha^2 - \beta^2}{2 \beta^2} + \left( \frac{\alpha - \beta}{2 \beta} \right)^2e^{-2\beta x}. \label{agex1} \end{equation} Subtracting (\ref{agex1}) from (\ref{agex3}) implies that: \begin{equation} \left( e^{ \beta x}_{\alpha - \beta} \right)^2- \left( e^{ \beta x}_{\beta - \alpha} \right)^2 = \frac{\alpha^2 - \beta^2}{\beta^2}. \label{diff} \end{equation} Taking the positive square root of each side leads to the desired result: \begin{equation} e^{\beta x}_{\alpha - \beta} = \sqrt{ \left( e^{\beta x}_{\beta - \alpha} \right)^2 + \frac{\alpha^2 - \beta^2}{\beta^2}}. \label{aswap} \end{equation} \newpage \begin {thebibliography}{99} \bibitem{b76} Black, F., 1976, ``The Pricing of Commodity Contracts'', {\em Journal of Financial Economics}, {\bf 3}, 167--179. \bibitem{c14} Carr P., 2014, ``First Order Calculus and Option Pricing'', {\em Journal of Financial Engineering} {\bf 1}, 1. \bibitem{g14} Guyon, J., 2014, ``Path-Dependent Volatility'', {\em Risk}, {\bf 10}. \bibitem{hr98} Hobson, D. G. and L. C. G. Rogers, 1998, ``Complete Models with Stochastic Volatility'', {\em Mathematical Finance}, {\bf 8}, 27-�48. \bibitem{m76} Merton, R.C., 1976, ``Option pricing when underlying stock returns are discontinuous'', {\em Journal of Financial Economics}, {\bf 3}, 125-�144. \bibitem{g83} Grabbe, J.O., 1983, ``The pricing of call and put options on foreign exchange'', {\em Journal of International Money and Finance}, {\bf 2}, 239-�253. \end{thebibliography} \end{document}

d6a62e1f0defca5a5b06673cfc880e3bc8dcc2cf

\subsubsection*{\bibname}} \usepackage{amssymb} \usepackage{amsmath} \usepackage[linesnumbered,boxed,ruled,vlined]{algorithm2e} \usepackage{algorithmic} \SetKwRepeat{DoWhile}{do}{while} \newcommand{\eat}[1]{} \newcommand{\topic}[1]{\vspace{5pt}\noindent{{\bf #1:}}} \newcommand{{\mathcal X}}{{\mathcal X}} \newcommand{{\mathcal A}}{{\mathcal A}} \newcommand{{\mathcal H}}{{\mathcal H}} \newcommand{{\mathcal C}}{{\mathcal C}} \newcommand{{\mathcal F}}{{\mathcal F}} \newcommand{{\mathcal P}}{{\mathcal P}} \newcommand{{\mathcal N}}{{\mathcal N}} \newcommand{{\mathcal I}}{{\mathcal I}} \newcommand{{\mathcal M}}{{\mathcal M}} \newcommand{{\mathbb R}}{{\mathbb R}} \newcommand{{\mathbf x}}{{\mathbf x}} \newcommand{{\mathbf y}}{{\mathbf y}} \newcommand{{\mathsf{Pr}}}{{\mathsf{Pr}}} \newcommand{{\mathbb{E}}}{{\mathbb{E}}} \newcommand{\mathrm{d}}{\mathrm{d}} \newcommand{\mathsf{rank}}{\mathsf{rank}} \newcommand{\mathsf{OPT}}{\mathsf{OPT}} \newcommand{\mathsf{SOL}}{\mathsf{SOL}} \newcommand{\mathsf{poly}}{\mathsf{poly}} \newcommand{{\mathcal G}}{{\mathcal G}} \newcommand{\red}[1]{\textcolor{red}{#1}} \newcommand{\blue}[1]{\textcolor{blue}{#1}} \newcommand{\inner}[2]{\langle #1, #2 \rangle} \newcommand{\nsl}[1]{\| #1 \|_2^2} \newcommand{\ns}[1]{\| #1 \|^2} \newcommand{\n}[1]{\| #1 \|} \newcommand{\hat{x}}{\hat{x}} \newcommand{\hat{y}}{\hat{y}} \newcommand{\hat{z}}{\hat{z}} \newcommand{\widetilde{x}}{\widetilde{x}} \newcommand{\tilde{y}}{\tilde{y}} \newcommand{\tilde{z}}{\tilde{z}} \newcommand{\bar{x}}{\bar{x}} \renewcommand{\algorithmicrequire}{ \textbf{Input:}} \renewcommand{\algorithmicensure}{ \textbf{Output:}} \newenvironment{proofsketch}{\noindent{\em Proof Sketch.}}% {\hspace*{\fill}$\Box$\par\vspace{4mm}} \newenvironment{proofof}[1]{\smallskip\noindent{\em Proof of #1.}}% {\hspace*{\fill}$\Box$\par} \newenvironment{proof}{\vspace{-0.05in}\noindent{\em Proof.}}% {\hspace*{\fill}$\Box$\par} \newcommand{\mathsf{AA}}{\mathsf{AA}} \newcommand{\mathsf{GMRES}}{\mathsf{GMRES}} \newcommand*{\hfill\ensuremath{\square}}{\hfill\ensuremath{\square}} \hyphenpenalty=700 \newtheorem{theorem}{Theorem} \newtheorem{claim}{Claim} \newtheorem{lemma}{Lemma} \newtheorem{corollary}{Corollary} \newtheorem{definition}{Definition} \newtheorem{assumption}{Assumption} \newtheorem{observation}{Observation} \usepackage{graphicx} \usepackage{hyperref} \hypersetup{ colorlinks=true, linkcolor=black, citecolor=black, filecolor=black, urlcolor=black } \begin{document} \runningtitle{A Fast Anderson-Chebyshev Acceleration for Nonlinear Optimization} \runningauthor{Zhize Li, Jian Li} \twocolumn[ \aistatstitle{A Fast Anderson-Chebyshev Acceleration for \\ Nonlinear Optimization} \aistatsauthor{ Zhize Li \And Jian Li } \aistatsaddress{ King Abdullah University of Science and Technology \And Tsinghua University } ] \begin{abstract} \emph{Anderson acceleration} (or Anderson mixing) is an efficient acceleration method for fixed point iterations $x_{t+1}=G(x_t)$, e.g., gradient descent can be viewed as iteratively applying the operation $G(x) \triangleq x-\alpha\nabla f(x)$. It is known that Anderson acceleration is quite efficient in practice and can be viewed as an extension of Krylov subspace methods for nonlinear problems. In this paper, we show that Anderson acceleration with Chebyshev polynomial can achieve the optimal convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$, which improves the previous result $O(\kappa\ln\frac{1}{\epsilon})$ provided by \citep{toth2015convergence} for quadratic functions. Moreover, we provide a convergence analysis for minimizing general nonlinear problems. Besides, if the hyperparameters (e.g., the Lipschitz smooth parameter $L$) are not available, we propose a \emph{guessing algorithm} for guessing them dynamically and also prove a similar convergence rate. Finally, the experimental results demonstrate that the proposed Anderson-Chebyshev acceleration method converges significantly faster than other algorithms, e.g., vanilla gradient descent (GD), Nesterov's Accelerated GD. Also, these algorithms combined with the proposed guessing algorithm (guessing the hyperparameters dynamically) achieve much better performance. \end{abstract} \section{Introduction} Machine learning problems are usually modeled as optimization problems, ranging from convex optimization to highly nonconvex optimization such as deep neural networks, e.g., \citep{nesterov2014introductory, bubeck2015convex, lecun2015deep, lei2017non, li2018simple, fang2018spider, zhou2018stochastic, li2019stochastic, ge2019stable, li2019ssrgd}. To solve an optimization problem $\min_x f(x)$, the classical method is gradient descent, i.e., $x_{t+1}= x_t-\alpha_t\nabla f(x_t)$. There exist several techniques to accelerate the standard gradient descent, e.g., momentum \citep{nesterov2014introductory, allen2017katyusha, lan2018random, lan2019unified}. There are also various vector sequence acceleration methods developed in the numerical analysis literature, e.g., \citep{brezinski2000convergence,sidi1986acceleration,smith1987extrapolation,brezinski1991extrapolation,brezinski2018shanks}. Roughly speaking, if a vector sequence converges very slowly to its limit, then one may apply such methods to accelerate the convergence of this sequence. Taking gradient descent as an example, the vector sequence are generated by $x_{t+1}= G(x_t) \triangleq x_t-\alpha_t\nabla f(x_t)$, where the limit is the fixed-point $G(x^*) = x^*$ (i.e. $\nabla f(x^*)=0)$. One notable advantage of such acceleration methods is that they usually do not require to know how the vector sequence is actually generated. Thus the applicability of those methods is very wide. Recently, \cite{scieur2016regularized} used the minimal polynomial extrapolation (MPE) method \citep{smith1987extrapolation} for convergence acceleration. This is a nice example of using sequence acceleration methods to optimization problems. In this paper, we are interested in another classical sequence acceleration method called \emph{Anderson acceleration} (or \emph{Anderson mixing}), which was proposed by Anderson in 1965 \citep{anderson1965iterative}. The method is known to be quite efficient in a variety of applications \citep{capehart1989techniques,pratapa2016anderson,higham2016anderson,loffeld2016considerations}. The idea of Anderson acceleration is to maintain $m$ recent iterations for determining the next iteration point, where $m$ is a parameter (typically a very small constant). Thus, it can be viewed as an extension of the existing momentum methods which usually use the last and current points to determine the next iteration point. Anderson acceleration with slight modifications is described in Algorithm \ref{alg:am}. \begin{algorithm}[h] \caption{Anderson Acceleration($m$)} \label{alg:am} \textbf{input:} $x_0, T, \lambda, \beta_t$\\ Define $G(x)\triangleq x+F \triangleq x - \lambda\nabla f(x)$\; $x_1=G(x_0)$, $F_0=G(x_0)-x_0$\; \For{$t= 1, 2,\ldots T$}{ $m_t=\min\{m,t\}$\; $F_t\triangleq G(x_t)-x_t$\; Solve $\min_{\alpha^t=(\alpha_0^t,\ldots,\alpha_{m_t}^t)^T} \left\|\sum_{i=0}^{m_t}{\alpha_i^t F_{t-i}}\right\|_2$ subject to $\sum_{i=0}^{m_t}{\alpha_i^t=1}$\; \label{line:ls} $x_{t+1}=(1-\beta_t)\sum_{i=0}^{m_t}{\alpha_i^tx_{t-i}} + \beta_t\sum_{i=0}^{m_t}{\alpha_i^tG(x_{t-i})}$\; } \Return $x_T$ \end{algorithm} Note that the step in Line \ref{line:ls} of Algorithm \ref{alg:am} can be transformed to an equivalent unconstrained least-squares problem: \begin{equation}\label{eq:transt} \min_{(\alpha_1^t,\ldots,\alpha_{m_t}^t)^T}\Big\|F_t-\sum_{i=1}^{m_t}{\alpha_i^t (F_t-F_{t-i})}\Big\|_2, \end{equation} then let $\alpha_0^t= 1-\sum_{i=1}^{m_t}{\alpha_i^t}$. Using QR decomposition, \eqref{eq:transt} can be solved in time $2m_t^2d$, where $d$ is the dimension. Moreover, the QR decomposition of \eqref{eq:transt} at iteration $t$ can be efficiently obtained from that of at iteration $t-1$ in $O(m_td)$ (see, e.g. \citep{golub1996matrix}). The constant $m_t\leq m$ is usually very small. We use $m=3$ and $5$ for the numerical experiments in Section \ref{sec:exp}. Hence, each iteration of Anderson acceleration can be implemented quite efficiently. Many studies showed the relations between Anderson acceleration and other optimization methods. In particular, for the quadratic case (linear problems), \cite{walker2011anderson} showed that it is related to the well-known Krylov subspace method GMRES (generalized minimal residual algorithm) \citep{saad1986gmres}. Furthermore, \cite{potra2013characterization} showed that GMRES is equivalent to Anderson acceleration with any mixing parameters under $m =\infty$ (see Line 5 of Algorithm \ref{alg:am}) for linear problems. Concretely, \cite{toth2015convergence} proved the first linear convergence rate $O(\kappa\ln\frac{1}{\epsilon})$ for linear problems with fixed parameter $\beta$, where $\kappa$ is the condition number. Besides, \cite{eyert1996comparative}, and \cite{fang2009two} showed that Anderson acceleration is related to the multisecant quasi-Newton methods (more concretely, the generalized Broyden's second method). Despite the above results, the convergence results for this efficient method are still limited (especially for general nonlinear problems and the case where $m$ is small). In this paper, we analyze the convergence for small $m$ which is the typical case in practice and also provide the convergence analysis for general nonlinear problems. \subsection{Our Contributions} There has been a growing number of applications of Anderson acceleration method \citep{pratapa2016anderson,higham2016anderson,loffeld2016considerations,scieur2018nonlinear}. Towards a better understanding of this efficient method, we make the following technical contributions: \begin{enumerate} \item We prove the optimal $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ convergence rate of the proposed Anderson-Chebyshev acceleration (i.e., Anderson acceleration with Chebyshev polynomial) for minimizing quadratic functions (see Theorem \ref{thm:opt}). Our result improves the previous result $O(\kappa\ln\frac{1}{\epsilon})$ given by \citep{toth2015convergence} and matches the lower bound $\Omega(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ provided by \citep{nesterov2014introductory}. Note that for ill-conditioned problems, the condition number $\kappa$ can be very large. \item Then, we prove the linear-quadratic convergence of Anderson acceleration for minimizing general nonlinear problems under some standard assumptions (see Theorem \ref{thm:gel}). Compared with Newton-like methods, it is more attractive since it does not require to compute (or approximate) Hessians, or Hessian-vector products. \item Besides, we propose a \emph{guessing algorithm} (Algorithm \ref{alg:guess}) for the case when the hyperparameters (e.g., $\mu, L$) are not available. We prove that it achieves a similar convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon}+\sqrt{\kappa}(\ln\kappa\ln B)^2)$ (see Theorem \ref{thm:guess}). This guessing algorithm can also be combined with other algorithms, e.g., Gradient Descent (GD), Nesterov's Accelerated GD (NAGD). The experimental results (see Section \ref{app:expga}) show that these algorithms combined with the proposed guessing algorithm achieve much better performance. \item Finally, the experimental results on the real-world UCI datasets and synthetic datasets demonstrate that Anderson acceleration methods converge significantly faster than other algorithms (see Section \ref{sec:exp}). Combined with our theoretical results, the experiments validate that Anderson acceleration methods (especially Anderson-Chebyshev acceleration) are efficient both in theory and practice. \end{enumerate} \subsection{Related Work} \label{sec:rw} As aforementioned, Anderson acceleration can be viewed as the extension of the momentum methods (e.g., NAGD) and the potential extension of Krylov subspace methods (e.g., GMRES) for nonlinear problems. In particular, GD is the special case of Anderson acceleration with $m=0$, and to some extent NAGD can be viewed as $m=1$. We also review the equivalence of GMRES and Anderson acceleration without truncation (i.e., $m =\infty$) in Appendix \ref{app:gam}. Besides, \cite{eyert1996comparative}, and \cite{fang2009two} showed that Anderson acceleration is related to the multisecant quasi-Newton methods. Note that Anderson acceleration has the advantage over the Newton-like methods since it does not require the computation of Hessians or approximation of Hessians or Hessian-vector products. There are many sequence acceleration methods in the numerical analysis literatures. In particular, the well-known Aitken's $\Delta^2$ process \citep{aitken1926bernoulli} accelerated the convergence of a sequence that is converging linearly. Shanks generalized the Aitken extrapolation which was known as Shanks transformation \citep{shanks1955non}. Recently, \cite{brezinski2018shanks} proposed a general framework for Shanks sequence transformations which includes many vector sequence acceleration methods. One fundamental difference between Anderson acceleration and other sequence acceleration methods (such as MPE, RRE (reduced rank extrapolation) \citep{sidi1986acceleration,smith1987extrapolation}, etc.) is that Anderson acceleration is a fully dynamic method \citep{capehart1989techniques}. Here \emph{dynamic} means all iterations are in the same sequence, and it does not require to restart the procedure. It can be seen from Algorithm \ref{alg:am} that all iterations are applied to the same sequence $\{x_t\}$. In fact, in Capehart's PhD thesis \citep{capehart1989techniques}, several experiments were conducted to demonstrate the superior performance of Anderson acceleration over other semi-dynamic methods such as MPE, RRE (semi-dynamic means that the algorithm maintains more than one sequences or needs to restart several times). More recently, Anderson acceleration with different variants and/or under different assumptions are widely studied (see e.g., \citep{zhang2018globally, evans2018proof, scieur2018nonlinearv2}). \section{The Quadratic Case} \label{sec:opt} In this section, we consider the problem of minimizing a quadratic function (also called least squares, or ridge regression \citep{boyd2004convex,hoerl1970ridge}). The formulation of the problem is \begin{equation}\label{prob:quad} \min_{x\in {\mathbb R}^d} f(x)=\frac{1}{2}x^TAx-b^Tx, \end{equation} where $\mu I_d\preceq\nabla^2f=A\preceq LI_d$. Note that $\mu$ and $L$ are usually called the strongly convex parameter and Lipschitz continuous gradient parameter, respectively (e.g. \citep{nesterov2014introductory, allen2017katyusha, lan2019unified}). There are many algorithms for optimizing this type of functions. See e.g.~\citep{bubeck2015convex} for more details. We analyze the problem of minimizing a more general function $f(x)$ in the next Section \ref{sec:gel}. We prove that Anderson acceleration with Chebyshev polynomial parameters $\{\beta_t\}$ achieves the optimal convergence rate, i.e., it obtains an $\epsilon$-approximate solution using $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ iterations. The convergence result is stated in the following Theorem \ref{thm:opt}. \begin{theorem}\label{thm:opt} The Anderson-Chebyshev acceleration method achieves the optimal convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ for obtaining an $\epsilon$-approximate solution of problem (\ref{prob:quad}) for any $0\leq m\leq k$, where $\kappa=L/\mu$ is the condition number, $k$ is defined in Definition \ref{def:k} and this method combines Anderson acceleration (Algorithm \ref{alg:am}) with the Chebyshev polynomial parameters $\beta_t=1/\big(\frac{L+\mu}{2}+\frac{L-\mu}{2}\cos(\frac{(2t-1)\pi}{2T})\big)$, for $t=1, 2, \ldots, T$. \end{theorem} \textbf{Remark:} In this quadratic case, we mention that \cite{toth2015convergence} proved the first convergence rate $O(\kappa\ln\frac{1}{\epsilon})$ for fixed parameter $\beta$. Here we use the Chebyshev polynomials to improve the result to the optimal $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ which matches the lower bound $\Omega(\sqrt{\kappa}\ln\frac{1}{\epsilon})$. Note that for ill-conditioned problems, the condition number $\kappa$ can be very large. Also note that in practice the constant $m$ is usually very small. Particularly, $m=3$ has already achieved a remarkable performance from our experimental results (see Figures \ref{fig:lreg}--\ref{fig:5} in Section \ref{sec:exp}). Before proving Theorem \ref{thm:opt}, we first define $k$ and then briefly review some properties of the Chebyshev polynomials. We refer to \citep{rivlin1974chebyshev,olshanskii2014iterative,hageman2012applied} for more details of Chebyshev polynomials. \begin{definition}\label{def:k} Let $v_i$'s be the unit eigenvectors of $A$, where $A$ is defined in \eqref{prob:quad}. Consider a unit vector $c\triangleq \sum_{i=1}^d{c_iv_i}$ and let $c'\triangleq \mathrm{Proj}_{B_k^\perp} c=\sum_{i=1}^d{c_i'v_i}$, where $\mathrm{Proj}_{B_k^\perp}$ denotes the projection to the orthogonal complement of the column space of $B_k \triangleq A[x_{t-k}-x_{t}, \dotsc, x_{t-1}-x_{t}] \in \mathbb{R}^{d\times k}$. Define $k$ to be the maximum integer such that $c_i'\leq (1+\frac{1}{\sqrt{\kappa}+1}){c_i}$ for any $i\in [d]$. \end{definition} Obviously, $k\geq 0$ since $c'=c$ due to $B_0=0$ and $\mathrm{Proj}_{B_0^\perp}=I$. Now we review the Chebyshev polynomials. The \emph{Chebyshev polynomials} are polynomials $P_k(x)$, where $k\geq 0$, $\deg(P_k)=k$, which is defined by the recursive relation: \begin{equation}\label{eq:cheb0} \begin{split} & P_0(x)=1,\\ & P_1(x)=x, \\ & P_{k+1}(x)=2xP_k(x)-P_{k-1}(x). \end{split} \end{equation} The key property is that $P_k(x)$ has minimal deviation from $0$ on $[-1,1]$ among all polynomials $Q_k$ with $\deg(Q_k)=k$ and leading coefficient $\alpha_k=2^{k-1}$ for the largest degree term $x^k$, i.e., \begin{equation}\label{eq:chebmin} \max_{x\in[-1,1]}|P_k(x)|\leq \max_{x\in[-1,1]}|Q_k(x)| \quad \mathrm{for~all}~ Q_k. \end{equation} In particular, for $|x|\leq 1$, Chebyshev polynomials can be written in an equivalent way: \begin{align} P_k(x)=\cos(k\arccos x). \label{eq:cheb} \end{align} In our proof, we use this equivalent form (\ref{eq:cheb}) instead of (\ref{eq:cheb0}). The equivalence can be verified as follows: \begin{align} P_k(x)&=2x\cos((k-1)\arccos x)-\cos((k-2)\arccos x) \notag\\ &= 2\cos \theta \cos((k-1)\theta)-\cos((k-2)\theta) \label{eq:xx11}\\ &=\cos(k\theta)+\cos((k-2)\theta)-\cos((k-2)\theta) \notag\\ &=\cos(k\arccos x), \label{eq:xx22} \end{align} where (\ref{eq:xx11}) and (\ref{eq:xx22}) use the transformation $x=\cos\theta$ due to $|x|\leq 1$. According to (\ref{eq:cheb}), $\max_{x\in[-1,1]}|P_k(x)|=1$ and the $k$ roots of $P_k$ are as follows: \begin{equation}\label{eq:root} x_i=\cos\Big(\frac{(2i-1)\pi}{2k}\Big),~ i=1,2,\ldots,k. \end{equation} To demonstrate it more clearly, we provide an example for $P_4(x)$ (W-shape curve) in Figure \ref{fig:cheb}. Since $k=4$ in this polynomial $P_4(x)$, the first root $x_1=\cos\left(\frac{(2i-1)\pi}{2k}\right)=\cos\left(\frac{\pi}{8}\right)\thickapprox 0.92$. The remaining three roots for $P_4(x)$ can be easily computed too. \begin{figure}[!h] \centering \includegraphics[width=\linewidth]{chebyshev.pdf} \caption{The Chebyshev polynomial $P_4(x)$} \label{fig:cheb \end{figure} \begin{proofof}{Theorem \ref{thm:opt}} For iteration $t+1$, the residual $F_{t+1}\triangleq -\lambda\nabla f(x_{t+1}) = -(Ax_{t+1}-b)$ (let $\lambda=1$) can be deduced as follows: \begin{align} F_{t+1} &=b-Ax_{t+1}\notag \\ &=b-A\biggl[(1-\beta_t)\sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}}+ \beta_t\sum_{i=0}^{m_t}{\alpha_i^t G(x_{t-i})}\biggr]\notag \\ &=b-A\biggl[\sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}}+ \beta_t\sum_{i=0}^{m_t}{\alpha_i^t \left(b-Ax_{t-i})\right)}\biggr] \label{eq:quadx1}\\ &=b-\beta_tAb-A\biggl[ \sum_{i=0}^{m_t}{\alpha_i^t \left((I-\beta_tA)x_{t-i})\right)}\biggr] \notag \\ &=(I-\beta_tA)\sum_{i=0}^{m_t}{\alpha_i^t \left(b-Ax_{t-i})\right)} \notag \\ &=(I-\beta_tA)\sum_{i=0}^{m_t}{\alpha_i^t F_{t-i}}, \label{eq:quadx2} \end{align} where (\ref{eq:quadx1}) uses $G(x_t)=x_t +F_t$. To bound $\|F_{t+1}\|_2$ (i.e., $\|\nabla f(x_{t+1})\|_2$), we first obtain the following lemma by using Singular Value Decomposition (SVD) to solve the least squares problem (\ref{eq:transt}) and then using several transformations. We defer the proof of Lemma \ref{lm:boundft1} to Appendix \ref{app:pflm2}. \begin{lemma}\label{lm:boundft1} Let $F_1=b-Ax_{1}$ and $F_{t+1}=b-Ax_{t+1}$, then \begin{equation}\label{eq:chebh} \|F_{t+1}\|_2/\|F_1\|_2 \leq \sqrt{2\min_{\beta}\max_{\lambda\in [\mu,L]}|H_t(\lambda)|} \end{equation} where $H_t(\lambda) =(1-\beta_t\lambda)\dotsm(1-\beta_1\lambda)$ is a degree $t$ polynomial. \end{lemma} According to Lemma \ref{lm:boundft1}, to bound $\|F_{t+1}\|_2$, it is sufficient to bound the right-hand-side (RHS) of (\ref{eq:chebh}) (i.e., $\min_{\beta}\max_{\lambda\in [\mu,L]}|H_t(\lambda)|$). So we want to choose parameter $\beta$ in order to make $\max_{\lambda\in [\mu,L]}|H_t(\lambda)|$ as small as possible. According to (\ref{eq:chebmin}) (the minimal deviation property of standard Chebyshev polynomials), hence a natural idea is to choose $\beta$ such that $H_t(\lambda) =(1-\beta_t\lambda)\dotsm(1-\beta_1\lambda)$ is a kind of modified Chebyshev polynomials. In order to do this, we first transform $[\mu,L]$ into $[-1,1]$, i.e., let $\lambda=\frac{L+\mu}{2}+\frac{L-\mu}{2}x$, where $x\in[-1,1]$. Also note that polynomial $H_t(\lambda) =(1-\beta_t\lambda)\dotsm(1-\beta_1\lambda)$ has (only) one constraint, i.e., $H_t(0)=1$. Thus we choose $\beta$ such that \begin{align} H_t(\lambda)&=P_t\Big(\frac{2\lambda-(L+\mu)}{L-\mu}\Big)\Big/P_t\Big(-\frac{L+\mu}{L-\mu}\Big) \notag\\ &=P_t(x)\big{/}P_t\Big(-\frac{L+\mu}{L-\mu}\Big), \label{eq:trh} \end{align} where $P_t(\cdot)$ is the standard Chebyshev polynomials. Now, the RHS of (\ref{eq:chebh}) can be bounded as follows: \begin{align} &\min_{\beta}\max_{\lambda\in [\mu,L]}|H_t(\lambda)| \notag\\ &\leq \max_{x\in [-1,1]}\Big|P_t(x)\big{/}P_t\Big(-\frac{L+\mu}{L-\mu}\Big)\Big| \label{eq:trhcall}\\ &\leq 1\Big{/}\Big|P_t\Big(-\frac{L+\mu}{L-\mu}\Big)\Big|, \label{eq:ptx} \end{align} where (\ref{eq:trhcall}) uses (\ref{eq:trh}), and (\ref{eq:ptx}) uses $\max_{x\in[-1,1]}|P_t(x)|=1$ (see (\ref{eq:cheb})). According to (\ref{eq:root}), it is not hard to see that $H_t(\lambda)$ is defined by the mixing parameters $\beta_i=1\Big/\Big(\frac{L+\mu}{2}+\frac{L-\mu}{2}\cos\big(\frac{(2i-1)\pi}{2t}\big)\Big)$ according to $\lambda=\frac{L+\mu}{2}+\frac{L-\mu}{2}x$, where $i=1,2,\ldots,t$. Note that the roots of standard Chebyshev polynomials (i.e., (\ref{eq:root})) can be found from many textbooks, e.g., Section 1.2 of \citep{rivlin1974chebyshev}. Now, we only need to bound $\big|P_t\big(-\frac{L+\mu}{L-\mu}\big)\big|$. First, we need to transform the form (\ref{eq:cheb}) of Chebyshev polynomials $P_t(x)$ as follows: \begin{align*} P_t(x)&=\cos(t\arccos x)\\ &=\cos(t\theta) \qquad \mathrm{Define}~x\triangleq\cos\theta\\ &=\left(e^{i\theta t}+e^{-i\theta t}\right)/2 \\ &=\left((\cos\theta+i\sin\theta)^t+(\cos\theta-i\sin\theta)^t\right)/2\\ &=\Big(\big(x+\sqrt{x^2-1}\big)^t+\big(x-\sqrt{x^2-1}\big)^t\Big)/2. \end{align*} Let $x=-\frac{L+\mu}{L-\mu}$, we get $\sqrt{x^2-1}=\sqrt{\frac{(L+\mu)^2-(L-\mu)^2}{(L-\mu)^2}}=\sqrt{\frac{4L\mu}{(L-\mu)^2}} =\frac{2\sqrt{L\mu}}{L-\mu}$. So we have \begin{align \Big|P_t\Big(-\frac{L+\mu}{L-\mu}\Big)\Big| &\geq \frac{1}{2}\Big(\frac{L+\mu}{L-\mu}+\frac{2\sqrt{L\mu}}{L-\mu}\Big)^t \notag\\ &\geq \frac{1}{2}\Big(\frac{\sqrt{L}+\sqrt{\mu}}{\sqrt{L}-\sqrt{\mu}}\Big)^t. \label{eq:thx1} \end{align} Now, the RHS of (\ref{eq:chebh}) can be bounded as \begin{align} \sqrt{2\min_{\beta}\max_{\lambda\in [\mu,L]}|H_t(\lambda)|} &\leq \sqrt{2\big{/}\Big|P_t\Big(-\frac{L+\mu}{L-\mu}\Big)\Big|} \label{eq:thx2}\\ & \leq 2\Big(\frac{\sqrt{L}-\sqrt{\mu}}{\sqrt{L}+\sqrt{\mu}}\Big)^{t/2}, \label{eq:thx3} \end{align} where (\ref{eq:thx2}) follows from (\ref{eq:ptx}), and (\ref{eq:thx3}) follows from (\ref{eq:thx1}). Then, according to (\ref{eq:chebh}), the gradient norm is bounded as $\|\nabla f(x_{t+1})\|_2=\|F_{t+1}\|_2 \leq 2\big(\frac{\sqrt{L}-\sqrt{\mu}}{\sqrt{L}+\sqrt{\mu}}\big)^{t/2} \|F_1\|_2 =2\big(\frac{\sqrt{\kappa}-1}{\sqrt{\kappa}+1}\big)^{t/2} \|\nabla f(x_1)\|_2$, where $\kappa=L/\mu$. Note that if the number of iterations $t=(\sqrt{\kappa}+1)\ln\frac{1}{\epsilon}$, then \begin{align*} \big(\frac{\sqrt{\kappa}-1}{\sqrt{\kappa}+1}\big)^{t/2}=\big(1-\frac{2}{\sqrt{\kappa}+1}\big)^{t/2} \leq \epsilon. \end{align*} Thus the Anderson-Chebyshev acceleration method achieves the optimal convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ for obtaining an $\epsilon$-approximate solution. \end{proofof} \section{The General Case} \label{sec:gel} In this section, we analyze the Anderson Acceleration (Algorithm \ref{alg:am}) in the general nonlinear case: \begin{equation}\label{prob:gel} \min_{x \in {\mathbb R}^d} f(x). \end{equation} We prove that Anderson acceleration method achieves the linear-quadratic convergence rate under the following standard Assumptions \ref{asp:1} and \ref{asp:2}, where $\|\cdot\|$ denotes the Euclidean norm. Let $\mathcal{B}_t$ denote the small matrix of the least-square problem in Line 7 of Algorithm \ref{alg:am}, i.e., $\mathcal{B}_t\triangleq[F_{t}-F_{t-1}, \dotsc, F_t-F_{t-m}] \in \mathbb{R}^{d\times m}$ (see problem (\ref{eq:transt})). Then, we define its condition number $\kappa_t\triangleq\|\nabla f(x_t)\|/\tilde{\mu}_t$ and $\tilde{\kappa} \triangleq \max_t\{\kappa_t\}$, where $\tilde{\mu}_t$ denotes the least non-zero singular value of $\mathcal{B}_t$. \begin{assumption}\label{asp:1} The Hessian $\nabla^2f$ satisfies $\mu\leq \|\nabla^2f\| \leq L$, where $0\leq \mu \leq L$. \end{assumption} \begin{assumption}\label{asp:2} The Hessian $\nabla^2f$ is $\gamma$-Lipschitz continuous, i.e., \begin{equation}\label{asp3} \|\nabla^2f(x) - \nabla^2f(y)\| \leq \gamma \|x-y\|. \end{equation} \end{assumption} \begin{theorem}\label{thm:gel} Suppose Assumption \ref{asp:1} and \ref{asp:2} hold. Let step-size $\lambda=\frac{2}{L+\mu}$. The convergence rate of Anderson Acceleration($m$) (Algorithm \ref{alg:am}) is linear-quadratic for problem (\ref{prob:gel}), i.e., \begin{equation}\label{eq:converge} \|\nabla f(x_{t+1})\| \leq c_1\Delta_t^2 + c_2\Delta_t\|\nabla f(x_t)\| + (1-c_3)\|\nabla f(x_t)\|, \end{equation} where $c_1=\frac{3\tilde{\kappa}^2\gamma m}{(L+\mu)^2}$, ~$c_2=\frac{2\tilde{\kappa}\beta_t\gamma\sqrt{m}}{(L+\mu)^2}$, ~$c_3=\beta_t\frac{2\mu}{L+\mu}$ and $\Delta_t \triangleq \max_{i\in[m]}\|x_t-x_{t-i}\|$. \end{theorem} \noindent\textbf{Remark:}\vspace{-3mm} \begin{enumerate} \item The constant $m\geq 0$ is usually very small. Particularly, we use $m=3$ and $5$ for the numerical experiments in Section \ref{sec:exp}. Hence $\Delta_t$ is very small and also decreases as the algorithm converges. \item Besides, one can also use $\|\nabla f(x_t)\|$ instead of $\Delta_t$ in (\ref{eq:converge}) according to the property of $f$ (Assumption \ref{asp:1}), i.e., $\mu\|x_t-x^*\| \leq\|\nabla f(x_t)-\nabla f(x^*)\| =\|\nabla f(x_t)\|$, and $\|x_t-x_{t-i}\| = \|x_t-x^*+x^*-x_{t-i}\| \leq \|x_t-x^*\|+\|x_{t-i}-x^*\|$. \item Note that the first two terms in RHS of (\ref{eq:converge}) converge quadratically and the last term converges linearly. Due to the fully dynamic property of Anderson acceleration as we discussed in Section \ref{sec:rw}, it turns out the exact convergence rate of Anderson acceleration in the general case is not easy to obtain. But we note that the convergence rate is roughly linear, i.e., $O(\frac{1}{c_3}\log\frac{1}{\epsilon})$ since the first two quadratic terms converge much faster than the last linear term in some neighborhood of optimum. In particular, if $f$ is a quadratic function, then $\gamma =0$ (Assumption \ref{asp:2}) and thus $c_1=c_2=0$ in (\ref{eq:converge}). Only the last linear term remained, thus it converges linearly (see the following corollary). \end{enumerate} \vspace{-2mm} \begin{corollary}\label{cor:quad} If $f$ is a quadratic function, let step-size $\lambda=\frac{2}{L+\mu}$ and $\beta_t=1$. Then the convergence rate of Anderson Acceleration is linear, i.e., $O(\kappa\ln\frac{1}{\epsilon})$, where $\kappa=L/\mu$ is the condition number. \end{corollary} \vspace{-1mm} Note that this corollary recovers the previous result (i.e., $O(\kappa\ln\frac{1}{\epsilon})$) obtained by \citep{toth2015convergence}, and we use \emph{Chebyshev polyniomial} to improve this result to the optimal convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ in our previous Section \ref{sec:opt} (see Theorem \ref{thm:opt}). Concretely, we transfer the weight of step-size $\lambda$ to the parameters $\beta_t$'s and use Chebyshev polynomial parameters $\beta_t$'s in our Theorem \ref{thm:opt} instead of using fixed parameter $\beta \equiv 1$. \vspace{1mm} Now, we provide a proof sketch for Theorem \ref{thm:gel}. The detailed proof can be found in Appendix \ref{app:pfthm1}. \noindent{\em Proof Sketch of Theorem \ref{thm:gel}.} Consider the iteration $t+1$, we have $F_t=-\frac{2}{L+\mu}\nabla f(x_t)$ according to $\lambda =\frac{2}{L+\mu}$. First, we need to demonstrate several useful forms of $x_{t+1}$ as follows: \begin{align} x_{t+1}&= (1-\beta_t)\sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}} + \beta_t\sum_{i=0}^{m_t}{\alpha_i^t G(x_{t-i})} \notag \\ &=\sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}} + \beta_t\sum_{i=0}^{m_t}{\alpha_i^t\Bigl(G(x_{t-i})-x_{t-i}\Bigr)} \notag \\ &=\sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}} + \beta_t\sum_{i=0}^{m_t}{\alpha_i^tF_{t-i}} \label{eq:x3}\\ &=x_t-\sum_{i=1}^{m_t}{\alpha_i^t (x_t-x_{t-i})} \notag\\ &\qquad\quad+ \beta_t\Big(F_t-\sum_{i=1}^{m_t}{\alpha_i^t(F_t-F_{t-i})}\Big), \label{eq:x2} \end{align} where (\ref{eq:x3}) holds due to the definition $G_t=G(x_t)=x_t+F_t$, and (\ref{eq:x2}) holds since $\sum_{i=0}^{m_t}{\alpha_i^t=1}$. Then, to bound $\|F_{t+1}\|_2$ (i.e., $\|\nabla f(x_{t+1})\|_2$), we deduce $F_{t+1}$ as follows: \begin{align} F_{t+1} &= G_{t+1} - x_{t+1} \notag \\ &= G_{t+1} - \sum_{i=0}^{m_t}{\alpha_i^t x_{t-i}} - \beta_t\sum_{i=0}^{m_t}{\alpha_i^tF_{t-i}} \notag \\ &= G_{t+1} - \sum_{i=0}^{m_t}{\alpha_i^t (G_{t-i}-F_{t-i})} - \beta_t\sum_{i=0}^{m_t}{\alpha_i^tF_{t-i}} \notag \\ &= G_{t+1} - \sum_{i=0}^{m_t}{\alpha_i^t G_{t-i}} + (1-\beta_t)\mathcal{F}, \label{eq:sf2} \end{align} where (\ref{eq:sf2}) uses the definition $\mathcal{F} \triangleq \sum_{i=0}^{m_t}{\alpha_i^tF_{t-i}}$. Now, we bound the first two terms of (\ref{eq:sf2}) as follows: \begin{align} &G_{t+1} - \sum_{i=0}^{m_t}{\alpha_i^t G_{t-i}} \notag\\ &= G_{t+1} - \bigl(G_{t} - \sum_{i=1}^{m_t}{\alpha_i^t}(G_{t}-G_{t-i})\bigr) \notag\\ &= \int_0^1G'\Big(x_t+u(x_{t+1}-x_t)\Big)(x_{t+1}-x_t)\,du \notag\\ &\quad~ -\sum_{i=1}^{m_t}{\alpha_i^t}\int_0^1G'\Bigl(x_t+u(x_{t-i}-x_t)\Bigr)(x_{t-i}-x_t)\,du \notag\\ &= \sum_{i=1}^{m_t}{\alpha_i^t}\int_0^1G'\Bigl(x_t+u(x_{t+1}-x_t)\Bigr)(x_{t-i}-x_t)\,du \notag\\ &\quad~ + \int_0^1G'\Bigl(x_t+u(x_{t+1}-x_t)\Bigr)\beta_t\mathcal{F}\,du \notag\\ &\quad~ -\sum_{i=1}^{m_t}{\alpha_i^t}\int_0^1G'\Bigl(x_t+u(x_{t-i}-x_t)\Bigr)(x_{t-i}-x_t)\,du, \label{eq:sint} \end{align} where (\ref{eq:sint}) is obtained by using (\ref{eq:x2}) to replace $x_{t+1}$. To bound (\ref{eq:sint}), we use Assumptions \ref{asp:1}, \ref{asp:2}, and the equation \begin{equation*} G_t' = I+F_t' = I - \frac{2}{L+\mu}\nabla^2f(x_t). \end{equation*} After some non-trivial calculations (details can be found in Appendix \ref{app:pfthm1}), we obtain \begin{align*} \|F_{t+1}\| &\leq \frac{\gamma(m\ns{\alpha}+\sqrt{m}\n{\alpha})\Delta_t^2}{L+\mu} + \frac{\gamma\sqrt{m}\n{\alpha}\beta_t\Delta_t\|\mathcal{F}\|}{L+\mu} \\ & \quad\qquad +\Bigl(1-\frac{2\mu}{L+\mu}\beta_t\Bigr)\|\mathcal{F}\|, \end{align*} where $\n{\alpha}$ denotes the Euclidean norm of $\alpha=(\alpha_1^t,\ldots,\alpha_{m_t}^t)^T$. Then, according to the problem (\ref{eq:transt}) and the definition of $\mathcal{F} \triangleq \sum_{i=0}^{m_t}{\alpha_i^tF_{t-i}}$, we have $\|\mathcal{F}\| \leq \|F_t\|$. Finally, we bound $\n{\alpha}\leq\frac{2\tilde{\kappa}}{L+\mu}$ using QR decomposition of problem (\ref{eq:transt}) and recall $F_t=-\frac{2}{L+\mu}\nabla f(x_t)$ to finish the proof of Theorem \ref{thm:gel}. \hfill\ensuremath{\square} \section{Guessing Algorithm} \label{sec:guess} In this section, we provide a \emph{Guessing Algorithm} (described in Algorithm \ref{alg:guess}) which guesses the parameters (e.g., $\mu, L$) dynamically. Intuitively, we guess the parameter $\mu$ and the condition number $\kappa$ in a doubling way. Note that in general these parameters are not available, since the time for computing these parameters is almost the same as (or even longer than) solving the original problem. Also note that the condition in Line 14 of Algorithm \ref{alg:guess} depends on the algorithm used in Line 12. \begin{algorithm}[!h] \caption{Guessing Algorithm} \label{alg:guess} \textbf{input:} $x_0, T, \delta, B$\\ Let $t=0$\; \For{$i= 1, 2,\ldots $}{ $\kappa_i=e^{i+2}$\; \For{$j=1,2,\ldots, \ln B$}{ $\mu_i=e^j\delta, L_i = \mu_i\kappa_i, t_i = 1$\; \DoWhile{ $\frac{\|\nabla f(x_t)\|}{\|\nabla f(x_{t-1})\|}\leq 2\left(\frac{\sqrt{\kappa_i}-1}{\sqrt{\kappa_i}+1}\right)^{t_i}$}{ $t_i = \lfloor et_i\rfloor$\; \If{$t+t_i>T$}{ break\; } $x_{t-1}= x_t$\; $x=$Anderson Acceleration($x_t,t_i,\mu_i, L_i$) //can be replaced by other algorithms\; $t = t+t_i, x_t = x$\; } \If{$\|\nabla f(x_t)\| > \|\nabla f(x_{t-1})\|$}{ $x_t = x_{t-1}$\; } } } \Return $x_t$ \end{algorithm} The convergence result of our Algorithm \ref{alg:guess} is stated in the following Theorem \ref{thm:guess}. The detailed proof is deferred to Appendix \ref{app:pfthm3}. Note that we only prove the quadratic case for Theorem \ref{thm:guess}, but it is possible to extend it to the general case. \begin{theorem} \label{thm:guess} Without knowing the parameters $\mu$ and $L$, Algorithm \ref{alg:guess} achieves $O(\sqrt{\kappa}\ln\frac{1}{\epsilon}+\sqrt{\kappa}(\ln\kappa\ln B)^2)$ convergence rate for obtaining an $\epsilon$-approximate solution of problem (\ref{prob:quad}), where $\kappa=L/\mu$, and $B$ can be any number as long as the eigenvalue spectrum belongs to $[\delta,B\delta]$. \end{theorem} \noindent\textbf{Remark:} We provide a simple example to show why this guessing algorithm is useful. Note that algorithms usually need the (exact) parameters $\mu$ and $L$ to set the step size. Without knowing the exact values $\mu$ and $L$, one needs to approximate these parameters once at the beginning. Let $\mu'=\frac{1}{c_1}\mu$ and $L'=c_2L$ denote the approximated values, where $c_1, c_2\geq 1$. Without guessing them dynamically, one fixes $\mu'$ and $L'$ all the time in its algorithm. According to the lower bound $\Omega(\sqrt{\kappa}\ln\frac{1}{\epsilon})$, we know that its convergence rate cannot be better than $O(\sqrt{\kappa'}\ln \frac{1}{\epsilon})=O(\sqrt{c_1c_2\kappa}\ln \frac{1}{\epsilon})$, where $\kappa'=L'/\mu'$. However, if one combines with our Algorithm \ref{alg:guess} (guessing the parameters dynamically), the convergence rate can be improved to $O(\sqrt{\kappa}\ln\frac{1}{\epsilon}+\sqrt{\kappa}(\ln\kappa\ln (c_1c_2\kappa))^2)$ according to our Theorem \ref{thm:guess} by letting $\delta=\mu'$ and $B\delta=L'$ (hence $B=c_1c_2\kappa$). Note that there is no $\epsilon$ (accuracy) in the second term $\sqrt{\kappa}(\ln\kappa\ln (c_1c_2\kappa))^2$. Thus the rate turns to the optimal $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$ when $\epsilon\rightarrow 0$. To achieve an $\epsilon$-approximate solution, our guessing algorithm can improve the convergence a lot especially for an imprecise estimate at the beginning (i.e., $c_1$ and $c_2$ are very large). The corresponding experimental results in Section \ref{app:expga} (see Figure~\ref{fig:7}) indeed validate our theoretical results. \section{Experiments} \label{sec:exp} In this section, we conduct the numerical experiments on the real-world UCI datasets\footnote{The UCI datasets can be downloaded from \url{https://archive.ics.uci.edu/ml}} and synthetic datasets. We compare the performance among these five algorithms: Anderson Acceleration (AA), Anderson-Chebyshev acceleration (AA-Cheby), vanilla Gradient Descent (GD), Nesterov's Accelerated Gradient Descent (NAGD) \citep{nesterov2014introductory} and Regularized Minimal Polynomial Extrapolation (RMPE) with $k=5$ (same as \citep{scieur2016regularized}). Regarding the hyperparameters, we directly set them from their corresponding theoretical results. See Proposition 1 of \citep{lessard2016analysis} for GD and NAGD. For RMPE5, we follow the same setting as in \citep{scieur2016regularized}. For our AA/AA-Cheby, we set them according to our Theorem \ref{thm:opt} and \ref{thm:gel}. Figure~\ref{fig:lreg} demonstrates the convergence performance of these algorithms in general nonlinear case and Figures~\ref{fig:1}--\ref{fig:5} demonstrate the convergence performance in quadratic case. The last Figure~\ref{fig:7} demonstrates the convergence performance of these algorithms combined with our guessing algorithm (Algorithm \ref{alg:guess}). The values of $m$ in the caption of figures denote the mixing parameter of Anderson acceleration algorithms (see Line 5 of Algorithm \ref{alg:am}). \begin{figure}[!htb] \centering \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{LReg_diabetes.pdf} \end{minipage}% \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{LReg_cancer.pdf} \end{minipage} \caption{Logistic regression, $m=3$} \label{fig:lreg} \end{figure} In Figure~\ref{fig:lreg}, we use the negative log-likelihood as the loss function $f$ (logistic regression), i.e., $f(\theta)= - \sum_{i=1}^n(y_i\log\phi(\theta^Tx_i) + (1-y_i)\log(1-\phi(\theta^Tx_i)))$, where $\phi(z)= 1/(1+\exp(-z))$. We run these five algorithms on real-world \emph{diabetes} and \emph{cancer} datasets which are standard UCI datasets. The x-axis and y-axis represent the number of iterations and the norm of the gradient of loss function respectively. \begin{figure}[!htb] \centering \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_1.pdf} \end{minipage}% \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_2.pdf} \end{minipage} \caption{$\kappa\in[0,500]$; $m=3$ (left), $m=5$ (right)} \label{fig:1}\vspace{-3mm} \end{figure} \begin{figure}[!htb] \centering \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_3.pdf} \end{minipage}% \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_4.pdf} \end{minipage}\\ \caption{$\kappa\in[500,2000]$; $m=3$ (left), $m=5$ (right)} \label{fig:3}\vspace{-3mm} \end{figure} \begin{figure}[!htb] \centering \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_5.pdf} \end{minipage}% \begin{minipage}[htb]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_6.pdf} \end{minipage}\\ \caption{$\kappa\in[2000,5000]$; $m=3$ (left), $m=5$ (right)} \label{fig:5}\vspace{-2mm} \end{figure} Figures~\ref{fig:1}--\ref{fig:5} demonstrate the convergence performance for the quadratic case, where $f(x)=\frac{1}{2}x^TAx-b^Tx$. Concretely, we compared the convergence performance among these algorithms when the condition number $\kappa(A)$ and the mixing parameter $m$ are varied, e.g., the left figure in Figure \ref{fig:1} is the case $\kappa\in[0,500]$ and $m =3$. Recall that $m$ is the mixing parameter for Anderson acceleration algorithms (see Line 5 of Algorithm \ref{alg:am}). We run these five algorithms on the synthetic datasets in which we randomly generate the $A$ and $b$ for the loss function $f$. Note that for randomly generated $A$ satisfying the property of $A\in \mathcal{S}_{++}^d$, we randomly generate $B$ instead and let $A \triangleq B^TB$. In conclusion, Anderson acceleration methods converge the fastest no matter it is a quadratic function or general function in all of our experiments. The efficient Anderson acceleration methods can be viewed as the extension of momentum methods (e.g., NAGD) since GD is the special case of Anderson Acceleration with $m=0$, and to some extent NAGD can be viewed as $m=1$. Combined with our theoretical results (i.e., optimal convergence rate in quadratic case and linear-quadratic convergence in general case), the experimental results validate that Anderson acceleration methods are efficient both in theory and practice. \subsection{Experiments for Guessing Algorithm} \label{app:expga} In this section, we conduct the experiments for guessing the hyperparameters (i.e., $\mu, L$) dynamically using our Algorithm \ref{alg:guess}. \begin{figure}[!htb] \centering \begin{minipage}[h]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_gd_guess.pdf} { (a) Gradient Descent} \end{minipage}% \begin{minipage}[h]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_nagd_guess.pdf} {\quad (b) Nesterov's AGD} \end{minipage}\\ \begin{minipage}[h]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_am_fix_guess.pdf} {\quad (c) Anderson Acceleration} \end{minipage}% \begin{minipage}[h]{0.25\textwidth} \includegraphics[width=\textwidth]{fig_am_cheby_guess.pdf} {\quad (d) Anderson-Chebyshev} \end{minipage} \caption{Algorithms with/without guessing algorithm} \label{fig:7 \end{figure} In Figure \ref{fig:7}, we separately consider these algorithms. For each of them, we compare its convergence performance between its original version and the one combined with our guessing algorithm (Algorithm \ref{alg:guess}). The experimental results show that all these four algorithms combined with our guessing algorithm achieve much better performance than their original versions. Thus it validates our theoretical results (see Theorem \ref{thm:guess} and its following Remark). \section{Conclusion} In this paper, we prove that Anderson acceleration with Chebyshev polynomial can achieve the optimal convergence rate $O(\sqrt{\kappa}\ln\frac{1}{\epsilon})$, which improves the previous result $O(\kappa\ln\frac{1}{\epsilon})$ provided by \citep{toth2015convergence}. Thus it can deal with ill-conditioned problems (condition number $\kappa$ is large) more efficiently. Furthermore, we also prove the linear-quadratic convergence of Anderson acceleration for minimizing general nonlinear problems. Besides, if the hyperparameters (e.g., the Lipschitz smooth parameter $L$) are not available, we propose a guessing algorithm for guessing them dynamically and also prove a similar convergence rate. Finally, the experimental results demonstrate that the efficient Anderson acceleration methods converge significantly faster than other algorithms. This validates that Anderson-Chebyshev acceleration is efficient both in theory and practice. \newpage \subsubsection*{Acknowledgements} Zhize was supported by the Office of Sponsored Research of KAUST, through the Baseline Research Fund of Prof. Peter Richt{\'a}rik. Jian was supported in part by the National Natural Science Foundation of China Grant 61822203, 61772297, 61632016, 61761146003, and the Zhongguancun Haihua Institute for Frontier Information Technology and Turing AI Institute of Nanjing. The authors also would like to thank Francis Bach, Claude Brezinski, Rong Ge, Damien Scieur, Le Zhang and anonymous reviewers for useful discussions and suggestions. \bibliographystyle{plainnat}

0b1b638ba1c2b62d5c77a239027a910f82c40198

\section{Introduction} Recently, Ozawa and Berthier studied a problem\cite{ozawa_does_2017} of diverging and discontinuous configurational entropy per particle $S_{\text{conf}} / N$ in systems with continuous polydispersity. Divergence happens due to the diverging and discontinuous (with respect to small compositional changes) mixing entropy per particle $S_{\text{mix}} / N$. Specifically, the total entropy $S_{\text{tot}}$ contains the mixing entropy $S_{\text{mix}}$, while the vibrational entropy $S_{\text{vib}}$ does not. Thus, the configurational entropy $S_{\text{conf}}$ from the relation\cite{speedy_hard_1998, stillinger_kauzmann_2001, sastry_relationship_2001, angelani_configurational_2007, donev_configurational_2007, foffi_short_2008, starr_relationship_2013} $S_{\text{tot}} = S_{\text{conf}} + S_{\text{vib}}$ shall contain $S_{\text{mix}}$. It is problematic because some observables are believed to depend on $S_{\text{conf}}$\cite{adam_temperature_1965, cavagna_supercooled_2009, starr_relationship_2013, kirkpatrick_random_2015, cavagna_supercooled_2009, starr_relationship_2013, bouchaud_adam_2004} but from physical considerations shall be continuous with slight updates of the system composition, as well as finite and non-zero for continuous polydispersity. Ozawa and Berthier\cite{ozawa_does_2017} suggested a redefinition of $S_{\text{vib}}$ (a ``scheme'' of merging the basins of attraction) so that $S_{\text{mix}}$ is moved into $S_{\text{vib}}$ and $S_{\text{conf}}$ thus remains finite. The authors refined their approach and provided a more robust and unambiguous procedure in a follow-up paper.\cite{ozawa_configurational_2018} Their approach essentially defines $S_{\text{conf}}$ through the free energy landscape (FEL).\cite{du_energy_2016, ikeda_note_2016} In an attempt to resolve this paradox in the potential energy landscape (PEL), we studied\cite{baranau_another_2017} $S_{\text{vib}}^{\text{PEL}}$ using a variant of thermodynamic integration\cite{frenkel_new_1984, frenkel_understanding_2002, stillinger_kauzmann_2001, donev_calculating_2007, donev_configurational_2007, asenjo_numerical_2014} and suggested that it is in fact normal for $S_{\text{conf}}^{\text{PEL}}$ to contain $S_{\text{mix}}$. We suggested that all the observables that depend on $S_{\text{conf}}^{\text{PEL}}$ and are believed to be well-behaving shall instead depend on $S_{\text{conf}}^{\text{PEL}} - S_{\text{mix}}$ or in general $S_{\text{conf}}^{\text{PEL}} - \min( S_{\text{conf}}^{\text{PEL}} )$. For this, we implicitly assumed that $\min( S_{\text{conf}}^{\text{PEL}} )$ can be non-zero, in particular $S_{\text{mix}}$, \textit{i.e.}, that $S_{\text{conf}}^{\text{PEL}}$ is bounded from below solely by the system composition. In this paper, we investigate the lower bound on $S_{\text{conf}}^{\text{PEL}}$ in more detail and explicitly demonstrate that for some systems $S_{\text{conf}}^{\text{PEL}}$ is indeed bounded from below by $S_{\text{mix}}$. Additionally, we investigate connections between $S_{\text{conf}}^{\text{PEL}}$ and $S_{\text{conf}}$ defined through the free energy landscape, $S_{\text{conf}}^{\text{FEL}}$. We show that for our model systems $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$ and that the condition $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$ implies for an arbitrary system $\min( S_{\text{conf}}^{\text{PEL}} ) = N A + S_{\text{mix}}$, where $A$ is some constant. We discuss which implications it has on the Adam--Gibbs (AG) and RFOT relations and show that they retain a physically meaningful shape for both configurational entropies, $S_{\text{conf}}^{\text{FEL}}$ and $S_{\text{conf}}^{\text{PEL}}$. We show that the two versions of the AG relation will differ by a well-behaving discrepancy term. \section{Preliminaries} \subsection{Different entropies} We study a canonical $NVT$ ensemble of classical systems and use a standard notation with $N$, $V$, $T$, $U$, $d$, and $\Lambda$ standing for the number of particles, volume, temperature, internal energy, dimensionality, and de Broglie wavelength, respectively. $S$, $F$, $Z$, and $U_N(\vec{r})$ stand for the entropy, Helmholtz free energy, partition function, and total potential energy in the configuration $\vec{r}$. We study polydisperse systems and denote with $N_m$ the number of particles of type $m$. One can routinely derive from standard equations $S = (U - F) / T$, $F = -k_B T \ln Z$, and $Z = \frac{1}{\Pi_{i=1}^M N_i!} \frac{1}{\Lambda^{d N}} \int_{V^N} e^{-U_N(\vec{r})/k_B T} \text{d} \vec{r}$ for the total entropy of the system.\cite{carnahan_thermodynamic_1970, adams_chemical_1974, speedy_hard_1998, donev_calculating_2007} It can be shown\cite{ozawa_does_2017, baranau_another_2017} that such an entropy will contain the mixing entropy $S_{\text{mix}} = - k_B N \sum_i^M \frac{N_i}{N} \ln \frac{N_i}{N}$. It can also be shown that $S_{\text{mix}} / N$ diverges for continuous particle type distributions (in the simplest case, $N_i = 1~\forall i$, $S_{\text{mix}} = k_B N \ln N$).\cite{ozawa_does_2017, baranau_another_2017} $S_{\text{mix}} / N$ is also obviously discontinuous if the system composition is abruptly changed. If we define the configurational entropy through the PEL, $S_{\text{conf}}^{\text{PEL}}$ counts (up to particle permutations) the density of potential energy minima\cite{donev_configurational_2007, ashwin_calculations_2012, asenjo_numerical_2014, martiniani_turning_2016, stillinger_systematic_1964, stillinger_topographic_1995, torquato_robust_2010} $N_{\text{min}}(N, u_{\text{min}})$ at a given energy per particle $u_{\text{min}} = U_{\text{min}} / N$. Specifically, $N_{\text{min}}(N, u_{\text{min}}) = e^{ S_{\text{conf}}^{\text{PEL}}(N, u_{\text{min}}) } / w$, where $w$ is the characteristic interval that comes from the saddle point approximation.\cite{speedy_hard_1998, parisi_ideal_2005, parisi_mean_field_2010, berthier_microscopic_2011} It is believed to be subexponential, \textit{i.e.}, $\ln(w)/N \to 0$ with $N \to \infty$. The vibrational partition function in the PEL is expressed as $Z_{\text{vib}}^{\text{PEL}} = \frac{\Pi_{i=1}^M N_i!}{\Pi_{i=1}^M N_i!} \frac{1}{\Lambda^{d N}} \int_{\text{basin}} e^{-U_N(\vec{r})/k_B T} \text{d} \vec{r}$, where integration is performed over a certain basin of attraction\cite{stillinger_systematic_1964, stillinger_topographic_1995, debenedetti_supercooled_2001, torquato_robust_2010} in the PEL. The $\Pi_{i=1}^M N_i!$ term in the denominator comes from particle indistinguishability as for the total partition function.\cite{frenkel_why_2014, meng_free_2010, asenjo_numerical_2014, martiniani_turning_2016} The same term in the nominator comes from basin multiplicity, \textit{i.e.}, indistinguishability of basins due to particle indistinguishability.\cite{stillinger_limiting_1969, speedy_entropy_1993, speedy_pressure_1998, ozawa_does_2017} Due to cancellation of these two terms, $S_{\text{vib}}^{\text{PEL}}$ does not contain $S_{\text{mix}}$. We showed\cite{baranau_another_2017} on the basis of thermodynamic integration that $S_{\text{mix}}$ is also not hidden in the integral over a basin and that $S_{\text{vib}}^{\text{PEL}}$ still contains the term $\ln \left( \frac{V}{\Lambda^d N} \right)$, which also enters $S_{\text{tot}}$. The three entropies are connected by the relation $S_{\text{tot}} = S_{\text{conf}}^{\text{PEL}} + S_{\text{vib}}^{\text{PEL}}$,\cite{speedy_hard_1998, stillinger_kauzmann_2001, sastry_relationship_2001, angelani_configurational_2007, donev_configurational_2007, foffi_short_2008, starr_relationship_2013} which comes from the saddle point approximation and is exact in the thermodynamic limit if the definitions from above are used.\cite{speedy_hard_1998, baranau_another_2017} In this relation, $S_{\text{conf}}^{\text{PEL}}$ is the equilibrium configurational entropy (equilibrium complexity), \textit{i.e.}, computed for the minima of those basins that are equilibrium (dominant) for the given temperature. At a given temperature $T$ basins with a certain average energy of minima $u_{\text{eq,min}}(T)$ dominate the phase space. Then, the relation for entropies shall fully be read as \begin{equation} \begin{aligned} S_{\text{tot}}(N, V, T) =& S_{\text{conf}}^{\text{PEL}}(N, V, u_{\text{eq,min}}(T)) \\ &+ S_{\text{vib}}^{\text{PEL}}(N, V, u_{\text{eq,min}}(T), T). \label{SeparationOfEntropiesFull} \end{aligned} \end{equation} Because $S_{\text{tot}}$ contains $S_{\text{mix}}$ and $S_{\text{vib}}^{\text{PEL}}$ does not, $S_{\text{mix}}$ shall be contained in $S_{\text{conf}}^{\text{PEL}}$. If the configurational and vibrational entropies are defined through the FEL, the equation also holds. The configurational entropy per particle $S_{\text{conf}} / N$ (however it is defined) enters several relations for the relaxation time $\tau$ of colloids and glassy systems at high volume fraction or low temperature. Both versions of $S_{\text{conf}}$ are used by different authors: either $S_{\text{conf}}^{\text{PEL}}$ or $S_{\text{conf}}^{\text{FEL}}$. The well-known Adam--Gibbs relation reads as $\tau = \tau_0 \exp \left( \frac{C}{S_{\text{conf}} / N} \right)$, where $\tau_0$ and $C$ are presumed to be approximately constant.\cite{adam_temperature_1965, cavagna_supercooled_2009, starr_relationship_2013} The other relation stems from the Random First Order Theory (RFOT).\cite{kirkpatrick_random_2015, cavagna_supercooled_2009, starr_relationship_2013, bouchaud_adam_2004} If $S_{\text{conf}} / N$ is divergent for continuous particle type distributions and discontinuous with a small system composition change, $\tau \equiv 0$ for systems with continuous particle type distributions and $\tau$ will be discontinuous when the system composition is slightly changed. It is paradoxical because $\tau$ shall be finite and continuous. Even worse, particle type distribution can be induced ``artificially'', by assigning arbitrary properties to particles without changing their interaction potential.\cite{frenkel_why_2014} For example, colloidal particles can be colored (dyed) arbitrarily. This will increase $S_{\text{conf}}^{\text{PEL}}$ arbitrarily without changing $\tau$. \paragraph{Hard spheres} A valid collection of $N$ frictionless hard spheres with predefined radii $R_i$ always has a zero potential energy, while an invalid collection has an infinite energy. Still, a pseudo-PEL can be introduced for hard spheres.\cite{torquato_robust_2010, baranau_random_close_2014, zinchenko_algorithm_1994} At any system configuration $\vec{x} \in \mathbb{R}^{3N}$ particle radii can be scaled proportionally to ensure that at least one pair of particles is in contact. The resulting solid volume fraction $\varphi$ defines the pseudo-PEL $U(\vec{x}) = -\varphi$ (or rather, $U(\vec{x}) = -N \varphi$ for $U(\vec{x})$ to be extensive). Local minima in this PEL correspond to mechanically stable (jammed) configurations with some jamming densities $\varphi_J$, while steepest descents correspond\cite{torquato_robust_2010, baranau_random_close_2014, zinchenko_algorithm_1994} to proportionally scaling sphere radii and moving the spheres as little as possible in the configuration space to ensure the absence of intersections. Eq. (\ref{SeparationOfEntropiesFull}) will then look like $S_{\text{tot}}(N, \varphi, T) = S_{\text{conf}}^{\text{PEL}}(N, \varphi_{\text{eq,J}}(\varphi)) + S_{\text{vib}}^{\text{PEL}}(N, \varphi_{\text{eq,J}}(\varphi), T)$. \subsection{Motivation} We argued\cite{baranau_another_2017} that it is normal for $S_{\text{conf}}^{\text{PEL}}$ to be divergent and discontinuous. Still, the relaxation time shall be finite for continuous particle type distributions and shall be continuous when slightly changing particle types or coloring the particles. We suggested that the equations for $\tau$ and for any other ``well-behaving'' observables shall be changed instead. If such an observable depends on $S_{\text{conf}}^{\text{PEL}}$, it shall always depend on $S_{\text{conf}}^{\text{PEL}} - S_{\text{mix}}$ and can not depend on $S_{\text{conf}}^{\text{PEL}}$ only. Thus, arbitrary changes in $S_{\text{conf}}^{\text{PEL}}$ will be compensated. Indeed, if $S_{\text{mix}}$ is the lower bound for $S_{\text{conf}}^{\text{PEL}}$, then only changes with respect to $S_{\text{mix}}$ can matter for physical observables. A more detailed derivation for the updated Adam--Gibbs equation can be found in our previous paper.\cite{baranau_another_2017} For this derivation, we implicitly assumed that $S_{\text{conf}}^{\text{PEL}}$ can not decrease below $S_{\text{mix}}$. In this paper, we provide a more detailed explanation and physical insight into this lower bound for $S_{\text{conf}}^{\text{PEL}}$. We do not discuss in this paper whether $S_{\text{conf}}^{\text{PEL}}$ is a better candidate for such equations than $S_{\text{conf}}^{\text{FEL}}$. But we point out that there indeed is a body of work showing that $S_{\text{conf}}^{\text{PEL}}$ can be a possible candidate for $S_{\text{conf}}$ in the Adam--Gibbs and RFOT relations.\cite{speedy_hard_1998, angelani_configurational_2007, starr_relationship_2013} \section{Main part} In the worst case, the lowest equilibrium complexity $S_{\text{conf}}^{\text{PEL}}(N, V, u_{\text{eq,min}}(T))$ at some low temperature $T_0$ will correspond to only one configuration of particles (up to particle permutations). This configuration will be the local minimum of a basin of attraction that is equilibrium at this $T_0$ and dominates the phase space. If particles are considered distinguishable and there are $N_m$ particles of each type $m$, there are $\Pi_m N_m!$ such local minima. Still, there are $N!$ particle permutations in total. Each of $N!$ configurations (if particles are considered distinguishable) lies in a certain basin of attraction in the potential energy landscape and each of these basins has a corresponding energy minimum (inherent structure). The question is whether these basins will also be equilibrium at $T_0$ or not. \subsection{Mean-field limit} To determine if a certain configuration among these $N!$ is in equilibrium at $T_0$, we compare its chemical potential $\mu$ with the chemical potential from the original equilibrium configuration. To compute the chemical potential, we rely on the Widom particle insertion method.\cite{widom_topics_1963, adams_chemical_1974, speedy_cavities_1981, speedy_cavities_1991, frenkel_understanding_2002, baranau_chemical_2016} This method relates the chemical potential $\mu$ to the potential $\Psi_{N+1, i}(\vec{r})$ that is experienced by a $(N+1)$th test particle of type $i$ if we try to insert this particle into the system at a position $\vec{r}$, sampled uniformly. Specifically, \begin{equation} - \frac{\Delta \mu_i}{kT} = \ln \left[ \left\langle \exp \left( - \frac{ \Psi_{N+1,i} }{k T} \right) \right\rangle \right], \end{equation} where $\Delta \mu_i$ is the excess chemical potential for particles of type $i$ and averaging is performed with respect to $\vec{r}$ of particle insertion positions.\cite{baranau_chemical_2016} If the original configuration is in equilibrium, then we estimate \textit{the} chemical potential (the equilibrium one). Thus, our main focus is how the field $\Psi_{N+1,i}$ changes when we perform any of the $N!$ particle permutations in the initial equilibrium configuration. We do not have an answer to this question in the general case, but this question can be resolved in the mean-field limit. We will denote the potential $\Psi_{N+1, i}(\vec{r})$ perturbed by a permutation $\Sigma_p$ as $\Psi_{\Sigma_p, N+1, i}$. In the mean-field limit, the potential $\Psi_{\Sigma_p, N+1, i}$ does not depend on $\Sigma_p$, \begin{equation} \Psi_{\Sigma_p, N+1, i} \equiv \Psi_{N+1, i}. \end{equation} Thus, all of the $N!$ configurations are equilibrium. The number of such configurations up to particle permutations is $\frac{N!}{\Pi_i N_i!}$ and the lowest possible equilibrium complexity is thus \begin{equation} \min( S_{\text{conf}}^{\text{PEL}} ) = k_B \ln \left( \frac{N!}{\Pi_i N_i!} \right). \end{equation} It can be easily shown with the help of the Stirling approximation $N! \sim N \ln N - N$ that $\min( S_{\text{conf}}^{\text{PEL}} )$ is \begin{equation} \min( S_{\text{conf}}^{\text{PEL}} ) = - k_B N \sum_i \frac{N_i}{N} \ln \frac{N_i}{N} = S_{\text{mix}}. \end{equation} \subsection{Infinitesimally polydisperse systems} Let us assume that a system is monodisperse and the worst case is realized for it, \textit{i.e.}, there is only one equilibrium basin (up to particle permutations) at the lowest temperature $T_0$ where the system can still be equilibrated. Thus, there are $N!$ equilibrium basins and respective minima if particles are considered distinguishable. Let us now assume that particle types are slightly changed, \textit{e.g.}, particle charges that govern pair potentials or radii of colloidal particles are infinitesimally changed. For sufficiently small particle type changes, the structure of the phase space and its basins will remain unchanged. Thus, all of the $N!$ basins will still be equilibrium at this $T_0$--- but there are now $\Pi_m N_m!$ particle permutations that keep the system in the same state, instead of $N!$. Hence, the minimum configurational entropy becomes $\min( S_{\text{conf}}^{\text{PEL}} ) = k_B \ln \left( \frac{N!}{\Pi_i N_i!} \right) = S_{\text{mix}}$. \subsection{$1D$ hard rods} Our last example is very brief and almost obvious. A collection of $1D$ hard rods has $N!$ jammed configurations (pseudo-PEL minima) if particles are considered distinguishable. Because there are $\Pi_m N_m!$ particle permutations if particles are considered indistinguishable, $S_{\text{conf}}^{\text{PEL}} = k_B \ln \left( \frac{N!}{\Pi_i N_i!} \right) = S_{\text{mix}}$. Particle types can stem either from actual rod lengths or from artificial labeling assigned by the observer arbitrarily. This is actually the only value of $S_{\text{conf}}^{\text{PEL}}$ available for $1D$ hard rods. \subsection{Remark: colloids and hard spheres} The procedure outlined for the mean-field case looks as follows for hard spheres. Assume that the largest density at which the system can be equilibrated is denoted as $\varphi_K$. The corresponding jamming density can be denoted as $\varphi_{\text{GCP}}$, the Glass Close Packing density.\cite{parisi_mean_field_2010} Then, one has to take an equilibrium configuration at $\varphi = \varphi_K$ (maybe even the one with $\varphi = \varphi_{\text{GCP}}$ but with particle radii proportionally decreased to achieve $\varphi_K$) and performs particle permutations. Some pairs of particles may intersect after permutations. To avoid permutations, one decreases particle radii proportionally until the closest pair of particles just touches each other; the density is decreased in this process ($\varphi < \varphi_K$). Then, one increases particle radii proportionally and simultaneously moves particles to ensure minimal possible movement in the configuration space (moves particles along contact normals). This movement corresponds to the steepest descent in the pseudo-PEL. One performs this particle scaling until $\varphi = \varphi_K$. Then, one has to check if the resulting configuration is equilibrium. Numerically, this can be done by employing the Widom particle insertion method. If the system becomes jammed before reaching $\varphi_K$, then the new basin (after permutation) was definitely not in equilibrium. \section{Relationship with $S_{\text{conf}}$ defined through the free energy landscape} We would like to discuss how $S_{\text{conf}}^{\text{PEL}}$ relates to the configurational entropy defined in the free energy landscape, $S_{\text{conf}}^{\text{FEL}}$. The free energy landscape to compute $S_{\text{conf}}^{\text{FEL}}$ can stem, \textit{e.g.}, from the density functional theory. We will show that for our model systems the result $\min (S_{\text{conf}}^{\text{PEL}} ) = S_{\text{mix}}$ complies with the requirement $\min (S_{\text{conf}}^{\text{FEL}} ) = 0$ for a certain definition of $S_{\text{conf}}^{\text{FEL}}$.\cite{ozawa_configurational_2018} \subsection{$S_{\text{conf}}^{\text{FEL}}$ of Ozawa \textit{et al.}\cite{ozawa_configurational_2018}} A recent advancement in understanding $S_{\text{conf}}^{\text{FEL}}$ can be found in a paper by Ozawa, Parisi, and Berthier.\cite{ozawa_configurational_2018} Their approach is explicit and mathematically precise and can be used in computer simulations. Essentially, while $S_{\text{vib}}^{\text{PEL}}$ is computed over a certain PEL basin, the authors compute $S_{\text{vib}}^{\text{FEL}}$ by including in the vibrational partition function all possible particle permutations around some reference configurations and weighting the permutations according to their statistical weights. Reference configurations are sampled from the entire configuration space, but are also taken with corresponding statistical weights. We will rely on this approach for the discussion below. We briefly repeat the essential equations from that paper (note that the paper uses $k_B = 1$, so we follow this convention in this section as well). The authors use the modified Frenkel--Ladd method of computing $S_{\text{vib}}$; \textit{i.e.}, they attach particles with springs to the reference positions and integrate over the spring constant $\alpha$. The authors utilize the following types of statistical averages: \begin{eqnarray} \left\langle (\cdots) \right\rangle_{\alpha}^{\rm T, S} &=& \frac{\frac{1}{N!} \sum_{\pi} \int_V \mathrm{d} {\bf r}^N (\cdots) e^{ -\beta U_{\alpha}(\Sigma_{\pi}^N, {\bf r}^N, {\bf r}_0^N)} }{\frac{1}{N!} \sum_{\pi} \int_V \mathrm{d} {\bf r}^N e^{-\beta U_{\alpha}(\Sigma_{\pi}^N, {\bf r}^N, {\bf r}_0^N)}}, \label{eq:T_S} \\ \left\langle (\cdots) \right\rangle_{\alpha}^{\rm T} &=& \frac{\int_V \mathrm{d} {\bf r}^N (\cdots) e^{ - \beta U_{\alpha}({\bf r}^N, {\bf r}_0^N)} }{\int_V \mathrm{d} {\bf r}^N e^{ - \beta U_{\alpha}({\bf r}^N, {\bf r}_0^N)} }, \label{eq:T} \\ \overline{(\cdots)} &=& \frac{\int_V \mathrm{d} {\bf r}_0^N (\cdots) e^{ -\beta U_0({\bf r}_0^N) }}{\int_V \mathrm{d} {\bf r}_0^N e^{ -\beta U_0({\bf r}_0^N) }}. \label{eq:template_average} \end{eqnarray} The superscripts T and S represent the statistical average over positions (T) and permutations (S), respectively. Here, $U_{\alpha}(\Sigma_{\pi}^N, {\bf r}^N, {\bf r}_0^N)$ is the potential energy at a configuration ${\bf r}^N$ around a reference configuration ${\bf r}_0^N$ when particles are additionally permuted by a permutation $\pi$ (note that $\Sigma_{\pi}$, not $\Sigma_{\pi}^N$, represents a usual sum over permutations). The glassy entropy (in our terms $S_{\text{vib}}^{\text{FEL}}$) is expressed according to their approach as \begin{eqnarray} S_{\text{vib}}^{\text{FEL}} &=& S_{\rm glass} = \frac{Nd}{2} - N \ln \Lambda^d - \frac{Nd}{2} \ln \left(\frac{\alpha_{\rm max}}{\pi}\right) \nonumber \\ &\quad& + N \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \Delta_{\alpha}^{\rm T,S} + S_{\rm mix} - \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)}, \nonumber \\ \label{eq:S_glass_final} \end{eqnarray} where $\Delta_{\alpha}^{\rm T, S}$ is a mean-squared displacement defined by \begin{equation} \Delta_{\alpha}^{\rm T, S} = \frac{1}{N} \overline{\left\langle \sum_{i=1}^{N} | {\bf r}_i - {\bf r}_{0 i} |^2 \right\rangle_{\alpha}^{\rm T,S}}, \end{equation} and $\mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)$ is a mixing entropy contribution defined by \begin{equation} \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta) = - \ln \left( \frac{1}{N!} \sum_{\pi} e^{-\beta \left( U_0(\Sigma_{\pi}^N, {\bf r}_0^N) - U_0({\bf r}_0^N) \right) } \right). \label{eq:s_mix_def_new} \end{equation} $S_{\text{conf}}^{\text{FEL}}$ is then given as usual by $S_{\text{tot}} - S_{\text{vib}}^{\text{FEL}}$. Compare Eq. (\ref{eq:S_glass_final}) to a ``usual'' expression for $S_{\text{vib}}^{\text{PEL}}$, when particle permutations are not included in the definition: \begin{eqnarray} S_{\text{vib}}^{\text{PEL}} &=& \frac{Nd}{2} - N \ln \Lambda^d - \frac{Nd}{2} \ln \left(\frac{\alpha_{\rm max}}{\pi}\right) \nonumber \\ &\quad& + N \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \Delta_{\alpha}^{\rm T}. \label{eq:S_vib_PEL} \end{eqnarray} Here, \begin{equation} \Delta_{\alpha}^{\rm T} = \frac{1}{N} \overline{\left\langle \sum_{i=1}^{N} | {\bf r}_i - {\bf r}_{0 i} |^2 \right\rangle_{\alpha}^{\rm T}}. \label{eq:pel_integrand} \end{equation} We note that a typical definition for $S_{\text{vib}}^{\text{PEL}}$ would use $\Delta_{\alpha}^{\rm T} = \frac{1}{N} \left\langle \sum_{i=1}^{N} | {\bf r}_i - {\bf r}_{0 i} |^2 \right\rangle_{\alpha}^{\rm T}$, where ${\bf r}_{0}$ represents a PEL minimum that is equilibrium at a given temperature. It is an approximation for averaging over all equilibrium PEL minima, which is itself an approximation for averaging over all PEL minima, given that they are taken with corresponding statistical weights. We note that integrals in Eqs. (\ref{eq:template_average}) and (\ref{eq:pel_integrand}) will be dominated by (equilibrium) PEL minima, so using Eq. (\ref{eq:pel_integrand}) for $S_{\text{vib}}^{\text{PEL}}$ is legitimate. \subsection{Model systems from our paper} For our examples (a mean-field system and an infinitesimally polydisperse system), particle permutations do not change the potential energy. Hence, $U_{\alpha}(\Sigma_{\pi}^N, {\bf r}^N, {\bf r}_0^N) = U_{\alpha}({\bf r}^N, {\bf r}_0^N)$. It then follows that $\mathcal{S}_{\rm mix}({\bf r}_0^N, \beta) = 0$ (\textit{cf}. Eq. (\ref{eq:s_mix_def_new})) and $\Delta_{\alpha}^{\rm T, S} = \Delta_{\alpha}^{\rm T}$ (\textit{cf}. Eqs. (\ref{eq:T_S}) and (\ref{eq:T})). Hence, for these systems $S_{\text{vib}}^{\text{FEL}} = S_{\text{vib}}^{\text{PEL}} + S_{\text{mix}}$ (\textit{cf}. Eqs. (\ref{eq:S_glass_final}) and (\ref{eq:S_vib_PEL})) and finally $S_{\text{conf}}^{\text{FEL}} = S_{\text{conf}}^{\text{PEL}} - S_{\text{mix}}$. Thus, our bound $\min( S_{\text{conf}}^{\text{PEL}} ) = S_{\text{mix}}$ is actually equivalent to the usual assumption $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$: \begin{equation} \min( S_{\text{conf}}^{\text{FEL}} ) = 0 \iff \min( S_{\text{conf}}^{\text{PEL}} ) = S_{\text{mix}}. \end{equation} \subsection{$\min( S_{\text{conf}}^{\text{PEL}} )$ in the general case} By comparing Eqs. (\ref{eq:S_glass_final}) and (\ref{eq:S_vib_PEL}) and utilizing $S_{\text{tot}} = S_{\text{conf}} + S_{\text{vib}}$, we get in the general case $S_{\text{conf}}^{\text{PEL}} = S_{\text{conf}}^{\text{FEL}} + S_{\text{vib}}^{\text{FEL}} - S_{\text{vib}}^{\text{PEL}}$ and finally \begin{equation} \begin{aligned} S_{\text{conf}}^{\text{PEL}} =& S_{\text{conf}}^{\text{FEL}} + N \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \left( \Delta_{\alpha}^{\rm T, S} - \Delta_{\alpha}^{\rm T} \right) \\ &+ S_{\rm mix} - \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)}. \label{eq:pel_through_fel_init} \end{aligned} \end{equation} One can presume that $\lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \left( \Delta_{\alpha}^{\rm T, S} - \Delta_{\alpha}^{\rm T} \right)$ and $\overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)} $ are decreasing functions of temperature. Indeed, the lower the temperature, the less statistical weight permuted configurations have and hence the closer $\Delta_{\alpha}^{\rm T, S}$ is to $\Delta_{\alpha}^{\rm T}$. The same applies to $\min( \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)} )$. It is usually believed that $S_{\text{conf}}^{\text{FEL}}$ reaches the minimum value when the phase space becomes non-ergodic (at $T = T_K$). It means that $S_{\text{conf}}^{\text{PEL}}$ is defined in the same range of temperatures and is not defined for $T < T_K$. Because Eq. (\ref{eq:pel_through_fel_init}) can be represented as a sum of terms that are either constant or decreasing with $T$ and because $S_{\text{conf}}^{\text{PEL}}$ is defined in the same range of temperatures as $S_{\text{conf}}^{\text{FEL}}$, we conclude that $\min( S_{\text{conf}}^{\text{PEL}} ) $ happens at the same temperature as $\min( S_{\text{conf}}^{\text{FEL}} )$. Hence, we write \begin{equation} \begin{aligned} \min&( S_{\text{conf}}^{\text{PEL}} ) = \min( S_{\text{conf}}^{\text{FEL}} ) \\ &+ N \min( \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \left( \Delta_{\alpha}^{\rm T, S} - \Delta_{\alpha}^{\rm T} \right) ) \\ &+ S_{\rm mix} - \min( \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)} ) \\ &= N A + S_{\rm mix}, \label{eq:s_pel_min_general} \end{aligned} \end{equation} where $A$ is some constant determined purely by particle interactions. A crucial feature in Eq. (\ref{eq:s_pel_min_general}) is that $S_{\rm mix}$ can be increased arbitrarily by infinitesimal changes in particle types or by artificial ``coloring'' of particles (assigning types that do not influence interaction potentials). If the distribution of ``colors'' is continuous, then $S_{\rm mix} / N \to \infty$. Thus, if we presume that $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$, then $\min( S_{\text{conf}}^{\text{PEL}} )$ can be increased arbitrarily for a general system just by additional labeling of particles: \begin{equation} \min( S_{\text{conf}}^{\text{FEL}} ) = 0 \implies \min( S_{\text{conf}}^{\text{PEL}} ) \text{ is arbitrary}. \end{equation} Additionally, $\min( S_{\text{conf}}^{\text{PEL}} ) / N$ in this case is discontinuous with introducing a slight polydispersity to a monodisperse system and divergent for a continuous polydispersity. \subsection{Adam--Gibbs and RFOT relations} As we argued in our previous paper,\cite{baranau_another_2017} arbitrary value of $\min( S_{\text{conf}}^{\text{PEL}})$ and other problems with it are actually not problematic for observables that may depend on $S_{\text{conf}}^{\text{PEL}}$, like the relaxation time $\tau_R$ of glassy systems. If one re-derives the Adam--Gibbs and RFOT relations between $\tau_R$ and $S_{\text{conf}}$ assuming that $\min( S_{\text{conf}} ) \neq 0$, one obtains \begin{equation} \tau_R = f\left( \frac{S_{\text{conf}} - \min( S_{\text{conf}} ) } {N} \right) \label{eq:RelaxationTimes} \end{equation} instead of the usual version $\tau_R = f(S_{\text{conf}} / N)$. For example, the Adam--Gibbs relation shall be transformed from a standard version $\tau_R^{\text{AG}} = \tau_0 \exp\left( \frac{C}{T S_{\text{conf}} / N} \right)$ into \begin{equation} \tau_R^{\text{AG}} = \tau_0 \exp\left( \frac{C}{T \left[ S_{\text{conf}} - \min( S_{\text{conf}}) \right] / N} \right), \end{equation} where $\tau_0$ and $C$ are (approximately) constants. It is interesting to compare what will be the difference between $\tau_R^{\text{PEL}}$ and $\tau_R^{\text{FEL}}$, given that $\tau_R^{\text{FEL}} = f(S_{\text{conf}}^{\text{FEL}} / N)$ and $\tau_R^{\text{PEL}} = f([S_{\text{conf}}^{\text{PEL}} - \min( S_{\text{conf}}^{\text{PEL}})] / N )$. By using Eqs. (\ref{eq:pel_through_fel_init}) and (\ref{eq:s_pel_min_general}) and again presuming $\min( S_{\text{conf}}^{\text{FEL}}) = 0$, we get for the Adam--Gibbs equation \begin{equation} \begin{aligned} \tau_R^{\text{FEL}} =& \tau_0 \exp\left( \frac{C}{T S_{\text{conf}}^{\text{FEL}} / N} \right),\\ \tau_R^{\text{PEL}} =& \tau_0 \exp\left( \frac{C}{T [S_{\text{conf}}^{\text{FEL}} / N + \delta]} \right), \label{eq:ag_two_forms} \end{aligned} \end{equation} where $\delta$ is expressed as \begin{equation} \begin{aligned} \delta =& \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \left( \Delta_{\alpha}^{\rm T, S} - \Delta_{\alpha}^{\rm T} \right) \\ &- \min( \lim_{\alpha_{\rm min} \to 0} \int_{\alpha_{\rm min}}^{\alpha_{\rm max}} \mathrm{d} \alpha \left( \Delta_{\alpha}^{\rm T, S} - \Delta_{\alpha}^{\rm T} \right) ) \\ &- \left( \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)} - \min( \overline{ \mathcal{S}_{\rm mix}({\bf r}_0^N, \beta)} ) \right) / N. \label{eq:ag_delta} \end{aligned} \end{equation} As can be seen from Eq. (\ref{eq:ag_delta}), $\delta$ possesses nice properties: (i) it is a quantity per particle and (ii) the mixing entropy per particle $S_{\text{mix}} / N$ is not present in $\delta$ (it was canceled out). Thus, $\delta$ will remain continuous with introduction of a small polydispersity, finite for a continuous polydispersity, and independent of artificial particle labeling (defined only by interaction potentials). Thus, both forms of $\tau_R^{\text{AG}}$ from Eq. (\ref{eq:ag_two_forms}) seem feasible. Which form better fits actual data can be tested through simulations. We plan to present in a follow-up paper results that demonstrate that $\tau_R^{\text{PEL}}$ can fit relaxation times for polydisperse hard spheres well. \section{Conclusion} We presented several examples of systems where the lower bound of the configurational entropy $S_{\text{conf}}^{\text{PEL}}$ is the mixing entropy $S_{\text{mix}}$. Our examples are mean-field systems, infinitesimally polydisperse systems, and $1D$ hard rods. We believe these examples provide a deep physical insight into why $S_{\text{conf}}^{\text{PEL}}$ shall in principle have a non-zero lower bound (which still does not preclude the existence of the ideal glass transition). Additionally, we demonstrated that this result is equivalent for our model systems to the condition $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$. Finally, we showed that the condition $\min( S_{\text{conf}}^{\text{FEL}} ) = 0$ implies for an arbitrary system that $\min( S_{\text{conf}}^{\text{PEL}} ) = N A + S_{\rm mix}$, where $A$ is a constant determined by particle interaction potentials, while $S_{\rm mix}$ can be increased arbitrarily by introducing ``artificial'' particle types (\textit{e.g.}, coloring colloidal particles). Thus, $\min( S_{\text{conf}}^{\text{PEL}} )$ can be changed arbitrarily. If the distribution of ``colors'' is continuous, $\min( S_{\text{conf}}^{\text{PEL}} ) / N \to \infty$. Also, $\min( S_{\text{conf}}^{\text{PEL}} ) / N$ will be diverging by introducing slight polydispersity to particle types. Nevertheless, we argued that the AG and RFOT relations will still retain a physically meaningful shape. We compared the AG relations derived through both versions of $S_{\text{conf}}$, $S_{\text{conf}}^{\text{FEL}}$ and $S_{\text{conf}}^{\text{PEL}}$ and demonstrated that they will be difference, but the discrepancy term will be well-behaving. We suggested that one can check which version better describes relaxation times through simulations. We plan to present in a follow-up paper results that demonstrate that $\tau_R^{\text{PEL}}$ can fit relaxation times for polydisperse hard spheres well. \section*{Acknowledgments} We thank Misaki Ozawa and Ludovic Berthier for helpful correspondence about the present manuscript (though they do not necessarily agree with all the ideas presented here).

59c3d8a9d40f87ba0fff3352948eb21fa8a55dc2

\section{Introduction} Coherent structures play a determinant role in several flows including the round jet. According to Lau \& Fisher (1975) the structure of the near-field of a round jet consists essentially in an serie of vortices moving downstream in the mixing layer of the jet. The regions located between the vortices are characterized by a high shearing which is at the source of the high level of turbulence in the shear-layer. This high level of turbulence leads to the spreading out of the jet. So, the coherent structures play a fundamental role in the expansion process of the jet. This is confirmed by the numerical results of Verzicco \& Orlandi (1994). The large scale vortices originate from the shear-layer instability (e.g. Michalke 1984). The Kelvin-Helmholtz instability involves the roll-up of the shear-layer and the formation of vortex rings observed by Liepmann \& Gharib (1992) and simulated by Verzicco \& Orlandi (1994). According to the review of Ho \& Huerre (1984) the roll-up process is predominantly a two-dimensional phenomenon. The vortex rings evolving in the jet shear-layer grow by pairings which have been simulated by Grinstein et al (1987). In a laminar or transitional jet the pairing speeds up the transition to the fully-developed turbulence (e.g. Verzicco \& Orlandi 1994). But the behavior of the round jet depends on the initial conditions of the flow. According to Sahr \& G\"{o}kalp (1991) a jet with a laminar initially shear-layer has a larger capacity of entrainment than a turbulent one. As the spreading out of a jet is related to the presence of coherent structures, this shows that the behavior of the jet depends on the instability wich should occured. In homogeneous incompressible jets the main parameter to influence instability is the initial shear-layer thickness. According to Michalke (1984) for homogeneous jets the coherent structures are involved by the induction of the vorticity occuring in the shear-layer. For Cohen \& Wygnanski (1986) the increase of the initial shear-layer thickness reduces the number of unstable modes. Therefore, the main objective of this paper is to specify the influence of the inlet conditions on the presence of large scale coherent structures in the flow. Previous numerical investigations allowed the simulation of natural unsteadiness in coaxial jets (e.g. Reynier \& Ha Minh 1996) and in compressible round jets (e.g. Reynier 1995) using turbulence models. This method which is an alternative to the LES is chosen for this study to evaluate the influence of the initial conditions on the organized unsteadiness. Finally, the impact of the modelling on the predictions will be investigated using both the classical model proposed by Launder \& Sharma (1974) and the model recalibrated by Ha Minh \& Kourta (1993). \section{Flow modelling} The main objective of this paper is the evaluation of the influence of inlet conditions on the coherent unsteadiness in a quasi-incompressible round jet. In this flow, the three-dimensional effects are weak up to five diameters (e.g. Grinstein et al 1987). Moreover, three-dimensional instabilities are strongly coupled with the random turbulence which becomes predominantly at the end of the potential core located at four or five diameters (e.g. Sokolov et al 1981). In the near-field of an homogeneous jet, coherent structures originate from the Kelvin-Helmholtz instability. This last involves some fluctuations in the mixing layer which are at the source of the large scale vortices. According to Ho \& Huerre (1984), the roll-up of the shear-layer is predominantly a two-dimensional process. Therefore, two-dimensional simulations were executed for this study. As the code works for compressible flows, the mass-weighted average of Favre (1965) is used. The governing equations are the Navier-Stokes equations, a state equation for a perfect gas and the equations of the turbulence model. As the flow to be computed is a round jet the numerical code used axisymmetric coordinates. The flow pattern studied is characterized by high Reynolds numbers, the DNS and the LES cannot be applied. As a consequence the semi-deterministic modelling (e.g. Ha Minh \& Kourta 1993), a method close to the LES, has been chosen for this numerical study. This method allows the simulation of the coherent unsteadiness using turbulence models. The turbulence model used for the predictions is the $k-\epsilon$ model. Two versions of this model are used for the simulations in order to evaluate the influence of the constant $C_{\mu}$ on the simulation of the natural unsteadiness. Firstly, the model proposed by Launder \& Sharma (1974) with the usual set of constants has been chosen. The second model retained is the version proposed by Ha Minh \& Kourta (1993) with a set of constants recalibrated on a backward-facing step to take into account the coherent structures evolving in this flow. If this model has not been recalibrated on a round jet, the low value of the constant $C_{\mu}$ (which has for value $0.02$ in this model and $0.09$ in the version of Launder \& Sharma) should make easier the simulation of the natural instability. Indeed, a lower value of this constant should involve a less diffusive turbulence model. \section{Computational aspects} The numerical scheme used for the calculations is the finite volume method proposed by MacCormack (1981). This explicit-implicit algorithm uses the prediction-correction step technique and resolved the Navier-Stokes equations in a conservative form. The method is accurate to the second order in time and space. The numerical code has been already successfully applied to the simulation of natural unsteadiness in coaxial jets by Reynier \& Ha Minh (1996) and validated with the experimental data of Ribeiro (1972). \begin{figure} \begin{center} \includegraphics[width=80mm]{fig1.jpg} \caption{Configuration of the round jet} \end{center} \end{figure} An air jet presented in Figure 1 is computed. The exit velocity is 104 m/s, the Mach number of the jet is equal to 0.3 and the Reynolds number is 52240. The pipe diameter D is $7.24$ mm. Initially, the temperature at the inlet and in the whole domain is equal to $300^{o}$K, the pressure $P_{e}$ at the exit and in the computational field is $0.101$ MPa and the density in all the field is $\rho_{o}=1.28$ kg.m$^{-3}$. The computational domain extends over 16.6 diameters in the streamwise direction and 8.3 diameters in the radial direction. The mesh uses 100 $\times$ 93 cells with a coarse grid in the radial direction outside the jet. It is uniform in the streamwise direction. \begin{figure} \begin{center} \includegraphics[width=80mm]{fig2.jpg} \caption{Profiles of velocity at the inlet} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=80mm]{fig3.jpg} \caption{Profiles of turbulent kinetic energy at the inlet} \end{center} \end{figure} To investigate the influence of the inlet conditions on the flow unsteadiness several profiles of velocity (see Figure 2) and turbulent kinetic energy (see Figure 3) are applied at the inlet. The different computed cases are reported in the table 1. The initial conditions are derived from the experiment of Dur\~{a}o (1971) for the under-developed turbulence and from the measurements of Chassaing (1979) for the fully-developed turbulence. \begin{table}[htb] \begin{center} \caption{Computed cases with different inlet conditions and turbulence models} \begin{tabular}{llll} \hline Mach number & Reynolds number & Inlet conditions & Turbulence model \\ \hline $0.3$ & $52240$ & Dur\~{a}o & Ha Minh \& Kourta \\ $0.3$ & $52240$ & Chassaing & Ha Minh \& Kourta \\ $0.3$ & $52240$ & Chassaing & Launder \& Sharma \\ \hline \end{tabular} \end{center} \end{table} Outside the jet, a wall is present (see Figure 1) in the transverse direction, so homogeneous Dirichlet conditions are applied on this boundary for velocity, turbulent kinetic energy and dissipation rate and homogeneous Neumann conditions are applied to density and pressure. The lower boundary is the jet axis, therefore symmetry conditions are assumed. The upper boundary is located far from the flow then homogeneous Neumann conditions are imposed on this boundary. In order to do not perturb the flow, non-reflective conditions are applied at the outlet. They are deducted from characteristic relationships. They originated from the theory of the characteristic analysis and they have been developed for the Euler equations by Thompson (1987). When the flow is subsonic the pressure must be specified at the exit, an homogeneous Neumann condition is applied for this quantity. For the turbulent kinetic energy and its dissipation rate the same condition is used on this boundary. \section{Results} \subsection{Influence of inlet conditions on natural instability} \begin{figure} \begin{center} \includegraphics[width=80mm]{fig4.jpg} \caption{Time-variations of the streamwise velocity in the near-field at x$=$1.5D and y$=$0.5D} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=80mm]{fig5.jpg} \caption{Spectrum of time-variations of the streamwise velocity in the near-field at x$=$1.5D and y$=$0.5D} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=120mm]{fig6.jpg} \caption{Profiles of the streamwise velocity for four sections located at x$=$0.33D, x$=$2D, x$=$5D and x$=$15D for four moments of a pseudo-period T/4, T/2, 3T/4 and T} \end{center} \end{figure} To evaluate the influence of initial conditions on the natural instability of the flow, the jet has been computed for fully-developed and under-developed turbulence at the inlet. If the velocity profiles (see Figure 2) are nearly similar it is not the case for the profiles of turbulent kinetic energy. The level of turbulence is lower for the under-developed case particularly in the central region of the flow where the level of turbulent kinetic energy is twice smaller that for the fully-developed turbulence. The turbulence model used for these simulations is the semi-deterministic model proposed by Ha Minh \& Kourta (1993). The executed computations with the inlet conditions derived from the experimental data of Dur\~{a}o (1971) lead to the simulation of instabilities in the near-field without any flow excitation. The figure 4 represents the time-dependent variations of the streamwise velocity for a point located in the shear-layer at x$=$1.5D and y$=$0.5D. The variations of the streamwise velocity are quasi-sinusoidal. The corresponding spectrum obtained by Fourier analysis over one hundred periods are presented in figure 5. The spectrum puts in evidence the presence of a dominant frequency equal to 5600 Hz. The associated Strouhal number (calculate from the diameter of the inlet pipe and the exit velocity) is equal to 0.39. This Strouhal number is in the range of values contained between 0.3 and 0.4 corresponding to the preferred mode (e.g. Michalke 1984). This preferred mode corresponds to the coherent structures which dominate the shear-layer of a round jet. In figure 6, the unsteady profiles of streamwise velocity are plotted. This figure shows the unsteady variations of the velocity for four sections of the mesh located at x$=$0.33D, x$=$2D, x$=$5D and x$=$15D, for four moments of a pseudo-period: T/4, T/2, 3T/4 and T. A high unsteadiness is active in the near region at x$=$0.33D and x$=$2D. This natural unsteadiness originates from the fluctuations in the shear-layer of the jet. The mixing layer becomes unstable near the inlet, due to the Kelvin-Helmholtz instability, then rolls up to form vortex rings. This phenomenon has been largely studied by Liepmann \& Gharib (1992) and Verzicco \& Orlandi (1994). The present results show the damping of the organized unsteadiness in the far field of the flow (see figure 6). At x$=$15D the unsteadiness has disappeared. \begin{figure} \begin{center} \includegraphics[width=120mm]{fig7.png} \caption{Streamwise velocity field predicts with the model of Ha Minh \& Kourta and a fully-developed turbulence at the inlet} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[width=120mm]{fig8.png} \caption{Radial velocity field predicts with the model of Ha Minh \& Kourta and a fully-developed turbulence at the inlet} \end{center} \end{figure} The simulation of the flow with a fully-developed turbulence at the inlet leads to the prediction of a steady field of streamwise velocity presented in figure 7. This lack of unsteadiness in the flow is due to a high turbulence level in the center of the jet at the inlet. According to Michalke (1984) the presence of coherent structures is involved by the induction of the vorticity occuring in the shear-layer. When the turbulence is fully-developed the vorticity occuring in the shear-layer is very weak or absent. As a consequence the organized unsteadiness evolving in the shear-layer damps for a fully-developed turbulence at the inlet. The field of radial velocity predicted by the semi-deterministic model shows some residual instabilities in the shear-layer (see figure 8) but they are very weak. With the inlet conditions derived from the experiment of Dur\~{a}o (1971) the maximum of the radial velocity, scaled by the exit velocity, $|V| / U_{o}$, is equal to 0.33. With the inlet conditions corresponding to a fully-developed turbulence the maximum of this quantity is equal to 0.07. This low value of the radial velocity in the shear-layer shows a weak entrainment of the jet for the fully-developed turbulence that agrees with the experimental results of Sahr \& G\"{o}kalp (1991). The weak instabilities can be interpreted as remnants of coherent structures at a great age of turbulence. \subsection{Influence of the modelling on the predictions} \begin{figure} \begin{center} \includegraphics[width=120mm]{fig9.png} \caption{Streamwise velocity field predicts with the model of Launder \& Sharma and a fully-developed turbulence at the inlet} \end{center} \end{figure} The flow has been computed with the initial conditions derived from the measurements of Chassaing (1979) for the two models retained for this study in order to evaluate the impact of the value of the constant $C_{\mu}$ on the predictions. The computations lead to the prediction of steady fields of streamwise velocity which are represented in figures 7 and 9. The visualizations show that the potential core is shorter for the simulation using the standard set of constants. With the classical model the expansion of the jet is larger than for the simulation with the recalibrated model. These discrepancies between the prediction of the two models are due to the difference of the value of the constant $C_{\mu}$. The low value of this constant in the model of Ha Minh \& Kourta (1993) involves a lower turbulent viscosity, therefore a less diffusive model. This explains the differences in the prediction of the streamwise velocity field. For the prediction of the radial velocity, the simulation using the classical model of Launder \& Sharma (1974) leads to steady results. But the results obtain using the semi-deterministic model (see figure 8) shows some weak unsteadiness in the shear-layer. This is a consequence of the smaller level of diffusion in the semi-deterministic model. \section{Conclusion} This numerical study shows that the coherent structures simulated with under-developed turbulence at the inlet are not found with fully-developed turbulence. Therefore the organized unsteadiness appears to be highly dependent on inlet conditions and particularly on the initial level of turbulent kinetic energy. Indeed, the coherent structures are characteristic of a young turbulence. The simulations with fully-developed initial conditions and the two turbulence models chosen for this study puts in evidence that a low value of the constant $C_{\mu}$ involves a weaker diffusion in the flow. This low diffusion allows the simulation of a weak unsteadiness in the shear-layer for a fully-developed turbulence at the inlet. These weak instabilities are interpreted as representative of remnants of coherent structures at a great age of turbulence.\\ \noindent \textbf{Acknowledgements} P. Reynier acknowledges with gratitude the support of the CNES through a post-doctorate grant.

a1a7c60bd509a469c512d52303ce6256a36cf3f7

\section{Conclusion} \inv This paper proposes a greatly improved translation-based graph embedding method that helps ontology population by way of relation prediction. The proposed \ontovec\ model can effectively address the learning issues on the two categories of comprehensive semantic relations in ontology graphs, and improves previous methods using two dedicated component models. Extensive experiments on four data sets show promising capability of \ontovec\ on predicting and verifying relation facts.\par The results here are very encouraging, but we also point out opportunities for further work and improvements. In particular, we should explore the effects of other possible forms of component-specific projections, such as dynamic mapping matrices and bilinear mappings. Encoding other information such as the domain and range information of concepts may also improve the precision of our tasks. More advanced applications may also be developed using \ontovec\, such as ontology-boosted question answering. Jointly training \ontovec\ with alignment models \cite{chen2017learning} is another meaningful direction since it provides a more generic embedding model that helps populating and aligning multilingual ontology graphs. \section{Experiments} In this section, we evaluate \ontovec\ on two tasks that answer two important questions for ontology population: (i) Relation prediction: what is the relation to be added between a given pair of concepts? (ii) Relation verification: is a candidate relation fact correct or not?\par The baselines that we compare against include the representative translation-based embedding methods TransE, TransH, TransR, and TransD \cite{bordes2013translating,wang2014knowledge,lin2015learning,ji2015knowledge}, and neural methos RESCAL and HolE \cite{nickel2011three,nickel2016holographic}. Experimental results are reported on four data sets extracted from DBpedia, ConceptNet, and Yago, for which comprehensive relation types have been predefined. Statistics of the data sets are shown in Table~\ref{tbl:stat}. All the meta relations that assign URIs and system timestamps are removed during the preparation of the data sets. To simplify the experiments, transitive relations are limited to four-hops. Relation facts for extra hops are hence discarded. Since DBpedia provides both ontology and instance-level graphs, we keep only the ontology view to obtain DB3.6k. CN30k and YG15k are extracted from English versions of ConceptNet and Yago respectively. These two graphs match the number of nodes with WN18 and FB15k respectively, which are two commonly-used instance-level graphs in related works~\cite{bordes2013translating,wang2014knowledge,lin2015learning,ji2015knowledge,bordes2011learning,bordes2012joint,yang2015network}. YG60k is a much larger data set that is about half of the entire English-version Yago after data cleaning. Each data set is randomly partitioned into training, validation, and test sets. \input{tbl2} \input{fig1} \subsection{Relation Prediction}~\label{sect:rel_pred} This task aims at extending an ontology graph by predicting the missing relations for given concept pairs. \par \stitle{Evaluation Protocol.} We evaluate our approach by way of held-out evaluation~\cite{wang2014joint,lin2016neural}. Each model is trained on the training set that represents the known ontology. Then, for each case in the test set, given the source and target concepts, the model predicts the relation that leads to the lowest dissimilarity score $S_d$ defined in Section~\ref{sect:CSM}. To evaluate with controlled variables, on each data set, we employ the same configuration for every models. On DB3.6k, we fix dimensionality $k=25$, margin $\gamma_1=2.0$, learning rate $\lambda=0.005$, $\alpha_2=0.5$, and $l_1$ norm. CN30k and YG15k shares the configuration as $k=50$, $\gamma_1=0.5$, $\lambda=0.001$, $\alpha_2=0.5$, and $l_2$ norm. Lastly, we use $k=100$, $\gamma_1=0.5$, $\lambda_1=0.001$, $\alpha_2=0.5$, and $l_2$ norm. $\gamma_2=0.5$ is configured for \ontovec. To test the effect of HM, we also provide two versions of \ontovec. One version (\ontovec\ w/ HM) is set with $\alpha_1=0.75$, which is empirically decided via the hyperparameter study in Section~\ref{sect:case}. The other version (\ontovec\ w/o HM) nullifies HM by setting $\alpha_1=0$. To enable batch sampling for HM, we implement the $\sigma$ function for hierarchical relation facts using hash trees. The learning process is stopped once the accuracy on the validation set stops improving. \par \stitle{Results.} The overall accuracy is reported per data set in Table~\ref{tbl:predict}. On each data set, we also aggregate respectively the accuracy on the test cases with relational properties, as well as the accuracy on those with hierarchical relations. We discover that, TransE, though has performed well on encoding instance-level knowledge graphs~\cite{bordes2013translating}, receives unsatisfactory results on predicting the comprehensive ontology relations. By learning each relation type on a different hyperplane, TransH notably solves the problem of TransE, but appears to fail on CN30k where the candidate space is larger than other graphs. TransR and TransD provide more robust characterization of relations than TransH, especially in TransR where relation-specific projections are implemented as linear transformations. However, the overall performance of both TransR and TransD is impaired by the two types of comprehensive relations. For neural models, HolE adapts better on the smaller DB3.6k data set, while it is at most comparable to TransR and TransD on larger ones, and RESCAL is less successful on all settings. As expected, \ontovec\ greatly outperforms the above baseline methods, regardless of whether HM is enabled or not. The \ontovec\ with HM thereof, outperforms the best runner-up baselines respectively in all settings by 3.72\%${\sim}$10.52\% of overall accuracy, 4.09\%${\sim}$11.69\% of accuracy on cases with relational properties, and 7.97\%${\sim}$14.24\% of accuracy on cases with hierarchical relations. We also discover that, when HM is enabled, it leverages the accuracy on hierarchical relations by up to 12.75\%, and overall accuracy by up to 7.27\%, and does not noticeably cause interference to the prediction for cases with relational properties. Though, the advantage of CSM alone (i.e. \ontovec\ w/o HM) is still significant over the baselines. Since the relation prediction accuracy of \ontovec\ is close to 90\% on all four data sets, this indicates that \ontovec\ achieves a promising level of performance in populating ontology graphs, and it is effective on both small and large graphs.\par We also perform precision-recall analysis on the two Yago data sets on translation-based models. To do so, we calculate the dissimilarity scores $S_d$ (Equation~\ref{eq:csm1}) for the possible predictions of each test case, and select those that are not ranked behind the correct prediction. Then a threshold is initiated as the minimum dissimilarity score. The answer set is inserted with predictions for which the dissimilarity scores fall below the threshold, and the answer set grows along with the increasing of the threshold, until all correct predictions are inserted. Therefore, we obtain the precision-recall curves in Fig.~\ref{fig:prc}, for which the area under curve is reported as: (i) For YG15k, \ontovec\ w/ HM: \textbf{0.9138}; \ontovec\ w/o HM: 0.8938; TransE: 0.0457; TransH: 0.4973; TransD: 0.8386; TransR: 0.8587. (ii) For YG60k, \ontovec\ w/ HM: \textbf{0.9005}; \ontovec\ w/o HM: 0.8703; TransE: 0.0313; TransH: 0.6688; TransD: 0.7275; TransR: 0.8372. This further indicates that \ontovec\ achieves better performance than other baselines, and HM improves the performance of On2Vec with CSM alone. \subsection{Relation Verification} \input{tbl3} Relation verification aims at judging whether a relation marked between two concepts is correct or not. It produces a classifier that helps to verify the candidate relation facts.\par \stitle{Evaluation Protocol.} Because this is a binary classification problem that needs positive and negative cases, we use a complete data set as the positive cases. Then, following the approach of \cite{socher2013reasoning}, we corrupt the data set to create negative cases. In detail, a negative case is created by (i) randomly replacing the relation of a positive case with another relation, or (ii) randomly assign a relation to a pair of unrelated concepts. Options (i) and (ii) respectively contribute negative cases that are as many as 100\% and 50\% of positive cases. We perform a 10-fold cross-validation. Within each fold, embeddings and the classifier are trained on the training data, and the classifier is evaluated on the remaining validation data.\par We use a threshold-based classifier, which is similar to the one for triple alignment verification in \cite{chen2017multi}. This simple classifier adequately relies on how precisely each model preserves the structure of the ontology graph in the embedding space. In detail, for each case, we calculate its dissimilarity score $S_d$ (Section~\ref{sect:CSM}). The classifier then finds a threshold $\tau$ such that $S_d < \tau$ implies positive, otherwise negative. The value of $\tau$ is determined to maximize the accuracy on the training data of each fold.\par We carry forward the corresponding configurations from the last experiment, in order to show the performance of each model under controlled variables.\par \input{tbl4} \stitle{Results.} We aggregate the mean accuracy for the two categories of comprehensive relation facts as well as the overall accuracy for each setting. The results are shown in Table~\ref{tbl:classification}, which has a maximum standard deviation of 0.005 in cross-validation for each setting. Thus, the results are statistically sufficient to reflect the performance of classifiers. Both versions of \ontovec\ again outperform the other models, especially on comprehensive relation facts. On all four data sets, \ontovec\ outperforms the best runner-up baselines by 2.98\%${\sim}$9.67\% of overall accuracy, 2.02\%${\sim}$12.57\% of accuracy for cases with relational properties, and 1.29${\sim}$8.15\% of accuracy on hierarchical relations. This indicates that \ontovec\ precisely encodes the ontology graph structures, and provides much accurate plausibility measurement to decide the correctness of unknown triples. We also discover that, \ontovec\ trained with HM has a drop of accuracy for up to 0.8\% on cases with relational properties from CN30k and YG15k. This is likely due to that the auxiliary learning process for hierarchical relations causes minor interference to the characterization of relational properties, while HM leverages the accuracy on hierarchical relations of these two data sets by at least 1.81\%, and the overall accuracy by 0.82\%${\sim}$3.83\%. This indicates that HM is helpful in relation verification. \subsection{Case Study}~\label{sect:case} Lastly, we provide some case studies on hyperparameter values, and some examples of relation prediction. \subsubsection{Hyperparameter study} We examine the hyperparameter $\alpha_1$, which is the trade-off between CSM and HM. The result based on relation prediction on YG15k is shown in Fig.~\ref{fig:alpha}. As we can see, although enabling HM with even a small value of $\alpha_1$ can noticeably leverage the performance of \ontovec, the influence of different values of $\alpha_1$ is not very notable, and the accuracy does not always go up along with the higher $\alpha_1$. In practice, $\alpha_1$ may be fine-tuned for marginal improvement, while $\alpha_1=0.75$ can be empirically selected. \subsubsection{Examples of relation prediction} Relation prediction is also performed for the complete data set of CN30k and YG60k. To do so, we randomly select 20 million pairs of unlinked concepts from these two data sets, and rank all the predictions based on the dissimilarity score $S_d$. Then top-ranked predictions are selected. Human evaluation is used in this procedure, since there is no ground truth for the relation facts that are not pre-existing. Like previous works~\cite{lin2016neural,zeng2015distant}, we aggregate $P@200$, i.e. the precision on the 200 predictions with highest confidence, which results in 73\% and 71\% respectively. Some examples of top-ranked predictions are shown in Table~\ref{tbl:new_rel}. \section{Introduction} Ontology graphs are a special category of knowledge graphs that support and augment the Semantic Web with comprehensive and transportable machine understanding \cite{maedche2001learning}. They store formal descriptions and specification of human knowledge in forms of relation facts (triples), making it semantically understandable and inferrable for the machine. Unlike other instance-level knowledge graphs~\cite{zaniolo2017user} that define simple and casual labeled relations for specified entities, ontology graphs define a fixed set of specialized semantic relations among generalized concepts. Such semantic relations of ontologies are typically very comprehensive in terms of relational properties and form hierarchies, which we are going to discuss shortly.\par Populating large ontologies has been a critical challenge to the Semantic \mbox{Web}. In the past decade, several well-known ontology graphs have been created and widely utilized, including Yago~\cite{mahdisoltani2014yago3}, ConceptNet~\cite{speer2017conceptnet}, and DBpedia OWL~\cite{lehmann2015dbpedia}. Although some of these graphs contain millions of relation facts, they still face the coverage and completeness issues that have been the subject of much research~\cite{quan2004automatic,mousavi2014mining}. This is because enriching such large structures of expertise knowledge requires levels of intelligence and labor that is hardly affordable to humans. Hence, some works have proposed to mine ontologies from text using parsing-based~\cite{culotta2004dependency,mousavi2014text,fundel2007relex} or fuzzy-logic-based~\cite{quan2004automatic,lau2009toward,widyantoro2001fuzzy} techniques. However, in practice, these techniques are often limited by the lack of high-quality reference corpora that are required for the harvest of the dedicated domain knowledge. Also, the precise recognition of relation facts for the ontology is another unsolved problem, since these relation facts are very high-level and are often not explicitly expressed in the corpora \cite{lau2009toward}. Hence, these methods merely help populate some small ontology graphs in narrow domains such as gene ontologies and scholarly ontologies~\cite{cheng2004netaffx,quan2004automatic}, but they have not been successfully used to improve the completeness of these large cross-domain ontology graphs such as Yago and ConceptNet.\par A more practical solution is to use translation-based graph embedding methods, which predict the missing relation facts using vector representations of the graph, without the need of additional information from any text corpus. Specifically, given a triple $(s, r, t)$ such that $s$, $t$ denote the source and the target entities (or concepts), and $r$ denotes the edge that marks the relation between $s$ and $t$, then $s$ and $t$ are represented as two $k$-dimensional vectors $\mathbf{s}$ and $\mathbf{t}$, respectively. An energy function $S_{r}(\mathbf{s},\mathbf{t})$ is used to measure the plausibility of the triple, which also implies the transformation $\mathbf{r}$ that characterizes $r$. Therefore, new triples with high plausibility (or low energy) are often induced. For example, TransE~\cite{bordes2013translating} uses the energy function $S_{r}(\mathbf{s},\mathbf{t})=\left \| \mathbf{s}+\mathbf{r}-\mathbf{t} \right \|$~\footnote{Hereafter, $\|\cdot\|$ means $l_1$ or $l_2$ norm unless specified.}, where $\mathbf{r}$ is characterized as a translation vector learnt from the latent connectivity patterns in the graph. Other representative works, such as TransH~\cite{wang2014knowledge}, TransR~\cite{lin2015learning}, and TransD~\cite{ji2015knowledge} improve TransE by specializing the encoding process for each relation type using a relation-specific projection on entities. \par While these methods help enrich instance-level knowledge graphs, they only focus on capturing the simple relations in instance-level knowledge graphs, paying less attention to the comprehensive relations in ontology graphs. In fact, relation facts in ontology graphs are often defined with relational properties, such as transitivity and symmetry, as well as form hierarchies. A typical example is provided by {\em Is-A}, which is both transitive and hierarchical, and is the most frequently appearing semantic relation in ontologies. We find that, in well-known ontology graphs, comprehensive relations usually comprise the majority: 85\% of the triples in Yago, 96\% of the triples in ConceptNet, and 47\% of the triples in DBpedia OWL enforce relational properties, while 60\%, 38\%, and 48\% of these triples are defined with hierarchical relations. However, existing \mbox{methods} fail to represent these comprehensive relations for several reasons: (i) These methods at most use the same relation-specific projection in the energy function, but fail to differentiate the components of triples. Therefore, they are ill-posed to characterize triples with relational properties. In fact, the encoding of a concept that serves as different components in such triples, i.e. either $s$ or $t$, must be differentiated so as to correctly preserve relational properties in the embedding spaces (as shown in Section~\ref{sect:pre}). (ii) These methods also lack a learning phase that is dedicated to hierarchical relations. This also impairs the preciseness of embeddings. We observe in our experiments that, above \mbox{limitations} largely hinder the effectiveness of existing methods for ontology graphs.\par Therefore, to support ontology population more effectively, we propose \ontovec, a translation-based graph embedding model that specializes in characterizing the comprehensive semantic relations in ontology graphs. \ontovec\ adopts two component models: the {\em Component-specific Model} which preserves the relational properties by applying component-specific projections on source and target concepts respectively, and the {\em Hierarchy Model} which performs an attentive learning process on hierarchical relations. We evaluate our model with the tasks of relation prediction and relation verification, which respond respectively to the following two questions: (i) What relation should be added between two concepts? (ii) Is the predicted relation correct? Experimental results on data sets extracted from Yago, ConceptNet, and DBpedia OWL show promising results and significant improvement on related methods. \par The rest of the paper is organized as follows. We first discuss the related work, and then introduce our approach in the section that follows. After that we present the experimental evaluation, and conclude the paper in the last section. \section{Embedding Ontology Graphs} In this section, we introduce the proposed method for learning ontology graph embeddings. We begin with the formalization of ontology graphs. \subsection{Preliminary} An ontology is a graph $G(C, R)$ where $C$ is the set of concepts, and $R$ is the set of semantic relations. $T = (s,r,t) \in G$ denotes a triple that represents a relation fact, for which $s, t \in C$ and $r \in R$. Boldfaced $\mathbf{s}$, $\mathbf{r}$, $\mathbf{t}$ respectively represent the embedding vectors of source $s$, relation $r$, and target $t$. Relations are further classified by $R=R_{tr} \cup R_{s} \cup R_h \cup R_{o}$, which respectively denote the sets of transitive, symmetric, hierarchical, and other simple relations. We do not specify reflexive relations here because such relations can be easily model as a zero vector by any translation-based model. $R_{tr}$ and $R_h$ thereof, are not required to be disjoint, while $R_o$ is disjoint with all the rest three. For transitive relations, that is to say, given $r \in R_{tr}$, and three different concepts $c_1, c_2, c_3 \in C$, if $(c_1, r, c_2), (c_2, r, c_3) \in G$, then $(c_1, r, c_3) \in G$. As for symmetric relations, that is to say, given $r \in R_s$, and two different concepts $c_1, c_2 \in C$, if $(c_1, r, c_2) \in G$, then $(c_2, r, c_1) \in G$. As for hierarchical relations, we further divide them into $R_h=R_r \cup R_c$ where $R_r$ denotes refinement relations that partition coarser concepts into finer ones, and $R_c$ denotes coercion relations that group finer concepts to coarser ones~\cite{camossi2006multigranular,chen2016converting,chen2016sac}. \subsection{Modeling}~\label{sect:pre} \ontovec\ adopts two component models that learn on the two facets of the ontology graph: the {\em Component-specific Model} (CSM) which encodes concepts and relations into low-dimensional embedding spaces without the loss of the relational properties, and the {\em Hierarchy Model} (HM) which strengthens the learning process on hierarchical relations with an auxiliary energy.\par \begin{figure} \centering \resizebox {\columnwidth} {!} { \begin{tikzpicture} [nd/.style={circle,draw=blue!50,fill=blue!20,thick}] \node (c1r) at (-4, 0.5) [nd,label=-180:{\tiny $C_{1r}$}] {}; \node (c2r) at (-2.5, 0) [nd,label=90:{\tiny $C_{2r}$}] {}; \node (c3r) at (-1, -0.5) [nd,label=90:{\tiny $C_{3r}$}] {}; \draw [->,draw=blue!50,thick] (c1r) to node [auto,swap,color=blue!80] {\tiny $\mathbf{c}_{1r}+\mathbf{r} \approx \mathbf{c}_{2r}$} (c2r); \draw [->,draw=blue!50,thick] (c2r) to node [auto,swap,color=blue!80] {\tiny $\mathbf{c}_{2r}+\mathbf{r} \approx \mathbf{c}_{3r}$} (c3r); \draw [->,draw=red!50,thick] (c1r) .. controls +(up:10mm) and +(right:20mm) .. node[auto,color=red!80] {\tiny $\mathbf{c}_{1r}+\mathbf{r} \not\approx \mathbf{c}_{3r}$} (c3r); \node (c4r) at (1, 0.5) [nd,label=-90:{\tiny $C_{1r}$}] {}; \node (c5r) at (4, -0.5) [nd,label=-90:{\tiny $C_{2r}$}] {}; \draw [->,draw=blue!50,thick] (c4r) to node [auto,swap,color=blue!80] {\tiny $\mathbf{c}_{1r}+\mathbf{r} \approx \mathbf{c}_{2r}$} (c5r); \draw [->,bend right,draw=red!50,thick] (c5r) to node[auto,swap,color=red!80] {\tiny $\mathbf{c}_{2r}+\mathbf{r} \not\approx \mathbf{c}_{1r}$} (c4r); \end{tikzpicture} } \caption{Depiction of the conflicts of the relation-specific projection for learning transitive relations (Case 1, left), and symmetric relations (Case 2, right).}\label{fig:conflict} \end{figure} \inv\inv \subsubsection{Component-specific Model.}~\label{sect:CSM} The reason that previous translation-based models fail to preserve relational properties is because the relation-specific projection $f_{r}$ place concepts involved in transitive or symmetric relations at conflict positions. Fig.~\ref{fig:conflict} depicts such conflicts, and a brief proof is given below: \begin{itemize}[noitemsep] \item \stitle{Case 1.} Consider $r \in R_{tr}$ and $c_1, c_2, c_3 \in C$ such that $(c_1, r, c_2)$, $(c_2, r, c_3)$, $(c_1, r, c_3) \in G$, where $c_1$, $c_2$, and $c_3$ are projected to $\mathbf{c}_{1r}$, $\mathbf{c}_{2r}$, and $\mathbf{c}_{3r}$ respectively by $f_{r}$. Then if $\mathbf{c}_{1r}+\mathbf{r}\approx \mathbf{c}_{2r}$ and $\mathbf{c}_{2r}+\mathbf{r}\approx \mathbf{c}_{3r}$ hold for the first and second triples, it is impossible for $\mathbf{c}_{1r}+\mathbf{r}\approx \mathbf{c}_{3r}$ to hold for the third triple, since $\mathbf{r} \neq 0$ (otherwise $\mathbf{r}$ does not provide a valid vector translation). \item \stitle{Case 2.} Consider $r \in R_{s}$ and $c_1, c_2 \in C$ such that $(c_1, r, c_2), (c_2, r, c_1) \in G$, where $c_1$ and $c_2$ are projected to $\mathbf{c}_{1r}$ and $\mathbf{c}_{2r}$ respectively by $f_{r}$. Then it is not possible for both $\mathbf{c}_{1r}+\mathbf{r}\approx \mathbf{c}_{2r}$ and $\mathbf{c}_{2r}+\mathbf{r}\approx \mathbf{c}_{1r}$ to hold, since $\mathbf{r} \neq 0$. \end{itemize} Hence, to solve the conflicts in the above two cases, CSM provides two component-specific (and also relation-specific) projections to differentiate the encoding of the same concept that serves as different components in triples. The general form of the energy function is given as below, \small \begin{equation*}~\label{eq:csm1} S_d(T)=\left \| f_{1,r}(\mathbf{s})+\mathbf{r}-f_{2,r}(\mathbf{t}) \right \| \end{equation*} \normalsize where $f_{1,r}$ and $f_{2,r}$ are respectively the component-specific projections for the source and the target concepts. It is easy to show that the component-specific projections are able to solve the conflicts in learning the relational properties, as $c_2$ in Case 1 is projected differently when it serves as the source of $(c_1, r, c_2)$ or the target of $(c_2, r, c_3)$, while both $c_1$ and $c_2$ in Case 2 can be learnt to be embedded in opposite positions respectively for $(c_1, r, c_2)$ and $(c_2, r, c_1)$ by the two projections. Corresponding conclusion can be easily extended to cases with more than three relation facts via mathematical induction.\par Besides measuring the plausibility (or the opposite: dissimilarity) of a given triple, $S_d$ is also the basis for predicting missing relation facts for an ontology. Given two concepts $s$ and $t$, we find the $r$ which leads to the lowest $S_d$. The forms of $f_{1,r}$ and $f_{2,r}$ are decided particularly by the techniques to differentiate the concept encoding under different contexts of relations. In this paper, we adopt the relation-specific linear transformations \cite{lin2015learning}. Hence, we have $f_{1,r}(\mathbf{s})=\mathbf{M}_{1,r}\mathbf{s}$ and $f_{2,r}(\mathbf{t})=\mathbf{M}_{2,r}\mathbf{t}$, such that $\mathbf{M}_{1,r},\mathbf{M}_{2,r} \in \mathbb{R}^{k \times k}$. Other techniques like hyperplane projections, dynamic matrices, and bilinear transformations may also be considered, which we leave as future work.\par The objective of CSM is to minimize the total $S_d$ energy of all triples. To achieve more efficient learning, we import negative sampling to the learning process, which is widely applied in previous works~\cite{bordes2013translating,wang2014knowledge,lin2015learning,ji2015knowledge}. Unlike these works that select negative samples on entities (or concepts), we perform negative sampling on semantic relations to better suit our tasks. Then the complete energy function of CSM is defined as the following hinge loss, \begin{align*} \begin{split \small S_{\mathrm{CSM}}(G)=\sum_{(s,r,t) \in G}[\left \| f_{1,r}(\mathbf{s})+\mathbf{r}-f_{2,r}(\mathbf{t}) \right \|\\ -\left \| f_{1,r}(\mathbf{s})+\mathbf{r'}-f_{2,r}(\mathbf{t}) \right \|+\gamma_1]_{+} \normalsize \end{split} \end{align*} for which $r'$ is a randomly sampled relation that does not hold between $s$ and $t$, $\gamma_1$ is a positive margin, and $[x]_{+}$ denotes the positive part of $x$ (i.e., $\max(x, 0)$). \subsubsection{Hierarchy Model.} For a hierarchical relation, we often have multiple finer concepts that apply this relation to a coarser one. In this case, we appreciate a good representation where all the embeddings of the finer concepts converge closely in a tight neighborhood, which corresponds to low dissimilarity of the embedded relation. However, it is very likely for the learning process to spread out the embeddings of the finer concepts. Because each of the finer concepts can participate in multiple relation facts, encoding of a concept in one relation fact can be easily interfered by that of many other relation facts. This no doubt indicates low plausibility measures of the triples, and imprecise vector translation for the corresponding relations. Therefore, HM is dedicated to converge closely the projected embeddings of every finer concepts for a hierarchical relation. \par To facilitate the definition of the energy function, we first define a {\em refine} operator denoted as $\sigma$: \begin{itemize}[noitemsep] \item Given $r \in R_r$, $c\in C$, then $\sigma(c, r)=\{c' | (c, r, c')\in G\}$ fetches all the finer concepts $c'$ that directly apply the refinement relation $r$ to the coarser $c$. \item Given $r \in R_c$, $c\in C$, then $\sigma(c, r)=\{c' | (c', r, c)\in G\}$ fetches all the finer concepts $c'$ that directly apply the coercion relation $r$ to the coarser $c$. \end{itemize} The energy function of HM is defined below, \begin{align*} \small \begin{split S_{hm}(G)=\sum_{r\in R_r} \sum_{s\in C} \sum_{t\in \sigma(s,r)} \omega\left( f_{1,r} (\mathbf{s}) + \mathbf{r}, f_{2,r} (\mathbf{t}) \right)\\ +\sum_{r\in R_c} \sum_{t\in C} \sum_{s\in \sigma(t,r)} \omega\left( f_{2,r} (\mathbf{t})-\mathbf{r}, f_{1,r} ( \mathbf{s}) \right) \end{split} \normalsize \end{align*} where $\omega$ is a function that monotonically increases w.r.t. the angle or the distance of the two argument vectors. In practice, $\omega$ can be easily implemented as cosine distance.\par Negative sampling is imported to rewrite $S_{hm}$ as below, \begin{align*} \small \begin{split}~\label{eq:hm} S_{\mathrm{HM}}(G)=\sum_{r\in R_r}\sum_{s\in C}\sum_{t\in \sigma(s,r) \land t'\notin \sigma(s,r)} S_{hr}\\ +\sum_{r\in R_c} \sum_{t\in C} \sum_{s\in \sigma(t,r) \land s'\notin \sigma(t,r)} S_{hc} \end{split} \normalsize \end{align*} such that $s'$ and $t'$ are negative samples of concepts, $S_{hr}$ and $S_{hc}$ are respectively the hinge loss for refinement and coercion relations defined as below, where $\gamma_2$ is a positive margin. \small \begin{equation*}~\label{eq:hr} S_{hr}=[\omega\left( f_{1,r} (\mathbf{s}) + \mathbf{r}, f_{2,r} (\mathbf{t}) \right)-\omega\left( f_{1,r} (\mathbf{s}) + \mathbf{r}, f_{2,r} (\mathbf{t}') \right)+\gamma_2]_{+} \end{equation*} \normalsize \small \begin{equation*}~\label{eq:hc} S_{hc}=[\omega\left( f_{2,r} (\mathbf{t})-\mathbf{r}, f_{1,r} ( \mathbf{s}) \right)-\omega\left( f_{2,r} (\mathbf{t}) - \mathbf{r}, f_{1,r} (\mathbf{s}') \right)+\gamma_2]_{+} \end{equation*} \normalsize \input{alg1} \input{tbl1} Table~\ref{tbl:model_complexity} gives the model complexity of \ontovec\ and some related models in terms of parameter sizes. We also give out the computational complexity of the relation prediction for a pair of concepts, which is the most frequent operation in our tasks. Although \ontovec\ unavoidably increases the parameter sizes due to additional projections, it keeps the computational complexity of relation prediction at the same magnitude as TransR, which is lower than TransD. \subsection{Learning Process} The objective of learning \ontovec\ is to minimize the combined energy of $S_{\mathrm{CSM}}$ and $S_{\mathrm{HM}}$. Meanwhile, norm constraints are enforced on embeddings and projections to prevent training from a trivial solution where vectors collapse to infinitely large~\cite{bordes2014semantic,chen2017multi,wang2014knowledge}. Such constraints are conjuncted below. \begin{align*} \footnotesize \begin{split \forall c \in C,\forall r \in R : \left \| \mathbf{c} \right \|\leq 1 \land \left \| f_{1,r}(\mathbf{c}) \right \|\leq 1 \land \left \| f_{2,r}(\mathbf{c}) \right \|\leq 1 \land \left \| \mathbf{r} \right \|\leq 2 \end{split} \normalsize \end{align*} In the learning process, these constraints are quantified as soft constraints: \begin{align*} \small \begin{split S_{\mathrm{N}}(C,R)=\sum_{c \in C}([\left \| \mathbf{c} \right \|-1]_{+}+[\left \| f_{1,r}(\mathbf{c}) \right \|-1]_{+}\\ +[\left \| f_{2,r}(\mathbf{c}) \right \|-1]_{+})+\sum_{r \in R}[\left \| \mathbf{r} \right \|-2]_{+} \end{split} \normalsize \end{align*} Finally, learning \ontovec\ is realized by using batch stochastic gradient descent \mbox{(SGD)} \cite{needell2014stochastic} to minimize the joint energy function given as below, \small \begin{equation*}~\label{eq:joint} J(\theta)=S_{\mathrm{CSM}}+\alpha_1 S_{\mathrm{HM}} + \alpha_2 S_{\mathrm{N}} \end{equation*} \normalsize where $\alpha_1$ and $\alpha_2$ are two non-negative hyperparameters, and $\theta$ is the set of model parameters that include embedding vectors and projection matrices. Empirically (as shown in~\cite{wang2014knowledge,lin2015learning}), $\alpha_2$ is assigned with a small value within (0, 1]. $\alpha_1$ is adjusted in experiments to weigh between the two component models. Instead of directly updating $J$, the learning process optimizes $S_{\mathrm{CSM}}$ and $\alpha_1 S_{\mathrm{HM}}$ in separated groups of batches, and the batches from both groups are used to optimize $\alpha_2 S_{\mathrm{N}}$. We initialize vectors by drawing from a uniform distribution on the unit spherical surface, and initialize matrices using random orthogonal initialization \cite{saxe2013exact}. The detailed optimization procedure is given in Algorithm~\ref{alg:training}.\par \section{Related Work} \inv In this section, we discuss three lines of works that are related to our topic. \stitle{Ontology Population.} Extensive human efforts are often put into the creation of ontology graphs. Thus, ontology population aims at automatically extending the graphs with the missing relation facts. A traditional strategy is to mine those facts from text corpora. Many works rely on parsing-based techniques to harvest relation facts ~\cite{culotta2004dependency,mousavi2014text,wang2006automatic,giuliano2008instance,fundel2007relex}. These approaches often construct hundreds or thousands of rules or parse-trees that are not reusable, and human involvement is indispensable to filter the frequently generated conflict candidates. Other works depend on fuzzy logic~\cite{quan2004automatic,lau2009toward,widyantoro2001fuzzy} to generate relation facts with uncertainty, which is more tractable than parsing-based techniques and do not generate conflict candidates. However, identifying or summarizing the concepts from text still requires human intelligence. Moreover, methods mentioned above suffer from the lack of reference corpora that are closely related to and highly cover the knowledge of the ontology. Moreover, to associate the right contexts of the corpora with corresponding relation facts creates another major challenge, as semantic relations in ontologies are often specialized and are not explicitly expressed in the text. Due to these issues, we have seen few successful applications of these traditional approaches in improving the coverage of large cross-domain ontology graphs like Yago and ConceptNet. These issues motivate us to consider the more flexible ``text-free'' methods based on translation-based graph embeddings \stitle{Translation-based Graph Embedding Methods.} Recently, significant advancements have been made in learning translation-based embeddings for knowledge graphs. To characterize a triple $(s, r, t)$, models of this family follow the common assumption that $\mathbf{s}_r + \mathbf{r} \approx \mathbf{t}_r$, where $\mathbf{s}_r$ and $\mathbf{t}_r$ are either the original vectors of $s$ and $t$, or the transformed vectors $f_{r}(s)$ and $f_{r}(t)$ under a certain transformation $f_{r}$ w.r.t. relation $r$. The forerunner TransE~\cite{bordes2013translating} sets $\mathbf{s}_r$ and $\mathbf{t}_r$ as the original $\mathbf{s}$ and $\mathbf{t}$. Later works improve TransE by introducing relation-specific projections on entities to obtain different $\mathbf{s}_r$ and $\mathbf{t}_r$, including projections on relation-specific hyperplanes in TransH~\cite{wang2014knowledge}, linear transformations to multiple relation spaces in TransR~\cite{lin2015learning}, dynamic matrices in TransD~\cite{ji2015knowledge}, and other forms~\cite{jia2016locally,nguyenstranse}. These variants of TransE specialize the encoding process for different relations, therefore they often achieve better representations of triples than TransE. Meanwhile translation-based models cooperate well with other models. For example, variants of TransE are trained in joint with word embeddings to enable synthesized word embeddings with relational inferences~\cite{wang2014joint,zhong2015aligning}, and are combined with alignment models to help cross-lingual knowledge alignment~\cite{chen2017multi,chen2017akbc}. However, existing translation-based models are not able to preserve triples with relational properties in the embedding spaces, because they do not differentiate the encoding of concepts that serve as different components in these triples. They also fail to provide a proper learning process for hierarchical relations. These are the major limitations we want to overcome.\par On the other hand, to enrich the knowledge in graphs, translation-based models proceed with {\em entity prediction} that predicts missing entities for triples. Since the candidate space of entities is extremely large, all these works seek to rank a set of candidates rather than acquiring the exact answers \cite{wang2014knowledge,lin2015learning,ji2015knowledge,jia2016locally,nguyenstranse}. We instead proceed with relation prediction, which practically obtains the exact answers, as the relations in ontology graphs are not very diverse \stitle{Other Knowledge Graph Embedding Methods.} There are non-translation-based methods that learn graph embeddings. UM~\cite{bordes2011learning}, SME~\cite{bordes2012joint} are simplified versions of TransE and TransR; LFM~\cite{jenatton2012latent} learns bilinear transformations among entities; {TADW}~\cite{yang2015network} learns context-based embeddings from random-walk generated contexts of the graphs (which is very similar to the recently introduced Rdf2Vec~\cite{ristoski2016rdf2vec}). These methods do not explicitly embed relations, thus do not apply to our tasks. Others include neural-based models SLM~\cite{collobert2008unified}, and NTN~\cite{socher2013reasoning} that were outperformed substantially by TransE and other translation-based methods on the tasks for populating instance-level knowledge graphs~\cite{bordes2013translating,wang2014knowledge,lin2015learning,ji2015knowledge}. There are some which perform comparably with translation-based methods, but at the cost of much higher parameter complexity, such as RESCAL~\cite{nickel2011three}, and HolE~\cite{nickel2016holographic}. We choose to compare with these two popular methods as well.

32e3305610f21ecb52df03f943ff6de9a6b022e5

\subsection*{Results} {\bf Magic wavelength determination.} The important step towards high resolution spectroscopy of the clock transition is determination of a magic wavelength of the optical lattice when dynamic polarizabilitites of the clock levels become equal~\cite{Takamoto2003}. We numerically calculated dynamic polarizabilities of the clock levels using time-dependent second order perturbation theory and transition data obtained using {\small COWAN} package \cite{cowan1981} (Model~1 described in Methods and in ref.\,\cite{Sukachev2016}), and predicted existence of the magic wavelength at 811.2\,nm (for the collinear magnetic field $\vec{B}$ and the lattice field polarization $\vec{\epsilon}$) near a narrow transition from the upper $J=5/2$ clock level at $809.5$\,nm. Calculated differential polarizability is shown in Fig.\,\ref{img:mwl} with the red solid line. This wavelength region is readily accessible by Ti:sapphire or powerful semiconductor lasers. \begin{figure}[t!] \includegraphics[width=.9\linewidth]{fig2} \caption{Magic wavelength determination. Calculated (Model\,1, red solid curve), measured (blue dots), and fitted (Model\,2, green dashed line) differential dynamic polarizability $\Delta\alpha$ (in atomic units, a.u.) between the upper ($J=5/2$) and the lower ($J=7/2$) clock levels; theoretical models are described further in the text and in Methods. Inset: zoom of the spectral region around the magic wavelength $\lambda_m=813.320(6)$\,nm. } \label{img:mwl} \end{figure} Using an approach described in~ref.\,\cite{Barber2008}, we experimentally searched for the magic wavelength for the clock transition in the spectral region of 810--815\,nm. The transition frequency shift $\Delta\nu$ as a function of optical lattice power $P$ was measured at different lattice wavelengths. The differential dynamic polarizability $\Delta\alpha$ between the clock levels was calculated using the expression $h\Delta\nu=-16 \,a_0^3 \Delta\alpha P /c \,w^2$. Here $h$ is the Plank constant, $c$ is the speed of light, $a_0$ is the Bohr radius, and $w=126.0(2.5)$\,$\mu$m is the lattice beam radius which was calculated from the enhancement cavity geometry. The intra-cavity power $P$ was determined by calibrated photodiodes placed after the cavity outcouplers M2 and M2$^\prime$. The details of the beam waist determination and power measurements are given in Methods. Figure\,\ref{img:mwl} shows the spectral dependency of $\Delta\alpha$ in atomic units~(a.u.). The magic wavelength of $\lambda_m=813.320(6)$\,nm was determined by zero crossing of the linear fit in the inset of Fig.\,\ref{img:mwl}. Trapping Tm atoms in the optical lattice at $\lambda_m$ drastically reduces inhomogeneous ac Stark broadening of the clock transition. Exciting with a 80\,ms-long Rabi $\pi-$pulses of the clock laser we recorded a spectrum with 10\,Hz full width at the half maximum shown in Fig.\,\ref{img:zeeman}(a). Non-unity excitation at the line center comes from a non-perfect initial polarization of atoms and finite lifetime of the upper clock level ($\tau=112$\,ms). \begin{figure} \includegraphics[width=1\linewidth]{fig3} \caption{Spectroscopy of the clock transition. a)~Spectral line shape of the clock transition in Tm. Every point is average of 6 measurements. The solid curve shows the fit calculated for a Fourier-limited $80$\,ms rectangular $\pi$-pulse. b)~Clock transition frequency shift $\Delta\nu$ depending on $B_y$ (dots) at $B_z=0$ and constant $B_x=225\,$mG; solid line is a parabolic fit. The dependence on $B_z$ is similar.} \label{img:zeeman} \end{figure} Using Ti:sapphire frequency comb, we determined the absolute frequency of $\ket{J=7/2,F=4}\rightarrow\ket{J=5/2,F=3}$ clock transition in Tm of $262\,954\,938\,269\,213(30)$\,Hz. The relative frequency uncertainty of $1.1\times10^{-13}$ mainly comes from instability and calibration accuracy of a GLONASS-calibrated passive hydrogen maser used as a frequency reference for the comb. {\bf Differential polarizability analysis.} In the second order approximation, the energy shift of an atomic level $\ket{J,F,m_F}$ in an external oscillating electromagnetic field with the wavelength $\lambda$ equals $-\alpha_{J,F,m_F}(\lambda)E^2/2$, where $E$ is the amplitude of the electric field. For linear field polarization, the dynamic polarizability $\alpha_{J,F,m_F}$ can be split in the scalar $\alpha_{J}^s$ and the tensor $\alpha_{J,F}^t$ parts as: \begin{equation} \label{eq:pol} \alpha_{J,F,m_F} = \alpha_{J}^s +\frac{3 \cos^2\Theta - 1}{2}\times\frac{3 m_F^2 - F (F+1)}{ F (2 F -1)} \alpha_{J,F}^t\,, \end{equation} where $\Theta$ is the angle between the quantization axis (here the direction of the external magnetic field $\vec B$) and the electric field polarization $\vec{\epsilon}$ of the optical lattice. In our case, the differential polarizability of the two clock levels equals \begin{equation} \label{eq:pol3} \Delta\alpha \equiv \alpha_{5/2,3,0}-\alpha_{7/2,4,0}= \Delta\alpha^s + \frac{3 \cos^2\Theta - 1}{2}\Delta\alpha^t\,, \end{equation} where $\Delta\alpha^s=\alpha_{5/2}^s-\alpha_{7/2}^s$ and $\Delta\alpha^t=\frac{5}{7}\alpha_{7/2,4}^t-\frac{4}{5}\alpha_{5/2,3}^t$. By definition, at the magic wavelength $\lambda_m$ the differential polarizability vanishes: $\Delta\alpha(\lambda_m)=0$. The frequency shift of the clock transition due to the optical lattice can be caused by (i) accuracy of the magic wavelength determination and (ii) angular dependency of the tensor part of the differential polarizability. The accuracy of the magic wavelength determination is related to the slope of $\Delta\alpha(\lambda)$ in the vicinity of $\lambda_m$, which is $-0.055(7)$\,a.u/nm for $\lambda_m\approx813$\,nm in Tm, as it is shown in the inset in Fig.\,\ref{img:mwl}. In more practical units, this corresponds to the clock transition frequency shift of $U\times\Delta f\times 0.30(4) $\,mHz for lattice frequency detuning $\Delta f$\,[GHz] from $\lambda_m$ and lattice depth $U$ in units of the recoil energy. This value is more than one order of magnitude smaller compared to the corresponding sensitivity of Sr and Yb lattice clocks~\cite{Barber2008,brown2017hyperpolarizability}. For $\Theta \ll 1$, the differential tensor polarizability $\Delta\alpha^t$ influences the clock transition frequency as: \begin{equation} \label{eq:tensor_shift} h\Delta\nu \approx -3/2\Delta\alpha^t \frac{E^2}{2}\Theta^2\,. \end{equation} To find $\Delta\alpha^t$, we measured dependency of the clock transition frequency shift $\Delta\nu$ on a small magnetic field $B_y$ at the constant bias fields $B_x=225\,$mG and $B_z=0$ ($\Theta\approx B_y/B_x$) shown in Fig.\,\ref{img:zeeman}(b). From the corresponding parabolic coefficient of 56(11)\,mHz/mG${}^2$, we get the differential tensor polarizability of $\Delta\alpha^t = 0.9(2)$\,a.u. at $\lambda_m$. The uncertainty comes from the absolute calibration of magnetic field and power calibration~(see~Methods). Both lattice frequency shifts mentioned in the previous paragraph can be readily reduced to mHz level by stabilizing the lattice wavelength with 0.1\,GHz accuracy and by maintaining $|\Theta| < 10^{-3}$. The frequency of $\ket{F=4,m_F=0}\rightarrow\ket{F=3,m_F=0}$ clock transition possesses a quadratic sensitivity to a dc magnetic field $B$ with a coefficient $\beta=-257.2$\,Hz/G${}^2$~\cite{Sukachev2016}. To provide uncertainty of the transition frequency below 1\,mHz, it would be necessary to stabilize magnetic field at the level of $20\,\mu$G at the bias field of $B=100$\,mG. Note, that the quadratic Zeeman shift in Tm can be fully canceled by measuring an averaged frequency of two clock transitions $\ket{F=4,m_F=0}\rightarrow\ket{F=3,m_F=0}$ and $\ket{F=3,m_F=0}\rightarrow\ket{F=2,m_F=0}$ [Fig.\,\ref{img:setup}(b)] possessing the quadratic Zeeman coefficients of the opposite signs. To implement this approach, one should provide the magic-wavelength condition for both transitions. This can be done by choosing $\Theta=\arccos{(1/\sqrt{3})}$ to cancel the tensor part in~Eq.\,(\ref{eq:pol3}) and tuning the lattice wavelength approximately to 850\,nm [Fig.\,\ref{img:pol_all}(b)]. At this wavelength the differential scalar polarizability vanishes for both transitions ($\Delta\alpha^s=0$ since $\alpha^s_{J,F}=\alpha^s_{J}$). {\bf Static differential polarizability and the BBR shift.} The BBR shift of the clock transition frequency can be accurately calculated from the static differential scalar polarizability $\Delta\alpha^s_\textup{DC}=\Delta\alpha^s(\lambda\rightarrow\infty)$ from a theoretical model based on the measured polarizability spectrum at the wavelengths of 810--860\,nm and at 1064\,nm. Measurements in the spectral region of 810--860\,nm were done by scanning the wavelength of the Ti:sapphire laser at two polarizations corresponding to $\Theta=0$ ($\vec\epsilon\parallel\vec x $) and $\Theta=\pi/2$ ($\vec\epsilon\parallel\vec y $) as shown in~Fig.\,\ref{img:pol_all}(a). The corresponding scalar $\Delta\alpha^s(\lambda)$ and tensor $\Delta\alpha^t(\lambda)$ differential polarizabilities calculated from Eq.\,(\ref{eq:pol3}) are shown in Fig.\,\ref{img:pol_all}(b). \begin{figure} [t!] \includegraphics[width=.8\linewidth]{fig4} \caption{Differential polarizabilities spectra. a) The differential dynamic polarizability $\Delta\alpha(\lambda)$ for $\Theta=\pi/2$ (red dots) and $\Theta=0$ (blue dots). b)~Corresponding scalar $\Delta\alpha^s(\lambda)$ (green dots) and tensor $\Delta\alpha^t(\lambda)$ (magenta dots) parts. Magenta cross is $\Delta\alpha^t(\lambda_m)$ determined from measurements in Fig.\,\ref{img:zeeman}(b). Solid and dotted curves are calculations based on Model 1 and Model 2, respectively~(see~Methods). } \label{img:pol_all} \end{figure} To measure dynamic polarizability at 1064\,nm, we used a slightly different procedure: Tm atoms were trapped in the optical lattice at $\lambda_m$ for which the differential polarizability vanishes, and the atomic cloud was illuminated along $y$-axis by a focused beam of a linearly-polarized single-frequency 1064\,nm fiber laser with the optical power up to 10\,W. Corresponding results for $\Theta=0$ ($\vec\epsilon\parallel\vec x $) and $\Theta=\pi/2$ ($\vec\epsilon\parallel\vec z $) are also shown in Fig.\,\ref{img:pol_all}. To compare with experimental data and to deduce $\Delta\alpha^s_\textup{DC}$, we use Model\,2 (see~Methods), which differs from Model\,1 by introducing four adjustable parameters: probabilities of the 806.7\,nm and 809.5\,nm transitions and two offsets for differential scalar and tensor polarizabilities. Corresponding fits based on Model\,2 are shown as dashed lines in Fig.\,\ref{img:pol_all}, while calculations with no free parameters (Model\,1) are shown as solid lines. From Model~2 we obtain \begin{equation}\label{eq:polresult} \Delta\alpha^s_\textup{DC}=-0.047(18)\,\textrm{a.u.} \end{equation} The differential static scalar polarizability from Model\,1 is $-0.062$\,a.u., which differs by only 1 standard deviation from~(\ref{eq:polresult}). Calculations of the BBR frequency shift can be readily done using the value of $\Delta\alpha^s_\textup{DC}$~\cite{Sukachev2016}. The differential scalar polarizability in the spectral region around $10\,\mu$m (the maximum of the BBR spectrum at the room temperature) differs by less than $10^{-3}$\,a.u. from $\Delta\alpha_\textrm{DC}^s$. Note, that there are no resonance transitions from the clock levels for $\lambda>1.5$\,$\mu$m. For our clock transition at $1.14\,\mu$m the room temperature BBR shift is $-0.45(18)$\,mHz. It is a few orders of magnitude smaller than for other neutral atoms and is comparable to the best ions species as shown in Table\,\ref{bbr_table}. This result quantitatively confirms the idea of strong shielding of inner-shell transitions in lanthanides from external electric fields. The clock transition frequency shifts due to a magnetic component of the BBR and electric field gradient of the optical lattice are less than $10^{-4}$\,Hz and can be neglected (see Methods). \begin{table}[t!] \caption{The fractional BBR shift at 300\,K for the clock transition frequencies in thulium and some other neutral atoms and ions.} \begin{ruledtabular} \begin{tabular}{cccc} &Element & $\Delta\nu^{BBR}/\nu, 10^{-17}$ \\ \hline &Tm (this work) & $-0.2$\\ &Sr \footnote[1]{ref.\,\cite{Ludlow2015}} & $-550$\\ &Yb \footnotemark[1] & $-270$ \\ &Hg \footnote[2]{ref.\,\cite{Bilicki2016}} & $-16$ \\ &Yb$^+$ (E3)\footnotemark[1] & $-11$\\ &Al$^+$ \footnotemark[1] & $-0.4$ \\ &Lu$^+$ \footnote[3]{transition ${}^1S_0 - {}^3D_1$, ref.\,\cite{Arnold2018}} & $-0.14$ \\ \end{tabular} \end{ruledtabular}\label{bbr_table} \end{table} \subsection*{Discussion} Specific shielding of the inner-shell magnetic-dipole clock transition in atomic thulium at 1.14\,$\mu$m by outer 5$s^2$ and 6$s^2$ electronic shells results in a very low sensitivity of its frequency to external electric fields. The differential static scalar polarizability of the two $J=7/2$ and $J=5/2$ clock levels is only $-0.047(18)$\,atomic units which corresponds to the fractional BBR frequency shift of the transition of $2\times 10^{-18}$ at the room temperature. It is by three orders of magnitude less compared to the prominent clock transitions in Sr and Yb (see Table\,\ref{bbr_table}). Taking into account that all major frequency shifts (the Zeeman shift, lattice shifts, collisional shifts) can be controlled at the low $10^{-17}$ level, these features make thulium a promising candidate for a transportable room-temperature optical atomic clock due to soft constrains on the ambient temperature stability. It combines advantages of unprecedented frequency stability of optical lattice clocks on neutral atoms and low sensitivity to BBR of ion optical clocks. Moreover, precision spectroscopy in Tm opens possibilities for sensitive tests of Lorentz invariance \cite{Shaniv2018} and for search of the fine structure constant variation \cite{kolachevsky2008high}. Optical clocks based on a $f$-$f$ transition in some other lanthanides with spinless nuclei could be even more attractive featuring the low sensitivity to magnetic fields due to the absence of the hyperfine structure and small BBR shift. For example, the fine-structure clock transition at the telecom-wavelength of 1.44\,$\mu$m in laser-cooled erbium atoms (e.g. $^{166}$Er)~\cite{McClelland2006,Kozlov2013} can be particularly interesting for optical frequency dissemination over fiber networks~\cite{Riehle2017}. \subsection*{Methods} {\bf Enhancement cavity}. Optical lattice is formed inside a $\Gamma$-shaped enhancement cavity, as shown in Fig.\,\ref{img:setup}(a). The reflectivity of the curved (the radius is $r=-250$\,mm) incoupler mirror M1 equals 87\% and matches losses introduced by the vacuum chamber viewports. Outcouplers M2 or M2$^\prime$ are identical flat mirrors with the reflectivity of $R>99\%$. For locking to the laser frequency, the cavity mirror reflecting the beam at $45^\circ$ is mounted on a piezo actuator. The intra-cavity polarization is defined by a broadband polarization beam splitter; depending on the polarization, either M2 or M2$^\prime$ outcoupler mirror is used. The intra-cavity lens has the focal length of $f=400$\,mm. Depending on experimental geometry ($\Theta=0$ or $\Theta=\pi/2$), we couple corresponding linearly polarized radiation from the Ti:sapphire laser through the incoupler mirror~M1. Intra-cavity polarization filtering by PBS defines the polarization angle and significantly improves polarization purity of the optical lattice. The angle between the laser field polarization and the bias magnetic field is adjusted with the accuracy of better than $1^\circ$. {\bf Measurement of differential dynamic polarizabilities.} Differential polarizability $\Delta\alpha$ of the clock levels is determined from the frequency shift of the corresponding transition $\Delta\nu$, circulating power $P$, and TEM$_{00}$ cavity mode radius $w$ at the atomic cloud position as \begin{equation}\label{eq4} \Delta\alpha =-\frac{hc w^2}{16 a_0^3}\frac{\Delta\nu}{P}\,. \end{equation} The dependency $\Delta\nu(P)$ was obtained for different wavelengths in the spectral range 810--860\,nm for the intra-cavity circulating power varying from 1\,W to 4\,W, as it is shown in Fig.\,\ref{img:s2}. Frequency shifts $\Delta\nu$ were measured relative to the laser frequency, which is stabilized to an ultra-stable ULE cavity with linear drift compensation. The slope coefficients of corresponding linear fits were substituted to Eq.\,(\ref{eq4}) to deduce $\Delta\alpha$ presented in Figs.\,\ref{img:mwl},\,\ref{img:pol_all}. \begin{figure}[h!] \includegraphics[width=.8\linewidth]{fig5} \caption{ Clock transition frequency shift $\Delta\nu$ as a function of optical lattice power $P$ in the vicinity of the magic wavelength $\lambda_m=813.32$\,nm. Experimental data are fitted by linear functions.} \label{img:s2} \end{figure} Uncertainty of the frequency shift $\Delta\nu$ comes from the residual instability of the reference cavity on time intervals of 1000\,s. To estimate it, we measured clock transition frequency relative to the clock laser frequency at the magic wavelength when the perturbations from the lattice are minimal. Results are shown in Fig.\,\ref{img:s3}. The standard deviation equals to 2.6\,Hz contributing $0.003$\,a.u. to the error budget of $\Delta\alpha$. For the lattice wavelength detuned from $\lambda_m$ contribution of the laser frequency instability is negligible. \begin{figure}[h!] \includegraphics[width=.8\linewidth]{fig6} \caption{Relative frequency of Tm clock transition and ULE cavity mode with linear drift compensation. Each data point and corresponding uncertainty comes from the fit of the clock transition spectrum. The shaded region corresponds to 1~standard deviation of the data set.} \label{img:s3} \end{figure} The intra-cavity power $P$ was determined by measuring power leaking through the cavity outcoupler M2 (or M2$^\prime$) using calibrated photodiodes. For each photodiode we measured the power-to-voltage transfer function $P(U) = \kappa U$, where $U$ is the voltage reading from the photodiode, and $\kappa$ is the coefficient measured using absolutely calibrated Thorlabs S121C power meter. To determine $\kappa$, we unlocked the cavity, slightly tilted the outcoupler, blocked the reflected beam to prevent feasible reflections from the incoupler, and measured the power before the outcoupler and corresponding voltage reading of the photodiode. The linearity of the photodiode response was checked separately and turned out to be better than $3\%$ in the working region. The photodiode calibration was done in the whole spectral range of 810--860\,nm to take into account the spectral response of the outcoupler and the photodiode. Although the specified uncertainty of the power sensor Thorlabs S121C equals 3\%, we ascribe the net uncertainty of power measurement of 10\% from comparison of readings from three different absolutely calibrated sensors. The beam radius $w$ at the atomic cloud position is deduced from the cavity geometry: distances from the vacuum chamber center (and target atomic cloud position) to M1, L and M2 (or M2$^\prime$) are 244\,mm, 384\,mm and 500\,mm, respectively, giving the beam radius of $w=126\,\mu$m at the position of atomic cloud. The uncertainty of $w$ comes from the position uncertainties of cavity elements and of the atomic cloud with respect to the chamber center, as well as from uncertainty of 1\,mm of the lens focal length. We conservatively evaluate position uncertainties of M1, L, and M2 (or PBS and M2$^\prime$) as 1\,mm, and the possible axial displacement of the atomic cloud of 2\,mm. The partial contributions to the beam radius uncertainty are $1.6\,\mu$m (the incoupler), $1.2\,\mu$m (the lens), $0.03\,\mu$m (the outcoupler), $0.3\,\mu$m (the cloud) and $1.5\,\mu$m (focal length). Adding up in quadratures, the total uncertainty of the beam radius $w_0$ equals $2.5\,\mu$m. The result is independently confirmed by measuring frequency intervals between cavity transversal modes. The finite temperature of the atomic cloud reduces averaged light intensity on the atomic cloud $I_\textrm{av}$ in respect to the on-axis antinode lattice intensity $I_0=8P/\pi w^2$. Assuming Boltzman distribution of the atoms with the temperature $T$ in the trap of depth $U_0$ one can calculate the parameter $\eta$ connecting the averaged and maximum intensity $I_\textrm{av} = \eta I_0$. The $\eta$ parameter is calculated from \begin{equation} \eta = \frac{\int\limits_{0}^{U_0} e^{-E/kT} \left(\frac{1}{2 r_0} \int\limits_{-r_0}^{r_0} e^{-2r^2}dr\right)\,dE}{\int\limits_{0}^{U_0} e^{-E/kT}dE}\,, \end{equation} where $r_0 = ({-\ln(1-E/U_0)/2})^{-1/2}$ is the classical turning point. The parameter $\eta$ equals $0.90(5)$ for $kT=0.3\,U_0$ (which corresponds to average experimental conditions). One standard deviation $\sigma=0.05$ of $\eta=0.9$ corresponds to $kT$ range from $0.14\,U_0$ to $0.6\,U_0$, while $2\,\sigma$ already covers ($0$\,-\,$3)\,U_0$ range. Summarizing, we evaluate the uncertainty of measured differential polarizability $\Delta\alpha$ to be 13\% with the dominating contribution from power calibration. In the vicinity of $\lambda_m$, the uncertainty is slightly higher by 0.003\,a.u. because of the reference clock laser frequency instability. To measure differential polarizability at 1064\,nm, atoms were trapped in the optical lattice at the magic wavelength and irradiated by focused slightly elliptic 1064\,nm laser beam with the waists of $w_x=320(20)\,\mu$m and $w_y=280(20)\,\mu$m in $x$ and $y$ directions, respectively. To adjust the 1064\,nm laser beam center to the atomic cloud, we maximized intensity of the beam on the atomic cloud by monitoring the frequency shift of the clock transition. To increase sensitivity, corresponding measurements were done by tuning the clock laser to the slope of 1.14\,$\mu$m transition. For the measurement session of $\Delta\alpha$ at $\Theta=0$, we performed three adjustments of 1064\,nm laser beam. The reproducibility corresponds to the frequency shift of 5\,Hz at the maximal frequency shift of 25\,Hz. The resulting linear coefficient is evaluated as $\Delta\nu/P=-3.1(6)$\,Hz/W including this uncertainty. In $\Theta=\pi/2$ configuration, we did not observe any significant effect from adjustment and the coefficient equals $0.04(24)$\,Hz/W. {\bf Theoretical analysis.} Theoretical approach to calculate polarizabilities of Tm atomic levels is described in our previous works~\cite{Sukachev2016,golovizin2017methods}. Calculations are based on the time-dependent second order perturbation theory with summation over known discrete transitions from the levels of interest. For calculations, we used transitions wavelengths and probabilities obtained with the {\small COWAN} package \cite{cowan1981} with some exceptions: for transitions with $\lambda>800$\,nm, experimental wavelengths from ref.\,\cite{NIST_ASD} are used. This approach allows to increase accuracy of the magic wavelength prediction in the corresponding spectral region of $\lambda>800$\,nm. According to calculations, the magic wavelength was expected at 811.2\,nm which motivated our experimental studies in the spectral region 810--815\,nm (see Fig.\,\ref{img:mwl}). We refer to this model as ``Model\,1'' and use for comparison with experimental results of this work as shown in Figs.\,\ref{img:mwl},\ref{img:pol_all}. Deviation of the experimental data from Model\,1 can be explained by two main factors. First, in this model we did not take into account transitions to the continuum. Together with uncertainties of {\small COWAN} calculations of transition amplitudes, it can result in a small offset of the infrared differential polarizability spectrum. Note, that although transitions to the continuum spectrum may significantly contribute to polarizabilities of the individual levels (up to 10\%), for the differential polarizability of the $f$-$f$ transition at 1.14\,$\mu$m in Tm contribution of the continuum is mostly canceled \cite{Sukachev2016}. \begin{table}[h] \caption{Uncertainty budget for the differential scalar polarizability $\Delta\alpha^s_\textup{DC}$.} \begin{ruledtabular} \begin{tabular}{l|c} Source & Uncertainty, a.u.\\ \hline Experimental results for 810--850\,nm & $0.013$ \\ Experimental result for $1064$\,nm & $0.006$ \\ Angle $\Theta$ & $0.002$\\ Transition probabilities for $\lambda>900$\,nm & $0.01$ \\ \hline \bf{Total} & $\bf{0.018}$ \\ \hline For the reference: & \\ Difference of Models 1 and 2& $0.015$\\ \end{tabular} \end{ruledtabular}\label{table2} \end{table} To fit experimental data we use Model\,2 which differs from Model\,1 by introducing four fit parameters. As parameters we use the probabilities of the 806.7\,nm and 809.5\,nm transitions which mostly effect polarizability spectrum in 810--860\,nm region and two offsets for the scalar and tensor polarizabilities. After fitting the experimental data (see Fig.\,\ref{img:pol_all}) by Model\,2, the probability of the 806.7\,nm transition is changed from $3473\,\text{s}^{-1}$ to $4208(298)\,\text{s}^{-1}$, the probability of the 809.5\,nm transition is changed from 149\,s$^{-1}$ to 357(109)\,s$^{-1}$, the fitted offsets for the differential scalar and tensor polarizabilities equal $0.012(6)$\,a.u. and $-0.028(12)$\,a.u. with reduced $\chi_s^2$ for the fits of 1.35 and 2.9, respectively. Transitions from the upper Tm clock level $J=5/2$ in the spectral range $\lambda>900$\,nm are weak and their probabilities are experimentally not measured. To calculate probabilities we used {\small COWAN} package. To estimate the impact of insufficient knowledge of transition probabilities on differential scalar polarizability $\Delta\alpha^s(\lambda)$, we assume possible variation of each transition probability by a factor of 2. After extrapolation of the fitted Model\,2 to $\lambda\rightarrow\infty$ we get the static differential polarizability $\Delta\alpha^s_\textup{DC}=-0.047^{+0.01}_{-0.005}$\,a.u., where uncertainty comes from variation of $\lambda>900$\,nm transition probabilities. We summarize all sources of uncertainties which contribute to the error of the static differential polarizability $\Delta\alpha^s_\textup{DC}$ in Table\,\ref{table2}. As discussed above, the experimental uncertainty for 810--860\,nm range is 13\% contributing $0.013$\,a.u. to the $\Delta\alpha^s_\textup{DC}$, while the measurement at 1064\,nm is less accurate (20\%) due to the laser beam adjustment and results in $0.006$\,a.u. variation of Model\,2 extrapolation. The uncertainty of the angle $\Theta$ adjustment contributes $0.002$\,a.u. The uncertainty coming from the poorly known transition probabilities from $J=5/2$ clock level in $\lambda>900$\,nm range contributes $0.01$\,a.u. Using extrapolation of Model\,2 and adding all uncertainties we come to the final result of $\Delta\alpha^s_\textup{DC}=-0.047(18)$\,a.u. Note, that this is fully consistent with the extrapolated value of -0.062\,a.u. obtained from Model\,1 and given for the reference in Table\,\ref{table2}. {\bf BBR magnetic field.} To estimate the clock transition frequency shift due to the magnetic component of BBR, we follow the analysis given in the work~\cite{gan2018oscillating}. The corresponding frequency shift of one of the clock levels coupled to another atomic level with magnetic-dipole transition at frequency $\omega_0$ can be found by integrating over the full BBR spectrum as \begin{equation} \begin{split} \Delta\nu^B_{bbr}(T) &= -\frac{\omega_0 }{2\pi}\frac{\mu_B^2}{2 \hbar \pi^2 c^5 \epsilon_0}\int_0^\infty \frac{1}{\omega_0^2-\omega^2}\frac{\omega^3}{e^{\hbar \omega/kT} -1 }d\omega \\ & = -\frac{\omega_0 }{2\pi}\frac{\gamma}{2} \left(\frac{T}{T_0}\right)^2 f(y), \end{split} \end{equation} where $\epsilon_0$ is the vacuum permittivity, $T_0=300$\,K, $y=\hbar\omega_0/k_B T$. Here \begin{gather} \gamma = \frac{\mu_B^2}{\hbar^2}\frac{\hbar}{6 c^5 \epsilon_0} \left(\frac{k_B T_0}{\hbar}\right)^2 \approx 9.78\times10^{-18}, \\ f(y) = \frac{6}{\pi^2}\int_0^\infty\frac{1}{y^2-x^2}\frac{x^3 dx}{e^x-1}. \end{gather} The hyperfine transition frequency $\omega_0$ of the $J=7/2$ ground level in Tm equals $2\pi\times1496$\,MHz, while for the $J=5/2$ clock level it equals $2\pi\times2115$\,MHz. For these values of $\omega_0$ $y \ll 1$, $f(y)\approx -1$, and the shift is on the order of $10^{-8}$\,Hz. To estimate the contribution from the optical transitions, we evaluate the shift from the lowest frequency magnetic-dipole transition at $\omega_0 \approx 2\pi\times263$\,THz, which is the clock transition itself: $y=42$, $f(y)=2.3\times10^{-3}$ and $\Delta\nu^B_{bbr}(T_0)=-3\times10^{-6}$\,Hz for the ground level (for the upper clock level the corresponding shift is $+3\times10^{-6}$\,Hz). Hence, we estimate total shift of the clock transition from the magnetic component of BBR to be less than $10^{-4}$\,Hz. {\bf Electric quadrupole shift.} Opposite to neutral Sr, Yb, and Hg atoms, Tm clock levels posses a non-zero electric quadrupole moment on the order of 1\,a.u.~\cite{Sukachev2016} and are coupled to an electric field gradient. Since the electric field gradient in an optical lattice oscillates at the optical frequency, the corresponding time-averaged frequency shift of the clock transition is zero. \subsection*{References}

860fc35ba7ac5da2ce3f1c654f3eee69b66c4610

\section{Introduction} Understanding the mechanisms underlying the emergence and persistence of new cell types is a central problem in the evolution and development of multicellular organisms. Whereas all cell types can in principle access the same genetic information, in practice, regulation of gene expression restricts this such that only a subset of an organism's total genomic information content is accessible to a given cell type at a given time, permitting differentiation of many phenotypes from a single genotype \cite{Slack2013}. Regulation of gene expression therefore plays a dominant role in establishing cell types. From a formal point of view, the question of how new cell types emerge, therefore reduces to the problem of understanding how new regulatory structures evolve that can specify and control the expression of novel phenotypes. The interplay among these regulatory genes, and their interaction with the other components of the cell governs the expression levels of both mRNA and proteins, where the set of interactions are described as a {\it gene regulatory network} (GRN). \rv{ If multicellular phenotypes are the product of developmental differentiation processes controlled by GRNs, as the raw material for evolutionary change, every phenotypic variant is expected to be the product of a corresponding change in those regulatory networks. In the case of evolution, continuous feedback in the form of selection or drift\cite{Lynch2007} leads to changes in the underlying network architecture controlling individual phenotypes, for instance by means of mutation, gene duplication or deletion, etc. In the case of developmental differentiation, continued extracellular feedback over the regulatory network assures that the steady gene activation patterns (i.e. the ID of the cell) follow a specific developmental sequence. Such extracellular signalling is not required to change the underlying structure of the GRN for steady activation pattern to change. Nonetheless, this signalling induced change can have the same phenotypic effect as a structural change to the network (as during evolutionary change). In this manuscript, we examine the possibility of achieving a common description of both evolutionary change and developmental processes in GRN models. When a GRN is described with the language of dynamical systems, extracellular feedback becomes formally equivalent to the notion of network controllability, and the interested reader is redirected to the excellent book of Iglesias and Ingalls for an extensive review of the subject \cite{Iglesias2010}. Things change drastically when the evolutionary timescale is involved.} Any attempt at trying to predict the steady states of the regulatory process --i.e. the cell types-- in terms of a dynamical model (coupled ordinary differential equations, \rv{Boolean} network models, stochastic gene networks, to name a few common approaches) faces the difficulty of having to reconcile the fixed number of genes in these models, whose expression level is representative of a given cell type \cite{Stegle2015,Shapiro2013,Schwartzman2015}, with the possibility for the size of the genotype to change over evolutionary timescales. Chromosome loss and gain, single gene and whole genome duplication, as well as horizontal gene transfer all alter the number of genes participating in the dynamics of a GRN. In doing so, these processes deprive the mapping of cell types to gene expression patterns of its original meaning. Therefore, even when successful in explaining developmental change, current GRN models must be redefined after each modification of the genotype for their use in evolutionary biology. In this manuscript, we generalize the well established framework of \rv{Boolean}, dynamical models of GRNs as proposed by Kauffman \cite{Kauffman1969} \rv{--based on the hypothesis that cell types represent the attractor states of the GRN dynamics \cite{Delbruck1949,Jacob1961}--} to include features of evolutionary biology. In \rv{\rv{Boolean}} models, the attractors of the network dynamics encode different, stable cellular phenotypes, permitting a model for how multiple cell type identities can be encoded in the same regulatory structure. The novelty of our approach consists in relaxing the restrictive one-to-one mapping between network attractors and cell phenotypes by redefining phenotypes as {\it collections of gene expression patterns} with a given subset of genes sharing the same pattern (section 2). While the traditional definition, assuming a one-to-one map between phenotype and genotype, yields increasingly fine-tuned specifications for the phenotype for progressively larger genotypes, our novel definition \rv{identifies} the phenotypes with a {\it macrostate}, as opposed to individual (micro)states, of a dynamical system, and as such does not require fine-tuning. We will show that, under the relaxed assumption of identifying phenotypes as macrostates of the underlying \rv{\rv{Boolean}} GRN, a fixed genotypic size is not necessary for specifying or retaining phenotypes through evolutionary processes. We will exploit this possibility to study the emergence of new cell types, as well as the consolidation or loss of old types \rv{Section 2}, as a consequence of the changing size and topology of the GRN over evolutionary timescales, and of shifting environmental conditions. As such, our model also addresses an inconsistency arising from considering concepts belonging to different levels of description of a well conceived ontology of biological objects \cite{Laubichler2017} as being modeled as {\it same-level} processes: for example, gene expression levels and phenotypes, as they were interchangeable. Our approach will instead assume gene expression to be at a lower level of the ontology than phenotype, while phenotype and environment will belong to the same, higher ontological level. In what follows, we focus on the case study of gene duplication, and use it as an example of genotype-changing evolutionary process. Gene duplication has occurred in all three domains of life \cite{Kondrashov2002}, and is an ancient mechanism dating to before the last universal common ancestor of all life on Earth \cite{Conant2008}. \rv{Most genomic evolutionary processes include at least some gene duplications. It} is, by far, the dominant force in creating new genes, and at least 50\% of genes in prokaryotes \cite{Brenner1995,Teichmann1998} and over 90\% of those in eukaryotes \cite{Gough2001} are the result of gene duplication. Nonetheless, with the exception of a few papers \cite{Wagner1994,Aldana2007,Crombach2008}, it has rarely been discussed as a mechanism for evolving new regulatory patterns for cell type identity. \rv{Our preliminary study seems to suggest} an interplay between gene duplication and natural selection as the driving force responsible for the assemblage of genetic {\it modules}, or {\it core sets of regulatory genes} \cite{Bonner1988,Wagner1996, Alberts1998,Hartwell1999,Pereira-Leal2006,Pereira-Leal2007,Achim2014}. Evidence in favor of modularity in biology has steadily increased over the last three decades, and, as of today, modules have been both reconstituted {\it in vitro} \cite{Khalil2010}, and transplanted from one cell type to another \cite{Hsu1993}. It is now well accepted that biological functions are only rarely attributed to individual molecules - the role of hemoglobin in transporting oxygen along the bloodstream being among the best examples. Far more often, a biological function results instead from the interaction among many different proteins, like in the transduction process converting pheromone detection into the act of mating in yeast \cite{Stock1996,Herskowitz1995,Posas1998}. Network controllability is usually reinterpreted in GRN dynamical models as the mathematical counterpart of extracellular signaling. Our macrostate interpretation of the phenotypes still allows us to adopt and assign biological meaning to network controllability techniques. For example, the control kernel of a GRN is defined in \cite{Kim2013} as the minimum number of genes/nodes whose expression it is necessary to regulate to steer the dynamics of the rest of the network toward a desired attractor (phenotype), {\it e.g.} an attractor associated with a functional cellular phenotype. Focusing on the developmental timescale we show it is still possible to easily rephrase, generalize, and adapt the notion of {\it control kernel} within our new framework. Therefore, our unified framework, while being suitable to mechanistic studies of GRN mutations over evolutionary timescales, is still able to describe developmental change. The manuscript is structured as follows: The next section contains a brief review of dynamical \rv{\rv{Boolean}} models of GRNs sufficient to orient those not familiar with the general theory. It reviews the main features of Kauffman's seminal theory, and of the identification between cell types and attractors. We then expose our core idea that the cell types are more properly described by collections of attractors, i.e. by macrostates of the dynamics. We then conclude the section by drawing an example network that we will use in the sections 3 and 4 as a toy model to explore the consequences of our hypothesis. Section 3 is devoted to the main consequence of our approach. We show how gene duplication and mutation events alter the nature of the cell types expressed as a consequence of selective pressure induced by a shifting environment. \rv{In particular, we show the connection between the presence of mutations and the open-ended evolution of the selected phenotype.} Section 4 shows how our generalized approach does not inhibit the possibility of adopting network controllability methods to describe epigenetically induced developmental change. We again use the same toy model to illustrate key concepts, and show extracellular control is now more easily achieved than in Kauffman's original framework. \section{Boolean Models} This section describes the mathematical details of the \rv{Boolean}, dynamical model we will assume in the rest of this manuscript. In living tissues, the intrinsic patterns of gene expression coupled with signaling input dictates cell fate. Both of these processes can readily be modeled by a \rv{\rv{Boolean}} network. \rv{Boolean} networks were originally proposed by Kauffman \cite{Kauffman1969} as a viable mathematical model of GRNs. They permit exploration of the complex steady-state dynamics of GRNs, where the attractors of the dynamics can be identified with different cell types/fates, e.g. quiescence, proliferation, apoptosis, differentiation, etc. \rv{In this manuscript, we anchor our discussion to Boolean networks as they represent the simplest mathematical model exhibiting biological and systemic properties of real GRNs \cite{Vladimir2005,Faure2006}. By virtue of this simplicity they are particularly easy to interpret biologically.} \subsection{Review of Boolean Models} At a given time $t$, the state of the \rv{Boolean} model of a GRN is known when the state $x_i(t)$ of every gene $i$ is known. The \rv{\rv{Boolean}} nature of the model assumes two possible states for gene $i$: active, which corresponds to $x_i(t)=1$, or inactive, with $x_i(t)=0$. For a network describing the interactions among $n$ genes, the state at time $t$ is specified by an $n-$dimensional \rv{Boolean} array $x_1(t),\dots,x_n(t)$. The dynamics of the GRN is then described by $n$ \rv{Boolean} functions $f_1,\dots,f_n$ which provide the state of the network at time $t+1$, given its state at time $t$ (synchronous update, but asynchronous models are also possible \cite{Garg2008}): \begin{eqnarray} x_1 (t+1) & = & f_1 (x_1(t),\dots,x_n(t)) \nonumber \\ \cdots \ \ \qquad & & \nonumber \\ x_n (t+1) & = & f_n (x_1(t),\dots,x_n(t)) \ , \label{system1} \end{eqnarray} For $n$ genes, $2^n$ possible states (gene expression patterns) exist, and the dynamics of the network is represented by a trajectory (a time series) in the discrete space containing the totality of these states. For deterministic functions $f_i$, and because of the finite size of this state space, these trajectories will eventually converge to either a fixed state or a cycle of states. These special activation patterns take the name of {\it attractors} of the \rv{Boolean} network, and were identified by Kauffman as corresponding to the stable phenotypes of the GRN. Nothing has been said so far about the functional form of the \rv{Boolean} functions $f_i$, and readers interested in an insightful analysis of how constrained such a specification is from experimental data are directed to \cite{Henry2013,Zhou2016}. For reasons that will become clearer in section 3, and given the conceptual nature of this manuscript, we will restrict ourselves to the case of \rv{\rv{Boolean}} network with thresholds: \begin{equation} x_i(t+1) = \textrm{sgn} \left( \sum_{j=1}^n a_{ji} x_j(t) - b_i \right) \ , \label{updating_rule} \end{equation} where $a_{ji}$ is the relative weight of the regulatory signal from gene $j$ to gene $i$ (activation when $a_{ji}$ is positive, inhibition otherwise), $b_i$ is the activation threshold of gene $i$, and sgn($x$) is a unitary step function, defined by $\textrm{sgn}(x)=0$ if $x \leq 0$ but $\textrm{sgn}(x)=1$ if $x>0$. In a more compact notation, we can write \[ X(t+1)=\textrm{sgn} \big(A^T \cdot X(t)-B\big) \ , \] where we have introduced the {\it adjacency matrix} $A=(a_{ij})$, and the columns $B=(b_i)$ and $X=(x_i)$. A useful feature of these models is that they convey the topology of the network, implicit in the definition of generic \rv{Boolean} functions $f_i$, in a very explicit way, with the edges of the network representing non-null entries of the adjacency matrix. Now that we have laid the groundwork for the numerical models we will be using in the rest of this manuscript, let us go back to the interpretation of the attractors of a \rv{\rv{Boolean}} network as the phenotypes of a GRN. Given this identification, robustness and evolvability of GRNs can be mapped directly to the \rv{evolutionary changes} of the attractor landscape of the corresponding \rv{Boolean} model \cite{Aldana2007,Ciliberti2007a,Ciliberti2007b}. In this framework, the emergence of new phenotypes has a precise mathematical meaning as the acquisition of new attractors, which can arise due to mutations in the network structure. Many steady-state attractors are permitted in a \rv{Boolean} network, making it an ideal model for describing how the genomic information contained within a single initial fertilized egg is differentially expressed in so many distinct cell types in response to different regional specification and morphogenetic histories \cite{Boveri1906}. In the absence of external regulation, the likelihood of a given cell type, or phenotype, is quantified in terms of the number of initial configurations of the network converging on the attractor state encoding that cell type. The higher the number of initial configurations leading to the same equilibrium dynamics, i.e. the production of a well defined set of proteins, the more likely the cell type that set of proteins represents will be. We will refer to these probabilities as the {\it basins of attraction} of the possible cell types encoded in the GRN. On the other hand, the effect that external regulation, in the form of extracellular signaling during development, might have in determining the expression of a specific subset of genes, is even more dramatic, as it might force the development toward cell types that were not even initially accessible to the unperturbed GRN. The minimum number of genes whose expression needs to be controlled, for the cell fate to be determined, is the {\it control kernel} associated to that cell type \cite{Kim2013}, as described in the introduction. \subsection{GRNs as Evolvable Systems} Most of the predictions derived within the framework we have just reviewed rely on the assumption that different cell phenotypes correspond to different attractors of the GRN. While the connection between cell types and attractors is both numerically and experimentally well-motivated, their one-to-one correspondence might be an \rv{artifact} due to the diminutive size of the mathematical models that were actually solvable in the not-so-distant past, when a small number of attractors were naturally identified with different phenotypes. Recent advances in computing power are finally making the study of much larger networks possible. One interesting example is the recent dynamical model developed by Fumi\~a and Martins \cite{Fumia} for the integration of the main signaling pathways involved in cancer. The signaling among almost 100 different genes is responsible for 63 different attractors, that eventually correspond to {\it only three} phenotypically distinct and incompatible cell fates: apoptotic, quiescent, and proliferative. An important observation is that these phenotypes are determined by just a small subset of values, e.g. the constant activation of the effector caspases in apoptotic cells, within a much larger gene activation pattern. A second example of several attractors all sharing the same phenotypic identity is already present in the much smaller GRN describing cell-fate determination during Arabidopsis thaliana flower development \cite{Espinosa-Soto2004}. Four among ten possible attractors all represent the same inflorescence meristematic cell type. Therefore, while most of the literature on network robustness and evolvability focuses on individual attractors, the concept of phenotype --that we redefine here as a macro-state of the cell characterized by the activation values (fixed or cycling) of only a subset of gene nodes-- seems to be better suited to the idealization of biological systems. While an attractor can represent a phenotype by itself, a phenotype can be compatible with multiple attractors. Our definition is exemplified in the following diagram, showing the case of a phenotype $\Phi$ defined by the expression of the two genes represented by dark gray boxes, and the \rv{non expression} of the two genes shown in light gray tone: \begin{center} \includegraphics[width=.9\columnwidth]{phenotype.png} \end{center} The expression of any other gene (boxes with a question mark) can take any value, as long as the four genes entering the definition of $\Phi$ exhibit the right expression pattern.\\ \begin{figure*}[!t] \centering \includegraphics[height=.55\textwidth]{g0.pdf} \caption{\rv{Boolean threshold network (as described in subsection 2.1) for the GRN of cell type 1. The subset of genes defining the phenotype are represented by nodes 1, 2, and 3. The characteristic pattern of cell type 1 is assumed to have genes 1 and 3 expressed, and inactive gene 2 (blue = active, white = inactive). In order to reproduce our example, the reader needs to assume equal weight edges $a_{ij}=\pm 1$ for activating/inhibiting links (continuous/dashed line), non-zero thresholds $b_1=-1$ and $b_{5,6,8}=1$.}} \label{GRN} \end{figure*} A critical motivation for this redefinition of the mathematical identity of a phenotype is a conceptual difficulty which naturally arises in Kauffman's original framework \cite{Kauffman1969}. Kauffman's definition loses its strict meaning in light of evolutionary biology: the identification of the attractors of a \rv{\rv{Boolean}} network with the phenotypes of a GRN requires --after the regulatory network has lost or acquired genes-- the comparison between genotypes of different lengths. An immediate, advantageous consequence of our definition is it keeps its meaning even after changes occur in the genotype. As an example, the following diagram, shows the case of a single gene duplication event enlarging the genome expressing $\Phi$: \begin{center} \includegraphics[width=.9\columnwidth]{duplication.png} \end{center} As this mutation is not affecting any of the defining genes, this change in the size of the genome does not alter our definition of $\Phi$. Therefore, it makes sense to study whether $\Phi$ is still expressed by the new GRN, and whether the gene replica is favoring or disfavoring $\Phi$'s basin. As we will show in section 3, these questions can be addressed in quantitative terms. The inclusion of different gene replicas will induce different changes in the relative size of the basin of attraction of $\Phi$, and a good dynamical model of the GRN will be enough to determine the duplication events that increase the basin of $\Phi$. When $\Phi$ is selected, these are likely to be the duplication events that are fixed more often. A second, less relevant difference between our approach and those outlined in previous literature on GRNs, is that we will not require the basin of attraction of a specific cell type to be exactly 100\% for the network to be representative of that specific type. Instead, we will only require one basin to be much larger than the remaining others. In the following example, aimed at making these ideas more concrete, we will assume 80\% as the threshold a basin needs to exceed for the remaining attractors to be treated as negligible. \subsection{A Sample Network} In the next two sections, we will explore the consequences of our approach in studying evolutionary and developmental processes \rv{within} the \rv{\rv{Boolean}} model framework. To anchor our discussion to a concrete example, we will introduce a small, tractable network that we will adopt as a toy model, and modify as needed. The example is shown in figure \ref{GRN}. It represents the GRN expressing a phenotype that we will call {\it cell type} 1. The subset of genes defining the phenotype are represented by nodes 1, 2, and 3. The characteristic pattern of cell type 1 is assumed to have genes 1 and 3 expressed, and inactive gene 2 (blue = active, white = inactive in figure \ref{GRN}). We will refer to this by saying that cell type 1 has expression pattern 101. In determining the basin of attraction of cell type 1, we will sum the basins of every attractor of the GRN which is compatible with the expression pattern 101, i.e. every pattern of the form $1,0,1,x_4,\dots, x_8$, where $x_4,\cdots,x_8$ (gray nodes) can take any possible (binary) value representing the expression/suppression of the remaining genes labeled from 4 to 8. The GRN of figure \ref{GRN} actually encodes two possible cell types, one being type 1, as we said, the other having the activation pattern 110. We will refer to this second phenotype as {\it cell type} 2. Despite encoding both phenotypes, the basin of attraction of cell type 1 includes more than 80\% of the possible initial activation patterns, all leading to the equilibrium dynamics characterized by the expression of genes 1 and 3, and the suppression of gene 2. This is why we assume this network to be a viable description of cell type 1. For the reader interested in reproducing our example: The GRN dynamics is modeled using a \rv{\rv{Boolean}} threshold network (as described in section 2.1), with equal weight edges $a_{ij}=\pm 1$ for activating/inhibiting links (continuous/dashed line in figure \ref{GRN}), non-zero thresholds $b_1=-1$ and $b_{5,6,8}=1$. But it is important to remark that none of the conclusions we will draw in the next sections depend on the mathematical details of the example we are using. \section{Evolution} We have seen how our re-definition of the phenotypes of a GRN in the framework of \rv{Boolean}, dynamical models is \rv{relatively robust towards mutations changing the size of the genome}. In this section, we seek to show an explicit example of how this allows the prediction of the genes whose duplication will reinforce a phenotype favored by natural selection. The underlying idea is exemplified in the following diagram (same example of phenotype $\Phi$ adopted in subsection 2.2), with arrows showing the regulation induced by an ideally connected \rv{gene} which activates exactly the \rv{genes} that are supposedly expressed, and suppresses just the genes that are inactive in our former definition of $\Phi$: \begin{center} \includegraphics[width=.9\columnwidth]{driver.png} \end{center} It is easy to foresee that the duplication of this ideal \rv{gene} would positively impact the change in the basin of attraction of $\Phi$. In a more realistic scenario, duplication would be followed by divergence, represented in the following diagram by a difference in one of the genes regulated by the replica: \begin{center} \includegraphics[width=.9\columnwidth]{mutated_driver.png} \end{center} In this section we will consider duplication events like the one just described. For each \rv{gene that does not enter the definition of the phenotypes} in the GRN, we will consider the entire spectrum of possible events of {\it duplication} + {\it divergence}, and show ideal pathways leading to cell type changes in a shifting environment. With reference to figure \ref{selection}, we will describe a simple evolutionary model for the differentiation of progenitor cells of type 1, as encoded in our network from subsection 2.3, into cells of type 2 (already encoded, but unlikely), and the newly born {\it cell type} 3, a novel phenotype induced by the modifications the network is going through. We are assuming here that the knowledge of the environment is equivalent to the knowledge of the selected cell type: they are on the same ontological level, and {\it share the same mathematical definition} in our theoretical model. The environmental pressure changes in both space and time, and selects cells that undergo mutation events that favor the emergence of the newly preferred cell type. In this example, sister cells \cite{Arendt2008} of type 1 start being selected for cell type 2, and undergo mutation events \circled{A} and \circled{B} until cell type 2 represents the dominant phenotype. In our model, the mutation events are represented by gene duplication and divergence of one of the \rv{genes not included in the definition of the phenotypes} (nodes 4--8 in fig. \ref{GRN}), during which the network acquires a non-mutated replica of a preexistent \rv{gene}, and then a mutation in the way the replica regulates the remaining genes and itself. Let us first provide a mathematical description of what we mean by a perfect replica of a preexistent gene. We will then introduce divergence in the form of a mutation affecting the replica's downstream signaling. \rv{Lastly, we will show that open-ended evolution is only possible if the network can acquire mutated gene replicas. This digression well highlight the general features of our model, before we show them in action on our specific example.} \\ \noindent {\bf Non-mutated gene replica\rv{s}:} Given a \rv{\rv{Boolean}} network $N$ with $n$ nodes, we want to study the dynamics of the $n+1$-node network $N'$ obtained by the inclusion of the perfect replica of a node already present in $N$. If the dynamics of $N$ is governed by the set of \rv{\rv{Boolean}} equations \begin{eqnarray} x'_1 & = & f_1 (x_1,\dots,x_n) \nonumber \\ \dots & & \nonumber \\ x'_n & = & f_n (x_1,\dots,x_n) \ , \label{system1} \end{eqnarray} then the updating rules of $N'$ will have the form \begin{eqnarray} x'_1 & = & \rv{g}_1 (x_1,\dots,x_n,x_{n+1}\rv{)} \nonumber \\ \dots & & \nonumber \\ x'_n & = & \rv{g}_n (x_1,\dots,x_n,x_{n+1}\rv{)} \nonumber \\ x'_{n+1} & = & \rv{g}_{n+1} (x_1,\dots,x_n,x_{n+1}\rv{)} \nonumber \label{system2} \end{eqnarray} In the previous equations, $x_i$ is the value (0 or 1) of node $i$ at time $t$, while $x'_i$ is the simplified notation for the value of the same node at time $t+1$. The $\rv{g}_i$ functions are {\it new} functions that we want to determine, given our knowledge of the functions $f_i$. Without loss of generality, we can assume node $n+1$ to be a perfect replica of node 1. Let us then consider the requirements we impose on $\rv{g}_i(x_1,\dots,x_n,x_{n+1})$, after we include node $n+1$: \begin{enumerate} \item The assertion that node ${n+1}$ is a {\it perfect replica} of node $1$ translates into the assumption that \[ \ \rv{g}_{n+1}=\rv{g}_1. \] Here we are just stating that node replicas obey the same updating rules as the original nodes. \item We also want the new node not to affect the system, when not expressed. Therefore, \[ \rv{g}_i (x_1,\dots, x_n, 0) = f_i (x_1,\dots, x_n) \qquad (i=1,\dots,n) \] \item Lastly, we want nodes $2,\dots,n$ to see node 1 and $n+1$ as {\it indistinguishable}: \begin{eqnarray} \rv{g}_i (0,x_2,\dots, x_n, x_{n+1}) = f_i (x_{n+1},x_2,\dots, x_n) \nonumber \\ (i=1,\dots,n) \nonumber \end{eqnarray} \end{enumerate} The previous conditions are enough to determine the values taken by the functions $\rv{g}_i$ for any dynamical state of $N'$ with the only exception being those cases where $x_1$ and $x_{n+1}$ are both 1. This is the only genuinely new scenario whose output cannot be predicted in terms of an equivalent configuration of $N$. The problem is easily solved in the case of \rv{\rv{Boolean}} networks with thresholds (subsection 2.1), as we know the prescription that gives the updating rule of a node in terms of the state of the network, eq. (\ref{updating_rule}). Building the functions $\rv{g}_i$ in a similar fashion, we deduce that it is both natural, and enough, to impose $a_{j+n,i}=a_{ji}$ and $b_{i+n}=b_i$ for the previous three conditions to be satisfied. \rv{This prescription, first adopted in \cite{Wagner1994}, is re-derive here, and contextualized into a more general framework of GRN described by arbitrary \rv{\rv{Boolean}} functions.} \\ \begin{figure*}[!t] \centering \includegraphics[width=.6\textwidth]{selection.pdf} \caption{Different environmental pressure in both space and time selects cells undergoing mutation events that favor the emergence of newly preferred cell types.} \label{selection} \end{figure*} \begin{figure*}[t] \centering \includegraphics[width=.5\textwidth]{multiple_replicas.pdf} \caption{\rv{Effect on the basin size of cell type 2 of multiple non-mutated gene replicas.}} \label{mul_repl} \end{figure*} \noindent {\bf Mutation\rv{s}:} The signaling of gene replica $i$ is represented by the string of numbers $a_{ij}=0,{\pm}1$, with $j=1,\dots,n$ ($n$~\,=~\,current number of genes in the network). A mutation is introduced by changing one of these numbers to \rv{any one} possible value \rv{(including its initial one, which corresponds to having a perfect replica)}. For example, changing $a_{ij}$ from 0 to $\pm1$ corresponds to the creation of a new edge. Changing it from $\pm1$ to 0 means deleting an edge which is instead departing from the original node. Changing it from $\pm1$ to $\mp1$ changes an activating link into an inhibiting one or vice-versa. Different models of mutation could be assumed, which in turn determine the minimum numbers of replicas the cell needs to acquire in order to change from one type to another. For simplicity, our examples show only the sequence of mutations that minimizes the number of steps needed to complete the phenotypic shift. Therefore, we consider all possible mutations of the kind previously described, and then select the one that maximizes the basin of attraction of the preferred cell type.\\ \rv{ \noindent {\bf Mutations and open-ended evolution:} In this section we want to show that allowing divergence (mutations of gene replicas) guarantees open-ended evolution of the GRN. While there are many definitions for open-ended evolution, here we restrict the concept to a simple one, and define a GRN as ``open-ended'' if the GRN can acquire any possible phenotype, given a high enough number of mutated gene replicas (this becomes open-ended in the sense the mechanism is robust to any evolutionary innovation). On the contrary, we will show that, if mutations were not possible, exact gene replicas would be unable to induce arbitrary phenotypic shifts. Let us begin by showing that mutations enable open-ended evolution. At the beginning of this section we have referred to the possibility that the GRN acquires replicas of genes optimally targeting the nodes entering the definition of a specific phenotype, and, therefore, reinforces it. Here we want to show a simple conceptual procedure able to create such optimal genes. We will focus on the existence of a path toward the expression of a novel phenotype, not on the optimal/fastest evolutionary path that leads to that. Such optimal path will depend on the specific network under study. And we will show with other example that it can be much shorter that the general construction illustrated here. Let us label $i_1, \dots, i_p$ the set of $p$ genes that specify the {\it arbitrary} phenotype $\Phi$. For example gene $i_1$ might need to be constantly active, $i_2$ constantly inactive, etc. Starting from an equally {\it arbitrary} gene $j$ with $m$ outgoing edges in the GRN, we can consider $m$ subsequent events of duplication and mutation of $j$ into genes $j_1, j_2, \dots j_{m}=j'$. Each mutation is affecting one of the downstream signals produced by $j$, until $j'$ activates only itself and $i_1$. Each additional, non-mutated replica of $j'$ will therefore activate itself, $i_1$, and each other replica of itself: \begin{center} \includegraphics[width=.9\columnwidth]{mut_j.pdf} \end{center} At this point it is possible to construct a clique of self-activating $j'$ nodes that keeps itself constantly on, regardless of the signaling the clique receives from the remaining genes in the GRN. For this to happen it is enough to have $n_1$ copies of $j'$ with $n_1 > b_1 + w_1$, where $b_1$ is the activation threshold of $i_1$, and $w_1$ is the number of inhibiting incoming edges of $i_1$: \begin{center} \includegraphics[width=.9\columnwidth]{clique.pdf} \end{center} A similar construction guarantees the possibility of keeping gene $i_2$ inactive by simply having $j'$ inhibiting it, as opposed to activating it. The previous construction is intentionally artificial, as we wanted to show that a complete shift of the GRN toward and arbitrary phenotype $\Phi$ can be achieved starting from the duplication and mutation of an arbitrary gene $j$ in the network. Even in the simplest examples, this kind of construction would require many events of duplication + mutations. Our aim here is to demonstrate the existence of an evolutionary path toward $\Phi$. In our example, we consider all possible duplication + divergence events, and only show the ones that induce the fastest shift, as we now know that many other, and slower paths, toward the same phenotypic shift exist. As a final remark, we point out less restrictive choices of the possible mutation events, as well as the inclusion of deletion events, just increase the number of paths towards phenotype $\Phi$.\\ \noindent We now will show open-ended evolution is not possible in the absence of mutations. To this aim, it will be enough to provide a counter-example. Starting from our reference network (figure \ref{GRN}), we consider the effect that the creation of any possible number of exact replicas would have, and we will show that it is not possible to choose an arbitrary phenotype $\Phi$, and convert it into the dominant phenotype. Let us assume $\Phi =$ cell type 2. With reference to figure \ref{mul_repl}, we can see that the relative basin size of cell type 2 does not depend on the number of additional, non-mutated replicas of gene 4. A single replica of gene 1 or 2 is instead enough to completely extinguish the phenotype. Additional replicas of the remaining nodes slowly decreases the size of the basin. Perfect replicas of any gene in the network are either negatively affect the likelihood of cell type 2, or not affect it at all. Therefore, it is not possible to induce a phenotypic shift from cell type 1 to cell type 2 by duplicating genes already present in this GRN. Analogously, it is not possible to create an arbitrary new phenotype through non-mutated gene replicas. As an example we can consider the cell type defined by the activation pattern $100$. Let us call it {\it cell type} 3. We know that this phenotype has an empty basin of attraction, and direct calculation shows that the exact duplication of {\it any} gene in the network leaves its basin unchanged. On the other hand, we are about to show that even adopting just our very restrictive prescription of divergence it is possible to mutate of example network until it undergoes a complete shift from cell type 1 to cell type 2, and even 3. (In agreement with the general argument just exposed.)\\ } \noindent \rv{Therefore, let us} return to the example shown in fig. 3 \rv{to} deduce that at least two mutation events are needed to convert sister cells of type 1 to type 2 (activation pattern: 110). \rv{These optimal mutations are derived as previously described in this section: All possible events of duplication + mutation are considered, and the one inducing the largest enhancement of the relative size of the basin of cell type 2 is selected. The network is then mutated accordingly, and this process iterated until cell type 2 is the dominant phenotype.} \begin{figure*} \centering \begin{tabular}{|c|c|} \hline & \\ \begin{tabular}{rl} \circled{A} & Environmental selection of {\bf cell type 2},\\ & characterized by the expression of gene 1\\ & and 2, and suppression of gene 3. Acqui-\\ & sition of a mutated replica of gene 7 that\\ & inactivates gene 3 instead of activating\\ & it. Mixed cell types (50\% type 1, 50\% \\ & type 2). \end{tabular} & \begin{tabular}{rl} \circled{B} & Environmental selection of cell type 2.\\ & Acquisition of a mutated replica of gene \\ & 6 that does not activate gene 7. Type 2 \\ & cells represent now the dominant cell \\ & type (88\%).\\ & \\ & \end{tabular} \\ \includegraphics[height=.30\textheight]{g1.pdf} & \includegraphics[height=.30\textheight]{g2.pdf} \\[2mm] \hline & \\ \begin{tabular}{rl} \circled{C} & Environmental selection of {\bf cell type 3}:\\ & Expression of gene 1, suppression of gene 2 \\ & and 3. Acquisition of a mutated replica of \\ & gene 8 that suppresses gene 3. Type 3 be-\\ & comes the dominant cell type (92\%).\\ & \\ & \end{tabular} & \begin{tabular}{rl} \circled{D} & After having been selected for type 2,\\ & cells are once again selected for {\bf type 1}. \\ & Acquisition of a mutated replica of gene\\ & 8 that activates gene 3. Type 1 is again\\ & the dominant cell type (96\%).\\ & \\ & \end{tabular} \\ \includegraphics[height=.34\textheight]{g3.pdf} & \includegraphics[height=.34\textheight]{g4.pdf} \\ & \\ \hline \end{tabular} \caption{Duplication and mutation events leading from cell type 1 to 2, and 3, and from cell type 2 back to 1, as described in section 3.} \label{ABCD} \end{figure*} By acquiring a mutated replica of \rv{gene} 7, event \circled{A}, that inactivates gene 3 instead of activating it, the dynamics of the GRN is equally likely to converge toward cell type 1 or 2 (figure \ref{ABCD}). The subsequent acquisition of a mutated replica of gene 6, event \circled{B}, that, differently from the original one, does not activate gene 7, determines the complete shift toward cell type 2 (88\% of the configuration space). The highly connected subset of nodes that appear in the B-panel of figure \ref{ABCD} is responsible for the convergence of the dynamics toward the characteristic activation pattern which defines cell type 2. The assumption that sister cells of cell type 2 start being selected for cell type 3 (activation pattern: 100), determines the selection of mutation events like \circled{C}, where the GRN acquires a mutated replica of \rv{gene} 8 that suppresses gene 3. One mutation event is now enough to determine the complete shift toward cell type 3. The last example we will consider here is the convergence of cell type 2 back to cell type 1. After the acquisition of mutations \circled{A} and \circled{B}, the preferred cell type is again type 1. This can be achieved (complete shift from less than 12\% to more than 96\%) with a single mutation event, marked \circled{D} in figure 3, consisting of a mutated replica of \rv{gene} 8 that activates gene 3. This last evolution of the GRN expresses the same phenotype as the initial network in figure \ref{GRN}, but carries memory of the evolutionary path that led it through its type 2 period in the form of the \rv{modular structure} visible in the D-panel of figure \ref{ABCD}. We hope this example is enough to convince the reader that, as long as a viable mathematical description of the GRN is available, as well as the knowledge of the preferred phenotypes for a shifting environment, our approach \rv{can identify the events of gene duplication and mutation with the highest causal effect on the phenotipic shift.} The possibility to retain cell type mathematical identities for evolving GRNs is the key feature enabling this. We have also shown that our method allows the identification of GRNs that differ for both genotype and network topology (like the GRN in figure 1 and the evolved network in the D-panel of figure 3) as encoding {\it the same cell type} (type 1 in our example). Therefore, apparently redundant topological features in the structure of a GRN might carry the imprint of their evolutionary history, and of the environment that induced them. \section{Development} \begin{table*}[t] \centering \includegraphics[width=.5\textwidth]{table.pdf} \label{pinning1} \caption{Phenotypic change induced by the controlled expression of \rv{genes 4-7}. This table shows the result of pinning just one \rv{gene} at a time (TS ON/OFF). Four new phenotypes, cell types $3\div6$, not initially encoded in the GRN of figure \ref{GRN}, are now possible equilibrium state of the gene expression dynamics, with basins of attraction sizes (likelihoods) listed as percentages.} \end{table*} Next we want to show how external control of the expression of accurately chosen genes -- e.g. as under the effect of drug therapies, or because of extracellular signalling during development-- can determine cell differentiation, and the expression of not only cell type 2, but also novel cell types that were not initially encoded in our sample network. For an explanation of cell differentiation not invoking extracellular signaling, and relying instead on cellular noise, the reader is redirected to \cite{Villani2011}. We will focus here on the possibility of epigenetic signaling to induce the appearance of new cell types. This approach has been extensively investigated as an alternative explanation \cite{Huang2012,Pisco2013} to the current paradigm of treatment-selected, drug-resistant clones in tumor progression \cite{Attolini2009,Diaz2012}. We generalize it here to the broader problem of cell differentiation over a developmental timescale. We also show the role played by our generalized definition of phenotype in reducing the size of the {\it control kernel}, as defined in \cite{Kim2013}. The original approach developed in \cite{Kim2013}, consists of taking every possible combination of any subset of genes and their possible expression patterns. For each combination, and each initial state, time series of the transient expression pattern of the remaining genes are generated, while the selected genes are \rv{set} to the specific selected values. When by \rv{setting the values of} the smallest \rv{possible} number of nodes this operation moves all possible initial conditions toward the basin of attraction of a specific attractor, Kim {\it et al.} define \rv{this} smallest subset of nodes as the {\it control kernel} of the attractor state the system is converging to. For consistency with our less constrained definition of what a phenotype is (subsection 2.2), we need to redefine the notion of control kernel accordingly. The generalized definition of control kernel we adopt here does not assume the forced terminal state to be a specific attractor, just a phenotype (i.e. any possible attractor compatible with the definition of that phenotype). As a result, our control kernels are significantly smaller than those found in \cite{Kim2013}. To show this explicitly, we have considered again our sample network from subsection 2.3, and the effect of \rv{setting the value of just one among genes 4-8} at a time. The result is shown in table 1. Forcing the expression of \rv{gene} 7 (referred to as ``7 is ON'' in the table) is enough to enlarge the basin of attraction of cell type 2 from less than 20\% to exactly 100\%. Likewise, forcing the expression of \rv{gene} 8 makes the network shift toward a phenotype characterized by the expression pattern 111, that was not even encoded in the network. Eventually, we might assume 80\% (or some other high percentage value) to be enough for the network to be considered locked in a certain cell phenotype. In this case \rv{gene} 4 would be an alternative control kernel for the cell type defined by the 111 pattern. This new definition of control kernel, focused on forcing the GRN into a less restrictive activation pattern than a single, specific attractor, seems to be much easier to achieve experimentally, better suited for large/realistic networks, and therefore relevant to drug treatments. Cell type differentiation in response to extracellular signaling, is still a key aspect of our generalized approach to dynamical GRN models. Even if mainly aimed at incorporating evolutionary features (section 3), our framework entirely preserves (and sometimes facilitates) the applicability of network control theory to GRN models in explaining both developmental or drug-induced change. \section{Conclusions} This manuscript represents a synthesis of mathematical models of gene regulatory networks within theoretical evolutionary biology, which also accounts for development. Despite of their predictive power, and their possibility to mathematically reinterpret cell fates as the steady states of the abstract, non-linear dynamics of the regulatory processes they describe, dynamical models (\rv{\rv{Boolean}} models in this manuscript) of GRNs have been limited by this identification between mathematical attractors and biological phenotypes. Evolutionary processes, like single gene and whole genome duplication, both chromosome gain and loss, horizontal gene transfer, etc., alter the size of the genome, and deprive the exact attractor-phenotype identification as a one-to-one map of its original meaning. The leading role played in evolution and speciation by the interplay between these processes and natural selection, makes current dynamical models of GRNs unfit to describe genetic network evolution. Motivated by this limitation, we have explored the consequences of relaxing this one-to-one identification between cell types and attractors. We have shown with a simple numerical model that, redefining a cell type as a {\it collection of attractors} of a GRN, all sharing a common, characterizing property, enables retaining the main features of Kauffman's original theory, and permitting easy generalization using network control theory. It is now no longer needed for the genome to retain a fixed size for the cell type type definition to preserve its meaning. This opens the possibility to quantitative studies of the effect that gene duplication, mutation, and natural selection have on cell types, by virtue of allowing the study of evolutionary processes within the well-established framework of dynamical systems theory. \rv{Lastly, by showing that our generalized framework does not inhibit traditional} network controllability, \rv{we have preserved} its biological interpretation as a viable mathematical description of extracellular/epigenetic control of gene expression in an evolutionary model. \end{multicols}

e4b53b2216cbeaff87e578fc61d9290ee61f171a

\section{Introduction} Cluster algebras are commutative algebras that come equipped with a rich structure: their generators, the \emph{cluster variables}, are grouped into finite sets called \emph{clusters}, each of which is endowed with a matrix (or a quiver) to form a \emph{seed}, in such a way that all seeds can be deduced from any one of them by a process called \emph{mutation}. Ever since their introduction and fundamental study by S.~Fomin and A.~Zelevinsky \cite{FZ02, FZ03, FZ07} (and also with A.~Berenstein in \cite{BFZ05}), cluster algebras have received a lot of attention due to their appearance in many fields of mathematics. Survey papers and books on cluster algebras include \cite{Marsh13, LW14, W14, GR17}; let us also mention \cite{GSV10}, which focuses on links with Poisson geometry, \cite{Reiten10, Amiot11, Keller12, Plamondon18} on links with representation theory of associative algebras, \cite{GLS13} on links with Lie theory, and \cite{Schiffler18} on links with triangulations of surfaces. In this paper, we are interested in structural questions for cluster algebras. If one forgets the cluster algebra structure given by the seeds, many cluster algebras become isomorphic as rings. In \cite{ASS12, ADS14}, a suitable notion of morphism of cluster algebras was introduced: roughly speaking, such a morphism should send seeds into seeds. An isomorphism of cluster algebra is then an isomorphism of rings satisfying the strong condition that it should preserve seeds. In \cite{ASS14} was introduced the notion of unistructurality: a cluster algebra is unistructural if the set of its cluster variables determines its seeds (see Section~\ref{sect::prelim}, where we discuss several variants of the definition). \begin{conjecture*}[1.2 of \cite{ASS14}] Any cluster algebra is unistructural. \end{conjecture*} The conjecture was proved in \cite{ASS14} for skew-symmetric cluster algebras of Dynkin type or of rank~$2$, and in~\cite{BM16} for cluster algebras of type~$\widetilde{\mathbb{A}}_n$ (all of these cluster algebras are taken with trivial coefficients). In the latter case, the first author of the present paper used the geometric model of~\cite{FST08} (see Section~\ref{sect::surfaces}), in which cluster algebras of type~$\widetilde{\mathbb{A}}_n$ are linked to triangulations of an annulus. In this paper, we extend these methods to any unpunctured surface. \begin{theorem*}[\ref{theo::main}] Any cluster algebra with trivial coefficients associated to a triangulation of an unpunctured surface is unistructural. \end{theorem*} The proof, given in Section~\ref{sect::proof}, relies on the existence of a positive basis (the \emph{bracelet basis} of \cite{MSW13}) and on the skein relations to analyze products of cluster variables. On the way, we obtain a reduction result (Proposition~\ref{prop::connected}) showing that it is sufficient to prove the unistructurality conjecture for cluster algebras defined from connected quivers. Shortly after this paper appeared, the conjecture was proved in full generality by P.~Cao and F.~Li \cite{CL18}. \section{Cluster algebras and unistructurality}\label{sect::prelim} \subsection{Cluster algebras} Let us recall the main definitions and properties of cluster algebras. We will mainly follow the exposition of~\cite{FZ07}, with the additional restrictions that we will only consider skew-symmetric cluster algebras (which allows us to use quivers instead of matrices) with trivial coefficients. Let us mention that examples of all notions recalled in this section can be easily generated using B.~Keller's applet \cite{KellerApp} or Sage \cite{Sage}. Our first ingredient is that of quiver mutation. A quiver is an oriented graph (possibly with multiple edges); if~$Q$ is a quiver, we denote by~$Q_0$ its set of vertices,~$Q_1$ its set of arrows, and by~$s,t:Q_1\to Q_0$ the maps sending each arrow to its source or its target, respectively. \begin{definition} Let~$Q$ be a finite quiver without oriented cycles of length~$1$ or~$2$, and let~$i$ be a vertex of~$Q$. The \emph{mutation of~$Q$ at~$i$} is the quiver~$\mu_i(Q)$ defined by modifying~$Q$ as follows: \begin{enumerate} \item for each path of length~$2$ of the form~$h\xrightarrow{\alpha} i \xrightarrow{\beta} j$ (recall that~$i$ is fixed), add an arrow~$h\xrightarrow{[\beta\alpha]} j$; \item reverse the orientation of each arrow having~$i$ as source or target; \item remove the arrows from a maximal set of disjoint oriented cycles of length~$2$ created in the two previous steps. \end{enumerate} \end{definition} It is immediate from the definition that mutation of quivers at a given vertex is an involution. \begin{definition} A \emph{seed} is a pair~$(\mathbf{u}, Q)$, where \begin{itemize} \item $Q$ is a finite quiver without oriented cycles of length~$1$ or~$2$, and with vertex set~$Q_0 = \{1, \ldots, n\}$; \item $\mathbf{u} = (u_1, \ldots, u_n)$ is a free generating set of the field~$\mathbb{Q}(x_1, \ldots, x_n)$. Note that the variables in~$\mathbf{u}$ are indexed by the vertices of~$Q_0$. \end{itemize} The \emph{mutation of~$(\mathbf{u}, Q)$ at~$i$} is the pair~$\mu_i(\mathbf{u}, Q) = (\mathbf{u}', Q')$, where \begin{itemize} \item $Q' = \mu_i(Q)$ is the mutation of~$Q$ at~$i$; \item $u'_{\ell} = u_{\ell}$ if~$\ell\neq i$, and~$u'_i$ satisfies the \emph{exchange relation} \[ u'_i = \frac{\prod_{\stackrel{\alpha\in Q_1}{t(\alpha) = i}} u_{s(\alpha)} + \prod_{\stackrel{\alpha\in Q_1}{s(\alpha) = i}} u_{t(\alpha)}}{u_i}. \] \end{itemize} Two seeds are \emph{mutation-equivalent} if one can be obtained from the other by a sequence of mutations. Two seeds are \emph{isomorphic} if they are the same up to reordering of their cluster variables and relabelling of the correponding vertices of their quivers. \end{definition} One readily checks that the mutation of a seed is still a seed, and that mutation of seeds at a given vertex is an involution. \begin{definition} Let~$(\mathbf{u}, Q)$ be a seed. Consider all seeds~$(\mathbf{u}', Q')$ obtained by iterated mutations of~$(\mathbf{u}, Q)$. \begin{itemize} \item The~$n$-tuples~$\mathbf{u}'$ appearing in those seeds are called \emph{clusters}. \item The elements of clusters are called \emph{cluster variables}. \item The \emph{cluster algebra}~$\mathscr{A}(\mathbf{u}, Q)$ is the subring of the \emph{ambient field}~$\mathcal{F}$ generated by all cluster variables, where $\mathcal{F} = \mathbb{Q}(x_1, \ldots, x_n)$. \item The integer~$n$ is called the \emph{rank of the cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$}. \item A \emph{cluster monomial} is a product of cluster variables belonging to the same cluster. \item The \emph{exchange graph} of~$\mathscr{A}(\mathbf{u}, Q)$ is the graph whose vertices are isomorphism classes of seeds of~$\mathscr{A}(\mathbf{u}, Q)$ and where two vertices are joined by an edge if one is obtained by the other by applying one mutation. \end{itemize} \end{definition} \begin{example} If the quiver~$Q$ has one vertex and no arrows, then the cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is~$\mathbb{Z}[u, \frac{2}{u}]$. Its exchange graph has two vertices joined by an edge. \end{example} Let us recollect some of the most important results on cluster algebras. \goodbreak \begin{theorem}\label{theo::recollectionsCluster} $\phantom{x}$ \begin{itemize} \item (Laurent Phenomenon) \cite{FZ02} A cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is contained in the Laurent polynomial ring~$\mathbb{Z}[u_1^{\pm 1}, \ldots, u_n^{\pm 1}]$. Equivalently, every cluster variable of~$\mathscr{A}(\mathbf{u}, Q)$ is a Laurent polynomial with integer coefficients in the cluster variables of any given cluster. \item (Positivity) \cite{LS15, GHKK18} Every cluster variable of~$\mathscr{A}(\mathbf{u}, Q)$ is a Laurent polynomial with \emph{non-negative} integer coefficients in the cluster variables of any given cluster. \item (Finite type) \cite{FZ03} A cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ has only finitely many cluster variables if and only if~$Q$ is mutation-equivalent to a disjoint union of quivers which are orientations of Dynkin diagrams of type~$ADE$. \item (Linear independence of cluster monomials) \cite{CKLP13, GHKK18, CaoLi} Cluster monomials are linearly independent over~$\mathbb{Z}$. \end{itemize} \end{theorem} We shall also need the following result on the structure of non-initial variables. \begin{lemma}[Lemma 3.7 of \cite{CKLP13}] Let~$\mathscr{A}(\mathbf{u}, Q)$ be a cluster algebra. If a cluster variable in~$\mathscr{A}(\mathbf{u}, Q)$ is a Laurent monomial in~$\mathbf{u}$ with coefficient~$\pm 1$, then it belongs to~$\mathbf{u}$. \end{lemma} \subsection{Unistructurality} Different cluster algebras can be isomorphic as rings without having the same exchange graphs. A stronger notion of morphism and isomorphism of cluster algebras was thus introduced and studied in \cite{ASS12, ADS14}. In this paper, we shall not use the full scope of this formalism, but merely the following notion from \cite{ASS14}. \begin{definition}[Section 5.1 of \cite{ASS14}]\label{defi::unistructurality} Let~$(\mathbf{u}, Q)$ be a seed, and let~$\mathscr{A}(\mathbf{u}, Q) \subseteq \mathbb{Q}(x_1, \ldots, x_n)$ be the corresponding cluster algebra. Denote by~$\mathscr{X}$ the set of its cluster variables. The cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is \emph{unistructural} if, for any~$n$-tuple~$\mathbf{v}$ which forms a free generating set of~$\mathbb{Q}(x_1, \ldots, x_n)$, the following condition is satisfied: \begin{itemize} \item [(U)] if~$(\mathbf{v}, R)$ is a seed such that the set of cluster variables of~$\mathscr{A}(\mathbf{v}, R)$ is equal to~$\mathscr{X}$, then there is an isomorphism between the exchange graphs of~$\mathscr{A}(\mathbf{u}, Q)$ and~$\mathscr{A}(\mathbf{v}, R)$ and the two cluster algebras have the same clusters (up to permutation of the cluster variables in each cluster). \end{itemize} \end{definition} In our setting, we can get rid of the condition on the exchange graphs. \begin{proposition}\label{prop::clustersAreSufficient} A cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ with set~$\mathscr{X}$ of cluster variables is unistructural if and only if the following condition is satisfied: \begin{itemize} \item [(U')] if~$(\mathbf{v}, R)$ is a seed such that the set of cluster variables of~$\mathscr{A}(\mathbf{v}, R)$ is equal to~$\mathscr{X}$, then the two cluster algebras have the same clusters (up to permutation of the cluster variables in each cluster). \end{itemize} \end{proposition} \begin{proof} Clearly, condition (U) implies (U'). To prove the converse, one uses the following results: \begin{enumerate} \item \cite[Theorem 4]{GSV08} for a (skew-symmetric) cluster algebra (with trivial coefficients), the same cluster (up to permutation of its variables) cannot appear in two different seeds (up to isomorphism); \item \cite[Theorem 5]{GSV08} in the same setting, two clusters (up to permutation) belong to adjacent seeds of the exchange graph if and only if they have exactly~$n-1$ cluster variables in common. \end{enumerate} Therefore, if the set of clusters of a cluster algebra is known, then its exchange graph can be recovered using the adjacency relations in (2). Thus two cluster algebras having the same set of clusters must have isomorphic exchange graphs. \end{proof} \begin{remark} Proposition~\ref{prop::clustersAreSufficient} is true in a greater generality, since the two results it uses, \cite[Theorems 4 and 5]{GSV08}, are proved under the following conditions. Let~$B$ be a skew-symmetrizable matrix, and~$\mathbf{y}$ be a set of coefficients in a semi-field. Then \cite[Theorems 4 and 5]{GSV08} are true for the cluster algebra with coefficients~$\mathscr{A}(\mathbf{u}, \mathbf{y}, B)$ if either \begin{itemize} \item the cluster algebra is of geometric type, or \item $B$ is of full rank. \end{itemize} Since all cluster algebras in this paper are of geometric type (they have trivial coefficients), we will not say more about this. \end{remark} \begin{remark}\label{rema::quivers or opposite} In the definition of unistructurality, we can also say something about seeds: a cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is unistructural if and only if the following condition is satisfied: \begin{itemize} \item[(U'')] if~$(\mathbf{v}, R)$ is a seed such that the set of cluster variables of~$\mathscr{A}(\mathbf{v}, R)$ is equal to~$\mathscr{X}$, and if~$R$ is a disjoint union of quivers~$R^1, \ldots, R^r$, then there is a seed~$(\mathbf{v} , Q^1 \sqcup \ldots \sqcup Q^r )$ which is mutation-equivalent to~$(\mathbf{u}, Q)$ (up to isomorphism), where each~$Q^j$ is either~$R^j$ or the opposite quiver~$(R^j)^{op}$. \end{itemize} To see this, notice that at the end of the proof of Proposition~\ref{prop::clustersAreSufficient}, the adjacent clusters in the exchange graph also allow one to recover the exchange relations, and thus to deduce the quiver of the associated seed up to a change in the orientation of all the arrows in each of its connected components. Hence, condition~(U') implies~(U''). The converse holds trivially, so the equivalence is proved. \end{remark} \begin{example} A cluster algebra of rank~$1$ is unistructural, since it only has two clusters and two cluster variables. \end{example} The above example is one of the few cluster algebras which are known to be unistructural. \begin{theorem} The following results are known for skew-symmetric cluster algebras with trivial coefficients. \begin{enumerate} \item \cite[Theorem 5.2]{ASS14} Cluster algebras of rank~$2$ are unistructural. \item \cite[Theorem 0.1]{ASS14-2} Cluster algebras of Dynkin type are unistructural. \item \cite[Theorem 3.2]{BM16} Cluster algebras of affine type~$\tilde{\mathbb{A}}_n$ are unistructural. \end{enumerate} \end{theorem} Moreover, there is no known counter-example to the following conjecture. \begin{conjecture} \cite[Conjecture 1.2]{ASS14}\label{conj::unistructurality} Any cluster algebra is unistructural. \end{conjecture} \subsection{Reduction to connected quivers} In this section, we show that it is enough to consider connected quivers when dealing with unistructurality. More precisely, we prove the following proposition. \begin{proposition}\label{prop::connected} Let~$Q$ be a quiver which is a disjoint union of quivers~$Q^1, \ldots, Q^r$. Let~$(\mathbf{u}, Q)$ be a seed, and let~$(\mathbf{u}^1, Q^1), \ldots, (\mathbf{u}^r, Q^r)$ be the seeds corresponding to the decomposition of~$Q$. Then the cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is unistructural if and only if the cluster algebra~$\mathscr{A}(\mathbf{u}^i, Q^i)$ is unistructural for all~$i\in \{1, \ldots, r\}$. \end{proposition} Before proving Proposition \ref{prop::connected}, let us state an immediate consequence. \begin{crl} Conjecture~\ref{conj::unistructurality} is true if and only if for any \emph{connected} quiver~$Q$ and any seed~$(\mathbf{u}, Q)$, the cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is unistructural. \end{crl} Let us now prove Proposition~\ref{prop::connected}. Assume that~$Q$ is a quiver which is a disjoint union of quivers~$Q^1, \ldots, Q^r$, and let~$(\mathbf{u}, Q),(\mathbf{u}^1, Q^1), \ldots, (\mathbf{u}^r, Q^r)$ be seeds as above. One implication of the proposition is clear. \begin{lemma} If the cluster algebra~$\mathscr{A}(\mathbf{u}, Q)$ is unistructural, then so is~$\mathscr{A}(\mathbf{u}^i, Q^i)$ for each~$i\in \{1, \ldots, r\}$. \end{lemma} \begin{proof} Suppose that~$\mathscr{A}(\mathbf{u}^i, Q^i)$ is not unistructural for a certain~$i$. Let $(\mathbf{v}^i, R^i)$ be a seed such that~$\mathscr{A}(\mathbf{v}^i, R^i)$ has the same cluster variables as~$\mathscr{A}(\mathbf{u}^i, Q^i)$, but such that~$(\mathbf{v}^i, R^i)$ has condition (U) in Definition \ref{defi::unistructurality} fail. Let~$\mathbf{u}' = \mathbf{u}^1 \sqcup \ldots \sqcup \mathbf{u}^{i-1} \sqcup \mathbf{v}^i \sqcup \mathbf{u}^{i+1} \sqcup \ldots \sqcup \mathbf{u}^r$ and~$Q' = Q^1 \sqcup \ldots \sqcup Q^{i-1} \sqcup R^i \sqcup Q^{i+1} \sqcup \ldots \sqcup Q^r$. Then~$\mathscr{A}(\mathbf{u}, Q)$ and~$\mathscr{A}(\mathbf{u}', Q')$ have the same cluster variables, but condition (U) fails. Thus~$\mathscr{A}(\mathbf{u}, Q)$ is not unistructural. \end{proof} Assume now that each~$\mathscr{A}(\mathbf{u}^i, Q^i)$ is unistructural. Note that each cluster variable of~$\mathscr{A}(\mathbf{u}, Q)$ lies in one of the fields~$\mathbb{Q}(\mathbf{v}^i)$, for some~$i\in \{1, \ldots, r\}$. Let~$(\mathbf{v}, R)$ be a seed such that~$\mathscr{A}(\mathbf{v}, R)$ and~$\mathscr{A}(\mathbf{u}, Q)$ have the same cluster variables. \begin{lemma}\label{lemm::disjointUnion} The quiver~$R$ is a disjoint union of quivers~$R^1, \ldots, R^r$ and~$\mathbf{v}$ is a corresponding disjoint union of tuples~$\mathbf{v}^1, \ldots, \mathbf{v}^r$ in such a way that each cluster algebra~$\mathscr{A}(\mathbf{v}^i, R^i)$ is contained in~$\mathbb{Q}(\mathbf{u}^i)$. \end{lemma} \begin{proof} By assumption, each element of~$\mathbf{v}$ is contained in one of the subfields~$\mathbb{Q}(\mathbf{u}^i)$ of~$\mathbb{Q}(\mathbf{u})$. Since $\mathbf{v}$ is a free generating set of~$\mathbb{Q}(\mathbf{u})$, all subsets of~$\mathbf{v}$ are algebraically free. Therefore, for each~$i$, there is exactly~$|\mathbf{u}^i|$ elements of~$\mathbf{v}$ in~$\mathbb{Q}(\mathbf{u}^i)$. This allows us to partition~$\mathbf{v}$ into sub-tuples~$\mathbf{v}^1, \ldots, \mathbf{v}^r$ in such a way that~$\mathbf{v}^i$ contains precisely the variables of~$\mathbf{v}$ contained in~$\mathbb{Q}(\mathbf{u}^i)$. Since~$\mathbf{v}$ freely generates~$\mathbb{Q}(\mathbf{u})$, this implies that each~$\mathbf{v}^i$ freely generates~$\mathbb{Q}(\mathbf{u}^i)$; in other words,~$\mathbb{Q}(\mathbf{u}^i)= \mathbb{Q}(\mathbf{v}^i)$ for each~$i$. Now, let~$v_j$ a cluster variable of~$\mathbf{v}$, and let~$v'_j$ be the variable obtained by mutation of~$(\mathbf{v}, R)$ at~$j$. By the above argument, if~$v_j$ is in~$\mathbb{Q}(\mathbf{u}^i)$, then so is~$v'_j$. Write the exchange relation as \[ v_jv'_j = \prod_{\stackrel{\alpha \in R_1}{t(\alpha) = j}} v_{s(\alpha)} + \prod_{\stackrel{\alpha \in R_1}{s(\alpha) = j}} v_{t(\alpha)}. \] Since~$v_jv'_j$ lies in~$\mathbb{Q}(\mathbf{u}^i) = \mathbb{Q}(\mathbf{v}^i)$ and each~$v_k$ lies in~$\mathbb{Q}(\mathbf{v})$, we must have that all variables occurring on the right hand side of the exchange relation are also in~$\mathbb{Q}(\mathbf{v}^i)$. Thus the vertex~$j$ of~$R$ is related by arrows only to vertices corresponding to variables in~$\mathbf{v}^i$. Repeating the argument for all vertices~$j$ of~$R$, we partition~$R$ into a disjoint union of quivers~$R^1, \ldots, R^r$ as required. \end{proof} As a consequence of Lemma~\ref{lemm::disjointUnion}, the set of cluster variables of~$\mathscr{A}(\mathbf{v}^i, R^i)$ is the same as that of~$\mathscr{A}(\mathbf{u}^i, Q^i)$. Since the latter is unistructural by hypothesis, we get that the two algebras have the same clusters. As~$R$ is the disjoint union of the~$R^i$ and~$\mathbf{v}$ that of the~$\mathbf{v}^i$, we thus get that~$\mathscr{A}(\mathbf{u}, Q)$ and~$\mathscr{A}(\mathbf{v}, R)$ also have the same clusters. Thus~$\mathscr{A}(\mathbf{u}, Q)$ is unistructural. This finishes the proof of Proposition~\ref{prop::connected}. \section{Cluster algebras arising from triangulations of surfaces}\label{sect::surfaces} \subsection{Cluster algebras and surfaces} We now recall how a cluster algebra can be associated to a triangulation of a surface. We mainly follow \cite{FST08}, but restrict ourselves to the setting of unpunctured surfaces. A \emph{marked surface} is a pair~$(S,M)$, where~$S$ is a connected oriented Riemann surface with boundary~$\partial S$ and~$M$ is a finite subset of~$\partial S$ such that~$M$ has at least one point on each connected component of~$\partial S$. We exclude the case where~$(S,M)$ is such that~$S$ is a disk and~$|M| \in \{1,2, 3\}$; these are the cases where there is either only one triangulation or no triangulation at all (in the sense that we recall below). An \emph{arc} on~$(S,M)$ is an isotopy class of curves on~$S$ with endpoints in~$M$. A \emph{boundary arc} is an arc isotopic to a curve contained in the boundary~$\partial S$ of~$S$; an \emph{internal arc} is an arc which is not a boundary arc. Finally, a \emph{closed curve} on~$(S,M)$ is a free isotopy class of curves on~$S$ whose starting point and ending point are the same point in the interior of~$S$ (a \emph{free isotopy} is an isotopy of curves that does not necessarily fix their endpoints). The figure below represents two triangulated surfaces: on the left, a sphere with two open disks removed (or annulus), and on the right, a sphere with three open disks removed. \begin{figure}[h!] \begin{displaymath} \begin{tikzpicture} \draw[thick] (0,0) circle (1); \filldraw[grey] (0,0) circle (0.3); \draw[thick] (0,0) circle (0.3); \filldraw[black] (0.3, 0) circle (0.05); \filldraw[black] (-0.3, 0) circle (0.05); \filldraw[black] (0, 1) circle (0.05); \filldraw[black] (0, -1) circle (0.05); \draw (0.3, 0) -- (0,1); \draw (-0.3, 0) -- (0,1); \draw (0.3, 0) -- (0,-1); \draw (-0.3, 0) -- (0,-1); \draw[thick] (5,0) ellipse (2cm and 1cm); \filldraw[grey] (4,0) circle (0.3); \draw[thick] (4,0) circle (0.3); \filldraw[grey] (6,0) circle (0.3); \draw[thick] (6,0) circle (0.3); \filldraw[black] (4.3, 0) circle (0.05); \filldraw[black] (3.7, 0) circle (0.05); \filldraw[black] (6.3, 0) circle (0.05); \filldraw[black] (5.7, 0) circle (0.05); \filldraw[black] (5,1) circle (0.05); \filldraw[black] (5,-1) circle (0.05); \draw (5,1) -- (5,-1); \draw (5,1) -- (5.7, 0); \draw (5,1) -- (4.3, 0); \draw (5,-1) -- (5.7, 0); \draw (5,-1) -- (4.3, 0); \draw (5,-1) to [bend left=40] (3.7,0); \draw (5,-1) to [bend right=40] (6.3,0); \draw (5,1) to [bend right=40] (3.7,0); \draw (5,1) to [bend left=40] (6.3,0); \end{tikzpicture} \end{displaymath} \end{figure} We say that two arcs or closed curves~$\gamma$ and~$\delta$ on~$(S,M)$ \emph{intersect} if, for all choices of isotopy representatives~$\bar \gamma$ of~$\gamma$ and~$\bar \delta$ of~$\delta$, the two representatives~$\bar\gamma$ and~$\bar\delta$ intersect in the interior of~$S$. We say that an arc or closed curve~\emph{self-intersects} if all isotopy representatives of it intersects itself in the interior of~$S$. A \emph{triangulation} of~$(S,M)$ is a maximal collection of non-self-intersecting and pairwise non-intersecting curves. For a triangulation~$\tau$ and an arc~$i$ in~$\tau$, the \emph{flip of~$\tau$ at~$i$} is the unique triangulation~$\mathfrak{f}_i(\tau)$ containing~$\tau\setminus i$ but not~$i$. Let~$\tau$ be a triangulation of~$(S,M)$. Define a quiver~$Q(\tau)$ in the following way: \begin{itemize} \item vertices of~$Q(\tau)$ are arcs in~$\tau$; \item there is an arrow $i\to j$ for each triangle~$\Delta$ of~$\tau$ in which both~$i$ and~$j$ occur, with~$j$ immediately following~$i$ in the clockwise order of the boundary of~$\Delta$. \end{itemize} \begin{definition} Let~$\tau$ be a triangulation of~$(S,M)$ with~$n$ arcs, and let~$\mathbf{u}$ be a free generating set of~$\mathbb{Q}(x_1, \ldots, x_n)$. The~\emph{cluster algebra associated with~$\tau$} (with set~$\mathbf{u}$ of initial variables) is the cluster algebra~$\mathscr{A}(\mathbf{u}, \tau) := \mathscr{A}(\mathbf{u}, Q(\tau))$. \end{definition} Let~$\tau$ be a triangulation of a marked surface~$(S,M)$. The following are consequences of \cite[Theorem 7.11]{FST08}. \begin{theorem}[Theorem 7.11 of \cite{FST08}] $\phantom{x}$ \begin{itemize} \item The cluster variables of~$\mathscr{A}(\mathbf{u}, \tau)$ are in bijection with non-self-intersecting internal arcs of~$(S,M)$. If~$\gamma$ is such an arc, denote by~$u_{\gamma}$ the corresponding cluster variable. Under this bijection, the arcs of~$\tau$ are sent to the cluster variables in $\mathbf{u}$. \item The above induces a bijection between triangulations of~$(S,M)$ and seeds of~$\mathscr{A}(\mathbf{u}, \tau)$: if~$\tau'$ is a triangulation of~$(S,M)$, then the associated seed is~$\big( (u_{\gamma} \ | \ \gamma\in \tau'), Q(\tau') \big)$. \item In particular, if~$\tau'$ and~$\tau''$ are related by a flip at~$i$, then the seed associated to~$\tau''$ is the mutation at~$i$ of the seed associated to~$\tau'$. \end{itemize} \end{theorem} \subsection{Bracelets} The following definitions are adapted from \cite{MSW13}. \begin{definition} Let $\gamma$ be an essential loop in $(S,M)$. The \df{bracelet} $\text{Brac}_m \gamma$ is the closed loop obtained by concatenating $\gamma$ with itself exactly $m$ times. We denote the polynomial in $\mathcal{F}$ associated to it by $b_m(\gamma)$. \end{definition} Note that $\text{Brac}_m \gamma$ has $m-1$ self-intersections. \begin{lemma}[Theorem 1.1 of \cite{MW13}]\label{LaurentBracelet} Every bracelet has a positive Laurent expansion with respect to any cluster of $\mathcal{A}(\mathbf{u},\tau).$ \end{lemma} \begin{comment} \begin{proposition}\label{bracelet} The polynomial in $\mathcal{F}$ associated to a bracelet is clusterisable. \end{proposition} \begin{proof} Let $B=\text{Brac}_m\gamma$ be a bracelet and $b=b_m(\gamma)$. A COMPLETET \end{proof} \end{comment} \begin{definition} A collection $c$ of arcs is \df{$\mathcal{C}$-compatible} if it contains no intersections (between two elements or self-intersections). Therefore, it excludes bracelets. A collection $c'$ of arcs and bracelets is \df{$\mathcal{C}'$-compatible} if: \begin{itemize} \item no two elements of $c'$ intersect each other, except for the self-intersections of a bracelet; \item there exists at least one essential loop $\gamma$ in $(S,M)$ such that the bracelet $\text{Brac}_m(\gamma)$ lies in $c'$, with $m \geq 1$. Moreover, there is exactly one copy of $\text{Brac}_m(\gamma)$ in $c'$ and if $m' \neq m$, then $\text{Brac}_{m'}(\gamma) \not\in c'$. \end{itemize} \end{definition} In other word, there are no bracelets in $\mathcal{C}$-compatible collections of arcs, but there are always bracelets in $\mathcal{C}'$-compatible collections of arcs and at most one per essential loop. Note that it does not completely agree with the definition if $\mathcal{C}$-compatibility from \cite{MSW13}: a collection of arcs is $\mathcal{C}$-compatible in \cite{MSW13} if it is $\mathcal{C}$-compatible or $\mathcal{C}'$-compatible. \begin{comment} \begin{definition} A collection $C$ of arcs \df{$\mathcal{C}$-compatible} no two elements of $C$ intersect each other. A collection $C'$ of arcs and bracelets is \df{$\mathcal{C}'$-compatible} if: \begin{itemize} \item no two elements of $C'$ intersect each other, except for the self-intersection of a bracelet \item given a essential loop $\gamma$ in $(S,M)$, there is at most one $m \geq 1$ such that the $m$-th bracelet lies in $C$, and, moreover, there is at most one copy of this bracelet $\text{Brac}_m(\gamma)$ in $C'$. \end{itemize} \end{definition} \end{comment} We define $\mathcal{C} (S,M)$ to be the set of all $\mathcal{C}$-compatible collections of arcs in $(S,M)$ and we define $\mathcal{C}'(S,M)$ to be the set of all $\mathcal{C}'$-compatible collections of arcs and bracelet in $(S,M)$. collections. We denote \[ \mathcal{B} = \left\{ \prod_{\gamma \in C} x_\gamma \mid c \in \mathcal{C} (S,M) \right\} \text{ and } \mathcal{B}' = \left\{ \prod_{\gamma \in C'} x_\gamma \mid c' \in \mathcal{C}' (S,M) \right\}.\] \begin{definition} Let $A$ be a $\mathbb{Z}$-algebra free as a $\mathbb{Z}$-module and $B$ be a basis of $A$. The basis $B$ is \df{positive}, if for every $b_1$, $b_2 \in B$, \[b_1 b_2 = \sum_{b \in B} m_b b \] with $m_b \geq 0$ for every $b \in B$. \end{definition} \begin{theorem}[\cite{MSW13,Thu14,CLS}]\label{positivebasis} The set $\mathcal{B} \cup \mathcal{B}'$ is a positive basis for the cluster algebra~$\mathscr{A}(\mathbf{u}, \tau)$. \end{theorem} It is proved in \cite[Theorem 1.1]{MSW13,CLS} that~$\mathcal{B} \cup \mathcal{B}'$ is a basis and in \cite[Theorem 1]{Thu14} that it is positive. \subsection{Products of cluster variables and skein relations} We recall here some definitions and results from \cite{MW13}. An \emph{multicurve} a multiset of arcs and closed curves on the surface. \begin{definition} [Definition 6.1 of \cite{MW13}] Let $M$ be a multicurve such that we have one of the following two cases: \begin{itemize} \item $M = \{ \gamma_1, \gamma_2\}$ where $\gamma_1$ and $\gamma_2$ are arcs or closed curves intersecting at a point~$p$; \item $M = \{ \gamma \}$ where $\gamma$ has a self-intersection at a point $p$. \end{itemize} The \emph{smoothing} of $M$ at point $p$ is $s_p(M) = \{ M_1, M_2\}$ where $M_1$ (respectively $M_2$) is the same as $M$ except for the local change at $p$ that replaces the intersection or self-intersection point $\times$ with the pair of curves segments {\Large$~_\cap^{\cup}$} (or {$\supset \subset$}, respectively). \end{definition} The skein relations are an important theorem from \cite{MW13} and are recalled in the following statement. \begin{theorem}[Skein relations, \cite{MW13}] \label{smoothing} Let $m$ be a multicurve with an intersection at point $p$ and let $s_p(m) = \{ M_1, M_2\}$. Denote by $x_C$, $x_1$ and $x_2$ the product of the polynomials in $\mathcal{F}$ associated to the curves in $C$, $M_1$ and $M_2$, respectively. Then, \[ x_C = \pm x_1 \pm x_2. \] \end{theorem} \begin{lemma}\label{noclosedloops} Let $\gamma_1, ..., \gamma_r$ be arcs on the surface. Then the smooth resolution of the multicurve $\{\gamma_1, ..., \gamma_r\}$ contains at least one multicurve with exactly $r$ arcs and no closed loops. \end{lemma} \begin{proof} We prove the result by induction on the number $n$ of intersection points of the multicurve $\{\gamma_1, ..., \gamma_r\}$. Let $p$ be an intersection point. Then either $p$ is an point of intersection between two distinct arcs $\gamma_i$ and $\gamma_j$, or $p$ is a point of self-intersection of an arc $\gamma_k$. In the first case, smoothing at $p$ will give rise to two multicurves, each made of exactly $r$ arcs and with at most $n-1$ intersection points. Locally, the situation is as in the following picture, where the two arcs intersecting at $P$ are depicted around $p$. In the picture, we draw a multicurve $\{\gamma_1, \ldots, \gamma_r\}$ instead of writing the product $x_{\gamma_1} \cdots x_{\gamma_r}$. \begin{displaymath} \begin{tikzpicture} \draw (-1,1) -- (1,-1); \draw (-1,-1) -- (1,1); \draw (2,0) node {$= \quad \pm$}; \draw (3,1) .. controls (3.9,0) and (4.1,0) .. (5,1); \draw (3,-1) .. controls (3.9,0) and (4.1,0) .. (5,-1); \draw (5.5,0) node {$\pm$}; \draw (6,1) .. controls (7,0.1) and (7,-0.1) .. (6,-1); \draw (8,1) .. controls (7,0.1) and (7,-0.1) .. (8,-1); \end{tikzpicture} \end{displaymath} In the second case, smoothing at $p$ gives rise to two multicurves: one is made of $r$ arcs and a closed curve, and the other is made of $r$ arcs. Both multicurves have at most $n-1$ intersection points. This is illustrated in the following picture. Note that the loop in the leftmost and rightmost terms may not be contractible. \begin{displaymath} \begin{tikzpicture} \draw (0,1) .. controls (1,-2) and (1,2) .. (0,-1); \draw (2,0) node {$= \quad \pm$}; \draw plot [smooth] coordinates {(3,1) (3.2,0.2) (3.8,0.5) (4,0) (3.8,-0.5) (3.2,-0.2) (3,-1)}; \draw (5,0) node {$\pm$}; \draw (6,1) .. controls (6.5,1) and (6.5,-1) .. (6,-1); \draw (6.8,0) circle (0.2); \end{tikzpicture} \end{displaymath} In each case, one of the multicurves obtained by smoothing at $p$ is made only of arcs. This finishes the proof. \end{proof} \begin{lemma} \label{termeva} Let $\mathscr{A}(\mathbf{x},Q)$ be a cluster algebra associated to a triangulation of a surface. Let $x_1$ and $x_2$ be cluster variables. Then \[ x_1x_2 = \sum_{b \in \mathcal{B}} m_b b + \sum_{b' \in \mathcal{B}'} m_{b'} b'\] with $m_b$, $m_{b'} \in \mathbb{N}$ for every $b \in B$, $b' \in B'$, and \[ \sum_{b \in B} m_b > 0. \] \end{lemma} \begin{proof} This follows from Theorem \ref{positivebasis} and Lemma \ref{noclosedloops}. \end{proof} \section{Unistructurality of cluster algebra from unpunctured surface}\label{sect::proof} \subsection{Proof of the main result} \begin{theorem}\label{theo::main} Cluster algebras with trivial coefficients from marked surfaces without punctures are unistructural. \end{theorem} \begin{proof} Let $\mathscr{A}(\mathbf{u},Q)$ be a cluster algebra associated to a triangulation of a marked surface $(S,M)$ without punctures, where $\mathbf{u}=\{u_1,\dots,u_{n}\}$. Denote by~$\mathscr{X}$ the set of cluster variables of~$\mathscr{A}(\mathbf{u}, Q)$. Let $(\mathbf{v},R)$ be a seed such that the cluster variables of $\mathscr{A}(\mathbf{u},Q)$ and of $\mathscr{A}(\mathbf{v},R)$ are the same; denote by~$\mathscr{Y}$ the set of cluster variables of~$\mathscr{A}(\mathbf{v},R)$. Note that a consequence of that hypothesis is that $\mathbf{u}$ and $\mathbf{v}$ have the same number of elements, which is the cardinality of a transcendence basis of the (common) ambient field. We show that $\mathbf{v}$ is a cluster of $\mathscr{A}(\mathbf{u},Q)$ by using a proof very similar to the one of Lemma 3.1 in \cite{BM16}. On the contrary, suppose that there exist $v_1$ and $v_2$ in $\mathscr{X}=\mathscr{Y}$ which are compatible in $\mathscr{A}(\mathbf{v},R)$, but not in $\mathscr{A}(\mathbf{u},Q)$. Without loss of generality, we can furthermore assume that $v_1$ and $v_2$ belong to $\mathbf{v}$. Since $v_1$ and $v_2$ are not compatible in $\mathscr{A}(\mathbf{u},Q)$, they are associated to arcs that intersect each other in $(S,M)$. We know from Lemma \ref{termeva} that \begin{equation} \label{produitvar} v_1v_2 = \sum_{b \in \mathcal{B}} m_b b + \sum_{b' \in \mathcal{B}'} m'_b b', \end{equation} with $\sum_{b \in \mathcal{B}} m_b b \neq 0$. Using Lemma \ref{LaurentBracelet}, we know that each $b$ is a cluster monomial and each $b'$ is a Laurent polynomial in cluster variables with positive coefficients. Since $\mathbf{v}$ is a cluster, each $b \in \mathcal{B}$ can be written as Laurent polynomial in $\mathbf{v}$. The same is true for $\sum_{b \in \mathcal{B}} m_b b$, so write $$ \sum_{b \in \mathcal{B}} m_b b = \frac{p_{\mathbf{v}}}{m_{\mathbf{v}}}, $$ where $p_{\mathbf{v}}$ (respectively $m_{\mathbf{v}}$) is a polynomial (respectively a monomial) in $\mathbf{v}$. If $\sum_{b' \in \mathcal{B}'} m'_b b' = 0$, we obtain that $$v_1v_2m_{\mathbf{v}} = p_{\mathbf{v}},$$ so that~$p_{\mathbf{v}}$ is a monomial in~$\mathbf{v}$. Now, each~$b$ in the sum~$\sum_{b \in \mathcal{B}} m_b b$ is a product of cluster variables in~$\mathscr{X}=\mathscr{Y}$; it is thus a Laurent polynomial in~$\mathbf{v}$ with positive coefficients by Theorem~\ref{theo::recollectionsCluster}. Moreover, the only cluster variables in~$\mathscr{Y}$ which are Laurent monomials in~$\mathbf{v}$ are the elements of~$\mathbf{v}$ itself by \cite[Lemma 3.7]{CKLP13}. Thus, if~$p_{\mathbf{v}}$ is a monomial in~$\mathbf{v}$, then the sum~$\sum_{b \in \mathcal{B}} m_b b$ has only one term, say~$b$, which has to be a Laurent monomial in~$\mathbf{v}$; since~$b$ is a product of elements of~$\mathscr{X} = \mathscr{Y}$, it is thus a monomial in~$\mathbf{v}$. By the linear independence of cluster monomials (see, in this generality, \cite{CL12}), this implies that~$b=v_1v_2$. This is then the decomposition of~$b$ as a product of compatible elements of~$\mathscr{X}$; therefore,~$v_1$ and~$v_2$ are compatible in~$\mathscr{A}(\mathbf{u}, Q)$, a contradiction. Now, suppose that $\sum_{b' \in \mathcal{B}'} m'_b b' \neq 0$. Because $\sum_{b' \in \mathcal{B}'} m'_b b'$ is a sum of Laurent polynomials in cluster variables with positive coefficients, we can write $$ \sum_{b' \in \mathcal{B}'} m'_b b' = \frac{p_\mathscr{X}}{m_\mathscr{X}} $$ where $p_\mathscr{X}$ (respectively $m_\mathscr{X}$) is a polynomial (respectively a monomial) with positive coefficients in $\mathscr{X}$. Now, as before, we change notation to write $p_\mathscr{X}$ and $m_\mathscr{X}$ as Laurent polynomials in $\mathbf{v}$ with positive coefficients: $$ p_\mathscr{X} = \frac{p'_{\mathbf{v}}}{m'_{\mathbf{v}}} \text{ and } m_\mathscr{X} = \frac{p''_{\mathbf{v}}}{m''_{\mathbf{v}}}.$$ Equation \ref{produitvar} becomes $$ v_1 v_2 = \frac{p_{\mathbf{v}}}{m_{\mathbf{v}}} + \frac{\left(\frac{p'_{\mathbf{v}}}{m'_{\mathbf{v}}}\right)}{\left(\frac{p''_{\mathbf{v}}}{m''_{\mathbf{v}}}\right)},$$ which is equivalent to $$ m_{\mathbf{v}}m''_{\mathbf{v}}p'_{\mathbf{v}} = m'_{\mathbf{v}}p''_{\mathbf{v}}(y_1y_2m_{\mathbf{v}} - p_{\mathbf{v}}). $$ Note that in the left-hand sight of the equation, there is a polynomial in $\mathbf{v}$ with positive coefficients, while there is a minus sign on the right-hand sight. Since $x_1x_2m'_y$ is a monomial, we deduce that $y_1y_2m'_{\mathbf{v}} = p_{\mathbf{v}}$, and we obtain as above that~$y_1$ and~$y_2$ are compatible in~$\mathscr{A}(\mathbf{u}, Q)$, a contradiction. Therefore, $v_1$ and $v_2$ are two compatible variables in $\mathscr{A}(\mathbf{u},Q)$ if they are compatible in $\mathscr{A}(\mathbf{v},R)$. It follows from the above that any cluster of~$\mathscr{A}(\mathbf{v},R)$ is a cluster of~$\mathscr{A}(\mathbf{u},R)$. This also implies the converse: any cluster of~$\mathscr{A}(\mathbf{u},R)$ is a cluster of~$\mathscr{A}(\mathbf{v},R)$. Indeed, let~$\mathbf{w}$ be a seed in~$\mathscr{A}(\mathbf{v},R)$, and let~$\mathbf{w}_1$, \ldots, $\mathbf{w}_n$ be the neighbouring clusters in the exchange graph of~$\mathscr{A}(\mathbf{v},R)$. Then, by~\cite[Theorem 5]{GSV08}, $\mathbf{w}, \mathbf{w}_1, \ldots, \mathbf{w}_n$ are neighbouring clusters in~$\mathscr{A}(\mathbf{u},R)$, since for every~$i$, $\mathbf{w}$ and~$\mathbf{w}_i$ only differ by one cluster variable. As in Remark~\ref{rema::quivers or opposite}, this allows us to recover the quiver~$Q$ in the seed~$(\mathbf{w},Q)$ containing~$\mathbf{w}$, up to a change of orientation of each of its connected components. This argument works in~$\mathscr{A}(\mathbf{u},R)$ as well as in~$\mathscr{A}(\mathbf{v},R)$, so we get a seed~$(\mathbf{w},Q')$ in~$\mathscr{A}(\mathbf{u},R)$. Since the connected components of~$Q$ and~$Q'$ are the same up to a change of orientation, the cluster algebras~$\mathscr{A}(\mathbf{u},R)$ and~$\mathscr{A}(\mathbf{v},R)$ have the same clusters. By Proposition~\ref{prop::clustersAreSufficient}, this shows that~$\mathscr{A}(\mathbf{u},Q)$ is unistructural. \end{proof} \subsection{Consequences for cluster automorphisms} The idea of unistructurality first appeared in \cite{ASS14} while studying automorphisms between cluster algebras. \begin{definition}[\cite{ASS12}] Let $\mathscr{A}(\mathbf{u},Q)$ be a cluster algebra, and let $f : \mathscr{A}(\mathbf{u},Q) \rightarrow \mathscr{A}(\mathbf{u},Q)$ be an automorphism of $\mathbb{Z}$-algebras. Then $f$ is called a \emph{cluster automorphism} if there exists a seed $(\mathbf{v},R)$ of $\mathscr{A}(\mathbf{u},Q)$, such that the following conditions are satisfied: \begin{itemize} \item $f(\mathbf{v})$ is a cluster in $\mathscr{A}(\mathbf{u},Q)$; \item $f$ is compatible with mutations, that is, for every $v \in \mathbf{v}$, we have \[f\left(\mu_{v}(\mathbf{v})\right) = \mu_{f(v)}(f(\mathbf{v})).\] \end{itemize} \end{definition} Remark that the existence of such a cluster $\mathbf{v}$ guarantees that the image of any cluster is a cluster. The following is a direct consequence of \cite[Lemma 5.2]{ASS14} and Theorem~\ref{theo::main}. \begin{crl} Let $\mathscr{A}(\mathbf{u},\tau)$ be a cluster algebra arising from a triangulation~$\tau$ of an unpunctured marked surface~$(S,M)$. Then a map $f:\mathscr{A}(\mathbf{u},\tau) \rightarrow \mathscr{A}(\mathbf{u},\tau)$ is a cluster automorphism if and only if $f$ is an automorphism of the ambient field which restricts to a permutation of the set of cluster variables. \end{crl} It is conjectured in \cite{ASS14} that this result is true for any cluster algebra, since, by \cite[Lemma 5.2]{ASS14}, it is true for any unistructural cluster algebra, and all cluster algebras are conjectured to be unistructural. \section*{Acknowledgements} The authors would like to thank Hugh Thomas for fruitful discussions and his comments on an earlier version of this work which resulted in a simplification of one of the arguments. \bibliographystyle{alpha}

647e773c46fa74526c9075330d7c9587d8f30613

\section{Introduction} Online advertising has grown exponentially over the past few years due to the wide spread usage of internet across the world. The marketers track customer journeys as they are exposed to different online media channels(e.g. email, display, paid search) before they make the conversion at the end. Companies allocate marketing budgets to promote their business through these multiple online campaigns among different channels. To get maximum return on investment on the spend of online ads, marketers have to optimize their budget allocation among different media channels based on their value. How to measure the value of ads spend, however, is not trivial for marketers. The problem of measuring the influence of each campaign or channel on a conversion is referred as attribution problem \cite{kannan2013Experiment} . \begin{figure} \centering \includegraphics[width=.8\linewidth]{human_dark_path.png} \caption{A possible behavioral customer journey in an online advertising system. Here, the user is exposed to display, paid search and email touch points, but he or she may choose to make conversion or not at the end } \label{fig:human_conversion_path} \end{figure} As shown in Figure[\ref{fig:human_conversion_path}], a user may be exposed to email, display, paid search ads before the users converts. Each ad has a relation with the user's final conversion decision. In such a case, the marketer faces a dilemma of assessing the contribution of each channel to user's conversion. Marketers have applied simple rule-based heuristics to solve attribution problem in the past. First or last touch point approach ignore the effects of other channels; equal weight approach assume equal contribution from each channel, which ignores the channel difference; time-decayed attribution algorithm assumes that the credit decays based on a decay parameter which is simply based on intuition without data support. In order to rectify the above pitfalls, data-driven attribution models have been introduced in recent years. In this paper, we propose a data-driven multi-touch attribution and conversion prediction model denoted as deep neural net with attention for multi-touch attribution (DNAMTA) that outperforms the other approaches in terms of both conversion prediction and attribution analysis. \section{Related Work} In order to overcome the drawbacks of rule-based heuristics, data-driven algorithmic models were proposed. Shao et al. \cite{Shao2011KDD} propose a bagged logistic regression method and compares it with a probabilistic model. They predict conversion rate using count of ad occurrences and uses weights as credits for attribution analysis. While bagging provides stable estimates and better accuracy than probabilistic model, they do not have an interpretable model and ignores temporal factor. Dalessandro et al. \cite{brian2012causal} propose causal inference methods to achieve interpretability. They used additional marginal lift of each ad as credits. Since their method was computationally difficult, under some assumptions, they were unable to estimate causal parameters. Ji et al. \cite{Ji2016Probabilistic} adopt a probabilistic framework to remove the presentation biases. However, they do not directly measure the effect of ad exposure. Zhang et al. \cite{Zhang2014Survival} propose data-driven multi-touch attribution with survival theory but do not consider user characteristics.Ji et al. \cite{Ji2017Survival} use hazard rate to reflect the influence of an ad exposure. However, they assume that the impact of ad exposures is additive and fades with time. Abhishek et al. \cite{abhishek2012HMM} propose multistage model of consumer response to advertising activity that addresses the problem of temporal dynamics of ad exposure. However, their framework is difficult to achieve model scalability, besides, higher order markov chains are hard to be implemented for better model accuracy. Deep Learning \cite{Goodfellow-et-al-2016} have been used extensively in image \cite{larochelle2010learning}, speech recognition and language translation \cite{bahdanau2014neural} to achieve state-of-art results. Attention mechanism embedded with Neural Network has been successfully applied in vision and NLP field \cite{denil2012learning,bahdanau2014neural}, as attention mechanism can emphasize the important features along the time-series observations. These novel ideas, however, are not yet used to tackle problems like attribution. \section{Notation and Problem Formulation} \label{notation} \par We formalize the attribution problem as follows. An event is either a conversion or a touchpoint. Each customer path consists of events from multiple advertising channels. Let $x_t$ denote the $t^{th}$ event the user is exposed to in the path and $x_t\in \mathbb{M}$, $\mathbb{M}$ is the set of all the touchpoints that we are interested in. Thus, a single customer sequence path $P_i$ can be represented as $P_i = \{x_1, x_2, \dots, x_T\}$, $T$ is the length of the sequence. $t$ represents the relative order of the event in the sequence, instead of the absolute event occurrence time. Beyond that, each event is also associated with some structure information, such as occurrence time, which can be formalized as another sequence $\{U_1, U_2, \dots, U_T\}$. In addition to these dynamic sequence information, some static information which is unlikely to be changed during the conversion journey, such as gender, age, sign-up date etc., are represented as control variables $C_i$. A customer path will be treated as positive if it ends with conversion($Y_i =1$), otherwise it's a negative(non-conversion) path($Y_i = 0$). Assuming each touchpoint $x_t$ has attribution value $a_t$, then $\sum_{t=1}^Ta_t = 1$. The objective of this attribution problem is to estimate attribution value $a_t$ which represents the touchpoint $x_t$'s contribution towards a successful conversion. \par To make this problem more mathematically well-defined, we use probabilistic reasoning to explain customer's conversion decision, i.e. we want to find how likely a path will end up with conversion if it is exposed to a sequence of touchpoints $P_i$ and its corresponding control variable $C_i$. We denote this as conditional probability $P(Y_i|P_i, C_i)$. According to Bayes formula, $P(Y_i|P_i, C_i) = P(Y_i, P_i | C_i)/P(P_i|C_i)$ and in order to get a good inference of this conditional probability we should have a good estimate of two components: $P(Y_i, P_i |C_i)$ and $P(P_i|C_i)$. $P(Y_i, P_i|C_i)$ can be estimated by maximum likelihood estimation(MLE) from the data. Since $P_i$ is a dynamic sequence observation with varying length, estimating $P(P_i|C_i)$ is difficult. Furthermore, if we use a naive one-hot representation by aggregating through time, it ignores the time variance information. Therefore, it's necessary to have a better representation of $P_i$, that helps to estimate probability $P(P_i|C_i)$ and $P(Y_i | P_i, C_i)$ easily. We use a learning function $f$ to approximate this conditional probability $P(Y_i|P_i, C_i) = f(C_i, \{x_t\}_{1:T})$. Thus the underlying structure for attribution of each touchpoint can be estimated from this learning function. \par Attribution problem is complex as hidden interactions between touchpoint needs to be modeled. Besides, contribution of touchpoint decreases with the increasing time lag(defined as the duration between the occurrence time and the end time) in a path. This typical time decay property is a common business assumption, which is unlikely to be captured by general linear model. Lastly, control variables like gender, age, sign up date etc. can also affect customer journey. \par We propose a general deep learning framework in order to solve the above three challenges: DNAMTA. This model has three advantages: 1) \textit{DNAMTA with attention} is a Long Short Term Memory (LSTM) based deep sequential model, which is well known for capturing the long time dependency of sequence observations\cite{hochreiter1997long}. Further, attention mechanism is used to capture the touchpoint contextual dependency. 2) Survival time-decay functions are introduced in \textit{DNAMTA with timedecay} to explicitly model the timedecay assumption. 3) \textit{DNAMTA fusion} model can combine static information of user as control variables with dynamic touchpoint observations. \vspace*{-.2cm} \section{ Deep Neural Network with Attention for Attribution} \par In a sequence of observations of touch points, same touchpoint may be differentially important at different time locations and at different frequency of occurrence. Our model introduces attention mechanism that lets the model to pay more or less attention to individual touchpoints when constructing the representation of the customer path. To demonstrate the idea, let's take a look at Figure(\ref{fig:att_heatmap}), which is a positive path where the customer finally made a conversion at the end. This customer has been exposed to a sequence of advertising events before a conversion decision is made. Each touchpoint is allocated different contribution value according to our model. The contribution of touchpoint "Email Sent" varies at different observation locations. Besides, touchpoint "Email Sent" has totally different importance compared with other touchpoints, such as "Display Impression". Details of our proposed model will be covered in Section[\ref{model}]. \par Attention serves two benefits: it not only provides us reasonable better performance, but also gives insight on how touchpoint contributes to the conversion decision at any specific time which is the most valuable part of an attribution conversion problem. LSTM could help us handle capture the hidden underlying complex interaction patterns. \begin{figure} \includegraphics[width=.9\linewidth]{attribution_heatmap.png} \caption{A heatmap visualizing the contribution of each touchpoint in a specific customer conversion path. From left to right, customer journey starts from the first event to the end of conversion, all events are coded by:display_click(DC), display_impression(DI), email_click(EC), email_open(EO), email_sent(ES), paid_search(PS). Y-axis indicates the conversion probability from the prediction model. The darker the color for a touchpoint, the higher influence of the corresponding touchpoint is. } \label{fig:att_heatmap} \end{figure} \vspace*{-.2cm} \section{Model} \label{model} \label{fig:att_model} \vspace*{-.2cm} \subsection{Touchpoint Input Layer} \par The input for the model is a touchpoint sequence path $P$ with one-hot representation of touch points $x_{t}, t\in [0, T], x_{t}\in \mathbb{R}^{v_{tp}}$, where $v_{tp}$ is the total number of all possible touchpoints we are interested in and $T$ is the length of the sequence, which varies for different sequences. Note that this sequence only considers absolute order, the real temporal difference between each touchpoint pair could be different. Detailed information about dealing with temporal relation will be discussed in later section [\ref{att_timedecay}]. \vspace*{-.2cm} \subsection{Touchpoint Embedding Layer} \par Given a path $P$ in the above format, we first transform the one-hot representation of the touchpoint at step $t$ to a dense vector through an embedding matrix $W_e\in \mathbb{R}^{v_e \times v_{tp}}$ by $e_{t} = W_ex_{t}$. Specifically, $t^{th}$ column of embedding matrix $W_e$, which is a vector of $v_e$ dimension, is the continuous representation vector of step $t$ touchpoint observation. \par Traditional one-hot representation or bag-of-words like feature representation are simply counting statistics, which ignore touchpoint contextual similarities and suffer from sparsity in representation. Touchpoint embedding aims to quantify and categorize hidden contextual similarities between each touchpoint based on their distribution in large samples of touchpoint paths. \vspace*{-.15cm} \subsection{Variable-depth LSTM Layer} \par We use LSTM \cite{hochreiter1997long} to obtain another level of representation of touchpoints by using embedding layer output $\{e_{1}, \dots, e_{T}\}$, and therefore incorporate the contextual information in the historical observations. Each LSTM block updates current hidden state output $h_{t}\in \mathbb{R}^{v_h}$ based on embedding output $e_{t}$ and previous hidden state output $h_{ t-1}$ through the formula \begin{eqnarray} h_{t}& = & \mathcal{H}(e_{t}, h_{t-1}), ~ t\in [0, T] \label{eq:hidden_state} \end{eqnarray} \par In Formula(\ref{eq:hidden_state}), $\mathcal{H}$ is a nonlinear transformation function, which has various definitions according to practical problems. \par Now each $h_t$ can be considered as a new representation of $t^{th}$ touchpoint by overviewing all historical touchpoint records, so conceptually $h_t$ is able to better describe the context meaning of touchpoint in the specific path compared with the raw embedding vector $e_t$, which is unaware of past information. This is important for customer conversion journey, since the order, frequency and long-term dependency of touchpoint exposure could have a high impact on their final conversion decision. \vspace*{-.2cm} \subsection{Touchpoint Attention Layer} \par We introduce attention mechanism to find touchpoints that are important to the conversion and aggregate the representation of those informative touchpoints to form a path vector. Yang et al. \cite{yang2016hierarchical} proposed hierarchical attention mechanism for text sentiment analysis. We shall leverage this idea in our case. Specifically, \begin{eqnarray} v_{t} & = & tanh(W_v h_{t} + b_v)\\ a_{t} & = & \frac{exp(v_{t}^Tu)}{\sum_{t}exp(v_{t}^Tu)} \label{eq:att}\\ s & = & \sum_t a_{t} h_{t} \end{eqnarray} We first feed the touchpoint representation $h_{t}$ through a one-layer multilayer perceptron(MLP) to get $v_{t}$ as a hidden representation of $h_{t}$, then we measure the importance of the word as the similarity of $v_{t}$ with touchpoint context vector $u$ and get a normalized importance weight $a_{t}$ through a softmax function. We can notice that by design $a_t > 0$. The advantage of this construction is that the contribution of every touchpoint is always positive. After that, we compute the path vector $s$ as the weighted sum of the touchpoint representation based on the non-negative weights. Actually, $s$ is the convex combination of all $h_{t}$. The context vector $u$ can also be seen as a high level representation of a fixed sequence based on our domain knowledge about touchpoint importance, campaign marketers can custom their attribution model by constraining vector $u$. The context vector $u$ can either be fixed or be randomly initialized and jointly learned during the process. We use the latter approach in our modeling. \subsection{Touchpoint Path Classification} \label{tp_classification} \par In our attribution conversion problem, some customer touchpoint journeys end up with conversions, these paths are treated as positive paths, otherwise, they are negative paths. With these labels, we can consider this attribution conversion learning problem as binary classification problem in the new path vector space. The path vector $s$ is a high level representation of the customer touchpoint journey by combining hidden outputs and attention weights. \begin{eqnarray} p& = & sigmoid(\sigma(W_c^T s)+ b_c) \end{eqnarray} where $W_c \in \mathbb{R}^{v_h}$ and $\sigma(\cdot)$ is nonlinear activation function ReLU $\sigma(x) = \max{(0, x)}$. In common binary classification problems, the probability for predicting the sequence observation path positive is usually the sigmoid function for linear combination of features. In attribution conversion problem, with some exposure of advertising channels, the probability for customers to make conversion decision is always greater than those without any exposure, which means the contribution of touchpoint for conversion is always nonnegative. Activation function ReLU is mathematically fit for this practical constraint. \subsection{Time Decay Attention Layer} \label{att_timedecay} \par Attention mechanism is widely used in NLP problems where the distance between each word is relative, depending on the word counts between them. We should consider exact time gap in attribution problem, since the time gaps between each touchpoint vary a lot, from hours to even months. This difference of time gaps could affect the connection strength of nearby touchpoints and further impact the final conversion. Therefore, we introduce the time decay attention layer by combining time decay information, inspired by the idea in \cite{wooff2015TimeWeighted}. Basically each touchpoint sequence observation has its occurrence time, the time gap difference between the occurrence time and the end time defined as $T_t$. The smaller $T_t$ is, nearer is the occurrence time to end time. We assume the touchpoint contribution decreases when the occurrence time is far away from the end time. We penalize each attention weight described in component in Formula(\ref{eq:att}) by non-increasing timedecay function. Detailed formula can be referred as below: \begin{eqnarray} v_{t} & = & tanh(W_v h_{t} + b_v)\\ a_{t} & = & \frac{exp(v_{t}^Tu - \lambda T_t)}{\sum_{t}exp(v_{t}^Tu - \lambda T_t)} \label{eq:att_decay}\\ s & = & \sum_t a_{t} h_{t} \end{eqnarray} where $\lambda > 0$ is the decay parameter, it can be predefined based on data analysis of past customer conversion trend, or it can also be randomly initialized in model and directly learned from data. \subsection{Fusion Model} \par As we have mentioned in previous section, attribution conversion models usually try to establish the relationship between advertising channels and final conversion. However, customer characteristic information such as gender, age and some other static information may affect the touchpoint exposure and the conversion engagement. \cite{rosenbaum1984reducing} points out that the confounding effects from these features could affect the distribution of conversion rate. For example, free signup is a promotion strategy from company to encourage customer to make conversion. Generally there are two situations when a conversion rate may peak: First, when customers free signup and second,when this free signup trial expires. Therefore, it's necessary for us to take these control variables into consideration, which helps us to minimize the potential bias inference effects. \par However, the number of control variables in real attribution conversion problems can be very large, which increases the difficulty of the variable selection among these control sets. Besides, a simple linear add-on may not fit for describing the complex relationship between the factors and conversion. In order to account for these two problems, we propose a fusion model, which is built on the original DNAMTA model by introducing another deep neural network for control variable learning. In Figure(\ref{fig:att_fusion}), deep neural network modeling control variables is on the right hand. It aims to learn a sophisticated feature vector representation by going through several dense fully connected layers, which can capture the underlying structure. Later we concatenate the customer touchpoint path representation vector and the control variable vector before we apply it to classification layer. The touchpoint path classification formula will be changed \begin{eqnarray} p& = & sigmoid(\sigma_1(W_{c_{tp}}^T s)+ \sigma_2(W_{c_{ntp}}^T v) + b_c) \label{eq:fusion} \end{eqnarray} where $\sigma_1$ is still the RELU function as mentioned in Section [\ref{tp_classification}], and $\sigma_2$ is just identical function. \begin{figure} \centering \includegraphics[width=1.0\linewidth]{rnn_attention_fusion.png} \caption{The structure of DNAMTA fusion model, including three parts: sequence encoder, control variable encoder and sequence classification} \label{fig:att_fusion} \end{figure} \section{Experiments} \par In this section, we present our experiments to compare different attribution models with DNAMTA. We also discuss the corresponding results and model interpretation for attribution. \subsection{Data} \par We ran our experiments on large event data set of a marketing organization with three primary channels display, email and paid search. We have 6 different touchpoints: display click, display impression, email click, email sent, email open and paid search. It contains 426853 records with history of 57 days including conversion day. Each record represents a customer's journey, if this journey ends up with conversion action before the given data collection time, it is regarded as positive path; otherwise it's labeled as negative. Even though customer may convert in the future, this kind journey is still not positive based on our definition. Due to the heavy imbalanced distribution of positive and negative paths in real dataset, we down sampled the negative path records to get the dataset with balanced labels. Each path record is associated with a free-signup date, a sequence of dates for each touchpoint event and a sequence of frequency of occurrence. A visit duration window is applied to multiple visits from the same advertising channel: subsequent visits are ignored if they occur within a short time. \par We randomly split this data into two sets: 80\% for training and 20\% for testing. All experiment comparison results are based on the test dataset. \subsection{Model Settings and Implementation} As mentioned in \cite{Eva2014Graph}, we will mainly focus on predictive accuracy(AUC) and interpretability. To demonstrate the performance of various attribution models, we compare our DNAMTA model with three commonly used attribution models i.e. last touch attribution, Logistic Regression and HMM \cite{abhishek2012HMM} in our experiments: \begin{itemize} \item \textbf{LSTM} is the fundamental LSTM model without attention mechanism. \item \textbf{DNAMTA} is the first version of our deep attribution model with attention mechanism. After getting the outputs from each LSTM module, we will calculate the attention weights based on Formula(\ref{eq:att}), later we will use the re-weighted LSTM outputs as a path representation vector for binary classification modeling. \item \textbf{DNAMTA with time decay} is the second version of our deep attribution model. Besides incorporating attention mechanism, it also accounts for temporal-effect in attribution. The time decay weighted attention calculation formula is followed by Formula(\ref{eq:att_decay}). For simplification, we assume time decay parameter stays the same for all channels, but each channel (e.g. page search) could have its own time decay parameter. \item \textbf{Fusion DNAMTA} is the third version of our deep attribution model. Built on the top of time decayed DNAMTA, control variables such as user activity will be learned as a feature representation vector in another neural network. A fused representation vector is generated by concatenating touchpoint path vector and control variable vector, and it will be used for classification task based on the Formula(\ref{eq:fusion}). \end{itemize} \par We use TensorFlow 1.2.0 \cite{abadi2016tensorflow} and Python 3.0 for all deep model implementation, and sklearn 0.18.1, pomegranate for baseline model implementation. All the comparison experiments are run on GPU Tesla K80 and CPU. For LSTM model we choose to use stochastic Adam gradient descent \cite{kingma2014adam} for training. In deep model, both the dimension of hidden size and attribution dimension (a.k.a. contextual vector $u$'s dimension) are 64. We use 3 hidden layers. During training process, a validation data set is hold out for hyper parameter tuning, and the model training process stops when the validation loss stops improving. \subsection{Results} \par In Table \ref{tab:prediction}, we report the prediction performance of all attribution models on testing dataset. We can observe that DNAMTA fusion model successfully utilizes both time and touchpoint dependent representation and confounding factors, and it achieves the highest prediction accuracy and AUC. Besides, on comparing with other models, we find that deep model with attention can generally improve the model prediction performance, which indicates the impact of attention mechanism in dynamic sequence path classification task, as attention can smartly reconsider touchpoint contextual dependencies and reallocate these touchpoint contributions. \par As we mentioned in previous section, model prediction is not the only goal for attribution modeling. From the perspective of representation learning, a good representation for dynamic path is good for future statistical inference and strategy decision making. The path representation vector from last touchpoint attribution model is simple without modeling, but it does not capture the time dependency between each touchpoint. If both a long touchpoint sequence and a short one ends with up the same touchpoint, these two paths will be considered same in the last touchpoint prediction model. For logistic regression, the path representation vector considers the touchpoint content information and time information, but the dimension of this vector can be dramatically high and sparse when our predefined observation time window grows. For our dataset that spans touchpoint data of 57 days, the feature dimension in logistic regression is 342. However, in DNAMTA model, the path representation dimension is only 64 and also achieves better prediction performance than logistic regression does, which shows us the efficiency of representation provided by DNAMTA. \par Similar to approach and arguments in \cite{mnih2014ram}, both the number of parameters in our model and the amount of computation it performs can be controlled independently of the size of the path if we fix the length of the customer path that is considered. Hence it is easily scalable with any size of data. In the case where we do not fix the path length, the computational demands scale linearly with the length of the path in consideration. \subsection{Modified Attribution Score with Attention} \begin{comment} \begin{table}[htb] \centering \caption{Fractional attribution values for different advertising channel} \label{frac} \small \begin{tabular}{|p{1.1cm}|p{.6cm}|p{.6cm}|p{.6cm}|p{1cm}|p{1cm}|p{1cm}|} \hline & LTA & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Display & 0.3919 & 0.5380 & 0.6423 & 0.4477 & 0.3985 & 0.4111 \\ \hline Email & 0.3826 & 0.2406 & 0.1742 & 0.3623 & 0.3836 & 0.3717 \\ \hline PaidSearch & 0.2253 & 0.2213 & 0.1835 & 0.1898 & 0.2177 & 0.2171 \\ \hline Total & 1 & 1 & 1 & 1 & 1 & 1 \\ \hline \end{tabular} \end{table} \begin{table}[htb] \centering \caption{Incremental attribution values for different advertising channel} \label{margin} \small \begin{tabular}{|p{1.1cm}|p{.6cm}|p{.6cm}|p{.6cm}|p{1cm}|p{1cm}|p{1cm}|} \hline & LTA & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Display & 0.3250& 0.3596 & 0.3915 & 0.3691 & 0.3258& 0.3410 \\ \hline Email &0.1334 & 0.1547 & 0.1580 & 0.1687 & 0.1829& 0.1795 \\ \hline PaidSearch & 0.2128& 0.1622 & 0.1310 & 0.1762 & 0.2064 & 0.2068 \\ \hline Total & 0.6714 & 0.6765& 0.6805 & 0.7141& 0.7151 & 0.7273 \\ \hline \end{tabular} \end{table} \begin{table}[htb] \centering \caption{Model prediction performance numerical values summary and comparison} \label{tab:prediction} \small \begin{tabular}{|p{.9cm}|p{.6cm}|p{.6cm}|p{.6cm}|p{.6cm}|p{1cm}|p{.8cm}|p{.8cm}|} \hline & LTA & HMM & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Accuracy & 0.7651 &0.7655& 0.7885 & 0.8072 & 0.8072 & 0.8072 & 0.8187 \\ \hline AUC & 0.8004 &0.8005& 0.8456 & 0.8406 & 0.8552 & 0.8513 & 0.8793 \\ \hline \end{tabular} \end{table} \end{comment} \par We propose a novel usage of the attention scores by incorporating it with traditional attribution score calculation \cite{kannan2013Experiment}: fractional attribution score and incremental attribution score. \vspace*{-.2cm} \begin{itemize} \item \textbf{Incremental score} We estimate the impact of a specific channel on the conversion by calculating the difference in conversion probabilities with and without the channel. \item \textbf{Fractional score} We normalize all incremental scores of each channel for each path, and aggregate all incremental contributions at channel level as the fractional score. \item \textbf{Attention based score} Attention values learned from deep model can be directly used as fractional score, as it serves as the contribution of each touchpoint after accounting for the interaction between each other. \end{itemize} \begin{table}[htb] \centering \caption{Fractional attribution values for different advertising channel} \label{frac} \small \begin{tabular}{|p{1.1cm}|p{.5cm}|p{.5cm}|p{.5cm}|p{1cm}|p{1cm}|p{1cm}|} \hline & LTA & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Display & 0.392 & 0.538 & 0.642 & 0.448 & 0.398 & 0.411 \\ \hline Email & 0.383 & 0.241 & 0.174 & 0.362 & 0.384 & 0.372 \\ \hline PaidSearch & 0.225 & 0.221 & 0.184 & 0.190 & 0.218 & 0.217 \\ \hline Total & 1 & 1 & 1 & 1 & 1 & 1 \\ \hline \end{tabular} \end{table} \begin{table}[htb] \centering \caption{Incremental attribution values for different advertising channel} \label{margin} \small \begin{tabular}{|p{1.1cm}|p{.5cm}|p{.5cm}|p{.5cm}|p{1cm}|p{1cm}|p{1cm}|} \hline & LTA & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Display & 0.325& 0.356 & 0.392 & 0.369 & 0.326& 0.341 \\ \hline Email &0.133 & 0.155 & 0.158 & 0.169 & 0.183& 0.180 \\ \hline PaidSearch & 0.213& 0.162 & 0.131 & 0.176 & 0.206 & 0.207 \\ \hline Total & 0.671 & 0.673& 0.681 & 0.714& 0.715 & 0.728 \\ \hline \end{tabular} \end{table} \begin{table}[htb] \centering \caption{Model prediction performance numerical values summary and comparison} \label{tab:prediction} \small \begin{tabular}{|p{.9cm}|p{.5cm}|p{.5cm}|p{.5cm}|p{.5cm}|p{1cm}|p{.5cm}|p{.5cm}|} \hline & LTA & HMM & LR & LSTM & DNAMTA & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ timedecay\end{tabular}}& \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}DNAMTA \\ fusion\end{tabular}}\\ \hline Accuracy & 0.765 &0.766& 0.789 & 0.807 & 0.807 & 0.807 & 0.819 \\ \hline AUC & 0.800 &0.801& 0.846 & 0.841 & 0.855 & 0.851 & 0.879 \\ \hline \end{tabular} \end{table} \vspace*{-.15cm} \par In all attribution models, display accounts for the most contribution for customer conversion. Especially for logistic regression and DNAMTA model, both fractional and incremental attribution scores for Display are very high. While after incorporating time decay property, DNAMTA with timedecay and fusion model lowers the display attribution scores. Indeed, customer has to be exposed to the product first before they can start their conversion journey. Display triggers the continuing advertising exposure, while display is usually less likely to show up closer to conversion. We didn't include HMM in the comparison table (\ref{margin}, \ref{frac}), because the attribution scores for HMM are quite similar to others. \par Figure(\ref{fig:time_dependency}) visualizes attribution density distributions for each touchpoint over various ad exposure lag. Overall, display accounts for the most conversion contribution, but among customers with different ads time exposure, touchpoint contribution distributions vary. For example, paid search has relatively high impact within the first week, but this contribution decreases for long time exposure of ads. As we mentioned in Section [\ref{notation}], DNAMTA is capable of capturing the underlying structure of touchpoints and their conversion contributions. \par Figure(\ref{fig:time}) shows the time decay of attribution for each touchpoint. As the time lag (the difference between observation time and the end time of path) increases, the attribution for each touchpoint decreases. It confirms the time decay property for attribution score. The variance of on-average attribution score at specific time lag also has a decreasing trend as the time lag increases. The most latest advertising exposure may contribute a lot to customer's final conversion decision. \begin{figure}[htb] \centering \includegraphics[width=1.\linewidth]{time_dependency.png} \caption{Attribution estimate density distributions for each ad channel vary over different ad exposure time. The area under the curve of a density function represents the probability of getting specific attribution values between a range. The number of days until customer convert ranges from top left to bottom right are: 0-7, 7-30, 30-56, 0-56. Paid search has relatively high impact within the first week,but this contribution decreases for long time exposure of ads. } \label{fig:time_dependency} \end{figure} \begin{figure}[htb] \centering \subfigure[Display Click]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionaldisplay_click.png} \label{fig:DC} } \subfigure[Display Impression]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionaldisplay_impression.png} \label{fig:DI} } \subfigure[Email Click]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionalemail_click.png} \label{fig:EC} } \subfigure[Email open]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionalemail_open.png} \label{fig:EO} } \subfigure[Email Sent]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionalemail_sent.png} \label{fig:ES} } \subfigure[Paid Search]{ \includegraphics[width=0.3\linewidth]{attribution_timelag_fractionalpaid_search.png} \label{fig:PS} } \caption{Mean fractional attribution measured on y-axis decreases as time lag increases indicating the time decay property for all the touchpoints. Variance in mean fractional attribution indicated by the grey shadow area also has decreasing trend.} \label{fig:time} \end{figure} \vspace{-.2cm} \section{Conclusion and Future Work} \par In this paper, we introduce DNAMTA, a deep neural network framework incorporating attention mechanism, by considering temporal effect and user characteristics through control variable adjustment. It aims to have a deeper understanding about the dynamic interactions between advertising channels and their contributions to customer conversion. For predictive task, DNAMTA surpasses some widely used attribution models as well as basic LSTM model. For interpretability, DNAMTA can also provide good insights of the relative touchpoint attribution estimates. \par Through the discussion in this paper, we also formalize attribution as a representation learning problem. Experiment results show that the dynamic path vector representation of dimension 64 from DNAMTA achieves better prediction performance compared to other attribution models. A good representation for dynamic advertising channels is not only good for prediction task and statistical inference, but also can be beneficial for transfer learning: transferring the domain knowledge and data-driven features to some other marketing problems with limited data observations. Marketers can thus allocate their budget spends on touchpoints in proportion to their contributions. \bibliographystyle{unsrt}

c1138e58dc548a365e664609da59ae3865505e8c

\section{Introduction and motivation}\label{S:Intro} Presumably, the binomial distribution $\operatorname{Bin}(n, p)$ is the most fundamental and simple model for the repetition of independent success/failure events. When both parameters $p$ and $n$ are unknown, which is the topic of this paper, it serves as a basic model for many applications. For example, $n$ corresponds to the population size of a certain species \citep{Otis, Royle2004, Raftery}, the number of defective appliances \citep{DG} or the number of faults in software reliability \citep{Basu}. In Section \ref{S:DataExample} we elaborate on a novel application where $n$ is the number of unknown fluorescent markers in quantitative super-resolution microscopy \citep{Hell, AEM2015}. Accordingly, joint estimation of the population size $n$ and the success probability $p$ of a binomial distribution from $k$ independent observations has a long history dating back to \citet{Fisher}. In contrast to the problem of estimating $p$ or $n$ when one of the parameters is known \citep{Lehmann}, this is a much more difficult issue. Fisher suggested the use of the sample maximum (which is a consistent estimator for $n$ as $k\to\infty$) and argued that the estimator is always "good", as long as the sample size is large enough. In fact, if $X_1,\dots, X_k\overset{i.i.d.}{\sim}\operatorname{Bin}(n,p) $ for fixed $n$ and $p$, the sample maximum converges exponentially fast to $n$ as $k\rightarrow \infty$ since \begin{align}\label{max_exp_fast} \mathbb{P}\big( \max_{i=1, \ldots, k} X_i = n \big) = 1-\mathbb{P}\big( \max_{i=1, \ldots, k} X_i <n \big) =1-(1-p^{n})^k. \end{align} While true asymptotically, the maximum very strongly underestimates the true $n$ even for relatively large sample size $k$ if the probability of success is small. This is explicitly quantified in \cite{Gupta}: if $p=0.1$ and $n=10$, then the sample size $k$ needs to be larger than $3635$ to ensure that $\mathbb{P}(\max_{i=1, \ldots, k} X_i\geq n/2)\geq1/2$. If $p=0.1$ and $n=20$, one would need a sample size of more than $k=900{,}000$ to guarantee the same probability statement as above. This fallacy of the sample maximum can be explicitly seen in a refined asymptotic analysis for $n$ and $p$ as well. By Bernoulli inequality and since $1-x \leq e^{-x},$ it follows from (\ref{max_exp_fast}) that \begin{align*} 1-e^{-kp^{n}} \leq \mathbb{P}\big( \max_{i=1, \ldots, k} X_i = n \big) \leq kp^{n}, \end{align*} which means that if $kp^{n}\rightarrow 0$, the sample maximum is no longer a consistent estimator of $n$. This occurs, for example, in the domain of attraction of the Poisson distribution, i.e., when $n\rightarrow\infty,\ p\rightarrow0$ and $np\rightarrow\mu\in(0,\infty)$ as $k\rightarrow\infty$ and $\log(k)\leq n$, since \begin{align*} kp^{n}&=\exp\{\log(k)+ n\log(p)\} \sim \exp\{\log(k)- n\log(n)\}\\ &\leq \exp\{\log(k)- \log(k)\log(\log(k))\} \rightarrow 0,\ \text{as } k\rightarrow\infty. \end{align*} In fact, when $np\rightarrow\mu$ both parameters become indistinguishable and this asymptotic scenario serves as a limiting benchmark for the $\operatorname{Bin}(n,p)$ problem to become solvable. However, in many applications the small $p$ regime (rare events) is the relevant one (see the references below and Section \ref{S:DataExample}), and this will be the topic of this paper. A variety of methods addressing this issue and improving over the sample maximum have been provided over the last decades but a final answer remains elusive until today. Broadly speaking, a major lesson from these attempts to obtain better estimators (see Section \ref{S:Frequentist} for a detailed discussion) seems that in this difficult regime further information on $n$ and $p$ is required to obtain estimators performing reasonably well. This asks for a Bayesian approach. An early Bayesian estimator of the binomial parameters $(N,P)$, now considered as random, dates back to \cite{DG}, who suggested the mode of the posterior distribution for a uniform prior on $\{1,\ldots,N_0\}$ for $N$ and a $\operatorname{Beta}(a,b)$ prior for $P$. Here, $N_0\in\mathbb{N}$ is fixed and the parameters $a,b>0$ are usually chosen as $a=b=1$, which yields the standard uniform distribution. Later \cite{Raftery}, \cite{Chilko}, \cite{Hamedani88} and \cite{Berger} provided further estimators, which mainly differ in their choices of loss functions and prior distributions for $N$ and $P$. A hierarchical Bayes approach is introduced in \cite{Raftery} with a Poisson prior on $N$ with mean $\mu$, which implies a Poisson distribution with parameter $\lambda=\mu p$ as the marginal distribution of each observation. The prior for the pair $(\lambda,P)$ is chosen proportional to $1/\lambda$, which is equivalent to a product prior for the pair $(N,P)$ with the prior for $N$ proportional to $1/n$ and the standard uniform prior for $P$. \cite{Raftery} suggested to minimize the Bayes risk with respect to the relative quadratic loss, which seems particularly suitable for estimating $n$ and will be adopted in this paper as well. From extensive simulation studies (see the afore mentioned references and Section \ref{S:SimulationStudy}), it is known that such Bayesian estimators of $n$ deliver numerically good results, in general. However, to the best of our knowledge, there is no rigorous theoretical underpinning of these findings. In particular, nothing is known about the posterior concentration of such estimators, and no systematic understanding of the role of the prior has been established. Our contribution to this topic is threefold: (i) we propose a new class of Bayesian estimators for $n$, generalizing the approach in \cite{Raftery}, and (ii) we prove the posterior contraction for $n$. The posterior contraction result holds for a wide class of priors for $n$ and does not depend on the choice of the loss function. It implies consistency in a general asymptotic setting of the introduced class of estimators as well as of many (and with small changes even all) Bayesian estimators mentioned above. Finally (iii), we extend the i.i.d.\ $\operatorname{Bin}(n,p)$ model to a regression setting and apply our Bayes approach to count the number of fluorophores from super-resolution images, which is considered a difficult task. \paragraph*{Ad (i).} For the new class of estimators, which we call the \emph{scale estimators}, we consider $k$ independent random variables $X_1,\ldots,X_k$ from a Bin$(N,P)$ distribution. Denote $\mathbf{X}^k := (X_1,\ldots,X_k)$ and $M_k:=\max_{i=1,\ldots,k}X_i$. We assume a product prior for the pair $(N,P)$, where the prior for $P$ is $\Pi_P\sim\operatorname{Beta}(a,b)$ for some $a,b>0$, and $\Pi_N$, the prior for $N$, satisfies $\Pi_N(n)\propto n^{-\gamma}$ for $\gamma>1$. Independence of $N$ and $P$ is a common assumption and also justified in our example (Section \ref{S:DataExample}) based on physical considerations. The scale estimator is then defined as the minimizer of the Bayes risk with respect to the relative quadratic loss, $ l(x,y)=( x/y-1)^2. $ Following \cite{Raftery}, it is given by \begin{equation}\label{E:ScaleBayes} \hat{n}:=\frac{\mathbb{E} \left[\frac{1}{N}|\mathbf{X}^k\right]}{\mathbb{E} \left[\frac{1}{N^2}|\mathbf{X}^k\right]}=\frac{\sum_{n=M_k}^\infty \frac{1}{n}L_{a,b}(n)\Pi_N(n)}{\sum_{n=M_k}^\infty \frac{1}{n^2}L_{a,b}(n)\Pi_N(n)}, \end{equation} where $L_{a,b}(n)$ is the beta-binomial likelihood, see, e.g., \cite{Carroll}. In existing literature \citep{Berger, Link}, the Bayesian estimator of $n$ with the prior $\Pi_N(n)\propto 1/n$ is often called the scale estimator. Even though we do not allow $\gamma=1$ in the above definition since it leads to an improper prior (see, however, Theorem \ref{improper_prior_mod} for a proper modification of these estimators for $0\leq\gamma\leq1$, which makes them accessible to our theory), we adopt this name for the new class of estimators. \paragraph*{Ad (ii).} We show posterior consistency in a quite general setting, where the prior distribution $\Pi_N$ can be chosen freely as long as it is a well-defined probability distribution satisfying \begin{equation}\label{a} \Pi_N(n) \geq \beta e^{-\alpha n^2},\ n\in\mathbb{N} \end{equation} for some positive constants $\alpha$ and $\beta$. In our asymptotic setting we consider sequences of parameters $(n_k, p_k)_k$ that may depend on the sample size $k$ and are described by the class \begin{align}\label{b} \mathcal{M}_{\lambda} : = \left\{ (n_k, p_k)_k : 1/\lambda \leq n_kp_k\leq\lambda, \ n_k \leq \lambda\sqrt[6]{k/\log(k)} \right\}. \end{align} for $\lambda>1$. We show that $$ \sup_{(n_k^0,p_k^0)\in\mathcal{M}_{\lambda}} \mathbb{E}_{n_k^0,p_k^0}\big[ \Pi\big( N \neq n_k^0\ |\, \mathbf{X}^k \big) \big] \rightarrow 0,\ \text{as } k\rightarrow\infty, $$ where $X_1,\dots,X_k\overset{i.i.d.}{\sim} \operatorname{Bin}(n_k^0,p_k^0)$. This is the main result of the paper and it will be formally stated as Theorem \ref{result} in Section \ref{S:PosteriorContraction}. The recent advances on posterior contraction focus mainly on nonparametric or semiparametric models \citep{Ghosal-2000, Ghosal-book} and posterior contraction for model selection in high-dimensional setups \citep{Castillo2012, Castillo2015, Gao}. Discrete models with complex structure have not yet been studied and it appears difficult to approach them by a general treatment. Our proof uses earlier work on maximum likelihood estimation by \cite{Hall} and opens another route to establish posterior consistency beyond the standard approach via testing, see \cite{Schwartz}. In the binomial model, posterior consistency for fixed parameters $n$ and $p$ with the priors above follows already by Doob's consistency theorem, see, e.g., \cite{Doob}. To the best of our knowledge, no refined asymptotic result for a Bayesian approach to estimate $n$ when $p$ is unknown exists. Our result shows consistency of the marginal posterior distribution of $N$ even in the challenging and relevant case of $n_k\rightarrow\infty$ and $p_k\rightarrow0$ as $k\rightarrow\infty$ as long as $(n_k,p_k)_k\in\mathcal{M}_{\lambda} $. The difficulty of this setup comes from the convergence of the binomial distribution to the Poisson distribution with parameter $\mu=\lim_{k\rightarrow\infty} n p$ as $n=n_k\rightarrow\infty$ and $p=p_k\rightarrow 0$. We have seen that the sample maximum is consistent as long as $kp^{n}\rightarrow\infty$, for which $e^n = o(k) $ is necessary (but not sufficient, see Lemma \ref{Consistency} for more details). In contrast, the definition of the class $\mathcal{M}_{\lambda}$ implies that $n^{6+\epsilon}=O(k)$ for $\epsilon>0$ is already sufficient for the posterior consistency of the suggested Bayes approach. We stress that a simulation study in \cite{SchneiderStaudt} suggests that the rate in Theorem \ref{result} cannot be relaxed significantly, as numerically posterior consistency is only observed up to $n^4=O(k)$. The posterior contraction result holds for the introduced scale estimators with $\Pi_N(n)\propto n^{-\gamma}$, $\gamma>1$. The improper priors with $0\leq\gamma\leq1$ satisfy the assumptions under slight modifications, which are described in Theorem \ref{improper_prior_mod} in Section \ref{S:PosteriorContraction}. With these modifications (restricting the support of $N$) the estimators of \cite{DG} and \cite{Raftery} are also covered by our theory. Our Theorems are applicable to many other Bayes estimators, as well. For example, Theorem \ref{result} holds for the estimator in \cite{Chilko}, where a Gamma prior for $N$ is suggested, and for the estimator in \cite{Hamedani88}, which suggests either a poisson prior on $N$ or an improper prior that can be considered via Theorem \ref{improper_prior_mod}. \paragraph*{Ad (iii).} Modern cell microscopy allows visualizing proteins and their modes of interaction during activity. It has become an indispensable tool for understanding biological function, transport and communication in the cell and its compartments, especially since the development of super-resolution nanoscopy (highlighted by the 2014 Nobel Prize in Chemistry). These techniques enable imaging of individual proteins through photon counts obtained from flourescent markers (fluorophores), which are tagged to the specific protein of interest and excited by a laser beam (see \cite{Hell2015} for a recent survey). In this paper, we are concerned with single marker switching (SMS) microscopy \citep{Betzig, Rust, Hess, Folling} where the emission of photons, which are then recorded, is inherently random: after laser excitation a fluorophore undergoes a complicated cycling through (typically unknown) quantum mechanical states on different time scales. This severely hinders a precise determination of the number of molecules at a certain spot in the specimen, see, e.g., \cite{Lee}, \cite{Rollins}, \cite{AEM2015}. In Section \ref{S:DataExample} we show how the number of fluorophores can be obtained from a modified $(n,p)$-Binomial model when they occur in clusters of similar size in the biological sample. A common difficulty in such experiments is that the number of active markers decreases over the measurement process due to bleaching effects. We show that the initial number $n^{(0)}$ can still be estimated from observations $X^{(t)}\sim \operatorname{Bin}(n^{(t)},p)$ at different time points $t$. We can link $n^{(0)}$ to $X^{(t)}$ by an exponential decay $n^{(t)}=n^{(0)}(1-B)^t$, which is known to be valid on physical grounds. This results now in a variant of the $(n,p)$-Binomial model, where the bleaching probability $B$ of a fluorophore can be estimated jointly with $n^{(0)}$ within this model. This allows us to determine the number of fluorophores $n^{(0)}$ on DNA origami test beds with high accuracy. This paper is organized as follows. Our main result on posterior contraction and the discussion on the asymptotics of other estimators for $n$ can be found in Section \ref{S:PosteriorContraction}. Section \ref{S:SimulationStudy} contains an extensive simulation study comparing the finite sample properties of those estimators and investigating robustness against model deviations from the Bin$(n,p)$ model relevant to our data example. In Section \ref{S:DataExample} the data example is presented. The proof of the posterior contraction and some auxiliary results about binomial random variables are stated in Section \ref{S:Proof}. Further auxiliary technicalities are deferred to the Appendix \ref{app}. \section{Posterior contraction for $\mathbf{n}$}\label{S:PosteriorContraction} Throughout the following $X_1,\dots,X_k$ are independent random variables with a $\operatorname{Bin}(N,P)$ distribution. We assume a product prior $\Pi_{(N,P)} = \Pi_N \Pi_P$ for the pair ($N$,$P$). For $P$ we choose a $\operatorname{Beta}(a,b)$ prior with parameters $a,b>0$. It is the conjugate prior suggested in \cite{DG} and widely used. The prior $\Pi_N$ for $N$ can be chosen as any proper probability distribution on the positive integers such that (\ref{a}) holds for some $\alpha, \beta>0$. Write $\mathbf{X}^k =(X_1, \ldots, X_k)$, $M_k = \max_{i=1, \ldots, k} X_i$ and $ S_k := \sum_{i=1}^k X_i$. For $A\subset [0,1]$ and $n\in\mathbb{N}$, the joint posterior distribution for $P$ and $N$ is then given by \begin{align*} \Pi\big( P \in A, N=n \, | \, \mathbf{X}^k \big)= \frac{\int_{A} t^{ S_k +a-1}(1-t)^{kn- S_k+b-1} \mathrm{d}t \cdot \prod_{i=1}^k \binom{n}{X_i}\cdot \Pi_N(n)}{\sum_{m=1}^{\infty} \int_0^1 t^{ S_k +a-1}(1-t)^{km- S_k+b-1} \mathrm{d}t \cdot \prod_{i=1}^k \binom{m}{X_i}\cdot \Pi_N(m)} \end{align*} if $n \geq M_k$ and $\Pi( P \in A, N=n \, | \, \mathbf{X}^k )=0$ otherwise. The marginal posterior likelihood function for $N$ is thus \begin{align*} \Pi\big( N= n \, | \, \mathbf{X}^k \big) \propto \prod_{i=1}^k \binom{n}{X_i} \frac{\Gamma(kn- S_k+b)\Gamma( S_k +a)}{\Gamma(kn+a+b)} \mathbf{1}( n \geq M_k ) \Pi_N(n)=: L_{a,b}(n) \Pi_N(n), \end{align*} where $\mathbf{1}(\cdot)$ denotes the indicator function and $L_{a,b}(\cdot)$ is the beta-binomial likelihood, see, e.g., \cite{Carroll}. The main result is stated in the following theorem and shows posterior contraction for $n$ in the asymptotic setting described by sequences of parameters $(n_k, p_k)_k\in\mathcal{M}_{\lambda}$ as defined in equation (\ref{b}). \begin{thm}\label{result} Conditionally on $N=n_k^0$ and $P=p_k^0$ let $X_1,\dots,X_k \overset{i.i.d.}{\sim} \operatorname{Bin}(n_k^0,p_k^0)$. For any prior distribution $\Pi_{(N,P)} = \Pi_N \Pi_P$ with $\Pi_P=\operatorname{Beta}(a,b)$, $a,b>0$, and where $\Pi_N$ is a probability distribution such that (\ref{a}) holds, we have uniform posterior contraction over $\mathcal{M}_{\lambda}$ in (\ref{b}) for $\lambda>1$, i.e., \begin{equation*}\label{E:Result_sup} \sup_{(n_k^0, p_k^0)_k \in \mathcal{M}_{\lambda}} \mathbb{E}_{n_k^0,p_k^0}\big[ \Pi\big( N \neq n_k^0\ |\, \mathbf{X}^k \big) \big] \rightarrow 0,\ \text{as } k\rightarrow\infty. \end{equation*} \end{thm} As mentioned in the introduction, from Theorem \ref{result} follows posterior consistency for the Bayesian estimators in equation (\ref{E:ScaleBayes}) and the ones in \cite{Hamedani88} and \cite{Chilko}, when considering parameter sequences in $\mathcal{M}_{\lambda}$. The estimator in \cite{DG} is based on a beta prior for $P$ and a uniform prior on $\{1,\ldots,N_0\}$ for some $N_0\in\mathbb{N}$ for $N$. Since $n_k > N_0$ cannot be excluded for $k$ large enough, assumption \eqref{a} is not fulfilled in this case. The estimators in \cite{Raftery}, \cite{Berger} and \cite{Link} violate the conditions of Theorem \ref{result} as well, since they are based on an improper prior on $N$ proportional to $1/n$. However, we can still extend our result to modifications of these estimators, where the support of $N$ is bounded but increases with $k$. \begin{thm}\label{improper_prior_mod} Theorem \ref{result} holds if we exchange $\Pi_N$ by $\Pi_{N,k}(n)\propto \frac{1}{n^{\gamma}} \mathbf{1}_{[1,T_k]}(n)$ with $\gamma\in[0,1]$, where $T_k$ satisfies \begin{equation}\label{E:Thm2} \lambda\sqrt[6]{k/\log(k)} \leq T_k <\begin{cases} \left( \exp\left\{\alpha k^{1/3}\right\}/\beta \right)^{\frac{1}{1-\gamma}}, \ &\gamma<1, \\ \exp\left\{\exp\left\{\alpha k^{1/3}\right\}/\beta\right\}, \ &\gamma=1, \end{cases} \end{equation} for all $k$ and some positive constants $\alpha$ and $\beta$. \end{thm} \begin{rem}\label{lambda_infinity} Theorem \ref{result} and Theorem \ref{improper_prior_mod} still hold true if we allow $\lambda$ in $\mathcal{M}_{\lambda}$ to increase with $k$, as long as $\lambda_k=o\left(\log(k)^{1/14}\right)$. This statement follows by verifying the conditions on the constants in the proof of Theorem \ref{result} and their dependence on $\lambda$. The strongest restriction results from equation (\ref{hm_integral}) and depends on Lemma \ref{Max_cases}. \end{rem} \subsection{Asymptotic results for frequentist methods: constrasted and compared}\label{S:Frequentist} In the following we present various existing asymptotic results for frequentist estimators and put them into perspective to Theorem \ref{result}, highlighting the differences of their respective asymptotic settings to the one described by the set $\mathcal{M}_{\lambda}$. Early estimators based on the method of moments and the maximum likelihood approach can be found in \cite{Haldane} and \cite{Blumen}. Their properties are further studied in \cite{OPZ}, where it is shown that the estimators for $n$, when both $n$ and $p$ are unknown and $p$ is small, are highly irregular and stabilized versions of the two estimators are proposed. Two estimators were introduced more recently in \cite{Gupta}: the first one is another modification of the method of moments estimator, and the second one is a bias correction of the sample maximum. The asymptotic behavior of these two estimators is also known. For the new moments estimator, $\hat{n}_{\mathrm{NME}}$, it holds that, as $k\to\infty$, $$\sqrt{k}(\hat{n}_{\mathrm{NME}}-n) \overset{\mathcal{D}}{\longrightarrow}\mathcal{N}(0,2\gamma^2n(n-1)),$$ where $n$ is fixed and $\gamma>0$ is a tuning parameter to be chosen by the practitioner. For the bias corrected sample maximum, $\hat{n}_{\mathrm{bias}}$ say, it holds for $n$ fixed, as $k\to\infty$: $$ (nk)^{1/(n-1)}(\hat{n}_{\mathrm{bias}}-n) \overset{\mathcal{D}}{\longrightarrow}\delta_1,$$ where $\delta_1$ denotes the Dirac measure at 1. In \cite{Carroll} a further modification of the maximum likelihood estimator is introduced. The estimator is the maximizer of the beta-binomial likelihood for $n$, where a beta density is assumed for $p$ and $p$ is integrated out. The Carroll-Lombard estimator is nearly equivalent to the Bayesian Draper-Guttman estimator (e.g., for $N_0$ large they produce the same estimates), since the Carroll-Lombard estimator can be understood as a maximum a posteriori (MAP) Bayesian estimator of $n$ with an improper uniform prior on $\mathbb{N}$. That means, if we bound the set of values where a maximum can be attained to $\{1,\ldots,T_k\}$, then Theorem \ref{improper_prior_mod} with $\gamma=0$ applies to the Caroll-Lombard estimator as well. This extends the classical asymptotic normality result of the Carroll-Lombard estimator $\hat{n}_{\mathrm{CL}}$, which holds for $p$ constant, $n\rightarrow\infty$ and $\sqrt{k}/n\rightarrow0$ as $k\rightarrow\infty$: \begin{align*} \sqrt{k} \left( \frac{\hat{n}_{\mathrm{CL}}- n}{n} \right) \overset{\mathcal{D}}{\longrightarrow} \mathcal{N}\big(0,2(1-p)^2/p^2\big). \end{align*} All of the results above hold for either $n$ or $p$ fixed and hence provide only limited insight into the situation when $p$ is small. A notable extension is discussed in \cite{Hall}. There, it is shown for $n=n_k\to\infty$ and $p=p_k\to0$ as $k\to\infty$ that \begin{align*} \frac{p \sqrt{k}}{2}\left( \frac{\hat{n}_{\mathrm{CL}}- n}{n} \right) \overset{\mathcal{D}}{\longrightarrow} \mathcal{N}(0,1), \end{align*} if $np\to\mu\in(0,\infty)$ and $kp^2\to\infty$ as $k\to\infty$. Note that this result, like the previous, studies the limiting distribution of the relative difference, where $\hat{n}-n$ is scaled by $n$. In contrast, we show posterior contraction to the exact value $\hat{n}=n$. This explains that \cite{Hall} can allow a faster rate ($n=o(k^2)$) than in our setting where $n = O\big( \sqrt[6]{k/\log(k)} \big)$. Also note that the above result is one specific scenario in a broader context and relies on further technical conditions, like $n$ to be lower bounded by some positive power of $k$. \section{Simulation study}\label{S:SimulationStudy} In this section we investigate the finite sample performance of Bayesian estimators numerically for different choices of priors $\Pi_P$ and $\Pi_N$. We compare the following estimators that we introduced in the previous sections. \begin{itemize} \item[(SE)] The scale estimator SE($\gamma$) with $\Pi_P = \mathrm{Beta}(a,b)$ and $\Pi_N(n)\propto n^{-\gamma}$. We consider both proper prior distributions ($\gamma>1$), and improper ones ($0\leq\gamma\leq 1$). The beta prior is chosen such that $P$ has expectation $\hat{p}$, where $\hat{p}\in(0,1]$ is a first guess for the probability of success, which might roughly be known beforehand. We select $a$ to be $1$ or $2$ and set $b=b(\hat{p}):=a/\hat{p}-a$. The scale factor $\gamma$ needs to be chosen. Note that the Raftery estimator is equivalent to the scale estimator with $\gamma=1$ and $a=b=1$. \item[(DGE)] The Draper-Guttman estimator DGE($N_0$). The parameters $a$ and $b$ of the beta distribution are selected in the same way as for the scale estimator. The upper bound $N_0$ should be selected sufficiently large to avoid underestimation. \end{itemize} We look at the SE with $\gamma\in\{0,0.5,1,2,3\}$ and the DGE with $N_0=500$. In case of an improper prior ($\gamma\leq1$), Theorem \ref{improper_prior_mod} applies, and the posterior distribution is well defined as long as $a+\gamma>1$ (see \citet{Kahn} for a cautionary note on this problem). We also employ the estimator SE$(0)$ with $a = 1$, for which the posterior does not exist (so it is no Bayes estimator), but which still produces finite estimates. \paragraph{General performance.} Our first simulation study is based on $1000$ samples of size $k\in\{30,100,300\}$ from a binomial distribution $\operatorname{Bin}(n_0,p_0)$ for $n_0\in\{20,50\}$ and $p_0\in\{0.05, 0.1, 0.3\}$. For all pairs $(n_0,p_0)$ and each estimator $\hat n$ we simulate: \begin{itemize} \item the relative mean squared error (RMSE), given by $\mathbb{E}\left[\left(\frac{\hat n}{n_0}-1 \right)^2\right]$, \item the bias $\mathbb{E}\left[\hat n\right] -n_0$ of the estimator. \end{itemize} We set $\hat{p}=p_0$ in the beta prior for this simulation and study the influence of the parameter $\gamma$. In Table \ref{Table:SummaryAlpha}, we present the estimators that have the lowest RMSE and the lowest bias for the different choices of $k$. The outcome advises to select smaller values of $\gamma$, the smaller $p_0$ is expected to be. Note that the DG estimator with large values $N_0$ is similar to the MAP estimator with the improper prior $\gamma=0$. Thus, it is not surprising that there is only little difference between the performance of DG(500) and SE(0) in the simulations. Both of them perform superior in the regime of very small $p_0$. Still, one should be aware that a small $\gamma$ increases the variance of the posterior and therefore of the estimates. For this reason, higher choices of $\gamma$ become preferable for low RMSEs as $k$ increases. The similarity of Table \ref{Table:SummaryAlpha} (A) and (B) for $n_0 = 20$ and $n_0 = 50$ suggests that the influence of $n_0$ is much weaker than the one of $p_0$ for the optimal estimator choice. \begin{table}[!htb] \begin{subtable}{.48\linewidth} \centering \caption{$n_0=20$} \begin{tabular}{rrll} \hline $p_0$&$k$& RMSE & bias\\ \hline \hline 0.05&30& DGE(500) & SE($0$)\\ 0.05&100& DGE(500) & SE($0$) \\ 0.05&300& SE($0.5$) & DGE(500) \\\hline 0.1&30& DGE(500) & SE($0$) \\ 0.1&100& SE($0.5$) & DGE(500) \\ 0.1&300& SE($1$) & DGE(500) \\\hline 0.3&30& SE($2$) & SE($1$) \\ 0.3&100& SE($3$) & DGE(500) \\ 0.3&300& SE($3$) & SE($2$) \\\hline \end{tabular} \end{subtable}% \quad \begin{subtable}{.48\linewidth} \centering \caption{$n_0=50$} \begin{tabular}{rrll} \hline $p_0$&$k$& RMSE & bias\\ \hline \hline 0.05&30& DGE(500) & SE($0$)\\ 0.05&100& DGE(500) & SE($0$)\\ 0.05&300& SE($0.5$)& DGE(500)\\\hline 0.1&30& DGE(500) &SE($0$)\\ 0.1&100& SE($0.5$) & DGE(500)\\ 0.1&300& SE($1$) & SE($0.5$) \\\hline 0.3&30& SE($1$) & SE($0.5$) \\ 0.3&100& SE($3$) & DGE(500) \\ 0.3&300& SE($3$) & DGE(500) \\\hline \end{tabular} \end{subtable} \caption{Overview of the estimators with the smallest RMSE and the smallest absolute bias for $a=2$ and $b=2/p_0-2$.}\label{Table:SummaryAlpha} \end{table} Our next numerical study covers a setting that is motivated by the data example in Section \ref{S:DataExample}, where $p_0\approx 0.0339$ and $k=94$. We therefore set $p_0=0.0339$ and $k=94$, and we select $n_0 = 15$. Our focus lies on the effect of the parameters $a$ and $b$, and particularly on the stability of the results with respect to misspecification of the guess $\hat{p}$. To this end, we consider four different scenarios: no information about $p_0$ (setting $\hat{p} = 0.5$), perfect information ($\hat{p} = p_0$), underestimation ($\hat{p} = 0.5\,p_0$), and overestimation ($\hat{p} = 1.5\,p_0$). The results in Table \ref{Table:SimsForExample} show that it is advantageous to choose a small $\gamma$ and a unimodal beta prior (i.e., $a=2$) if $p_0$ is known. If we have no information or are overestimating, it is again advisable to select $\gamma=0$, while choosing a less confident prior for $P$ with $a = 1$. In contrast, underestimation of $p_0$ leads to high instabilities and substantial overestimation of $n_0$ if $\gamma$ is small. Here, estimators with proper priors for $\gamma = 1$ and $2$ perform very well: the tendency for overestimation caused by the choice $\hat{p} = 0.5\,p_0$ is compensated by the tendency for underestimation in case of higher values of $\gamma$. \begin{table}[!htb] \begin{subtable}{.48\linewidth} \centering \begin{tabular}{cllll} \hline $\hat{p}$ & $a$ &estimator & RMSE & bias\\ \hline \hline $0.5$ & 1 & SE$(0.5)$ & 0.478 & -10.17 \\ & 1 & SE$(0)$& 0.395 & -9 \\ \hline $p_0$ & 2 & DGE$(500)$ & 0.034 & -0.266 \\ & 2 & SE$(0)$ & 0.036 & -0.043 \\ \hline \end{tabular} \end{subtable}% \quad \begin{subtable}{.48\linewidth} \centering \begin{tabular}{cllll} \hline $\hat{p}$ & $a$ & estimator & RMSE & bias\\ \hline \hline $1.5\,p_0$ & $1$ & SE$(0)$ & 0.12 & -3.73 \\ & $2$ & SE$(0)$ & 0.121 & -4.69 \\\hline $0.5\,p_0$ & $1$ & SE$(1)$& 0.036 & -0.032 \\ & $2$ & SE$(2)$& 0.025 & -0.55 \\\hline \end{tabular} \end{subtable} \caption{The two estimators that perform best under different choices of $\hat{p}$ for $n_0=15$, $p_0 = 0.0339$, and $k=94$. The respective values of $b$ are given by $b(\hat{p}) = a/\hat{p} - a$.}\label{Table:SimsForExample} \end{table} The general lesson seems to be that the smaller $p_0$, the more difficult it becomes to estimate $n_0$ and the smaller we want to choose $\gamma$. A smaller $\gamma$, however, increases the variance of the posterior distribution and leads to estimators that are more sensitive against misspecification of $\hat{p}$ in the beta prior. This is investigated in Table \ref{Table:SimsSensitivity}, where we compare the sensitivity of estimators corresponding to $\gamma=0$ and $\gamma=1$. We see that misspecifying $\hat{p} = 0.5\,p_0$ leads to severe overestimates $\mathbb{E}\left[\hat{n}\right] \approx 2\,n_0$ for DGE$(500)$, while SE$(1)$ is less sensitive. Selecting $\gamma=0$ can therefore help to estimate $n_0$ in very difficult scenarios, but it can also lead to heavily biased results if $\hat{p}$ is chosen too small. \begin{table}[!htb] \centering \begin{tabular}{l@{\qquad}cll} \hline estimator & $\hat{p}$ & RMSE & bias \\ \hline \hline SE$(1)$ & $p_0$ & 0.122 & -4.85 \\ & $0.5\,p_0$ & 0.129 & 4.43 \\ & $1.5\,p_0$ & 0.279 & -7.73 \\\hline DGE$(500)$ & $p_0$ & 0.034 & -0.27 \\ & $0.5\,p_0$ & 1.002 & 14.32 \\ & $1.5\,p_0$ & 0.139 & -5.09 \\\hline \end{tabular} \caption{Sensitivity of SE$(1)$ and DGE$(500)$ against misspecification of $\hat{p}$. The value $a$ is set to 2. All other parameters are selected like in Table \ref{Table:SimsForExample}. Note that the behavior of DGE$(500)$ and SE$(0)$ is comparable in this setting.}\label{Table:SimsSensitivity} \end{table} \paragraph*{Robustness.} Motivated by our data example in Section \ref{S:DataExample}, we also investigate the situation where $n$ may slightly vary within the sample. This appears to be relevant in many other situations as well, e.g., the (unknown) population size of a species may vary from experiment to experiment in the capture-recapture method. Whereas varying probabilities $p$ have been investigated in \cite{Basu}, models with a varying population size $n$ have not received any attention in the previous research. We consider $1000$ repetitions of size $k=100$, where each observation $X_i$, $i=1,\dots,k$, is generated from a $\operatorname{Bin}(n_i,p_0)$ distribution. Each $n_i$ is in turn a realization of a binomial random variable $N\sim\operatorname{Bin}(\tilde{n},\tilde{p})$. For each sample, $p_0$ is drawn from a Beta$(2,38)$ distribution with expectation $0.05$. To test the influence of the varying parameter $n_i$, we compare the performance of the estimators in the described scenario to their performance on binomial samples with a constant $n_0$ (chosen as the integer nearest to $\mathbb{E}[N] = \tilde{n}\tilde{p}$) and the same realizations $p_0$. We calculate the RMSE with respect to $n_0$ for both scenarios and present the RMSE for $X_i\sim\operatorname{Bin}(n_i,p)$ divided by the RMSE in the i.i.d.\ case. The ratios in Table \ref{Table:SimsForRobustness} verify a stable performance of the estimators in this setting since all values are close to $1$. The parameters in Table \ref{Table:SimsForRobustness} are chosen close to the data example in Section \ref{S:DataExample} with $\tilde{n}\in\{8,22\}$ and $\tilde{p}=0.7$, but further simulations (not shown) confirmed the stability for other parameter choices, like $\tilde{p}=0.5$ or $\tilde{p}=0.9$, as well. Hence, in summary, we find that for inhomogeneous (random) $N$ all estimators perform quite similar to the situation of a homogeneous (constant) $n_0$ ($\approx\mathbb{E}[N]$). \begin{table}[!htb] \centering \begin{tabular}{lll} & $\tilde{n}=8 $ & $\tilde{n}=22 $ \\ \hline estimator & RMSE-R & RMSE-R \\ \hline \hline SE(0.5) & 1.022 & 1.130 \\ SE(1) & 1.011 & 1.067 \\ SE(2) & 1.020 & 1.010 \\ DGE(500) & 1.032 & 1.073 \\ RE & 0.988 & 0.981 \\ \hline \end{tabular} \caption{Ratios of the RMSE for i.i.d.\ and non-i.i.d.\ samples (RMSE-R) for the estimators SE($\gamma$), DGE($N_0$), and the Raftery estimator RE. The beta prior in SE and DGE is defined by $a=2$ and $b=38$.}\label{Table:SimsForRobustness} \end{table} \section{Data example} \label{S:DataExample} In this section we extend the previously described Bayesian estimation methods to quantify the number of fluorescent molecules in a specimen recorded with super-resolution microscopy. Reliable methods to count such molecules are highly relevant to quantitative cell biology, for example, to determine the number of proteins of interest in a compartment of the cell, see, e.g., \citet{Lee}, \cite{Rollins}, \citet{TaH}, \citet{AEM2015} or \citet{Karathanasis} and references therein. \paragraph*{Experimental setup.} Data has been recorded at the Laser-Laboratorium G\"{o}ttingen e.V. During experimental preparation so called DNA origami \citep{Schmied}, tagged with the fluorescent marker Alexa647, were dispersed on a cover slip. DNA origami are nucleotide sequences which have been engineered in such a way that the origami folds itself into a desired shape (see Fig.\ \ref{fig:origami}A). Fluorescent molecules (fluorophores), which are equipped with an ``anchor'' that sticks to a specific region of the origami, are attached to the origami molecules. In the experiment, Alexa647 fluorophores with 22 different types of anchors were used, each one matching a different anchor position on the origami (see Fig.\ \ref{fig:origami}A). Therefore, at most 22 fluorophores can be attached to a single origami. The pairing itself is random (so not every possible anchor position needs to be occupied) and is expected to occur with a probability between 0.6 and 0.75, according to producer specifications Fundamental to super-resolution microscopy is the switching behavior of the fluorophores. A fluorophore can be in two different states (``on'' or ``off'') but only emits light in the ``on'' state. When excited with a laser beam, it switches between these ``on'' and ``off'' states until it bleaches, i.e., reaches an irreversible ``off'' state. During the course of the experiment, an image sequence of several origami distributed on a cover slip is recorded over a period of a few minutes (see the movie supplement material). The exposure time for one image (denoted as frame) is $15\,\mathrm{ms}$. Switching of fluorophores between ``on'' and ``off'' states is necessary to achieve super-resolution, which denotes the ability to discern markers with distance below the diffraction limit achievable with visible light of about $250$ -- $500\,\mathrm{nm}$ \citep{Hell}. Such fluorophores could not be discerned by conventional microscopy. Super-resolution becomes possible by separating photon emissions of spatially close molecules in time. This is realized by applying a low laser intensity, such that only a small fraction of fluorophores switches in the ``on'' state for a given frame. Hence, it is very unlikely that nearby fluorophores emit photons at the same time (see, e.g., \citet{Betzig}, \citet{Rust}, \citet{Hess}, \citet{Folling} for different variants of this principle). By this method, an increased resolution of up to $20$ -- $30\,\mathrm{nm}$ can be achieved. The experiment was prepared in such a way that most fluorophores are guaranteed to be ``on'' in the first frame, and all origami are thus visible as bright spots in Fig.\ \ref{fig:origami}B. Note that individual fluorophores occupying the same origami cannot be discerned in this frame. This becomes possible only when most of the fluorophores are switched ``off'' at later times, such that markers show up individually (see the supplementary movie for illustration). \begin{figure} \centering \begin{subfigure}[b]{0.49\linewidth} \centering\includegraphics[width=1.0\linewidth]{schematic_origami.pdf} \caption{} \end{subfigure} \begin{subfigure}[b]{0.49\linewidth} \centering\includegraphics[width=0.95\linewidth]{frame_1941_rois_trimmed.pdf} \caption{} \end{subfigure} \caption{(A) Schematic drawing of the DNA-origami used in the experiment. The origami is a tube-like structure that consists of 12 suitably folded DNA helices. In each of the two highlighted green regions up to 11 fluorescence markers can anchor. (B) First frame from the sequence of microscopic images. The 94 regions of interest (ROIs) that were chosen for analysis are identified by white boxes. The selection was done algorithmically. No overlapp between ROIs was allowed, and it was made sure that no excessive background noise and disturbances affected the ROI during the course of the experiment.} \label{fig:origami} \end{figure} \paragraph*{Quantitative biology.} Quantitative biology addresses the issue of counting the number of fluorophores from measurements like the one described above. The brightness of each spot is proportional to the number of fluorophores in the ``on'' state within the respective origami. An origami is invisible if all of its fluorophores are ``off'', but its location is still known from the first frame, which allows us to register 94 regions of interest (ROIs) in a preparational step (see Figure \ref{fig:origami}B). Exemplarily, six microscopic frames (out of $14{,}060$) recorded at different times $t\in\{1500, 3000, 4500, 6000, 7500, 9000\}$, which show the influence of switching and bleaching on the observations, are visualized in Figure \ref{fig:frames}. \begin{figure}[tb!] \centering \includegraphics[width=0.95\linewidth]{frames} \caption{Six selected frames from the dataset of recorded origami. The (physical) time difference between two consecutive images in this figure is roughly 22.5 seconds. Bleaching causes the number of visible origami to decrease with increasing frame number, and switching causes that unbleached origami are visible only in some frames.} \label{fig:frames} \end{figure} We aim to estimate the number of fluorophores attached to each origami, which we expect to be between $13$ and $16$ according to the producer specification. In order to make our model accessible to the data, we assume for simplicity that each origami carries the same number $n^{(0)}$ of fluorophores and we only model the mean number $n^{(t)}$ of unbleached fluorophores at time $t$. The physical relation between $n^{(0)}$ and $n^{(t)}$ is given by \begin{equation}\label{E:N0_Bleaching} n^{(t)} = n^{(0)}(1-B)^t, \end{equation} where $B$ denotes the bleaching probability. Now, the brightness observed for a spot in frame $t$ is proportional to the number $X^{(t)}$ of ``on'' fluorophores during the frame's exposure. This number $X^{(t)}$ is binomially distributed $\mathrm{Bin}(n^{(t)}, p)$, where $p$ denotes the (time-independent) probability that an unbleached fluorophore is in its ``on'' state. We can estimate $n^{(0)}$ and $B$ by fitting a log-linear model to equation \eqref{E:N0_Bleaching}, where the respective population sizes $n^{(t)}$ are in turn estimated from the 94 realizations of $X^{(t)}$ observed in frame $t$. To get a sense for the magnitude of $p$, we use data from a similar experiment: in this case, each origami has been designed in such a way that it carries exactly one fluorophore. We estimate $p$ as the average ratio $\hat{p}$ of the number of frames where the fluorophore is ``on'' (a bright spot is seen) and the total number of observed frames before bleaching (no spot is seen for any time in the future), and we find $\hat{p}\approx 0.0339$. This indicates that we are in the difficult ``small $p$'' regime of the $\mathrm{Bin}(n, p)$ problem, and we will therefore apply the Bayesian estimators introduced in Section \ref{S:SimulationStudy} (SE, DGE) to estimate $n^{(t)}$. The beta prior for SE and DGE uses the parameters $a = 2$ and $b = 2/\hat{p} - 2 \approx 56.99$. We choose the unimodal prior with $a=2$, as suggested by Table \ref{Table:SimsForExample}, since we assume that our guess $\hat{p}$ is reasonably accurate. Note that a finer degree of modeling would require to view $n^{(0)}$, $n^{(t)}$ and $p$ as random variables (with small variances) instead of constants. However, as shown in Section \ref{S:SimulationStudy}, the Bayesian estimators we consider are robust against fluctuations in the parameters and are therefore suited to estimate the respective mean values. Since most fluorophores are deliberately forced to be ``on'' in the first frame, the relation $X^{(t)} \sim \mathrm{Bin}(n^{(t)}, p)$ does not hold initially. It only becomes valid after the initial state has relaxed to an equilibrium, which is why we only take into account data after frame 1500 ($\approx 22.5$ seconds). To mitigate the influence of correlations between observations (i.e., $X^{(t)}$ and $X^{(t+1)}$ for a spot cannot be considered independent), we also add a waiting time of 1500 frames between the frames we use for our analysis. In total, we use the six frames at six time points with $t \in \{1500, 3000, 4500, 6000, 7500, 9000\}$ depicted in Figure \ref{fig:frames}. The 94 realizations of $X^{(t)}$ are extracted from the image data as follows: at each registered origami position, represented by a $6\times6$ pixel ROI, the total brightness is measured and then divided by the brightness of a single fluorophore. We determined the brightness of a single fluorophore from the late frames of the experiment, where at most one fluorophore of each origami is active with high probability. \begin{table}[!htb] \centering \begin{tabular}{lcc} \hline estimator & $n^{(0)}$ & $B\cdot 10^3$\\ \hline \hline SE$(0)$ & \textbf{16} & \textbf{0.152}\\ SE$(0.5)$ & \textbf{13} & \textbf{0.148}\\ SE$(1)$ & 11 & 0.139\\ SE$(2)$ & 9 & 0.163\\ SE$(3)$ & 6 & 0.123\\ SE$(5)$ & 5 & 0.114\\ DGE$(500)$ & \textbf{16} & \textbf{0.167}\\ \hline \end{tabular} \caption{Estimates of the bleaching probability $B$ and the number $n^{(0)}$ of fluorophore molecules on single DNA origami.}\label{Table:OrigamiEstimates-TimeFrame} \end{table} The results for the scale estimator with $\gamma = 0.5$ are depicted in Figure \ref{Fig:Fit}, which also shows the log-linear fit for model \eqref{E:N0_Bleaching}. This provides us with estimates for $n^{(0)}$ and $B$. The point estimates of $n^{(0)}$ and $B$ for different estimators are summarized in Table \ref{Table:OrigamiEstimates-TimeFrame}. Given that it is to be expected that the true $n^{(0)}$ in this experiment lies between 13 and 16, we can see in Table \ref{Table:OrigamiEstimates-TimeFrame} that the SEs with an improper prior ($\gamma\leq1$) produce reasonable results, and the DGE also performs well. This confirms that we are indeed in the critical case of Bin$(n,p)$ with small $p$, so that the prior putting a lot of weight on large values on $n$ gives best results by correcting for the usual tendency to underestimate, see also the results of the simulation study performed under comparable conditions in Table \ref{Table:SimsForExample}. To illustrate the difficulty of this problem, Figure \ref{Fig:BarChart} shows exemplary counting results we obtained for $t \in \{1500, 7500\}$. Note that the final estimates for $n^{(0)}$ are exclusively based on observations $X^{(t)}\leq 3$, where a great majority of these observations is already 0. \begin{figure}[t!] \centering \includegraphics[width=0.66\linewidth]{LogLinearFit}% \caption{The log-linear fit described by $n^{(t)} = n^{(0)}(1-B)^t$ for the SE with $\gamma = 0.5$.}\label{Fig:Fit} \end{figure} \begin{figure}[h] \centering \includegraphics[width=0.55\linewidth]{BarChart} \caption{Bar charts of the observed numbers of fluorophore molecules for time frames 1500 and 7500.}\label{Fig:BarChart} \end{figure} \section{Proofs}\label{S:Proof} \subsection{Proof of the main theorems} Here we present the proofs of our posterior contraction results for $n$ (Theorem \ref{result} and \ref{improper_prior_mod}). These require further technical results, e.g., fine moment estimates of a binomial random variable and bounds on the maximum of a triangular array of independent binomials, see Lemma \ref{moment_bound} - \ref{Consistency}. Auxiliary technicalities are postponed to the appendix. Throughout the proof of Theorem \ref{result} we will be concerned with an exemplary sequence in $\mathcal{M}_{\lambda}$. We call this sequence $(n_k, p_k)_k$ instead of $(n^0_k, p^0_k)_k$ for notational simplification. Our arguments do not depend on the specific choice of $(n_k, p_k)_k$ but only rely on the parameters $\lambda, a$ and $b$. \begin{proof}[Proof of Theorem \ref{result}] First observe that \begin{equation*} \Pi\big(N \neq n_k \,|\, X^k\big) = \frac{\sum_{n\neq n_k, n\ge M_k} L_{a,b}(n) \,\Pi_N(n)}{\sum_{n=M_k}^\infty L_{a,b}(n) \,\Pi_N(n)} \le \sum_{n\neq n_k, n \ge M_k} \frac{L_{a,b}(n)\,\Pi_N(n)}{L_{a,b}(n_k)\,\Pi_N(n_k)}. \end{equation*} Under the assumption that $S_k \ge 2$ (which we justify below), we can apply Lemma \ref{Monotonicity_a} and find \begin{equation*} \frac{L_{a,b}(n)}{L_{a,b}(n_k)} \le c_1 kn_k\,\frac{L_{\lfloor a\rfloor, b}(n)}{L_{\lfloor a\rfloor, b}(n_k)} \end{equation*} for $c_1 = 1 + \lceil a \rceil + b$, where $\lceil\cdot \rceil$ and $\lfloor\cdot\rfloor$ denote the ceil and floor functions, respectively. It follows that \begin{equation}\label{posterior-bound} \Pi\big(N\neq n_k\,|X^k\big) \le c_1 kn_k\,\sum_{n\neq n_k, n\ge M_k} \exp\left(k\int_{n_k}^{n} f'(m) \,\mathrm{d}m\right) \,\frac{\Pi_N(n)}{\Pi_N(n_k)} \end{equation} with $f(m) = \frac{1}{k}\log L_{\lfloor a\rfloor, b}(m)$. If $n < n_k$, we can write $\int_{n_k}^{n} f'(m)\,\mathrm{d}m = -\int_{n}^{n_k} f'(m)\,\mathrm{d}m$. For an upper bound on the posterior we thus need a lower bound of $f'(m)$ if $m\le n_k$ and an upper bound if $m \ge n_k$. Since $f$ only depends on $a$ via $\lfloor a \rfloor$, we assume that $a \in\mathbb{N}_0$ in the following. Then we can apply Lemma 4.1 from \cite{Hall} and find \begin{equation}\label{fprime} f'(m)= \frac{1}{k}\sum_{i=1}^k \sum_{j=1}^{X_i} \frac 1{m-j+1} - \sum_{j=1}^{ S_k +a} \frac 1{km+a+b-j} = \sum_{j=1}^{M_k}\frac{ T_j-U_j}{j} -\sum_{j=M_k+1}^{ S_k+a}\frac{U_j}{j} \end{equation} with \begin{equation*} T_j := \frac{1}{k}\sum_{i=1}^k \frac{(X_i)_j}{(m)_j}\qquad \text{and}\qquad U_j := \frac{( S_k +a)_j}{(km+a+b-1)_j} \end{equation*} for $j \le M_k$ and $j \le S_k + a$ respectively, where $(t)_j = t\,(t-1)\cdots(t-j+1)$ denotes the falling factorial for $t > 0$. For convenience we define $T_j := 0$ for all $j > M_k$. Next, we introduce the events \begin{align*} \mathcal{U}_k &:= \Big\{ M_k = n_k \text{~~or~~} M_k \ge l_k\Big\}, \qquad \mathcal{R}_k := \Big\{ M_k \le 2\log(k) \Big\}, \\[0.1cm] \mathcal{T}_{kj} &:= \Big\{ (m)_j\big|T_j - \mathbb{E}T_j\big| \le \sqrt{(c_2j)^j\,l_k\log(k)/k}\Big\}, \qquad \mathcal{T}_k := \bigcap_{j\in\mathbb{N}} \mathcal{T}_{kj}, \\[0.1cm] \mathcal{S}_k &:= \Big\{ \big|S_k - kn_kp_k\big| \le \sqrt{\lambda\,k\log(k)}\Big\}, \end{align*} and denote the intersection $\,\mathcal{U}_k\, \cap\, \mathcal{R}_k\, \cap\, \mathcal{T}_k\, \cap\, \mathcal{S}_k\,$ by $\mathcal{A}_k$. The constant $c_2 = 2\,\lambda\,(\lambda + 2)$ is chosen to satisfy Lemma \ref{Tj-concentration} for each sequence $(n_k, p_k)_k \in\mathcal{M}_{\lambda}$, and $l_k$ is a fixed sequence with $l_k=o(\sqrt{\log(k)})$. Note that the sets $\mathcal{T}_{kj}$ are in fact independent of $m$ due to the definition of $T_j$. On the event $\mathcal{S}_k$, Lemma \ref{Uj-concentration} grants us the additional property \begin{equation*} \big|U_j - \widetilde{U}_j\big| \le j \,\sqrt{\frac{\lambda \log(k)}{k}} \left(\frac{c_3}{m}\right)^j \qquad\text{with}\qquad\widetilde{U}_j := \frac{(kn_kp_k + a)_j}{(km + a + b - 1)_j} \end{equation*} for $j \le S_k + a$ and $c_3 = 2e^2(3\lambda + a + 1)$. Also note that $S_k \le 2k\lambda$ holds and that $S_k \ge 2$ is guaranteed for $k/\lambda - \sqrt{\lambda \,k\log(k)} \ge 2$ on $\mathcal{S}_k$. Thus, equations \eqref{posterior-bound} and \eqref{fprime} apply on $\mathcal{A}_k$ if $k$ is sufficiently large. Indeed, we can restrict our attention to $\mathcal{A}_k$, since \begin{equation}\label{restriction} \mathbb{E}_{n_k, p_k}\big[\Pi\big(N\neq n_k\,|\,X^k\big)\big] - \mathbb{E}_{n_k, p_k}\big[\mathbf{1}_{\mathcal{A}_k}\Pi\big(N\neq n_k\,|\,X^k\big)\big] \le \mathbb{P}_{n_k, p_k}\big(\mathcal{A}_k^\mathrm{c}\big) \longrightarrow 0 \end{equation} for $k \to \infty$. To see this, one can bound $\mathbb{P}_k\big(\mathcal{A}_k^\mathrm{c}\big) := \mathbb{P}_{n_k, p_k}\big(\mathcal{A}_k^\mathrm{c}\big)$ by \begin{equation*} \mathbb{P}_k\big(\mathcal{A}_k^\mathrm{c}\big) \le \mathbb{P}_k\big(\mathcal{S}_k^\mathrm{c}\big) + \mathbb{P}_k\big(\mathcal{U}_k^\mathrm{c}\big) + 2\,\mathbb{P}_k\big(\mathcal{R}_k^\mathrm{c}\big) + \mathbb{P}_k\big(\mathcal{T}_k^\mathrm{c} \cap \mathcal{R}_k\big). \end{equation*} The first contribution vanishes due to Chebyshev's inequality (see e.g., \cite{degroot2012}), and the second and third terms are controlled by Lemma \ref{Max_cases} and Lemma \ref{Max} respectively. The last contribution satisfies \begin{equation*} \mathbb{P}_k\big(\mathcal{T}_k^\mathrm{c}\cap\mathcal{R}_k\big) = \mathbb{P}_k\left(\bigcup\nolimits_{j=1}^{\lfloor 2\log(k)\rfloor} \mathcal{T}_{kj}^\mathrm{c}\right) \le \frac{\lfloor 2\log(k)\rfloor}{l_k\log(k)} \longrightarrow 0 \end{equation*} for $k \to \infty$ due to Lemma \ref{Tj-concentration}. It is important to note that the upper bounds in these considerations only depend on $\lambda$ if suitable choices for the involved constants are made. In the following, we always assume that $X^k\in\mathcal{A}_k$. \paragraph{Auxiliary lower bound.} For $M_k \le m < n_k$, we prove a lower bound of $f'(m)$. We may assume that $M_k \ge l_k \to \infty$ for $k\to\infty$ in this case, since $X^k\in\mathcal{U}_k$. For $k$ such that $l_k \ge 4$ we can bound equation \eqref{fprime} by \begin{equation}\label{lower_bound_f} f'(m) \ge \sum_{j=1}^4 \frac{T_j - U_j}{j} - \sum_{5}^{S_k + a} \frac{U_j}{j}, \end{equation} as $T_j \ge 0$ for all $j$. In case of $j = 1$ we obtain \begin{equation*} T_1 - U_1 = \frac{S_k}{km} - \frac{S_k + a}{km + a + b - 1} \ge - \frac{a+1}{km-1} \ge - 2\,\frac{(a+1)}{km} \ge - 2\,\frac{\lambda(a+1)}{m^2}\sqrt{\frac{\log(k)}{k}}, \end{equation*} where we used the upper bound $m < n_k \le \lambda\sqrt{k\log(k)}$ in the last inequality, which is guaranteed in $\mathcal{M}_{\lambda}$. To handle the terms with $j \ge 2$, we exploit that $X^k\in\mathcal{T}_k$ and apply $(m)_j \ge \big(m/e^2\big)^j$ from Lemma \ref{lem.falling_factorial} in order to derive \begin{equation*} \sum_{j=2}^4 \frac{\big|T_j - \mathbb{E}T_j\big|}{j} \le \sqrt{\frac{l_k\log(k)}{k}}~ \sum_{j=2}^4 \left(\frac{\sqrt{c_2j}}{m/e^2}\right)^j \le 2\,\frac{4\,c_2\,e^4}{m^2}~\sqrt{\frac{l_k\log(k)}{k}}, \end{equation*} for $k$ large enough such that $\sqrt{4\,c_2}\,e^2 / l_k < 1/2$. Similarly, we find \begin{equation*} \sum_{j=2}^{S_k+a} \frac{\big|U_j - \widetilde{U}_j\big|}{j} \le \sqrt{\frac{\lambda\log(k)}{k}}~ \sum_{j=2}^{S_k + a} \left(\frac{c_3}{m}\right)^j \le 2\,\frac{\sqrt{\lambda}\,c_3^2}{m^2}~\sqrt{\frac{\log(k)}{k}} \end{equation*} and \begin{equation*} \sum_{j=5}^{S_k+a} \frac{\widetilde{U}_j}{j} \le \sum_{j=5}^{S_k+a} \frac{1}{j} \left(\frac{c_4}{m}\right)^j \le 2\,\left(\frac{c_4}{m}\right)^5 \end{equation*} for $k$ (and thus $m$) sufficiently large. In the latter equation, we applied the first part of Lemma \ref{Utilde-bound} with $c_4 = 6e^2(\lambda + a)$, using $S_k \le 2\,k\lambda$ on $\mathcal{S}_k$. Next, we note that $\mathbb{E}(X_1)_j = (n_k)_j\, p^j$ and consult the first result of Lemma \ref{Utilde-Tj-bound} to establish \begin{equation*} \sum_{j=2}^4 \frac{\mathbb{E}T_j - \widetilde{U}_j}{j} \ge \frac{1}{2\lambda^2} \frac{n_k - m}{n_km^3} - \frac{3\,c_5^4}{k} \ge \frac{1}{2\lambda^2} \frac{n_k - m}{n_km^3} - 2\,\frac{2\,\lambda^2c_5^4}{m^2} \sqrt{\frac{\log(k)}{k}}, \end{equation*} where $c_5 = 3\,\lambda + 2\,a + 2$. Similar to the case $j = 1$, we applied $m^2 < \lambda^2\sqrt{k\log(k)}$ in the last inequality. All bounds calculated above can be inserted into inequality \eqref{lower_bound_f}, yielding \begin{align} f'(m) &\ge (T_1 - U_1) + \sum_{j=2}^4 \frac{T_j - \mathbb{E}T_j}{j} + \sum_{j=2}^{S_k+a} \frac{\widetilde{U}_j - U_j}{j} + \sum_{j=2}^4 \frac{\mathbb{E}T_j - \widetilde{U}_j}{j} - \sum_{j=5}^{S_k+a} \frac{\widetilde{U}_j}{j} \nonumber \\ &\ge \frac{1}{4\lambda^2} \frac{n_k - m}{n_km^3} - \frac{C_2}{m^2}\,\sqrt{\frac{l_k\log(k)}{k}} + \left[\frac{1}{4\lambda^2}\frac{n_k-m}{n_km^3} - 2\,\left(\frac{c_4}{m}\right)^5\right] \nonumber \\[0.2cm] &\ge \frac{C_1}{m^3}\frac{n_k - m}{n_k} - \frac{C_2}{m^2}\,\sqrt{\frac{l_k\log(k)}{k}} + h(m), \label{auxiliary_lower_bound} \end{align} where $h(m)$ is defined as the expression in square brackets. The constants in this bound are $C_1 = 1/4\lambda^2$ and $C_2 = 2\,\big(\lambda(a+1) + 4\,c_2\,e^4 + \sqrt{\lambda}\,c_3^2 + 2\,\lambda^2c_5^4\big)$. \paragraph{Auxiliary upper bound.} We next provide an upper bound for $f'(m)$ if $m > n_k \ge M_k$. Unlike for the lower bound, we cannot assume that $m$ becomes larger than any given constant with increasing $k$. Since $U_j$ is nonnegative, we can bound \begin{equation*} f'(m) \le \sum_{j=1}^{M_k} \frac{T_j - U_j}{j}. \end{equation*} in equation \eqref{fprime}. We look at the case $j = 1$ first, and see \begin{equation*} T_1 - U_1 = \frac{S_k}{km} - \frac{S_k + a}{km + a + b - 1} \le \frac{S_k\,(a + b)}{km\,(km - 1)} \le \frac{4\lambda\,(a + b)}{km^2} \le \frac{4\lambda\,(a+b)}{m^2} \sqrt{\frac{\log^3(k)}{k}}, \end{equation*} where we used that $S_k \le 2\lambda\,k$ on the event $\mathcal{S}_k$. Next we set $\widetilde{m} := 4\,c_2\,e^4$ and derive \begin{align*} \sum_{j=2}^{M_k} \frac{\big|T_j - \mathbb{E}T_j\big|}{j} &\le \sqrt{\frac{l_k\log(k)}{k}}~ \sum_{j=2}^{M_k} \left(\frac{\sqrt{c_2j}}{m/e^2}\right)^j \\ &\le \frac{c_2M_k\,e^4}{m^2}~\sqrt{\frac{l_k\log(k)}{k}}~ \sum_{j=0}^{\lfloor m\rfloor} \left(\frac{e^2\sqrt{c_2}}{\sqrt{m}}\right)^j \\ &\le \frac{c_2M_k\,e^4}{m^2}~\sqrt{\frac{l_k\log(k)}{k}} \cdot \begin{cases} 2 &\text{if}~ m > \widetilde{m} \\ \widetilde{m} \,(e^2\sqrt{c_2} + 1)^{\widetilde{m}} &\text{if}~ m \le \widetilde{m} \end{cases} \\ &\le \frac{c_6}{m^2}~\sqrt{\frac{l_k\log^3(k)}{k}} \end{align*} for $c_6 = 2\,c_2\,e^4\big(\widetilde{m}\,(e^2\sqrt{c_2} + 1)^{\widetilde{m}} + 2\big)$. In the last step, we used that $M_k \le 2\log(k)$ on the event $\mathcal{R}_k$. In a similar fashion, we can establish the bound \begin{equation*} \sum_{j=2}^{M_k} \frac{\big|U_j - \widetilde{U}_j\big|}{j} \le \sqrt{\frac{\lambda\log(k)}{k}}~ \sum_{j=2}^{M_k} \left(\frac{c_3}{m}\right)^j \le \frac{c_7}{m^2}~\sqrt{\frac{\log^3(k)}{k}}, \end{equation*} where $c_7 = 4\,\sqrt{\lambda}\,c_3^2\,\big(c_3^{2c_3+1} + 1\big)$. Finally, we apply the second claim of Lemma \ref{Utilde-Tj-bound} and obtain \begin{equation*} \sum_{j=2}^{M_k} \frac{\mathbb{E}T_j - \widetilde{U}_j}{j} \le - C_1'\,\frac{m - n_k}{n_km^3} \end{equation*} with $C_1' = \big(2\lambda^2\left(a+b+1\right)^2 \big)^{-1}$ for sufficiently large $k$. We conclude \begin{align} f'(m) &\le (T_1 - U_1) + \sum_{j=2}^{M_k} \frac{T_j - \mathbb{E}T_j}{j} + \sum_{j=2}^{M_k} \frac{\widetilde{U}_j - U_j}{j} + \sum_{j=2}^{M_k}\frac{\mathbb{E}T_j - \widetilde{U}_j}{j} \nonumber \\ &\le -\frac{C_1'}{m^3}\frac{m - n_k}{n_k} + \frac{C_2'}{m^2}~\sqrt{\frac{l_k\,\log^3(k)}{k}} \label{auxiliary_upper_bound} \end{align} for $C_2' = 4\lambda\,(a+b) + c_6 + c_7$. \paragraph{Posterior bound.} By applying the two inequalities \eqref{auxiliary_lower_bound} and \eqref{auxiliary_upper_bound} for $m < n_k$ and $m > n_k$, we can now bound the posterior probability $\Pi\big(N\neq n_k\,|\,X^k\big)$ on the event $\mathcal{A}_k$ through equation \eqref{posterior-bound}. First, we observe that for $n\in\mathbb{N}$ with $n\neq n_k$ \begin{equation*} \int_{n_k}^{n} \frac{m - n_k}{n_km^3}\,\mathrm{d}m = \frac{1}{2} \frac{(n-n_k)^2}{(n_kn)^2} \qquad\text{and}\qquad\int_{n_k}^{n} \frac{1}{m^2}\,\mathrm{d}m = \frac{1}{2}\frac{(n-n_k)^2}{(n_kn)^2} \frac{2n_kn}{n - n_k}. \end{equation*} It also holds for $n\neq n_k$ that \begin{equation}\label{n_over_nk_ineq} \left|\frac{n}{n_k} - 1\right| \ge \frac{1}{2n}\,. \end{equation} Therefore, if $l_k \le n < n_k$, the function $h(m)$ introduced in equation \eqref{auxiliary_lower_bound} satisfies \begin{align}\label{hm_integral} \int_n^{n_k}h(m) \,\mathrm{d}m &= \frac{C_1}{2}\,\frac{(n-n_k)^2}{(n_kn)^2} - \frac{c_4^5}{2}\,\frac{n_k^4 - n^4}{(n_kn)^4} \nonumber \\ &\ge \frac{C_1}{2} \,\frac{(n-n_k)^2}{(n_kn)^2}\left(1 - \frac{4c_4^5}{C_1}\frac{1}{1-n/n_k}\frac{1}{n^2}\right) \ge 0 \end{align} for $k$ such that $l_k \ge 8\,c_4^5/C_1$. Employing bound \eqref{auxiliary_lower_bound} thus yields \begin{align*} -k \int_n^{n_k} f'(m)\,\mathrm{d}m &\le -k \frac{C_1}{2} \frac{(n-n_k)^2}{(n_kn)^2} \left( 1 - C \frac{n_kn}{n_k-n}\sqrt{\frac{l_k\log(k)}{k}}\right), \intertext{ where the constant $C$ is given by $2\,C_2/C_1$. On the other hand, for $n_k < n$, bound \eqref{auxiliary_upper_bound} similarly leads to } k \int_{n_k}^n f'(m)\,\mathrm{d}m &\le -k \frac{C_1'}{2} \frac{(n-n_k)^2}{(n_kn)^2} \left( 1 - C' \frac{n_kn}{n-n_k}\sqrt{\frac{l_k\log^3(k)}{k}}\right) \end{align*} for $C' = 2\,C_2'/C_1'$. Finally, let $\widetilde{C}_1 = \min\big\{C_1, C_1'\big\}$ and $\widetilde{C} = \max\big\{C, C'\big\}$. We can apply inequality \eqref{n_over_nk_ineq} (with $n$ and $n_k$ switched) to find for any $n\in\mathbb{N}$ with $n\neq n_k$ and $n \ge M_k$ that \begin{align*} k\int_{n_k}^n f'(m)\,\mathrm{d}m &\le -k \frac{\widetilde{C}_1}{2\,n_k^2}\left(\frac{n_k}{n} - 1\right)^2 \left( 1 - \frac{\widetilde{C}\,n_k}{|1 - n_k/n|} \sqrt{\frac{l_k\log^3(k)}{k}} \right) \\ &\le -k \frac{\widetilde{C}_1}{8\,n_k^4} \left( 1 - 2\,\widetilde{C}\,n_k^2 \sqrt{\frac{l_k\log^3(k)}{k}} \right) \le -\frac{\widetilde{C}_1}{16} \frac{k}{n_k^4} \end{align*} for $k$ large enough such that $n_k^2 \le \sqrt{k/\big(l_k\log^3(k)\big)} / 4\widetilde{C}$ for each sequence in $\mathcal{M}_{\lambda}$. Consulting inequality \eqref{posterior-bound} and using the constraint $\Pi_N(n_k) \ge \beta\exp\big(-\alpha n_k^2\big)$ of the prior yields \begin{align}\label{bounding_prior} \mathbf{1}_{\mathcal{A}_k}\Pi\big(N\neq n_k\,|X^k\big) &\le c_1kn_k \sum_{n\neq n_k} \exp\left(-\frac{\widetilde{C}_1}{16}\frac{k}{n_k^4}\right) \frac{\Pi_N(n)}{\Pi_N(n_k)} \\ &\le \frac{c_1}{\beta}\exp\left(-\frac{\widetilde{C}_1}{16}\frac{k}{n_k^4} + \alpha n_k^2 + \log\big(kn_k\big)\right) \nonumber \\ &\le \frac{c_1}{\beta}\exp\left(-\frac{\widetilde{C}_1}{16\,\lambda^4}\, k^{1/3} \log(k)^{2/3} + \alpha \lambda^2\,\frac{k^{1/3}}{\log(k)^{1/3}} + \log\big(\lambda k^2\big)\right) \longrightarrow 0 \nonumber \end{align} for $k\to \infty$. Due to statement \eqref{restriction} this is sufficient to prove Theorem \ref{result}. \end{proof} \begin{proof}[Proof of Theorem \ref{improper_prior_mod}] The result follows from the proof of Theorem \ref{result}, where we only need to handle the inequalities in (\ref{bounding_prior}) differently. Bounding the sum over the prior in (\ref{bounding_prior}) by an integral, one sees that the upper bound on $T_k$ in (\ref{E:Thm2}) is sufficient to ensure convergence towards $0$, if $\Pi_{N,k}$ is considered instead of $\Pi_N$. \end{proof} \subsection{Auxiliary results for binomial random variables}\label{S:aux_results} \begin{lem}\label{moment_bound} Let $X$ be a binomial random variable, $X\sim\operatorname{Bin}(n,p)$, for $n\in\mathbb{N}$ and $p\in(0, 1)$. Then \begin{equation*} \mathbb{E}[X^r]\leq B_r \cdot \max\{np, (np)^{r}\}, \end{equation*} where $B_r$ is the $r$-th Bell number. \end{lem} \begin{proof} Let $q=(1-p)$ and let $M_{n,p}$ be the moment generating function of the binomial distribution, \begin{equation*} M_{n,p}(t) = (pe^t+q)^n = f(g(t)), \end{equation*} where $f(s)=s^n$ and $g(t)=pe^t+q$. To obtain the moments of $X$, we look at the derivatives of $M_{n,p}$ at $t=0$. The $r$-th derivatives of $f$ and $g$ are \begin{align*} f^{(r)}(s)=(n)_r\, s^{n-r}\quad \text{ and }\quad g^{(r)}(t)=p e^{t} \end{align*} for $r\in\mathbb{N}$. Since $g(0)=1$, it holds that $f^{(r)}\big(g(0)\big)=(n)_r$. Furthermore, $g^{(r)}(0)=p$ for all $r$. We employ the Bell polynomial version of Fa\`a di Bruno's formula, see \cite{Bruno} equation (2.2), which is \begin{align}\label{faa} (f\circ g)^{(r)}(t)= \sum_{k=1}^r f^{(k)}\big(g(t)\big)\, B_{r,k}\left(g^{(1)}(t),g^{(2)}(t),\dots, g^{(r-k+1)}(t)\right). \end{align} The Bell polynomials $B_{r,k}$ are homogeneous of degree $k$. Therefore, \begin{align*} \mathbb{E}[X^r]= (f\circ g)^{(r)}(0)&= \sum_{k=1}^{r} f^{(k)}(g(0)) \,B_{r,k}\,\left(g^{(1)}(0),\dots, g^{(r-k+1)}(0)\right)\\ &=\sum_{k=1}^r B_{r,k}(1, \dots, 1)\,(n)_kp^k\\ &\leq B_r \cdot \max\{np, (np)^{r}\}, \end{align*} where $B_r = \sum_{k=1}^r B_{r,k}$ is the $r$-th Bell number. \end{proof} \begin{lem}\label{Tj-concentration} Let $n\in\mathbb{N}$ and $p\in(0, 1)$. Define $k$ i.i.d.\ binomial random variables $X_1,\dots,X_k$ with distribution $\operatorname{Bin}(n,p)$. For each $j\in\mathbb{N}$ with $j \le n$, the inequality \begin{equation*} \mathbb{P}\left( \bigg| \frac{1}{k}\sum_{i=1}^k(X_i)_j - \mathbb{E}[(X_1)_j] \bigg| > \sqrt{\frac{l\,(cj)^j}{k}}\right) \le \frac{1}{l} \end{equation*} holds for any $l > 0$ and $c \ge 2\,np\,(np + 2)$. \end{lem} \begin{proof} We define the random variable $\widetilde{X} \sim \operatorname{Bin}(n - j, p)$ and note that $\mathbb{E}[(X_i)_j] = (n)_jp^k$. Invoking Lemma \ref{moment_bound}, we derive the upper bound \begin{align*} \mathrm{Var}[(X_i)_j] &\le \mathbb{E}[(X_i)^2_j] \\ &\le (n)_jp^j \, \mathbb{E}[(\widetilde{X} + j)_j] \\ &\le (np)^j \mathbb{E}[(\widetilde{X} + j)^j] \\ &\le 2^j(np)^j \big(\mathbb{E}[\widetilde{X}^j] + j^j\big) \\ &\le 2^j(np)^j \big(B_j\, (np+1)^j + j^j\big) \\ &\le \big(2j\,np\,(np+2)\big)^j \end{align*} on the variance of $(X_i)_j$. The second inequality becomes transparent from expanding the expectation as a sum, and the last inequality is valid due to the relation $B_j \le j^j$ that can be found in \cite{bell-bound}. For $c \ge 2\,np\,(np + 2)$, we obtain by Chebyshev's inequality that \begin{equation*} \mathbb{P}\left( \bigg| \frac{1}{k}\sum_{i=1}^k(X_i)_j - \mathbb{E}[(X_1)_j] \bigg| > \sqrt{\frac{l\,(cj)^j}{k}}\right) \le \frac{\mathrm{Var}[(X_1)_j]/k}{l\,(cj)^j/k} \le \frac{1}{l}. \end{equation*} \end{proof} \begin{lem}\label{Max_cases} Let $n\in\mathbb{N}$ and $p\in(0,1)$, and let $M_k$ denote the maximum of $k$ independent binomial variables $X_1, \dots, X_k \sim\operatorname{Bin}(n,p)$. Let $(l_k)_{k\in\mathbb{N}}$ be such that $l_k\rightarrow\infty$ and $l_k = o( \sqrt{\log(k)} )$. Then, for each $k$ with $l_k > \max\{1, 4\,np\}$, \begin{equation*} \mathbb{P}\big(M_k < \min\{l_k, n\}\big) \le e^{-d_k}, \end{equation*} where \begin{equation*} d_k = \min\left\{\frac{k}{e^{l_k\log(l_k/np)}}, \frac{k\,np}{8\pi\,l^2_k e^{l_k^2/np}}\right\} \to +\infty \quad \text{as} \ k \to \infty. \end{equation*} \end{lem} \begin{proof} We have that \begin{equation*} \mathbb{P}\big(M_k < \min\{l_k, n\}\big) = \begin{cases} \mathbb{P}(M_k < n)\quad &\text{if}~n \le l_k, \\ \mathbb{P}(M_k < l_k)\quad &\text{if}~n > l_k. \end{cases} \end{equation*} In case of $l_k \ge n$, we derive the upper bound \begin{equation*} \log\mathbb{P}(M_k < n) \le -kp^n \le -k\,e^{- l_k\log(l_k/np)} \longrightarrow -\infty,\ \text{as } k\rightarrow\infty, \end{equation*} by applying Bernoulli's inequality and for $l_k = o( \sqrt{\log(k)} )$. If $n > l_k$, we find that $p \le 1/4$, and thus Slud's bound from \cite{Tele} can be applied to yield \begin{align*} \mathbb{P}(M_k < l_k) &= \big(1 - \mathbb{P}(X_1 \ge l_k) \big)^k \\ &\le \Phi\left(\frac{l_k}{\sqrt{np(1-p)}}\right)^k \le \Phi\left(\frac{\sqrt{2}\,l_k}{\sqrt{np}\,}\right)^k, \end{align*} where $\Phi$ is the cumulative standard normal distribution function. By the lower tail bound in \cite{Gordon1941}, which states that $1 - \Phi(t) \ge \frac{1}{2\pi} \frac{t}{t^2+1}\,e^{-t^2/2}$ for $t > 0$, we obtain \begin{align*} \Phi\left(\frac{\sqrt{2}\,l_k}{\sqrt{np}\,}\right)^k &\le \left(1 - \frac{np}{8\pi\,l_k^2}\,e^{-\frac{l_k^2}{np}}\right)^k \\ &\le \exp\left(-\frac{k}{8\pi}\,\frac{np}{l_k^2\,e^{l_k^2/np}}\right)\longrightarrow 0,\ \text{as } k\rightarrow\infty, \end{align*} where we set $t = \sqrt{2} \,l_k / \sqrt{np} > 1$ (by assumption) and used $t/(t^2+1) \ge 1/2t^2$ for $t \ge 1$ and $l_k = o\left( \sqrt{\log(k)} \right)$. \end{proof} \begin{lem}\label{Max} Let $n\in \mathbb{N}$ and $p\in(0, 1)$. Define k i.i.d.\ binomial random variables $X_1, \ldots, X_k$ with distribution $\operatorname{Bin}(n, p)$, and let $M_k := \max_{i=1,\dots,k}X_i$. Then \begin{equation*} \mathbb{P}\big(M_k\leq 2 \log(k)\big) \ge \left(1 - \frac{1}{k^2}\right)^k \end{equation*} if $k \ge e^{3np}$. Consequently, $\mathbb{P}\big(M_k > 2 \log(k)\big) \to 0$, as $k\to \infty$. \end{lem} \begin{proof} We can write $X_1$ as a sum of $n$ i.i.d.\ Bernoulli random variables bounded by $1$. By Bernstein's inequality (see e.g., \cite{vdvaart1996}) \begin{align*} \mathbb{P}\left( M_k \leq 2\log(k) \right) &= \big( 1-\mathbb{P}\big( X_1-np > 2\log(k) - np \big)\big)^k\\ &\geq \left(1 - \exp\left\{-\frac{ \big(2\log(k) - np\big)^2}{2\big(np(1-p) + \log(k)/3\big)} \right\}\right)^k\\ &\geq \big(1 - e^{-2\log(k)}\big)^k, \end{align*} where the last inequality holds for $\log(k) \ge 3np$. \end{proof} \begin{lem}\label{Consistency} Let $(n_k,p_k)_{k\in\mathbb{N}}$ be a sequences with $n_k\in\mathbb{N}$, $p_k\in(0, 1)$ and $n_kp_k\to \mu > 0$. Define the independent random variables $X_1,\dots,X_k \sim\operatorname{Bin}(n_k,p_k)$ and let $M_k:= \underset{i=1,\dots,k}{\max}X_i$. \begin{tabular}{ll} (i) & \quad If $n_k\log(n_k)<c\log(k)$ for $c < 1$, then $\mathbb{P}(M_k=n_k)\rightarrow 1, \ \text{as } k\rightarrow\infty.$ \\ (ii) & \quad If $n_k\log(n_k)>c\log(k)$ for $c > 1$, then $\mathbb{P}(M_k=n_k)\rightarrow 0, \ \text{as } k\rightarrow\infty.$ \end{tabular} \end{lem} \begin{proof} \textit{(i):} We have convergence of the sample maximum towards the parameter $n_k$ if $$ \mathbb{P}(M_k=n_k)= 1-\left(1-p_k^{n_k}\right)^k\geq 1-e^{-kp_k^{n_k}} \rightarrow 1, \ \text{as } k\rightarrow\infty, $$ where we applied Bernoulli's inequality. This holds if $\log(k) - n_k\log(n_k/n_kp_k) \to \infty$, which follows from \begin{equation*} \frac{n_k\log(n_k)}{\log(k)} < c < 1\qquad \text{and}\qquad \frac{n_k\,|\log(n_kp_k)|}{\log(k)} \le \frac{c\,|\log(n_kp_k) |}{~\log\big(c\log(k)\big)}. \end{equation*} \textit{(ii):} It holds that $P(M_k = n_k) \le kp_k^{n_k} \le \exp\big(\log(k) - n_k\log(n_k/n_kp_k)\big)$. Similar to the argument above, the right hand side in this inequality converges to $0$ since \begin{equation*} \frac{n_k\log(n_k)}{\log(k)} > c > 1\qquad \text{and}\qquad \frac{n_k\,|\log(n_kp_k)|}{\log(k)} \le \frac{c\,|\log(n_kp_k) |}{~\log\big(c\log(k)\big)}. \end{equation*} \end{proof}

120df5edff434247fe8b934c7aa0560a7d88e482

\section*{Introduction}\label{sec:intro} A \textit{path} is a map $X:[0,1]\to\mathbb{R}^d$. Classically, the components $X_1,\dots,X_d$ of $X$ are assumed to be sufficiently smooth. For every positive integer $k$, it is therefore possible to define an order $k$ tensor $\sigma^{(k)}(X)\in(\mathbb{R}^d)^{\otimes k}$, whose $(i_1\ldots i_k)$-th entry is \[\int_{0}^{1}\int_{0}^{t_k}\dots\int_{0}^{t_3}\int_{0}^{t_2}\dot{X}_{i_1}(t_1)\cdot\ldots\cdot\dot{X}_{i_k}(t_k)dt_1\dots dt_k. \] By convention, we define $\sigma^{(0)}(X)=1$. The sequence \[\sigma(X)=(\sigma^{(k)}(X)\mid k\in\mathbb{N})\] is called the \textit{signature} of $X$. Signatures were first defined in \cite{Chen}, and they enjoy many interesting properties. For instance, the signature allows to uniquely recover a sufficiently smooth path up to a mild equivalence relation (see \cite[Theorem 4.1]{chenuniqueness}). Many physical behaviors and experiments can be modeled by using paths, and signatures are useful tools to encode in a compact form the information carried by paths. In \cite{AFS}, the authors consider signature tensors from an algebraic geometry perspective. If we fix a certain class of paths and the order $k$ of the tensors, then the $k$-th signature $\sigma^{(k)}$ is a polynomial map into $(\mathbb{R}^d)^{\otimes k}$. Its image variety parametrizes the closure of the set of all $k$-th signatures of paths of the chosen class. The study of this map, this \textit{signature variety} and its geometric properties is interesting for many reasons. For instance, in applied problems one sometimes has a signature, coming from empirical data, and wants to know if there is a path of a certain kind (say, piecewise linear) having that given signature. Knowing whether the map is injective, or at least finitely many to one, tells us if there are chances to solve this inverse problem. Another issue is the study of singularities. What does it mean, for a path, to have a signature which is a singular point in the variety? In \cite{AFS} we find a detailed study of the signature varieties of polynomial paths, piecewise linear paths and also random paths arising from Brownian motion. These three classes of paths have a common generalization: the class of rough paths. Rough paths have a number of applications, for instance the study of controlled ODEs and stochastic PDEs (see \cite{Friz}), as well as sound compression (see \cite{soundcompression}). While they are not necessarily smooth, it is possible to define their signature, and therefore to study their signature variety. Even at a first glance, such a variety exhibits analogies with the Veronese variety, and it is therefore named the \textit{Rough Veronese variety} in \cite[Section 5.4]{AFS}. The main purpose of this paper is to study its geometry. We will prove that the Rough Veronese variety is toric, and we will characterize the monomials parameterizing it. Sometimes it is useful to perform explicit computations in order to handle concrete examples. In these cases, we use the software Macaulay2, freely available at {\tt{www.math.uiuc.edu/Macaulay2}}. \section*{Notations and preliminaries}\label{sec:notation} The $k$-th signature of a path $X$ belongs to $(\mathbb{R}^d)^{\otimes k}$, but we need a space to store the whole signature $\sigma(X)$. In this Section we define such a space, which has a rich algebraic structure. We will recall the features we need, but we do not attempt to any extent to describe all its properties. Every definition and result of this section can be found in \cite{bible}. \begin{dfn} The \textit{tensor algebra} over $\mathbb{R}^d$ is the graded $\mathbb{R}$-vector space \[T((\mathbb{R}^d))= \mathbb{R}\times\mathbb{R}^d\times(\mathbb{R}^d)^{\otimes 2}\times\ldots\] of formal power series in the non-commuting variables $x_1,\dots,x_d$. It is an $\mathbb{R}$-algebra with respect to the tensor product, and we denote by $p_k:T((\mathbb{R}^d))\to(\mathbb{R}^d)^{\otimes k}$ the projection. The algebraic dual of $T((\mathbb{R}^d))$ is the graded algebra \[T(\mathbb{R}^d)=\mathbb{R}\langle x_1,\dots,x_d\rangle\] of polynomials in the non-commuting variables $x_1,\dots,x_d$. It is the free $\mathbb{R}$-algebra over $x_1,\dots,x_d$. \end{dfn} \begin{notaz} Given an element $T\in T((\mathbb{R}^d))$, we denote by $T_{i_1\dots i_k}$ the $(i_1\dots i_k)$-th entry of the order $k$ element of $T$. For $y\in\mathbb{R}$, we denote $T_y((\mathbb{R}^d))=\{T\in T((\mathbb{R}^d))\mid T_0=y\}$. It will often be convenient to identify a degree $k$ monomial $x_{i_1}\cdot\ldots\cdot x_{i_k}$ with the word $w={\bf i_1}\ldots {\bf i_k}$ in the alphabet $\{{\bf 1},\dots,{\bf d}\}$. The number $k$ is called the \textit{length} of $w$ and it is denoted by $|w|$. The degree 0 monomial corresponds to the empty word ${\bf e}$. We will write letters in bold in order to distinguish the number 1 from the letter ${\bf 1}=x_1$. In this way, the product of two words $v$ and $w$ is simply the word obtained by writing $v$ followed by $w$. It is called the \textit{concatenation product} and sometimes denoted by $v\cdot w$. The natural duality pairing \[\langle-,-\rangle:T((\mathbb{R}^d))\times T(\mathbb{R}^d)\to\mathbb{R} \] is given by $\langle T,{\bf i_1}\ldots {\bf i_k}\rangle=T_{i_1\dots i_k}$, extended by linearity. \end{notaz} Besides the concatenation of words, there is another product on $T(\mathbb{R}^d)$. It will play a very important role in this paper. \begin{dfn} The \textit{shuffle product} of two words $v$ and $w$, denoted by $v\shuffle w$, is the sum of all order-preserving interleavings of them. More precisely, we define the shuffle recursively. If ${\bf i}$ and ${\bf j}$ are letters, then \begin{align*} &{\bf e} \shuffle w = w \shuffle {\bf e} = w \text{ and }\\ &(v\cdot {\bf i}) \shuffle (w\cdot {\bf j}) = \left(v\shuffle (w\cdot {\bf j})\right)\cdot {\bf i} + \left((v\cdot {\bf i} )\shuffle w\right)\cdot {\bf j}. \end{align*} As an example, \[{\bf 12}\shuffle{\bf 34}={\bf 1234}+{\bf 1324}+{\bf 1342}+{\bf 3124}+{\bf 3142}+{\bf 3412}. \] Again, the shuffle product is extended by linearity to $T(\mathbb{R}^d)$. We will sometimes use the notation \[v^{\shuffle n}=\underbrace{v\shuffle\ldots\shuffle v}_{n\mbox{\begin{Small} times\end{Small} }}.\] \end{dfn} Despite its apparently complicated definition, the shuffle product enjoys good properties. For instance, the space $(T(\mathbb{R}^d),\shuffle,{\bf e})$ is a commutative algebra. Moreover, shuffle behaves nicely with respect to the signatures. In order to be precise, we need to specify the class of paths we deal with. Intuitively, we want to enlarge the class of smooth paths to comprehend those paths for which the iterated integrals are well-defined. \begin{dfn} A path $X:[0,1]\to\mathbb{R}^d$ is \textit{regular} if $\dot{X}:[0,1]\to\mathbb{R}^d$ is well defined, continuous and non-vanishing on $[0,1]$. It is \textit{piecewise regular} if there are $0=a_0<a_1<\ldots<a_{r-1}<a_r=1$ such that $X_{|[a_{i-1},a_i]}$ is regular on $[a_{i-1},a_i]$ for every $i\in\{1,\dots,r\}$. In particular, a piecewise regular path is piecewise continuous and its finitely many discontinuity points are jumps. \end{dfn} The class of piecewise regular path is the one considered by Chen in \cite{chenuniqueness}. These paths have a nice interplay with the shuffle product. \begin{lem}[Shuffle identity]\label{lem:shuffle identity} If $X:[0,1]\to\mathbb{R}^d$ is piecewise regular, then \[\langle \sigma(X),v\rangle\cdot\langle \sigma(X),w\rangle=\langle \sigma(X),v\shuffle w\rangle \] for all words $v,w\in T(\mathbb{R}^d)$. \end{lem} The shuffle identity is proved in \cite[Proof of Corollary 3.5]{bible} for a more general class of paths. Hence signatures do not fill the whole tensor algebra $T((\mathbb{R}^d))$, but rather they live in the subset of elements with constant term 1 and satisfying the shuffle identity. This is one of the many possible motivations for the next definition. \begin{dfn} We will denote \[\mathcal{G}(\mathbb{R}^d)=\{T\in T_1((\mathbb{R}^d))\mid \langle T,v\rangle\cdot\langle T,w\rangle=\langle T,v\shuffle w\rangle \mbox{ for all words } v,w\in T(\mathbb{R}^d)\}.\] \end{dfn} The object we have just defined is worth a few remarks. It not only contains the signatures of all piecewise regular paths, but it is also a group with respect to the tensor product. This is why its elements are sometimes called \textit{group-like} elements in the literature. $\mathcal{G}(\mathbb{R}^d)$ is not linear, but it is closely related to a linear space. \begin{dfn} On $T((\mathbb{R}^d))$ there is a Lie bracketing $[T,S]=TS-ST$. Then we can define $\Lie(\mathbb{R}^d)$ to be the free Lie algebra generated by $x_1,\dots,x_d$, that is, the smallest linear subspace of $T((\mathbb{R}^d))$ that contains $x_1,\dots,x_d$ and is closed with respect to the bracketing. \end{dfn} This Lie algebra and $\mathcal{G}(\mathbb{R}^d)$ are linked by two maps. \begin{dfn} Define $\exp:T_0((\mathbb{R}^d))\to T_1((\mathbb{R}^d))$ by the formal power series \[\exp(T)=\sum_{n=0}^{\infty}\frac{T^{\otimes n}}{n!}.\] Not surprisingly, $\exp$ has a two-sided inverse $\log:T_1((\mathbb{R}^d))\to T_0((\mathbb{R}^d))$ defined by \[\log(S)=\sum_{n=1}^\infty\frac{(-1)^{n+1}}{n}(S-1)^{\otimes n}.\] \end{dfn} For our purposes, we need to point out that all the definitions we have recalled have a truncated version. Namely, one can fix $m\in\mathbb{N}$ and consider \[T^m(\mathbb{R}^d)=\bigoplus_{k=0}^m(\mathbb{R}^d)^{\otimes k},\] where tensors of order greater than $m$ are set to zero. Inside $T^m(\mathbb{R}^d)$ there are $\mathcal{G}^m(\mathbb{R}^d)$ and $\Lie^m(\mathbb{R}^d)$. The maps $\exp$ and $\log$ are defined in the same way and they restrict to a bijection between $\Lie^m(\mathbb{R}^d)$ and $\mathcal{G}^m(\mathbb{R}^d)$. In order to avoid confusion, we will write $\exp^{(m)}$ to denote the map $T_0^m((\mathbb{R}^d))\to T_1^m((\mathbb{R}^d))$. We need a last definition before we move to our Rough Veronese variety. \begin{dfn} A non-empty word $w$ is a \textit{Lyndon word} if, whenever we write $w=pq$ as the concatenation of two nonempty words, we have $w<q$ in the lexicographic order. We denote by $W_{d,m}$ the set of Lyndon words of length at most $m$ in the alphabet $\{{\bf 1},\dots,{\bf d}\}$. For every word $w$, there exists a unique pair $(p,q)$ of nonempty words such that $w=pq$ and $q$ is minimal with respect to lexicographic order. The \textit{bracketing} of $w$ is $[p,q]=pq-qp$. \end{dfn} We care about Lyndon words because $\Lie^m(\mathbb{R}^d)$ has a basis consisting of all bracketings of Lyndon words of length at most $m$. Recall that the M\"{o}bius function $\mu:\mathbb{N}\to\mathbb{N}$ sends a natural number $t$ to \[\mu(t)=\begin{cases} 0 & \mbox{ if $t$ is divisible by the square of a prime},\\ 1 & \mbox{ if $t$ is the product of an even number of distinct primes},\\ -1 & \mbox{ if $t$ is the product of an odd number of distinct primes}.\\ \end{cases} \] Then the number of length $l$ Lyndon words in the alphabet $\{{\bf 1},\dots,{\bf d}\}$, denoted by $\mu_{l,d}$, is \[\mu_{l,d}=\sum_{t\mid l}\frac{\mu(t)}{l}d^{\frac{l}{t}}\] and therefore, as a vector space, $\Lie^m(\mathbb{R}^d)$ has dimension \[\dim\Lie^m(\mathbb{R}^d)=\sum_{l=1}^m \sum_{t\mid l}\frac{\mu(t)}{l}d^{\frac{l}{t}}.\] \section*{Signatures of rough paths}\label{sec:enter roughveronese} In stochastic analysis, people often deal with paths that are far from being smooth, not even piecewise regular. Therefore we want to further expand our class of paths. Given their capital importance in stochastics, it is natural to consider rough paths. In this section we recall the basic definitions and we introduce their signature varieties. The main reference for rough paths is \cite{Multidim}. Consider a piecewise regular path $X$ and let $t\in [0,1]$. In the definition of $k$-th signature we can replace the integral on $[0,1]$ with the integral on $[0,t]$. This is the same as restricting $X$ to the sub-interval $[0,t]$, hence we will denote it as $\sigma^{(k)}(X_{|[0,t]})$. As an example, \[\sigma^{(1)}(X_{|[0,t]})_i=\int_{0}^{t}\dot{X_i}(\lambda)d\lambda=X_i(t)-X_i(0).\] For every $k$, we notice that $\sigma^{(k)}(X_{|[0,t]})$, as a function of $t$, is a path $[0,1]\to(\mathbb{R}^d)^{\otimes k}$. If we look at the full signature $\sigma(X_{|[0,t]})$, we get a path $[0,1]\to\mathcal{G}(\mathbb{R}^d)$ such that its endpoint is $\sigma(X)$. Moreover, this $\mathcal{G}(\mathbb{R}^d)$-valued path satisfies a H\"{o}lder-like inequality. We will use the symbol $f(t)\apprle g(t)$ to indicate that there is a constant $c$ such that $f(t)\le c\cdot g(t)$ for every $t$. \begin{lem} Let $X:[0,1]\to\mathbb{R}^d$ be a piecewise regular path and let $k\in\mathbb{N}$. If $s,t\in [0,1]$, then \begin{equation}\label{eq:holder} \left| \sigma^{(k)}(X_{|[s,t]})\right| \apprle |t-s|^k. \end{equation} \begin{proof} Let $S=\sigma^{(k)}(X_{|[s,t]})$. Since $k$ is fixed, in order to conclude it is enough to bound every entry of $S$. By definition \begin{align} |S_{i_1\ldots i_k}|&=\left| \int_{s}^{t}\int_{s}^{t_k}\dots\int_{s}^{t_3}\int_{s}^{t_2}\dot{X}_{i_1}(t_1)\cdot\ldots\cdot\dot{X}_{i_k}(t_k)dt_1\dots dt_k\right| \nonumber\\ &\le\int_{s}^{t}\int_{s}^{t_k}\dots\int_{s}^{t_3}\int_{s}^{t_2}\left| \dot{X}_{i_1}(t_1)\cdot\ldots\cdot\dot{X}_{i_k}(t_k)\right| dt_1\dots dt_k\nonumber\\ &\le\sup_{t_1\in[0,1]}| \dot{X}_{i_1}(t_1)| \cdot\ldots\cdot\sup_{t_k\in[0,1]}| \dot{X}_{i_k}(t_k)|\int_{s}^{t}\int_{s}^{t_k}\dots\int_{s}^{t_3}\int_{s}^{t_2} dt_1\dots dt_k\nonumber\\ &=\sup_{t_1\in[0,1]}| \dot{X}_{i_1}(t_1)| \cdot\ldots\cdot\sup_{t_k\in[0,1]}| \dot{X}_{i_k}(t_k)|\cdot\frac{|t-s|^k}{k!}.\nonumber \end{align} Since $X$ is piecewise regular, all the suprema are finite. \end{proof} \end{lem} So we see that a piecewise regular path $X:[0,1]\to\mathbb{R}^d$ induces a path $\sigma(X_{|[0,\textendash]}):[0,1]\to\mathcal{G}(\mathbb{R}^d)$ satisfying inequality (\ref{eq:holder}). If we want a rough path to be a generalization of a piecewise regular path, we can define it in a similar flavor, also allowing different exponents. Recall that $p_k:T((\mathbb{R}^d))\to(\mathbb{R}^d)^{\otimes k}$ is the projection. \begin{dfn}\label{dfn:roughpaths} A \textit{rough path} of order $m$ is a path ${\bf X}:[0,1]\to\mathcal{G}^m(\mathbb{R}^d)$ such that $|p_k({\bf X}(s)^{-1}\otimes {\bf X}(t))|\apprle |t-s|^\frac{k}{m}$ for every $k\in\{1,\dots,m\}$ and every $s,t\in [0,1]$. The inverse is taken in the group $\mathcal{G}^m(\mathbb{R}^d)$. \end{dfn} In the literature, there are other classes of paths that are called rough, see for instance \cite[Definition 9.15]{Multidim} and \cite[Section 2.6]{Friz}. However, in this paper we only consider rough paths as in Definition \ref{dfn:roughpaths}, sometimes called \textit{weakly geometric H\"older} $m$-rough paths. Moreover, following \cite[Section 5.4]{AFS}, we will focus on a special subclass of rough paths of order $m$, indexed by elements of $\Lie^m(\mathbb{R}^d)$. \begin{dfn} For $L\in\Lie^m(\mathbb{R}^d)$, consider the path ${\bf X}_L:[0,1]\to\mathcal{G}^m(\mathbb{R}^d)$ sending $t$ to $\exp^{(m)}(tL)$. By \cite[Exercise 9.17]{Multidim}, this is indeed an order $m$ rough path. In analogy with the piecewise regular case, we define the \textit{signature} of ${\bf X}_L$ to be its endpoint $\sigma({\bf X}_L)=\exp(L)\in\mathcal{G}(\mathbb{R}^d)$. \end{dfn} We want to study the set parameterizing the $k$-th signatures of ${\bf X}_L$, when $L$ ranges over $\Lie^m(\mathbb{R}^d)$. Such set is the image of $p_k\circ\exp$. A priori, this is just a semialgebraic subset of $(\mathbb{R}^d)^{\otimes k}$, that is, it is described by a finite number of polynomial equations and inequalities. Semialgebraic sets are usually hard to handle. In order to make our analysis simpler, we will follow a common approach in applied algebraic geometry and take the Zariski closure of this set, which means that we only look at the equations without considering the inequalities. Furthermore, from a geometric viewpoint it is convenient to work over an algebraically closed field, so we look at the variety that these equations define in $(\mathbb{C}^d)^{\otimes k}$, instead of $(\mathbb{R}^d)^{\otimes k}$. Finally, we want to work up to scalar multiples, so we pass to the projectivization and hence we deal with a projective variety. \begin{dfn} Let $f_{d,k,m}$ be the complexification of the map \[p_k\circ\exp:\Lie^m(\mathbb{R}^d)\rightarrow(\mathbb{R}^{d})^{\otimes k}.\] The \textit{Rough Veronese variety} $\mathcal{R}_{d,k,m}\subset\mathbb{P}^{d^k-1}$ is the projectivization of the closure of the image of $f_{d,k,m}$. \end{dfn} First of all observe that, being the image of a map, $\mathcal{R}_{d,k,m}$ is irreducible. There are several reasons to compare $\mathcal{R}_{d,k,m}$ to a Veronese variety. Since $\Lie^1(\mathbb{R}^d)=\mathbb{R}^d$, an element $L\in\Lie^1(\mathbb{R}^d)$ is just a vector and therefore, up to a multiplicative constant, $p_k(\exp (L))=L^{\otimes k}$ can be viewed as the Veronese embedding of $\mathbb{R}^d$ into $\mathop{{\rm Sym}}\nolimits^k\mathbb{R}^d\subset (\mathbb{R}^{d})^{\otimes k}$. In other words, $\mathcal{R}_{d,k,1}=\mathcal{V}_{d-1,k}$. Moreover, $\mathcal{V}_{d,k}$ is toric, parametrized by all degree $k$ monomials in $d+1$ variables. We will see that $\mathcal{R}_{d,k,m}$ is toric as well, and it is defined by monomials of weighted degree $k$, for a suitable choice of weights imposed by the structure of $\Lie^m(\mathbb{R}^d)$. Unlike $\mathcal{V}_{d,k}$, however, in general $\mathcal{R}_{d,k,m}$ fails to be smooth. The inclusions $\Lie^i(\mathbb{R}^d)\subset\Lie^{i+1}(\mathbb{R}^d)$ show that the Rough Veronese varieties are neste . On the other hand, this chain stabilizes. Indeed, when we apply $p_k$ and project onto the order $k$ summand, we do not see anything of order greater than $k$. So $\mathcal{R}_{d,k,k+i}=\mathcal{R}_{d,k,k}$ for every $i\in\mathbb{N}$. Being the image of $\mathcal{G}^k(\mathbb{R}^d)$ under $p_k$, $\mathcal{R}_{d,k,k}$ contains all the $k$-th signatures of piecewise regular paths. For this reason, in \cite[Section 4.3]{AFS} it is called \textit{universal variety} and denoted by $\mathcal{U}_{d,k}$. Summarizing, there is a chain of strict inclusions \[\mathcal{V}_{d-1,k}=\mathcal{R}_{d,k,1}\subset\mathcal{R}_{d,k,2}\subset\ldots\subset\mathcal{R}_{d,k,k}=\mathcal{R}_{d,k,k+1}=\ldots=\mathcal{U}_{d,k}.\] It is not restrictive to assume $d\ge 2$ and $m\le k$. The first thing we want to do is to determine $\dim\mathcal{R}_{d,k,m}$. \begin{pro} If $m\le k$, then the general point of $\mathcal{R}_{d,k,m}$ has $k$ preimages under $f_{d,k,m}$. In particular, $\dim\mathcal{R}_{d,k,m}=\dim(\mathbb{P}(\Lie^m(\mathbb{R}^d)))$. \end{pro} This was already noted in \cite[Remark 6.5]{AFS}. We now want to understand the geometry of $\mathcal{R}_{d,k,m}$, and we start by looking at the simplest example. \begin{es}\label{ex:R222} Consider $d=k=m=2$. We want to write down $f_{2,2,2}$. The Lyndon words of length at most 2 in the alphabet $\{{\bf 1},{\bf 2}\}$ are {\bf 1}, {\bf 2} and {\bf 12}, hence $\dim\Lie^2(\mathbb{R}^2)=3$. An element of $\Lie^2(\mathbb{R}^2)$ can therefore be written as $x_1{\bf 1}+x_2{\bf 2}+a({\bf 12}-{\bf 21})$. If we look at it in tensor terms, we see a vector $x=(x_1,x_2)$ as order 1 summand, and a $2\times 2$ matrix $A$ as order 2 summand. Being a multiple of $({\bf 12}-{\bf 21})$, $A$ is skew-symmetric. Then \begin{align} p_2(\exp(x+A)) &=p_2\left(1+x+A+\frac{(x+A)^2}{2}+\ldots\right)\nonumber\\ &=p_2\left(1+x+A+\frac{x^2+xA+Ax+A^2}{2}+\ldots\right)\nonumber\\ &=A+\frac{x^2}{2}=\left(\begin{array}{cc} 0 & a\\ -a & 0\end{array}\right)+\frac{1}{2}\left(\begin{array}{cc} x_1^2 & x_1x_2\\ x_1x_2 & x_2^2\end{array}\right).\nonumber \end{align} Then we can write $f_{2,2,2}:\mathbb{C}^3\rightarrow\mathbb{C}^{2\times 2}$ as \[(x_1,x_2,a)\mapsto\left(\begin{array}{cc} \frac{x_1^2}{2} & \frac{x_1x_2}{2}+a\\ \frac{x_1x_2}{2}-a & \frac{x_2^2}{2}\end{array}\right).\] Up to a linear change of coordinates, it becomes \[(x_1,x_2,a)\mapsto\left(\begin{array}{cc} x_1^2 & x_1x_2\\ a & x_2^2\end{array}\right).\] We can make several important remarks. First of all, the map is defined by monomials, so $\mathcal{R}_{2,2,2}$ is a toric variety. It is a cone over the Veronese variety $\mathcal{V}_{1,2}$ and it spans the whole $\mathbb{P}^3$. Its ideal is generated by one quadratic polynomial. Finally, it can be seen as the embedding \[\mathbb{P}(1,1,2)\hookrightarrow\mathbb{P}^3 \] of a weighted projective plane, defined by all monomials of (weighted) degree 2. \end{es} Our main goal is to generalize these remarks to all values of $d,k,m$. \section*{$\mathcal{R}_{d,k,m}$ as a toric variety} Roughly speaking, a variety is toric if it is the image of a monomial map. A toric variety not only has nice properties - for instance, it is irreducible, rational and its ideal is generated by binomials - but it can be associated to a polytope that completely encodes its geometry. This makes toric varieties accessible from a theoretical, combinatorial and computational viewpoint. A good reference on toric varieties is \cite{cox}. This section is devoted to the proof that the Rough Veronese is indeed toric and to provide an explicit way to make computations on it. If we write an element $L\in\Lie^m(\mathbb{R}^d)$ as the sum $L=L_1+\ldots+L_m$ of terms of order $1,\ldots,m$, then $p_k(\exp(L))\in(\mathbb{R}^d)^{\otimes k}$ is a linear combination of all possible ways to get an order $k$ tensor by multiplying $L_1,\ldots,L_m$. If we consider coordinates $T_{i_1\ldots i_k}$ on $(\mathbb{C}^d)^{\otimes k}$, we can rephrase this observation by saying that every coordinate $T_{i_1\ldots i_k}$ of $f_{d,k,m}$ is a linear combination of weighted degree $k$ monomials. The weight of a variable corresponding to a length $i$ Lyndon word is $i$. We can define a map \[\Lie^m(\mathbb{R}^d)\to(\mathbb{R}^d)^{\otimes k}\] by using all such weighted monomials, and consider its complexification $g_{d,k,m}$. By our observation, there is a linear change of coordinates sending the image of $g_{d,k,m}$ to the image of $f_{d,k,m}$. We want to prove that such change of coordinates is invertible, that is, every weighted monomial can be obtained as a linear combination of the entries of $f_{d,k,m}$, and we also want to do so as explicitly as possible. This section is based on \cite[Section IV]{exponential}. Let us start with two definitions. \begin{dfn}\label{def:S} Let $v$ be a word. If $v={\bf e}$ is the empty word, set $S_{{\bf e}}={\bf e}$. Otherwise, we define $S_v$ in the following recursive way. \begin{enumerate} \item If $v$ is Lyndon, write $v=l\cdot w$, where $l$ is a letter, and define $S_{v}=l\cdot S_w$. \item Otherwise, write $v=w_1^{ i_1}\cdot\ldots\cdot w_k^{i_k}$ as concatenation of decreasing Lyndon words. This can be done uniquely by \cite[Section 7.4]{bible}. Define \[S_v=\frac{S_{w_1}^{\shuffle i_1}\shuffle\ldots\shuffle S_{w_k}^{\shuffle i_k}}{i_1!\ldots i_k!}.\] \end{enumerate} \end{dfn} The next ingredient we need is the following. \begin{dfn}\label{def:pi} Define a linear map $\psi:T(\mathbb{R}^d)\to T(\mathbb{R}^d)$ that acts on a word $v$ by \[v\mapsto\sum_{n=1}^{\infty}\frac{(-1)^{n+1}}{n}\sum_{\substack{u_1,\dots,u_n \\ \mbox{\begin{tiny} nonempty \end{tiny}} \\ \mbox{\begin{tiny} words \end{tiny}}}} \langle v,u_1\cdot\ldots\cdot u_n\rangle u_1\shuffle\ldots\shuffle u_n.\] \end{dfn} In \cite{exponential}, the map $\psi$ is called $\pi'_1 . Observe that only finitely many terms of the sum are non-zero. Indeed, an element in $T(\mathbb{R}^d)$ is the linear combination of finitely many words, and if $v$ is a word then there are only finitely many ways to write it as a concatenation $u_1\cdot\ldots\cdot u_n$. Moreover, if $v$ is a word of length $l$ and $v=u_1\cdot\ldots\cdot u_n$, then $u_1\shuffle\ldots\shuffle u_n$ is a sum of length $l$ words. This means that $\psi$ preserves the grading of $T(\mathbb{R}^d)$. The following result about the exponential map will be of great help. Recall that $W_{d,m}$ is the set of Lyndon words of length at most $m$ in the alphabet with $d$ letters. \begin{lem} \label{lem:explicit change of coord} Let $\{P_{w}\mid w\in W_{d,m}\}$ be the basis of $\Lie^m(\mathbb{R}^d)$ indexed by the Lyndon words and let $w_1,\dots,w_r\in W_{d,m}$. Then \[\left\langle\exp\left( \sum_{w\in W_{d,m}}\alpha_wP_{w}\right) ,\psi(S_{w_1})\shuffle\ldots\shuffle \psi(S_{w_r})\right\rangle=\alpha_{w_1}\cdot\ldots\cdot \alpha_{w_r}.\] \begin{proof} By \cite[Theorem 1]{exponential}, for every $i\in\{1,\dots,r\}$ we have \[\left\langle\exp\left( \sum_{w\in W_{d,m}}\alpha_wP_{w}\right) ,\psi(S_{w_i})\right\rangle=\alpha_{w_i}.\] Now the statement follows by Lemma \ref{lem:shuffle identity}. \end{proof} \end{lem} Clearly every entry of $f_{d,k,m}$ is a linear combination of monomials. Lemma \ref{lem:explicit change of coord} shows that the converse holds. Every monomial can be obtained as linear combination of entries of $f_{d,k,m}$. In other words, we can use the map $\psi$ to build the linear forms we needed to pass from $f_{d,k,m}$ to $g_{d,k,m}$, allowing us to identify them. Let us summarize the conclusion. \begin{pro}\label{pro:toric} For every $w\in W_{d,m}$, define a variable $x_w$ and assign it the weight $|w|$. Then, up to a linear change of coordinates, $f_{d,k,m}$ is defined by all monomials of weighed degree $k$ in the variables $\{x_w\mid w\in W_{d,m}\}$. More explicitly, if we set $$J=\{(w_1,\dots, w_r)\mid r\in\mathbb{N}, w_i\in W_{d,m}\mbox{ and } |w_1|+\ldots+|w_r|=k\},$$ then $\mathcal{R}_{d,k,m}$ is isomorphic to the image of \[[x_w]_{w\in W_{d,m}}\mapsto [x_{w_1}\cdot\ldots\cdot x_{w_r}]_{(w_1,\dots, w_r)\in J}. \] In particular, it is a toric variety. \end{pro} \begin{rmk} We could phrase Proposition \ref{pro:toric} in a slightly different way by saying that, up to a linear change of coordinates, $p_k\circ\exp$ is a monomial map. More precisely, for every $d,k,m\in\mathbb{N}$ there exists a linear change of coordinates $\lambda:(\mathbb{R}^d)^{\otimes k}\to (\mathbb{R}^d)^{\otimes k}$ such that the map \[\lambda\circ p_k\circ \exp:\Lie^m(\mathbb{R}^d)\to (\mathbb{R}^d)^{\otimes k}\] is defined by all monomials of weighted degree $k$. \end{rmk} Notice that the image of the map defined in Proposition \ref{pro:toric} lies in a possibly proper linear subspace of $\mathbb{P}^{d^k-1}$. We will discuss the linear span of $\mathcal{R}_{d,k,m}$ in Lemma \ref{pro:countmonomials}. For the purpose of practical applications, we want to explicitly describe a linear change of coordinates in $\mathbb{P}^{d^k-1}$ that makes $\mathcal{R}_{d,k,m}$ toric. If $m<k$, then every element of $L\in\Lie^{m+1}(\mathbb{R}^d)$ can be written uniquely as $L=(L_1,L_2)$, where $L_1\in\Lie^m(\mathbb{R}^d)$. In this case, $f_{d,k,m}(L_1)=f_{d,k,m+1}(L_1,0)$. This means that every change of coordinates that makes $f_{d,k,m+1}$ a monomial map also makes $f_{d,k,m}$ monomial. Therefore we can assume $m=k$, because the change of coordinates that will make $\mathcal{R}_{d,k,k}$ toric will also work on $\mathcal{R}_{d,k,m}$ for every $m\le k$. In the notation of Proposition \ref{pro:toric}, the change of coordinates in $\mathbb{P}^{d^k-1}$ is \begin{align}\label{eq:changeofcoordinates} T\mapsto \left[ \langle T,\psi(S_{w_1})\shuffle\ldots\shuffle \psi(S_{w_r}) \rangle\right]_{(w_1,\dots, w_r)\in J}. \end{align} This is indeed well defined because, as we will show in Proposition \ref{pro:nondegen}, $J$ has exactly $d^k$ elements when $m=k$. \begin{es} \label{ex:change of coordinates U24} Let us consider $\mathcal{U}_{2,4}=\mathcal{R}_{2,4,4}$ and let $T_{ijkl}$ be the coordinates of $(\mathbb{R}^2)^{\otimes 4}$. The set of Lyndon words of length at most 4 in the alphabet $\{{\bf 1},{\bf 2}\}$ is \[W_{2,4}=\{{\bf 1},{\bf 2},{\bf 12},{\bf 112},{\bf 122},{\bf 1112},{\bf 1122},{\bf 1222}\}.\] By using Definition \ref{def:S}, in this case we get $S_w=w$ for every $w\in W_{2,4}$. The first 5 entries of the change of variables (\ref{eq:changeofcoordinates}) are then \begin{align} \langle T,\psi(S_{{\bf i}})\shuffle\psi(S_{{\bf j}})\shuffle \psi(S_{{\bf k}})\shuffle \psi(S_{{\bf l}}) \rangle&=\langle T,\psi({\bf i})\shuffle\psi({\bf j})\shuffle \psi({\bf k})\shuffle \psi({\bf l}) \rangle\nonumber\\ &=\langle T,{\bf i}\shuffle{\bf j}\shuffle {\bf k}\shuffle {\bf l} \rangle\nonumber\\ &=\sum_{\sigma}T_{\sigma(i)\sigma(j)\sigma(k)\sigma(l)},\nonumber \end{align} where $1\le i\le j\le k\le l\le 2$ and $\sigma$ ranges among the permutations of $\{i,j,k,l\}$. The next 3 entries are of the form \begin{align} \langle T,\psi(S_{{\bf 12}})\shuffle\psi(S_{{\bf 1}})\shuffle \psi(S_{{\bf 1}})\rangle&=\langle T,\psi({\bf 12})\shuffle\psi({\bf 1})\shuffle \psi({\bf 1})\rangle\nonumber\\ &=\left\langle T,\frac{1}{2}({\bf 12}-{\bf 21})\shuffle{\bf 1}\shuffle {\bf 1}\right\rangle \nonumber\\ &=\langle T, 3\cdot {\bf 1112}+ {\bf 1121}- {\bf 1211}-3\cdot {\bf 2111}\rangle \nonumber\\ &=3T_{1112}+T_{1121}-T_{1211}-3T_{2111}.\nonumber \end{align} We continue in this way until we reach length 4 Lyndon words \begin{align} \langle T,\psi(S_{{\bf 1112}})\rangle=\langle T,\psi({\bf 1112})\rangle=\left\langle T,\frac{1}{6}({\bf 1211}-{\bf 1121}) \right\rangle=\frac{1}{6}(T_{1211}-T_{1121})\nonumber \end{align} and, in the same way, \begin{align} \langle T,\psi(S_{{\bf 1122}})\rangle&=\frac{1}{6}(T_{1122}-T_{1212}+T_{2121}-T_{2211}),\nonumber\\ \langle T,\psi(S_{{\bf 1222}})\rangle&=\frac16(T_{2212}-T_{2122}).\nonumber \end{align} Compare our computations to the ones in \cite[Section IV]{exponential}. If we denote by $Q_{ijkl}$ the new coordinates on $(\mathbb{R}^2)^{\otimes 4}$, then our change of coordinates in $\mathbb{P}^{15}$ is \[\left[ \begin{matrix} Q_{1111}\\ Q_{1112}\\ Q_{1121}\\ Q_{1122}\\ Q_{1211}\\ Q_{1212}\\ Q_{1221}\\ Q_{1222}\\ Q_{2111}\\ Q_{2112}\\ Q_{2121}\\ Q_{2122}\\ Q_{2211}\\ Q_{2212}\\ Q_{2221}\\ Q_{2222}\\ \end{matrix}\right]=\left[ \begin{matrix} 24T_{1111}\\ 6(T_{1112}+T_{1121}+T_{1211}+T_{2111})\\ 4(T_{1122}+T_{1212}+T_{1221}+T_{2211}+T_{2112}+T_{2121})\\ 6(T_{1222}+T_{2122}+T_{2212}+T_{2221})\\ 24T_{2222}\\ 3T_{1112}+T_{1121}-T_{1211}-3T_{2111}\\ 2T_{1122}+T_{1212}-T_{2121}-2T_{2211}\\ 3T_{1222}+T_{2122}-T_{2212}-3T_{2221}\\ \frac12(T_{1112}-T_{1121}-T_{1211}+T_{2111})\\ \frac13(T_{1122}+T_{2112}+T_{2211})-\frac16(T_{1212}+T_{2121}+4T_{1221})\\ \frac13(T_{1122}+T_{1221}+T_{2211})-\frac16(T_{1212}+T_{2121}+4T_{2112})\\ \frac12(T_{1222}-T_{2122}-T_{2212}+T_{2221})\\ T_{1122}-T_{1221}-T_{2112}+T_{2211}\\ \frac{1}{6}(T_{1211}-T_{1121})\\ \frac{1}{6}(T_{1122}-T_{1212}+T_{2121}-T_{2211})\\ \frac16(T_{2212}-T_{2122}) \end{matrix}\right]. \] \end{es} In Example \ref{ex:change of coordinates U24}, every $S_w$ constructed via Definition \ref{def:S} coincides with the corresponding word $w$. However, this is not true in general, as the case $d=k=3$ shows. \begin{es}\label{ex:change of coordinates U33} Let us consider $\mathcal{U}_{3,3}=\mathcal{R}_{3,3,3} . The set of Lyndon words of length at most 3 in the alphabet $\{{\bf 1},{\bf 2},{\bf 3}\}$ is \[W_{3,3}=\{{\bf 1},{\bf 2},{\bf 3},{\bf 12},{\bf 13},{\bf 23},{\bf 112},{\bf 113},{\bf 122},{\bf 133},{\bf 223},{\bf 233},{\bf 123},{\bf 132}\}. \] Here we observe that $S_{{\bf 132}}={\bf 123}+{\bf 132}$. On the other hand, $S_w=w$ for every $w\in W_{3,3}\setminus\{{\bf 132}\}$. By carrying out the computations as in Example \ref{ex:change of coordinates U24}, we find that the first 10 entries of the change of variables (\ref{eq:changeofcoordinates}) are \begin{align} \langle T,\psi(S_{{\bf i}})\shuffle\psi(S_{{\bf j}})\shuffle \psi(S_{{\bf k}}) \rangl &=T_{kij}+T_{ikj}+T_{ijk}+T_{kji}+T_{jki}+T_{jik}\nonumber \end{align} for every $1\le i\le j\le k\le 3$. Then we have 9 more entries of the form \begin{align} \langle T,\psi(S_{{\bf i}})\shuffle\psi(S_{{\bf jk}})\rangl &=\frac{1}{2}(T_{ijk}+T_{jik}+T_{jki}-T_{ikj}-T_{kij}-T_{kji}),\nonumber \end{align}for every $i\in\{1,2,3\}$ and $1\le j<k\le 3$. Finally, we tackle length 3 Lyndon words. We have \begin{align} \langle T,\psi(S_{{\bf 132}})\rangl &=\frac{1}{6}(T_{123}+T_{132}+T_{231}+T_{321})-\frac{1}{3}(T_{213}+ T_{312}),\nonumber \end{align}while the last 7 entries are \begin{align} \langle T,\psi(S_{{\bf ijk}})\rangl &=\frac{1}{3}(T_{ijk}+T_{kji})-\frac{1}{6}(T_{kij}+T_{ikj}+T_{jki}+T_{jik}),\nonumber \end{align}for ${\bf ijk}\in\{{\bf 112},{\bf 113},{\bf 122},{\bf 133},{\bf 223},{\bf 233},{\bf 123}\}$. It is interesting to point out that some of these linear forms are used in \cite[Section 5]{annaemax} in order to recover the path of a given signature.\end{es} \begin{rmk} The universal variety contains many interesting subvarieties besides $\mathcal{R}_{d,k,m}$. Examples include the signature variety $\mathcal{L}_{d,k,m}$ of piecewise linear paths with $m$ steps and its subvariety $\mathcal{A}_{\nu,k}$ of axis-parallel paths, both studied in \cite{AFS}. The change of coordinates given by Lemma \ref{lem:explicit change of coord} proves that $\mathcal{R}_{d,k,m}$ and therefore $\mathcal{U}_{d,k}$ are toric, but it does not necessarily work as nicely with other subvarieties. For instance, a software computation shows that, after our change of coordinates, the ideals of both $\mathcal{L}_{2,3,2}$ and $\mathcal{A}_{1212,3}$ are not generated by binomials. The question of whether $\mathcal{A}_{\nu,k}$ is toric is addressed in \cite[Section 3.1]{CGM}, but the answer is not known in full generality. \end{rmk} \section*{Further geometric properties} In this section we will generalize the remarks we made for Example \ref{ex:R222}. While $\mathcal{R}_{2,2,2}\subset\mathbb{P}^3$ is nondegenerate, for other values of $d,k,m$ the Rough Veronese may be contained in a smaller linear subspace, as we already observed in Proposition \ref{pro:toric}. \begin{pro}\label{pro:countmonomials} The affine dimension of the linear span of $\mathcal{R}_{d,k,m}\subset\mathbb{P}^{d^k-1}$ is \[\sum_{\lambda\vdash k,\lambda_1\le m}\left( \prod_{i=1}^{m}\binom{\mu_{i,d}+\sharp\{j\mid\lambda_j=i\}-1}{\mu_{i,d}-1}\right) .\] This number equals the coefficient of $t^k$ of the expansion of the generating function \[\prod_{i=1}^{m}\frac{1}{(1-t^i)^{\mu_{i,d}}}.\] \begin{proof} As in Proposition \ref{pro:toric}, we define a variable $x_w$ for every $w\in W_{d,m}$ and we assign it the weight $|w|$. Given a weighted degree $k$ monomial $\varphi\in\mathbb{C}[x_w\mid w\in W_{d,m}]$, we can write it in reverse lexicographic order as a string of possibly repeated variables. Define a partition $\lambda$ of $k$ by \[\lambda_j=i\Leftrightarrow\mbox{ the $j$-th entry of the string is } x_w \mbox{ for some } |w|=i.\] Now $\varphi$ is the product of monomials $\varphi=\varphi_1\cdot\ldots\cdot \varphi_m$, where $\varphi_i$ is a monomial in $\mathbb{C}[x_w\mid |w|=i]$. The degree of $\varphi_i$ is the number of times $i$ appears as an entry of $\lambda$. Hence, for every $\varphi_i$ there are $\binom{\mu_{i,d}+\sharp\{j\mid\lambda_j=i\}-1}{\mu_{i,d}-1}$ choices. For the generating function, see \cite[Section 1.8]{combinat}. \end{proof} \end{pro} Nonetheless, the universal variety $\mathcal{U}_{d,k}=\mathcal{R}_{d,k,k}$ is indeed nondegenerate for every $d$ and $k$. \begin{pro}\label{pro:nondegen} Under our assumption $k\ge m$, the rough Veronese $\mathcal{R}_{d,k,m}\subset\mathbb{P}^{d^k-1}$ is nondegenerate if and only if $m=k$. \begin{proof} By Proposition \ref{pro:countmonomials}, the dimension of the linear span of $\mathcal{R}_{d,k,m}$ is strictly increasing with respect to $m$, hence it is enough to show that $\mathcal{R}_{d,k,k}$ is nondegenerate. Thanks to Proposition \ref{pro:toric}, we only have to prove that there are $d^k$ distinct monomials of weighted degree $k$. Proposition \ref{pro:countmonomials} may be difficult to apply, so instead we want to define a bijection between these monomials and the set of length $k$ words in the alphabet with $d$ letters. Since our variables are indexed by Lyndon words, we can think of a monomial as a product of Lyndon words such that the sum of the lengths is $k$. Observe that the monomial remains the same if we permute the variables, i.e.~the Lyndon words. In order to avoid redundancy, we can fix an order. Now we only have to prove that, after fixing an order among Lyndon words, every length $k$ word in the alphabet with $d$ letters can be written uniquely as an ordered product of Lyndon words whose lengths sum to $k$. This is a well-known fact, and a proof can be found for instance in \cite[Section 7.4]{bible}. \end{proof} \end{pro} The next result shows another feature of rough paths. Not only does the universal variety $\mathcal{U}_{d,k}$ coincide with the last Rough Veronese $\mathcal{R}_{d,k,k}$, but its structure is already determined by the second to last one, $\mathcal{R}_{d,k,k-1}$. \begin{pro}\label{pro:universal} The universal variety $\mathcal{U}_{d,k}=\mathcal{R}_{d,k,k}$ is a cone over $\mathcal{R}_{d,k,k-1}$ with vertex $\mathbb{P}^{\mu_{k,d}-1}$. The preimage of the vertex is defined by the vanishing of the first $k-1$ entries. \begin{proof} Let $V=\mathbb{R}^{\mu_{d,k}}$ be the vector subspace of $\Lie^k(\mathbb{R}^d)$ defined by the vanishing of the first $k-1$ entries. Thanks to Proposition \ref{pro:toric}, we know that $\mathcal{R}_{d,k,k}$ is defined by the degree $k$ weighted monomials. Exactly $\mu_{k,d}$ of them are $\{x_w\mid w\in W_{d,m}\mbox{ and } |w|=k\}$. In other words, the map $p_k\circ\exp$ restricts to the identity on $V$. It follows that $$p_k(\exp(\Lie^k(\mathbb{R}^d)))=p_k(\exp(\Lie^{k-1}(\mathbb{R}^d)))\times V$$ is a cylinder. When we pass to the projectivization, we get a cone with base $\mathcal{R}_{d,k,k-1}$ and vertex $\mathbb{P}(p_k(\exp(V))\otimes\mathbb{C})=\mathbb{P}(V\otimes\mathbb{C})$. \end{proof} \end{pro} \begin{rmk} It is interesting to try to classify the paths whose signatures lie in the vertex of the universal variety. Equivalently, we wonder what it means for a rough path to have zeroes in the first $k-1$ entries. It is straightforward to check that for a smooth path $X$, the first signature \[\sigma^{(1)}(X)=X(1)-X(0)\] is just the vector joining the endpoint of $X$ to its starting point. The second signature also has a geometric interpretation: if we take the projection of $X$ onto the $i,j$ plane, the signed area of the region bounded by $X$ and the segment between $X(0)$ and $X(1)$ is $\frac{1}{2}(\sigma(X)_{ij}-\sigma(X)_{ji})$. For instance, if $X$ is a loop then $\sigma^{(1)}(X)=0$ but $\sigma^{(2)}(X)$ may be nonzero. However, if we allow not only smooth paths, but also rough paths, we find more interesting examples. There are rough paths ${\bf X}$ such that $p_1({\bf X}(t))=p_1({\bf X}(0))$ for every $t$, but with a nonzero second entry. If we try to imagine it, our intuition suggests a constant path that nonetheless describes an area. This is not possible for a piecewise regular path. One example of these \textit{pure area paths} is \cite[Exercise 2.17]{Friz}, where it is built as a limit of smaller and smaller loops. In the same fashion, we can think of the vanishing of the first two signatures as a loop with some symmetry that makes the signed area zero, such as two circles in $\mathbb{R}^2$ meeting at a point. Then again, there are more exotic examples of rough paths with the same property. However, the geometric meaning of the third and higher order signatures is not yet clear, as it is an open research area. \end{rmk} We now turn our attention to the ideal of $\mathcal{R}_{d,k,m}$. In \cite[Section 4]{annaemax}, it is proved that the ideal of $\mathcal{R}_{d,3,3}$ is generated by the $2\times 2$ minors of a suitable Hankel matrix, in particular by quadrics. In many other examples this ideal is generated in degree 2. However, this is not true in general, not even if $k=m$. The following counterexample, suggested by Micha\l{}ek, answers a question posed in \cite[Section 4.3]{AFS}. \begin{pro}\label{pro:not generated by quadrics} For $14\le m\le k=20$, the ideal of the Rough Veronese variety $\mathcal{R}_{d,20,m}$ is not generated by quadrics. \begin{proof} Let $N$ be the number of weighted degree $k$ monomials. By Proposition \ref{pro:toric}, up to change of coordinates $\mathcal{R}_{d,20,m}$ lies in $\mathbb{P}^{N-1}$. The coordinates $T$ of $\mathbb{P}^{N-1}$ are indexed by the $N$ weighted monomials. For every $i\in\{1,\dots,m\}$, let $x^{(i)}$ be one of the variables of weight $i$. Let $I$ be the ideal of $\mathcal{R}_{d,20,m}$ and define \begin{align} t_1&=T_{x^{(1)}x^{(9)}x^{(10)}},& t_2 &= T_{x^{(5)}x^{(7)}x^{(8)}},\nonumber\\ t_3&=T_{x^{(2)}x^{(4)}x^{(14)}}, & t_4&=T_{x^{(1)}x^{(5)}x^{(14)}},\nonumber\\ t_5&=T_{x^{(4)}x^{(7)}x^{(9)}},& t_6&=T_{x^{(2)}x^{(8)}x^{(10)}}.\nonumber \end{align} Then $f=t_1t_2t_3-t_4t_5t_6$ is a degree 3 element of $I$. Let us show that $f$ is not generated by quadrics. Let $q_1,\dots,q_r$ be the degree two generators of $I$. Since $\mathcal{R}_{d,20,m}$ is toric, they are binomials and so we can write $q_i=g_i-h_i$, where each $g_i$ and each $h_i$ is a degree 2 monomial. Assume by contradiction that $f$ can be algebraically generated by $q_1,\dots,q_r$. Then there exist degree 1 polynomials $l_1,\dots,l_r$ such that, up to order, $f$ is a sum of nonzero polynomials $f=l_1(g_1-h_1)+\ldots+l_t(g_r-h_r)$. It follows that each term in the sum is a multiple of some $g_i$ or some $h_i$. For instance, there is a degree 2 monomial dividing $t_1t_2t_3$. The only three possibilities are $t_1t_2$, $t_1t_3$ and $t_2t_3$. Suppose then that $g_i=t_1t_2$. Since $g_i-h_i\in I$, the product of the two monomials indexing the variables of $h_i$ equals $t_1t_2$. But it is easy to see that $1+9+10$ and $5+7+8$ are the only ways to obtain 20 by sums of non-repeated elements of $\{1,5,7,8,9,10\}$. Hence $g_i-h_i=0$, a contradiction. \end{proof} \end{pro} Among the features of $\mathcal{R}_{2,2,2}$ we pointed out in Example \ref{ex:R222}, there is one we still have to check. We saw that $\mathcal{R}_{2,2,2}$ is the embedding of $\mathbb{P}(1,1,2)$ given by its weighted quadrics. If we define the sequence of weights \[s=(\underbrace{1,\dots,1}_{\mu_{1,d}\mbox{ {\Small times}}},\underbrace{2,\dots,2}_{\mu_{2,d}\mbox{ {\Small times}}},\dots,\underbrace{m,\dots,m}_{\mu_{m,d}\mbox{ {\Small times}}}), \] then by Proposition \ref{pro:toric} $f_{d,k,m}$ induces a rational map $\mathbb{P}(s)\dashrightarrow\mathbb{P}^{d^k-1}$, that we will still call in the same way by abuse of notation. However, $f_{d,k,m}$ does not need to be an embedding. Actually, it does not even need to be defined everywhere. For instance, $f_{2,3,2}:\mathbb{P}(1,1,2)\dashrightarrow\mathbb{P}^7$ is defined by \[[x_1,x_2,a]\mapsto \left[\begin{array}{cc} x_1^3 & x_1^2x_2\\ x_1x_2^2 & x_2^3\\\hline x_1a & x_2a\\ 0 & 0\end{array}\right],\] so $[0,0,1]$ is a base point. This is a general behavior. \begin{pro} The map $f_{d,k,m}:\mathbb{P}(s)\dashrightarrow\mathbb{P}^{d^k-1}$ is base point free if and only if every entry of $s$ divides $k$. \begin{proof} Assume that $k$ is a multiple of every entry of $s$. Then for every variable $x_w$, there is a power of $x_w$ appearing among the monomials of weighted degree $k$. Therefore the only way for all weighted monomials to vanish is setting all variables to zero. This means there are no base points. On the other hand, assume that $k$ is not a multiple of one of the entries of $s$, say $i$, and consider a variable $x^{(i)}$ of weight $i$. Since $i\nmid k$, no power of $x^{(i)}$ appears among the monomials defining $f_{d,k,m}$. Then $[0,\dots,0,1,0\dots,0]\in\mathbb{P}(s)$, with a 1 in the entry corresponding to $x^{(i)}$, is a base point for $f_{d,k,m}$. \end{proof} \end{pro} \section*{Table of invariants}\label{sec:table} We collect some of the geometric invariants of $\mathcal{R}_{d,k,m}$, obtained with the software Macaulay2. We compute the projective dimension of the linear span, the dimension of $\mathcal{R}_{d,k,m}$ and its degree. Despite Proposition \ref{pro:not generated by quadrics}, all the ideals in the examples we present are generated by quadrics. In the last column \textquotedblleft gen\textquotedblright\ we record the number of generators. \begin{center} \begin{tabular}{|c c c |c c c|c|} \hline $d$ & $k$ & $m$ & span & $\dim$ & $\deg$ & gen \\ \hline $2$ & $2$ & $2$ & 3 & 2 & 2 & 1 \\ $3$ & $2$ & $2$ & 8 & 5 & 4 & 6 \\ $4$ & $2$ & $2$ & 15 & 9 & 8 & 20 \\ $5$ & $2$ & $2$ & 24 & 14 & 16 & 50 \\ $6$ & $2$ & $2$ & 35 & 20 & 32 & 105 \\ $2$ & $3$ & $2$ & 5 & 2 & 4 & 6 \\ $3$ & $3$ & $2$ & 18 & 5 & 24 & 81 \\ $4$ & $3$ & $2$ & 43 & 9 & 200 & 486 \\ $5$ & $3$ & $2$ & 84 & 14 & 2221 & 1920 \\ $2$ & $4$ & $2$ & 8 & 2 & 8 & 27 \\ $3$ & $4$ & $2$ & 38 & 5 & 128 & 528 \\ $2$ & $5$ & $2$ & 11 & 2 & 12 & 43 \\ $3$ & $5$ & $2$ & 68 & 5 & 368 & 1806\\ $2$ & $6$ & $2$ & 15 & 2 & 18 & 87 \\ \hline $2$ & $3$ & $3$ & 7 & 4 & 4 & 6 \\ $3$ & $3$ & $3$ & 26 & 13 & 24 & 81 \\ $4$ & $3$ & $3$ & 63 & 29 & 200 & 486 \\ $2$ & $4$ & $3$ & 12 & 4 & 12 & 33 \\ $3$ & $4$ & $3$ & 62 & 13 & 672 & 954 \\ $2$ & $5$ & $3$ & 19 & 4 & 28 & 108 \\ \hline $2$ & $4$ & $4$ & 15 & 7 & 12 & 33 \\ $2$ & $5$ & $4$ & 25 & 7 & 40 & 150\\ \hline $2$ & $5$ & $5$ & 31 & 13 & 40 & 150\\ $2$ & $6$ & $5$ & 54& 13 & 336 & 694 \\ \hline \end{tabular} \end{center}

fb05ae554e202dd34b39eb5a35862129c7a72262

\section{Introduction} The purpose of this article is to survey some of the context, achievements, challenges and mysteries of the field of {\em metric dimension reduction}, including new perspectives on major older results as well as recent advances. From the point of view of theoretical computer science, mathematicians "stumbled upon" metric dimension reduction in the early 1980s, as exemplified by the following quote~\cite{Vem04}. \blockquote{{ \small \em Two decades ago, analysts stumbled upon a surprising fact [...], the Johnson--Lindenstrauss Lemma, as a crucial tool in their project of extending functions in continuous ways. This result [...] says that, if you project $n$ points in some high dimensional space down to a random $O(\log n)$-dimensional plane, the chances are overwhelming that all distances will be preserved within a small relative error. So, if distance is all you care about, there is no reason to stay in high dimensions! \begin{flushright}{\small \em C.~Papadimitriou, 2004 (forward to {\em The random projection method} by S.~Vempala).}\end{flushright}}} The above use of the term "stumbled upon" is justified, because it would be fair to say that at the inception of this research direction mathematicians did not anticipate the remarkable swath of its later impact on algorithms. However, rather than being discovered accidentally, the investigations that will be surveyed here can be motivated by classical issues in metric geometry. From the internal perspective of pure mathematics, it would be more befitting to state that the aforementioned early work stumbled upon the unexpected depth, difficulty and richness of basic questions on the relation between "rough quantitative geometry" and dimension. Despite major efforts by many mathematicians over the past four decades, such questions remain by and large stubbornly open. We will explain below key ideas of major developments in metric dimension reduction, and also describe the larger mathematical landscape that partially motivates these investigations, most notably the {\em bi-Lipschitz embedding problem into $\mathbb R^n$} and the {\em Ribe program}. By choosing to focus on aspects of this area within pure mathematics, we will put aside the large (and growing) literature that investigates algorithmic ramifications of metric dimension reduction. Such applications warrant a separate treatment that is far beyond the scope of the present exposition; some aspects of that material are covered in the monographs~\cite{Mat02,Vem04,Har11} and the surveys~\cite{Ind01,Lin02}, as well as the articles of Andoni--Indyk--Razenshteyn and Arora in the present volume. \begin{remark} The broader term {\em dimension reduction} is used ubiquitously in statistics and machine learning, with striking applications whose full rigorous understanding sometimes awaits the scrutiny of mathematicians (see e.g.~\cite{Bur10}). A common (purposefully vague) description of this term is the desire to decrease the degrees of freedom of a high-dimensional data set while approximately preserving some of its pertinent features; formulated in such great generality, the area includes topics such as neural networks (see e.g.~\cite{HS06}). The commonly used term {\em curse of dimensionality} refers to the perceived impossibility of this goal in many settings, and that the performance (running time, storage space) of certain algorithmic tasks must deteriorate exponentially as the underlying dimension grows. But, sometimes it does seem that certain high-dimensional data sets can be realized faithfully using a small number of latent variables as auxiliary "coordinates." Here we restrict ourselves exclusively to {\em metric} dimension reduction, i.e., to notions of faithfulness of low-dimensional representations that require the (perhaps quite rough) preservation of pairwise distances, including ways to prove the impossibility thereof. \end{remark} \subsection*{Roadmap} The rest of the Introduction is an extensive and detailed account of the area of metric dimension reduction, including statements of most of the main known results, background and context, and many important open questions. The Introduction is thus an expository account of the field, so those readers who do not wish to delve into some proofs, could read it separately from the rest of the text. The remaining sections contain further details and complete justifications of those statements that have not appeared in the literature. \subsection{Bi-Lipschitz embeddings}\label{sec:bilip} Fix $\alpha\ge 1$. A metric space $(\mathcal{M},d_\mathcal{M})$ is said to embed with distortion $\alpha$ into a metric space $(\mathcal{N},d_\mathcal{N})$ if there is a mapping (an embedding) $f:\mathcal{M}\to \mathcal{N}$ and (a scaling factor) $\tau>0$ such that \begin{equation}\label{eq:def embedding} \forall\, x,y\in \mathcal{M},\qquad \tau d_{\mathcal{M}}(x,y)\le d_\mathcal{N}\big(f(x),f(y)\big)\le \alpha \tau d_\mathcal{M}(x,y). \end{equation} The infimum over $\alpha\in [1,\infty]$ for which $(\mathcal{M},d_\mathcal{M})$ embeds with distortion $\alpha$ into $(\mathcal{N},d_\mathcal{N})$ is denoted $\mathsf{c}_{(\mathcal{N},d_\mathcal{N})}(\mathcal{M},d_\mathcal{M})$, or $\mathsf{c}_\mathcal{N}(\mathcal{M})$ if the underlying metrics are clear from the context. If $\mathsf{c}_\mathcal{N}(\mathcal{M})<\infty$, then $(\mathcal{M},d_\mathcal{M})$ is said to admit a bi-Lipschitz embedding into $(\mathcal{N},d_\mathcal{N})$. Given $p\in [1,\infty)$, if $\mathcal{N}$ is an $L_p(\mu)$ space into which $\mathcal{M}$ admits a bi-Lipschitz embedding, then we use the notation $\mathsf{c}_{L_p(\mu)}(\mathcal{M})=\mathsf{c}_p(\mathcal{M})$. The numerical invariant $\mathsf{c}_2(\mathcal{M})$, which measures the extent to which $\mathcal{M}$ is close to being a (subset of a) Euclidean geometry, is called the {\em Euclidean distortion} of $\mathcal{M}$. A century of intensive research into bi-Lipschitz embeddings led to a rich theory with many deep achievements, but the following problem, which is one of the first questions that arise naturally in this direction, remains a major longstanding mystery; see e.g.~\cite{Sem99,LP01,Hei03,NN12}. Analogous issues in the context of topological dimension, differentiable manifolds and Riemannian manifolds were famously settled by Menger~\cite{Men28} and N\"obeling~\cite{Nob31}, Whitney~\cite{Whi36} and Nash~\cite{Nas54}, respectively. \begin{problem}[the bi-Lipschitz embedding problem into $\mathbb R^k$]\label{Q:bilip Rk} Obtain an intrinsic characterization of those metric spaces $(\mathcal{M},d_\mathcal{M})$ that admit a bi-Lipschitz embedding into $\mathbb R^k$ for some $k\in \mathbb N$. \end{problem} Problem~\ref{Q:bilip Rk} is one of the qualitative underpinnings of the issues that will be surveyed here. We say that it is "qualitative" because it ignores the magnitude of the distortion altogether, and therefore one does not need to specify which norm on $\mathbb R^k$ is considered in Problem~\ref{Q:bilip Rk}, since all the norms on $\mathbb R^k$ are (bi-Lipschitz) equivalent. Problem~\ref{Q:bilip Rk} is also (purposefully) somewhat vague, because the notion of "intrinsic characterization" is not well-defined. We will return to this matter in Section~\ref{sec:doubling} below, where candidates for such a characterization are discussed. At this juncture, it suffices to illustrate what Problem~\ref{Q:bilip Rk} aims to achieve through the following useful example. If one does not impose any restriction on the target dimension and allows for a bi-Lipschitz embedding into an infinite dimensional Hilbert space, then the following intrinsic characterization is available. A metric space $(\mathcal{M},d_\mathcal{M})$ admits a bi-Lipschitz embedding into a Hilbert space if and only if there exists $C=C_\mathcal{M}\in [0,1)$ such that for every $n\in \mathbb N$ and every positive semidefinite symmetric matrix $\mathsf{A}=(a_{ij})\in \mathsf{M}_n(\mathbb R)$ all of whose rows sum to zero (i.e., $\sum_{j=1}^na_{ij}=0$ for every $i\in \{1,\ldots,n\}$), the following quadratic distance inequality holds true. \begin{equation}\label{eq:K LLR} \forall\, x_1,\ldots,x_n\in \mathcal{M},\qquad \sum_{i=1}^n\sum_{j=1}^na_{ij}d_\mathcal{M}(x_i,x_j)^2\le C\sum_{i=1}^n\sum_{j=1}^n|a_{ij}|d_\mathcal{M}(x_i,x_j)^2. \end{equation} In fact, one can refine this statement quantitatively as follows. A metric space embeds with distortion $\alpha\in [1,\infty)$ into a Hilbert space if and only if in the setting of~\eqref{eq:K LLR} we have \begin{equation}\label{eq:alpha LLR} \forall\, x_1,\ldots,x_n\in \mathcal{M},\qquad \sum_{i=1}^n\sum_{j=1}^na_{ij}d_\mathcal{M}(x_i,x_j)^2\le \frac{\alpha^2-1}{\alpha^2+1}\sum_{i=1}^n\sum_{j=1}^n|a_{ij}|d_\mathcal{M}(x_i,x_j)^2. \end{equation} The case $\alpha=1$ of~\eqref{eq:alpha LLR}, i.e., the case of isometric embeddings, is a famous classical theorem of Schoenberg~\cite{Sch38}, and the general case of~\eqref{eq:alpha LLR} is due to Linial, London and Rabinovich~\cite[Corollary~3.5]{LLR}. The above characterization is clearly intrinsic, as it is a family of finitary distance inequalities among points of $\mathcal{M}$ that do not make reference to any other auxiliary/external object. With such a characterization at hand, one could examine the internal structure of a given metric space so as to determine if it can be represented faithfully as a subset of a Hilbert space. Indeed, \cite{LLR} uses~\eqref{eq:alpha LLR} to obtain an algorithm that takes as input an $n$-point metric space $(\mathcal{M},d_\mathcal{M})$ and outputs in polynomial time an arbitrarily good approximation to its Euclidean distortion $\mathsf{c}_2(\mathcal{M})$. A meaningful answer to Problem~\ref{Q:bilip Rk} could in principle lead to a method for determining if a member of a family $\mathcal{F}$ of metric spaces admits an embedding with specified distortion into a member of a family $\mathcal{G}$ of low dimensional normed spaces. Formulated in such great generality, this type of question encompasses all of the investigations into metric dimension reduction that will be discussed in what follows, except that we will also examine analogous issues for embeddings with guarantees that are substantially weaker (though natural and useful) than the "vanilla" bi-Lipschitz requirement~\eqref{eq:def embedding}. \begin{remark}Analogues of the above questions are very natural also when the target low-dimensional geometries are not necessarily normed spaces. Formulating meaningful goals in such a setting is not as straightforward as it is for normed spaces, e.g.~requiring that the target space is a manifold of low topological dimension is not very useful, so one must impose geometric restrictions on the target manifold. As another example (to which we will briefly return later), one could ask about embeddings into spaces of probability measures that are equipped with a Wasserstein (transportation cost) metric, with control on the size of the underlying metric space. At present, issues of this type are largely an uncharted terrain whose exploration is likely to be interesting and useful. \end{remark} \subsection{Local theory and the Ribe program} Besides being motivated by the bi-Lipschitz embedding problem into $\mathbb R^k$, much of the inspiration for the studies that will be presented below comes from a major endeavour in metric geometry called the {\em Ribe program}. This is a large and active research area that has been (partially) surveyed in~\cite{Kal08,Nao12,Bal13,Ost13,God17}. It would be highly unrealistic to attempt to cover it comprehensively here, but we will next present a self-contained general introduction to the Ribe program that is intended for non-experts, including aspects that are relevant to the ensuing discussion on metric dimension reduction. Martin Ribe was a mathematician who in the 1970s obtained a few beautiful results in functional analysis, prior to leaving mathematics. Among his achievements is a very influential rigidity theorem~\cite{Ribe76} which shows that the local linear theory of Banach spaces could in principle be described using only distances between points, and hence it could potentially apply to general metric spaces. Before formulating the above statement precisely, it is instructive to consider a key example. The {\em infimal cotype}~\cite{Mau03} $q_X$ of a Banach space $(X,\|\cdot\|)$ is the infimum over those $q\in [2,\infty]$ for which\footnote{In addition to the standard $``O"$ notation, we will use throughout this article the following standard and convenient asymptotic notation. Given two quantities $Q,Q'>0$, the notations $Q\lesssim Q'$ and $Q'\gtrsim Q$ mean that $Q\le CQ'$ for some universal constant $C>0$. The notation $Q\asymp Q'$ stands for $(Q\lesssim Q') \wedge (Q'\lesssim Q)$. If we need to allow for dependence on parameters, we indicate this by subscripts. For example, in the presence of auxiliary objects (e.g.~numbers or spaces) $\phi,\mathfrak{Z}$, the notation $Q\lesssim_{\phi,\mathfrak{Z}} Q'$ means that $Q\le C(\phi,\mathfrak{Z})Q' $, where $C(\phi,\mathfrak{Z}) >0$ is allowed to depend only on $\phi,\mathfrak{Z}$; similarly for the notations $Q\gtrsim_{\phi,\mathfrak{Z}} Q'$ and $Q\asymp_{\phi,\mathfrak{Z}} Q'$.} \begin{equation}\label{def:cotype} \forall\, n\in \mathbb N,\ \forall\, x_1,\ldots,x_n\in X,\qquad \sum_{i=1}^n \|x_i\|^2\lesssim_{X,q} \frac{n^{1-\frac2{q}}}{2^n}\sum_{\varepsilon\in \{-1,1\}^n}\bigg\|\sum_{i=1}^n \varepsilon_ix_i\bigg\|^2. \end{equation} In the special case $x_1=\ldots=x_n=x\in X\setminus\{0\}$, the left hand side of~\eqref{def:cotype} is equal to $n\|x\|^2$ and by expanding the squares one computes that the right hand side of~\eqref{def:cotype} is equal to $n^{2(1-1/q)}\|x\|^2$. Hence~\eqref{def:cotype} necessitates that $q\ge 2$, which explains why we imposed this restriction on $q$ at the outset. Note also that~\eqref{def:cotype} holds true in any Banach space when $q=\infty$. This is a quick consequence of the convexity of the mapping $x\mapsto \|x\|^2$, since for every $\varepsilon\in \{-1,1\}^n$ and $i \in \{1,\ldots,n\}$ we have \begin{multline}\label{convex priove cotype} \|x_i\|^2=\bigg\|\frac{(\varepsilon_1x_1+\ldots+\varepsilon_nx_n)+(-\varepsilon_1x_1-\ldots-\varepsilon_{i-1}x_{i-1}+\varepsilon_i x_i-\varepsilon_{i+1}x_{i+1}\ldots-\varepsilon_n x_n)}{2}\bigg\|^2\\ \le \frac{\|\varepsilon_1x_1+\ldots+\varepsilon_nx_n\|^2+\|-\varepsilon_1x_1-\ldots-\varepsilon_{i-1}x_{i-1}+\varepsilon_i x_i-\varepsilon_{i+1}x_{i+1}\ldots-\varepsilon_n x_n\|^2}{2}. \end{multline} By averaging~\eqref{convex priove cotype} over $\varepsilon\in \{-1,1\}^n$ and $i \in \{1,\ldots,n\}$ we see that~\eqref{def:cotype} holds if $q=\infty$. So, one could view~\eqref{def:cotype} for $q<\infty$ as a requirement that the norm $\|\cdot\|:X\to [0,\infty)$ has a property that is asymptotically stronger than mere convexity. When $X=\ell_\infty$, this requirement does not hold for any $q<\infty$, since if $\{x_i\}_{i=1}^n$ are the first $n$ elements of the coordinate basis, then the left hand side of~\eqref{def:cotype} equals $n$ while its right hand side equals $n^{1-1/q}$. Maurey and Pisier proved~\cite{MP73} that the above obstruction to having $q_X<\infty$ is actually the {\em only possible} such obstruction. Thus, by ruling out the presence of copies of $\{\ell_\infty^n\}_{n=1}^\infty$ in $X$ one immediately deduces the "upgraded" (asymptotically stronger as $n\to \infty$) randomized convexity inequality~\eqref{def:cotype} for some $q<\infty$. \begin{theorem}\label{thm:MP cotype} The following conditions are equivalent for every Banach space $(X,\|\cdot\|)$. \begin{itemize} \item There is no $\alpha\in [1,\infty)$ such that $\ell_\infty^n$ is $\alpha$-isomorphic to a subspace of $X$ for every $n\in \mathbb N$. \item $q_X<\infty$. \end{itemize} \end{theorem} The (standard) terminology that is used in Theorem~\ref{thm:MP cotype} is that given $\alpha\in [1,\infty)$, a Banach space $(Y,\|\cdot\|_Y)$ is said to be $\alpha$-isomorphic to a subspace of a Banach space $(Z,\|\cdot\|_Z)$ if there is a linear operator $T:Y\to Z$ satisfying $\|y\|_Y\le \|Ty\|_Z\le \alpha\|y\|_Y$ for every $y\in Y$; this is the same as saying that $Y$ embeds into $Z$ with distortion $\alpha$ via an embedding that is a linear operator. Suppose that $X$ and $Y$ are Banach spaces that are uniformly homeomorphic, i.e., there is a bijection $f:X\to Y$ such that both $f$ and $f^{-1}$ are uniformly continuous. By the aforementioned rigidity theorem of Ribe (which will be formulated below in full generality), this implies in particular that $q_X=q_Y$. So, despite the fact that the requirement~\eqref{def:cotype} involves linear operations (summation and sign changes) that do not make sense in general metric spaces, it is in fact preserved by purely metric (quantitatively continuous, though potentially very complicated) deformations. Therefore, in principle~\eqref{def:cotype} could be characterized while only making reference to distances between points in $X$. More generally, Ribe's rigidity theorem makes an analogous assertion for {\em any} isomorphic local linear property of a Banach space; we will define formally those properties in a moment, but, informally, they are requirements in the spirit of~\eqref{def:cotype} that depend only on the finite dimensional subspaces of the given Banach space and are stable under linear isomorphisms that could potentially incur a large error. The purely metric reformulation of~\eqref{def:cotype} about which we speculated above is only suggested but not guaranteed by Ribe's theorem. From Ribe's statement we will only infer an indication that there might be a "hidden dictionary" for translating certain linear properties into metric properties, but we will not be certain that any specific "entry" of this dictionary (e.g.~the entry for, say, "$q_X=\pi$") does in fact exist, and even if it does exist, we will not have an indication what it says. A hallmark of the Ribe program is that at its core it is a search for a family of analogies and definitions, rather than being a collection of specific conjectures. Once such analogies are made and the corresponding questions are formulated, their value is of course determined by the usefulness/depth of the phenomena that they uncover and the theorems that could be proved about them. Thus far, not all of the steps of this endeavour turned out to have a positive answer, but the vast majority did. This had major impact on the study of metric spaces that a priori have nothing to do with Banach spaces, such as graphs, manifolds, groups, and metrics that arise in algorithms (e.g.~as continuous relaxations). The first written formulation of the plan to uncover a hidden dictionary between normed spaces and metric spaces is the following quote of Bourgain~\cite{Bourgain-trees}, a decade after Ribe's theorem appeared. \blockquote{{ \small \em It follows in particular from Ribe's result [...] that the notions from local theory of normed spaces are determined by the metric structure of the space and thus have a purely metrical formulation. The next step consists in studying these metrical concepts in general metric spaces in an attempt to develop an analogue of the linear theory. A detailed exposition of this program will appear in J. Lindenstrauss's forthcoming survey [...] in our "dictionary" linear operators are translated in Lipschitz maps, the operator norm by the Lipschitz constant of the map [...] The translations of "Banach-Mazur distance" and "finite-representability" in linear theory are immediate. At the roots of the local theory of normed spaces are properties such as type, cotype, superreflexivity [...] The analogue of type in the geometry of metric spaces is [...] A simple metrical invariant replacing the notion of cotype was not yet discovered. \begin{flushright}{\small \em J.~Bourgain, 1986.}\end{flushright}}} Unfortunately, the survey of Lindenstrauss that is mentioned above never appeared. Nonetheless, Lindenstrauss had massive impact on this area as a leader who helped set the course of the Ribe program, as well as due to the important theorems that he proved in this direction. In fact, the article~\cite{JL84} of Johnson and Lindenstrauss, where the aforementioned metric dimension reduction lemma was proved, appeared a few years before~\cite{Bourgain-trees} and contained inspirational (even prophetic) ideas that had major subsequent impact on the Ribe program (including on Bourgain's works in this area). In the above quote, we removed the text describing "the analogue of type in the geometry of metric spaces" so as to not digress; it refers to the influential work of Bourgain, Milman and Wolfson~\cite{BMW} (see also the earlier work of Enflo~\cite{Enf78} and the subsequent work of Pisier~\cite{Pisier-type}). "Superreflexivity" was the main focus of~\cite{Bourgain-trees}, where the corresponding step of the Ribe program was completed (we will later discuss and use a refinement of this solution). An answer to the above mentioned question on cotype, which we will soon describe, was subsequently found by Mendel and the author~\cite{MN08-cotype}. We will next explain the terminology "finite-representability" in the above quote, so as to facilitate the ensuing discussion. \subsubsection{Finite representability} The first decades of work on the geometry of Banach spaces focused almost entirely on an inherently infinite dimensional theory. This was governed by Banach's partial ordering of Banach spaces~\cite[Chapter~7]{Ban32}, which declares that a Banach space $X$ has "linear dimension" at most that of a Banach space $Y$ if there exists $\alpha\ge 1$ such that $X$ is $\alpha$-isomorphic to a subspace of $Y$. In a remarkable feat of foresight, the following quote of Grothendieck~\cite{Gro53-dvo-conj} heralded the {\em local theory of Banach spaces}, by shifting attention to the geometry of the finite dimensional subspaces of a Banach space as a way to understand its global structure. \blockquote{{ \small\em assouplissons la notion de "dimension lin\'eaire" de Banach, en disant que l'espace norm\'e E a un {\em type lin\'eaire} inf\'erieur \`a celui d'un espace norm\'e F, si on peut trouver un $M>0$ fixe tel que tout sous-espace de dimension finie $E_1$ de E soit isomorphe "\`a M pr\`es" \`a un sous-espace $F_1$ de $F$ (i.e. il existe une application lin\'eaire biunivoque de $E_1$ sur $F_1$ , de norme $\le 1$, dont l\'application inverse a une norme $\le 1+M$). \begin{flushright}{\small \em A.~Grothendieck, 1953.}\end{flushright}}} Grothendieck's work in the 1950s exhibited astounding (technical and conceptual) ingenuity and insight that go well-beyond merely defining a key concept, as he did above. In particular, in~\cite{Gro53-dvo-conj} he conjectured an important phenomenon\footnote{This phenomenon was situated within the Ribe program by Bourgain, Figiel and Milman~\cite{BFM86}, and as such it eventually had ramifications to a well-studied (algorithmic) form of metric dimension reduction through its use to "compress" a finite metric space into a data structure called an {\em approximate distance oracle}~\cite{TZ05}. To date, the only known way to construct such a data structure with constant query time (and even, by now, conjecturally sharp approximation factor~\cite{Che14}) is via the nonlinear Dvoretzky theorem of~\cite{MN07-ramsey}, and thus through the Ribe program. For lack of space, we will not discuss this direction here; see the survey~\cite{Nao12}.} that was later proved by Dvoretzky~\cite{Dvo60} (see the discussion in~\cite{Sch17}), and his contributions in~\cite{Gro53} were transformative (e.g.~\cite{LP68,DFS08,KN12,Pis12}). The above definition set the stage for decades of (still ongoing) work on the local theory of Banach spaces which had major impact on a wide range of mathematical areas. The above "softening" of Banach's "linear dimension" is called today {\em finite representability}, following the terminology of James~\cite{Jam72} (and his important contributions on this topic). Given $\alpha\in [1,\infty)$, a Banach space $X$ is said to be $\alpha$-finitely representable in a Banach space $Y$ if for any $\beta>\alpha$, any finite dimensional subspace of $X$ is $\beta$-isomorphic to a subspace of $Y$ (in the notation of the above quote, $\beta=1+M$); $X$ is (crudely) finitely representable in $Y$ if there is some $\alpha\in [1,\infty)$ such that $X$ is $\alpha$-finitely representable in $Y$. This means that the finite dimensional subspaces of $X$ are not very different from subspaces of $Y$; if each of $X$ and $Y$ is finitely representable in the other, then this should be viewed as saying that $X$ and $Y$ have the same finite dimensional subspaces (up to a global allowable error that does not depend on the finite dimensional subspace in question). As an important example of the "taming power" of this definition, the {\em principle of local reflexivity} of Lindenstrauss and Rosenthal~\cite{LR69} asserts that even though sometimes $X^{**}\neq X$, it is always true that $X^{**}$ is $1$-finitely representable in $X$. Thus, while in infinite dimensions $X^{**}$ can be much larger than $X$, passing to the bidual cannot produce substantially new finite dimensional structures. The aforementioned Dvoretzky theorem~\cite{Dvo60} asserts that $\ell_2$ is $1$-finitely representable in any infinite dimensional Banach space. As another example of a landmark theorem on finite representability, Maurey and Pisier strengthened~\cite{MP-type-cotype} Theorem~\ref{thm:MP cotype} by showing that $\ell_{q_X}$ is $1$-finitely representable in any infinite dimensional Banach space $X$. Isomorphic local linear properties of Banach spaces are defined to be those properties that are preserved under finite representability. As an example, one should keep in mind finitary inequalities such as the cotype condition~\eqref{def:cotype}. The formal statement of Ribe's rigidity theorem~\cite{Ribe76} is \begin{theorem}\label{thm:ribe} Uniformly homeomorphic Banach spaces $X$ and $Y$ are finitely representable in each other. \end{theorem} The "immediate translation" of finite representability in the above quoted text from~\cite{Bourgain-trees} is to define that for $\alpha\in [1,\infty)$ a metric space $\mathcal{M}$ is $\alpha$-finitely representable in a metric space $\mathcal{N}$ if $\mathsf{c}_\mathcal{N}(\mathscr{C})\le \alpha$ for {\em every} finite subset $\mathscr{C}\subset \mathcal{M}$. By doing so one does not induce any terminological conflict, because one can show that a Banach space $X$ is (linearly) $\alpha$-finitely representable in a Banach space $Y$ if and only if $X$ is $\alpha$-finitely representable in $Y$ when $X$ and $Y$ are viewed as metric spaces. This statement follows from "soft" reasoning that is explained in~\cite{GNS12} (relying on a $w^*$-differentiation argument of Heinrich and Mankiewicz~\cite{HM82} as well as properties of ultrapowers of Banach spaces and the aforementioned principle of local reflexivity), though it also follows from Ribe's original proof of Theorem~\ref{thm:ribe} in~\cite{Ribe76}, and a different (quantitative) approach to this statement was obtained in~\cite{Bou87}. \subsubsection{Universality and dichotomies} Say that a metric space $\mathcal{M}$ is (finitarily) universal if there is $\alpha\ge 1$ such that $\mathsf{c}_\mathcal{M}(\mathscr{F})\le \alpha$ for {\em every} finite metric space $\mathscr{F}$. By~\cite{Mat92-ramsey,MN08-cotype}, this requirement holds for some $\alpha\ge 1$ if and only if it holds for $\alpha=1$ (almost-isometric embeddings), so the notion of universality turns out to be insensitive to the underlying distortion bound. Since for every $n\in \mathbb N$, any $n$-point metric space $(\mathscr{F}=\{x_1,\ldots,x_n\},d_\mathscr{F})$ is isometric to a subset of $\ell_\infty$ via the embedding $(x\in \mathscr{F})\mapsto (d_\mathscr{F}(x,x_i))_{i=1}^n$ (Fr\'echet embedding~\cite{Fre10}), a different way to state the notion of universality is to say that $\mathcal{M}$ is universal if $\ell_\infty$ is finitely representable in $\mathcal{M}$. Determining whether a given metric space is universal is a subtle matter. By Theorem~\ref{thm:MP cotype}, for a Banach space $X$ this is the same as asking to determine whether $q_X=\infty$. Such questions include major difficult issues in functional analysis that have been studied for a long time; as notable examples, see the works~\cite{Pel77,Bou84} on the (non)universality of the dual of the disc algebra, and the characterization~\cite{BM85} of Sidon subsets of the dual of a compact Abelian group $G$ in terms of the universality of their span is the space of continuous functions on $G$. Here are three famous concrete situations in which it is unknown if a certain specific space is universal. \begin{question}[Pisier's dichotomy problem] For each $n\in \mathbb N$ let $X_n$ be an arbitrary linear subspace of $\ell_\infty^n$ satisfying \begin{equation}\label{eq:Pisier dichotomy} \limsup_{n\to \infty}\frac{\dim(X_n)}{\log n}=\infty. \end{equation} Pisier conjectured~\cite{Pis81} that~\eqref{eq:Pisier dichotomy} forces the $\ell_2$ (Pythagorean) direct sum $(X_1\oplus X_2\oplus\ldots)_2$ to be universal. By duality, a positive answer to this question is equivalent to the following appealing statement on the geometry of polytopes. For $n\in \mathbb N$, suppose that $K\subset \mathbb R^n$ is an origin-symmetric polytope with $e^{o(n)}$ faces. Then, for each $\d>0$ there is $k=k(n,\d)\in \{1,\ldots,n\}$ with $\lim_{n\to \infty} k(n,\d)=\infty$, a subspace $F=F(n,\d)$ of $\mathbb R^n$ with $\dim(F)=k$ and a parallelepiped $Q\subset F$ (thus, $Q$ is an image of $[-1,1]^{k}$ under an invertible linear transformation) such that $Q\subset K\cap F\subset (1+\d)Q$. Hence, a positive answer to Pisier's dichotomy conjecture implies that every centrally symmetric polytope with $e^{o(n)}$ faces has a central section of dimension $k$ going to $\infty$ (as a function of the specific $o(n)$ dependence in the underlying assumption), which is $(1+\d)$-close to a polytope (a parallelepiped) with only $O(k)$ faces. The use of "dichotomy" in the name of this conjecture is due to the fact that this conclusion does not hold with $o(n)$ replaced by $O(n)$, as seen by considering polytopes that approximate the Euclidean ball. More generally, by the "isomorphic version" of the Dvoretzky theorem due to Milman and Schechtman~\cite{MS95}, for every sequence of normed spaces $\{Y_n\}_{n=1}^\infty$ with $\dim(Y_n)=n$ (not only $Y_n=\ell_\infty^n$, which is the case of interest above), and every $k(n)\in\{1,\ldots,n\}$ with $k(n)=O(\log n)$, there is a subspace $X_n\subset Y_n$ with $\dim(X_n)=k(n)$ such that the space $(X_1\oplus X_2\oplus\ldots)_2$ is isomorphic to a Hilbert space, and hence in particular it is not universal. The best-known bound in Pisier's dichotomy conjecture appears in the forthcoming work of Schechtman and Tomczak-Jaegermann~\cite{ST18}, where it is shown that the desired conclusion does indeed hold true if~\eqref{eq:Pisier dichotomy} is replaced by the stronger assumption $\limsup_{n\to\infty} \dim(X_n)/((\log n)^2(\log\log n)^2)=\infty$; this is achieved~\cite{ST18} by building on ideas of Bourgain~\cite{Bou84-pisier}, who obtained the same conclusion if $\limsup_{n\to\infty} \dim(X_n)/(\log n)^4>0$. Thus, due to~\cite{ST18} the above statement about almost-parallelepiped central sections of centrally symmetric polytopes does hold true if the initial polytope is assumed to have $\exp(o(\sqrt{n}/\log n))$ faces. \end{question} Prior to stating the next question on universality (which, apart from its intrinsic interest, plays a role in the ensuing discussion on metric dimension reduction), we need to very briefly recall some basic notation and terminology from optimal transport (see e.g.~\cite{AGS08,Vil09}). Suppose that $(\mathcal{M},d_\mathcal{M})$ is a separable complete metric space and fix $p\in [1,\infty)$. Denote by $\mathsf{P}_1(\mathcal{M})$ the set of all Borel probability measures $\mu$ on $\mathcal{M}$ of finite $p$'th moment, i.e., those Borel probability measure $\mu$ on $\mathcal{M}$ for which $\int_\mathcal{M} d_\mathcal{M}(x,y)^p\ud \mu(y)<\infty$ for all $x\in \mathcal{M}$. A probability measure $\pi\in \mathsf{P}_p(\mathcal{M}\times \mathcal{M})$ is a called a coupling of $\mu,\nu\in \mathsf{P}_p(\mathcal{M})$ if $\mu(A)=\pi(A\times\mathcal{M})$ and $\nu(A)=\pi(\mathcal{M}\times A)$ for every Borel measurable subset $A\subset \mathcal{M}$. The Wasserstein-$p$ distance between $\mu,\nu\in \mathsf{P}_p(\mathcal{M})$, denoted $\mathsf{W}_p(\mu,\nu)$, is defined to be the infimum of $(\iint_{\mathcal{M}\times \mathcal{M}} d_\mathcal{M}(x,y)^p\ud \pi(x,y))^{1/p}$ over all couplings $\pi$ of $\mu,\nu$. Below, $\mathsf{P}_p(\mathcal{M})$ is always assumed to be endowed with the metric $\mathsf{W}_p$. The following question is from~\cite{Bourgain-trees}. \begin{question}[Bourgain's universality problem]\label{Q:bourgain universality} Is $\mathsf{P}_1(\mathbb R^2)$ universal? This formulation may seem different from the way it is asked in~\cite{Bourgain-trees}, but, as explained~\cite[Section~1.5]{ANN15}, it is equivalent to it. More generally, is $\mathsf{P}_1(\mathbb R^k)$ universal for some integer $k\ge 2$ (it is simple to see that $\mathsf{P}_1(\mathbb R)$ is not universal)? In~\cite{Bourgain-trees} it was proved that $\mathsf{P}_1(\ell_1)$ is universal (see also the exposition in~\cite{Ost13}). So, it is important here that the underlying space is finite dimensional, though to the best of our knowledge it is also unknown whether $\mathsf{P}_1(\ell_2)$ is universal, or, for that matter, if $\mathsf{P}_1(\ell_p)$ is universal for any $p\in (1,\infty)$. See~\cite{ANN15} for a (sharp) universality property of $\mathsf{P}_p(\mathbb R^3)$ if $p\in (1,2]$. \end{question} For the following open question about universality (which will also play a role in the subsequent discussion on metric dimension reduction), recall the notion~\cite{Gro52,Gro55} of {\em projective tensor product} of Banach spaces. Given two Banach spaces $(X,\|\cdot\|_X)$ and $(Y,\|\cdot\|_Y)$, their projective tensor product $X\widehat{\otimes} Y$ is the completion of their algebraic tensor product $X\otimes Y$ under the norm whose unit ball is the convex hull of the simple tensors of vectors of norm at most $1$, i.e., the convex hull of the set $\{x\otimes y\in X\otimes Y:\ \|x\|_X,\|y\|_Y\le 1\}$. For example, $\ell_1\widehat{\otimes} X$ can be naturally identified with $\ell_1(X)$, and $\ell_2\widehat{\otimes} \ell_2$ can be naturally identified with Schatten--von Neumann trace class $\S_1$ (recall that for $p\in [1,\infty]$, the Schatten--von Neumann trace class $\S_p$ is the Banach space~\cite{vN37} of those compact linear operators $T:\ell_2\to \ell_2$ for which $\|T\|_{\S_p}=(\sum_{j=1}^\infty\mathsf{s}_j(T)^p)^{1/p}<\infty$, where $\{\mathsf{s}_j(T)\}_{j=1}$ are the singular values of $T$); see the monograph~\cite{Rya02} for much more on tensor products of Banach spaces. It is a longstanding endeavour in Banach space theory to understand which properties of Banach spaces are preserved under projective tensor products; see~\cite{DFS03,BNR12} and the references therein for more on this research direction. Deep work of Pisier~\cite{Pis83} shows that there exist two Banach spaces $X,Y$ that are not universal (even of cotype $2$) such that $X\widehat{\otimes} Y$ is universal. The following question was posed by Pisier in~\cite{Pis-92-clapem}. \begin{question}[universality of projective tensor products] Suppose that $p\in (1,2)$. Is $\ell_p\widehat{\otimes} \ell_2$ universal? We restricted the range of $p$ here because it is simple to check that $\ell_1\widehat{\otimes} \ell_2\cong\ell_1(\ell_2)$ is not universal, Tomczak-Jaegermann~\cite{Tom74} proved that $\ell_2\otimes \ell_2\cong \S_1$ is not universal, and Pisier proved~\cite{Pis92} that $\ell_p\widehat{\otimes} \ell_q$ is not universal when $p,q\in [2,\infty)$. It was also asked in~\cite{Pis-92-clapem} if $\ell_2\widehat{\otimes}\ell_2\widehat{\otimes}\ell_2$ is universal. The best currently available result in this direction (which will be used below) is that, using the local theory of Banach spaces and recent work on locally decodable codes, it was shown in~\cite{BNR12} that if $a,b,c\in (1,\infty)$ satisfy $\frac{1}{a}+\frac{1}{b}+\frac{1}{c}\le 1$, then $\ell_a\widehat{\otimes}\ell_b\widehat{\otimes}\ell_c$ is universal. \end{question} The following theorem is a union of several results of~\cite{MN08-cotype}. \begin{theorem}\label{thm:union MN} The following conditions are equivalent for a metric space $(\mathcal{M},d)$. \begin{itemize} \item $\mathcal{M}$ is not universal. \item There exists $q=q(\mathcal{M})\in (0,\infty)$ with the following property. For every $n\in \mathbb N$ there is $m=m(n,\mathcal{M},q)\in \mathbb N$ such that any collection of points $\{x_w\}_{w\in \mathbb Z_{2m}^n}$ in $\mathcal{M}$ satisfies the inequality \begin{equation}\label{eq:metric cotype} \sum_{i=1}^n \sum_{w\in \mathbb Z_{2m}^n} \frac{d(x_{w+me_i},x_w)^2}{m^2}\lesssim_{X,q} \frac{n^{1-\frac{2}{q}}}{3^n}\sum_{\varepsilon\in \{-1,0,1\}^n}\sum_{w\in \mathbb Z_{2m}^n}d(x_{w+\varepsilon},x)^2. \end{equation} Here $e_1,\ldots,e_n$ are the standard basis of $\mathbb R^n$ and addition (in the indices) is modulo $2m$. \item There is $\theta(\mathcal{M})\in (0,\infty)$ with the following property. For arbitrarily large $n\in \mathbb N$ there exists an $n$-point metric space $(\mathcal{B}_n,d_n)$ such that \begin{equation} \mathsf{c}_\mathcal{M}(\mathcal{B}_n)\gtrsim_\mathcal{M} (\log n)^{\theta(\mathcal{M})}. \end{equation} \end{itemize} Moreover, if we assume that $\mathcal{M}$ is a Banach space rather than an arbitrary metric space, then for any $q\in (0,\infty]$ the validity of~\eqref{eq:metric cotype} (as stated, i.e., for each $n\in \mathbb N$ there is $m\in \mathbb N$ for which~\eqref{eq:metric cotype} holds for every configuration $\{x_w\}_{w\in \mathbb Z_{2m}^n}$ of points in $\mathcal{M}$) is equivalent to~\eqref{def:cotype}, and hence, in particular, the infimum over those $q\in (0,\infty]$ for which~\eqref{eq:metric cotype} holds true is equal to the infimal cotype $q_\mathcal{M}$ of $\mathcal{M}$. \end{theorem} The final sentence of Theorem~\ref{thm:union MN} is an example of a successful step in the Ribe program, because it reformulates the local linear invariant~\eqref{def:cotype} in purely metric terms, namely as the quadratic geometric inequality~\eqref{eq:metric cotype} that imposes a restriction on the behavior of pairwise distances within any configuration of $(2m)^n$ points $\{x_w\}_{w\in \mathbb Z_{2m}^n}$ (indexed by the discrete torus $\mathbb Z_{2m}^n$) in the given Banach space. With this at hand, one can consider~\eqref{eq:metric cotype} to be a property of a metric space, while initially~\eqref{def:cotype} made sense only for a normed space. As in the case of~\eqref{def:cotype}, if $q=\infty$, then~\eqref{eq:metric cotype} holds in any metric space $(\mathcal{M},d_\mathcal{M})$ (for any $m\in \mathbb Z$, with the implicit constant in~\eqref{eq:metric cotype} being universal); by its general nature, such a statement must of course be nothing more than a formal consequence of the triangle inequality, as carried out in~\cite{MN08-cotype}. So, the validity of~\eqref{eq:metric cotype} for $q<\infty$ could be viewed as an asymptotic (randomized) enhancement of the triangle inequality in $(\mathcal{M},d_\mathcal{M})$; by considering the the canonical realization of $\mathbb Z_{2m}^n$ in $\mathbb C^n$, namely the points $\{(\exp(\pi iw_1/m),\ldots, \exp(\pi iw_n/m))\}_{w\in \mathbb Z_{{2m}}^n}$, equipped with the metric inherited from $\ell_\infty^n(\mathbb C)$, one checks that not every metric space satisfies this requirement. The equivalence of the first two bullet points in Theorem~\ref{thm:union MN} shows that once one knows that a metric space is not universal, one deduces the validity of such an enhancement of the triangle inequality. This is an analogue of Theorem~\ref{thm:MP cotype} of Maurey and Pisier for general metric spaces. The equivalence of the first and third bullet points in Theorem~\ref{thm:union MN} yields the following dichotomy. If one finds a finite metric space $\mathscr{F}$ such that $\mathsf{c}_{\mathscr{F}}(\mathcal{M})>1$, then there are arbitrary large finite metric spaces whose minimal distortion in $\mathcal{M}$ is at least a fixed positive power (depending on $\mathcal{M}$) of the logarithm their cardinality. Hence, for example, if every $n$-point metric space embeds into $\mathcal{M}$ with distortion $O(\log\log n)$, then actually for any $\d>0$, every finite metric space embeds into $\mathcal{M}$ with distortion $1+\d$. See~\cite{Men09,MN11-arxiv,MN13,ANN15} for more on metric dichotomies of this nature, as well as a quite delicate counterexample~\cite{MN13} for a natural variant for trees (originally asked by C. Fefferman). It remains a mystery~\cite{MN08-cotype} if the power of the logarithm $\theta(\mathcal{M})$ in Theorem~\ref{thm:union MN} could be bounded from below by a universal positive constant, as formulated in the following open question. \begin{question}[metric cotype dichotomy problem] Is there a universal constant $\theta>0$ such that in Theorem~\ref{thm:union MN} one could take $\theta(\mathcal{M})>\theta$. All examples that have been computed thus far leave the possibility that even $\theta(\mathcal{M})\ge 1$, which would be sharp (for $\mathcal{M}=\ell_2$) by Bourgain's embedding theorem~\cite{Bou85}. Note, however, that in~\cite{ANN15} it is asked whether for the Wasserstein space $\mathsf{P}_p(\mathbb R^3)$ we have $\liminf_{p\to 1} \theta(\mathsf{P}_p(\mathbb R^3))=0$. If this were true, then it would resolve the metric cotype dichotomy problem negatively. It would be interesting to understand the bi-Lipschitz structure of these spaces of measures on $\mathbb R^3$ regardless of this context, due to their independent importance. \end{question} Theorem~\ref{thm:union MN} is a good illustration of a "vanilla" accomplishment of the Ribe program, since it obtains a metric reformulation of a key isomorphic linear property of metric spaces, and also proves statements about general metric spaces which are inspired by the analogies with the linear theory that the Ribe program is aiming for. However, even in this particular setting of metric cotype, Theorem~\ref{thm:union MN} is only a part of the full picture, as it has additional purely metric ramifications. Most of these rely on a delicate issue that has been suppressed in the above statement of Theorem~\ref{thm:union MN}, namely that of understanding the asymptotic behavior of $m=m(n,\mathcal{M},q)$ in~\eqref{eq:metric cotype}. This matter is not yet fully resolved even when $\mathcal{M}$ is a Banach space~\cite{MN08-cotype,GNS11}, and generally such questions seem to be quite challenging (see~\cite{MN07-scaled,GN10,Nao16-riesz} for related issues). Thus far, whenever this question was answered for specific (classes of) metric spaces, it led to interesting geometric applications; e.g.~its resolution for certain Banach spaces in~\cite{MN08-cotype} was used in~\cite{Nao12-quasi} to answer a longstanding question~\cite{Vai99} about quasisymmetric embeddings, and its resolution for Alexandrov spaces of (global) nonpositive curvature~\cite{Ale51} (see e.g.~\cite{BGP92,Stu03} for the relevant background) in the forthcoming work~\cite{EMN17} is used there to answer a longstanding question about the coarse geometry of such Alexandrov spaces. \begin{comment} \begin{theorem} There exists a metric space $(\mathscr{B},d_\mathscr{B})$ that does not admit a coarse embedding into any Alexandorv space of global nonpositive curvature $(\mathcal{M},d_\mathcal{M})$, i.e., if $\omega,\Omega:[0,\infty)\to [0,\infty)$ are any two nondecreasing moduli with $\lim_{s\to \infty}\omega(s)=\infty$, then there does not exist $f:\mathscr{B}\to \mathcal{M}$ satisfying \begin{equation}\label{eq:coarse defined in theorem} \forall\, x,y\in \mathscr{B},\qquad \omega\big(d_\mathscr{B}(x,y)\big)\le d_\mathcal{M}\big(f(x),f(y)\big)\le \Omega\big(d_\mathscr{B}(x,y)\big). \end{equation} \end{theorem} \end{comment} \subsection{Metric dimension reduction}\label{sec:metric dim reduction intro} By its nature, many aspects of the local theory of Banach spaces involve describing phenomena that rely on dimension-dependent estimates. In the context of the Ribe program, the goal is to formulate/conjecture analogous phenomena for metric spaces, which is traditionally governed by asking Banach space-inspired questions about a finite metric space $(\mathcal{M},d_\mathcal{M})$ in which $\log |\mathcal{M}|$ serves as a replacement for the dimension. This analogy arises naturally also in the context of the bi-Lipschitz embedding problem into $\mathbb R^k$ (Problem~\ref{Q:bilip Rk}); see Remark~\ref{rem:doub log n} below. Early successful instances of this analogy can be found in the work of Marcus and Pisier~\cite{MP84}, as well as the aforementioned work of Johnson and Lindentrauss~\cite{JL84}. However, it should be stated at the outset that over the years it became clear that while making this analogy is the right way to get "on track" toward the discovery of fundamental metric phenomena, from the perspective of the Ribe program the reality is much more nuanced and, at times, even unexpected and surprising. \begin{comment} This endeavour/hope is well-described in the following quote~\cite{Mat96}. \blockquote{{ \small \em This investigation started in the context of the local Banach space theory, where the general idea was to obtain some analogs for general metric spaces of notions and results dealing with the structure of finite dimensional subspaces of Banach spaces. The distortion of a mapping should play the role of the norm of a linear operator, and the quantity $\log n$, where $n$ is the number of points in a metric space, would serve as an analog of the dimension of a normed space. Parts of this programme have been carried out by [...] \begin{flushright}{\small \em J.~Matou\v{s}ek, 1996.}\end{flushright}}} \end{comment} Johnson and Lindenstrauss asked~\cite[Problem~3]{JL84} whether every finite metric space $\mathcal{M}$ embeds with distortion $O(1)$ into some normed space $X_\mathcal{M}$ (which is allowed to depend on $\mathcal{M}$) of dimension $\dim(X_\mathcal{M})\lesssim \log |\mathcal{M}|$. In addition to arising from the above background, this question is motivated by~\cite[Problem~4]{JL84}, which asks if the Euclidean distortion of every finite metric space $\mathcal{M}$ satisfies $\mathsf{c}_2(\mathcal{M})\lesssim \sqrt{\log |\mathcal{M}|}$. If so, this would have served as a very satisfactory metric analogue of John's theorem~\cite{Joh48}, which asserts that any finite dimensional normed space $X$ is $\sqrt{\dim(X)}$-isomorphic to a subspace of $\ell_2$. Of course, John's theorem shows that a positive answer to the former question~\cite[Problem~3]{JL84} formally implies a positive answer to the latter question~\cite[Problem~4]{JL84}. The aforementioned Johnson--Lindenstrauss lemma~\cite{JL84} (JL lemma, in short) shows that, at least for finite subsets of a Hilbert space, the answer to the above stated~\cite[Problem~3]{JL84} is positive. \begin{theorem}[JL lemma]\label{thm:JL alpha} For each $n\in \mathbb N$ and $\alpha\in (1,\infty)$, there is $k\in \{1,\ldots,n\}$ with $k\lesssim_\alpha \log n$ such that any $n$-point subset of $\ell_2$ embeds into $\ell_2^k$ with distortion $\alpha$. \end{theorem} We postpone discussion of this fundamental geometric fact to Section~\ref{sec:JL} below, where it is examined in detail and its proof is presented. Beyond Hilbert spaces, there is only one other example (and variants thereof) of a Banach space for which it is currently known that~\cite[Problem~3]{JL84} has a positive answer for any of its finite subsets, as shown in the following theorem from~\cite{JN10}. \begin{theorem}\label{thm:tsirelson} There is a Banach space $\mathscr{T}^{(2)}$ which is not isomorphic to a Hilbert space yet it has the following property. For every finite subset $\mathscr{C}\subset \mathscr{T}^{(2)}$ there is $k\in \{1,\ldots,n\}$ with $k\lesssim\log |\mathscr{C}|$ and a $k$-dimensional linear subspace $F$ of $\mathscr{T}^{(2)}$ such that $\mathscr{C}$ embeds into $F$ with $O(1)$ distortion. \end{theorem} The space $\mathscr{T}^{(2)}$ of Theorem~\ref{thm:tsirelson} is not very quick to describe, so we refer to~\cite{JN10} for the details. It suffices to say here that this space is the $2$-convexification of the classical Tsirelson space~\cite{Cir74,FJ74}, and that the proof that it satisfies the stated dimension reduction result is obtained in~\cite{JN10} via a concatenation of several (substantial) structural results in the literature; see Section~4 in~\cite{JN10} for a discussion of variants of this construction, as well as related open questions. The space $\mathscr{T}^{(2)}$ of Theorem~\ref{thm:tsirelson} is not isomorphic to a Hilbert space, but barely so: it is explained in~\cite{JN10} that for every $n\in \mathbb N$ there exists an $n$-dimensional subspace $F_n$ of $\mathscr{T}^{(2)}$ with $\mathsf{c}_2(F_n)\ge e^{c\mathsf{Ack}^{-1}(n)}$, where $c>0$ is a universal constant and $\mathsf{Ack}^{-1}(\cdot)$ is the inverse of the Ackermann function from computability theory (see e.g.~\cite[Appendix~B]{AKNSS08}). So, indeed $\lim_{n\to \infty} \mathsf{c}_2(F_n)=\infty$, but at a tremendously slow rate. Remarkably, despite major scrutiny for over $3$ decades, it remains unknown if~\cite[Problem~3]{JL84} has a positive answer for subsets of {\em any} non-universal classical Banach space. In particular, the following question is open. \begin{question}\label{Q:lp into some space} Suppose that $p\in [1,\infty)\setminus\{2\}$. Are there $\alpha=\alpha(p),\beta=\beta(p)\in [1,\infty)$ such that for any $n\in \mathbb N$, every $n$-point subset of $\ell_p$ embeds with distortion $\alpha$ into some $k$-dimensional normed space with $k\le \beta\log n$? \end{question} It is even open if in Question~\ref{Q:lp into some space} one could obtain a bound of $k=o(n)$ for any fixed $p\in [1,\infty)\setminus\{2\}$. Using John's theorem as above, a positive answer to Question~\ref{Q:lp into some space} would imply that $\mathsf{c}_2(\mathscr{C})\lesssim_p \sqrt{\log |\mathscr{C}|}$ for any finite subset $\mathscr{C}$ of $\ell_p$. At present, such an embedding statement is not known for {\em any} $p\in [1,\infty)\setminus\{2\}$, though for $p\in [1,2]$ it is known~\cite{ALN08} that any $n$-point subset of $\ell_p$ embeds into $\ell_2$ with distortion $(\log n)^{1/2+o(1)}$; it would be interesting to obtain any $o(\log n)$ bound here for any fixed $p\in (2,\infty)$, which would be a "nontrivial" asymptotic behavior in light of the following general theorem~\cite{Bou85}. \begin{theorem}[Bourgain's embedding theorem]\label{thm:bourgain embedding} $\mathsf{c}_2(\mathcal{M})\lesssim \log |\mathcal{M}|$ for every finite metric space $\mathcal{M}$. \end{theorem} The above questions from~\cite{JL84} were the motivation for the influential work~\cite{Bou85}, where Theorem~\ref{thm:bourgain embedding} was proved. Using a probabilistic construction and the JL lemma, it was shown in~\cite{Bou85} that Theorem~\ref{thm:bourgain embedding} is almost sharp in the sense that there are arbitrarily large $n$-point metric spaces $\mathcal{M}_n$ for which $\mathsf{c}_2(\mathcal{M}_n)\gtrsim (\log n)/\log\log n$. By John's theorem, for every $\alpha\ge 1$, if $X$ is a finite dimensional normed space and $\mathsf{c}_X(\mathcal{M}_n)\le \alpha$, then $\mathsf{c}_2(\mathcal{M}_n)\le \alpha\sqrt{\dim(X)}$. Therefore the above lower bound on $\mathsf{c}_2(\mathcal{M}_n)$ implies that $\dim(X)\gtrsim (\log n)^2/(\alpha^2(\log\log n)^2)$. The achievement of~\cite{Bou85} is thus twofold. Firstly, it discovered Theorem~\ref{thm:bourgain embedding} (via the introduction of an influential randomized embedding method), which is the "correct" metric version of John's theorem in the Ribe program. The reality turned out to be more nuanced in the sense that the answer is not quite as good as the $O(\sqrt{\log n})$ that was predicted in~\cite{JL84}, but the $O(\log n)$ of Theorem~\ref{thm:bourgain embedding} is still a strong and useful phenomenon that was discovered through the analogy that the Ribe program provided. Secondly, we saw above that~\cite[Problem~3]{JL84} was {\em disproved} in~\cite{Bou85}, though the "bad news" that follows from~\cite{Bou85} is only mildly worse than the $O(\log n)$ dimension bound that~\cite[Problem~3]{JL84} predicted, namely a dimension lower bound that grows quite slowly, not faster than $(\log n)^{O(1)}$. Curiously, the very availability of strong dimension reduction in $\ell_2$ through the JL lemma is what was harnessed in~\cite{Bou85} to deduce that any "host normed space" that contains $\mathcal{M}_n$ with $O(1)$ distortion must have dimension at least of order $(\log n/\log\log n)^2\gg \log n$. Naturally, in light of these developments, the question of understanding what is the correct asymptotic behavior of the smallest $k(n)\in \mathbb N$ such that any $n$-point metric space embeds with distortion $O(1)$ into a $k(n)$-dimensional normed space was raised in~\cite{Bou85}. In order to proceed, it would be convenient to introduce some notation and terminology. \begin{definition}[metric dimension reduction modulus]\label{def:modulus bilip} Fix $n\in \mathbb N$ and $\alpha\in [1,\infty)$. Suppose that $(X,\|\cdot\|_X)$ is a normed space. Denote by $\mathsf{k}_n^\alpha(X)$ the minimum $k\in \mathbb N$ such that for {\em any} $\mathscr{C}\subset X$ with $|\mathscr{C}|=n$ there exists a $k$-dimensional linear subspace $F_\mathscr{C}$ of $X$ into which $\mathscr{C}$ embeds with distortion $\alpha$. \end{definition} The quantity $\mathsf{k}_n^\alpha(\ell_\infty)$ was introduced by Bourgain~\cite{Bou85} under the notation $\psi_\alpha(n)= \mathsf{k}_n^\alpha(\ell_\infty)$; see also~\cite{AR92,Mat96} where this different notation persists, though for the sake of uniformity of the ensuing discussion we prefer not to use it here because we will treat $X\neq\ell_\infty$ extensively. \cite{Bou85} focused for concreteness on the arbitrary value $\alpha=2$, and asked for the asymptotic behavior $\mathsf{k}_n^2(\ell_\infty)$ as $n\to \infty$. An $n$-point subset of $X=\ell_\infty$ is nothing more than a general $n$-point metric space, via the aforementioned isometric Fr\'echet embedding. In the same vein, a $k$-dimensional linear subspace of $\ell_\infty$ is nothing more than a general $k$-dimensional normed space $(F,\|\cdot\|_F)$ via the linear isometric embedding $(x\in F)\mapsto (x_i^*(x))_{i=1}^\infty$, where $\{x_i^*\}_{i=1}^\infty$ is an arbitrary sequence of linear functionals on $F$ that are dense in the unit sphere of the dual space $F^*$. Thus, the quantity $\mathsf{k}_n^\alpha(\ell_\infty)$ is the smallest $k\in \mathbb N$ such that every $n$-point metric space $(\mathcal{M},d_\mathcal{M})$ can be realized with distortion at most $\alpha$ as a subset of $(\mathbb R^k,\|\cdot\|_\mathcal{M})$ for some norm $\|\cdot\|_\mathcal{M}:\mathbb R^k\to \mathbb R^k$ on $\mathbb R^k$ (which, importantly, is allowed to be adapted to the initial metric space $\mathcal{M}$), i.e., $\mathsf{c}_{(\mathbb R^k,\|\cdot\|_\mathcal{M})}(\mathcal{M},d_\mathcal{M})\le \alpha$. This is precisely the quantity that~\cite[Problem~3]{JL84} asks about, and above we have seen that~\cite{Bou85} gives the lower bound \begin{equation}\label{eq:bourgain displayed log log} \mathsf{k}_n^\alpha(\ell_\infty)\gtrsim \left(\frac{\log n}{\alpha\log\log n}\right)^2. \end{equation} Theorem~\ref{thm:summary} below is a summary of the main nontrivial\footnote{Trivially $\mathsf{k}_n^\alpha(X)\le n-1$, by considering in Definition~\ref{def:modulus bilip} the subspace $F_\mathscr{C}=\mathrm{\bf span}(\mathscr{C}-x_0)$ for any fixed $x_0\in \mathscr{C}$.} bounds on the modulus $\mathsf{k}_n^\alpha(X)$ that are currently known for specific Banach spaces $X$. Since such a "combined statement" contains a large amount information and covers a lot of the literature on this topic, we suggest reading it in tandem with the subsequent discussion, which includes further clarifications and explanation of the history of the respective results. A "take home" message from the statements below is that despite major efforts by many researchers, apart from information on metric dimension reduction for $\ell_2$, the space $\mathscr{T}^{(2)}$ of Theorem~\ref{thm:tsirelson}, $\ell_\infty$, $\ell_1$ and $\S_1$, nothing is known for other spaces (even for $\ell_1$ and $\S_1$ more remains to be done, notably with respect to bounding $\mathsf{k}_n^\alpha(\ell_1),\mathsf{k}_n^\alpha(\S_1)$ from above). \begin{theorem}[summary of the currently known upper and lower bounds on metric dimension reduction]\label{thm:summary} There exist universal constants $c,C>0$ such that the following assertions hold true for every integer $n\ge 20$. \begin{enumerate} \item In the Hilbertian setting, we have the sharp bounds \begin{equation}\label{eq:JL in theporem} \forall\, \alpha\ge 1+\frac{1}{\sqrt[3]{n}},\qquad \mathsf{k}_n^\alpha(\ell_2)\asymp \frac{\log n}{\log(1+(\alpha-1)^2)}\asymp \max\left\{\frac{\log n}{(\alpha-1)^2},\frac{\log n}{\log \alpha}\right\}. \end{equation} \item For the space $\mathscr{T}^{(2)}$ of Theorem~\ref{thm:tsirelson}, there exists $\alpha_0\in [1,\infty)$ for which $\mathsf{k}_n^{\alpha_0}(\mathscr{T}^{(2)})\asymp \log n$. \item For $\ell_\infty$, namely in the setting of~\cite[Problem~3]{JL84}, we have \begin{equation}\label{eq:alpha<2} \forall\, \alpha\in [1,2),\qquad \mathsf{k}_n^\alpha(\ell_\infty)\asymp n, \end{equation} and \begin{equation}\label{eq:linfty bounds in theorem} \forall\, \alpha\ge 2,\qquad n^{\frac{c}{\alpha}}+\frac{\log n}{\log\left(\frac{\log n}{\alpha}+\frac{\alpha\log\log n}{\log n}\right)}\lesssim \mathsf{k}_n^\alpha(\ell_\infty)\lesssim \frac{n^{\frac{C}{\alpha}}\log n}{\log\left(1+\frac{\alpha}{\log n}\right)}. \end{equation} \item For $\ell_1$, we have \begin{equation}\label{eq:ell1 in theorem} \forall\, \alpha\ge 1,\qquad n^{\frac{c}{\alpha^2}}+\frac{\log n}{\log(\alpha+1)}\lesssim \mathsf{k}_{n}^\alpha(\ell_1)\lesssim \frac{n}{\alpha}. \end{equation} Moreover, if $\alpha\ge 2C\sqrt{\log n}\log\log n$, then we have the better upper bound \begin{equation}\label{eq:use ALN} \mathsf{k}_n^\alpha(\ell_1)\lesssim \frac{\log n}{\log\left(\frac{\alpha}{C\sqrt{\log n}\log\log n}\right)}. \end{equation} \item For the Schatten--von Neumann trace class $\S_1$, we have \begin{equation}\label{eq:S1 in theorem} \forall\, \alpha\ge 1,\qquad \mathsf{k}_n^\alpha(\S_1)\gtrsim n^{\frac{c}{\alpha^2}}+\frac{\log n}{\log(\alpha+1)}. \end{equation} \end{enumerate} \end{theorem} The bound $\mathsf{k}_n^\alpha(\ell_2)\lesssim (\log n)/\log(1+(\alpha-1)^2)$ in~\eqref{eq:JL in theporem} restates Theorem~\ref{thm:JL alpha} (the JL lemma) with the implicit dependence on $\alpha$ now stated explicitly; it actually holds for every $\alpha> 1$, as follows from the original proof in~\cite{JL84} and explained in Section~\ref{sec:JL} below. The restriction $\alpha\ge 1+1/\sqrt[3]{n}$ in~\eqref{eq:JL in theporem} pertains only to the corresponding lower bound on $\mathsf{k}_n^\alpha(\ell_2)$, which exhibits different behaviors in the low-distortion and high-distortion regimes. Despite scrutiny of many researchers over the past 3 decades, only very recently the dependence on $\alpha$ in the JL lemma when $\alpha$ is arbitrarily close to $1$ but independent of $n$ was proved to be sharp by Larsen and Nelson~\cite{GN17} (see~\cite{Alo03,Alo09,GN16} for earlier results in this direction, as well as the subsequent work~\cite{AK17}). This is so even when $\alpha$ is allowed to tend to $1$ with $n$, and even in a somewhat larger range than the requirement $\alpha\ge 1+1/\sqrt[3]{n}$ in~\eqref{eq:JL in theporem} (see~\cite{GN17} for the details), though there remains a small range of values of $\alpha$ ($n$-dependent, very close to $1$) for which it isn't currently known what is the behavior of $\mathsf{k}_n^\alpha(\ell_2)$. The present article is focused on embeddings that permit large errors, and in particular in ways to prove impossibility results even if large errors are allowed. For this reason, we will not describe here the ideas of the proof in~\cite{GN17} that pertains to the almost-isometric regime. For, say, $\alpha\ge 2$, it is much simpler to see that the $\mathsf{k}_n^\alpha(\ell_2)\gtrsim (\log n)/\log\alpha$, in even greater generality that also explains the appearance of the term $(\log n)/\log(\alpha+1)$ in~\eqref{eq:ell1 in theorem} and~\eqref{eq:S1 in theorem}. One could naturally generalize Definition~\ref{def:modulus bilip} so as to introduce the following notation for {\em relative metric dimension reduction moduli}. Let $\mathcal{F}$ be a family of metric spaces and $\mathcal{Y}$ be a family of normed spaces. For $n\in \mathbb N$ and $\alpha\in [1,\infty)$, denote by $\mathsf{k}_n^\alpha(\mathcal{F},\mathcal{Y})$ the minimum $k\in \mathbb N$ such that for every $\mathcal{M}\in \mathcal{F}$ with $|\mathcal{M}|=n$ there exists $Y\in \mathcal{Y}$ with $\dim(Y)=k$ such that $\mathsf{c}_Y(\mathcal{M})\le \alpha$. When $\mathcal{F}$ is the collection of all the finite subsets of a fixed Banach space $X$, and $\mathcal{Y}$ is the collection of all the finite-dimensional linear subspaces of a fixed Banach space $Y$, we use the simpler notation $\mathsf{k}_n^\alpha(\mathcal{F},\mathcal{Y})=\mathsf{k}_n^\alpha(X,Y)$. Thus, the modulus $\mathsf{k}_n^\alpha(X)$ of Definition~\ref{def:modulus bilip} coincides with $\mathsf{k}_n^\alpha(X,X)$. Also, under this notation Question~\eqref{Q:lp into some space} asks if for $p\in [1,\infty)\setminus \{2\}$ we have $\mathsf{k}_n^\alpha(\ell_p,\ell_\infty)\lesssim_p \log n$ for some $1\le \alpha\lesssim_p 1$. The study of the modulus $\mathsf{k}_n^\alpha(\mathcal{F},\mathcal{Y})$ is essentially a completely unexplored area, partially because even our understanding of the "vanilla" dimension reduction modulus $\mathsf{k}_n^\alpha(X)$ is currently very limited. By a short volumetric argument that is presented in Section~\ref{sec:JL} below, every infinite dimensional Banach space $X$ satisfies \begin{equation}\label{lower brunnminkowksi intro} \forall (n,\alpha)\in \mathbb N\times [1,\infty),\qquad \mathsf{k}_n^\alpha(X,\ell_\infty)\ge \frac{\log n}{\log(\alpha+1)}. \end{equation} Hence also $\mathsf{k}_n^\alpha(X)\ge (\log n)/\log(\alpha+1)$, since~\eqref{lower brunnminkowksi intro} rules out embeddings into any normed space of dimension less than $(\log n)/\log(\alpha+1)$, rather than only into such spaces that are also subspaces of $X$. Using an elegant Fourier-analytic argument, Arias-de-Reyna and Rodr\'iguez-Piazza proved in~\cite{AR92} that for every $\alpha\in [1,2)$ we have $\mathsf{k}_n^\alpha(\ell_\infty)\gtrsim (2-\alpha)n$. This was slightly improved by Matou\v{s}ek~\cite{Mat96} to $\mathsf{k}_n^\alpha(\ell_\infty)\gtrsim n$, i.e., he showed that the constant multiple of $n$ actually remains bounded below by a positive constant as $\alpha\to 2^-$ (curiously, the asymptotic behavior of $\mathsf{k}_n^2(\ell_\infty)$ remains unknown). These results establish~\eqref{eq:alpha<2}. So, for sufficiently small distortions one cannot hope to embed every $n$-point metric space into some normed space of dimension $o(n)$. For larger distortions (our main interest), it was conjectured in~\cite{AR92} that $\mathsf{k}_n^\alpha(\ell_\infty)\lesssim (\log n)^{O(1)}$ if $\alpha>2$. The bounds in~\eqref{eq:linfty bounds in theorem} refute this conjecture of~\cite{AR92}, since they include the lower bound $\mathsf{k}_n^\alpha(\ell_\infty)\gtrsim n^{c/\alpha}$, which is a landmark achievement of Matou\v{s}ek~\cite{Mat96} (obtained a decade after Bourgain asked about the asymptotics here and over a decade after Johnson and Lindenstrauss posed the question whether $\mathsf{k}_n^\alpha(\ell_\infty)\lesssim_\alpha\log n$). It is, of course, an exponential improvement over Bourgain's bound~\eqref{eq:bourgain displayed log log}. Actually, in the intervening period Linial, London and Rabinovich~\cite{LLR} removed the iterated logarithm in the lower bound of~\cite{Bou85} by showing that Theorem~\ref{thm:bourgain embedding} (Bourgain's embedding) is sharp up to the value of the implicit universal constant. By the same reasoning as above (using John's theorem), this also removed the iterated logarithm from the denominator in~\eqref{eq:bourgain displayed log log}, i.e., \cite{LLR} established that $\mathsf{k}_n^\alpha(\ell_\infty)\gtrsim (\log n)^2/\alpha^2$. This was the best-known bound prior to~\cite{Mat96}. Beyond proving a fundamental geometric theorem, which, as seen in~\eqref{eq:linfty bounds in theorem}, is optimal up to the constant in the exponent, this work of Matou\v{s}ek is important because it injected a refreshing approach from real algebraic geometry into this area, which was previously governed by considerations from analysis, geometry, probability and combinatorics. Section~\ref{sec:matousek} covers this outstanding contribution in detail, and obtains the following stronger statement that wasn't previously noticed in the literature but follows from an adaptation of Matou\v{s}ek's ideas. \begin{theorem}[impossibility of coarse dimension reduction]\label{thm:coarse matousek} There is a universal constant $c\in (0,\infty)$ with the following property. Suppose that $\omega,\Omega:[0,\infty)\to [0,\infty)$ are increasing functions that satisfy $\omega(s)\le \Omega(s)$ for all $s\in [0,\infty)$, as well as $\lim_{s\to\infty}\omega(s)=\infty$. Define \begin{equation}\label{eq:def beta} \beta(\omega,\Omega)\stackrel{\mathrm{def}}{=} \sup_{s\in (0,\infty)} \frac{s}{\omega^{-1}\big(2\Omega(s)\big)}\in (0,1). \end{equation} For arbitrarily large $n\in \mathbb N$ there is a metric space $(\mathcal{M},d_{\mathcal{M}})=\left(\mathcal{M}(n,\omega,\Omega),d_{\mathcal{M}(n,\omega,\Omega)}\right)$ with $|\mathcal{M}|=3n$ such that for any normed space $(X,\|\cdot\|_X)$, if there exists $f:\mathcal{M}\to X$ which satisfies \begin{equation}\label{eq:coarse condition} \forall\,x,y\in \mathcal{M},\qquad \omega\big(d_\mathcal{M}(x,y)\big)\le \|f(x)-f(y)\|_X\le \Omega\big(d_\mathcal{M}(x,y)\big), \end{equation} then necessarily \begin{equation}\label{eq:dim lower coarse} \dim(X)\gtrsim n^{c\beta(\omega,\Omega)}. \end{equation} \end{theorem} A mapping that satisfies~\eqref{eq:coarse condition} is called a {\em coarse embedding} (with moduli $\omega,\Omega$), as introduced in Gromov's seminal work~\cite{Gro93} and studied extensively ever since, with a variety of interesting applications (see the monographs~\cite{Roe03,NY12,Ost13} and the references therein). The bi-Lipschitz requirement~\eqref{eq:def embedding} corresponds to $\omega(s)=\tau s$ and $\Omega(s)=\alpha\tau s$ in~\eqref{eq:coarse condition}, in which case~\eqref{eq:dim lower coarse} becomes Matou\v{s}ek's aforementioned lower bound on $\mathsf{k}_n^\alpha(\ell_\infty)$. Theorem~\ref{thm:coarse matousek} asserts that there exist arbitrarily large finite metric spaces that cannot be embedded even with a very weak (coarse) guarantee into any low-dimensional normed space, with the dimension of the host space being forced to be at least a power of their cardinality, which is exponentially larger than the logarithmic behavior that one would predict from the natural ball-covering requirement that is induced by low-dimensionality (see the discussion of the doubling condition in Section~\ref{sec:doubling}, as well as the proof of~\eqref{lower brunnminkowksi intro} in Section~\ref{sec:JL}). \begin{remark}\label{rem:snowflake} Consider the following special case of Theorem~\ref{thm:coarse matousek}. Fix $\theta\in (0,1]$ and let $(\mathcal{M},d_\mathcal{M})$ be a metric space. It is straightforward to check that $d_\mathcal{M}^\theta:\mathcal{M}\times \mathcal{M}\to [0,\infty)$ is also a metric on $\mathcal{M}$. The metric space $(\mathcal{M},d_\mathcal{M}^\theta)$ is commonly called the {\em $\theta$-snowflake} of $\mathcal{M}$ (in reference to the von Koch snowflake curve; see e.g.~\cite{DS97}) and it is denoted $\mathcal{M}^\theta$. Given $\alpha\ge 1$, the statement that $\mathcal{M}^\theta$ embeds with distortion $\alpha$ into a normed space $(X,\|\cdot\|_X)$ is the same as the requirement~\eqref{eq:coarse condition} with $\omega(s)=s^\theta$ and $\Omega(s)=\alpha s^\theta$. Hence, by Theorem~\ref{thm:coarse matousek} there exist arbitrarily large $n$-point metric spaces $\mathcal{M}_n=\mathcal{M}_n(\alpha,\theta)$ such that if $\mathcal{M}_n^\theta$ embeds with distortion $\alpha$ into some $k$-dimensional normed space, then $k\ge n^{c/(2\alpha)^{1/\theta}}$. Conversely, Remark~\ref{rem:snowflake of ell infty} below shows that for every $n\in \mathbb N$ and $\alpha>1$, the $\theta$-snowflake of any $n$-point metric space embeds with distortion $\alpha$ into a normed space $X$ with $\dim(X)\lesssim_{\alpha,\theta} n^{C/\alpha^{1/\theta}}$. So, the bound~\eqref{eq:dim lower coarse} of Theorem~\ref{thm:coarse matousek} is quite sharp even for embeddings that are not bi-Lipschitz, though we did not investigate the extent of its sharpness for more general moduli $\omega,\Omega:[0,\infty)\to [0,\infty)$. \end{remark} At this juncture, it is natural to complement the (coarse) strengthening in Theorem~\ref{thm:coarse matousek} of Matou\v{s}ek's bound $\mathsf{k}_n^\alpha(\ell_\infty)\ge n^{c/\alpha}$ by stating the following different type of strengthening, which we recently obtained in~\cite{Nao17}. \begin{theorem}[impossibility of average dimension reduction]\label{thm:average distortion} There is a universal constant $c\in (0,\infty)$ with the following property. For arbitrarily large $n\in \mathbb N$ there is an $n$-point metric space $(\mathcal{M},d_{\mathcal{M}})$ such that for any normed space $(X,\|\cdot\|_X)$ and any $\alpha\in [1,\infty)$, if there exists $f:\mathcal{M}\to X$ which satisfies $\|f(x)-f(y)\|_X\le\alpha d_\mathcal{M}(x,y)$ for all $x,y\in X$, yet $\frac{1}{n^2}\sum_{x,y\in X} \|f(x)-f(y)\|_X\ge \frac{1}{n^2}\sum_{x,y\in X} d_\mathcal{M}(x,y)$, then necessarily $\dim(X)\ge n^{c/\alpha}$. \end{theorem} An $n$-point metric space $\mathcal{M}$ as in Theorem~\ref{thm:average distortion} is intrinsically high dimensional even on average, in the sense that if one wishes to assign in an $\alpha$-Lipschitz manner to each point in $\mathcal{M}$ a vector in some normed space $X$ such that the average distance in the image is the same as the average distance in $\mathcal{M}$, then this forces the abient dimension to satisfy $\dim(X)\ge n^{c/\alpha}$. Prior to Theorem~\ref{thm:average distortion}, the best-known bound here was $\dim(X)\gtrsim (\log n)^2/\alpha^2$, namely the aforementioned lower bound $\mathsf{k}_n^\alpha(\ell_\infty)\gtrsim (\log n)^2/\alpha^2$ of Linial, London and Rabinovich~\cite{LLR} actually treated the above "average distortion" requirement rather than only the (pairwise) bi-Lipschitz requirement. \begin{remark} The significance of Theorem~\ref{thm:average distortion} will be discussed further in Section~\ref{sec:average} below; see also~\cite{ANNRW17,ANNRW18}. In Section~\ref{sec:average} we will present a new proof of Theorem~\ref{thm:average distortion} that is different from (though inspired by) its proof in~\cite{Nao17}. It suffices to say here that the proof of Theorem~\ref{thm:average distortion} is conceptually different from Matou\v{s}ek's approach~\cite{Mat96}. Namely, in contrast to the algebraic/topological argument of~\cite{Mat96}, the proof of Theorem~\ref{thm:average distortion} relies on the theory of nonlinear spectral gaps, which is also an outgrowth of the Ribe program; doing justice to this theory and its ramifications is beyond the scope of the present article (see~\cite{MN14} and the references therein), but the basics are recalled in Section~\ref{sec:average}. Importantly, the proof of Theorem~\ref{thm:average distortion} obtains a criterion for determining if a given metric space $\mathcal{M}$ satisfies its conclusion, namely $\mathcal{M}$ can be taken to be the shortest-path metric of any bounded degree graph with a spectral gap. This information is harnessed in the forthcoming work~\cite{ANNRW18} to imply that finite-dimensional normed spaces have a structural proprty (a new type of hierarchical partitioning scheme) which has implications to the design of efficient data structures for approximate nearest neighbor search, demonstrating that the omnipresent "curse of dimensionality" is to some extent absent from this fundamental algorithmic task. \end{remark} In the intervening period between Bourgain's work~\cite{Bou85} and Matousek's solution~\cite{Mat96}, the question of determining the asymptotic behavior of $\mathsf{k}_n^\alpha(\ell_\infty)$ was pursued by Johnson, Lindenstrauss and Schechtman, who proved in~\cite{JLS87} that $\mathsf{k}_n^\alpha(\ell_\infty)\lesssim_\alpha n^{C/\alpha}$ for some universal constant $C>0$. They demonstrated this by constructing for every $n$-point metric space $\mathcal{M}$ a normed space $X_\mathcal{M}$, which they (probabilistically) tailored to the given metric space $\mathcal{M}$, with $\dim(X_\mathcal{M})\lesssim_\alpha n^{C/\alpha}$ and such that $\mathcal{M}$ embeds into $X_\mathcal{M}$ with distortion $\alpha$. Subsequently, Matou\v{s}ek showed~\cite{Mat92} via a different argument that one could actually work here with $X_\mathcal{M}=\ell_\infty^k$ for $k\in \mathbb N$ satisfying $k\lesssim_\alpha n^{C/\alpha}$, i.e., in order to obtain this type of upper bound on the asymptotic behavior of $\mathsf{k}_n^\alpha(\ell_\infty)$ one does not need to adapt the target normed space to the metric space $\mathcal{M}$ that is being embedded. The implicit dependence on $\alpha$ here, as well as the constant $C$ in the exponent, were further improved in~\cite{Mat96}. For $\alpha=O((\log n)/\log\log n)$, the upper bound on $\mathsf{k}_n^\alpha(\ell_\infty)$ that appears in~\eqref{eq:linfty bounds in theorem} is that of~\cite{Mat96}, and for the remaining values of $\alpha$ it is due to a more recent improvement over~\cite{Mat96} by Abraham, Bartal and Neiman~\cite{ABN11} (specifically, the upper bound in~\eqref{eq:linfty bounds in theorem} is a combination of Theorem~5 and Theorem~6 of~\cite{ABN11}). \begin{remark}\label{rem:snowflake of ell infty} An advantage of the fact~\cite{Mat92} that one could take $X_\mathcal{M}=\ell_\infty^k$ rather than the more general normed space of~\cite{JLS87} is that it quickly implies the optimality of the lower bound from Remark~\ref{rem:snowflake} on dimension reduction of snowflakes. Fix $n\in \mathbb N$, $\alpha>1$ and $\theta\in (0,1]$. Denote $\d=\min\{\sqrt{\alpha}-1,1\}$, so that $\alpha\asymp \alpha/(1+\d)>1$. By~\cite{Mat92}, given an $n$-point metric space $(\mathcal{M},d_\mathcal{M})$ there is an integer $k\lesssim n^{c/\alpha^{1/\theta}}$ and $f=(f_1,\ldots,f_k):\mathcal{M}\to \mathbb R^k$ such that $$ \forall\, x,y\in \mathcal{M},\qquad d_\mathcal{M}(x,y)\le \|f(x)-f(y)\|_{\ell_\infty^k}\le \left(\frac{\alpha}{1+\d}\right)^{\frac{1}{\theta}}d_\mathcal{M}(x,y). $$ Hence (here it becomes useful that we are dealing with the $\ell_\infty^k$ norm, as it commutes with powering), $$ \forall\, x,y\in \mathcal{M},\qquad d_\mathcal{M}(x,y)^\theta\le \max_{i\in \k} |f_i(x)-f_i(y)|^\theta \le \frac{\alpha}{1+\d}d_\mathcal{M}(x,y)^\theta. $$ By works of Kahane~\cite{Kah81} and Talagrand~\cite{Tal92-helix}, there is $m=m(\d,\theta)$ and a mapping (a quasi-helix) $h:\mathbb R\to \mathbb R^m$ such that $|s-t|^\theta\le\|h(s)-h(t)\|_{\ell_\infty^m}\le (1+\d)|s-t|^\theta$ for all $s,t\in \mathbb R$. The mapping $$(x\in \mathcal{M})\mapsto \bigoplus_{i=1}^k h\circ f_i(x)\in \bigoplus_{i=1}^k \ell_\infty^m$$ is a distortion-$\alpha$ embedding of the $\theta$-snowflake $(\mathcal{M},d_\mathcal{M}^\theta)$ into a normed space of dimension $mk\lesssim_{\alpha,\theta} n^{c/\alpha^{1/\theta}}$. The implicit dependence on $\alpha,\theta$ that~\cite{Kah81,Tal92-helix} imply here is quite good, but likely not sharp as $\alpha\to 1^+$ when $\theta\neq\frac12$. \end{remark} Since the expressions in~\eqref{eq:linfty bounds in theorem} are somewhat involved, it is beneficial to restate them on a case-by-case basis as follows. For sufficiently large $\alpha$, we have a bound\footnote{One can alternatively justify the upper bound in~\eqref{eq:sharp for large alpha} (for sufficiently large $n$) by first using Theorem~\ref{thm:bourgain embedding} (Bourgain's embedding theorem) to embed an $n$-point metric space $\mathcal{M}$ into $\ell_2$ with distortion $A\log n$ for some universal constant $A\ge 1$, and then using Theorem~\ref{thm:JL alpha} (the JL lemma) with the dependence on the distortion as stated in~\eqref{eq:JL in theporem} to reduce the dimension of the image of $\mathcal{M}$ under Bourgain's embedding to $O((\log n)/\log(\alpha/(A\log n)))$ while incurring a further distortion of $\alpha/(A\log n)$, thus making the overall distortion be at most $\alpha$. The right hand side of~\eqref{eq:sharp for large alpha} is therefore in fact an upper bound on $\mathsf{k}_n^\alpha(\ell_\infty,\ell_2)$; see also Corollary~\ref{cor:nontrivial type}.} that is sharp up to universal constant factors. \begin{equation}\label{eq:sharp for large alpha} \alpha\ge (\log n)\log\log n\implies \mathsf{k}_n^\alpha(\ell_\infty)\asymp \frac{\log n}{\log\left(\frac{\alpha}{\log n}\right)}. \end{equation} For a range of smaller values of $\alpha$, including those $\alpha$ that do not tend to $\infty$ with $n$, we have \begin{equation}\label{eq:sharp for small alpha} 1\le \alpha\le \frac{\log n}{\log\log n}\implies n^{\frac{c}{\alpha}}\lesssim \mathsf{k}_n^\alpha(\ell_\infty)\lesssim n^{\frac{C}{\alpha}}. \end{equation} \eqref{eq:sharp for small alpha} satisfactorily shows that the asymptotic behavior of $\mathsf{k}_n^\alpha(\ell_\infty)$ is of power-type, but it is not as sharp as~\eqref{eq:sharp for large alpha}. We suspect that determining the correct exponent of $n$ in the power-type dependence of $\mathsf{k}_n^\alpha(\ell_\infty)$ would be challenging (there is indication~\cite{Mat96,Mat02}, partially assuming a positive answer to a difficult conjecture of Erd\H{o}s~\cite{Erd64,Bol78}, that this exponent has infinitely many jump discontinuities as a function of $\alpha$). In an intermediate range $(\log n)/\log\log n\lesssim \alpha\lesssim (\log n)\log\log n$ the bounds~\eqref{eq:linfty bounds in theorem} are less satisfactory. The case $\alpha\asymp \log n$, corresponding to the distortion in Bourgain's embedding theorem, is especially intriguing, with~\eqref{eq:linfty bounds in theorem} becoming \begin{equation}\label{eq:log log log} \frac{\log n}{\log\log \log n}\lesssim \mathsf{k}_n^{\Theta(\log n)}(\ell_\infty)\lesssim \log n. \end{equation} The first inequality in~\eqref{eq:log log log} has not been stated in the literature, and we justify it in Section~\ref{sec:average} below. A more natural lower bound here would be a constant multiple of $(\log n)/\log\log n$, as this corresponds to the volumetric restriction~\eqref{lower brunnminkowksi intro}, and moreover by the upper bound~\eqref{eq:JL in theporem} in the JL lemma we know that any $n$-point subset of a Hilbert space does in fact embed with distortion $\log n$ into $\ell_2^k$ with $k\lesssim (\log n)/\log\log n$. The triple logarithm in~\eqref{eq:log log log} is therefore quite intriguing/surprising, thus leading to the following open question. \begin{question}\label{Q:log n} Given an integer $n\ge 2$, what is the asymptotic behavior of the smallest $k=k_n\in \mathbb N$ for which any $n$-point metric space $\mathcal{M}$ embeds with distortion $O(\log n)$ into some $k$-dimensional normed space $X_\mathcal{M}$. \end{question} There is a dearth of available upper bounds on $\mathsf{k}_n^\alpha(\cdot)$, i.e., positive results establishing that metric dimension reduction is possible. This is especially striking in the case of $\mathsf{k}_n^\alpha(\ell_1)$, due to the importance of $\ell_1$ from the perspective of pure mathematics and algorithms. The upper bound on $\mathsf{k}_n^\alpha(\ell_1)$ in the large distortion regime~\eqref{eq:use ALN} follows from combining the Euclidean embedding of~\cite{ALN08} with the JL lemma. The only general dimension reduction result in $\ell_1$ that lowers the dimension below the trivial bound $\mathsf{k}_n^\alpha(\ell_1)\le n-1$ is the forthcoming work~\cite{AN17}, where the estimate $\mathsf{k}_n^\alpha(\ell_1)\lesssim n/\alpha$ in~\eqref{eq:ell1 in theorem} is obtained; even this modest statement requires effort (among other things, it relies on the sparsification method of Batson, Spielman and Srivastava~\cite{BSS12}). The bound $\mathsf{k}_n^\alpha(\ell_1)\ge n^{c/\alpha^2}$ in~\eqref{eq:ell1 in theorem} is a remarkable theorem of Brinkman and Charikar~\cite{BC05} which answered a question that was at the time open for many years. To avoid any possible confusion, it is important to note that~\cite{BC05} actually exhibits an $n$-point subset $\mathscr{C}_{\mathsf{BC}}$ of $\ell_1$ for which it is shown in~\cite{BC05} that if $\mathscr{C}_{\mathsf{BC}}$ embeds with distortion $\alpha$ into $\ell_1^k$, then necessarily $k\ge n^{c/\alpha^2}$. On the face of it, this seems weaker than~\eqref{eq:ell1 in theorem}, because the lower bound on $\mathsf{k}_n^\alpha(\ell_1)$ in~\eqref{eq:ell1 in theorem} requires showing that if $\mathscr{C}_{\mathsf{BC}}$ embeds into an {\em arbitrary} finite-dimensional linear subspace $F$ of $\ell_1$, then necessarily $\dim(F)\ge n^{c/\alpha^2}$. However, Talagrand proved~\cite{Tal90} that in this setting for every $\beta> 1$ the subspace $F$ embeds with distortion $\beta$ into $\ell_1^k$, where $k\lesssim_\beta \dim(F)\log \dim(F)$. From this, an application of the above stated result of~\cite{BC05} gives that $\dim(F)\log \dim(F)\gtrsim n^{c/\alpha^2}$, and so the lower bound in~\eqref{eq:ell1 in theorem} follows from the formulation in~\cite{BC05}. Satisfactory analogues of the above theorem of Talagrand are known~\cite{Sch87,BLM89,Tal95} (see also the survey~\cite{JS01} for more on this subtle issue) when $\ell_1$ is replaced by $\ell_p$ for some $p\in (1,\infty)$, but such reductions to "canonical" linear subspaces are not available elsewhere, so the above reasoning is a rare "luxury" and in general one must treat arbitrary low-dimensional linear subspaces of the Banach space in question. The above difficulty was overcome for $\mathsf{S}_1$ in~\cite{NPS18}, where~\eqref{eq:S1 in theorem} was proven. The similarity of the lower bounds in~\eqref{eq:ell1 in theorem} and~\eqref{eq:S1 in theorem} is not coincidental. One can view the Brinkman--Charikar example $\mathscr{C}_{\mathsf{BC}}\subset \ell_1$ also as a collection of diagonal matrices in $\S_1$, and~\cite{NPS18} treats this very same subset by strengthening the assertion of~\cite{BC05} that $\mathscr{C}_{\mathsf{BC}}$ does not well-embed into low-dimensional subspaces of $\mathsf{S_1}$ which consist entirely of diagonal matrices, to the same assertion for low-dimensional subspaces of $\mathsf{S}_1$ which are now allowed to consist of any matrices whatsoever. Using our notation for relative dimension reduction moduli, this gives the stronger assertion $\mathsf{k}_n^\alpha(\ell_1,\mathsf{S}_1)\ge n^{c/\alpha^2}$. A geometric challenge of the above discussion is that, even after one isolates a candidate $n$-point subset $\mathscr{C}$ of $\ell_1$ that is suspected not to be realizable with $O(1)$ distortion in low-dimensions (finding such a suspected intrinsically high-dimensional set is of course a major challenge in itself), one needs to devise a way to somehow argue that if one could find a configuration of $n$ points in a low-dimensional subspace $F$ of $\ell_1$ (or $\mathsf{S}_1$) whose pairwise distances are within a fixed, but potentially very large, factor $\alpha\ge 1$ of the corresponding pairwise distances within $\mathscr{C}$ itself, then this would force the ambient dimension $\dim(F)$ to be very large. In~\cite{BC05} this was achieved via a clever proof that relies on linear programming; see also~\cite{ACNNH11} for a variant of this linear programming approach in the almost isometric regime $\alpha\to 1^+$. In~\cite{Reg13} a different proof of the Brinkman--Charikar theorem was found, based on information-theoretic reasoning. Another entirely different geometric method to prove that theorem was devised in~\cite{LN04}; see also~\cite{LMN05,JS09} for more applications of the approach of~\cite{LN04}. Very recently, a further geometric approach was obtained in~\cite{NPS18}, where it was used to derive a stronger statement that, as shown in~\cite{NPS18}, cannot follow from the method of~\cite{LN04} (the statement is that the $n$-point subset $\mathscr{C}_{\mathsf{BC}}\subset \ell_1$ is not even an $\alpha$-Lipschitz quotient of any subset of a low dimensional subspace of $\mathsf{S}_1$; see~\cite{NPS18} for the relevant definition an a complete discussion). The approach of~\cite{NPS18} relies on an invariant that arose in the Ribe program and is called {\em Markov convexity}. Fix $q>0$. Let $\{\chi_t\}_{t\in \mathbb Z}$ be a Markov chain on a state space $\Omega$. Given an integer $k\ge 0$, denote by $\{\widetilde{\chi}_t(k)\}_{t\in \mathbb Z}$ the process that equals $\chi_t$ for time $t\le k$, and evolves independently of $\chi_t$ (with respect to the same transition probabilities) for time $t>k$. Following~\cite{LNP09}, the Markov $q$-convexity constant of a metric space $(\mathcal{M},d_\mathcal{M})$, denoted $\Pi_q(\mathcal{M})$, is the infimum over those $\Pi\in [0,\infty]$ such that for every Markov chain $\{\chi_t\}_{t\in \mathbb Z}$ on a state space $\Omega$ and every $f:\Omega\to \mathcal{M}$ we have \begin{equation*}\label{eq:def markov convexity} \bigg(\sum_{k=1}^\infty\sum_{t\in \mathbb Z}\frac{1}{2^{qk}}\mathbb{E}\left[d_\mathcal{M} \big(f\big(\widetilde{\chi}_t(t-2^k)\big),f(\chi_t)\big)^q\right]\bigg)^{\frac1{q}}\le \Pi\bigg(\sum_{t\in \mathbb Z} \mathbb{E}\Big[d_\mathcal{M}\big(f(\chi_t),f(\chi_{t-1})\big)^q\Big]\bigg)^{\frac1{q}}. \end{equation*} By~\cite{LNP09,MN13}, a Banach space $X$ satisfies $\Pi_q(X)<\infty$ if and only if it has an equivalent norm $|||\cdot|||:X\to [0,\infty)$ whose modulus of uniform convexity has power type $q$, i.e., $|||x+y|||\le 2-\Omega_X(|||x-y|||^q)$ for every $x,y\in X$ with $|||x|||=|||y|||=1$. This completes the step in the Ribe program which corresponds to the local linear property "$X$ admits an equivalent norm whose modulus of uniform convexity has power type $q$," and it is a refinement of the aforementioned characterization of superreflexivity in~\cite{Bourgain-trees} (which by deep results of~\cite{Enf72,Pisier-martingales} corresponds to the cruder local linear property "there is a finite $q\ge 2$ for which $X$ admits an equivalent norm whose modulus of uniform convexity has power type $q$"). By~\cite{MN13,EMN17-bifurcation}, the Brinkman--Charikar subset $\mathscr{C}_{\mathsf{BC}}\subset \ell_1$ (as well as a variant of it due to Laakso~\cite{Laakso} which has~\cite{LMN05} the same non-embeddability property into low-dimensional subspaces of $\ell_1$) satisfies $\Pi_q(\mathscr{C}_{\mathsf{BC}})\gtrsim (\log n)^{1/q}$ for every $q\ge 2$ (recall that in our notation $|\mathscr{C}_{\mathsf{BC}}|=n$). At the same time, it is proved in~\cite{NPS18} that $\Pi_2(F)\lesssim \sqrt{\log \dim(F)}$ for every finite dimensional subset of $\mathsf{S}_1$. It remains to contrast these asymptotic behaviors (for $q=2$) to deduce that if $\mathsf{c}_F(\mathscr{C}_{\mathsf{BC}})\le \alpha$, then necessarily $\dim(F)\ge n^{c/\alpha^2}$. Prior to the forthcoming work~\cite{NY17-3dim}, the set $\mathscr{C}_{\mathsf{BC}}$ (and variants thereof of a similar nature) was the only known example that demonstrates that there is no $\ell_1$ analogue of the JL-lemma. The following theorem is from~\cite{NY17-3dim}. \begin{theorem}\label{thm:heisenberg} There is a universal constant $c\in (0,\infty)$ with the following property. For arbitrarily large $n\in \mathbb N$ there exists an $n$-point $O(1)$-doubling subset $\mathscr{H}_n$ of $\ell_1$ satisfying $\mathsf{c}_4(\mathscr{H}_n)\lesssim 1$, such that for every $\alpha\in [1,\infty)$ and every finite-dimensional linear subspace $F$ of $\S_1$, if $\mathscr{H}_n$ embeds into $F$ with distortion $\alpha$, then necessarily \begin{equation}\label{eq:heisenberg lower} \dim(F)\ge \exp\Big(\frac{c}{\alpha^2}\sqrt{\log n}\Big). \end{equation} \end{theorem} See Section~\ref{sec:doubling} for the (standard) terminology "doubling" that is used in Theorem~\ref{thm:heisenberg}. While~\eqref{eq:heisenberg lower} is weaker than the lower bound of Brinkman--Charikar in terms of the dependence on $n$, it nevertheless rules out metric dimension reduction in $\ell_1$ (or $\mathsf{S}_1$) in which the target dimension is, say, a power of $\log n$. The example $\mathscr{H}_n$ of Theorem~\ref{thm:heisenberg} embeds with distortion $O(1)$ into $\ell_4$, and hence in particular $\sup_{n\in \mathbb N}\Pi_4(\mathscr{H}_n)\lesssim \Pi_4(\ell_4)<\infty$, by~\cite{LNP09}. This makes $\mathscr{H}_n$ be qualitatively different from all the previously known examples which exhibit the impossibility of metric dimension reduction in $\ell_1$, and as such its existence has further ramifications that answer longstanding questions; see~\cite{NY17-3dim} for a detailed discussion. The proof of Theorem~\ref{thm:heisenberg} is markedly different from (and more involved than) previous proofs of impossibility of dimension reduction in $\ell_1$, as it relies on new geometric input (a subtle property of the $3$-dimensional Heisenberg group which fails for the $5$-dimensional Heisenberg group) that is obtained in~\cite{NY17-3dim}, in combination with results from~\cite{ANT13,LN14-poincare,LN14,HN16}; full details appear in~\cite{NY17-3dim}. \subsection{Spaces admitting bi-Lipschitz and average metric dimension reduction}\label{sec:spaces} Say that an infinite dimensional Banach space $(X,\|\cdot\|_X)$ {\em admits metric dimension reduction} if there is $\alpha=\alpha_X\in [1,\infty)$ such that \begin{equation*}\label{eq:def admits} \lim_{n\to \infty} \frac{\log \mathsf{k}_n^\alpha(X)}{\log n}=0. \end{equation*} In other words, the requirement here is that for some $\alpha=\alpha_X\in [1,\infty)$ and every $n\in \mathbb N$, any $n$-point subset $\mathscr{C}\subset X$ embeds with (bi-Lipschitz) distortion $\alpha$ into some linear subspace $F$ of $X$ with $\dim(F)=n^{o_X(1)}$. Analogously, we say that $(X,\|\cdot\|_X)$ {\em admits average metric dimension reduction} if there is $\alpha=\alpha_X\in [1,\infty)$ such that for any $n\in \mathbb N$ there is $k_n=n^{o_X(1)}$, i.e., $\lim_{n\to\infty} (\log k_n)/\log n=0$, such that for any $n$-point subset $\mathscr{C}$ of $X$ there is a linear subspace $F$ of $X$ with $\dim(F)=k_n$ and a mapping $f:\mathscr{C}\to F$ which is $\alpha$-Lipschitz, i.e., $\|f(x)-f(y)\|_X\le \alpha \|x-y\|_X$ for all all $x,y\in \mathscr{C}$, yet \begin{equation}\label{eq:av dist requirement} \frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}}\|f(x)-f(y)\|_X\ge \frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}}\|x-y\|_X. \end{equation} Our choice here of the behavior $n^{o_X(1)}$ for the target dimension is partially motivated by the available results, based on which this type of asymptotic behavior appears to be a benchmark. We stress, however, that since the repertoire of spaces that are known to admit metric dimension reduction is currently very limited, finding any new setting in which one could prove that reducing dimension to $o_X(n)$ is possible would be a highly sought after achievement. In the same vein, finding new spaces for which one could prove a metric dimension reduction lower bound that tends to $\infty$ faster than $\log n$ (impossibility of a JL-style guarantee) would be very interesting. \begin{remark} In the above definition of spaces admitting average metric dimension reduction we imposed the requirement~\eqref{eq:av dist requirement} following the terminology that was introduced by Rabinovich in~\cite{Rab08}, and due to the algorithmic usefulness of this notion of embedding. However, one could also consider natural variants such as $(\frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}}\|f(x)-f(y)\|_X^p)^{1/p}\ge (\frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}}\|x-y\|_X^q)^{1/q}$ for any $p,q\in (0,\infty]$, and much of the ensuing discussion holds mutatis mutandis in this setting as well. \end{remark} The only Banach spaces that are currently known to admit metric dimension reduction are Hilbert spaces and the space $\mathscr{T}^{(2)}$ of Theorem~\ref{thm:tsirelson} (and variants thereof). These examples allow for the possibility that if $(X,\|\cdot\|_X)$ admits metric dimension reduction, i.e., $\mathsf{k}_n^{O_X(1)}(X)=n^{o_X(1)}$, then actually $\mathsf{k}_n^{O_X(1)}(X)=O_X(\log n)$. Such a dichotomy would of course be remarkable, but there is insufficient evidence to conjecture that this is so. The available repertoire of spaces that admit average metric dimension reduction is larger, since if $p\in [2,\infty)$, then $\ell_p$ and even $\mathsf{S}_p$ satisfy the assumption of the following theorem, by~\cite{Maz29} and~\cite{Ric15}, respectively. \begin{theorem}\label{thm:average in lp} Let $(X,\|\cdot\|_X)$ be an infinite dimensional Banach space with unit ball $B_X=\{x\in X:\ \|x\|_X\le 1\}$. Suppose that there is a Hilbert space $(H,\|\cdot\|_H)$ and a one-to-one mapping $f:B_X\to H$ such that $f$ is Lipschitz and $f^{-1}:f(B_X)\to X$ is uniformly continuous. Then $X$ admits average metric dimension reduction. In fact, this holds for embeddings into a subspace of logarithmic dimension, i.e., there is $\alpha=\alpha_X\in [1,\infty)$ such that for any $n\in \mathbb N$ and any $n$-point subset $\mathscr{C}$ of $X$ there is a linear subspace $F$ of $X$ with $\dim(F)\lesssim \log n$ and a mapping $f:\mathscr{C}\to F$ which satisfies both~\eqref{eq:av dist requirement} and $\|f(x)-f(y)\|_X\le \alpha\|x-y\|_X$ for all $x,y\in \mathscr{C}$. \end{theorem} \begin{proof}This statement is implicit in~\cite{Nao14}. By combining~\cite[Proposition~7.5]{Nao14} and~\cite[Lemma~7.6]{Nao14} there is a $O_X(1)$-Lipschitz mapping $f:\mathscr{C}\to \ell_2$ which satisfies $\frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}} \|f(x)-f(y)\|_{2}\ge \frac{1}{n^2}\sum_{x\in \mathscr{C}}\sum_{y\in \mathscr{C}} \|x-y\|_X$. By the JL lemma we may assume that $f$ actually takes values in $\ell_2^k$ for some $k\lesssim \log n$. Since $X$ is infinite dimensional, Dvoretzky's theorem~\cite{Dvo60} ensures that $\ell_2^k$ is $2$-isomorphic to a $k$-dimensional subspace $F$ of $X$. \end{proof} \begin{remark} By~\cite{Maz29}, for $p\in [2,\infty)$ the assumption of Theorem~\ref{thm:average in lp} holds for $X=\ell_p$. An inspection of the proofs in~\cite{Nao14} reveals that the dependence of the Lipschitz constant $\alpha=\alpha_p$ on $p$ that Theorem~\ref{thm:average in lp} provides for $X=\ell_p$ grows to $\infty$ exponentially with $p$. As argued in~\cite[Section~5.1]{Nao14} (using metric cotype), this exponential behavior is unavoidable using the above proof. However, in this special case a more sophisticated argument of~\cite{Nao14} yields $\alpha_p\lesssim p^{5/2}$; see equation (7.40) in~\cite{Nao14}. Motivated by~\cite[Corollary~1.6]{Nao14}, we conjecture that this could be improved to $\alpha_p\lesssim p$, and there is some indication (see~\cite[Lemma~1.11]{Nao14}) that this would be sharp. \end{remark} Prior to~\cite{Nao17}, it was not known if there exists a Banach space which fails to admit average metric dimension reduction. Now we know (Theorem~\ref{thm:average distortion}) that $\ell_\infty$ fails to admit average metric dimension reduction, and therefore also any universal Banach space fails to admit average metric dimension reduction. A fortiori, the same is true also for (non-average) metric dimension reduction, but this statement follows from the older work~\cite{Mat96}. Failure of average metric dimension reduction is not known for {\em any} non-universal (finite cotype) Banach space, and it would be very interesting to provide such an example. By~\cite{BC05,NPS18} we know that $\ell_1$ and $\mathsf{S}_1$ fail to admit metric dimension reduction, but this is not known for average distortion, thus leading to the following question. \begin{question}\label{Q:l1 average} Does $\ell_1$ admit average metric dimension reduction? Does $\ell_p$ have this property for any $p\in [1,2)$? \end{question} All of the available examples of $n$-point subsets of $\ell_1$ for which the $\ell_1$ analogue of the JL lemma fails (namely if $k=O(\log n)$, then they do not embed with $O(1)$ distortion into $\ell_1^{k}$) actually embed into the real line $\mathbb R$ with $O(1)$ average distortion; this follows from~\cite{Rab08}. Specifically, the examples in~\cite{BC05,LMN05} are the shortest-path metric on planar graphs, and the example in Theorem~\ref{thm:heisenberg} is $O(1)$-doubling, and both of these classes of metric spaces are covered by~\cite{Rab08}; see also~\cite[Section~7]{Nao14} for generalizations. Thus, the various known proofs which demonstrate that the available examples cannot be embedded into a low dimensional subspace of $\ell_1$ argue that any such low-dimensional embedding must highly distort some distance, but this is not so for a typical distance. A negative answer to Question~\ref{Q:l1 average} would therefore require a substantially new type of construction which exhibits a much more "diffuse" intrinsic high-dimensionality despite it being a subset of $\ell_1$. In the reverse direction, a positive answer to Question~\ref{Q:l1 average}, beyond its intrinsic geometric/structural interest, could have algorithmic applications. \subsubsection{Lack of stability under projective tensor products} Prior to the recent work~\cite{NPS18}, it was unknown whether the property of admitting metric dimension reduction is preserved under projective tensor products. \begin{corollary}\label{coro:S1} There exist Banach spaces $X,Y$ that admit metric dimension reduction yet $X\widehat{\otimes} Y$ does not. \end{corollary} Since $\S_1$ is isometric to $\ell_2\widehat{\otimes}\ell_2$ and~\cite{NPS18} establishes that $\S_1$ fails to admit metric dimension reduction, together with the JL lemma this implies Corollary~\ref{coro:S1} (we can thus even have $X=Y$ and $\mathsf{k}_n^\alpha(n)\lesssim_\alpha \log n$ for all $\alpha>1$). Since we do not know whether $\S_1$ admits average metric dimension reduction (the above comments pertaining to Question~\ref{Q:l1 average} are valid also for $\S_1$), the analogue of Corollary~\ref{coro:S1} for average metric dimension reduction was previously unknown. Here we note the following statement, whose proof is a somewhat curious argument. \begin{theorem}\label{thm:tensor average} There exist Banach spaces $X,Y$ that admit average metric dimension reduction yet $X\widehat{\otimes} Y$ does not. Moreover, for every $p\in (2,\infty)$ we can take here $X=\ell_p$. \end{theorem} \begin{proof} By~\cite{BNR12} (which relies on major input from the theory of locally decodable codes~\cite{Efr09} and an important inequality of Pisier~\cite{Pis80}), the $3$-fold product $\ell_3\widehat{\otimes}\ell_3\widehat{\otimes}\ell_3$ is universal. So, by the recent work~\cite{Nao17} (Theorem~\ref{thm:average distortion}), $\ell_3\widehat{\otimes}\ell_3\widehat{\otimes}\ell_3$ does not admit average metric dimension reduction. At the same time, by Theorem~\ref{thm:average in lp} we know that $\ell_3$ admits average metric dimension reduction. So, if $\ell_3\widehat{\otimes}\ell_3$ fails to admit average metric dimension reduction, then we can take $X=Y=\ell_3$ in Theorem~\ref{thm:tensor average}. Otherwise, $\ell_3\widehat{\otimes}\ell_3$ does admit average metric dimension reduction, in which case we can take $X=\ell_3$ and $Y=\ell_3\widehat{\otimes}\ell_3$. Thus, in either of the above two cases, the first assertion of Theorem~\ref{thm:tensor average} holds true. The second assertion of Theorem~\ref{thm:tensor average} follows by repeating this argument using the fact~\cite{BNR12} that $\ell_p\widehat{\otimes} \ell_p\widehat{\otimes} \ell_{q}$ is universal if $2/p+1/q\le 1$, or equivalently $q\ge p/(p-2)$. If we choose, say, $q=\max\{2,p/(p-2)\}$, then by Theorem~\ref{thm:average in lp} we know that both $\ell_p$ and $\ell_q$ admit average metric dimension reduction, while $\ell_p\widehat{\otimes} \ell_p\widehat{\otimes} \ell_{q}$ does not. So, the second assertion of Theorem~\ref{thm:tensor average} holds for either $Y=\ell_p$ or $Y=\ell_p\widehat{\otimes}\ell_q$. \end{proof} The proof of Theorem~\ref{thm:tensor average} establishes that at least one of the pairs $(X=\ell_3,Y=\ell_3)$ or $(X=\ell_3,Y=\ell_3\widehat{\otimes}\ell_3)$ satisfies its conclusion, but it gives no indication which of these two options occurs. This naturally leads to \begin{question}\label{Q:l3} Does $\ell_3\widehat{\otimes}\ell_3$ admit average metric dimension reduction? \end{question} A positive answer to Question~\ref{Q:l3} would yield a new space that admits average metric dimension reduction. In order to claim that $\ell_3\widehat{\otimes}\ell_3$ is indeed new in this context, one must show that it does not satisfy the assumption of Theorem~\ref{thm:average in lp}. This is so because $\S_1$ (hence also $\ell_1$) is finitely representable in $\ell_3\widehat{\otimes} \ell_3$; see e.g.~\cite[page~61]{DFS03}. The fact that no Banach space in which $\ell_1$ is finitely representably satisfies the assumption of Theorem~\ref{thm:average in lp} follows by combining~\cite[Lemma~1.12]{Nao14}, \cite[Proposition~7.5]{Nao14}, and~\cite[Lemma~7.6]{Nao14}. This also shows that a positive answer to Question~\ref{Q:l3} would imply that any $n$-point subset of $\ell_1$ (or $\S_1$) embeds with $O(1)$ average distortion into some normed space (a subspace of $\ell_3\widehat{\otimes}\ell_3$) of dimension $n^{o(1)}$, which is a statement in the spirit of Question~\ref{Q:l1 average}. If the answer to Question~\ref{Q:l3} were negative, then $\ell_3\widehat{\otimes}\ell_3$ would be the first example of a non-universal space which fails to admit average metric dimension reduction, because Pisier proved~\cite{Pis92,Pis-92-clapem} that $\ell_3\widehat{\otimes}\ell_3$ is not universal. Another question that arises naturally from Theorem~\ref{thm:tensor average} is whether its conclusion holds true also for $p=2$. \begin{question} Is there a Banach space $Y$ that admits average metric dimension reduction yet $\ell_2\widehat{\otimes} Y$ does not? \end{question} \subsubsection{Wasserstein spaces} Let $(\mathcal{M},d_\mathcal{M})$ be a metric space and $p\in [1,\infty)$. The Wasserstein space $\mathsf{P}_p(\mathcal{M})$ is not a Banach space, but there is a natural version of the metric dimension reduction question in this context. \begin{question}\label{Q:wasserstein} Fix $\alpha>1$, $n\in \mathbb N$ and $\mu_1,\ldots,\mu_n\in \mathsf{P}_p(\mathcal{M})$. What is the asymptotic behavior of the smallest $k\in \mathbb N$ for which there is $\mathcal{S}\subset \mathcal{M}$ with $|\mathcal{S}|\le k$ such that $(\{\mu_1,\ldots,\mu_n\},\mathsf{W}_p)$ embeds with distortion $\alpha$ into $\mathsf{P}_p(\mathcal{S})$? \end{question} Spaces of measures with the Wasserstein metric $\mathsf{W}_p$ are of major importance in pure and applied mathematics, as well as in computer science (mainly for $p=1$, where they are used in graphics and vision, but also for other values of $p$; see e.g.~the discussion in~\cite{ANN16-ICALP}). However, their bi-Lipschitz structure is poorly understood, especially so in the above context of metric dimension reduction. If $k$ were small in Question~\ref{Q:wasserstein}, then this would give a way to "compress" collections of measures using measures with small support while approximately preserving Wasserstein distances. In the context of, say, image retrieval (mainly $\mathcal{M}=\{1,\ldots,n\}^2\subset \mathbb R^2$ and $p=1$), this could be viewed as obtaining representations of images using a small number of "pixels." Charikar~\cite{Cha02} and Indyk--Thaper~\cite{IT03} proved that if $\mathcal{M}$ is a finite metric space, then $\mathsf{P}_1(\mathcal{M})$ embeds into $\ell_1$ with distortion $O(\log |\mathcal{M}|)$. Hence, if the answer to Question~\ref{Q:wasserstein} were $k=n^{o(1)}$ for some $\alpha=O(1)$, then it would follow that any $n$-point subset of $\mathsf{P}_1(\mathcal{M})$ embeds into $\ell_1$ with distortion $o(\log n)$, i.e., better distortion than the general bound that is provided by Bourgain's embedding theorem (actually the $\ell_1$-variant of that theorem, which is also known to be sharp in general~\cite{LLR}). This shows that one cannot hope to answer Question~\ref{Q:wasserstein} with $k=n^{o(1)}$ and $\alpha=O(1)$ without imposing geometric restrictions on the underlying metric space $\mathcal{M}$, since if $(\mathcal{M}=\{x_1,\ldots,x_n\},d_\mathcal{M})$ is a metric space for which $\mathsf{c}_1(\mathcal{M})\asymp \log n$, then we can take $\mu_1,\ldots,\mu_n$ to be the point masses $\d_{x_1},\ldots,\d_{x_n}$, so that $(\{\mu_1,\ldots,\mu_n\},\mathsf{W}_1)$ is isometric to $(\mathcal{M},d_\mathcal{M})$. The pertinent issue is therefore to study Question~\ref{Q:wasserstein} when the $\mathcal{M}$ is "nice." For example, sufficiently good bounds here for $\mathcal{M}=\mathbb R^2$ would be relevant to Question~\ref{Q:bourgain universality}, but at this juncture such a potential approach to Question~\ref{Q:bourgain universality} is quite speculative. The above "problematic" example relied inherently on the fact that the underlying metric space $\mathcal{M}$ is itself far from being embeddable in $\ell_1$, but the difficulty persists even when $\mathcal{M}=\ell_1$. Indeed, we recalled in Question~\ref{Q:bourgain universality} that Bourgain proved~~\cite{Bourgain-trees} that $\mathsf{P}_1(\ell_1)$ is universal, and hence the spaces of either Theorem~\ref{thm:coarse matousek} or Theorem~\ref{thm:average distortion} embed into $\mathsf{P}_1(\ell_1)$ with $O(1)$ distortion. So, for arbitrarily large $n\in \mathbb N$ we can find probability measures $\mu_1,\ldots,\mu_n$ on $\ell_1$ (actually on a sufficiently high dimensional Hamming cube $\{0,1\}^N$) such that $(\{\mu_1,\ldots,\mu_n\},\mathsf{W}_1)$ does not admit a good embedding into any normed space of dimension $n^{o(1)}$. This rules out an answer of $k=n^{o(1)}$ to Question~\ref{Q:wasserstein} (even for average distortion) for $(\mu_1,\ldots,\mu_n,\mathsf{W}_1)$, because in the setting of Question~\ref{Q:wasserstein}, while $\mathsf{P}_1(\mathcal{S})$ is not a normed space, it embeds isometrically into a normed space of dimension $|\mathcal{S}|-1$ (the dual of the mean-zero Lipschitz functions on $(\mathcal{S},d_\mathcal{M})$; see e.g.~\cite{NS07,Vil09} for an explanation of this standard fact). In the case of average distortion, one could see this using a different approach of Khot and the author~\cite{KN06} which constructs a collection of $n=e^{O(d)}$ probability measures on the Hamming cube $\{0,1\}^d$ that satisfy the conclusion of Theorem~\ref{thm:average distortion}, as explained in~\cite[Remark~5]{Nao17}. This shows that even though these $n$ probability measures reside on a Hamming cube of dimension $O(\log n)$, one cannot realize their Wasserstein-1 geometry with $O(1)$ distortion (even on average) in any normed space of dimension $n^{o(1)}$, let alone in $\mathsf{P}_1(\mathcal{S})$ with $|\mathcal{S}|=n^{o(1)}$. It is therefore natural to investigate Question~\ref{Q:wasserstein} when $\mathcal{M}$ is low-dimensional. When $p=1$, this remains an (important) uncharted terrain. When $p>1$ and $\mathcal{M}=\mathbb R^3$, partial information on Question~\ref{Q:wasserstein} follows from~\cite{ANN15}. To see this, focus for concreteness on the case $p=2$. Fix $\alpha\ge 1$ and $n\in \mathbb N$. Suppose that $(\mathcal{N},d_\mathcal{N})$ is an $n$-point metric space for which the conclusion of Theorem~\ref{thm:coarse matousek} holds true with $\omega(t)=\sqrt{t}$ and $\Omega(t)=2\alpha\sqrt{t}$. By~\cite{ANN15}, the metric space $(\mathcal{N},\sqrt{d_\mathcal{N}})$ embeds with distortion $2$ into $\mathsf{P}_2(\mathbb R^3)$, where $\mathbb R^3$ is equipped with the standard Euclidean metric. Hence, if the image under this embedding of $\mathcal{N}$ in $\mathsf{P}_2(\mathbb R^3)$ embedded into some $k$-dimensional normed space with distortion $\alpha$, then by Theorem~\ref{thm:coarse matousek} necessarily $k\ge n^{c/\alpha^2}$ for some universal constant $c$. This does not address Question~\ref{Q:wasserstein} as stated, because to the best of our knowledge it is not known whether $\mathsf{P}_2(\mathcal{S})$ embeds with $O(1)$ distortion into some "low-dimensional" normed space for every "small" $\mathcal{S}\subset \mathbb R^3$ (the relation between "small" and "low-dimensional" remains to be studied). In the case of average distortion, repeat this argument with $\mathcal{N}$ now being the metric space of Theorem~\ref{thm:average distortion}. By Remark~\ref{rem:for average snowflake} below, if the image in $\mathsf{P}_2(\mathbb R^3)$ of $(\mathcal{N},\sqrt{d_\mathcal{N}})$ embedded with average distortion $\alpha$ into some $k$-dimensional normed space, then necessarily $k\ge\exp(\frac{c}{\alpha}\sqrt{\log n})$. \section{Finite subsets of Hilbert space}\label{sec:JL} The article~\cite{JL84} of Johnson and Lindenstrauss is devoted to proving a theorem on the extension of Lipschitz functions from finite subsets of metric spaces.\footnote{Stating this theorem here would be an unnecessary digression, but we highly recommend examining the accessible geometric result of~\cite{JL84}; see~\cite{NR17} for a review of the current state of the art on Lipschitz extension from finite subsets.} Over the ensuing decades, the classic~\cite{JL84} attained widespread prominence outside the rich literature on the {\em Lipschitz extension problem}, due to two components of~\cite{JL84} that had major conceptual significance and influence, but are technically simpler than the proof of its main theorem. The first of these components is the JL lemma, which we already stated in the Introduction. Despite its wide acclaim and applicability, this result is commonly called a "lemma" rather than a "theorem" because within the context of~\cite{JL84} it was just that, i.e., a relatively simple step toward the proof of the main theorem of~\cite{JL84}. The second of these components is a section of~\cite{JL84} that is devoted to formulating open problems in the context of the Ribe program; we already described a couple of the questions that were raised there, but it contains more questions that proved to be remarkably insightful and had major impact on subsequent research (see e.g.~\cite{Bal92,NPSS06}). Despite its importance, the impact of~\cite{JL84} on the Ribe program will not be pursued further in the present article, but we will next proceed to study the JL lemma in detail (including some new observations). Recalling Theorem~\ref{thm:JL alpha}, the JL lemma~\cite{JL84} asserts that for every integer $n\ge 2$ and (distortion/error tolerance) $\alpha\in (1,\infty)$, if $x_1,\ldots,x_n$ are distinct vectors in a Hilbert space $(H,\|\cdot\|_H)$, then there exists (a target dimension) $k\in \{1,\ldots,n\}$ and a new $n$-tuple of $k$-dimensional vectors $y_1,\ldots,y_n\in \mathbb R^k$ such that \begin{equation}\label{eq:JL conclusion1} k\lesssim_\alpha \log n, \end{equation} and the assignment $x_i\mapsto y_i$, viewed as a mapping into $\ell_2^k$, has distortion at most $\alpha$, i.e., \begin{equation}\label{eq:JL conclusion2} \forall\, i,j\in\{1,\ldots,n\},\qquad\|x_i-x_j\|_H\le \|y_i-y_j\|_{\ell_2^k}\le \alpha\|x_i-x_j\|_H. \end{equation} It is instructive to take note of the "compression" that this statement entails. By tracking the numerical value of the target dimension $k$ that the proof in Section~\ref{sec:JL opt} below yields (see Remark~\ref{rem:constants}), one concludes that given an arbitrary collection of, say, a billion vectors of length a billion (i.e., $1\,000\,000\,000$ elements of $\mathbb R^{1\,000\,000\,000}$), one can find a billion vectors of length $329$ (i.e., elements of $\mathbb R^{329}$), all of whose pairwise distances are within a factor $2$ of the corresponding pairwise distances among the initial configuration of billion-dimensional vectors. Furthermore, if one wishes to maintain the pairwise distances of those billion vectors within a somewhat larger constant factor, say, a factor of $10$ or $450$, then one could do so in dimension $37$ or $9$, respectively. The logarithmic dependence on $n$ in~\eqref{eq:JL conclusion1} is optimal, up to the value of the implicit ($\alpha$-dependent) constant factor. This is so even when one considers the special case when $x_1,\ldots,x_n\in H$ are the vertices of the standard $(n-1)$-simplex, i.e., $\|x_i-x_j\|_H=1$ for all distinct $i,j\in \{1,\ldots,n\}$, and even when one allows the Euclidean norm in~\eqref{eq:JL conclusion2} to be replaced by any norm $\|\cdot\|:\mathbb R^k\to [0,\infty)$, namely if instead of~\eqref{eq:JL conclusion2} we have $1\le \|y_i-y_j\|\le \alpha$ for all distinct $i,j\in \{1,\ldots,n\}$. Indeed, denote the unit ball of $\|\cdot\|$ by $B=\{z\in \mathbb R^k:\ \|z\|\le 1\}$ and let $\mathrm{\bf vol}_k(\cdot)$ be the Lebesgue measure on $\mathbb R^k$. If $i,j\in \{1,\ldots,n\}$ are distinct, then by the triangle inequality the assumed lower bound $\|y_i-y_j\|\ge 1$ implies that the interiors of $y_i+\frac12 B$ and $y_j+\frac12 B$ are disjoint. Hence, if we denote $A=\bigcup_{i=1}^n(y_i+\frac12 B)$, then $\mathrm{\bf vol}_k(A)=\sum_{i=1}^n \mathrm{\bf vol}_k(y_i+\frac12 B)=\frac{n}{2^k}\mathrm{\bf vol}_k(B)$. At the same time, for every $u,v\in A$ there are $i,j\in \{1,\ldots,n\}$ for which $u\in y_i+\frac12 B$ and $v\in y_j+\frac12 B$, so by another application of the triangle inequality we have $\|u-v\|\le \|y_i-y_j\|+1\le \alpha+1$. This implies that $A-A\subset (\alpha+1)B$. Hence, \begin{equation*}\label{eq:use BM} (\alpha+1)\sqrt[k]{\mathrm{\bf vol}_k(B)}=\sqrt[k]{\mathrm{\bf vol}_k((\alpha+1)B)}\ge \sqrt[k]{\mathrm{\bf vol}_k(A-A)}\ge 2\sqrt[k]{\mathrm{\bf vol}_k(A)} =\sqrt[k]{n\mathrm{\bf vol}_k(B)}, \end{equation*} where the penultimate step uses the Brunn--Minkowski inequality~\cite{Sch14}. This simplifies to give \begin{equation}\label{eq:k lower log general norm} k\ge \frac{\log n}{\log(\alpha+1)}. \end{equation} By~\cite{Bra99,Dek00}, the vertices of $(n-1)$-simplex embed isometrically into any infinite dimensional Banach space, so we have thus justified the bound~\eqref{lower brunnminkowksi intro}, and hence in particular the first lower bound on $\mathsf{k}_n^\alpha(\ell_2)$ in~\eqref{eq:JL in theporem}. As we already explained, the second lower bound (for the almost-isometric regime) on $\mathsf{k}_n^\alpha(\ell_2)$ in~\eqref{eq:JL in theporem} is due to the very recent work~\cite{GN17}. The upper bound on $\mathsf{k}_n^\alpha(\ell_2)$ in~\eqref{eq:JL in theporem}, namely that in~\eqref{eq:JL conclusion1} we can take \begin{equation}\label{eq:JL const best known} k\lesssim \frac{\log n}{\log\big(1+(\alpha-1)^2\big)}\asymp \max\left\{\frac{\log n}{(\alpha-1)^2},\frac{\log n}{\log \alpha}\right\}, \end{equation} follows from the original proof of the JL lemma in~\cite{JL84}. A justification of~\eqref{eq:JL const best known} appears in Section~\ref{sec:JL opt} below. \begin{question}[dimension reduction for the vertices of the simplex]Fix $\d\in (0,\frac12)$. What is the order of magnitude (up to universal constant factors) of the smallest $\mathfrak{S}(\d)\in (0,\infty)$ such that for every $n\in \mathbb N$ there is $k\in \mathbb N$ with $k\le \mathfrak{S}(\d)\log n$ and $y_1,\ldots,y_{n}\in \mathbb R^k$ that satisfy $1\le \|y_i-y_j\|_{2}\le 1+\d$ for all distinct $i,j\in \{1,\ldots,n\}$? By~\eqref{eq:JL const best known} we have $\mathfrak{S}(\d)\lesssim 1/\d^{2}$. The best-known lower bound here is $\mathfrak{S}(\d)\gtrsim 1/(\d^2\log(1/\d))$, due to Alon~\cite{Alo03}. \end{question} \begin{remark} The upper bound~\eqref{eq:JL const best known} treats the target dimension in the JL lemma for an {\em arbitrary} subset of a Hilbert space. The lower bound~\eqref{eq:k lower log general norm} was derived in the special case of the vertices of the regular simplex, but it is also more general as it is valid for embeddings of these vertices into an {\em arbitrary} $k$-dimensional norm. In this (both special, and more general) setting, the bound~\eqref{eq:k lower log general norm} is quite sharp for large $\alpha$. Indeed, by~\cite{ABV98} (see also~\cite[Corollary~2.4]{OR16}), for each $n\in \mathbb N$ and $\alpha>\sqrt{2}$, if we write $ k=\lceil(\log(4n))/\log(\alpha^2/(2\sqrt{\alpha^2-1}))\rceil, $ then for {\em every} norm $\|\cdot\|$ on $\mathbb R^k$ there exist $y_1,\ldots,y_n\in \mathbb R^k$ satisfying $1\le \|y_i-y_j\|\le \alpha$ for distinct $i,j\in \{1,\ldots,n\}$. See~\cite[Theorem~4.3]{FL94} for an earlier result in this direction. See also~\cite{ABV98} and the references therein (as well as~\cite[Problem~2.5]{OR16}) for partial results towards understanding the analogous issue (which is a longstanding open question) in the small distortion regime $\alpha\in (1,\sqrt{2}]$. \end{remark} \subsection{Optimality of re-scaled random projections}\label{sec:JL opt} To set the stage for the proof of the JL lemma, note that by translation-invariance we may assume without loss of generality that one of the vectors $\{x_i\}_{i=1}^n$ vanishes, and then by replacing the Hilbert space $H$ with the span of $\{x_i\}_{i=1}^n$, we may further assume that $H=\mathbb R^{n-1}$. Let $\mathsf{Proj}_{\mathbb R^k}\in \mathsf{M}_{k\times (n-1)}(\mathbb R)$ be the $k$ by $n-1$ matrix of the orthogonal projection from $\mathbb R^{n-1}$ onto $\mathbb R^k$, i.e., $\mathsf{Proj}_{\mathbb R^k} z=(z_1,\ldots,z_k)\in\mathbb R^k$ is the first $k$ coordinates of $z=(z_1,\ldots,z_{n-1})\in \mathbb R^{n-1}$. One could attempt to simply truncate the vectors vectors $x_1,\ldots,x_n$ so as to obtain $k$-dimensional vectors, namely to consider the vectors $\{y_i=\mathsf{Proj}_{\mathbb R^k} x_i\}_{i=1}^n$ in~\eqref{eq:JL conclusion2}. This naive (and heavy-handed) way of forcing low-dimensionality can obviously fail miserably, e.g.~we could have $\mathsf{Proj}_kx_i=0$ for all $i\in \{1,\ldots,n\}$. Such a simplistic idea performs poorly because it makes two arbitrary and unnatural choices, namely it does not take advantage of rotation-invariance and scale-invariance. To remedy this, let $\mathsf{O}_{n-1}\subset \mathsf{M}_{n-1}(\mathbb R)$ denote the group of $n-1$ by $n-1$ orthogonal matrices, and fix (a scaling factor) $\sigma\in (0,\infty)$. Let $\mathsf{O}\in \mathsf{O}_{n-1}$ be a random orthogonal matrix distributed according to the Haar probability measure on $\mathsf{O}_{n-1}$. In~\cite{JL84} it was shown that if $k$ is sufficiently large (yet satisfying~\eqref{eq:JL conclusion1}), then for an appropriate $\sigma>0$ with positive probability~\eqref{eq:JL conclusion2} holds for the following random vectors. \begin{equation}\label{eq:choice of yi} \{y_i =\sigma \mathsf{Proj}_{\mathbb R^k}\mathsf{O} x_i\}_{i=1}^n\subset \mathbb R^k. \end{equation} We will do more than merely explain why the randomly projected vectors in~\eqref{eq:choice of yi} satisfy the desired conclusion~\eqref{eq:JL conclusion2} of the JL lemma with positive probability. We shall next demonstrate that such a procedure is the {\em best possible} (in a certain sense that will be made precise) among all the possible choices of random assignments of $x_1,\ldots,x_n$ to $y_1,\ldots,y_n$ via multiplication by a random matrix in $M_{k\times(n-1)}(\mathbb R)$, provided that we optimize so as to use the best scaling factor $\sigma\in (0,\infty)$ in~\eqref{eq:choice of yi}. Let $\mu$ be any Borel probability measure on $\mathsf{M}_{k\times (n-1)}(\mathbb R)$, i.e., $\mu$ represents an arbitrary (reasonably measurable) distribution over $k\times (n-1)$ random matrices $\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R)$. For $\alpha\in (1,\infty)$ define \begin{equation}\label{eqLdef frak p mu} \mathfrak{p}_\mu^\alpha\stackrel{\mathrm{def}}{=} \inf_{z\in \mathbf{S}^{n-2}} \mu\Big[\big\{\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R):\ 1\le \|\mathsf{A} z\|_{\ell_2^k}\le \alpha\big\} \Big], \end{equation} where $\mathbf{S}^{n-2}=\{z\in \mathbb R^{n-1}:\ \|z\|_{\ell_2^{n-1}}=1\}$ denotes the unit Euclidean sphere in $\mathbb R^{n-1}$. Then \begin{align}\label{eq:uninion JL} \nonumber \mu \bigg[&\bigcap_{i,j\in \{1,\ldots,n\}}\Big\{\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R):\|x_i-x_j\|_{\ell_2^{n-1}}\le \|\mathsf{A} x_i-\mathsf{A} x_j\|_{\ell_2^k}\le \alpha\|x_i-x_j\|_{\ell_2^{n-1}}\Big\}\bigg]\\ \nonumber &= 1- \mu\bigg[\bigcup_{i=1}^n\bigcup_{j=i+1}^n \bigg(\mathsf{M}_{k\times (n-1)}(\mathbb R)\setminus \Big\{\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R):1 \le \Big\|\mathsf{A} \frac{x_i- x_j}{\|x_i-x_j\|_{\ell_2^{n-1}}}\Big\|_{\ell_2^k}\le \alpha\Big\}\bigg)\bigg]\\ \nonumber &\ge 1- \sum_{i=1}^n\sum_{j=i+1}^n \bigg(1-\mu \bigg[\Big\{\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R):1 \le \Big\|\mathsf{A} \frac{x_i- x_j}{\|x_i-x_j\|_{\ell_2^{n-1}}}\Big\|_{\ell_2^k}\le \alpha\Big\}\bigg]\bigg)\\ &\ge 1- \binom{n}{2}\big(1-\mathfrak{p}_\mu^\alpha\big). \end{align} Hence, the random vectors $\{y_i=\mathsf{A} x_i\}_{i=1}^n$ will satisfy~\eqref{eq:JL conclusion2} with positive probability if $\mathfrak{p}_\mu^\alpha>1-\frac{2}{n(n-1)}$. In order to succeed to embed the largest possible number of vectors into $\mathbb R^k$ via the above randomized procedure while using the estimate~\eqref{eq:uninion JL}, it is in our best interest to work with a probability measure $\mu$ on $\mathsf{M}_{k\times (n-1)}(\mathbb R)$ for which $\mathfrak{p}_\mu^\alpha$ is as large as possible. To this end, define \begin{equation}\label{eq:def pnk} \mathfrak{p}^\alpha_{n,k}\stackrel{\mathrm{def}}{=} \sup \Big\{\mathfrak{p}_\mu^\alpha:\ \mu\ \mathrm{is\ a \ Borel\ probability\ measure\ on\ } \mathsf{M}_{k\times (n-1)}(\mathbb R)\Big\}. \end{equation} Then, the conclusion~\eqref{eq:JL conclusion2} of the JL lemma will be valid provided $k\in \{1,\ldots,n\}$ satisfies \begin{equation}\label{eq:JL condition random} \mathfrak{p}^\alpha_{n,k}>1-\frac{2}{n(n-1)}. \end{equation} The following proposition asserts that the supremum in the definition~\eqref{eq:def pnk} of $\mathfrak{p}_{n,k}^\alpha$ is attained at a distribution over random matrices that has the aforementioned structure~\eqref{eq:choice of yi}. \begin{proposition}[multiples of random orthogonal projections are JL-optimal]\label{prop:sigma nk} Fix $\alpha\in (1,\infty)$, an integer $n\ge 4$ and $k\in \{1,\ldots,n-3\}$. Let $\mu=\mu_{n,k}^\alpha$ be the probability distribution on $\mathsf{M}_{k\times (n-1)}(\mathbb R)$ of the random matrix \begin{equation}\label{eq:optimal rescaled projection} \sqrt{\frac{\alpha^{\frac{2n-6}{n-k-3}}-1}{\alpha^{\frac{2k}{n-k-3}}-1}}\cdot \mathsf{Proj}_{\mathbb R^k}\mathsf{O}, \end{equation} that is obtained by choosing $\mathsf{O}\in \mathsf{O}_{n-1}$ according to the normalized Haar measure on $\mathsf{O}_{n-1}$. Then $\mathfrak{p}_\mu^\alpha=\mathfrak{p}_{n,k}^\alpha$. \end{proposition} Obviously~\eqref{eq:optimal rescaled projection} is {\em not} a multiple of a uniformly random rank $k$ orthogonal projection $\mathsf{Proj}:\mathbb R^{n-1}\to \mathbb R^{n-1}$ (chosen according to the normalized Haar measure on the appropriate Grassmannian). To obtain such a distribution, one should multiply the matrix in~\eqref{eq:optimal rescaled projection} on the left by $\mathsf{O}^*$. That additional rotation does not influence the Euclidean length of the image, and hence it does not affect the quantity~\eqref{eqLdef frak p mu}. For this reason and for simplicity of notation, we prefer to work with~\eqref{eq:optimal rescaled projection} rather than random projections as was done in~\cite{JL84}. \begin{proof}[Proof of Proposition~\ref{prop:sigma nk}] Given $\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R)$, denote its singular values by $\mathsf{s}_1(\mathsf{A})\ge\ldots\ge \mathsf{s}_k(\mathsf{A})$, i.e., they are the eigenvalues (with multiplicity) of the symmetric matrix $\sqrt{\mathsf{A}\mA^*}\in \mathsf{M}_k(\mathbb R)$. Then, \begin{equation}\label{eq:haar identity} \mathfrak{H}^{\mathsf{O}_{n-1}} \Big[\big\{\mathsf{O}\in \mathsf{O}_{n-1}:\ 1\le \|\mathsf{A}\mathsf{O} z\|_{\ell_2^k}\le \alpha\big\}\Big]= \int_{\mathbf{S}^{k-1}}\psi_{n,k}^\alpha\bigg(\Big(\sum_{i=1}^k\mathsf{s}_i(\mathsf{A})^2 \omega_i^2\Big)^{\frac12}\bigg)\ud \mathfrak{H}^{\mathbf{S}^{k-1}}(\omega), \end{equation} where $\mathfrak{H}^{\mathsf{O}_{n-1}}$ and $\mathfrak{H}^{\mathbf{S}^{k-1}}$ are the Haar probability measures on the orthogonal group $\mathsf{O}_{n-1}$ and the unit Euclidean sphere $\mathbf{S}^{k-1}$, respectively, and the function $\psi_{n,k}^\alpha:[0,\infty)\to \mathbb R$ is defined by \begin{equation}\label{eq:def psi nk modified} \forall\, \sigma\in [0,\infty),\qquad \psi_{n,k}^\alpha(\sigma)\stackrel{\mathrm{def}}{=} \frac{2\pi^{\frac{k}{2}}}{\Gamma\big(\frac{k}{2}\big)}\int^{\max\left\{1,\sigma\right\}}_{\max\left\{1,\frac{\sigma}{\alpha}\right\}}\frac{(s^2-1)^{\frac{n-k-3}{2}}}{s^{n-2}}\ud s. \end{equation} To verify the identity~\eqref{eq:haar identity}, consider the singular value decomposition \begin{equation}\label{eq:write SVD}\mathsf{A}= \mathsf{U}\begin{pmatrix} \mathsf{s}_1(\mathsf{A}) & 0 & \dots& \dots&0 \\ 0 & \mathsf{s}_2(\mathsf{A})& \ddots& \ddots & \vdots\\ \vdots & \ddots & \ddots & \ddots & \vdots\\ \vdots & \ddots & \ddots& \ddots &0\\ 0 & \dots & \dots &0&\mathsf{s}_k(\mathsf{A}) \end{pmatrix}\mathsf{Proj}_{\mathbb R^k}\mathsf{V},\end{equation} where $\mathsf{U}\in \mathsf{O}_k$ and $\mathsf{V}\in \mathsf{O}_{n-1}$. If $\mathsf{O}\in \mathsf{O}_{n-1}$ is distributed according to $\mathfrak{H}^{\mathsf{O}_{n-1}}$, then by the left-invariance of $\mathfrak{H}^{\mathsf{O}_{n-1}}$ we know that $\mathsf{V}\mathsf{O}$ is distributed according to $\mathfrak{H}^{\mathsf{O}_{n-1}}$. By rotation-invariance and uniqueness of Haar measure on $\mathbf{S}^{n-2}$ (e.g.~\cite{MS}), it follows that for every $z\in \mathbf{S}^{n-1}$ the random vector $\mathsf{V}\mathsf{O} z$ is distributed according to the normalized Haar measure on $\mathbf{S}^{n-2}$. So, $\mathsf{Proj}_{\mathbb R^k} \mathsf{V}\mathsf{O} z$ is distributed on the Euclidean unit ball of $\mathbb R^k$, with density \begin{equation}\label{eq:projected density} \big(u\in \mathbb R^k\big)\mapsto \frac{\Gamma\big(\frac{n-1}{2}\big)}{\pi^{\frac{k}{2}}\Gamma\big(\frac{n-1-k}{2}\big)} \left(1-\|u\|_{\ell_2^k}^2\right)^{\frac{n-k-3}{2}}\mathbf{1}_{\big\{\|u\|_{\ell_2^k}\le 1\big\}}. \end{equation} See~\cite{Sta82} for a proof of this distributional identity (or~\cite[Corollary~4]{BGMN05} for a more general derivation); in codimension $2$, namely $k=n-3$, this is a higher-dimensional analogue of Archimedes' theorem that the projection to $\mathbb R$ of the uniform surface area measure on the unit Euclidean sphere in $\mathbb R^{3}$ is the Lebesgue measure on $[-1,1]$. Recalling~\eqref{eq:write SVD}, it follows from this discussion that the Euclidean norm of $\mathsf{A}\mathsf{O} z$ has the same distribution as $(\sum_{i=1}^k \mathsf{s}_i(\mathsf{A})^2u_i^2)^{1/2}$, where $u=(u_1,\ldots,u_k)\in \mathbb R^k$ is distributed according to the density~\eqref{eq:projected density}. The identity~\eqref{eq:haar identity} now follows by integration in polar coordinates $(\omega,r)\in \mathbf{S}^{k-1}\times [0,\infty)$, followed by the change of variable $s=1/r$. Next, $\psi_{n,k}^\alpha$ vanishes on $[0,1]$, increases on $[1,\alpha]$, and is smooth on $[\alpha,\infty)$. The integrand in~\eqref{eq:def psi nk modified} is at most $s^{-k-1}$, so $\lim_{\sigma\to \infty}\psi_{n,k}^\alpha(\sigma)=0$. By directly differentiating~\eqref{eq:def psi nk modified} and simplifying the resulting expression, one sees that if $\sigma\in [\alpha,\infty)$, then $(\psi_{n,k}^\alpha)'(\sigma)=0$ if and only if $\sigma=\sigma_{\max}(n,k,\alpha)$, where \begin{equation}\label{eq:def sigma max} \sigma_{\max}(n,k,\alpha)\stackrel{\mathrm{def}}{=} \sqrt{\frac{\alpha^{\frac{2n-6}{n-k-3}}-1}{\alpha^{\frac{2k}{n-k-3}}-1}}. \end{equation} Therefore, the global maximum of $\psi_{n,k}^\alpha$ is attained at $\sigma_{\max}(n,k,\alpha)$, and by~\eqref{eq:haar identity} we have \begin{equation}\label{eq:use identity agaoin for maximizer} \forall\, \mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R),\qquad \mathfrak{H}^{\mathsf{O}_{n-1}} \Big[\big\{\mathsf{O}\in \mathsf{O}_{n-1}:\ 1\le \|\mathsf{A}\mathsf{O} z\|_{\ell_2^k}\le \alpha\big\}\Big]\le \psi_{n,k}^\alpha\big(\sigma_{\max}(n,k,\alpha)\big)=\mathfrak{p}_\mu^\alpha. \end{equation} The final step of~\eqref{eq:use identity agaoin for maximizer} is another application~\eqref{eq:haar identity}, this time in the special case $\mathsf{A}=\sigma_{\max}(n,k,\alpha)\mathsf{Proj}_{\mathbb R^k}$, while recalling~\eqref{eqLdef frak p mu} and~\eqref{eq:def sigma max}, and that $\mu$ is the distribution of the random matrix appearing in~\eqref{eq:optimal rescaled projection}. To conclude the proof of Proposition~\eqref{prop:sigma nk}, take any Borel probability measure $\nu$ on $\mathsf{M}_{k\times (n-1)}(\mathbb R)$ and integrate~\eqref{eq:use identity agaoin for maximizer} while using Fubini's theorem to obtain the estimate \begin{multline*} \mathfrak{p}_\mu^\alpha\ge\int_{\mathsf{M}_{k\times (n-1)}(\mathbb R)} \mathfrak{H}^{\mathsf{O}_{n-1}}\Big[\big\{\mathsf{O}\in \mathsf{O}_{n-1}:\ 1\le \|\mathsf{A}\mathsf{O} z_0\|_{\ell_2^k}\le \alpha\big\}\Big]\ud\nu(\mathsf{A})\\ = \int_{\mathsf{O}_{n-1}} \nu\Big[\big\{\mathsf{A}\in \mathsf{M}_{k\times (n-1)}(\mathbb R):\ 1\le \|\mathsf{A}\mathsf{O} z_0\|_{\ell_2^k}\le \alpha\big\}\Big] \ud\mathfrak{H}^{\mathsf{O}_{n-1}}(\mathsf{O})\stackrel{\eqref{eqLdef frak p mu}}{\ge} \int_{\mathsf{O}_{n-1}} \mathfrak{p}_\nu^\alpha \ud\mathfrak{H}^{\mathsf{O}_{n-1}}(\mathsf{O})=\mathfrak{p}_\nu^\alpha. \end{multline*} So, the maximum of $\mathfrak{p}_\nu^\alpha$ over the Borel probability measures $\nu$ on $\mathsf{M}_{k\times (n-1)}(\mathbb R)$ is attained at $\mu$.\end{proof} \begin{remark} Recalling~\eqref{eq:JL condition random}, due to~\eqref{eq:use identity agaoin for maximizer} the conclusion~\eqref{eq:JL conclusion2} of the JL lemma holds if $k$ satisfies \begin{equation}\label{eq:beta JL} \frac{2\pi^{\frac{k}{2}}}{\Gamma\big(\frac{k}{2}\big)}\int^{\sigma_{\max}(n,k,\alpha)}_{\frac{1}{\alpha} \sigma_{\max}(n,k,\alpha)}\frac{(s^2-1)^{\frac{n-k-3}{2}}}{s^{n-2}}\ud s>1-\frac{2}{n(n-1)}, \end{equation} where $\sigma_{\max}(n,k,\alpha)$ is given in~\eqref{eq:def sigma max}. This is the best-known bound on $k$ in the JL lemma, which, due to Proposition~\ref{prop:sigma nk}, is the best-possible bound that is obtainable through the reasoning~\eqref{eq:uninion JL}. In particular, the asymptotic estimate~\eqref{eq:JL const best known} follows from~\eqref{eq:beta JL} via straightforward elementary calculus. \end{remark} \begin{remark} The JL lemma was reproved many times; see~\cite{FM88,Gor88,IM99,AV99,DG03,Ach03,KM05,IN07,Mat08,AC09,DG09,KW11,AL13,DG14,KN14,BDN15,Dir16}, though we make no claim that this is a comprehensive list of references. There were several motivations for these further investigations, ranging from the desire to obtain an overall better understanding of the JL phenomenon, to obtain better bounds, and to obtain distributions on random matrices $\mathsf{A}$ as in~\eqref{eq:uninion JL} with certain additional properties that are favorable from the computational perspective, such as ease of simulation, use of fewer random bits, sparsity, and the ability to evaluate the mapping $(z\in \mathbb R^{n-1})\mapsto \mathsf{A} z$ quickly (akin to the fast Fourier transform). This body of work represents ongoing efforts by computer scientists and applied mathematicians to further develop improved "JL transforms," driven by their usefulness as a tool for data-compression. We will not survey these ideas here, partially because we established that using random projections yields the best-possible bound on the target dimension $k$ (moreover, this procedure is natural and simple). We speculate that working with the Haar measure on the orthogonal group $\mathsf{O}_{n-1}$ as in~\eqref{eq:choice of yi} could have benefits that address the above computational issues, but leave this as an interesting open-ended direction for further research. A specific conjecture towards this goal appears in~\cite[page~320]{AC09}, and we suspect that the more recent work~\cite{BG12} on the spectral gap of Hecke operators of orthogonal Cayley graphs should be relevant in this context as well (e.g. for derandomization and fast implementation of~\eqref{eq:choice of yi}; see~\cite{BHH16,KM15} for steps in this direction). \end{remark} \begin{remark}\label{rem:constants} In the literature there is often a preference to use random matrices with independent entries in the context of the JL lemma, partially because they are simple to generate, though see the works~\cite{Ste80,Gen00,Mez07} on generating elements of the orthogonal group $\mathsf{O}_{n-1}$ that are distributed according to its Haar measure. In particular, the best bound on $k$ in~\eqref{eq:JL conclusion1} that was previously available in the literature~\cite{DG03} arose from applying~\eqref{eq:uninion JL} when $\mathsf{A}$ is replaced by the random matrix $\sigma\mathsf{G}$, where $\sigma=1/\sqrt{k}$ and the entries of $\mathsf{G}=(\mathsf{g}_{ij})\in M_{k\times (n-1)}(\mathbb R)$ are independent standard Gaussian random variables. We can, however, optimize over the scaling factor $\sigma$ in this setting as well, in analogy to the above optimization over the scaling factor in~\eqref{eq:choice of yi}, despite the fact that we know that working with the Gaussian matrix $\mathsf{G}$ is inferior to using a random rotation. A short calculation reveals that the optimal scaling factor is now $\sqrt{(\alpha^2-1)/(2k\log \alpha)}$, i.e., the best possible re-scaled Gaussian matrix for the purpose of reasoning as in~\eqref{eq:uninion JL} is not $\frac{1}{\sqrt{k}}\mathsf{G}$ but rather the random matrix \begin{equation}\label{eq:rescaled gaussian} \mathsf{G}_k^\alpha\stackrel{\mathrm{def}}{=} \sqrt{\frac{\alpha^2-1}{2k\log \alpha}}\cdot\mathsf{G}. \end{equation} For this optimal multiple of a Gaussian matrix, one computes that for every $z\in \mathbf{S}^{n-2}$ we have \begin{multline}\label{eq:gaussian optimizer identity and estimate} 1-\Pr \Big[1\le \|\mathsf{G}_k^\alpha z\|_{\ell_2^k}\le \alpha\Big]=\frac{2k^{\frac{k}{2}}}{\Gamma\big(\frac{k}{2}\big)}\int_{\log \alpha}^\infty \left(\frac{\beta}{e^{2\beta}-1}\right)^{\frac{k}{2}}\exp\bigg(-\frac{k\beta}{e^{2\beta}-1}\bigg)\ud \beta\\ < \frac{4k^{\frac{k}{2}-1}}{\Gamma\big(\frac{k}{2}\big)}\left(\frac{\alpha^2-1}{\log \alpha}\alpha^{\frac{2}{\alpha^2-1}}\right)^{-\frac{k}{2}}\frac{(\alpha^2-1)^2\log\alpha }{2\alpha^4\log\alpha+2\alpha^2-\alpha^4-4\alpha^2(\log\alpha)^2-2\log\alpha-1}. \end{multline} The first step in~\eqref{eq:gaussian optimizer identity and estimate} follows from a straightforward computation using the fact that the squared Euclidean length of $\mathsf{G}_k^\alpha z$ is distributed according to a multiple of the $\chi^2$ distribution with $k$ degrees of freedom (see e.g.~\cite{Dur10}), i.e., one can write the leftmost term of~\eqref{eq:gaussian optimizer identity and estimate} explicitly as a definite integral, and then check that it indeed equals the middle term of~\eqref{eq:gaussian optimizer identity and estimate}, e.g., by verifying the the derivatives with respect to $\alpha$ of both expressions coincide. The final estimate in~\eqref{eq:gaussian optimizer identity and estimate} can be justified via a modicum of straightforward calculus. We deduce from this that the conclusion~\eqref{eq:JL conclusion2} of the JL lemma is holds with positive probability if for each $i\in \{1,\ldots,n\}$ we take $y_i$ to be the image of $x_i$ under the re-scaled Gaussian matrix in~\eqref{eq:rescaled gaussian}, provided that $k$ is sufficiently large so as to ensure that \begin{equation}\label{eq:k condition improved dg} \frac{\Gamma\big(\frac{k}{2}\big)}{k^{\frac{k}{2}-1}}\left(\frac{\alpha^2-1}{\log \alpha}\alpha^{\frac{2}{\alpha^2-1}}\right)^{\frac{k}{2}}\ge \frac{2n^2(\alpha^2-1)^2\log\alpha }{2\alpha^4\log\alpha+2\alpha^2-\alpha^4-4\alpha^2(\log\alpha)^2-2\log\alpha-1}. \end{equation} The values that we stated for the target dimension $k$ in the JL lemma with a billion vectors were obtained by using~\eqref{eq:k condition improved dg}, though even better bounds arise from an evaluation of the integral in~\eqref{eq:gaussian optimizer identity and estimate} numerically, which is what we recommend to do for particular settings of the parameters. As $\alpha\to 1$, the above bounds improve over those of~\cite{DG03} only in the second-order terms. For larger $\alpha$ these bounds yield substantial improvements that might matter in practice, e.g.~for embedding a billion vectors with distortion $2$, the target dimension that is required using the best-available estimate in the literature~\cite{DG03} is $k=768$, while~\eqref{eq:k condition improved dg} shows that $k=329$ suffices. \end{remark} \section{Infinite subsets of Hilbert space}\label{sec:doubling} The JL lemma provides a quite complete understanding of the metric dimension reduction problem for finite subsets of Hilbert space. For infinite subsets of Hilbert space, the research splits into two strands. The first is to understand those subsets $\mathscr{C}\subset \mathbb R^n$ for which certain random matrices in $\mathsf{M}_{k\times n}(\mathbb R)$ (e.g.~random projections, or matrices whose entries are i.i.d. independent sub-Gaussian random variables) yield with positive probability an embedding of $\mathscr{C}$ into $\mathbb R^k$ of a certain pre-specified distortion; this was pursued in~\cite{Gor88,KM05,IN07,MPT07,MT08,BDN15,Dir16,PDG17}, yielding a satisfactory answer which relies on multi-scale chaining criteria~\cite{Tal14,Nel16} . The second (and older) research strand focuses on the mere {\em existence} of a low-dimensional embedding rather than on the success of the specific embedding approach of (all the known proofs of) the JL lemma. Specifically, given a subset $\mathscr{C}$ of a Hilbert space and $\alpha\in [1,\infty)$, could one understand when does $\mathscr{C}$ admit an embedding with distortion $\alpha$ into $\ell_2^k$ for some $k\in \mathbb N$? If one ignores the dependence on the distortion $\alpha$, then this qualitative question coincides with Problem~\ref{Q:bilip Rk} (the bi-Lipschitz embedding problem into $\mathbb R^k$), since if a metric space $(\mathcal{M},d_\mathcal{M})$ satisfies $\inf_{k\in \mathbb N}\mathsf{c}_{\mathbb R^k}(\mathcal{M})<\infty$, then in particular it admits a bi-Lipschitz embedding into a Hilbert space. We shall next describe an obvious necessary condition for bi-Lipschitz embeddability into $\mathbb R^k$ for some $k\in\mathbb N$. In what follows, all balls in a metric space $(\mathcal{M},d_\mathcal{M})$ will be closed balls, i.e., for $x\in \mathcal{M}$ and $r\in [0,\infty)$ we write $B_\mathcal{M}(x,r)=\{y\in \mathcal{M}:\ d_\mathcal{M}(x,y)\le r\}$. Given $K\in [2,\infty)$, a metric space $(\mathcal{M},d_\mathcal{M})$ is said to be $K$-{\em doubling} (e.g.~\cite{Bou28,CW71}) if every ball in $\mathcal{M}$ (centered anywhere in $\mathcal{M}$ and of any radius) can be covered by at most $K$ balls of half its radius, i.e., for every $x\in \mathcal{M}$ and $r\in [0,\infty)$ there is $m\in \mathbb N$ with $m\le K$ and $y_1,\ldots,y_m\in \mathcal{M}$ such that $B_\mathcal{M}(x,r)\subset B_\mathcal{M}(y_1,\frac12 r)\cup\ldots\cup B_\mathcal{M}(y_m,\frac12 r)$. A metric space is doubling if it is $K$-doubling for some $K\in [2,\infty)$. Fix $k\in \mathbb N$ and $\alpha\ge 1$. If a metric space $(\mathcal{M},d_\mathcal{M})$ embeds with distortion $\alpha$ into a normed space $(\mathbb R^k,\|\cdot\|)$, then $\mathcal{M}$ is $(4\alpha+1)^k$-doubling. Indeed, fix $x\in \mathcal{M}$ and $r>0$. Let $\{z_1,\ldots,z_n\}\subset B_\mathcal{M}(x,r)$ be a maximal subset (with respect to inclusion) of $B_\mathcal{M}(x,r)$ satisfying $d_\mathcal{M}(z_i,z_j)>\frac12 r$ for distinct $i,j\in \{1,\ldots,n\}$. The maximality of $\{z_1,\ldots,z_n\}$ ensures that for any $w\in B_\mathcal{M}(x,r)\setminus \{z_1,\ldots,z_n\}$ we have $\min_{i\in \{1,\ldots,n\}} d_\mathcal{M}(w,z_i)\le \frac12 r$, i.e., $B_\mathcal{M}(x,r)\subset B_\mathcal{M}(z_1,\frac12 r)\cup\ldots\cup B_\mathcal{M}(z_n,\frac12 r)$. We are assuming that there is an embedding $f:\mathcal{M}\to \mathbb R^k$ that satisfies $d_\mathcal{M}(u,v)\le \|f(u)-f(v)\|\le \alpha d_\mathcal{M}(u,v)$ for all $u,v\in \mathcal{M}$. So, for distinct $i,j\in \{1,\ldots,n\}$ we have $\frac{r}{2}<d_\mathcal{M}(z_i,z_j)\le \|f(z_i)-f(z_j)\|\le \alpha d_\mathcal{M}(z_i,z_j)\le \alpha (d_\mathcal{M}(z_i,x)+ d_\mathcal{M}(x,z_j))\le2\alpha r$. The reasoning that led to~\eqref{eq:k lower log general norm} with $y_1=\frac{2}{r}f(z_1),\ldots,y_n=\frac{2}{r}f(z_n)$ and $\alpha$ replaced by $4\alpha$ gives $k\ge (\log n)/\log(4\alpha+1)$, i.e., $n\le (4\alpha+1)^k$. \begin{remark}\label{rem:doub log n} In Section~\ref{sec:metric dim reduction intro} we recalled that in the context of the Ribe program $\log |\mathcal{M}|$ was the initial (in hindsight somewhat naive, though still very useful) replacement for the "dimension" of a finite metric space $\mathcal{M}$. This arises naturally also from the above discussion. Indeed, $\mathcal{M}$ is trivially $|\mathcal{M}|$-doubling (simply cover each ball in $\mathcal{M}$ by singletons), and this is the best bound that one could give on the doubling constant of $\mathcal{M}$ in terms of $|\mathcal{M}|$. So, from the perspective of the doubling property, the natural restriction on $k\in \mathbb N$ for which there exists an embedding of $\mathcal{M}$ into some $k$-dimensional normed space with $O(1)$ distortion is that $k\gtrsim \log |\mathcal{M}|$. \end{remark} Using terminology that was recalled in Remark~\ref{rem:snowflake}, the definition of the doubling property directly implies that for every $\theta\in (0,1)$ a metric space $\mathcal{M}$ is doubling if and only if its $\theta$-snowflake $\mathcal{M}^\theta$ is doubling. With this in mind, Theorem~\ref{thm:assouad} below is a very important classical achievement of Assouad~\cite{Ass83}. \begin{theorem}\label{thm:assouad} The following assertions are equivalent for every metric space $(\mathcal{M},d_\mathcal{M})$. \begin{itemize} \item $\mathcal{M}$ is doubling. \item For every $\theta\in (0,1)$ there is $k\in \mathbb N$ such that $\mathcal{M}^\theta$ admits a bi-Lipschitz embedding into $\mathbb R^k$. \item Some snowflake of $\mathcal{M}$ admits a bi-Lipschitz embedding into $\mathbb R^k$ for some $k\in \mathbb N$. \end{itemize} \end{theorem} Theorem~\ref{thm:assouad} is a qualitative statement, but its proof in~\cite{Ass83} shows that for every $K\in [2,\infty)$ and $\theta\in (0,1)$, there are $\alpha(K,\theta)\in [1,\infty)$ and $k(K,\theta)\in \mathbb N$ such that if $\mathcal{M}$ is $K$-doubling, then $\mathcal{M}^\theta$ embeds into $\mathbb R^{k(K,\theta)}$ with distortion $\alpha(K,\theta)$; the argument of~\cite{Ass83} inherently gives that as $\theta\to 1$, i.e., as the $\theta$-snowflake $\mathcal{M}^\theta$ approaches the initial metric space $\mathcal{M}$, we have $\alpha(K,\theta)\to \infty$ and $k(K,\theta)\to \infty$. A meaningful study of the best-possible asymptotic behavior of the distortion $\alpha(K,\theta)$ here would require specifying which norm on $\mathbb R^k$ is being considered. Characterizing the quantitative dependence in terms of geometric properties of the target norm on $\mathbb R^k$ has not been carried out yet (it isn't even clear what should the pertinent geometric properties be), though see~\cite{HM06} for an almost isometric version when one considers the $\ell_\infty$ norm on $\mathbb R^k$ (with the dimension $k$ tending to $\infty$ as the distortion approaches $1$); see also~\cite{GK15} for a further partial step in this direction. In~\cite{NN12} it was shown that for $\theta\in [\frac12,1)$ one could take $k(K,\theta)\le k(K)$ to be bounded by a constant that depends only on $K$; the proof of this fact in~\cite{NN12} relies on a probabilistic construction, but in~\cite{DS13} a clever and instructive deterministic proof of this phenomenon was found (though, yielding asymptotically worse estimates on $\alpha(K,\theta),k(K)$ than those of~\cite{NN12}). Assouad's theorem is a satisfactory characterization of the doubling property in terms of embeddability into finite-dimensional Euclidean space. However, it is a "near miss" as an answer to Problem~\ref{Q:bilip Rk}: the same statement with $\theta=1$ would have been a wonderful resolution of the bi-Lipschitz embedding problem into $\mathbb R^k$, showing that a simple intrinsic ball covering property is equivalent to bi-Lipschitz embeddability into some $\mathbb R^k$. It is important to note that while the snowflaking procedure does in some sense "tend to" the initial metric space as $\theta\to 1$, for $\theta<1$ it deforms the initial metric space substantially (e.g.~such a $\theta$-snowflake does not contain any non-constant rectifiable curve). So, while Assouad's theorem with the stated snowflaking is useful (examples of nice applications appear in~\cite{BS00,HM06}), its failure to address the bi-Lipschitz category is a major drawback. Alas, more than a decade after the publication of Assouad's theorem, it was shown in~\cite{Sem96} (relying a on a rigidity theorem of~\cite{Pan89}) that Assouad's theorem does not hold with $\theta=1$, namely there exists a doubling metric space that does not admit a bi-Lipschitz embedding into $\mathbb R^k$ for any $k\in \mathbb N$. From the qualitative perspective, we now know that the case $\theta=1$ of Assouad's theorem fails badly in the sense that there exists a doubling metric space (the continuous $3$-dimensional Heisenberg group, equipped with the Carnot--Carath\'eodory metric) that does not admit a bi-Lipschitz embedding into any Banach space with the Radon--Nikod\'ym property~\cite{LN06,CK06} (in particular, it does not admit a bi-Lipschitz embedding into any reflexive or separable dual Banach space, let alone a finite dimensional Banach space), into any $L_1(\mu)$ space~\cite{CK10}, or into any Alexandrov space of curvature bounded above or below~\cite{Pau01} (a further strengthening appears in the forthcoming work~\cite{AN17}). From the quantitative perspective, by now we know that balls in the discrete $5$-dimensional Heisenberg group equipped with the word metric (which is doubling) have the asymptotically worst-possible bi-Lipschitz distortion (as a function of their cardinality) in uniformly convex Banach spaces~\cite{LN14-poincare} (see also~\cite{ANT13}) and $L_1(\mu)$ spaces~\cite{NY17,NY17-versus}; interestingly, the latter assertion is not true for the $3$-dimensional Heisenberg group~\cite{NY17-3dim}, while the former assertion does hold true for the $3$-dimensional Heisenberg group~\cite{LN14-poincare}. All of the known "bad examples" (including, in addition to the Heisenberg group, those that were subsequently found in~\cite{Laakso,Laa02,BP99,Che99}) which show that the doubling property is not the sought-after answer to Problem~\ref{Q:bilip Rk} do not even embed into an infinite-dimensional Hilbert space. This leads to the following natural and intriguing question that was stated by Lang and Plaut in~\cite{LP01}. \begin{question}\label{Q:LP} Does every doubling subset of a Hilbert admit a bi-Lipschitz embedding into $\mathbb R^k$ for some $k\in \mathbb N$? \end{question} As stated, Question~\ref{Q:LP} is qualitative, but by a compactness argument (see~\cite[Section~4]{NN12}) if its answer were positive, then for every $K\in [2,\infty)$ there would exist $d_K\in \mathbb N$ and $\alpha_K\in [1,\infty)$ such that any $K$-doubling subset of a Hilbert space would embed into $\ell_2^{d_K}$ with distortion $\alpha_K$. If Question~\ref{Q:LP} had a positive answer, then it would be very interesting to determine the asymptotic behavior of $d_K$ and $\alpha_K$ as $K\to \infty$. A positive answer to Question~\ref{Q:LP} would be a solution of Problem~\ref{Q:bilip Rk}, though the intrinsic criterion that it would provide would be quite complicated, namely it would say that a metric space $(\mathcal{M},d_\mathcal{M})$ admits a bi-Lipschitz embedding into $\mathbb R^k$ for some $k\in \mathbb N$ if and only if it is doubling and satisfies the family of quadratic distance inequalities~\eqref{eq:K LLR}. More importantly, it seems that any positive answer to Question~\ref{Q:LP} would devise a procedure that starts with a subset in a very high-dimensional Euclidean space and, if that subset is $O(1)$-doubling, produce a bi-Lipschitz embedding into $\mathbb R^{O(1)}$; such a procedure, if possible, would be a quintessential metric dimension reduction result that is bound to be of major importance. It should be noted that, as proved in~\cite[Remark~4.1]{IN07}, any such general procedure cannot be an embedding into low-dimensions via a linear mapping as in the JL lemma, i.e., Question~\ref{Q:LP} calls for a genuinely nonlinear dimension reduction technique.\footnote{On its own, the established necessity of obtaining a genuinely nonlinear embedding method into low dimensions should not discourage attempts to answer Question~\ref{Q:LP}, because some rigorous nonlinear dimension reduction methods have been devised in the literature; see e.g.~\cite{Ass83,Sem99,CS02,GKL03,BLMN04,BM04,KLMN05,LS05,BKL07,LNP09,CGT10,ABN11,BRS11,GT11,NN12,DS13,LdMM13,NR13,GK15,BG16,Nei16,OR16,ANN17}. However, all of these approaches seem far from addressing Question~\ref{Q:LP}.} Despite the above reasons why a positive answer to Question~\ref{Q:LP} would be very worthwhile, we suspect that Question~\ref{Q:LP} has a negative answer. A specific doubling subset of a Hilbert space which is a potential counterexample to Question~\ref{Q:LP} was constructed in~\cite[Question~3]{NN12}, but to date it remains unknown whether or not this subset admits a bi-Lipschitz embedding into $\mathbb R^{O(1)}$. If the answer to Question~\ref{Q:LP} is indeed negative, then the next challenge would be to formulate a candidate conjectural characterization for resolving the bi-Lipschitz embedding problem into $\mathbb R^k$. The analogue of Question~\ref{Q:LP} is known to fail in some non-Hilbertian settings. Specifically, it follows from~\cite{LN14,NY17,NY17-versus} that for every $p\in (2,\infty)$ there exists a doubling subset $\mathscr{D}_p$ of $L_p(\mathbb R)$ that does not admit a bi-Lipschitz embedding into any $L_q(\mu)$ space for any $q\in [1,p)$. So, in particular there is no bi-Lipschitz embedding of $\mathscr{D}_p$ into any finite-dimensional normed space, and a fortiori there is no such embedding into any finite-dimensional subspace of $L_p(\mathbb R)$. Note that in~\cite{LN14} this statement is made for embeddings of $\mathscr{D}_p$ into $L_q(\mu)$ in the reflexive range $q\in (1,p)$, and the case $q=1$ is treated in~\cite{LN14} only when $p\ge p_0$ for some universal constant $p_0>2$. The fact that $\mathscr{D}_p$ does not admit a bi-Lipschitz embedding into any $L_1(\mu)$ space follows by combining the argument of~\cite{LN14} with the more recent result\footnote{When~\cite{LN14} was written, only a weaker bound of~\cite{CKN11} was known.} of~\cite{NY17,NY17-versus} when the underlying group in the construction of~\cite{LN14} is the $5$-dimensional Heisenberg group; interestingly we now know~\cite{NY17-3dim} that if one carries out the construction of~\cite{LN14} for the $3$-dimensional Heisenberg group, then the reasoning of~\cite{LN14} would yield the above conclusion only when $p>4$. A different example of a doubling subset of $L_p(\mathbb R)$ that fails to embed bi-Lipschitzly into $\ell_p^k$ for any $k\in \mathbb N$ was found in~\cite{BGN15}. In $L_1(\mathbb R)$, there is an even stronger counterexample~\cite[Remark~1.4]{LN14}: By~\cite{GNRS04}, the spaces considered in~\cite{Laakso,Laa02} yields a doubling subset of $L_1(\mathbb R)$ that by~\cite{CK09-RNP} (see also~\cite{Ost11}) does not admit a bi-Lipschitz embedding into any Banach space with the Radon--Nikod\'ym property~\cite{LN06,CK06}, hence it does not admit a bi-Lipschitz embedding into any reflexive or separable dual Banach space. The potential validity of the above statement for $p\in (1,2)$ remains an intriguing open problem, and the case $p=2$ is of course Question~\ref{Q:LP}. \section{Matou\v{s}ek's random metrics, Milnor--Thom, and coarse dimension reduction}\label{sec:matousek} Fix two moduli $\omega,\Omega:[0,\infty)\to [0,\infty)$ as in Theorem~\ref{thm:coarse matousek}, i.e., they are increasing functions and $\omega\le \Omega$ point-wise. For a metric space $(\mathcal{M},d_\mathcal{M})$ define $\dim_{(\omega,\Omega)}(\mathcal{M},d_\mathcal{M})$ to be the smallest dimension $k\in \mathbb N$ for which there exists a $k$-dimensional normed space $(X,\|\cdot\|_X)=(X(\mathcal{M}),\|\cdot\|_{X(\mathcal{M})})$ and a mapping $f:\mathcal{M}\to X$ that satisfies~\eqref{eq:coarse condition}. If no such $k\in \mathbb N$ exists, then write $\dim_{(\omega,\Omega)}(\mathcal{M},d_\mathcal{M})=\infty$. For $\alpha\in [1,\infty)$, this naturally generalizes the notation $\dim_\alpha(\mathcal{M},d_\mathcal{M})$ of~\cite{LLR} in the bi-Lipschitz setting, which coincides with $\dim_{(t,\alpha t)}(\mathcal{M},d_\mathcal{M})$. Recalling~\eqref{eq:def beta}, the goal of this section is to show that $\dim_{(\omega,\Omega)}(\mathcal{M},d_\mathcal{M})\ge n^{c\beta(\omega,\Omega)}$ for arbitrarily large $n\in \mathbb N$, some universal constant $c\in (0,\infty)$ and some metric space $(\mathcal{M},d_\mathcal{M})$ with $|\mathcal{M}|=3n$, thus proving Theorem~\ref{thm:coarse matousek}. We will do so by following Matou\v{s}ek's beautiful ideas in~\cite{Mat96}, yielding a probabilistic argument for the existence of such an intrinsically (coarsely) high-dimensional metric space $(\mathcal{M},d_\mathcal{M})$. The collections of subsets of a set $S$ of size $\ell\in \mathbb N$ will be denoted below $\binom{S}{\ell}=\{\mathfrak{e}\subset S:\ |\mathfrak{e}|=\ell\}$. Fix $n\in \mathbb N$ and a bipartite graph $\mathsf{G}=(\mathsf{L},\mathsf{R},\mathsf{E})$ with $|\mathsf{L}|=|\mathsf{R}|=n$. Thus, $\mathsf{L}$ and $\mathsf{R}$ are disjoint $n$-point sets (the "left side" and "right side" of $\mathsf{G}$) and $\mathsf{E}$ is a subset of $\binom{\mathsf{L}\cup\mathsf{R}}{2}$ consisting only of $\mathfrak{e}\subset \mathsf{L}\cup\mathsf{R}$ such that $|\mathsf{L}\cap \mathfrak{e}|=|\mathsf{R}\cap \mathfrak{e}|=1$. Following Matou\v{s}ek~\cite{Mat96}, any such graph $\mathsf{G}$ can used as follows as a "template" for obtaining a family $2^{|\mathsf{E}|}$ graphs, each of which having $3n$ vertices. For each $\lambda \in \mathsf{L}$ introduce two new elements $\lambda^{\!+},\lambda^{\!-}$. Denote $\mathsf{L}^{\!+}=\{\lambda^{\!+}:\ \lambda\in \mathsf{L}\}$ and $\mathsf{L}^{\!-}=\{\lambda^{\!-}:\ \lambda\in \mathsf{L}\}$. Assume that the sets $\mathsf{L}^{\!+},\mathsf{L}^{\!-},\mathsf{R}$ are disjoint. For every $\sigma:\mathsf{E}\to \{-,+\}$ define \begin{equation}\label{eq:def signed edges} \mathsf{E}_\sigma\stackrel{\mathrm{def}}{=} \left\{\big\{\lambda^{\sigma(\{\lambda,\rho\})},\rho\big\}:\ (\lambda,\rho)\in \mathsf{L}\times\mathsf{R}\ \wedge\ \{\lambda,\rho\}\in \mathsf{E}\right\}\subset \binom{\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R}}{2}. \end{equation} We thus obtain a bipartite graph $\mathsf{G}_\sigma=(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-},\mathsf{R},\mathsf{E}_\sigma)$. By choosing $\sigma:\mathsf{E}\to \{-,+\}$ uniformly at random, we think of $\mathsf{G}_\sigma$ as a random graph; let $\Pr$ denote the uniform probability measure on the set of all such $\sigma$. In other words, consider $\sigma:\mathsf{E}\to \{-,+\}$ to be independent tosses of a fair coin, one for each edge of $\mathsf{G}$. Given an outcome of the coin tosses $\sigma$, each edge $\mathfrak{e}\in \mathsf{E}$ of $\mathsf{G}$ induces an element of $\mathsf{E}_\sigma$ as follows. If $\lambda$ is the endpoint of $\mathfrak{e}$ in $\mathsf{L}$ and $\rho$ is the endpoint of $\mathfrak{e}$ in $\mathsf{R}$, then $\mathsf{E}_\sigma$ contains exactly one of the unordered pairs $\{\lambda^{\!+},\rho\}, \{\lambda^{\!-},\rho\}$ depending on whether $\sigma(\mathfrak{e})=+$ or $\sigma(\mathfrak{e})=-$, respectively; see Figure~\ref{fig:mat graphs} below for a schematic depiction of this construction. \begin{figure}[h] \centering \begin{minipage}{7.3in} \centering \includegraphics[height=3.3in]{random-girth.png} \smallskip \caption{The random bipartite graph $\mathsf{G}_\sigma=(\mathsf{L}^{\!+}\cup \mathsf{L}^{\!-},\mathsf{R},\mathsf{E}_\sigma)$ that is associated to the bipartite graph $\mathsf{G}=(\mathsf{L},\mathsf{R},\mathsf{E})$ and coin flips $\sigma:\mathsf{E}\to \{-,+\}$. Suppose that $(\lambda,\rho)\in \mathsf{L}\times \mathsf{R}$ and $\mathfrak{e}=\{\lambda,\rho\}\in \mathsf{E}$. If the outcome of the coin that was flipped for the edge $\mathfrak{e}$ is $+$, then include in $\mathsf{E}_\sigma$ the red edge between $\lambda^{\!+}$ and $\rho$ and do not include an edge between $\lambda^{\!-}$ and $\rho$. If the outcome of the coin that was flipped for the edge $\mathfrak{e}$ is $-$, then include in $\mathsf{E}_\sigma$ the blue edge between $\lambda^{\!-}$ and $\rho$ and do not include an edge between $\lambda^{\!+}$ and $\rho$. } \label{fig:mat graphs} \end{minipage} \end{figure} Let $d_{\mathsf{G_\sigma}}:(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R})\times (\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R})\to [0,\infty]$ be the shortest-path metric corresponding to $\mathsf{G}_\sigma$, with the convention that $d_{\mathsf{G}_\sigma}(x,y)=\infty$ if $x,y\in \mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R}$ belong to different connected components of $\mathsf{G}_\sigma$. We record for convenience of later use the following very simple observation. \begin{claim}\label{claim:cycle compress} Fix $\lambda\in \mathsf{L}$ and $\sigma:\mathsf{E}\to \{-,+\}$. Suppose that $k\stackrel{\mathrm{def}}{=} d_{\mathsf{G}_\sigma}(\lambda^{\!+},\lambda^{\!-})<\infty$. Then the original "template graph" $\mathsf{G}$ contains a cycle of length at most $k$. \end{claim} \begin{proof} Denote by $\pi:\mathsf{L}^{\!+}\cup \mathsf{L}^{\!-}\cup\mathsf{R}\to \mathsf{L}\cup\mathsf{R}$ the canonical "projection," i.e., $\pi$ is the identity mapping on $\mathsf{R}$ and $\pi(\lambda^{\!+})=\pi(\lambda^{\!-})=\lambda$ for every $\lambda\in \mathsf{L}$. The natural induced mapping $\pi:\mathsf{E}_\sigma\to \mathsf{E}$ (given by $\pi(\{x,y\})=\{\pi(x),\pi(y)\}$ for each $\{x,y\}\in \mathsf{E}_\sigma$) is one-to-one, because by construction $\mathsf{E}_\sigma$ contains one and only one of the unordered pairs $\{\mu^+,\rho\},\{\mu^-,\rho\}$ for each $(\mu,\rho)\in \mathsf{L}\times \mathsf{R}$ with $\{\mu,\rho\}\in \mathsf{E}$. Let $\gamma: \{0,\ldots, k\}\to \mathsf{L}^{\!+}\cup \mathsf{L}^{\!-}\cup\mathsf{R}$ be a geodesic in $\mathsf{G}_\sigma$ that joins $\lambda^{\!+}$ and $\lambda^{\!-}$. Thus $\gamma(0)= \lambda^{\!+}$, $\gamma(k)= \lambda^{\!-}$ and $\{\{\gamma(i-1),\gamma(i)\}\}_{i=1}^k$ are distinct edges in $\mathsf{E}_\sigma$ (they are distinct because $\gamma$ is a shortest path joining $\lambda^{\!+}$ and $\lambda^{\!-}$ in $\mathsf{G}_\sigma$). By the injectivity of $\pi$ on $\mathsf{E}_\sigma$, the unordered pairs $\{\{\pi(\gamma(i-1)),\pi(\gamma(i))\}\}_{i=1}^k$ are distinct edges in $\mathsf{E}$. So, the subgraph $\mathsf{H}$ of $\mathsf{G}$ that is induced on the vertices $\{\pi(\gamma(i))\}_{i=0}^k$ has at least $k$ edges. But, $\mathsf{H}$ has at most $k$ vertices, since $\pi(\gamma(0))=\pi(\gamma(k))=\lambda$. Hence $\mathsf{H}$ is not a forest, i.e., it contains a cycle of length at most $k$. \end{proof} Even though $d_{\mathsf{G_\sigma}}$ is not necessarily a metric due to its possible infinite values, for every $s,T\in (0,\infty)$ we can re-scale and truncate it so as to obtain a (random) metric $d_\sigma^{s,T}:(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R})\times (\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R})\to [0,\infty]$ by defining \begin{equation}\label{eq:truncated} \forall\, x,y\in \mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},\qquad d_\sigma^{s,T}(x,y)\stackrel{\mathrm{def}}{=} \min\big\{sd_{\mathsf{G}_\sigma}(x,y),T\big\}. \end{equation} The following lemma shows that if in the above construction $\mathsf{G}$ has many edges and no short cycles, then with overwhelmingly high probability the random metric in~\eqref{eq:truncated} has large coarse metric dimension. \begin{lemma}\label{lem:prob matousek} There is a universal constant $\eta>0$ with the following property. For every $\omega,\Omega:[0,\infty)\to [0,\infty)$ as above, every $n\in \mathbb N$ and every template graph $\mathsf{G}=(\mathsf{L},\mathsf{R},\mathsf{E})$ as above, suppose that $g\in \mathbb N$ and $s,T>0$ satisfy \begin{equation}\label{eq:girth assumtpions} \frac{\omega^{-1}(2\Omega(s))}{s}<g\le \frac{T}{s}, \end{equation} and that the shortest cycle in $\mathsf{G}$ has length at least $g$. Then for every $\d\in (0,\frac13]$ we have \begin{equation}\label{eq:probabilistic version'} \Pr\left[\sigma:\mathsf{E}\to \{-,+\}:\ \dim_{\omega,\Omega}\Big(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T}\Big)\le \d\eta\frac{|\mathsf{E}|}{n}\right] < \big(2\d^{\d}\big)^{-|\mathsf{E}|}. \end{equation} In particular, by choosing $\d=\frac13$ in~\eqref{eq:probabilistic version'} we deduce that \begin{equation}\label{eq:probabilistic version} \Pr\left[\sigma:\mathsf{E}\to \{-,+\}:\ \dim_{\omega,\Omega}\Big(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T}\Big)> \frac{\eta|\mathsf{E}|}{3n}\right] > 1-e^{-\frac15|\mathsf{E}|}. \end{equation} \end{lemma} Prior to proving Lemma~\ref{lem:prob matousek}, we shall now explain how it implies Theorem~\ref{thm:coarse matousek}. \begin{proof}[Proof of Theorem~\ref{thm:coarse matousek} assuming Lemma~\ref{lem:prob matousek}] Recalling~\eqref{eq:def beta}, we can fix $s\in (0,\infty)$ such that \begin{equation}\label{eq:def girth} g\stackrel{\mathrm{def}}{=} \left\lfloor \frac{\omega^{-1}(2\Omega(s))}{s} \right\rfloor+1\le \frac{2}{\beta(\omega,\Omega)}. \end{equation} There is a universal constant $\kappa\in (0,\infty)$ such that for arbitrarily large $n\in \mathbb N$ there exists a bipartite graph $\mathsf{G}=(\mathsf{L},\mathsf{R},\mathsf{E})$ with $|\mathsf{L}|=|\mathsf{R}|=n$, girth at least $g$ (i.e., $\mathsf{G}$ does not contain any cycle of length smaller than $g$) and $|\mathsf{E}|\ge n^{1+\kappa/g}$. Determining the largest possible value of $\kappa$ here is a well-studied and longstanding open problem in graph theory (see e.g.~the discussions in~\cite{Bol01,Mat02,Ost13}), but for the present purposes any value of $\kappa$ suffices. For the latter (much more modest) requirement, one can obtain $\mathsf{G}$ via a simple probabilistic construction (choosing each of the edges independently at random and deleting an arbitrary edge from each cycle of length at most $g-1$). See~\cite{LUW95} for the best known lower bound on $\kappa$ here (arising from an algebraic construction). We shall use the above graph $\mathsf{G}$ as the template graph for the random graphs $\{\mathsf{G}_\sigma\}_{\sigma:\mathsf{E}\to \{-,+\}}$. Our choice of $g$ in~\eqref{eq:def girth} ensures that if we write $T=sg$, then~\eqref{eq:girth assumtpions} holds true and we obtain a distribution over metric spaces $(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T})$ for which the conclusion~\eqref{eq:probabilistic version} of Lemma~\ref{lem:prob matousek} holds true. Hence, by choosing $c=\kappa/2$ and substituting the bound $|\mathsf{E}|\ge n^{1+\kappa/g}$ into~\eqref{eq:probabilistic version} while using~\eqref{eq:def girth} we have \begin{equation*} \Pr\left[\sigma:\mathsf{E}\to \{-,+\}:\ \dim_{\omega,\Omega}\Big(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T}\Big) \gtrsim n^{c\beta(\omega,\Omega)}\right] \ge 1-\exp\left(-\frac15n^{1+c\beta(\omega,\Omega)}\right). \end{equation*} Consequently, by the definition of $\dim_{\omega,\Omega}(\cdot)$, with probability exponentially close to $1$ the random metric space $(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T})$ satisfies the assertion of Theorem~\ref{thm:coarse matousek}. \end{proof} The proof of Lemma~\ref{lem:prob matousek} relies on the following lemma that was obtained implicitly by Matou\v{s}ek~\cite{Mat96}. Its proof takes as input a clever argument of Alon~\cite{Al86} which uses the classical bound of Milnor~\cite{Mil64} and Thom~\cite{Tho65} on the number of connected components of a real algebraic variety. \begin{lemma}\label{lem:rank} Fix $m,n\in \mathbb N$ and $\mathsf{E}\subset \{1,\ldots,n\}^2$. Suppose that $\mathsf{A}_1=(a_{ij}^1),\ldots,\mathsf{A}_m=(a_{ij}^m)\in \mathsf{M}_n(\mathbb R)$ are matrices that satisfy $a_{ij}^k\neq 0$ for all $(i,j)\in \mathsf{E}$ and $k\in \{1,\ldots,m\}$, and that the sign vectors \begin{equation}\label{eq:sign patterns} \big(\mathrm{\bf sign}(a^1_{ij})\big)_{(i,j)\in \mathsf{E}},\big(\mathrm{\bf sign}(a^2_{ij})\big)_{(i,j)\in \mathsf{E}}\ldots,\big(\mathrm{\bf sign}(a^m_{ij})\big)_{(i,j)\in \mathsf{E}}\subset \{-1,1\}^{\mathsf{E}} \end{equation} are distinct. Then there exists $k\in \{1,\ldots,m\}$ such that \begin{equation}\label{eq:rank lower optimized} \mathrm{\bf rank}(\mathsf{A}_k)\gtrsim \frac{\log m}{n\log\left(\frac{|\mathsf{E}|}{\log m}\right)}. \end{equation} \end{lemma} \begin{proof} Let $\alpha\in \mathbb N$ be an auxiliary parameter that will be specified later so as to optimize the ensuing argument. Write $h=\lceil |\mathsf{E}|/\alpha\rceil$ and fix any partition of $\mathsf{E}$ into subsets $\mathsf{J}_1,\ldots,\mathsf{J}_h\neq \emptyset$ (i.e., $\mathsf{J}_1,\ldots,\mathsf{J}_h\subset\mathsf{E}$ are pairwise disjoint and $\mathsf{E}=\mathsf{J}_1\cup\ldots \cup\mathsf{J}_h$) that satisfy $|\mathsf{J}_u|\le \alpha$ for all $u\in \{1,\ldots,h\}$. Denote \begin{equation}\label{eq:def mu rho} \mu\stackrel{\mathrm{def}}{=} \min_{\substack{(i,j)\in \mathsf{E}\\ k\in \{1,\ldots,m\}}}\big|a_{ij}^k\big|\qquad\mathrm{and}\qquad r\stackrel{\mathrm{def}}{=} \max_{k\in \{1,\ldots,m\}} \mathrm{\bf rank}(\mathsf{A}_k). \end{equation} Lemma~\ref{lem:rank} assumes that $\mu>0$, and its goal is to show that $r$ is at least a universal constant multiple of the quantity that appears in the right hand side on~\eqref{eq:rank lower optimized}. The definition of $r$ means that for every $k\in \{1,\ldots,m\}$ there exist $n$-by-$r$ and $r$-by-$n$ matrices $\mathsf{B}_k\in \mathsf{M}_{n\times r}(\mathbb R)$ and $\mathsf{C}_k\in \mathsf{M}_{r\times n}(\mathbb R)$, respectively, such that $\mathsf{A}_k=\mathsf{B}_k\mathsf{C}_k$. Define vectors $\{\zeta_k=(\zeta_1^k,\ldots,\zeta^k_h)\in \mathbb R^h\}_{k=1}^m$ by setting \begin{equation}\label{eq:square root} \forall\, (k,u)\in \{1,\ldots,m\}\times \{1,\ldots,h\},\qquad \zeta_u^k\stackrel{\mathrm{def}}{=} \sqrt{\prod_{(i,j)\in\mathsf{J}_u}(a_{ij}^k)^2-\frac12 \mu^{2|\mathsf{J}_u|}}. \end{equation} Observe that the definition of $\mu$ in~\eqref{eq:def mu rho} ensures that the quantity under the square root in~\eqref{eq:square root} is positive, so $\zeta_u\in (0,\infty)$. Define polynomials $\{\mathfrak{p}_{u}:\mathsf{M}_{n\times r}(\mathbb R)\times \mathsf{M}_{r\times n}(\mathbb R)\times \mathbb R^h\to \mathbb R\}_{u=1}^h$ by setting \begin{equation}\label{eq:def pu} \mathfrak{p}_{u}(X,Y,z)\stackrel{\mathrm{def}}{=} \prod_{(i,j)\in\mathsf{J}_u} (XY)_{ij}^2-z_u^2-\frac{1}{2}\mu^{2|\mathsf{J}_u|}= \prod_{(i,j)\in\mathsf{J}_u} \bigg(\sum_{k=1}^n x_{ik}y_{kj}\bigg)^2-z_u^2-\frac{1}{2}\mu^{2|\mathsf{J}_u|}, \end{equation} for all $u\in \{1,\ldots,h\}$, $X=(x_{is})\in \mathsf{M}_{n\times r}(\mathbb R)$, $Y=(y_{sj})\in \mathsf{M}_{r\times n}(\mathbb R)$ and $z=(z_i)\in \mathbb R^h$. The above notation ensures that $\mathfrak{p}_u(\mathsf{B}_k,\mathsf{C}_k,\zeta_k)=0$ for all $k\in \{1,\ldots,m\}$ and $u\in \{1,\ldots,h\}$. In other words, $\{(\mathsf{B}_k,\mathsf{C}_k,\zeta_k)\}_{k=1}^m\subset \mathscr{V}$, where $\mathscr{V}\subset \mathsf{M}_{n\times r}(\mathbb R)\times \mathsf{M}_{r\times n}(\mathbb R)\times \mathbb R^h$ is the variety \begin{equation}\label{eq:def variety} \mathscr{V}\stackrel{\mathrm{def}}{=} \bigcap_{u=1}^h \left\{(X,Y,z)\in \mathsf{M}_{n\times r}(\mathbb R)\times \mathsf{M}_{r\times n}(\mathbb R)\times \mathbb R^h;\ \mathfrak{p}_{u}(X,Y,z)=0\right\}. \end{equation} We claim that $\mathscr{V}$ has at least $m$ connected components. In fact, if $k,\ell\in \{1,\ldots,m\}$ are distinct, then $(\mathsf{B}_k,\mathsf{C}_k,\zeta_k)$ and $(\mathsf{B}_\ell,\mathsf{C}_\ell,\zeta_\ell)$ belong to different connected component of $\mathscr{V}$. Indeed, suppose for the sake of obtaining a contradiction that $\mathcal{C}\subset \mathscr{V}$ is a connected subset of $\mathscr{V}$ and $(\mathsf{B}_k,\mathsf{C}_k,\zeta_k),(\mathsf{B}_\ell,\mathsf{C}_\ell,\zeta_\ell)\in \mathcal{C}$. Since $k\neq \ell$, by switching the roles of $k$ and $\ell$ if necessary, the assumption of Lemma~\ref{lem:rank} ensures that there exists $(i,j)\in \mathsf{E}$ such that $(\mathsf{B}_k\mathsf{C}_k)_{ij}=a_{ij}^k<0<a_{ij}^\ell=(\mathsf{B}_\ell \mathsf{C}_\ell)_{ij}$. So, if we denote $\psi:\mathcal{C}\to \mathbb R$ by $\psi(X,Y,z)=(XY)_{ij}$, then $\psi(\mathsf{B}_k,\mathsf{C}_k,\zeta_k)<0<\psi(\mathsf{B}_\ell,\mathsf{C}_\ell,\zeta_\ell)$. Since $\mathcal{C}$ is connected and $\psi$ is continuous, it follows that $\psi(X,Y,z)=0$ for some $(X,Y,z)\in \mathcal{C}$. Let $u\in \{1,\ldots,h\}$ be the index for which $(i,j)\in \mathsf{J}_u$. By the definition~\eqref{eq:def pu} of $\mathfrak{p}_u$, the fact that $\psi(X,Y,z)=(XY)_{ij}=0$ implies that $\mathfrak{p}_u(X,Y,z)=-z_u^2-\frac12\mu^{|J_u|}\le -\frac12\mu^{|J_u|}<0$, since $\mu>0$. Hence $(X,Y,z)\notin \mathscr{V}$, in contradiction to our choice of $(X,Y,z)$ as an element of $\mathcal{C}\subset \mathscr{V}$. Recalling~\eqref{eq:def pu}, for all $u\in \{1,\ldots,h\}$ the degree of $\mathfrak{p}_u$ is $4|\mathsf{J}_u|\le 4\alpha$. So, the variety $\mathscr{V}$ in~\eqref{eq:def variety} is defined using $h$ polynomials of degree at most $4\alpha$ in $2nr+h$ variables. By (a special case of) a theorem of Milnor~\cite{Mil64} and Thom~\cite{Tho65}, the number of connected components of $\mathscr{V}$ is at most $4\alpha(8\alpha-1)^{2nr+h-1}=4\alpha(8\alpha-1)^{2nr+\lceil |\mathsf{E}|/\alpha\rceil-1}$. Since we already established that this number of connected components is at least $m$, it follows that \begin{equation}\label{eq:alpha to optimize} m\le 4\alpha(8\alpha-1)^{2nr+\left\lceil \frac{|\mathsf{E}|}{\alpha}\right\rceil-1}\iff r\ge \frac{1}{2n} \left(\frac{\log\left(\frac{m}{4\alpha}\right)}{\log(8\alpha-1)}-\left\lceil \frac{|\mathsf{E}|}{\alpha}\right\rceil+1\right). \end{equation} The value of $\alpha\in \mathbb N$ that maximizes the rightmost quantity in~\eqref{eq:alpha to optimize} satisfies $$ \alpha\asymp \frac{|\mathsf{E}|}{\log (2m)}\left(\log\left(\frac{|\mathsf{E}|}{\log (2m)}\right)\right)^2. $$ For this $\alpha$ the estimate~\eqref{eq:alpha to optimize} simplifies to imply the desired bound $r\gtrsim \log m/(n\log(|\mathsf{E}|/\log m))$. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:prob matousek}] For notational convenience, write $\mathsf{L}=\{\lambda_1,\ldots,\lambda_n\}$ and $\mathsf{R}=\{\rho_1,\ldots,\rho_n\}$, and think of $\mathsf{E}$ as a subset of $\{1,\ldots,n\}^2$ (i.e., $(i,j)\in \mathsf{E}$ if and only if $\{\lambda_i,\rho_j\}$ is an edge of $\mathsf{G}$). For every $\Delta\in \mathbb N$ denote \begin{equation}\label{eq:def bad event} \mathscr{B}_{\!\Delta}\stackrel{\mathrm{def}}{=} \left\{\sigma:\mathsf{E}\to \{-,+\}:\ \dim_{\omega,\Omega}\Big(\mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},d_\sigma^{s,T}\Big)\le \Delta\right\}. \end{equation} Then, by the definition of $\dim_{\omega,\Omega}(\cdot)$, if $\sigma\in \mathscr{B}_{\!\Delta}$, we can fix a normed space $(X_\sigma,\|\cdot\|_{X_\sigma})$ with $\dim(X_\sigma)=\Delta$ and a mapping $f_\sigma: \mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R}\to X_\sigma$ that satisfies \begin{equation}\label{eq:omega Omega truncated} \forall\, x,y\in \mathsf{L}^{\!+}\cup\mathsf{L}^{\!-}\cup \mathsf{R},\qquad \omega\big(d_\sigma^{s,T}(x,y)\big)\le \|f_\sigma(x)-f_\sigma(y)\|_{X_\sigma}\le \Omega\big(d_\sigma^{s,T}(x,y)\big). \end{equation} Using the Hahn--Banach theorem, for each $i\in \{1,\ldots,n\}$ and $\sigma\in \mathscr{B}_{\!\Delta}$ we can fix a linear functional $z_{\sigma,i}^*\in X_\sigma^*$ of unit norm that normalizes the vector $f_\sigma(\lambda_i^{\!+})-f_\sigma(\lambda_i^{\!-})\in X_\sigma$, i.e., \begin{equation}\label{eq:HB-mat} z^*_{\sigma,i}\big(f_\sigma(\lambda^{\!+}_i)-f_\sigma(\lambda^{\!-}_i)\big)=\big\|f_\sigma(\lambda^{\!+}_i)-f_\sigma(\lambda^{\!+}_i)\big\|_{X_\sigma} \qquad\mathrm{and}\qquad \|z^*_{\sigma,i}\|_{X_\sigma^*}=\sup_{w\in X_\sigma\setminus \{0\}}\frac{\big|z_{\sigma,i}^*(w)\big|}{\|w\|_{X_\sigma}}=1. \end{equation} Using these linear functionals, define an $n\times n$ matrix $\mathsf{A}_\sigma=(a^\sigma_{ij})\in \mathsf{M}_n(\mathbb R)$ by setting \begin{equation}\label{eq:our gram} \forall(i,j)\in \{1,\ldots,n\}^2,\qquad a_{ij}^\sigma\stackrel{\mathrm{def}}{=} z_{\sigma,i}^*\big(f_\sigma(\rho_j)\big)-\frac12 z_{\sigma,i}^*\big( f_\sigma(\lambda_i^{\!+})+f_\sigma(\lambda_i^{\!-})\big). \end{equation} Observe in passing that the following identity holds true for every $i,j\in \{1,\ldots,n\}$ and $\sigma \in \mathscr{B}_{\!\Delta}$. \begin{equation}\label{eq:AR identity} \sigma(i,j)a_{ij}^\sigma=\frac12 z^*_{\sigma,i}\big(f_\sigma(\lambda^{\!+}_i)-f_\sigma(\lambda^{\!-}_i)\big) +\sigma(i,j)z_{\sigma,i}^*\Big(f_\sigma(\rho_j)-f_\sigma\Big(\lambda_i^{\sigma(i,j)}\Big)\Big). \end{equation} (Simply verify~\eqref{eq:AR identity} for the cases $\sigma(i,j)=+$ and $\sigma(i,j)=-$ separately, using the linearity of $z^*_{\sigma,i}$.) Since we are assuming in Lemma~\ref{lem:prob matousek} that the shortest cycle in the template graph $\mathsf{G}$ has length at least $g$, it follows from Claim~\ref{claim:cycle compress} that $d_{\mathsf{G}_\sigma}(\lambda_i^{\!+},\lambda_i^{\!-})\ge g$ for all $i\in \{1,\ldots,n\}$ and $\sigma:\mathsf{E}\to \{-,+\}$. So, \begin{equation}\label{eq:plus minus dist big} d_\sigma^{s,T}(\lambda_i^{\!+},\lambda_i^{\!-})\stackrel{\eqref{eq:truncated}}{=}\min\left\{sd_{\mathsf{G}_\sigma}(\lambda_i^{\!+},\lambda_i^{\!-}),T\right\}\ge \min\{sg,T\}\stackrel{\eqref{eq:girth assumtpions}}{=}sg. \end{equation} Recalling~\eqref{eq:def signed edges}, we have $\{\lambda_i^{\sigma(i,j)},\rho_j\}\in \mathsf{E}_\sigma$ for all $(i,j)\in \mathsf{E}$. Hence $d_{\mathsf{G}_\sigma}(\lambda_i^{\sigma(i,j)},\rho_j)=1$ and therefore \begin{equation}\label{eq:truncated upper edge} d_\sigma^{s,T}\big(\lambda_i^{\sigma(i,j)},\rho_j\big)\stackrel{\eqref{eq:truncated}}{\le} sd_{\mathsf{G}_\sigma}\big(\lambda_i^{\sigma(i,j)},\rho_j\big)=s. \end{equation} Consequently, for every $(i,j)\in \mathsf{E}$ and $\sigma\in \mathscr{B}_{\!\Delta}$ we have \begin{eqnarray*} \sigma(i,j)a_{ij}^\sigma&\stackrel{\eqref{eq:AR identity}}{\ge} & \frac12 z^*_{\sigma,i}\big(f_\sigma(\lambda^{\!+}_i)-f_\sigma(\lambda^{\!-}_i)\big)-\left\|z_{\sigma,i}^*\right\|_{X_\sigma^*}\cdot \Big\|f_\sigma(\rho_j)-f_\sigma\Big(\lambda_i^{\sigma(i,j)}\Big)\Big\|_{X_\sigma}\\&\stackrel{\eqref{eq:HB-mat}}{=}&\frac12 \big\|f_\sigma(\lambda^{\!+}_i)-f_\sigma(\lambda^{\!-}_i)\big\|_{X_\sigma}-\Big\|f_\sigma(\rho_j)-f_\sigma\Big(\lambda_i^{\sigma(i,j)}\Big)\Big\|_{X_\sigma}\\ &\stackrel{\eqref{eq:omega Omega truncated}}{\ge} &\frac12\omega \big(d_\sigma^{s,T}(\lambda_i^{\!+},\lambda_i^{\!-})\big)-\Omega \Big(d_\sigma^{s,T}\big(\lambda_i^{\sigma(i,j)},\rho_j\big)\Big)\\ &\stackrel{\eqref{eq:plus minus dist big}\wedge \eqref{eq:truncated upper edge}}{\ge}& \frac12 \omega(sg)-\Omega(s)\stackrel{\eqref{eq:girth assumtpions}}{>}0. \end{eqnarray*} Hence, $a_{ij}^\sigma\neq 0$ and $\mathrm{\bf sign}(a_{ij}^\sigma)=\sigma(i,j)$ for all $(i,j)\in \mathsf{E}$ and $\sigma\in \mathscr{B}_{\!\Delta}$. This is precisely the setting of Lemma~\ref{lem:rank} (with $m=|\mathscr{B}_{\!\Delta}|$), from which we conclude that there exists $\tau\in \mathscr{B}_{\!\Delta}$ such that \begin{equation}\label{eq:rank lower use} \mathrm{\bf rank}(\mathsf{A}_\tau)\ge \frac{c\log |\mathscr{B}_{\!\Delta}|}{n\log\left(\frac{|\mathsf{E}|}{\log |\mathscr{B}_{\!\Delta}|}\right)}, \end{equation} where $c\in (0,\infty)$ is a universal constant. Henceforth, we shall fix a specific $\tau\in \mathscr{B}_{\!\Delta}$ as in~\eqref{eq:rank lower use}. Since $\tau\in \mathscr{B}_{\!\Delta}$ we have $\dim(X_\tau)=\Delta$, so we can fix a basis $e_{\tau}^1,\ldots,e_{\tau}^\Delta$ of $X_\tau$ and for every $j\in \{1,\ldots,n\}$ write $f_\tau(\rho_j)=\gamma_{\tau,j}^1e_{\tau}^1+\ldots+\gamma_{\tau,j}^\Delta e_{\tau}^\Delta$ for some scalars $\gamma_{\tau,j}^1,\ldots,\gamma_{\tau,j}^\Delta\in \mathbb R$. Hence, $$ (a_{ij}^\tau)_{i=1}^n\stackrel{\eqref{eq:our gram}}{=}\gamma_{\tau,j}^1 \big(z_{\tau,i}^*(e_\tau^1)\big)_{i=1}^n+\ldots+\gamma_{\tau,j}^\Delta \big(z_{\tau,i}^*(e_\tau^\Delta)\big)_{i=1}^n-\frac12\big(z_{\sigma,i}^*\big( f_\tau(\lambda_i^{\!+})+f_\tau(\lambda_i^{\!-})\big)\big)_{i=1}^n. $$ We have thus expressed the columns of the matrix $\mathsf{A}_\tau$ as elements of the span of the $\Delta+1$ vectors $$ \big(z_{\tau,i}^*(e_\tau^1)\big)_{i=1}^n,\big(z_{\tau,i}^*(e_\tau^2)\big)_{i=1}^n,\ldots,\big(z_{\tau,i}^*(e_\tau^\Delta)\big)_{i=1}^n,\big(z_{\sigma,i}^*\big( f_\tau(\lambda_i^{\!+})+f_\tau(\lambda_i^{\!-})\big)\big)_{i=1}^n\in \mathbb R^n. $$ Consequently, the rank of $\mathsf{A}_\tau$ is at most $\Delta+1\le 2\Delta$. By contrasting this with~\eqref{eq:rank lower use}, we see that \begin{equation}\label{eq:almost matousek punchline} \frac{|\mathsf{E}|}{\log |\mathscr{B}_{\!\Delta}|}\log\left(\frac{|\mathsf{E}|}{\log |\mathscr{B}_{\!\Delta}|}\right)\ge \frac{c|\mathsf{E}|}{2\Delta n}. \end{equation} We shall now conclude by showing that Lemma~\ref{lem:prob matousek} holds true with $\eta=c/2$. Indeed, fix $\delta\in (0,\frac13]$ and observe that we may assume also that $\d\eta|\mathsf{E}|/n\ge 1$, since otherwise the left hand side of~\eqref{eq:probabilistic version'} vanishes. Then, by choosing $\Delta=\lfloor \d\eta|\mathsf{E}|/n\rfloor\in \mathbb N$ in the above reasoning it follows from~\eqref{eq:almost matousek punchline} that $$ \frac{|\mathsf{E}|}{\log |\mathscr{B}_{\!\Delta}|}\log\left(\frac{|\mathsf{E}|}{\log |\mathscr{B}_{\!\Delta}|}\right)\ge \frac{c}{2\d\eta}=\frac{1}{\d}\ge 3>e. $$ This implies that $|\mathscr{B}_{\!\Delta}|\le \d^{-\d|\mathsf{E}|}$. Equivalently, $\Pr[\mathscr{B}_{\!\Delta}]\le (2\d)^{-\d|\mathsf{E}|}$, which is the desired bound~\eqref{eq:probabilistic version'}. \end{proof} \section{Nonlinear spectral gaps and impossibility of average dimension reduction}\label{sec:average} Fix $n\in \mathbb N$ and an irreducible reversible row-stochastic matrix $\mathsf{A}=(a_{ij})\in \mathsf{M}_n(\mathbb R)$. This implies that there is a unique\footnote{We are assuming irreducibility only for notational convenience, namely so that $\pi$ will be unique and could therefore be suppressed in the ensuing notation. Our arguments work for any stochastic matrix and any probability measure $\pi$ on $\{1,\ldots,n\}$ with respect to which $\mathsf{A}$ is reversible. We suggest focusing initially on the case when $\mathsf{A}$ is symmetric and $\pi$ is the uniform measure on $\{1,\ldots,n\}$, though the general case is useful for treating graphs that are not regular, e.g.~those of Section~\ref{sec:matousek}. See~\cite{LPW17} for the relevant background.} $\AA$-stationary probability measure $\pi=(\pi_1,\ldots,\pi_n)\in [0,1]^n$ on $\{1,\ldots,n\}$, namely $\pi\AA=\pi$, and we have the reversibility condition $\pi_ia_{ij}=\pi_ja_{ji}$ for all $i,j\in \{1,\ldots,n\}$. Then $\AA$ is a self-adjoint contraction on $L_2(\pi)$, and we denote by $1=\lambda_1(\AA)\ge \lambda_2(\AA)\ge\ldots\ge \lambda_n(\AA)\ge -1$ the decreasing rearrangement of its eigenvalues. Given a metric space $(\mathcal{M},d_\mathcal{M})$ and $p>0$, define $\gamma(\AA,d_\mathcal{M}^p)$ to be the infimum over those $\gamma>0$ such that \begin{equation}\label{eq:def nonlinear gap} \forall\, x_1,\ldots,x_n\in \mathcal{M},\qquad \sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j d_\mathcal{M}(x_i,x_j)^p\le \gamma\sum_{i=1}^n\sum_{j=1}^n \pi_ia_{ij} d_\mathcal{M}(x_i,x_j)^p. \end{equation} This definition is implicit in~\cite{Gro03}, and appeared explicitly in~\cite{NS11}; see~\cite{MN14,Nao14,MN15} for a detailed treatment. It suffices to note here that if $(H,\|\cdot\|_H)$ is a Hilbert space and $p=2$, then by expanding the squares one directly sees that $\gamma(\AA,\|\cdot\|_H^2)=1/(1-\lambda_2(\AA))$ is the reciprocal of the {\em spectral gap} of $\AA$. In general, we think of $\gamma(\AA,d_\mathcal{M}^p)$ as measuring the magnitude of the nonlinear spectral gap of $\AA$ with respect to the kernel $d_\mathcal{M}^p:\mathcal{M}\times \mathcal{M}\to [0,\infty)$. Using the notation that was recalled in Section~\ref{sec:bilip}, the definition~\eqref{eq:def nonlinear gap} immediately implies that nonlinear spectral gaps are bi-Lipschitz invariants in the sense that $\gamma(\AA,d_\mathcal{M}^p)\le \mathsf{c}_\mathcal{N}(\mathcal{M})^p\gamma(\AA,d_\mathcal{N}^p)$ for every two metric spaces $(\mathcal{M},d_\mathcal{M})$ and $(\mathcal{N},d_\mathcal{N})$, every matrix $\AA$ as above and every $p>0$. In particular, if $(H,\|\cdot\|_H)$ is a Hilbert space into which $(\mathcal{M},d_\mathcal{M})$ admits a bi-Lipschitz embedding, then we have the following general (trivial) bound. \begin{equation}\label{eq:trivial spectral gap bound} \sqrt{\gamma(\AA,d_\mathcal{M}^2)}\le \mathsf{c}_2(\mathcal{M})\sqrt{\gamma(\AA,\|\cdot\|_H^2)}. \end{equation} In the recent work~\cite{Nao17} we proved the following theorem, which improves over~\eqref{eq:trivial spectral gap bound} when $\mathcal{M}$ is a Banach space. \begin{theorem}\label{thm:our gamma} Suppose that $(X,\|\cdot\|_X)$ is a Banach space and that $(H,\|\cdot\|_H)$ is a Hilbert space. Then for every $M\in (0,\infty)$ and every matrix $\AA$ as above for which $\lambda_2(\AA)\le 1-M^2/\mathsf{c}_2(X)^2$ we have \begin{equation}\label{eq:M version} \sqrt{\gamma\big(\mathsf{A},\|\cdot\|_X^2\big)}\lesssim \frac{\log (M+1)}{M}\mathsf{c}_2(X)\sqrt{\gamma(\AA,\|\cdot\|_H^2)}. \end{equation} \end{theorem} In the setting of Theorem~\ref{thm:our gamma}, since $\gamma(\AA,\|\cdot\|_H^2)=1/(1-\lambda_2(\AA))$, the bound~\eqref{eq:M version} can be rewritten as \begin{equation*} \gamma\big(\mathsf{A},\|\cdot\|_X^2\big)\lesssim \bigg(\frac{\log(\mathsf{c}_2(X)\sqrt{1-\lambda_2(\mathsf{A})}+1)}{1-\lambda_2(\mathsf{A})}\bigg)^2, \end{equation*} which is how Theorem~\ref{thm:our gamma} was stated in~\cite{Nao17}. Note that~\eqref{eq:M version} coincides (up to the implicit constant factor) with the trivial bound~\eqref{eq:trivial spectral gap bound} if $M=O(1)$, but~\eqref{eq:M version} is an asymptotic improvement over~\eqref{eq:trivial spectral gap bound} as $M\to\infty$. The proof of Theorem~\ref{thm:our gamma} in~\cite{Nao17} is a short interpolation argument that takes as input a theorem from~\cite{Nao14}. While we do not know of a different proof of~\eqref{eq:M version}, below we will present a new and self-contained derivation of the following weaker estimate (using the same notation as in Theorem~\ref{thm:our gamma}) that suffices for deducing Theorem~\ref{thm:average distortion}. \begin{equation}\label{eq:M-d version} \sqrt{\gamma\big(\mathsf{A},\|\cdot\|_X^2\big)}\lesssim \frac{\log (\mathsf{c}_2(X)+1)}{M}\mathsf{c}_2(X)\sqrt{\gamma(\AA,\|\cdot\|_H^2)}. \end{equation} The proof of~\eqref{eq:M-d version} appears in Section~\ref{sec:ray} below. We will next show how~\eqref{eq:M-d version} implies Theorem~\ref{thm:average distortion} as well as the leftmost inequality in~\eqref{eq:linfty bounds in theorem}, which includes as a special case the triple logarithmic estimate in~\eqref{eq:log log log}. Given $n\in \mathbb N$ and an $n$-vertex connected graph $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_{\mathsf{G}})$, let $\AA_\mathsf{G}$ be its random walk matrix, i.e., if $\deg_\mathsf{G}(i)$ is the degree in $\mathsf{G}$ of the vertex $i\in \{1,\ldots,n\}$, then $(\AA_\mathsf{G})_{ij}=\mathbf{1}_{\{i,j\}\in \mathsf{E}_\mathsf{G}}/\deg_\mathsf{G}(i)$ for all $i,j\in \{1,\ldots,n\}$. For $i\in \{1,\ldots,n\}$ we also denote $\pi_i^\mathsf{G}=\deg_\mathsf{G}(i)/(2|\mathsf{E}_\mathsf{G}|)$. Thus, $\pi^\mathsf{G}\in \mathbb R^n$ is the probability measure on $\{1,\ldots,n\}$ with respect to which $\AA_\mathsf{G}$ is reversible. We will use the simpler notation $\lambda_i(\AA_\mathsf{G})=\lambda_i(\mathsf{G})$ for every $i\in \{1,\ldots,n\}$. For $p\in (0,\infty)$and a metric space $(\mathcal{M},d_\mathcal{M})$, we will write $\gamma(\mathsf{G},d_\mathcal{M}^2)=\gamma(\AA_\mathsf{G},d_\mathcal{M}^2)$. The shortest-path metric that is induced by $\mathsf{G}$ on $\{1,\ldots,n\}$ will be denoted $d_\mathsf{G}:\{1,\ldots,n\}\times \{1,\ldots,n\}\to \mathbb N\cup\{0\}$. \begin{theorem}\label{thm:pi version average embedding} There is a universal constant $K>1$ with the following property. Fix $n\in \mathbb N$ and $\alpha\ge 1$. Let $\AA\in \mathsf{M}_n(\mathbb R)$ and $\pi\in [0,1]^n$ be as above. For every normed space $(X,\|\cdot\|_X)$, if $f:\{1,\ldots,n\}\to X$ satisfies \begin{equation}\label{eq:two averages} \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i a_{ij}\|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le \alpha, \end{equation} then necessarily \begin{equation}\label{eq:dim condition} \dim(X)\gtrsim K^{\frac{1-\lambda_2(\AA)}{\alpha}\sqrt{\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j \|f(i)-f(j)\|_X^2}}. \end{equation} In particular, in the special case when $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ is a connected graph we have \begin{equation*}\label{eq:two averages-graph} \bigg(\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le \alpha\implies \dim(X)\gtrsim K^{\frac{1-\lambda_2(\mathsf{G})}{\alpha}\sqrt{\sum_{i=1}^n\sum_{j=1}^n \pi_i^\mathsf{G}\pi_j^\mathsf{G} \|f(i)-f(j)\|_X^2}}. \end{equation*} \begin{comment} \begin{equation}\label{eq:upper W12 norm} \bigg(\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le \alpha, \end{equation} and \begin{equation}\label{eq:lower quadratic aaverage dist} \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\ge \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j d_\mathsf{G}(i,j)^2\bigg)^{\frac12}, \end{equation} then necessarily \begin{equation*}\label{eq:dim condition-graph} \dim(X)\gtrsim K^{\frac{1-\lambda_2(\mathsf{G})}{\alpha}\sqrt{\sum_{i=1}^n\sum_{j=1}^n \pi_i^\mathsf{G}\pi_j^\mathsf{G} \|f(i)-f(j)\|_X^2}}. \end{equation*} \end{comment} \end{theorem} In the case of regular graphs with a spectral gap, Theorem~\ref{thm:pi version average embedding} has the following corollary. \begin{corollary}\label{cor:regular case} Fix two integers $n,r\ge 3$ and let $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ be a connected $r$-regular graph. If $(X,\|\cdot\|_X)$ is a normed space into which there is a mapping $f:\{1,\ldots,n\}\to X$ that satisfies \begin{equation}\label{eq:two averages-regular} \bigg(\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le \alpha\qquad\mathrm{and}\qquad \bigg(\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\ge \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n d_\mathsf{G}(i,j), \end{equation} then necessarily $$ \dim(X)\gtrsim n^{\frac{c(1-\lambda_2(\mathsf{G}))}{\alpha\log r}}, $$ where $c\in (0,\infty)$ is a universal constant. \end{corollary} \begin{proof}This is nothing more than a special case of Theorem~\ref{thm:pi version average embedding} once we note that by a straightforward and standard counting argument (see e.g.~\cite{Mat97}) we have $\frac{1}{n}\sum_{i=1}^n\sum_{j=1}^n d_\mathsf{G}(i,j)\gtrsim \log_r n$. \end{proof} For every integer $r\ge 3$ there exist arbitrarily large $r$-regular graphs $\mathsf{G}$ with $\lambda_2(\mathsf{G})=1-\Omega(1)$; see~\cite{HLW} for this and much more on such {\em spectral expanders}. Corollary~\ref{cor:regular case} shows that the shortest-path metric on any such graph with $r=O(1)$ satisfies the conclusion of Theorem~\ref{thm:average distortion}, because the $\alpha$-Lipschitz assumption of Theorem~\ref{thm:average distortion} implies the first inequality in~\eqref{eq:two averages-regular} and the assumption $\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X\ge \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n d_\mathsf{G}(i,j)$ of Theorem~\ref{thm:average distortion} implies the second inequality in~\eqref{eq:two averages-regular} (using Jensen's inequality). Note that we actually proved above that any expander is "metrically high dimensional" in a stronger sense. Specifically, if $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ is a $O(1)$-spectral expander, i.e., it is $O(1)$-regular and $\lambda_2(\mathsf{G})\le 1-\Omega(1)$, and one finds vectors $x_1,\ldots,x_n$ in a normed space $(X,\|\cdot\|_X)$ for which the averages $\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|x_i-x_j\|_X^2$ and $\frac{1}{n^2}\sum_{i,j=1}^n \|x_i-x_j\|_X^2$ are within a $O(1)$ factor of the averages $\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} d_\mathsf{G}(i,j)^2=1$ and $\frac{1}{n^2}\sum_{i,j=1}^n d_\mathsf{G}(i,j)^2$, respectively, then this "finitary average distance information" (up to a fixed but potentially very large multiplicative error) forces the ambient space $X$ to be very high (worst-possible) dimensional, namely $\dim(X)\ge n^{\Omega(1)}$. \begin{remark}\label{rem:for average snowflake} If one replaces~\eqref{eq:two averages-regular} by the requirement that for an increasing modulus $\omega:[0,\infty)\to [0,\infty)$ we have, \begin{equation}\label{eq:two averages-regular-omega} \bigg(\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le 1\qquad\mathrm{and}\qquad \bigg(\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\ge \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \omega\big(d_\mathsf{G}(i,j)\big), \end{equation} then the above argument applies mutatis mutandis to yield the conclusion \begin{equation}\label{eq:omega lower dimension} \dim(X)\gtrsim e^{c(1-\lambda_2(\mathsf{G}))\omega(c\log_r n)}. \end{equation} Indeed, the aforementioned counting argument shows that least $50\%$ of the pairs $(i,j)\in \{1,\ldots,n\}^2$ satisfy $d_\mathsf{G}(i,j)\gtrsim \log_r n$. Compare~\eqref{eq:omega lower dimension} to Theorem~\ref{thm:coarse matousek} which provides a stronger bound if the average requirement~\eqref{eq:two averages-regular-omega} is replaced by its pairwise counterpart~\eqref{eq:coarse condition}. Nevertheless, the bound~\eqref{eq:omega lower dimension} is quite sharp (at least when $r=O(1)$ and $\lambda_2(\mathsf{G})=1-\Omega(1)$), in the sense that there is a normed space $(X,\|\cdot\|_X)$ for which~\eqref{eq:two averages-regular-omega} holds and \begin{equation}\label{eq:dim upper compression expander} \dim(X)\lesssim e^{C(\log r)\omega\left(\frac{C\log n}{\sqrt{1-\lambda_2(\mathsf{G})}}\right)}\log n, \end{equation} where $C>0$ is a universal constant. Indeed, by~\cite{CFM94} the diameter of the metric space $(\{1,\ldots,n\},d_\mathsf{G})$ satisfies $\diam(\mathsf{G})\lesssim (\log n)/\sqrt{1-\lambda_2(\mathsf{G})}$. By an application of~\eqref{eq:linfty bounds in theorem} with $\alpha \asymp (\log_r n)/\omega(\diam(\mathsf{G}))$ there exists a normed space $(X,\|\cdot\|_X)$ with $\dim(X)\lesssim n^{O(1/\alpha)}\log n$, thus~\eqref{eq:dim upper compression expander} holds, and a mapping $f:\{1,\ldots,n\}\to X$ that satisfies $d_\mathsf{G}(i,j)/\alpha\le\|f(i)-f(j)\|_X\le d_\mathsf{G}(i,j)$ for all $i,j\in \{1,\ldots,n\}$. Hence, the first inequality in~\eqref{eq:two averages-regular-omega} holds, and $$ \bigg(\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\ge \frac{1}{\alpha} \bigg(\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n d_\mathsf{G}(i,j)^2\bigg)^{\frac12}\gtrsim\frac{\log_r n}{\alpha}\asymp\omega\big(\diam(\mathsf{G})\big)\ge \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \omega\big(d_\mathsf{G}(i,j)\big). $$ \end{remark} \begin{proof}[Proof of Theorem~\ref{thm:pi version average embedding} assuming~\eqref{eq:M-d version}] Let $C\in (0,\infty)$ be the implicit universal constant in~\eqref{eq:M-d version}. Then \begin{multline}\label{eq:use log gap} \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\stackrel{\eqref{eq:def nonlinear gap}}{\le} \sqrt{\gamma(\AA,\|\cdot\|_X^2)}\bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i a_{ij} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\\\stackrel{\eqref{eq:two averages}}{\le} \alpha\sqrt{\gamma(\AA,\|\cdot\|_X^2)}\stackrel{\eqref{eq:M-d version}}{\le} \frac{C\alpha\log(\mathsf{c}_2(X)+1)}{1-\lambda_2(\AA)}, \end{multline} where the last step of~\eqref{eq:use log gap} is an application of~\eqref{eq:M-d version} with $M=\mathsf{c}_2(X)\sqrt{1-\lambda_2(\AA)}$, while using that for a Hilbert space $H$ we have $\gamma(\AA,\|\cdot\|_H^2)=1/(1-\lambda_2(\AA))$. It follows that \begin{equation}\label{eq:use john} 2\sqrt{\dim(X)}\ge 2\mathsf{c}_2(X)\ge\mathsf{c}_2(X)+1\stackrel{\eqref{eq:use log gap}}{\ge} e^{\frac{1-\lambda_2(\AA)}{C\alpha}\sqrt{\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j \|f(i)-f(j)\|_X^2}}, \end{equation} where the first step of~\eqref{eq:use john} uses John's theorem~\cite{Joh48}. This establishes~\eqref{eq:dim condition} with $K=e^{2/C}>1$. \end{proof} For non-contracting embeddings (in particular, for bi-Lipschitz embedding), the proof of the following lemma is an adaptation of the proof of~\cite[Theorem~13]{ABN11}. \begin{lemma}\label{lem:ABN referee}Fix two integers $n,r\ge 3$ and let $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ be a connected $r$-regular graph. If $(X,\|\cdot\|_X)$ is a normed space into which there is a mapping $f:\{1,\ldots,n\}\to X$ that satisfies \begin{equation}\label{eq:non contracting} \min_{\substack{i,j\in \{1,\ldots,n\}\\ i\neq j}}\frac{\|f(i)-f(j)\|_X}{d_\mathsf{G}(i,j)}\ge 1,\qquad \mathrm{and}\qquad \bigg(\frac{1}{|\mathsf{E}_\mathsf{G}|}\sum_{\{i,j\}\in \mathsf{E}_\mathsf{G}} \|f(i)-f(j)\|_X^2\bigg)^{\frac12}\le \alpha. \end{equation} Then necessarily \begin{equation}\label{eq:ball net conclusion} \frac{\log n}{\log r} n^{\frac{1}{2\dim(X)}}\lesssim \alpha \sqrt{\gamma(\mathsf{G},\|\cdot\|_X^2)}. \end{equation} \end{lemma} Prior to proving Lemma~\ref{lem:ABN referee}, we will derive some of its corollaries. For the terminology of Corollary~\ref{cor:nontrivial type} below, recall that a Banach space $Y$ is said to be $B$-convex~\cite{Bec62} if $\ell_1$ is {\em not} finitely representable in $Y$; see the survey~\cite{Mau03} for more on this important notion, including useful analytic, geometric and probabilistic characterizations. \begin{corollary}\label{cor:nontrivial type} There is a universal constant $C\in (0,\infty)$ for which the following assertion holds true. Let $Y$ be an infinite dimensional $B$-convex Banach space. For arbitrarily large $n\in \mathbb N$, if $\alpha\ge C\log n$, then we have \begin{equation}\label{eq:K-convex dim reduction} \mathsf{k}_n^\alpha(\ell_\infty,Y)\asymp_Y \frac{\log n}{\log\left(\frac{\alpha}{\log n}\right)}. \end{equation} Thus, we have in particular $\mathsf{k}_n^{C\log n}(\ell_\infty,Y)\asymp_Y\log n$. \end{corollary} \begin{proof} The upper bound $\mathsf{k}_n^\alpha(\ell_\infty,Y)\lesssim(\log n)/\log(\alpha/\log n)$ actually holds for {\em any} infinite dimensional Banach space $Y$. Indeed, by Bourgain's embedding theorem~\cite{Bou85} any $n$-point metric space $\mathcal{M}$ admits an embedding $f$ into $\ell_2$ with distortion $A\log n$, where $A\in (0,\infty)$ is a universal constant. If $\alpha\ge 4A\log n$, then by applying the JL-Lemma~\cite{JL84} we know that $f(\mathcal{M})$ embeds with distortion $\alpha/(2A\log n)$ into $\ell_2^k$, where $k\lesssim (\log n)/\log(\alpha/\log n)$. By Dvoretzky's theorem~\cite{Dvo60}, we know that $\ell_2^k$ embeds with distortion $2$ into $Y$, so overall we obtain an embedding of $\mathcal{M}$ into a $k$-dimensional subspace of $Y$ with distortion at most $2(A\log n)(\alpha/(2A\log n))=\alpha$. Conversely, suppose that $\alpha\ge 2\log n$ and that $Y$ is a $B$-convex Banach space. By a theorem of V. Lafforgue~\cite{Laf09} (see also~\cite{MN14} for a different approach), for arbitrarily large $n\in \mathbb N$ there is a $O(1)$-regular graph $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ such that $\gamma(\mathsf{G},\|\cdot\|_Y^2)\lesssim_Y 1$. If $(\{1,\ldots,n\},d_\mathsf{G})$ embeds with distortion $\alpha$ into a $k$-dimensional subspace of $Y$, then by Lemma~\ref{lem:ABN referee} we have $n^{1/(2k)}\lesssim_Y \alpha/\log n$. Thus $k\gtrsim_Y (\log n)/\log(\alpha/\log n)$, as required. \end{proof} \begin{question} Is the assumption of $B$-convexity needed for the conclusion~\eqref{eq:K-convex dim reduction} of Corollary~\ref{cor:nontrivial type}? Perhaps finite cotype suffices for this purpose? This matter is of course closely related to Question~\ref{Q:log n}. \end{question} \begin{corollary}\label{coro:combine ball with spectral} Under the assumptions and notation of Lemma~\ref{lem:ABN referee}, we have $$ \frac{\log n}{\log r} n^{\frac{1}{2\dim(X)}}\lesssim \frac{\alpha\log(\mathsf{c}_2(X)+1)}{1-\lambda_2(\mathsf{G})}. $$ \end{corollary} \begin{proof} This is nothing more than a substitution of~\eqref{eq:M-d version} with $M=\mathsf{c}_2(X)\sqrt{1-\lambda_2(\mathsf{G})}$ into~\eqref{eq:ball net conclusion}. \end{proof} Since by John's theorem~\cite{Joh48} we have $\mathsf{c}_2(X)\le \sqrt{\dim(X)}$ and for every $n\in \mathbb N$ there exists a graph $\mathsf{G}$ as in Lemma~\ref{lem:ABN referee} with $r=O(1)$ and $\lambda_2(\mathsf{G})=1-\Omega(1)$, it follows from Corollary~\ref{coro:combine ball with spectral} that \begin{equation*} \alpha \log(\mathsf{k}_n^\alpha(\ell_\infty)+1)\gtrsim n^{\frac{1}{2\mathsf{k}_n^\alpha(\ell_\infty)}}\log n. \end{equation*} This implies the lower bound on $\mathsf{k}_n^\alpha(\ell_\infty)$ in~\eqref{eq:linfty bounds in theorem}. In particular, for $\alpha\asymp \log n$ it gives the first inequality in~\eqref{eq:log log log}. \begin{proof}[Proof of Lemma~\ref{lem:ABN referee}] Denote $\gamma=\gamma(\mathsf{G},\|\cdot\|_X^2)$. For $i\in \{1,\ldots,n\}$ write \begin{equation}\label{eq:def Ui} \mathscr{U}_i=f^{-1}\Big(B_X\big(f(i),\alpha\sqrt{2\gamma}\big)\Big)=\Big\{j\in \{1,\ldots,n\}:\ \|f(i)-f(j)\|_X\le \alpha\sqrt{2\gamma}\Big\}. \end{equation} Let $m\in \{1,\ldots,n\}$ satisfy $|\mathscr{U}_m|=\max_{i\in \{1,\ldots,n\}}|\mathscr{U}_i|$. Then \begin{multline}\label{eq:use point again rho} n^2\gamma\alpha^2\stackrel{\eqref{eq:def nonlinear gap}\wedge \eqref{eq:non contracting}}{\ge} \sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X^2\ge \sum_{i=1}^n\sum_{j\in \{1,\ldots,n\}\setminus \mathscr{U}_i} \|f(i)-f(j)\|_X^2\\\stackrel{\eqref{eq:def Ui}}> \sum_{i=1}^n \big(\alpha\sqrt{2\gamma}\big)^2(n-|\mathscr{U}_i|)\ge 2n\gamma\alpha^2(n-|\mathscr{U}_m|). \end{multline} This simplifies to $|\mathscr{U}_m|\ge \frac12n$. Also, since $\frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n \|f(i)-f(j)\|_X^2\ge \frac{1}{n^2}\sum_{i=1}^n\sum_{j=1}^n d_\mathsf{G}(i,j)^2\gtrsim (\log_r n)^2$, the first inequality in~\eqref{eq:use point again rho} implies the a priori lower bound $\alpha\sqrt{\gamma}\gtrsim \log _r n$. Next, fix $\rho\in (0,\infty)$ and let $\mathscr{N}_{2\rho}\subset \mathscr{U}_m$ be a maximal (with respect to inclusion) $2\rho$-separated subset of $\mathscr{U}_m$. Then $\mathscr{U}_m\subset \cup_{i\in \mathscr{N}_\Delta} B_\mathsf{G}(i,2\rho)$, where $B_\mathsf{G}(i,2\rho)$ denotes the ball centered at $i$ of radius $2\rho$ in the shortest-path metric $d_\mathsf{G}$. Since $\mathsf{G}$ is $r$-regular, for each $i\in \{1,\ldots,n\}$ we have the (crude) bound $|B_\mathsf{G}(i,2\rho)|\le 2r^{2\rho}$. Hence, $\frac12n\le|\mathscr{U}_m|\le 2r^{2\rho} |\mathscr{N}_{2\rho}|$. So, if we choose $\rho= \frac14\log_r n$, then $|\mathscr{N}_{2\rho}|\gtrsim \sqrt{n}$. Since by~\eqref{eq:non contracting} distinct $i,j\in \mathscr{N}_{2\rho}$ satisfy $\|f(i)-f(j)\|_X\ge d_\mathsf{G}(i,j)\ge 2\rho$, the $X$-balls $\{B_X(f(i),\rho):\ i\in\mathscr{N}_{2\rho}\}$ have pairwise disjoint interiors. At the same time, since each $i\in \mathscr{N}_{2\rho}$ belongs to $\mathscr{U}_m$, we have $\|f(i)-f(m)\|_X\le \alpha\sqrt{2\gamma}$ (by the definition of $\mathscr{U}_m$), and hence $B_X(f(i), \rho)\subset B_X(f(m),\alpha\sqrt{2\gamma}+\rho)$. So, writing $\dim(X)=k$, we have the following volume comparison. \begin{multline*} (\alpha\sqrt{2\gamma}+\rho)^{k}\mathrm{\bf vol}_k\big(B_X(0,1)\big)=\mathrm{\bf vol}_k\big(B_X(f(m),\alpha\sqrt{2\gamma}+\rho)\big)\ge \mathrm{\bf vol}_k\bigg(\bigcup_{i\in \mathscr{N}_{2\rho}} B_X(f(i),\rho)\bigg) \\=\sum_{i\in \mathscr{N}_{2\rho}}\mathrm{\bf vol}_k\big(B_X(f(i),\rho)\big)=\rho^k\mathrm{\bf vol}_k\big(B_X(0,1)\big)|\mathscr{N}_{2\rho}|\gtrsim \rho^k\mathrm{\bf vol}_k\big(B_X(0,1)\big)\sqrt{n}. \end{multline*} This simplifies to give $ n^{\frac{1}{2k}}\lesssim \frac{\alpha\sqrt{2\gamma}}{\rho}+1\asymp \frac{\alpha\sqrt{\gamma}}{\log_r n}, $ where we used the definition of $\rho$, and that $\alpha\sqrt{\gamma}\gtrsim \log_r n$. \end{proof} \subsection{Nonlinear Rayleigh quotient inequalities}\label{sec:ray} Our goal in this section is to present a proof of~\eqref{eq:M-d version}. As we stated earlier, the proof that appears below is different from the proof of Theorem~\ref{thm:our gamma} in~\cite{Nao17}. However, the reason that underlies its validity is the same as that of the original argument in~\cite{Nao17}. Specifically, we arrived at the ensuing proof because we were driven by an algorithmic need that arose in~\cite{ANNRW18}. This need required proving a point-wise strengthening of an upper bound on nonlinear spectral gaps, which is called in~\cite{ANNRW18} a "nonlinear Rayleigh quotient inequality." We will clarify what we mean by this later; a detailed discussion appears in~\cite{ANNRW18}. The need to make the interpolation-based proof in~\cite{Nao17} constructive/algorithmic led us to merge the argument in~\cite{Nao17} with the {\em proof of} a theorem from~\cite{Nao14}, rather than quoting and using the latter as a "black box" as we did in~\cite{Nao17}. In doing so, we realized that for the purpose of obtaining only the weaker bound~\eqref{eq:M-d version} one could more efficiently combine~\cite{Nao14} and~\cite{Nao17} so as to skip the use of complex interpolation and to obtain the estimate~\eqref{eq:M-d version} as well as its nonlinear Rayleigh quotient counterpart. Thus, despite superficial differences, the argument below amounts to unravelling the proofs in~\cite{Nao14,Nao17} and removing steps that are needed elsewhere but not for~\eqref{eq:M-d version}. At present, we do not have a proof of the stronger inequality~\eqref{eq:M version} that differs from its proof in~\cite{Nao17}, and the interpolation-based approach of~\cite{Nao17} is used for more refined algorithmic results in the forthcoming work~\cite{ANNRW18-uniform}. We will continue using the notation/conventions that were set at the beginning of Section~\ref{sec:average}. Fix $p\ge 1$ and a metric space $(\mathcal{M},d_\mathcal{M})$. Let $L_p(\pi;\mathcal{M})$ be the metric space $(\mathcal{M}^n,d_{L_p(\pi;\mathcal{M})})$, where $d_{L_p(\pi;\mathcal{M})}:\mathcal{M}^n\times \mathcal{M}^n\to [0,\infty)$ is $$ \forall\,x=(x_1,\ldots,x_n),y=(y_1,\ldots,y_n)\in \mathcal{M}^n,\qquad d_{L_p(\pi;\mathcal{M})}(x,y)\stackrel{\mathrm{def}}{=} \bigg(\sum_{i=1}^n \pi_id_\mathcal{M}(x_i,y_i)^p \bigg)^{\frac{1}{p}}. $$ Throughout what follows, it will be notationally convenient to slightly abuse notation by considering $\mathcal{M}$ as a subset of $L_p(\pi;\mathcal{M})$ through its identification with the {\em diagonal} subset of $\mathcal{M}^n$, which is an isometric copy of $\mathcal{M}$ in $L_p(\pi;\mathcal{M})$. Namely, we identify each $x\in \mathcal{M}$ with the $n$-tuple $(x,x\ldots,x)\in \mathcal{M}^n$. If $x=(x_1,\ldots,x_n)\in L_p(\pi;\mathcal{M})\setminus \mathcal{M}$, then the corresponding {\em nonlinear Rayleigh quotient} is defined to be \begin{equation}\label{eq:ray def} \mathscr{R}(x;\mathsf{A},d_\mathcal{M}^p)\stackrel{\mathrm{def}}{=} \frac{\sum_{i=1}^n\sum_{j=1}^n\pi_i a_{ij}d_\mathcal{M}(x_i,x_j)^p}{\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j d_\mathcal{M}(x_i,x_j)^p}. \end{equation} The restriction $x\notin\mathcal{M}$ was made here only to ensure that the denominator in~\eqref{eq:ray def} does not vanish. By definition, \begin{equation}\label{eq:gamma ray relation} \gamma(\mathsf{A},d_\mathcal{M}^p)=\sup_{x\in L_p(\pi;\mathcal{M})\setminus \mathcal{M}} \frac{1}{\mathscr{R}(x;\mathsf{A},d_\mathcal{M}^p)}. \end{equation} Note that $L_p(\pi;X)$ is a Banach space for every Banach space $(X,\|\cdot\|_X)$. In this case, the matrix $\AA\in \mathsf{M}_n(\mathbb R)$ induces a linear operator $\AA\otimes \mathsf{Id}_X:L_p(\pi;X)\to L_p(\pi;X)$ that is given by $(\AA\otimes \mathsf{Id}_X)(x_1,\ldots x_n)=(\sum_{j=1}^n a_{ij}x_j)_{i=1}^n$. The following lemma records some simple and elementary general properties of nonlinear Rayleigh quotients. \begin{comment} Its third assertion is a simple a priori upper bound. We won't use this upper bound in the ensuing proof, but it is worthwhile to be aware of it so as to indicate that the pertinent issue is to bound $\mathscr{R}(x;A,d_\mathcal{M}^p)$ from below, which by~\eqref{eq:gamma ray relation} corresponds to bounding $\gamma(\mathsf{A},d_\mathsf{M}^p)$ from above. Nevertheless, it might be meaningful to obtain better a priori upper bounds under additional geometric assumptions on the underlying metric space $(\mathcal{M},d_\mathcal{M})$; e.g., if the underlying space is a Hilbert space and $p=2$, then one can improve the stated bound using the spectral interpretation of Hilbertain Rayleigh quotients. \end{comment} \begin{lemma}\label{lem:Mtype2} Suppose that $(\mathcal{M},d_\mathcal{M})$ is a metric space, $n\in \mathbb N$, $p\in [1,\infty)$ and $\d\in [0,1]$. Let $\pi=(\pi_1,\ldots,\pi_n)$ be a probability measure on $\{1,\ldots,n\}$ and $\mathsf{A},\mathsf{B}\in \mathsf{M}_n(\mathbb R)$ be row-stochastic matrices that are reversible with respect to $\pi$. For any $x\in L_p^n(\pi;\mathcal{M})\setminus \mathcal{M}$ we have \begin{enumerate} \item $\mathscr{R} \big(x;\d\mathsf{A}+(1-\d)\mathsf{B},d_\mathcal{M}^p\big)=\d\mathscr{R} \big(x;\mathsf{A},d_\mathcal{M}^p\big)+(1-\d)\mathscr{R} \big(x;\mathsf{B},d_\mathcal{M}^p\big)$. \item $ \mathscr{R} \big(x;(1-\d)\mathsf{Id}_n+\d\mathsf{A},d_\mathcal{M}^p\big)=\d\mathscr{R} \big(x;\mathsf{A},d_\mathcal{M}^p\big)$, where $\mathsf{Id}_n\in \mathsf{M}_n(\mathbb R)$ is the identity matrix. \item $ \mathscr{R}(x;\mathsf{A},d_\mathcal{M}^p)\le 2^p$. \item $ \mathscr{R}(x;\mathsf{A}\mathsf{B},d_\mathcal{M}^p)^{\frac{1}{p}}\le \mathscr{R}(x;\mathsf{A},d_\mathcal{M}^p)^{\frac{1}{p}}+\mathscr{R}(x;\mathsf{B},d_\mathcal{M}^p)^{\frac{1}{p}}$. \item $\mathscr{R} \big(x;\mathsf{A}^\tt,d_\mathcal{M}^p\big)\le \tt^p\mathscr{R} \big(x;\mathsf{A},d_\mathcal{M}^p\big)$ for every $\tt\in \mathbb N$. \end{enumerate} \end{lemma} \begin{proof} The first assertion is an immediate consequence of the definition of nonlinear Rayleigh quotients. The second assertion is a special case of the first assertion, since by definition $\mathscr{R}(x;\mathsf{Id}_n,d_\mathcal{M}^p)=0$. The third assertion is justified by noting that by the triangle inequality, for every $i,j,k\in \{1,\ldots,n\}$ we have \begin{equation}\label{eq:to bound rayleigh lower} d_\mathcal{M}(x_i,x_j)^p\le \big(d_\mathcal{M}(x_i,x_k)+d_\mathcal{M}(x_k,x_j)\big)^p \le 2^{p-1}d_\mathcal{M}(x_i,x_k)^p+2^{p-1}d_\mathcal{M}(x_k,x_j)^p. \end{equation} where the last step of~\eqref{eq:to bound rayleigh lower} uses the convexity of the function $(t\in [0,\infty))\mapsto t^p$. By multiplying~\eqref{eq:to bound rayleigh lower} by $\pi_i\pi_k a_{ij}$, summing over $i,j,k\in \{1,\ldots,n\}$ and using the fact that $\mathsf{A}$ is reversible with respect to $\pi$, we get $$ \sum_{i=1}^n\sum_{j=1}^n \pi_i a_{ij} d_\mathcal{M}(x_i,x_j)^p\le 2^p\sum_{i=1}^n\sum_{j=1}^n\pi_i\pi_j d_\mathcal{M}(x_i,x_j)^p. $$ Recalling the notation~\eqref{eq:ray def}, this is precisely the third assertion of Lemma~\ref{lem:Mtype2}. It remains to justify the fourth assertion of Lemma~\ref{lem:Mtype2}, because its fifth assertion follows from iterating its fourth assertion $\tt-1$ times (with $\mathsf{B}$ a power of $\mathsf{A}$). To this end, writing $\mathsf{A}=(a_{ij})$ and $\mathsf{B}=(b_{ij})$, we have \begin{align} \nonumber \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i(\mathsf{A}\mathsf{B})_{ij}d_\mathcal{M}(x_i,x_j)^p\bigg)^{\frac{1}{p}} &\le \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\bigg(\sum_{k=1}^n a_{ik}b_{kj}\bigg)\big(d_\mathcal{M}(x_i,x_k)+d_\mathcal{M}(x_k,x_j)\big)^p\bigg)^{\frac{1}{p}}\\ \nonumber &\le \bigg(\sum_{i=1}^n\sum_{j=1}^n \sum_{k=1}^n \pi_i a_{ik}b_{kj}d_\mathcal{M}(x_i,x_k)^p\bigg)^{\frac{1}{p}}+\bigg(\sum_{i=1}^n\sum_{j=1}^n \sum_{k=1}^n \pi_i a_{ik}b_{kj}d_\mathcal{M}(x_k,x_j)^p\bigg)^{\frac{1}{p}} \\&= \bigg(\sum_{i=1}^n \sum_{k=1}^n\pi_i a_{ik}d_\mathcal{M}(x_i,x_k)^p\bigg)^{\frac{1}{p}}+\bigg(\sum_{j=1}^n \sum_{k=1}^n \pi_kb_{kj}d_\mathcal{M}(x_k,x_j)^p\bigg)^{\frac{1}{p}},\label{remove A and B resp} \end{align} where the first step of~\eqref{remove A and B resp} uses the triangle inequality in $\mathcal{M}$, the second step of~\eqref{remove A and B resp} uses the triangle inequality in $L_p(\mu)$ with $\mu$ being the measure on $\{1,\ldots,n\}^3$ given by $\mu(i,j,k)=\pi_ia_{ik}b_{kj}$ for all $i,j,k\in \{1,\ldots,n\}$, and the final step of~\eqref{remove A and B resp} uses the fact that $\mathsf{A}$ and $\mathsf{B}$ are both row-stochastic and reversible with respect to $\pi$. \end{proof} The identity in the following claim is a consequence of a very simple and standard Hilbertian computation that we record here for ease of later references. \begin{claim}\label{claim:hilbertial rayleigh} For every Hilbert space $(H,\|\cdot\|_H)$ and every $x\in L_2(\pi;H)\setminus H$ we have $\mathscr{R}(x;\mathsf{A}^2,\|\cdot\|_H^2)\le 1$. Moreover, if $\sum_{i=1}^n \pi_i x_i=0$, then \begin{equation*}\label{eq:square identity} \frac{\left\|(\mathsf{A}\otimes \mathsf{Id}_H)x\right\|_{L_2(\pi;H)}}{\|x\|_{L_2(\pi;H)}}=\sqrt{1-\mathscr{R}(x;\mathsf{A}^2,\|\cdot\|_H^2)}. \end{equation*} \end{claim} \begin{proof} Let $\langle\cdot,\cdot\rangle:H\times H\to \mathbb R$ be the scalar product that induces the Hilbertian norm $\|\cdot\|_H$. Then, the scalar product that induces the norm $\|\cdot\|_{L_2(\pi;H)}$ is given by $\langle y,z\rangle_{L_2(\pi;H)} =\sum_{i=1}^n \pi_i\langle y_i,z_i\rangle$. By expanding the squares while using the fact that $\mathsf{A}$ is row-stochastic, reversible relative to $\pi$, and $\sum_{i=1}^n \pi_i x_i=0$, we get that \begin{multline*} \sum_{i=1}^n\sum_{j=1}^n \pi_i(\mathsf{A}^2)_{ij}\|x_i-x_j\|_H^2=2\|x\|_{L_2(\pi;H)}^2-2\sum_{i=1}^n \pi_i\bigg\langle x_i,\sum_{j=1}^n (\mathsf{A}^2)_{ij}x_j\bigg\rangle\\ =2\|x\|_{L_2(\pi;H)}^2-2 \big\langle x,(\mathsf{A}^2\otimes \mathsf{Id}_H)x\big\rangle_{L_2(\pi;H)}=2\|x\|_{L_2(\pi;H)}^2-2\big\|(\mathsf{A}\otimes \mathsf{Id}_H)x\big\|_{L_2(\pi;H)}^2, \end{multline*} and $$ \sum_{i=1}^n\sum_{j=1}^n\pi_i\pi_j\|x_i-x_j\|_H^2=2\sum_{i=1}^n \pi_i\|x_i\|_H^2-2\bigg\|\sum_{i=1}^n \pi_i x_i\bigg\|_H^2=2\|x\|_{L_2(\pi;H)}^2. $$ Therefore, recalling the definition~\eqref{eq:ray def}, we have \begin{equation*} \mathscr{R}(x;\mathsf{A}^2,\|\cdot\|_H^2)=1-\frac{\big\|(\mathsf{A}\otimes \mathsf{Id}_H)x\big\|_{L_2(\pi;H)}^2}{\|x\|_{L_2(\pi;H)}^2}\le 1. \qedhere \end{equation*} \end{proof} \begin{lemma}[point-wise Rayleigh quotient estimate for Hilbert isomorphs]\label{lem:power Rayleigh}Let $(X,\|\cdot\|_X)$ be a normed space and fix $\mathsf{d}\in [1,\infty)$. Suppose that $\|\cdot\|_H:X\to [0,\infty)$ is a Hilbertian norm on $X$ that satisfies \begin{equation}\label{eq:d assumption} \forall\, y\in X,\qquad \|y\|_H\le \|y\|_X\le \mathsf{d}\|y\|_H. \end{equation} For every $x\in L_2(\pi;X)\setminus X$ define a quantity $\tt(x,\mathsf{A})=\tt(x;\mathsf{A},\|\cdot\|_H,\mathsf{d})$ to be the minimum $\tt\in \mathbb N$ such that \begin{equation}\label{eq:Euclidean ray condition} \mathscr{R}\bigg(x;\Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\!2\tt},\|\cdot\|_H^2\bigg)\ge 1-\frac{1}{4\mathsf{d}^2}, \end{equation} with the convention that $\tt(x;\mathsf{A})=\infty$ if no such $\tt$ exists. Then, \begin{equation}\label{eq:t2} \frac{1}{\mathscr{R}(x;\mathsf{A},\|\cdot\|_X^2)}\lesssim\tt(x;\mathsf{A})^2. \end{equation} \end{lemma} \begin{proof} We may assume without loss of generality that $\sum_{i=1}^n\pi_i x_i=0$ and $\tt(x;\mathsf{A})<\infty$. Define a matrix \begin{equation}\label{eq:def mB} \mathsf{B}_x\stackrel{\mathrm{def}}{=} \Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\!\tt(x;\mathsf{A})}\in \mathsf{M}_n(\mathbb R). \end{equation} Then $\mathsf{B}_x$ is also a row-stochastic matrix which is reversible with respect to $\pi$, and, by the definition of $\tt(x;\mathsf{A})$, $$ \mathscr{R}(x;\mathsf{B}^2_x,\|\cdot\|_H^2)\ge 1-\frac{1}{4\mathsf{d}^2}. $$ By Claim~\ref{claim:hilbertial rayleigh}, since $\sum_{i=1}^n\pi_ix_i=0$, this implies that \begin{equation}\label{eq:hilbertian norm quotient} \frac{\left\|(\mathsf{B}_x\otimes \mathsf{Id}_H)x\right\|_{L_2(\pi;H)}}{\|x\|_{L_2(\pi;H)}}\le \sqrt{1-\left(1-\frac{1}{4\mathsf{d}^2}\right)}=\frac{1}{2\mathsf{d}}. \end{equation} At the same time, due to~\eqref{eq:d assumption} we have \begin{equation}\label{eq:comparison of norm quotients} \frac{\left\|(\mathsf{B}_x\otimes \mathsf{Id}_X)x\right\|_{L_2(\pi;X)}}{\|x\|_{L_2(\pi;X)}}\le \mathsf{d}\frac{\left\|(\mathsf{B}_x\otimes \mathsf{Id}_H)x\right\|_{L_2(\pi;H)}}{\|x\|_{L_2(\pi;H)}}. \end{equation} By combining~\eqref{eq:hilbertian norm quotient} and~\eqref{eq:comparison of norm quotients} we see that $\left\|(\mathsf{B}_x\otimes \mathsf{Id}_H)x\right\|_{L_2(\pi;X)}\le\frac12 \|x\|_{L_2(\pi;X)}$. Consequently, \begin{equation}\label{eq:lower I-B} \left\|x-\left(\mathsf{B}_x\otimes \mathsf{Id}_X\right)x\right\|_{L_2(\pi;X)}\ge \|x\|_{L_2(\pi;X)}-\left\|\left(\mathsf{B}_x\otimes \mathsf{Id}_X\right)x\right\|_{L_2(\pi;X)}\ge \frac12 \|x\|_{L_2(\pi;X)}. \end{equation} Observe that \begin{equation}\label{eq:variance} \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j\|x_i-x_j\|_X^2\bigg)^{\frac12}\le \bigg(\sum_{i=1}^n\sum_{j=1}^n \pi_i\pi_j \big(\|x_i\|_X+\|x_j\|_X\big)^2\bigg)^{\frac12}\le 2\|x\|_{L_2(\pi;X)}, \end{equation} where in the first step of~\eqref{eq:variance} we used the triangle inequality in $X$ and the second step of~\eqref{eq:variance} is an application of the triangle inequality in $L_2(\pi\otimes \pi)$. Also, since $\mathsf{B}_x$ is row-stochastic, \begin{equation}\label{eq:convexity B} \left\|x-\left(\mathsf{B}_x\otimes \mathsf{Id}_X\right)x\right\|_{L_2(\pi;X)}=\bigg(\sum_{i=1}^n \pi_i \bigg\|\sum_{j=1}^n (\mathsf{B}_x)_{ij}(x_i-x_j)\bigg\|_X^2\bigg)^{\frac12}\le \bigg(\sum_{i=1}^n \sum_{j=1}^n \pi_i(\mathsf{B}_x)_{ij}\|x_i-x_j\|_X^2\bigg)^{\frac12}, \end{equation} where in the final step of~\eqref{eq:convexity B} we used the convexity of the function $\|\cdot\|_X^2:X\to \mathbb R$. Recalling the definition~\eqref{eq:ray def}, by substituting~\eqref{eq:variance} and~\eqref{eq:convexity B} into~\eqref{eq:lower I-B} we see that \begin{equation}\label{eq:power is expander} \mathscr{R}\bigg(x;\Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\tt(x;\mathsf{A})},\|\cdot\|_X^2\bigg)\stackrel{\eqref{eq:def mB}}{=}\mathscr{R}\big(x;\mathsf{B}_x,\|\cdot\|_X^2\big)\ge \frac{1}{16}. \end{equation} We now conclude the proof of the desired estimate~\eqref{eq:t2} as follows. \begin{equation}\label{eq:use mtype} 1\stackrel{\eqref{eq:power is expander}}{\lesssim}\mathscr{R}\bigg(x;\Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\tt(x;\mathsf{A})},\|\cdot\|_X^2\bigg)\le \tt(x;\mathsf{A})^2\mathscr{R}\bigg(x;\frac12\mathsf{Id}_n+\frac12\mathsf{A},\|\cdot\|_X^2\bigg)=\frac12\tt(x;\mathsf{A})^2\mathscr{R}\big(x;\mathsf{A},\|\cdot\|_X^2\big), \end{equation} where the second step of~\eqref{eq:use mtype} uses the fifth assertion of Lemma~\ref{lem:Mtype2}, and the final step uses its second assertion. \end{proof} The quantity $\tt(x;\mathsf{A})$ of Lemma~\ref{lem:power Rayleigh} can be bounded as follows in terms of the spectral gap of $\AA$. \begin{lemma} Continuing with the notation of Lemma~\ref{lem:power Rayleigh}, the following estimate holds true. \begin{equation}\label{eq:t upper} \tt(x;\mathsf{A})\le \left\lceil \frac{\log(2\mathsf{d})}{\log\left(\frac{2}{1+\lambda_2(\mathsf{A})}\right)}\right\rceil\lesssim \frac{\log(2\mathsf{d})}{1-\lambda_2(\mathsf{A})}. \end{equation} \end{lemma} \begin{proof} Since $\AA$ is row-stochastic, $\lambda_n(\AA)\ge -1$. Therefore $\frac12\mathsf{Id}_n+\frac12\AA$ is a positive semidefinite self-adjoint operator on $L_2(\pi)$ that preserves the hyperplane $L_2^0(\pi)=\{u\in \mathbb R^n;\ \sum_{i=1}^n\pi_i u_i=0\}$. The largest eigenvalue of $\frac12\mathsf{Id}_n+\frac12\AA$ on $L_2^0(\pi)$ is $\frac12 +\frac12 \lambda_2(\AA)$, and therefore $\|(\frac12\mathsf{Id}_n+\frac12\AA)^\tt u\|_{L_2(\pi)}\le (\frac12 +\frac12 \lambda_2(\AA))^\tt\|u\|_{L_2(\pi)}$ for $u\in L_2^0(\pi)$ and $\tt\in \mathbb N$. If $x\in L_2(\pi;X)\setminus X$ satisfies $\sum_{i=1}^n\pi_ix_i=0$, then we may apply the above observation to the coordinates of $x$ with respect to some orthonormal basis of $H$, each of which is an element of $L_2^0(\pi)$, and deduce that \begin{equation}\label{eq:in L_2H} \Big(\frac12 +\frac12 \lambda_2(\AA)\Big)^\tt\ge \frac{\left\|\left(\left(\frac12\mathsf{Id}_n+\frac12\AA\right)^\tt \otimes \mathsf{Id}_H\right) x\right\|_{L_2(\pi;H)}}{\|x\|_{L_2(\pi;H)}}=\sqrt{1-\mathscr{R}\bigg(x;\Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\!2\tt},\|\cdot\|_H^2\bigg)}, \end{equation} where in the second step of~\eqref{eq:in L_2H} we applied Lemma~\ref{claim:hilbertial rayleigh} with $\AA$ replaced by $(\frac12\mathsf{Id}_n+\frac12\AA)^\tt$. Hence $$ \mathscr{R}\bigg(x;\Big(\frac12\mathsf{Id}_n+\frac12\mathsf{A}\Big)^{\!2\tt},\|\cdot\|_H^2\bigg)\ge 1- \Big(\frac12 +\frac12 \lambda_2(\AA)\Big)^{\!2\tt}. $$ Consequently, if $\tt\ge (\log(2\mathsf{d}))/\log(2/(\lambda_2(\AA)+1))$, then $ \mathscr{R}(x;(\frac12\mathsf{Id}_n+\frac12\mathsf{A})^{\!2\tt},\|\cdot\|_H^2)\ge 1-\frac{1}{4\mathsf{d}^2}$. By the definition of $\tt(x;\AA)$, this implies the first inequality in~\eqref{eq:t upper}. The second inequality in~\eqref{eq:t upper} follows by elementary calculus. \end{proof} \begin{proof}[Proof of~\eqref{eq:M-d version}] By a classical linearization argument~\cite{Enf70} (see~\cite[Chapter~7]{BL} for a modern treatment), for every $\mathsf{d}>\mathsf{c}_2(X)$ there is a Hilbertian norm $\|\cdot\|_H$ on $X$ that satisfies~\eqref{eq:d assumption}. We therefore see that for every $\AA$ as above \begin{multline}\label{eq:deduce slightly weaker} \sqrt{\gamma(\AA,\|\cdot\|_X^2)}\stackrel{\eqref{eq:gamma ray relation}}{=}\sup_{x\in L_2(\pi;X)\setminus X} \frac{1}{\sqrt{\mathscr{R}(x;\mathsf{A},\|\cdot\|_X^2)}}\stackrel{\eqref{eq:t2}}{\le} \sup_{x\in L_2(\pi;X)\setminus X} \tt(x;\AA)\\\stackrel{\eqref{eq:t upper}}{\lesssim} \frac{\log(\mathsf{c}_2(X)+1)}{1-\lambda_2(\AA)}=\frac{\log(\mathsf{c}_2(X)+1)}{\sqrt{1-\lambda_2(\AA)}}\sqrt{\gamma(\AA,\|\cdot\|_H^2)}\le \frac{\log (\mathsf{c}_2(X)+1)}{M}\mathsf{c}_2(X)\sqrt{\gamma(\AA,\|\cdot\|_H^2)}, \end{multline} where, for the final step of~\eqref{eq:deduce slightly weaker} recall that in the context of~\eqref{eq:M-d version} we assume that $\lambda_2(\AA)\le 1-M^2/\mathsf{c}_2(X)^2$. \end{proof} \subsubsection{Structural implications of nonlinear Rayleigh quotient inequalities} Fix integers $n,k,r\ge 3$ (think of $n$ as much larger than $k$). Let $(X,\|\cdot\|_X)$ be a $k$-dimensional normed space. Suppose that $\mathsf{G}=(\{1,\ldots,n\},\mathsf{E}_\mathsf{G})$ is a connected $r$-regular graph. Although we phrased (and used) Corollary~\ref{cor:regular case} as an impossibility result that provides an obstruction (spectral gap) for faithfully realizing (on average) the metric space $(\{1,\ldots,n\},d_\mathsf{G})$ in $X$, a key insight of the recent work~\cite{ANNRW18} by Andoni, Nikolov, Razenshteyn, Waingarten and the author is that one could "flip" this point of view to deduce from Corollary~\ref{cor:regular case} useful information on those graphs that do happen to admit such a faithful geometric realization in $X$, namely they satisfy~\eqref{eq:two averages-regular}. Clearly there are plenty of graphs with this property, including those graphs that arise from discrete approximations of subsets of $X$ (as a "vanilla" example to keep in mind, fix a small parameter $\d>0$, consider a $\d$-net in the unit ball of $X$ as the vertices, and join two net points by an edge if their distance in $X$ is $O(\d)$). The conclusion of Corollary~\ref{cor:regular case} for any such graph is that it cannot have a large spectral gap, and by Cheeger's inequality~\cite{Che70,Tan84,AM85,SJ89} it follows that this graph can be partitioned into two pieces with a small (relative) "discrete boundary." On the other hand, if we are given a mapping $f:\{1,\ldots,n\}\to X$ that satisfies the first condition in~\eqref{eq:two averages-regular} but not the second condition in~\eqref{eq:two averages-regular}, then there must be a ball in $X$ of relatively small radius that contains a substantial fraction of the vectors $\{f(i)\}_{i=1}^n$. The partition of $\{1,\ldots,n\}$ that corresponds to this dense ball and its complement encodes useful geometric "clustering" information. We have thus observed a dichotomic behavior that allows one to partition geometrically-induced graphs using either a "spectral partition" or a "dense ball partition." In~\cite{ANNRW18}, the above idea is used iteratively to construct a hierarchical partition of $X$. Our overview suppresses important technical steps, which include both randomization and a re-weighting procedure of the graphs that arise at later stages of the construction (the start of the construction is indeed the above "net graph"); see~\cite{ANNRW18} for the full details. In particular, one needs to use general row-stochastic matrices due to the re-weighting procedure, i.e., one uses the full strength of Theorem~\ref{thm:pi version average embedding} rather than only the case of graphs as in Corollary~\ref{cor:regular case}. In summary, one can use the bound~\eqref{eq:M-d version} on nonlinear spectral gaps to provide a "cutting rule" that governs an iterative partitioning procedure in which each inductive step is either geometric (a ball and its complement) or a less explicit existential step that follows from spectral information which is deduced from a contrapositive assumption of (rough, average) embeddability. This structural information is used in~\cite{ANNRW18} to design a new data structure for approximate nearest neighbor search in arbitrary norms (see the article of Andoni, Indyk and Razenshteyn in the present volume for an extensive account of approximate nearest neighbor search). Although this yields important (and arguably unexpected) progress on an algorithmic question of central importance, the non-explicitness and potential high complexity of the spectral partitioning step raises issues of efficiency that are not yet fully resolved. Specifically, the most general data structure that is designed in~\cite{ANNRW18} is efficient only in the so-called "cell probe model," but not in the full polynomial-time sense; we refer to~\cite{ANNRW18} for an explanation of these complexity-theoretic issues and their significance, because they are beyond the scope of the present article. While the above issue of efficiency does not occur in our initial investigation within pure mathematics, it is very important from the algorithmic perspective. This is what initially led to the desire to obtain a nonlinear Rayleigh quotient inequality rather than to merely bound the nonlinear spectral gap, though (in hindsight) such inequalities are interesting from the mathematical perspective as well. We did not formally define what we mean by a "nonlinear Rayleigh quotient inequality" because there is some flexibility here, but the basic desire is, given $x_1,\ldots,x_n\in X$, to bound their Rayleigh quotient in $X$ by a Rayleigh quotient of points in a {\em Euclidean space}. The inequality~\eqref{eq:t2} of Lemma~\ref{lem:power Rayleigh} is of the above form, because the parameter $\tt(x;\AA)$ defined by~\eqref{eq:Euclidean ray condition} involves only examining a certain Rayleigh quotient in a Hilbert space. It should be noted, however, that to date we have not succeeded to use the specific nonlinear Rayleigh quotient inequality of Lemma~\ref{lem:power Rayleigh} for algorithmic purposes (though with more work this may be possible), despite the fact that it was found with this motivation in mind. Other nonlinear Rayleigh quotient inequalities were obtained in~\cite{ANNRW18,ANNRW18-uniform} and used to address issues of algorithmic efficiency. Very roughly, the drawback of~\eqref{eq:t2} is that the matrix $\AA$ is changed in the Hilbertian Rayleigh quotient of~\eqref{eq:Euclidean ray condition} (the main problem is the potentially high power $2\tt$). A more directly algorithmically-useful nonlinear Rayleigh quotient inequality would be to change the point $x\in L_2(\pi;X)$ but not change the matrix $\AA$. Namely, suppose that we could control $\mathscr{R}(x;\mathsf{A},\|\cdot\|_X^2)$ from below by a function of $\mathscr{R}(\phi_\AA(x);\mathsf{A},\|\cdot\|_H^2)$, for some mapping $\phi_\AA:L_2(\pi;X)\to L_2(\pi;H)$. Nonlinear Rayleigh quotient inequalities of this type are proved in~\cite{ANNRW18,ANNRW18-uniform}, though the associated mappings $\phi_\AA$ turn out to be highly nonlinear and quite complicated.\footnote{Specifically, in~\cite{ANNRW18} such a mapping $\phi_\AA$ is constructed for Schatten-von Neumann trace classes using the Brouwer fixed-point theorem and estimates from~\cite{Ric15}. In~\cite{ANNRW18-uniform}, $\phi_\AA$ is constructed for general normed spaces using, in addition to Brouwer's theorem, convex programming and (algorithmic variants of) complex interpolation. These lead to data structures that are efficient in all respects other than the "preprocessing stage," which at present remains potentially time-consuming due to the complexity of $\phi_\AA$.} The upshot of the latter type of nonlinear Rayleigh quotient inequality is that if (due to existence of a faithful embedding into $X$) we know that $\mathscr{R}(x;\mathsf{A},\|\cdot\|_X^2)$ is small, then it follows that also $\mathscr{R}(\phi_\AA(x);\mathsf{A},\|\cdot\|_H^2)$ is small. The {\em proof of} Cheeger's inequality (via examination of level sets of the second eigenvector) would now provide a sparse "spectral partition" of $\{1,\ldots,n\}$ that has the following auxiliary structure: The partition is determined by thresholding one of the coordinates of $H$ (in some fixed orthonormal basis), namely the part to which each $i\in \{1,\ldots,n\}$ belongs depends only on whether the coordinate in question of the transformed vector $\phi_\AA(x)_i\in H$ is above or below a certain value. If in addition $(\AA,x)\mapsto \phi_\AA(x)$ has favorable computational properties (see~\cite{ANNRW18} for a formulation; roughly, what is important here is that after a "preprocessing step" one can decide quickly to which piece of the partition each $i\in \{1,\ldots,n\}$ belongs), then this would lead to fast "query time." The above description of the algorithmic role of nonlinear Rayleigh quotient inequalities is impressionistic, but it conveys the core ideas while not delving into (substantial) details. Such inequalities are interesting in their own right, partially because they necessitate making mathematical arguments constructive, thus leading to new proofs, as we did for~\eqref{eq:M-d version}, and also leading to intrinsically meaningful studies, such as obtaining~\cite{ANNRW18-uniform} algorithmic versions of existential statements that arise from the use of the maximum principle in complex interpolation. \bigskip \noindent{\bf Acknowledgements.} I am grateful to Daniel Kane for a conversation that led to a simplification of Proposition~\ref{prop:sigma nk}. I also thank Alexandros Eskenazis, Manor Mendel, Ilya Razenshteyn and Gideon Schechtman for their helpful suggestions. \bibliographystyle{abbrv}

e9ea9035ec2a4c962f5fbd2b5959bf772486b870

\section{Introduction} Our main object in this article is a valued hyperfield. A hyperfield is a field-like algebraic structure whose addition is multivalued, and a valued hyperfield is a hyperfield equipped with a valuation. A typical example is the quotient of a valued field by a multiplicative subgroup of the form $1+\mathfrak{m}^n$ for the maximal ideal $\mathfrak{m}$ of the valuation ring, which is called the $n$-th valued hyperfield of a valued field. In this article, we consider a lifting problem of homomorphisms of the $n$-th valued hyperfields to homomorphisms of valued fields(See Corollary \ref{cor:general_tamely_ramifieid} and Theorem \ref{thm:mainhomlifting}). Also we study relationships between certain categories of valued hyperfields, truncated discrete valuation rings, and discrete valued fields of mixed characteristic(See Theorem \ref{thm:equiv_DVR_hvf} and Theorem \ref{thm:lifting_vhf_valuedfield}). At last, we prove a relative completeness theorem of AKE-style in terms of valued hyperfields(See Theorem \ref{thm:AKE_hyperfield}).\\ M. Krasner in \cite{Kr} introduced a notion of valued hyperfield and used it to do a theory of limits of local fields. In \cite{De}, P. Deligne did the theory of limits of local fields in a different way by defining a notion of triple, which consists of truncated discrete valuation rings and some additional data. Typical examples of a valued hyperfield and a truncated discrete valuation ring are the $n$-th valued hyperfield and the $n$-th residue ring of a valued field respectively, where the $n$-th residue ring is a quotient of a valuation ring by the $n$-th power of the maximal ideal. J. Tolliver in \cite{To} showed that discrete valued hyperfields and triples are essentially same, stated by P. Deligne in \cite{De} without a proof. In \cite{LL}, W. Lee and the author showed that given complete discrete valued fields of mixed characteristic with perfect residue fields, any homomorphism between the $n$-th residue rings of the valued fields is lifted to a homomorphism between the valued fields for large enough $n$. From this, we ask the following question. \begin{question}\label{question:main_question_1} Let $K_1$ and $K_2$ be discrete valued fields of mixed characteristic $(0,p)$ having perfect residue fields. Suppose the $n$-th valued hyperfields of $K_1$ and $K_2$ are isomorphic for every $n\ge 1$. Then are $K_1$ and $K_2$ are isomorphic? Moreover, is there $N>0$ such that if the $N$-th valued hyperfiels of $K_1$ and $K_2$ are isomorphic, are $K_1$ and $K_2$ isomorphic? \end{question} The $n$-th valued hyperfield $\mathcal H_n(K):=K/(1+\mathfrak{m}^n)$ of a valued field $K$ contains automatically information of the multiplicative group $K^{\times}/(1+\mathfrak{m}^n)$. By the multivalued addition of $\mathcal H_n(K)$, it also contains information of the $n$-residue ring of $K$ too(See Fact \ref{fact:vhf_to_triple}). We compute the multivalued addition of the $n$-th valued hyperfields rather explicitly, where the multivalued addition of given two elements in $\mathcal H_n(K)$ is given as a ball and we can compute its center and its radius of the ball explicitly(See Lemma \ref{lem:basic_on_hyperfield}). In \cite{BaKuh} and \cite{Kuh}, it was considered some structures, called {\em amc-structures}, consisted with information of the $n$-th residue rings and the multiplicative groups $K^{\times}/(1+\mathfrak{m}^n)$. For $n\ge 1$, the $n$-th amc structure of a valued field $K$ is a triple $K_n:=(R_K^{2n}, G_K^n, \Theta_n)$, where $R_K^n:=R/t^{2(n-1)}\mathfrak{m}$, $G_K^{n}:=K^{\times }/(1+t^{n-1}\mathfrak{m})$, and a map $\Theta_{n}:\{x\in R_K^{2n}:\ x^2\neq 0\}\rightarrow G_K^{n},a+t^{2n}\mathfrak{m}\mapsto a(1+t^n\mathfrak{m})$. Here, $R$ is the valuation ring of $K$, and $t=1$ if the residue field is of characteristic $0$ and $t=p$ if the residue field is of characteristic $p>0$. In \cite{BaKuh}, S. A. Basrab and F. V. Kuhlmann showed that for a valued field $K$, and for henselian valued fields $L$ and $F$ which are algebraic extension of $K$, if $L_n$ and $F_n$ are isomorphic over $K_n$ for each $n$, then $L$ and $F$ are isomorphic over $K$(See \cite[Corollary 1.4]{BaKuh}). Moreover, if $L$ and $F$ are finite extension of $K$, then it is enough to consider whether $L_n$ and $F_n$ are isomorphic over $K_n$ for large enough $n$. In \cite{Kuh}, F. V. Kuhlmann showed that if $L$ and $F$ are additionally tame extensions of $K$, then $L$ and $F$ are isomorphic over $K$ if $L_1$ and $F_1$ are isomorphic over $K_1$(See \cite[Lemma 3.1]{Kuh}). We also study relationships between discrete valued hyperfields, truncated discrete valuation rings, and complete discrete valued fields of mixed characteristic. Fix a prime number $p$ and a positive integer $e$. Let ${\mathcal C}_{p,e}$ be the category of complete discrete valuation rings of mixed characteristic $(0,p)$ having perfect residue fields and ramification index $e$. Let ${\mathcal R}_{p,e}^n$ be the category of truncated discrete valuation rings of length $n$ with perfect residue fields such that $p$ is in the $e$-th power of the maximal ideal but not in the $e+1$-th power of the maximal ideal. \begin{fact}\label{fact:motivation_question_2}\cite[Theorem 4.7, Example 3.7(2)]{LL} For large enough $n$, there is a lifting functor $\operatorname{L}:{\mathcal R}_{p,e}^n\rightarrow {\mathcal C}_{p,e}$ satisfying several natural conditions. But two categories ${\mathcal C}_{p,e}$ and ${\mathcal R}_{p,e}^n$ are not equivalent. \end{fact} \noindent We ask a question analogous to Fact \ref{fact:motivation_question_2} for valued hyperfields. \begin{problem}\label{prob:main_problem_1} Fix a prime number $p$ and a positive integer $e$. Let $n$ be a positive integer. Find a suitable category $\mathcal H_{p,e}^n$ of valued hyperfields whose has a lifting functor $\operatorname{L}^{\mathcal H}: \mathcal H_{p,e}^n\rightarrow {\mathcal C}_{p,e}$ satisfying some natural conditions, and which makes $\mathcal H_{p,e}^n$ and ${\mathcal C}_{p,e}$ equivalent. \end{problem} At last, we concern a question of relative completeness of AKE-style in terms of hyper valuedfields. S. A. Basarab in \cite{Ba} showed that the theory of finitely ramified henselian valued fields of mixed characteristic is determined by the theory of the $n$-th residue ring for each $n$ and the theory of value group. \begin{fact}\cite[Corollay 3.1]{Ba}\label{fact:Basarab_result} Let $K_1$ and $K_2$ be finitely ramified henselian valued fields of mixed characteristic. Let $R_{1,n}$ and $R_{2,n}$ be the $n$-th residue rings of $K_1$ and $K_2$ respectively. Let $\Gamma_1$ and $\Gamma_2$ be the value groups of $K_1$ and $K_2$. Then the following are equivalent: \begin{enumerate} \item $K_1\equiv K_2$. \item $R_{1,n}\equiv R_{2,n}$ for every $n\ge 1$ and $\Gamma_1\equiv \Gamma_2$. \end{enumerate} \end{fact} \begin{question}\label{question:main_question2} Let $K_1$ and $K_2$ be finitely ramified henselian valued fields mixed characteristic. Let $\mathcal H_n(K_1)$ and $\mathcal H_n(K_2)$ be the $n$-th valued hyperfields rings of $K_1$ and $K_2$ respectively. If $\mathcal H_n(K_1)\equiv \mathcal H_{n}(K_2)$ for every $n\ge 1$, $K_1\equiv K_2$? \end{question} \bigskip In Section \ref{preliminaries}, we recall basic notations and facts. In Section \ref{section:lifting}, we answer Question \ref{question:main_question_1} positively. We also compute such $N$ effectively. To lift a homomorphism of the $n$-th valued hyperfields, we consider homomorphisms of valued hyperfields which are over $p$. In Subsection \ref{subsection:tamely_ramifieid}, it is enough to check whether there is an isomorphism over $p$ between two first valued hyperfields in the tamely ramified case. More precisely, any homomorphism over $p$ between the first valued hyperfields is induced from a unique homomorphism between given tamely ramified complete discrete valued fields. In Subsection \ref{subsection:generally_ramified}, we compute such $N$ in Question \ref{question:main_question_1} depending only on ramification indices of given complete discrete valued fields. To compute $N$ effectively, we first show that a homomorphism over $p$ of the $n$-th valued gives a unique homomorphism between Witt rings of given complete discrete valued fields of mixed characteristic with perfect residue fields. Using the structure theorem of totally ramified extension and Krasner's lemma, we compute $N$ effectively, which is depending only on the ramification index. In general, we can not drop the condition of being over $p$. In Section \ref{section:functoriality}, we suggest a suitable category $\widehat{\H}_{p,e}^n$ of valued hyperfields whose morphisms are isometric homomorphism, and show that there is a lifting functor $\operatorname{L}^{{\mathcal H}}:\widehat{\H}_{p,e}^n\rightarrow {\mathcal C}_{p,e}$ satisfying proper conditions for large enough $n$. Actually, we show that this category $\widehat{\H}_{p,e}^n$ and the category ${\mathcal R}_{p,e}^n$ are equivalent for large enough $n$. This essentially comes from the result of J. Tolliver in \cite{To}. We also show that the a subcategory $\mathcal H_{p,e}^n$ of $\widehat{\H}_{p,e}^n$ whose morphisms are over $p$ is equivalent to ${\mathcal C}_{p,e}$ for every $n>e$ if $p$ does not divide $e$. In Section \ref{section:AKE_hyperfields}, using lifting result of homomorphisms of $n$-th valued hyperfields in Section \ref{section:lifting}, we give a positive answer of Question \ref{question:main_question2} for the case of perfect residue fields. In this case, it is enough to check whether $N$-th valued hyperfields are elementary equivalent for large enough $N$. Specially, if $K_1$ and $K_2$ are tamely ramified and the first valued hyperfields are elementary equivalent, then $K_1$ and $K_2$ are elementary equivalent. \section{Preliminaries}\label{preliminaries} We introduce basic notations and terminologies which will be used in this paper. We denote a valued field by a tuple $(K,R(K),\mathfrak{m}_K,\nu_K,k(K),\Gamma_K)$ consisting of the following data : $K$ is the underlying field, $R(K)$ is the valuation ring, $\mathfrak{m}_K$ is the maximal ideal of $R(K)$, $\nu_K$ is the valuation map, $k_K$ is the residue field, and $\Gamma_{K}$ is the value group considering as an additive group. If there is no confusion, we omit $K$. Hereafter, the full tuple $(K,R,\mathfrak{m},\nu,k,\Gamma)$ will be abbreviated in accordance with the situational need for the components. For $\gamma\in \Gamma$, we write $\Gamma_{* \gamma}:=\{x\in \Gamma|\ x* \gamma\}$ for $*\in\{\ge,>\}$ and $\mathfrak{m}^{\gamma}:=\{x\in K|\ \nu(x)\in \Gamma_{> \gamma}\}$. Note that for each $\gamma\in \Gamma_{\ge 0}$, $1+\mathfrak{m}^{\gamma}$ is a multiplicative subgroup of $K^{\times}$. For $\gamma\in \Gamma_{\ge 0}$, let $R_{\gamma}:=R/\mathfrak{m}^{\gamma}$. We recall the definitions of ramified valued fields. \begin{definition}\label{def:unramified_valuedfield} We say that a valued field $(K,\nu,\Gamma)$ is {\em (absolutely) unramified} if $\ch(k)=0$, or $\ch(k)=p$ and $\nu(p)$ is the minimal positive element in $\Gamma$ for $p>0$. We say $(K,\nu)$ is {\em (absolutely) ramified} if it is not absolutely unramified. \end{definition} \begin{definition}\label{def:ramfication_index} Let $(K,\nu,\Gamma)$ be a valued field whose residue field has prime characteristic $p$. \begin{enumerate} \item We say $(K,\nu,\Gamma)$ is {\em (absolutely) finitely ramified} if the set $\{\gamma\in \Gamma|\ 0< \gamma\le \nu(p)\}$ is finite. The cardinality of $\{\gamma\in \Gamma|\ 0< \gamma\le \nu(p)\}$ is called {\em the (absolute) ramification index} of $(K,\nu)$, denoted by $e(K,\nu)$ or $e(R)$. If $K$ or $\nu$ is clear from context, we write $e(K)$ or $e$ for $e(K,\nu)$. For $x\in R$, we write $e_{\nu}(x):=|\{\gamma\in \Gamma|\ 0<\gamma\le \nu(x)\}|$. If there is no confusion, we write $e(x)$ for $e_{\nu}(x)$ \item Let $(K,\nu,\Gamma)$ be finitely ramified. If $p$ does not divide $e_{\nu}(p)$, we say $(K,\nu)$ is \emph{(absolutely) tamely ramified}. Otherwise, we say $(K,\nu)$ is \emph{(absolutely) wildly ramified}. \end{enumerate} \end{definition} \noindent Note that a discrete valued field having a residue field of characteristic $p>0$ is finitely ramified. We say that a discrete valued field $(K,\nu,\Gamma)$ with the residue field having characteristic $p>0$ is {\em normalized} if $\Gamma$ is a subgroup of ${\mathbb R}$ and $\nu(p)=1$. From now on, we mean a homomorphism between valued fields is an isometric homomorphism, where a field homomorphism $f:K_1\rightarrow K_2$ is called {\em isometric} if for $a,b\in K_1$, $$\nu_{K_1}(a)<\nu_{K_1}(b)\Leftrightarrow \nu_{K_2}(f(a))<\nu_{K_2}(f(b)).$$ We recall some facts on the Witt ring and Teichm\"{u}ller representatives of complete discrete valued fields of mixed characteristic(c.f. \cite{S}). \begin{fact}\label{fact:witt_teichmuller} \begin{enumerate} \item Let $W(k_1)$ and $W(k_2)$ be Witt rings of perfect residue fields $k_1$ and $k_2$ of characteristic $p>0$ respectively. Suppose that there is a homomorphism $\phi : k_1 \longrightarrow k_2 $. Then there is a unique lifting homomorphism $g:R_1 \longrightarrow R_2 $ such that $g$ induces $\phi$. \item Let $R$ be a complete discrete valuation ring of characteristic $0$ with perfect residue field $k$ of characteristic $p$ and corresponding valuation $\nu$. Then $\operatorname{W}(k)$ can be embedded as a subring of $R$ and $R$ is a free $\operatorname{W}(k)$-module of rank $\nu(p)$. Moreover, $R=\operatorname{W}(k) [\pi]$ where $\pi$ is a uniformizer of $R$. \item Let $A$ be a ring that is Hausdorff and complete for a topology defined by a decreasing sequence $\mathfrak{a}_1 \supset \mathfrak{a}_2 \supset ...$ of ideals such that $\mathfrak{a}_n \cdot \mathfrak{a}_m \subset \mathfrak{a}_{n+m}$. Assume that the residue ring $A_1=A/\mathfrak{a}_1 $ is a perfect field of characteristic $p$. Then: \begin{enumerate} \item There exists one and only one system of representatives $h: A_1 \longrightarrow A$ which commutes with $p$-th powers: $h(\lambda^p ) = h(\lambda)^p $. This system of representatives is called the set of Teichm\"{u}ller representatives. \item In order that $ a\in A $ belong to $S= h(A_1)$, it is necessary and sufficient that $a$ be a $p^n$-th power for all $n\geq 0$. \item This system of representatives is multiplicative which means $$ h( \lambda \mu ) = h(\lambda)h(\mu) $$ for all $\lambda, \mu \in A_1$. \item $S$ contains $0$ and $1$. \item $S\setminus\{0\}$ is a subgroup of the unit group of $A$. \end{enumerate} \end{enumerate} \end{fact} Next we introduce a notion of valued hyperfields. \begin{definition}\label{def:vhf}\cite[Definition 1.2 and 1.4]{Kr} \begin{enumerate} \item A {\em hyeprfield} is an algebraic structure $(H,+,\cdot,0,1)$ such that $(H^{\times},\cdot,1)$, where $H^{\times}:=H\setminus\{0\}$, is an abelian group and there is a multivalued operation $+:\ H\times H\rightarrow 2^{H}$ for the power set $2^H$ of $H$ satisfying the followings: \begin{enumerate} \item $0\cdot \alpha=0$ for all $\alpha\in H$. \item (Associative) $(\alpha+\beta)+\gamma=\alpha+(\beta+\gamma)$ for all $\alpha,\beta,\gamma\in H$. \item (Commutative) $\alpha+\beta=\beta+\alpha$ for all $\alpha,\beta\in H$. \item (Distributive) $(\alpha+\beta)\cdot \gamma\subset \alpha\cdot\gamma+\beta\cdot\gamma$ for all $\alpha,\beta,\gamma\in H$. \item (Identity) $\alpha+0=\{\alpha\}$ for all $\alpha\in H$. \item (Inverse) For any $\alpha\in H$, there is a unique $-\alpha\in H$ such that $0\in \alpha+(-\alpha)$. \item For all $\alpha,\beta,\gamma\in H$, $\alpha\in \beta+(-\gamma)$ if and only if $\beta\in \alpha+\gamma$. \end{enumerate} \item A {\em valued hyperfield} is a hyperfield $(H,+,\cdot,0,1)$ equipped with a map $\nu$ from $H$ to $\Gamma\cup\{\infty\}$ for an ordered abelian group $\Gamma$ such that \begin{enumerate} \item For $\alpha\in H$, $\nu(\alpha)=\infty$ if and only if $\alpha=0$; \item For all $\alpha,\beta\in H$, $\nu(\alpha\cdot\beta)=\nu(\alpha)+\nu(\beta)$; \item For all $\alpha,\beta\in H$, $\nu(\alpha+\beta)\ge \min\{\nu(\alpha),\nu(\beta)\}$; \item For all $\alpha,\beta\in H$, $\nu(\alpha+\beta)$ consists of single element unless $0\in \alpha+\beta$; and \item There is $\rho_H\in \Gamma$ such that either $\alpha+\beta$ is a closed ball of radius $\rho_H+\min\{\nu(\alpha),\nu(\beta)\}$ for all $\alpha,\beta\in H$, or $\alpha+\beta$ is a open ball of radius $\rho_H+\min\{\nu(\alpha),\nu(\beta)\}$ for all $\alpha,\beta\in H^{\times}$. \end{enumerate} \end{enumerate} For $B\subset H$ and $\alpha\in H$, define $\alpha+ B:=\bigcup_{\beta \in B}\alpha+ \beta$(*). The associativity of $+$ means that given $\alpha,\beta,\gamma\in H$, we have $(\alpha+\beta),(\beta+ \gamma)\subset H$ and $\alpha+(\beta+\gamma)=(\alpha+\beta)+\gamma$ in the sense of $(*)$. We say that $H$ is {\em discrete} if $\Gamma$ is a discrete subgroup of ${\mathbb R}$. \end{definition} \noindent For $\alpha_0,\ldots, \alpha_k \in H$, we write $\sum^{\mathcal H} \alpha_i$ for $(\alpha_0+\cdots+ \alpha_k)\subset H$. Since the multivalued operation $+$ is associative, the notion of $\sum^{\mathcal H}$ is well-defined. \begin{definition}\label{def:morphism_vhf} Let $(H_i,+_i,\cdot_i,0_i,1_i,\nu_i)$ be a valued hyperfield for $i=1,2$. A map $f$ from $H_1$ to $H_2$ is called a {\em homomorphsim} if the followings hold: \begin{enumerate} \item $f(0_1)=0_2$ and $f(1_2)=1_2$. \item $f(\alpha\cdot_1 \beta)=f(\alpha)\cdot_2 f(\beta)$ for all $\alpha,\beta \in H_1$. \item $f(\alpha+_1\beta)\subset f(\alpha)+_2 f(\beta)$ for all $\alpha,\beta\in H_1$. \item For all $\alpha,\beta \in H_1$, $$\nu_1(\alpha)\le \nu_1(\beta)\Leftrightarrow\nu_2(f(\alpha_1))\le \nu_2(f(\beta)).$$ \end{enumerate} Let $\mathrm{Hom}(H_1,H_2)$ be the set of homomorphism from $H_1$ to $H_2$. \end{definition} \noindent Note that the definition of homomorphisms in Definition \ref{def:morphism_vhf} is weaker than one in \cite[Definition 1.5]{To}. We call homomorphisms in \cite[Definition 1.5]{To} an {\em isometric homomorphism}. \begin{definition}\cite[Definition 1.5]{To}\label{def:isometry} Let $(H_i,+_i,\cdot_i,0_i,1_i,\nu_i)$ be a valued hyperfield for $i=1,2$. Suppose $\nu_1(H_1^{\times})=\nu_2(H_2^{\times})$. A map $f$ from $H_1$ to $H_2$ is called an {\em isometric homomorphism} if the followings hold: \begin{enumerate} \item $f(0_1)=0_2$ and $f(1_2)=1_2$. \item $f(\alpha\cdot_1 \beta)=f(\alpha)\cdot_2 f(\beta)$ for all $\alpha,\beta \in H_1$. \item $f^{-1}(\alpha+_2 \beta)= f^{-1}(\alpha)+_1f^{-1}(\beta)$ for all $\alpha,\beta \in f(H_1)$. \item For all $\alpha\in H_1$, $\nu_1(\alpha)=\nu_2(f(\alpha))$. \item $f^{-1}(1_2)$ is a ball. \end{enumerate} Let $\mathrm{Iso}(H_1,H_2)$ be the set of isometric homomorphisms from $H_1$ to $H_2$. Note that $(3)$ implies that for $\alpha,\beta\in H_1$ and $f\in \mathrm{Hom}(H_1,H_2)$, $f(\alpha+_1\beta)\subset f(\alpha)+_2 f(\beta)$ so that $\mathrm{Iso}(H_1,H_2)\subset \mathrm{Hom}(H_1,H_2)$. \end{definition} \begin{definition}\label{def:vhf_valuedhyperfield} Let $\gamma\in\Gamma_{\ge 0}$. The {$\gamma$-hyperfield} of $K$ is a hyperfield $(K/(1+\mathfrak{m}^{\gamma}),+_{\mathcal H},\cdot_{\mathcal H})$, denoted by $\mathcal H_{\gamma}(K)$, such that for each $a(1+\mathfrak{m}^{\gamma}),b(1+\mathfrak{m}^n)\in K/1+\mathfrak{m}^n$, \begin{enumerate} \item $a(1+\mathfrak{m}^{\gamma})\cdot_{\mathcal H} b(1+\mathfrak{m}^{\gamma}):=ab(1+\mathfrak{m}^{\gamma})$, \item $a(1+\mathfrak{m}^{\gamma})+_{\mathcal H} b(1+\mathfrak{m}^{\gamma}):=\{(x+y)(1+\mathfrak{m}^{\gamma})|\ x\in a(1+\mathfrak{m}^{\gamma}),y\in b(1+\mathfrak{m}^{\gamma})\}$, and \end{enumerate} Conventionally, we write $0$ for $0(1+\mathfrak{m}^{\gamma})$ and $1$ for $1(1+\mathfrak{m}^{\gamma})$. The valuation $\nu$ of $K$ induces a map $\nu_{\mathcal H}$ on $K/(1+\mathfrak{m}^{\gamma})$ sending $\nu_{\mathcal H}(a(1+\mathfrak{m}^{\gamma}))$ to $\nu(a)$. We call $(K/(1+\mathfrak{m}^{\gamma}),+_{\mathcal H},\cdot_{\mathcal H},\nu_{\mathcal H})$ the {\em valued $\gamma$-hyperfield}. For $\gamma\le \lambda\in \Gamma_{>0}$, we have that $(1+\mathfrak{m}^{\lambda})\subset (1+\mathfrak{m}^{\gamma})$ and it induces a projection $\mathcal H^{\lambda}_{\gamma}:\ \mathcal H_{\lambda}(K)\rightarrow\mathcal H_{\gamma}(K),a(1+\mathfrak{m}^{\lambda})\mapsto a(1+\mathfrak{m}^{\gamma})$. We write $\mathcal H_{\gamma}:K\rightarrow \mathcal H_{\gamma}(K),x\mapsto [x]_{\gamma}$. \end{definition} \noindent For $A\subset K$, let $\mathcal H_{\gamma}(A):=\{a(1+\mathfrak{m}^{\gamma}):\ a\in A\}$ and for $a\in K$, we write $[a]_n$ for $a(1+\mathfrak{m}^{\gamma})\in \mathcal H_{\gamma}(K)$. If $\gamma$ is obvious, we write $[a]$ for $[a]_{\gamma}$. Note that the valuation $\nu_{\mathcal H} : \mathcal H_{\gamma}^{\times}(K)\rightarrow \Gamma$ is a group epimorphism with the kernel $\mathcal H_{\gamma}(R^{\times})$. \begin{remark}\label{rem:vhf_valuegroup} The quotient group $\mathcal H_{\gamma}^{\times}(K)/\mathcal H_{\gamma}(R^{\times})$ is an ordered group isomorphic to $\Gamma$. \end{remark} \begin{definition}\label{def:over_p} Let $K_1$ and $K_2$ be valued fields whose residue fields are of characteristic $p>0$. Let $\mathcal H_{\gamma}(K_1)$ and $\mathcal H_{\lambda}(K_2)$ be valued hyper fields of $K_1$ and $K_2$ respectively. We say a homomorphism $f:\mathcal H_{\gamma}(K_1)\rightarrow \mathcal H_{\lambda}(K_2)$ is {\em over $p$} if $f([p])=[p]$. \end{definition} \noindent Note that there is an isometric isomorphism which is not over $p$(See Example \ref{ex:Hom_and_HomoverZ}). In Section \ref{section:functoriality}, we see how two sets of isometric homomorphisms and of isometric homomorphisms over $p$ are different(See Theorem \ref{thm:equiv_DVR_hvf} and Theorem \ref{thm:tame_equiv_vhf_valuedfields}). \begin{notation}\label{notation:nth_residuering_vhf} Let $(K,\nu)$ be a finitely ramified valued field and let $\pi$ be a uniformizer of $K$. Let $n\ge m$ be positive integers. We write $\mathfrak{m}^n:=\mathfrak{m}^{\nu(\pi^{n-1})}$. We write $R_{n}(K):=R(K)/\mathfrak{m}^n$ and $\mathcal H_n(K):=K/(1+\mathfrak{m}^n)$. We call $R_n(K)$ {\em the $n$-th residue ring} of $K$ and $\mathcal H_n(K)$ {\em the $n$-th valued hyperfield} of $K$. We write $[x]_n:=x(1+\mathfrak{m}^n)$ for $x\in K$. We denote $\mathcal H^n_m$ and $\mathcal H_m$ for $\mathcal H^{\nu(\pi^{n-1})}_{\nu(\pi^{m-1})}$ and $\mathcal H_{\nu(\pi^{m-1})}$ respectively. \end{notation} We now recall some results on hyperfields and (Deligne's) triples in \cite{To}. \begin{definition}\cite[Definition 1.8]{To}\label{def:DVR} A {\em truncated discrete valuation ring}, in short a truncated DVR, is a principal Artinian local ring. Let $R$ be a truncated DVR of length $l$. For $x\in R$, we define $\nu_R(x)=\sup\{ i\in {\mathbb N}|\ x\in \mathfrak{m}_R^i \}$, where $\mathfrak{m}_R$ is the maximal ideal of $R$. Then $\nu_R(R)=\{ 0,1,\ldots,l-1 \}\cup \{ \infty \}$, and $\nu_R(x)=\infty$ if and only if $x=0$. \end{definition} \begin{definition}\cite[Definition 1.9]{To}\label{def:triples} A {\em (Deligne's) triple} $(R,M,\epsilon)$ consists of a truncated DVR $R$, a free $R$-module $M$ of rank $1$, and a $R$-module homomorphism $\epsilon:\ M\rightarrow R$ whose image is the maximal ideal of $R$. \end{definition} \begin{definition}\cite[Definition 1.10]{To}\label{def:morphism_triples} Let $T_1=(R_1,M_1,\epsilon_1)$ and $T_2=(R_2,M_2,\epsilon_2)$ be triples. A morphism of triples $(r,f,\eta):T_1\rightarrow T_2$ consists of a homomorphism $f:R_1\rightarrow R_2$, an integer $r$, called the ramification index, and an $R_1$-module homomorphism $\eta :M_1\rightarrow M_2^{\otimes r}$ such that \begin{itemize} \item $f\circ \epsilon_1=\epsilon_2\circ \eta$; and \item $\eta$ induces a $R_2$-module isomorphism from $M_1\otimes_{R_1} R_2$ to $M_2^{\otimes r}$. \end{itemize} The composition of morphisms $(r_1,f_2,\eta_1)$ and $(r_2,f_2,\eta_2)$ of triples is given by $$(r_1,f_2,\eta_1)\circ (r_2,f_2,\eta_2)=(r_1r_2,f_1\circ f_2, \eta_1^{\otimes r_2}\circ \eta_2).$$ \end{definition} \begin{remark}\cite[Remark 4.5]{To} Let $T=(R,M,\epsilon)$ be a triple. Let $M$ be a free $R$-module of rank $1$ and $\Pi$ be a generator. For each $k\in {\mathbb Z}$, the tensor power $M^{\otimes k}$ is a well-defined $R$-module of rank $1$. More precisely, $$M^{\otimes k}=\begin{cases} R(\Pi^{\otimes k})& \mbox{if }k>0\\ R& \mbox{if }k=0\\ \mathrm{Hom}_R(M^{\otimes (-k)}, R)= R(\Pi^{\otimes k})& \mbox{if }k<0 \end{cases},$$ where for $k<0$, $\Pi^{\otimes k}\in \mathrm{Hom}_R(M^{\otimes (-k)}, R)$ is a unique homomorphism sending $\Pi^{\otimes (-k)}$ to $1$. Define a map $\nu_T:\ \bigcup_{k\in {\mathbb Z}}\limits M^{\otimes k}\rightarrow {\mathbb Z}\cup \{\infty\}$, called a {\em valuation map} of $T$, as follows: For $x=r\Pi^{\otimes k}\in M^{\otimes k}$, $\nu_T(x)=\nu_{R}(r)+k$. \end{remark} \begin{notation} Let $(H,\nu)$ be a discrete valued hyperfield. Denote $\theta_H:=\min\{\nu(x)|\ \nu(x)>0,\ x\in H\}$. \end{notation} \begin{rem/def}\cite[Remark 4.1]{To}\label{rem/def:length_vhf} Let $H$ be a discrete valued hyperfield. There is a positive integer $l$ such that $\rho_H=l\theta_H$. Such $l$ is called the {\em length} of $H$, denoted by $l(H)$. \end{rem/def} \begin{example} For a discrete valued field $K$, we have that $l(\mathcal H_n(K))=n$. \end{example} \begin{definition}\label{def:distance_equiv} Let $(H,\nu)$ be a discrete valued hyperfield. For $\eta\in {\mathbb R}$, define an equivalence relation $\equiv_{\eta}$ on $H$ as follows: For $\alpha,\beta\in H$, $\alpha\equiv_{\eta} \beta$ if and only if $\nu(\alpha-\beta)\ge \eta$. For $\alpha\in H$, we write $[\alpha]_{\eta}$ for the $\equiv_{\eta}$-class of $\alpha$. If $\eta$ is obvious, we omit it. Denote $\mathcal{O}_H:=\{\alpha\in H|\ \nu(\alpha)>0\}$ and $\mathfrak{m}_H:=\{\alpha\in H|\ \nu(\alpha)\ge \theta_H\}$. \end{definition} \begin{fact}\cite[Sections 4 and 5]{To}\label{fact:vhf_to_triple} Let $H$ be a discrete valued hyperfield. Let $R=\mathcal{O}_H/\equiv_{\rho_H}$, $M=\mathfrak{m}_H/\equiv_{\rho_H+\theta_H}$, and $\epsilon:\ M\rightarrow R,\ [\alpha]_{\rho_H+\theta_H}\mapsto [\alpha]_{\rho_H}$ for $\alpha\in H$. Then, \begin{enumerate} \item $\operatorname{Tr}(H)=(R,M,\epsilon)$ is a triple, and \item $R$ is of length $l(H)$ and the maximal ideal $\mathfrak{m}_R$ is $M/\equiv_{\rho_H}$. \end{enumerate} We write $R_l(H)$ for $R$. If $H$ or $l$ is obvious, we write $R_l$ or $R$ for $R_l(H)$. For discrete valued hyperfields $H'$ and $H"$, and for $f\in \mathrm{Iso}(H,H')$ and $g\in \mathrm{Iso}(H',H'')$, we have that \begin{enumerate} \item $\operatorname{Tr}(f)$ and $\operatorname{Tr}(g)$ are homomorphisms between $\operatorname{Tr}(H)$, $\operatorname{Tr}(H')$, and $\operatorname{Tr}(H'')$ respectively, and \item $\operatorname{Tr}(g\circ f)=\operatorname{Tr}(g)\circ \operatorname{Tr}(f)$. \end{enumerate} \end{fact} \begin{rem/def}\label{rem/def:residuefield_vhf} Let $H$ be a discrete valued hyperfield and $\operatorname{Tr}(H)=(R,M,\epsilon)$. The set $\mathcal{O}_H/\equiv_{\theta_H}$ forms a field and it is isomorphic to $R/\mathfrak{m}_R$. The field $\mathcal{O}_H/\equiv_{\theta_H}$ is called the {\em residue field} of $H$, denoted by $k(H)$. Moreover, for each $g\in \mathrm{Iso}(H_1,H_2)$, it induces a homomorphism $k(g):k(H_1)\rightarrow k(H_2)$. \end{rem/def} \begin{proof} To show $\mathcal{O}_H/\equiv_{\theta_H}$ forms a field, mimic the proof of \cite[Lemma 4.2]{To}. And consider a map $f:\mathcal{O}_H/\equiv_{\theta_H}\rightarrow R$ sending $[\alpha]_{\theta_H}$ to $[\alpha]_{\rho_H}+\mathfrak{m}_R$. Since $\theta_H\le \rho_H$, $f$ is well-defined and it induces an isomorphism. Let $H_1$ and $H_2$ be discrete valued hyperfield and let $g:H_1\rightarrow H_2$ be an isometric homomorphism. Then it induces $\operatorname{Tr}(g):\operatorname{Tr}(H_1)\rightarrow \operatorname{Tr}(H_2)$, which induces a homomorphism $k(g)$ from $k(H_1)$ to $k(H_2)$. \end{proof} \begin{remark}\cite[Section 6]{To}\label{rem:triple_to_vhf} Let $T=(R,M,\epsilon)$ be a triple with the valuation map $\nu_T$. Define $\operatorname{U}(T):=\{0\}\cup\bigcup_{i\in {\mathbb Z}}\limits \{x\in M^{\otimes i}|\ \nu_T(x)=i\}$. Then $\operatorname{U}(T)$ is a discrete valued hyperfield. Moreover, if $T=\operatorname{Tr}(H)$ for a discrete valued hyperfield $H$, then $\operatorname{U}(T)\cong H$ by an isometric isomorphism after rescaling $\nu_T(\Pi)=\theta_H$ for a generator $\Pi$ of $M$. Moreover the assignments $\operatorname{Tr}$ and $\operatorname{U}$ are funtorial. For discrete valued hyperfields $H_1$ and $H_2$, each $f\in\mathrm{Iso}(H_1,H_2)$ induces a morphism $\operatorname{Tr}(f)\in\mathrm{Hom}(\operatorname{Tr}(H_1),\operatorname{Tr}(H_2))$. For triples $T_1$ and $T_2$, each $g\in \mathrm{Hom}(T_1,T_2)$ induces a morphism $\operatorname{U}(g)\in \mathrm{Iso}(\operatorname{U}(T_1),\operatorname{U}(T_2))$. \end{remark} \begin{example}\cite[Example 4.9]{To}\label{ex:vhf_of_valuedfield} Let $K$ be a discrete valued field. Let $\mathcal H_n(K)$ be the $n$-th valued hyperfield of $K$ for a positive integer $n$. Then $\operatorname{Tr}(\mathcal H_n(K))\cong (R(K)/\mathfrak{m}_K^n,\mathfrak{m}_K/\mathfrak{m}_K^{n+1},\epsilon)$, where $\mathfrak{m}_K$ is the maximal ideal of the valuation ring $R(K)$ and the map $\epsilon$ is induced by the inclusion $\mathfrak{m}_K\subset R(K)$. \end{example} We introduce some algebraic and model theoretic structural theorems in terms of the $n$-th valued rings for finitely ramified valued fields. \begin{rem/def}\label{rem/def:krasnernumber} Let $R$ be a complete discrete valuation ring of mixed characteristic. Let $\pi$ be a uniformizer of $R$ and $\nu$ corresponding valuation of $R$. Let $L$ and $K$ be the fraction fields of $R$ and $\operatorname{W}(k)$ respectively. We denote the maximal number $$ \max\left\{\nu\big(\pi- \sigma(\pi) \big): \sigma\in \mathrm{Hom}_{K}(L,L^{alg}), \sigma (\pi) \neq \pi \right\} $$ by $M(R)_{\pi}$ or $M(L)_{\pi}$. This $M(R)_{\pi}$ does not depend on the choice of $\pi$ and we write $M(R)=M(L)$. Let $(F,\nu_F)$ be a complete discrete valuation ring of mixed characteristic having the same ramification index with $L$. Suppose there is an homomorphism from $L$ to $F$. Then $M(F)=M(L)$ \end{rem/def} \begin{remark}\cite[Theorem 3.9]{LL} Let $(K_1,\nu_1)$ and $(K_2,\nu_2)$ be complete discrete valued fields of mixed characteristic $(0,p)$. Then we have that $$M(K_1)e_{\nu_1}(p)e_{\nu_2}(p)\le e_{\nu_2}(p)(1+e_{\nu_1}^2(p)).$$ \end{remark} \noindent To determine whether two complete discrete valued fields of mixed characteristic with perfect residue fields are isomorphic, it is enough to check whether the $n$-th residue rings are isomorphic for large enough $n$. \begin{fact}\cite{LL}\label{fact:lift_n-th_residue} Let $K_1$ and $K_2$ be complete discrete valued fields of mixed characteristic with perfect residue fields. Let $R_{1,n}$ and $R_{2,n}$ be the $n$-th residue rings of $K_1$ and $K_2$ respectively. For $n> e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$, if $R_{1,n}$ and $R_{2,n}$ are isomorphic, then $K_1$ and $K_2$ are isomorphic. More generally, for positive integers $n_1$ and $n_2$ with $n_2>e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$, any homomorphism from $R_{1,n_1}$ to $R_{2,n_2}$ is lifted to a homomorphism from $K_1$ to $K_2$. \end{fact} For model theory of valued fields, we take the following languages for valued fields and their related structures. Let ${\mathcal L}_{K}=\{+,-,\cdot;0,1;|\}$ be a ring language with a binary relation $|$ for valued fields, where we interpret the binary relation $|$ as $a|b$ if $\nu(a)\le \nu(b)$ for $a,b\in K$. Let ${\mathcal L}_{k}=\{+',-',\cdot';0',1'\}$ be the ring language for residue fields, and ${\mathcal L}_{\Gamma}=\{+^*;0^*;<\}$ be the ordered group language for value groups. For each $n\le 1$, let ${\mathcal L}_{R_n}=\{+_n,-_n,\cdot_n;0_n,1_n\}$ be the ring language for the $n$-th residue ring. For $n=1$, we identify ${\mathcal L}_{R_1}={\mathcal L}_{k}$. We use a language ${\mathcal L}_{vhf}:=\{0,1,\cdot,+,|\}$ for valued hyperfields, where $0,1$ are constant symbols, $\cdot$ is a binary function symbol, $+$ is a ternary predicate, and $|$ is a binary relation. For a valued hyperfield $(H,\cdot,+,\nu)$, $1^H$ is interpreted as the identity of the multiplication, $0^H$ as the identity of the addition, $\cdot^H$ as the multiplication function on $H$. For $\alpha,\beta,\gamma\in H$, $(\alpha,\beta,\gamma)\in +^H$ if and only if $\gamma\in \alpha+\beta$, and $(\alpha,\beta)\in |^H$ if and only if $\nu(\alpha)\le \nu(\beta)$. For the convention, $(\alpha,0^H)\in |^H$ for all $\alpha$ in $H$. J. Ax and S. Kochen in \cite{AK}, and Y. Ershov in \cite{E} independently proved that the first order theories of unramifieid valued fields of characteristic $0$ are determined by the first order theories of their residue fields and valued groups. \begin{fact}\label{fact:AKE}\cite{AK,E}{(Ax-Kochen-Ershov principle)} Let $(K_1,k_1,\Gamma_1)$ and $(K_2,k_2,\Gamma_2)$ be unramified henselian valued fields of characteristic zero. Then we have that \begin{center} $K_1\equiv K_2$ if and only if $k_1\equiv k_2$ and $\Gamma_1\equiv \Gamma_2$. \end{center} \end{fact} \noindent In \cite{Ba}, S. A. Basarab generalized Fact \ref{fact:AKE} to the case of finitely ramified valued fields and in \cite{LL}, W. Lee and the author improved the result of Basarab in the case of perfect residue fields. \begin{fact}\cite{Ba,LL}\label{fact:fined_AKE_Rn} Let $(K_1,\nu_1,k_1,\Gamma_1)$ and $(K_2,\nu_2,k_2,\Gamma_1)$ be henselian valued fields of mixed characteristic $(0,p)$ with finite ramification indices. Suppose $k_1$ and $k_2$ are perfect fields. Let $n> e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$. The following are equivalent: \begin{enumerate} \item $K_1\equiv K_2$; \item $\Gamma_1\equiv \Gamma_2$ and $R_n(K_1)\equiv R_n(K_2)$. \end{enumerate} \end{fact} \section{Lifting}\label{section:lifting} In this section, we aim to show that for large enough $n$, any homomorphism over $p$ between the $n$-th valued hyperfields of discrete complete valued fields can be lifted to a homomorphism of given valued fields. \begin{lemma}\label{lem:basic_on_hyperfield} Let $a,b\in K$ and $a_0,\ldots,a_k\in K$. Fix $\gamma\in \Gamma_{\ge 0}$. \begin{enumerate} \item $\bigcup [a]_{\gamma}=a(1+\mathfrak{m}^{\gamma})=\{x|\ \nu(x-a)\ge \gamma+\nu(a)\}$. \item $\bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}=\{x\in K|\ \nu(x-(a+b))> \gamma+\min\{\nu(a),\nu(b)\}\}$. \item $0\in \bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}$ if and only if $\bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}=\mathfrak{m}^{\gamma+\min\{\nu(a),\nu(b)\}}$. \item $(a_0+\ldots+a_k)\in \sum^{\mathcal H} [a_i]_{\gamma}$. \item Suppose $b\in \bigcup \sum^{\mathcal H} [a_i]_{\gamma}$ and $a_0,\ldots,a_k\in R$. Then $b=(a_0+\ldots+a_k)+d$ for some $d\in \mathfrak{m}^{\gamma}$. \end{enumerate} \end{lemma} \begin{proof} Let $a,b\in K$ and $a_0,\ldots,a_k\in K$. (1) For $x\in K$, we have that \begin{align*} x\in a(1+\mathfrak{m}^{\gamma})&\Leftrightarrow x=a+ad,\ d\in \mathfrak{m}^{\gamma}\\ &\Leftrightarrow \nu(x-a)=\nu(ad)=\nu(a)+\nu(d), d\in \mathfrak{m}^{\gamma}\\ &\Leftrightarrow \nu(x-a)> \nu(a)+\gamma. \end{align*}\\ (2) ($\subseteq$) Let $x\in \bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}$ so that thee are $c,d\in \mathfrak{m}^{\gamma}$ such that $x=a+b+ac+bd$. Then we have that $$\nu(x-(a+b))=\nu(ac+bd)\ge \min\{\nu(ac),\nu(bd)\}> \gamma+\min\{\nu(a),\nu(b)\}.$$ ($\supseteq$) Take $x\in K$ such that $\nu(x-(a+b))>\gamma+\min\{\nu(a),\nu(b)\}$. WLOG we may assume that $\nu(a)\ge \nu(b)$. Then there is $c\in \mathfrak{m}^{\gamma}$ such that $x=(a+b)+bc=a+b(1+c)$, and $[x]_{\gamma}\in[a]_{\gamma}+_{\mathcal H}[b]$. So, we conclude that $x\in \bigcup [a]_{\gamma}+_G [b]_{\gamma}$.\\ (3) By (2), $\bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}=(a+b)+\mathfrak{m}^{{\gamma}+\min\{\nu(a),\nu(b)\}}$. Thus we have that \begin{align*} 0\in \bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}&\Leftrightarrow \nu(a+b)> {\gamma}+\min\{\nu(a),\nu(b)\}\\ &\Leftrightarrow (a+b)\in \mathfrak{m}^{{\gamma}+\min\{\nu(a),\nu(b)\}}\\ &\Leftrightarrow \bigcup [a]_{\gamma}+_{\mathcal H} [b]_{\gamma}=\mathfrak{m}^{{\gamma}+\min\{\nu(a),\nu(b)\}}. \end{align*} (4) and (5) come from (2). \end{proof} \begin{proposition}\label{prop:over_p=over_Z} Let $K_1$ and $K_2$ be valued fields whose residue fields are of characteristic $p>0$. Let $\mathcal H_{\gamma}(K_1)$ and $\mathcal H_{\lambda}(K_2)$ be valued hyper fields of $K_1$ and $K_2$ respectively. Let $f\in \mathrm{Hom}(\mathcal H_{\gamma}(K_1),\mathcal H_{\lambda}(K_2))$ be over $p$. Then for all $n\in {\mathbb Z}$, $f([n])=[n]$. We denote $\mathrm{Hom}_{{\mathbb Z}}(\mathcal H_{\gamma}(K_1),\mathcal H_{\lambda}(K_2))$ for the set of all homomorphisms over $p$. \end{proposition} \begin{proof} Let $f\in \mathrm{Hom}(\mathcal H_{\gamma}(K_1),\mathcal H_{\lambda}(K_2))$ be over $p$. First, we have $f([1])=[1]$ because $[1]$ is the multiplicative identity. Since $[-1]$ is the additive inverse of $[1]$, $f([-1])=[-1]$. By Lemma \ref{lem:basic_on_hyperfield}(2), we have $f([k])=[k]$ for $1\le k\le p-1$ inductively. Now choose an integer $n$ arbitrary. Since $f([-1])=[-1]$ and $f([0])=[0]$, we may assume that $n> 0$. Suppose $n$ and $p$ are relatively prime. We write $n=a_0+a_1p+a_2+\ldots+a_mp^m$ with $0\le a_i\le p-1$ and $a_0\neq 0$ for some $m\ge 0$. Suppose $f([n])=[n']$ for some $n'\in K_2$. Then we have that \begin{align*} [n']&=f([n])=f([\sum_{i}\limits a_ip^i])\\ &\in \sum_i^{\mathcal H}f([a_ip^i])= \sum_i^{\mathcal H}f([a_i])f([p])^i=\sum_i^{\mathcal H}[a_i][p]^i=\sum_i^{\mathcal H}[a_ip^i]. \end{align*} By Lemma \ref{lem:basic_on_hyperfield}(5), we have that $n'=\sum_i a_ip^i + d=n+d$ for some $d\in \mathfrak{m}_2^{\lambda}$. Since $n\notin \mathfrak{m}_2$, we have that $n'/n=1+d/n\in 1+\mathfrak{m}_2^{\lambda}$ and $n'(1+\mathfrak{m}_2^{\lambda})=n(1+\mathfrak{m}_2^{\lambda})$. Now suppose $p$ divides $n$. Write $n=n_0p^l$ for some $l>0$ and for some $n_0$ coprime to $p$. Then we have that $f([n])=f([n_0p^l])=f([n_0])f([p^l])=[n_0][p]^l=[n_0p^l]=[n]$. \end{proof} \begin{remark}\label{rem:vf_rigid_hom} Let $(K_1,\nu_1,k_1)$ and $(K_2,\nu_2, k_2)$ be finitely ramified valued fields having the same ramification index $e$. Let $p=\ch(k_1)=\ch(k_2)>0$. Suppose $\nu_1$ and $\nu_2$ are normalized, that is, $\nu_1(p)=\nu_2(p)=1$. Then for any $n,m\ge 1$ and $f\in \mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$, we have that $\nu_1(\alpha)=\nu_2(f(\alpha))$ for every $\alpha\in \mathcal H_n(K_1)$. And any homomorphism from $K_1$ to $K_2$ induces an isometric homomorphism from $\mathcal H_n(K_1)$ to $\mathcal H_n(K_2)$. \end{remark} In the remaining part of this section, {\bf we assume that a complete discrete complete valued field of mixed characteristic $(0,p)$ has the normalized valuation so that $\nu(K^{\times})\subset {\mathbb R}$ and $\nu(p)=1$}. For a henelian valued field $(K,\nu)$ of characteristic $0$, there is a unique valuation on $K^{alg}$ extending $\nu$ and we use the same notion $\nu$ for this valuation on $K^{alg}$. \begin{remark}\label{rem:H1(S)_resideufield} Let $S$ be the set of Teichm\"{u}ller representatives of a complete discrete valued field $K$ of mixed characteristic having a perfect residue field. Then, $\mathcal H_1(S)$ is a field which is isomorphic to the residue field of $K$. \end{remark} \begin{proof} Let $H=(\mathcal H_1(K),+_1,\times_1)$ be the first valued hyperfield of $K$. For each $a,b\in S$, there is a unique $c\in S$ such that $[c]\in [a]+_1[b]$. So, $(\mathcal H_1(S),+,\times_1)$ forms a field where for $a,b\in S$, $[a]+[b]=[c]$ if $[c]\in [a]+_1[b]$. Consider a map sending $[a]$ to $a+\mathfrak{m}$, where $\mathfrak{m}$ is the maximal ideal of the valuation ring of $K$ and this induces an isomorphism from $\mathcal H_1(S)$ to the residue field of $K$. \end{proof} \begin{lemma}\label{lem:preserving_Teichmuller} Let $K_1$ and $K_2$ be complete discrete valued fields of mixed characteristic having perfect residue fields. Let $S_1$ and $S_2$ be the set of Teichm\"{u}ller representatives of $K_1$ and $K_2$ respectively, and let $f\in \mathrm{Hom}(\mathcal H_{n}(K_1),\mathcal H_{m}(K_2))$. \begin{enumerate} \item We have $f(\mathcal H_{n}(S_1))\subset \mathcal H_{m}(S_2)$ and $f\restriction_{ \mathcal H_n(S_1)}$ is injective. \item If $f$ is over $p$, then $f\restriction_{\mathcal H_n(\operatorname{W}(k_1))}$ is induced by a unique homomorphism from $\operatorname{W}(k_1)$ to $\operatorname{W}(k_2)$, where $\operatorname{W}(k_1)$ and $\operatorname{W}(k_2)$ are Witt subrings of $K_1$ and $K_2$ respectively. \end{enumerate} \end{lemma} \begin{proof} (1) The proof is similar to the proof of \cite[Lemma 3.3]{LL}. Let $R_1$ and $R_2$ be the valuation rings of $K_1$ and $K_2$ respectively. By Fact \ref{fact:witt_teichmuller}(3), $S_i$ is contained in $\operatorname{W}(k_i)^{\times}$ where $k_i$ is the residue field of $K_i$ for $i=1,2$. For each $\lambda \in S_1$, let $\eta_{s}$ be any representative of $f(\lambda^{1/p^s}(1+\mathfrak{m}_1^n))$ so that $(\eta_s)^{p^s}(1+\mathfrak{m}_2^m)=f(\lambda(1+\mathfrak{m}_1^n))$. Since $\eta_s$ is in $\operatorname{W}(k_2)^{\times}$, we have that $\eta_s(1+\mathfrak{m}_2^m)=\eta_s+\mathfrak{m}_2^m$ and $(\eta_s)^{p^s}(1+\mathfrak{m}_2^m)=(\eta_s)^{p^s}+\mathfrak{m}_2^m$ as a set. For any other representative $\theta_s$ of $f(\lambda^{1/p^s}(1+\mathfrak{m}_1^n))$, we have that $\eta_s+\mathfrak{m}_2^m=\theta_s+\mathfrak{m}_2^m$. If we write $\eta_s= \theta_s + \pi_2 ^{n_2}a$ for some $a$ in $R_2$, the following binomial expansion \begin{align*} \eta_s ^{p^s} &=(\theta_s + \pi_2 ^{n_2}a)^{p^s}\\ {}&=\theta_s ^{p^s} + p^s \theta_s ^{ p^s -1} \pi_2^{n}a+...+(\pi_2 ^{n}a)^{p^s} \end{align*} shows $\eta_s ^{p^s} -\theta_s ^{p^s} \in \mathfrak{m}_2 ^{s} $. Since $\eta_{s+1} ^p$ is a representative of $ f(\lambda ^{1/{p^s}}(1 +\mathfrak{m}_1^{n}))$, the calculation above shows that $(\eta_s ^{p^s})$ is a Cauchy sequence and $\lim_{s\rightarrow \infty} \eta_s ^{p^s}$ is well-defined in $R_2$. Since $ \eta_s ^{p^s}(1 +\mathfrak{m}_2^{m}) =f(\lambda(1+\mathfrak{m}_1^{n}))$ and $1+\mathfrak{m}_2^{m}$ is topologically closed in $R_2$, $$ f\left(\lambda(1+\mathfrak{m}_1^{n}) \right)= \left(\lim_{s\rightarrow \infty} \eta_s ^{p^s}\right)(1 +\mathfrak{m}_2^{m}). $$ Similarly, we have $$ f\left((\lambda)^{1/p}(1+\mathfrak{m}_1^{n}) \right)= \left(\lim_{s\rightarrow \infty} \eta_s ^{p^{s-1}}\right)(1 +\mathfrak{m}_2^{m}). $$ Since $$ \lim_{s\rightarrow \infty} \eta_s ^{p^s} = \left(\lim_{s\rightarrow \infty} \eta_s ^{p^{s-1}}\right)^p, $$ we obtain $$ \lim_{s\rightarrow \infty} \eta_s ^{p^s}\in S_2 $$ by \ref{fact:witt_teichmuller}(3). Therefore, we have that $f(\mathcal H_n(S_1))\subset \mathcal H_m(S_2)$. By Remark \ref{rem:H1(S)_resideufield}, the restriction map $f\restriction \mathcal H_1(S_1)$ induces an homomorphism between residue fields of $K_1$ and $K_2$. So, we have that $f\restriction \mathcal H_1(S_1)$ is trivial or injective. Since $1\in S_1$, we have that $f([1]_n)=[1]_m$ and $f\restriction \mathcal H_1(S_1)$ is not trivial. Therefore, $f\restriction \mathcal H_1(S_1)$ must be injective.\\ (2) Suppose $f$ is over $p$, that is, $f([p]_n)=[p]_m$. Note that each $a\in \operatorname{W}(k_i)$ is uniquely written as $\sum_{k\ge 0}\limits a_kp^k$ for $a_k\in S$ for $i=1,2$. By Fact \ref{fact:witt_teichmuller}(1), Remark \ref{rem:H1(S)_resideufield}, and $(1)$, we have a homomorphism $\bar f:\operatorname{W}(k_1)\rightarrow \operatorname{W}(k_2)$ such that $f([a]_n)=[\bar f(a)]_m$ for $a\in S_1$. Take $a \in \operatorname{W}(k_1)$ and write $a=p^l\sum_{k\ge 0}\limits a_k p^k$ for $l\ge 0$ and $a_k\in S_1$ with $a_0\neq 0$. We have that \begin{align*} f([a]_n)&=[p]_m^lf([\sum_{k\ge 0} a_k p^k])_m\\ &=[p]_m^lf([a_0+a_1p+a_2p^2+\cdots+a_sp^s])_m, \end{align*} where $s=\max\{n,m\}$. It is enough to show that $$f([\sum_{0\le k\le s}\limits a_kp^k]_n)=[\sum_{0\le k\le s}\limits \bar f(a_k)p^k]_m$$ for $a_0\neq 0$ and $s\ge \max\{n,m\}$. Take $a=\sum_{0\le k\le s}\limits a_kp^k\in \operatorname{W}(k_1)$ with $a_0\neq 0$ and $s\ge \max\{n,m\}$. Then, \begin{align*} f([a])&\in \sum_{0\le k\le s}^{\mathcal H}\limits f([a_k]_n)[p]_m^k\\ &=\sum_{0\le k\le s}^{\mathcal H}\limits [\bar f(a_k)]_m[p]_m^k\\ &=\sum_{0\le k\le s}^{\mathcal H}\limits [\bar f(a_k)p^k]_m. \end{align*} So, we have that \begin{align*} f([a]_n)&=[\sum_{0\le k\le s}\limits \bar f(a_k)p^k+d]_m\\ &=[\bar f(a)+d] \end{align*} for some $d\in \mathfrak{m}_2^m$ by Lemma \ref{lem:basic_on_hyperfield}(5). Since $(\bar f(a)+d)/\bar f(a)=1+d/\bar f(a)\in 1+\mathfrak{m}_2^m$, we have that $[\bar f(a)+d]_m=[\bar f(a)]_m$. Therefore, $f([a]_n)=[\bar f(a)]_m$ for each $a\in \operatorname{W}(k_1)$. \end{proof} \subsection{Tamely ramified case}\label{subsection:tamely_ramifieid} We first recall some embedding lemma in \cite{Kuh} for tame algebraic extensions of valued fields. We say that an algebraic extension $(L,\nu_L)$ of a henselian valued field $(K,\nu_K)$ is {\em tame} if the following conditions hold: For every finite subextension $(F,\nu_F)$ of $(L,\nu_L)$ over $(K,\nu)$, \begin{itemize} \item the residue field extension $k_L$ over $k_K$ is separable; \item if the characteristic of $k_K$ is $p>0$, then the ramification index $(\Gamma_F:\Gamma_K)$ is prime to $p$; and \item $[F:K]=[k_F:k_K](\Gamma_F:\Gamma_K)$. \end{itemize} If $K$ is a complete discrete tamely ramified valued field with a perfect residue field $k$, then $K$ is the tame extension of the fraction field of $\operatorname{W}(k)$. \begin{fact}\cite[Lemma 3.1]{Kuh}\label{fact:embedding_tame} Let $K$ be an arbitrary valued field. Let $L$ be an algebraic tame extension of some henselization of $K$ and $F$ be an arbitrary henselian extension of $K$. Any $K_{1}$-embedding from $L_{1}$ to $F_{1}$ is induced from a $K$-embedding from $L$ to $F$. Furthermore if $k_K=k_L$, then any (group) $G_K^1$-embedding from $G_L^1$ to $G_F^1$ is induced from a $K$-embedding from $L$ to $F$. \end{fact} \noindent By adapting the ideas of the proofs of \cite[Lemma 3.1]{Kuh}, we prove the following theorems. \begin{theorem}\label{thm:tame_vf_vhf} Let $K_1$ and $K_2$ be complete discrete valued fields of mixed characteristic $(0,p)$ with perfect residue fields. Suppose $K_1$ is tamely ramified. Any homomorphism over $p$ from $\mathcal H_n(K_1)$ to $\mathcal H_m(K_2)$ is induced from a unique homomorphism from $K_1$ to $K_2$. From this, we conclude that there is one-to-one correspondence between $\mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$ and $\mathrm{Hom}(K_1,K_2)$. \end{theorem} \begin{proof} Take $f\in \mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$. Let $F_i$ be the fraction field of $\operatorname{W}(k_i)$ where $k_i$ is the residue field of $K_i$ for $i=1,2$. By Lemma \ref{lem:preserving_Teichmuller}(2), $f\restriction_{\mathcal H_1(F_1)}$ is induced from a homomorphism $\bar f:F_1\rightarrow F_2$. Since $K_1$ is a totally tamely ramified extension of $F_1$, $K_1=F_1(\sqrt[e]{pa})$ for some $a$ in $\operatorname{W}(k_1)^{\times}$ where $e$ is the ramification index of $K_1$ (c.f. Chapter $2$ of \cite{L}). Let $\pi_1=\sqrt[e]{pa}$ and $f([\pi_1]_n)=[\pi_2']_m$ for some $\pi_2'\in K_2$ so that $[\pi_2']_m^e=[p\bar f(a)]_m$. Consider a polynomial $P(X)=X^e-(p\bar f(u))/\pi_2'^e\in K_2[X]$. Since $P(1)\in \mathfrak{m}_2^m$ and $P'(1)\notin \mathfrak{m}_2$, by Hensel's lemma, there is unique $b\in R_2^{\times}$ such that $b^e=(p\bar f(a))/\pi_2'^e$ and $(b-1)\in \mathfrak{m}_2^m$. Let $\pi_2:= b\pi_2'$ so that $\pi_2^e=p\bar f(a)$. Note that $[\pi_2]_m=[\pi_2']_m$ and $\pi_2$ is such a unique zero of the polynomial $X^e-p\bar f(a)$. So we have a homomorphism $\tilde{f}:K_1\rightarrow K_2$ extending $\bar f\cup\{(\pi_1,\pi_2)\}$ and it induces $f$. \end{proof} We generalize Theorem \ref{thm:tame_vf_vhf} to the case of infinitely tamely ramified valued fields. To do this, we first recall the Ax-Sen-Tate theorem. \begin{definition}\label{def:Ax_diameter_conjugate}\cite{A} Let $(K,\nu)$ be a henselian valued field and $(K^{alg},\nu)$ be the algebraic closure of $K$. For $a\in K^{alg}$, define $$\Delta_K(a):=\min \{\nu(\sigma(a)-a)|\ \sigma\in G_K \},$$ where $G_K$ is the Galois group of $K^{alg}$ over $K$. \end{definition} \begin{fact}\label{fact:Ax_zeros_approximation}\cite[Proposition 1, Proposition 2']{A} Let $(K,\nu)$ be a complete valued field and $(K^{alg},\nu)$ be the algebraic closure of $K$. Suppose $\Gamma_{K^{alg}}$ is archimedean. Let $K\subset F\subset K^{alg}$. \begin{enumerate} \item Suppose $K$ is of mixed characteristic $(0,p)$. Then for all $a\in K^{alg}$, there exists $b\in F$ such that $$\nu(a-b)\ge \Delta_F(a)-(p/(p-1)^2)\nu(p).$$ \item Suppose $K$ is of equal characteristic $p\ge 0$. Then for all $a\in K^{alg}$ and for all $\gamma\in (\Gamma_{K^{alg}})_{> 0}$, there exists $b$ in the perfect closure of $F$ such that $$\nu(a-b)\ge \Delta_F(a)-\gamma.$$ \end{enumerate} \end{fact} \begin{fact}[Ax-Sen-Tate Theorem]\label{fact:structure_complete_subfields}\cite[Proposition 3.8]{FO} Let $(K,\nu)$ be a complete valued field and $(K^{alg},\nu)$ be the algebraic closure of $K$ with $\Gamma_{K^{alg}}$ archimedean. Let $C$ be the completion of $K^{alg}$ which is algebraically closed and let $L$ be a perfect complete subfield of $C$ containing $K$. Then $L$ is the completion of $L\cap K^{alg}$. \end{fact} \begin{proof} See \cite[Proposition 3.8]{FO} with Fact \ref{fact:Ax_zeros_approximation}. \end{proof} \begin{corollary}\label{cor:general_tamely_ramifieid} Let $p$ be a prime number. Let $K$ be a tamely ramified valued field of mixed characteristic $(0,p)$ with a perfect residue field $k$, and $F$ be the fraction field of $\operatorname{W}(k)$. Suppose $K$ is a subfield of the completion of $F^{alg}$, and either \begin{itemize} \item $K$ is an algebraic extension of $F$, or \item complete. \end{itemize} For a complete valued field $L$ of mixed characteristic $(0,p)$ with a perfect residue field, any homomorphism over $p$ from $\mathcal H_\gamma(K)$ to $\mathcal H_\lambda(L)$ is induced from a unique homomorphism $K$ to $L$ for any $\gamma\in (\Gamma_{K})_{\ge 0}$ and $\lambda\in (\Gamma_{L})_{\ge 0}$. \end{corollary} \begin{proof} Let $L$ be a complete valued field of mixed characteristic $(0,p)$ with a perfect residue field. Let $F$ be the fraction field of $\operatorname{W}(k)$. Fix $\gamma\in (\Gamma_{K})_{\ge 0}$ and $\lambda\in (\Gamma_{L})_{\ge 0}$.\\ Suppose $K$ is an algebraic extension of $F$. Since $K$ is tamely ramified, $K=\bigcup K_i$ where $K_i$ is a tamely totally ramified finite extension of $F$. Take $f\in \mathrm{Hom}_{\BZ}(\mathcal H_{\gamma}(K),\mathcal H_{\lambda}(L))$. Set $f_i:=f\restriction_{\mathcal H_{\gamma}(K_i)}$ and we have $f:=\lim_{\longleftarrow K_i}\limits f_i=\bigcup f_i$. By Theorem \ref{thm:tame_vf_vhf}, $f_i$ is induced from a unique homomorphism $\sigma_i:K_i\rightarrow L$. Take $\sigma:=\lim_{\longleftarrow K_i}\limits \sigma_i=\bigcup \sigma_i:K\rightarrow L$ and it is a unique homomorphism inducing $f$.\\ Suppose $K$ is complete. By Fact \ref{fact:structure_complete_subfields}, $K$ is the completion of $K':=K\cap F^{alg}$. Let $f'=f\restriction_{\mathcal H_{\gamma}(K')}$. By above result, $f'$ is induced from a unique homomorphism $\tilde f':K'\rightarrow L$. Since $K$ is a completion of $K'$, $\tilde f'$ induces a unique homomorphism $\tilde{f}\rightarrow L$, which induces $f$. \end{proof} \noindent We can not drop the condition of being over $p$ in Theorem \ref{thm:tame_vf_vhf}. \begin{example}\label{ex:criteria_via_vhf_2} Consider $K_1={\mathbb Q}_3(\sqrt{3})$ and $K_2={\mathbb Q}_3(\sqrt{-3})$, which are not isomorphic by Kummer Theory. Note that their residue fields are isomorphic to ${\mathbb F}_3$ and so their Teichm\"{u}ller representatives are $\{-1,0,1\}$. Let $\pi_1=\sqrt{3}$ and $\pi_2=\sqrt{-3}$. Every elements of $K_1$ is of the form $\sum_{i\ge n}a_i \pi_1^n$ for some integer $n$ and $a_i\in\{-1,0,1\}$ with $a_n\neq 0$. So each elements in $\mathcal H_1(K_1)$ is of the form $\pi_1^n a_n(1+\mathfrak{m}_1)$ for some integer $n$ and $a_n\in\{-1,1\}$($\dagger$). The same formulas hold for $K_2$. We will show that $\mathrm{Hom}(\mathcal H_1(K_1),\mathcal H_1(K_2))=\{f_0,f_1\}$, where $f_i$ sends $[a]$ to $[a]$ for $a\in\{-1,1\}$, $[\pi_1]$ to $[(-1)^i\pi_2]$, and $[3]$ to $[-3]$ for $i=1,2$, so that $\mathrm{Hom}_{\BZ}(\mathcal H_1(K_1),\mathcal H_1(K_2))=\emptyset$. Note that $[1]\neq [-1]$. Let $f\in \mathrm{Hom}(\mathcal H_1(K_1),\mathcal H_1(K_2))$. Then $f([-1])=[-1]$ because $f([1])=[1]$ and $0\in f([1]+[-1])$. Since $\pi_1$ is an uniformizer, $f([\pi_1])=[\pi_2]$ or $=[-\pi_2]$. In both cases, $f([3])=f([\pi_1]^2)=[\pi_2]^2=[\pi_2^2]=[-3]$. By ($\dagger$), such $f$ induces an isomorphism between $\mathcal H_1(K_1)$ and $\mathcal H_1(K_2)$. \end{example} \noindent Without base fields, even the residue fields of $K_1$ and $K_2$ are primes fields so that the residue fields are equal, we can not lift a group homomorphism from $\mathcal H_1^{\times}(K_1)$ to $\mathcal H_1^{\times}(K_2)$ to a homomorphism from $K_1$ to $K_2$(c.f. Fact \ref{fact:embedding_tame}). \begin{example}\label{ex:nobase_nogroup} Let $K_1=K_2={\mathbb Q}_5$. The set of Teichm\"{u}ller representatives of ${\mathbb Q}_5$ is $\{0,i,i^2,i^3,i^4\}$ where $i=\sqrt{-1}$. Consider a group isomorphism $f:\mathcal H_1^{\times}({\mathbb Q}_5)\rightarrow \mathcal H_1^{\times}({\mathbb Q}_5)$ by mapping $[i]\mapsto [-i]$ and $[5]\mapsto[5]$. Then $f$ is never induced from an automorphism of ${\mathbb Q}_5$ since any automorphism of ${\mathbb Q}_5$ sends $i$ to $i$. \end{example} \subsection{Generally ramified case}\label{subsection:generally_ramified} We first introduce of a notion of lifting map of homomorphisms of the $n$-th valued hyperfields, which is an analogy to a lifting map of homomorphisms of the $n$-th residue rings in \cite[Definition 3.1]{LL}. \begin{definition}\label{def:lifting homomorphim} Let $K_1$ and $K_2$ be complete discrete valuation rings of characteristic $0$ with perfect residue fields $k_1$ and $k_2$ of characteristic $p$ respectively. Let $\pi_i$ be a uniformizer of $K_i$ and $\nu_i$ be a corresponding valuation of $K_i$ for $i=1,2$. For any homomorphism $\phi:\mathcal H_n(K_1)\rightarrow \mathcal H_m(K_2)$, we say that a homomorphism $g:K_1\rightarrow K_2$ is a \emph{$(n,m)$-lifting} of $\phi$ at $\pi_1$ if $g$ satisfies the following: \begin{itemize} \item There exists a representaive $b$ of $\phi([\pi_1])$ which satisfies $$ \nu_2\big(g(\pi_1)-b\big)> M(K_1). $$ \item $\phi_{red,1} \circ \mathcal H_1=\mathcal H_1 \circ g$ where $\phi_{red,1}:\mathcal H_1(K_1) \rightarrow \mathcal H_1(K_2) $ denotes the natural reduction map of $\phi$. \end{itemize} When such $g$ is unique, we denote $g$ by $\operatorname{L}_{\pi_1, n,m}^{\mathcal H}(\phi)$. When $\operatorname{L}_{\pi_1, n,m}^{\mathcal H}(\phi)$ exists for all $\phi:\mathcal H_n(K_1)\rightarrow \mathcal H_m(K_2)$, we write $\operatorname{L}_{\pi_1, n,m}^{\mathcal H}:\mathrm{Hom}_{{\mathbb Z}}(\mathcal H_n(K_1),\mathcal H_m(K_2))\rightarrow \mathrm{Hom}(K_1,K_2)$. When $n=m$, we briefly write $\operatorname{L}_{\pi_1, n,m}^{\mathcal H}=\operatorname{L}_{\pi_1, n}^{\mathcal H}$ and say that $\operatorname{L}_{\pi_1, n}^{\mathcal H}$ is an \emph{$n$-lifting at $\pi_1$}. \end{definition} \noindent The following result is analogous to Proposition 2.9(2) for $n$-th residue rings in \cite{LL}. \begin{proposition}\label{prop:lifting_independent_uniformizer} Let $K_1$ and $K_2$ be complete discrete valued fields of characteristic $0$ with perfect residue fields $k_1$ and $k_2$ of characteristic $p$ respectively. Let $\pi_i$ be a uniformizer of $K_i$ and $\nu_i$ be a corresponding valuation of $K_i$ for $i=1,2$. Let $R_i$ be the valuation ring of $K_i$ for $i=1,2$. The definition of liftings is independent of the choice of uniformizer of $K_1$. More precisely, saying that $g:K_1\rightarrow K_2$ is a $(n,m)$-lifting of $\phi:\mathcal H_n(K_1)\rightarrow \mathcal H_m(K_2)$ at $\pi_1$ is equivalent to the following: \begin{enumerate} \item For any $x$ in $R_1$, there exists a representative $b_x$ of $\phi(x(1+\mathfrak{m}_1^n))$ which satisfies $$ \nu_2\big(g(x)-b_x\big) > M(K_1). $$ \item $\phi_{red,1} \circ \mathcal H_1=\mathcal H_1 \circ g$ \end{enumerate} We write $\operatorname{L}_{\pi_1, n, m}^{\mathcal H}=\operatorname{L}_{n, m}^{\mathcal H}$ and say that $\operatorname{L}_{n, m}^{\mathcal H}$ is a \emph{$(n,m)$-lifting}. Moreover, there is at most one $(n,m)$-lifting for $m>M(K_1)e_2$, where $e_2$ is the ramification index of $K_2$. \end{proposition} \begin{proof} The proof is similar to the proof of \cite[Proposition 3.5 (2)]{LL}. Let $S_i$ be the Teichm\"{u}ller representatives of $K_i$ for $i=1,2$. Fix a uniformizer $\pi$ of $K_1$. Let $g:K_1\rightarrow K_2$ be a $(n,m)$-lifting of $\phi:\mathcal H_n(K_1)\rightarrow \mathcal H_m(K_2)$ at a uniformizer $\pi$. Let $b \in K_2$ be a representative of $\phi([\pi]))$ such that $\nu_2(g(x)-b)>M(K_1)$. Note that $\nu_2(b)>0$. Take $x\in R_1$. Then $x=\sum_{i\ge 0}\lambda_i \pi^i$. Take $l>0$ such that $\nu_1(\sum_{i>l}\lambda_i\pi^i)>M(K_1)$. Denote $x^{\le l}:=\sum_{i\le l}\lambda_i \pi^i$ and $x^{>l}\sum_{i>l}\lambda_i \pi^i$ so that $x=x^{\le l}+x^{>l}$. Then we have that \begin{align*} \phi([x])&=[x^{\le l}+x^{>l}])\\ &\in \big ( \phi([x^{\le l}]+_{\mathcal H}\phi([x^{>l}]) \big )\\ &\subset \big( \sum_{i\le l}^{\mathcal H} \phi([\lambda_i])\phi([\pi]^i)+\phi([x^{>l}]) \big)\\ &\subset \big( \sum_{i\le l}^{\mathcal H} [g(\lambda_i)][b]^i+\phi([x^{>l}]) \big). \end{align*} There is a representative $b_x$ of $\phi([x])$ of the form:$$b_x=\sum_{i\le l}g(\lambda_i)b^i +d$$ for some $\nu_2(d)>M(K_1)$. Compute \begin{align*} \nu_2(g(x)-b_x)&=\nu_2\big( \sum_i (g(\lambda_i)g(\pi)^i)- b_x \big)\\ &=\nu_2\big(\sum_{0<i\le l} g(\lambda_i)(g(\pi)^i-b^i)+(\sum_{i>l} g(\lambda_i)g(\pi)^i-d) \big)\\ &\ge \min\{\nu_2(\sum_{0<i\le l} g(\lambda)(g(\pi)^i-b^i)),\nu_2(\sum_{i>l} g(\lambda_i)g(\pi)^i-d)\}\\ &>M(K_1) \end{align*} because $\nu_2(g(\pi)^i-\beta^i)=\nu_2(g(\pi)-\beta)+\nu_2(g(\pi)^{i-1}+\cdots + \beta^{i-1})>M(K_1)$, and $\nu_2(\sum_{i>l} g(\lambda_i)g(\pi)^i),\nu_2(d)>M(K_1)$.\\ Now we show moreover part. Assume $m>M(K_1)e_2$ and there are two $(n,m)$-liftings $\operatorname{L},\operatorname{L}': \mathrm{Hom}_{{\mathbb Z}}(\mathcal H_n(K_1),\mathcal H_m(K_2))\rightarrow \mathrm{Hom}(K_1,K_2)$. Fix $f\in \mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$. By Lemma \ref{lem:preserving_Teichmuller} and $(2)$, we have that $\operatorname{L}(f)\restriction \operatorname{W}(k_1)=\operatorname{L}'(f)\restriction \operatorname{W}(k_1)$ $(\dagger)$. It remains to show that $\operatorname{L}(f)(\pi_1)=\operatorname{L}'(f)(\pi_1)$. Set $\pi:=\operatorname{L}(f)(\pi_1)$ and $\pi':=\operatorname{L}'(f)(\pi_1)$. By $(\dagger)$, $\pi$ and $\pi'$ are conjugates over the fraction field of $\operatorname{W}(k_2)$ $(\ddagger)$. By $(1)$, there are two representatives $b$ and $b'$ of $f([\pi_1])$ such that $$\nu_2(\pi-\beta),\nu_2(\pi'-\beta')>M(K_1).$$ Since $\nu_2(b-b')>m/e_2>M(K_1)$, we have that \begin{align*} \nu_2(\pi-\pi')&>\max\{\nu_2(\pi-b),\nu_2(b-b'),\nu_2(\pi'-b')\}\\ &>M(K_1). \end{align*} By $(\ddagger)$, we conclude that $\nu_2(\pi-\pi')=\infty$ and $\pi=\pi'$. \end{proof} \begin{remark} Let $K_1$ and $K_2$ be complete discrete valued field of mixed characteristic $(0,p)$ with perfect residue fields. Suppose $K_1$ is tamely ramified. By Theorem \ref{thm:tame_vf_vhf}, for every $n\ge 1$ there is a unique bijective $n$-lifting map $$\operatorname{L}_n^{\mathcal H}:\mathrm{Hom}_{{\mathbb Z}}(\mathcal H_n(K_1),\mathcal H_n(K_2))\rightarrow \mathrm{Hom}(K_1,K_2)$$ such that $f([a]_n)=[\operatorname{L}_n^{\mathcal H}(f)(a)]$ for $f\in \mathrm{Hom}_{{\mathbb Z}}(\mathcal H_n(K_1),\mathcal H_n(K_2))$ and $a\in K_1$. \end{remark} \begin{fact}[Krasner's lemma]\label{lem:Krasner} Let $(K,\nu)$ be henseilan valued field whose value group is contained in ${\mathbb R}$ and let $a,b\in K^{alg}$. Suppose $a$ is separable over $K(b)$. Suppose that for all embeddings $\sigma(\neq id)$ of $K(a)$ over $K$, we have $$\widetilde{\nu}(b-a)>\widetilde{\nu}\big(\sigma (a)-a\big). $$ Then $K(a)\subset K(b)$. \end{fact} \noindent We show that a lifting map of homomorphisms of $n$-th valued hyperfields for large enough $n$ \begin{theorem}\label{thm:mainhomlifting} Let $K_1$ and $K_2$ be complete discrete valued field of mixed characteristic $(0,p)$ with perfect residue fields. Let $e_1$ and $e_2$ be ramification indices of $K_1$ and $K_2$ respectively. Let $\mathcal H_n(K_1)$ and $\mathcal H_m(K_2)$ be valued hyper fields of $K_1$ and $K_2$ respectively. Suppose $m>M(K_1)e_1e_2$. There is a unique lifting map $\operatorname{L}_{n,m}^{\mathcal H}:\mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))\rightarrow \mathrm{Hom}(K_1,K_2)$. \end{theorem} \begin{proof} Fix a homomorphism $f$ in $\mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$. By Lemma \ref{lem:preserving_Teichmuller}, $f$ induces a map $f\restriction_{\mathcal H_n(S_1)}$ from $\mathcal H_n(S_1)$ to $\mathcal H_m(S_2)$. This map $f\restriction_{\mathcal H_n(S_1)}$ induces a homomorphism from $k_1$ to $k_2$ and by the functoriality of Witt ring, we have a homomorphism $\bar f:\operatorname{W}(k_1)\rightarrow \operatorname{W}(k_2)$. Note that $\bar f$ induces a homomorphism from $\operatorname{W}(k_1)[X]$ to $\operatorname{W}(k_2)[X]$ by acting on coefficients and we denote this homomorphism by $\bar f$ also. By Fact \ref{fact:witt_teichmuller}(2), there is a uniformizer $\pi_1$ of $K_1$ such that $R_1=\operatorname{W}(k_1)[\pi_1]$. Let $q(X)$ be the irreducible polynomial of $\pi_1$ over the fraction field of $\operatorname{W}(k_1)$, which is in $\operatorname{W}(k_1)[X]$. Write $q(X)=X^e+a_{e-1}X^{e-1}+\ldots+a_0$ where $e=\nu_1(p)$ is the ramification index of $K_1$. Let $\pi_2'\in R_2$ such that $[\pi_2']_m=f([\pi_1]_n)$. For $a\in \operatorname{W}(k_1)$, we can uniquely write $a=a_0+a_1p+a_2p^2+\ldots$ with $a_i\in S_1$. For $l\ge 0$, define $a^{\le l}:=a_0+a_1p+\ldots+a_lp^l$ and $a^{>l}:=a-a^{\le l}$. And define $q^{\le l}(X):=X^e+a_{e-1}^{\le l}X^{e-1}+\ldots+a_0^{\le l}$ and $q^{>l}(X)=q(X)-q^{\le l}(X)$, which are in $\operatorname{W}(k_1)[X]$. Then we have that \begin{align*} 0&=f([q(\pi_1)]_n)\\ &=f([q^{\le m}(\pi_1)+q^{>m}(\pi_1)]_n)\\ &\in f([q^{\le m}(\pi_1)]_n)+_{\mathcal H} f([q^{> m}(\pi_1)]_n)\\ &\subset [\pi_2'^e]_m+_{\mathcal H} f[a_{e-1}^{\le m}\pi_2'^{e-1}]_m+_{\mathcal H}\ldots+_{\mathcal H} f[a_0^{\le m}]_m +_{\mathcal H} f([q^{>m}(\pi_1)]_n)\\ &\subset [\pi_2'^e]_m+_{\mathcal H}\sum^{\mathcal H}_{0\le i<e}\limits \sum^{\mathcal H}_{0\le j\le m}\limits [\pi_2'^i\bar f(a_i^j)p^j]_m+_{\mathcal H} f([q^{>m}(\pi_1)]_n). \end{align*} Note that $\nu_2(q^{>m}(\pi_1))\ge m/e_2$. By Lemma \ref{lem:basic_on_hyperfield}(2), we have that $$0=\bar f(q^{\le m})(\pi_2')+d (*)$$ for some $d\in \mathfrak{m}_2^m$. Since $q=q^{\le m}+q^{>m}$, we have that $\bar f(q)(\pi_2')=\bar f(q^{\le m})(\pi_2')+\bar f(q^{> m})(\pi_2')$ and $\nu_2(\bar f(q^{> m})(\pi_2'))\ge m/e_2$. From $(*)$ we have that $$0=\bar f(q)(\pi_2')+d'$$ for some $d'\in \mathfrak{m}_2^m$, and $\nu_2(\bar f(q)(\pi_2'))\ge m/e_2>M(K_1)e_1$. By Krasner's lemma, there is $\pi_2\in K_2$ such that $\bar f(q)(\pi_2)=0$ with $\nu_2(\pi_2-\pi_2')>M(K_1)$. Define $\tilde f:K_1\rightarrow K_2,\pi_1\mapsto \pi_2$ extending $\bar f$, and set $\operatorname{L}_{n,m}^{\mathcal H}(f)=\tilde f$. \end{proof} \begin{corollary}\label{cor:criteria_isom} Let $(K_1,\nu_1,k_1,\Gamma_1)$ and $(K_2,\nu_2,k_2,\Gamma_1)$ be finitely ramified complete valued fields of mixed characteristic. Let $n> e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$. The following are equivalent: \begin{enumerate} \item $K_1\cong K_2$; \item $R_{n}(K_1)\cong R_{n}(K_2)$; and \item $\mathcal H_{n}(K_1)\cong_{\mathcal H_n(\{p\})} \mathcal H_{n}(K)$. \end{enumerate} \end{corollary} The following is an analogy of \cite[Proposition 4.4]{LL} for a lifting map of homomorphisms of $n$-th valued hyperfields, which gives a funtoriality of lifting map in Section \ref{section:functoriality}. \begin{proposition}\label{prop:funtoriality} Let $(K_1,\nu_1)$, $(K_2,\nu_2)$, and $(K_3,\nu_3)$ be complete discrete valued fields of mixed characteristic $(0,p)$ with perfect residue fields. Suppose $K_1$ and $K_2$ have the same ramification index. Suppose $m,k> \max\{ e_{\nu_2}(p)(1+e_{\nu_1}^2(p), e_{\nu_3}(p)(1+e_{\nu_2}^2(p))\} $. Let $f\in \mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$ and $g\in \mathrm{Hom}_{\BZ}(\mathcal H_m(K_2),\mathcal H_k(K_3))$. Then $\operatorname{L}_{n,k}^{\mathcal H}(g\circ f)=\operatorname{L}_{m,k}^{\mathcal H}(g)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)$. \end{proposition} \begin{proof} Take $f\in \mathrm{Hom}_{\BZ}(\mathcal H_n(K_1),\mathcal H_m(K_2))$ and $g\in \mathrm{Hom}_{\BZ}(\mathcal H_m(K_2),\mathcal H_k(K_3))$. First, we have that \begin{align*} (g\circ f)\circ \mathcal H_1 &= g\circ (f\circ \mathcal H_1)\\ &= g\circ (\mathcal H_1\circ \operatorname{L}_{n,m}^{\mathcal H}(f))\\ &= (g\circ \mathcal H_1)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)\\ &= (\mathcal H_1\circ \operatorname{L}_{m,k}^{\mathcal H}(g))\circ \operatorname{L}_{n,m}^{\mathcal H}(f)\\ &= \mathcal H_1\circ(\operatorname{L}_{m,k}^{\mathcal H}(g)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)). \end{align*} It remains to show that for a uniformizer $\pi_1$ of $K_1$, there is a representative $b$ of $g\circ f([\pi_1])$ such that $$\nu_3(\operatorname{L}_{m,k}^{\mathcal H}(g)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)(\pi_1)-b)>M(K_1).$$ Take a representative $b_1$ of $f([\pi_1])$ such that $\nu_2(\operatorname{L}_{n,m}^{\mathcal H}(f)(\pi_1)-b_1)>M(K_1)$, and we have that $\nu_3(\operatorname{L}_{m,k}^{\mathcal H}(g)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)(\pi_1)-\operatorname{L}_{m,k}^{\mathcal H}(g)(b_1))>M(K_1)$. Next consider a representative $b_2$ of $g([b_1])( =g\left (f([\pi_1])\right) )$ such that $\nu_3(\operatorname{L}_{m,k}^{\mathcal H}(g)(b_1)-b_2)>M(K_2)$. Since $K_1$ and $K_2$ have the same ramification indices, we have that $M(K_1)=M(K_2)$. Thus we conclude that $\nu_3(\operatorname{L}_{m,k}^{\mathcal H}(g)\circ \operatorname{L}_{n,m}^{\mathcal H}(f)(\pi_1)-b_2)>M(K_1)$, and set $b:=b_2$. \end{proof} \section{Valued hyperfields, truncated DVRs, and valued fields}\label{section:functoriality} In this section, we figure out relationships between valued hyperfields, truncated DVRs, and complete discrete valued fields of mixed characteristic. We first define a notion of finitely ramified valued hyperfields. \begin{definition}\label{def:char_ramificationindex_vhf} Let $(H,\nu)$ be a discrete valued hyperfield of length $l$. Let $R_l(H)=R$. Suppose the residue field $k(H)$ has the characteristic $p\ge 0$. \begin{enumerate} \item We say $H$ is of {\em equal characteristic $(p,p)$} if $p=0$ in $R$. Otherwise, we say $H$ is of {\em mixed characteristic $(0,p)$}. \item Suppose $H$ is {\em of mixed characteristic $(0,p)$}. We say $H$ is {\em finitely ramified} if $\nu_R(p)=e<\infty$, and $e$ is called the {\em ramification index} of $H$. \item Suppose $H$ is finitely ramified. We say $H$ is normalized if $\theta_H=1/e$ where $e$ is the ramification index of $H$. \end{enumerate} \end{definition} \noindent Note that any discrete valued hyperfield of mixed characteristic need not be finitely ramified. \begin{proposition}\label{prop:DVR_enough} Let $K$ be a discrete valued field. Let $K_1$, $K_2$, and $K_3$ be discrete valued fields of mixed characteristic $(0,p)$ having the same ramification index $e$. \begin{enumerate} \item $T_n(K)=(R_n(K),M_n(K),\epsilon_n)$ is a triple, where $M_n(K):=\mathfrak{m}_K/\mathfrak{m}_K^{n+1}$ and the map $\epsilon_n$ is induced by the inclusion $\mathfrak{m}_K\subset R(K)$. \item For every triple $T=(R_n(K),M,\epsilon)$, we have $T\cong T_n(K)$. \item Let $n>e$. Each morphism $f\in \mathrm{Hom}(R_n(K_1), R_n(K_2))$ induces a morphism $T_n(f)\in \mathrm{Hom}(T_n(K_1),T_n(K_2))$. For $f\in \mathrm{Hom}(R_n(K_1),R_n(K_2))$ and $g\in \mathrm{Hom}(R_n(K_2),R_n(K_3))$, we have $T_{n}(g\circ f)=T_{n}(g)\circ T_{n}(f)$. \end{enumerate} \end{proposition} \begin{proof} (1) It is clear.\\ (2) Let $R:=R_n(K)$. Given a triple $T=(R,M,\epsilon)$, let $\Pi$ be a generator of $M$. Since $\epsilon(M)=\mathfrak{m}_R$, $\epsilon(\Pi)=\pi+\mathfrak{m}_K$ for a uniformizer $\pi$ of $K$. Define a map $\eta:M\rightarrow \mathfrak{m}_K/\mathfrak{m}_K^{n+1},\ \Pi\mapsto \pi/\mathfrak{m}_K^{n+1}$. Then the triple $(1,\operatorname{Id}_R,\eta)$ gives an isomorphism from $T$ to $T_n(K)$.\\ (3) Let $f\in \mathrm{Hom}(R_n(K_1),R_n(K_2))$. Let $\pi_1$ and $\pi_2$ be uniformizers of $K_1$ and $K_2$ respectively. By the choice of $n$, $f(\pi_1/\mathfrak{m}_{K_1}^n)$ generates the maximal ideal of $R_n(K_2)$ so that $f(\pi_1/\mathfrak{m}_{K_1}^n)=(a_f\pi_2)/\mathfrak{m}_{K_2}^{n+1}$ for a unit $a_f\in R(K_2)^{\times}$. Define a map $\eta_f:M_n(K_1)\rightarrow M_n(K_2)$ sending $\pi_1/\mathfrak{m}_{K_1}^{n+1}$ to $(a_f\pi_2)/\mathfrak{m}_{K_2}^{n+1}$. Then the triple $T_n(f):=(1,f,\eta_f)$ gives a morphism from $T_n(K_1)$ to $T_n(K_2)$. Note that $T_n(f)$ does not depend on the choices of $\pi_1$ and $\pi_2$ because $M_n(K_1)$ and $M_n(K_2)$ are free $R_n(K_1)$ and $R_n(K_2)$-modules of rank $1$ respectively. Now we show $T_n$ is commute with the composition. Let $\pi_3$ be a uniformizer of $K_3$. Fix $f\in \mathrm{Hom}(R_n(K_1),R_n(K_2))$ and $g\in \mathrm{Hom}(R_n(K_2),R_n(K_3))$. There are $a_f\in R(K_2)^{\times}$, $b_g, b_{g\circ f}\in R(K_3)^{\times}$ such that $f(\pi_1/\mathfrak{m}_{K_1}^n)=(a_f\pi_2)/\mathfrak{m}_{K_2}^n$, $g(\pi_2/\mathfrak{m}_{K_1}^n)=(b_g\pi_3)/\mathfrak{m}_{K_3}^n$, and $g\circ f(\pi_1/\mathfrak{m}_{K_1}^n)=(b_{g\circ f}\pi_3)/\mathfrak{m}_{K_3}^n$. Let $b_f\in R(K_3)^{\times}$ such that $g(a_f/\mathfrak{m}_{K_2}^n)=b_f/\mathfrak{m}_{K_3}^n$. Then we have that $b_{g\circ f}/\mathfrak{m}_{K_3}^n=(b_gb_f)/\mathfrak{m}_{K_3}^n$. By choosing $b_g$ and $b_f$ properly, we may assume that $b_gb_f-b_{g\circ f}\in \mathfrak{m}_{K_3}^{n+1}$ $(\dagger)$. Define $\eta_{g\circ f}: M_n(K_1)\rightarrow M_n(K_3), \pi_1/\mathfrak{m}_{K_1}^{n+1}\mapsto (b_{g\circ f}\pi_3)/\mathfrak{m}_{K_3}^{n+1}$. By $(\dagger)$, we have that $\eta_{g\circ f}=\eta_g\circ \eta_f$. Therefore, \begin{align*} T_n(g\circ f)&= (1,g\circ f, \eta_{g\circ f})\\ &= (1,g\circ f, \eta_g\circ \eta_f)\\ &= (1,g,\eta_g)\circ (1,f,\eta_f)\\ &= T_n(g)\circ T_n(f). \end{align*} \end{proof} \begin{corollary}\label{cor:vhf_unique_realizable} Let $H$ be a finitely ramified discrete valued hyperfield of mixed characteristic with the perfect residue field and let $e$ be the ramification index. Suppose $l(:=l(H))>e(1+\nu_R(e))$ so that $e$ is not zero in $R(:=R_l(H))$. Then there is a unique complete discrete valued field $K$ (up to isomorphic) such that $\mathcal H_l(K)\cong H$. \end{corollary} \begin{rem/def}\label{rem/def:general_HomoverZ} Let $H_1$ and $H_2$ be finitely ramified discrete valued hyperfield of mixed characteristic $(0,p)$ with perfect residue fields. Let $\operatorname{Tr}(H_1)=(R_1,M_1,\epsilon_1)$ and $\operatorname{Tr}(H_2)=(R_2,M_2,\epsilon_2)$, and we identify $H_1=\operatorname{U}(\operatorname{Tr}(H_1))$ and $H_2=\operatorname{U}(\operatorname{Tr}(H_2))$. Suppose they have the same length $l$ and the same ramification index $e$ so that $n=\nu_1(e)=\nu_2(e)\in \{0,1,\ldots, l-1\}\cup\{\infty\}$, where $\nu_1=\nu_{R_1}$ and $\nu_2=\nu_{R_2}$. Suppose $l>e$. Since $l>e$, $p$ is not zero in $R_l$ and $R_2$ and $\nu_1(p)=\nu_2(p)=e$. So $p=a_1\pi_1^e$ in $R_1$ and $p=a_2\pi_2^e$ in $R_2$ for some units $a_1,a_2$ and some uniformizers $\pi_1,\pi_2$ in $R_1$ and $R_2$ respectively. Take $\Pi_i\in M_i$ such that $\epsilon_i(\Pi_i)=\pi_i$ for $i=1,2$. Then $\nu_{\operatorname{Tr}(H_i)}(a_i \Pi_i^{\otimes e})=e$ for $i=1,2$. We say a homomorphism $f:H_1\rightarrow H_2$ is {\em over $p$} if $f(a_1\Pi_1^{\otimes e})=a_2\Pi_2^{\otimes e}$. Denote $\mathrm{Hom}_{\BZ}(H_1,H_2)$ for the set of all homomorphisms from $H_1$ to $H_2$, which are over $p$. \end{rem/def} \begin{example}\label{ex:Hom_and_HomoverZ} Let $K_1=K_2={\mathbb Q}_3(\sqrt{3})$ and $\mathfrak{m}$ be the maximal ideal of the valuation ring ${\mathbb Z}_3[\sqrt{3}]$. Note that $R:=R_4(K_1)=R_4(K_2)\cong ({\mathbb Z}_3/9{\mathbb Z}_3)[x]/(x^2-3)$. Then there is an isomorphism $f:a+bx\mapsto a+4bx$ in $\mathrm{Hom}(R_4(K_1),R_4(K_2))$. Then $f$ induces an isomorphism $T_4(f):(R,\mathfrak{m}/\mathfrak{m}^4,\epsilon_4)\rightarrow (R,\mathfrak{m}/\mathfrak{m}^4,\epsilon_4)$ and it induces an isometric isomorphism $U(T_4(f)):\mathcal H_4(K_1)\rightarrow \mathcal H_4(K_2),\sqrt{3}(1+\mathfrak{m}^4)\mapsto4\sqrt{3}(1+\mathfrak{m}^4)$, which is not in $\mathrm{Hom}_{\BZ}(\mathcal H_4(K_1),\mathcal H_4(K_2))$. Suppose $U(T_4(f))$ is over $p$. Then $3(1+\mathfrak{m}^4)=U(T_4(f))(3(1+\mathfrak{m}^4))=U(T_4(f))((\sqrt{3})^2(1+\mathfrak{m}^4))=(4\sqrt{3})^2(1+\mathfrak{m}^4)=4^2 3(1+\mathfrak{m}^4)$. So, we have $(1+\mathfrak{m}^4)=4^2 (1+\mathfrak{m}^4)$, which is impossible, because $15\notin \mathfrak{m}^4=9{\mathbb Z}_3$. \end{example} \bigskip Now we introduce some categories of valued hyperfields, truncated DVRs, and valuation rings and we study relationships between them. We recall two categories of truncated DVRs and valuation rings, which were used to generalize the functoriality of unramified valuation rings in \cite[Section 4]{LL}. For a prime number $p$ and a positive integer $e$, let ${\mathcal C}_{p,e}$ be a category consisting of the following data : \begin{itemize} \item $\operatorname{Ob}({\mathcal C}_{p,e})$ is the family of complete discrete valuation rings of mixed characteristic having perfect residue fields of characteristic $p$ and the ramification index $e$; and \item $\operatorname{Mor}_{{\mathcal C}_{p,e}}(R_1,R_2):=\mathrm{Hom}(R_1,R_2)$ for $R_1$ and $R_2$ in $\operatorname{Ob}({\mathcal C}_{p,e})$. \end{itemize} Let ${\mathcal R}_{p,e}^{n}$ be a category consisting of the following data : \begin{itemize} \item For $n\le e$, $\operatorname{Ob}({\mathcal R}_{p,e}^{n})$ is the family of truncated DVRs $\overline{R}$ of length $n$ with perfect residue fields of characteristic $p$, and for $n>e$, $\operatorname{Ob}({\mathcal R}_{p,e}^{n})$ is the family of truncated DVRs $\overline{R}$ of length $n$ with perfect residue fields of characteristic $p$ such that $p\in \overline{\mathfrak{m}}^{e}\setminus\overline{\mathfrak{m}}^{e+1}$ where $\overline{\mathfrak{m}}$ is the maximal ideal of $\overline{R}$; and \item $\operatorname{Mor}_{{\mathcal R}_{p,e}^{n}}(\overline{R_1},\overline{R_2}) :=\mathrm{Hom}(\overline{R_1},\overline{R_2})$ for $\overline{R_1}$ and $\overline{R_2}$ in $\operatorname{Ob}({\mathcal R}_{p,e,}^{n})$, \end{itemize} Note that for $e_1,e_2\ge 1$ and for $n\le e_1,e_2$, two categories ${\mathcal R}_{p,e_1}^n$, ${\mathcal R}_{p,e_2}^n$ are the same. For each $m>n$, let $\Pr_n:\ {\mathcal C}_{p,e}\rightarrow {\mathcal R}_{p,e}^n$ and $\Pr^m_n:\ {\mathcal R}_{p,e}^m\rightarrow {\mathcal R}_{p,e}^n$ be the canonical projection functors respectively. Given a prime number $p$ and a positive integer $e$, let $l_{p,e}:=e(1+\nu(e))$ for some(every) $R\in \operatorname{Ob}({\mathcal C}_{p,e})$. Note that for $p\not| e$, we have that $l_{p,e}=e$. \begin{fact}\cite[Definition 4.2, Theorem 4.7]{LL}\label{fact:lifting_DVR_valuedfield} Fix a prime number $p$ and a positive integer $e$. For every $n>l_{p,e}$, there is a functor $\operatorname{L}:{\mathcal R}_{p,e}^n \longrightarrow {\mathcal C}_{p,e}$, called an {\em $n$-th lifting functor}, which satisfies the following: \begin{enumerate} \item $(\Pr_n\circ \operatorname{L})(\overline{R})\cong \overline{R}$ for each $\overline{R}$ in $\operatorname{Ob}({\mathcal R}_{p,e}^n)$. \item $\Pr_1 \circ\operatorname{L}$ is equivalent to $\Pr^{n}_1$. \item $\operatorname{L}\circ \Pr_n$ is equivalent to $\operatorname{Id}_{{\mathcal C}_{p,e}}$. \end{enumerate} Moreover, there is a unique $n$-th lifting functor $\operatorname{L}$ satisfying \begin{enumerate} \item[(4)] For each $g\in \operatorname{Mor}_{{\mathcal R}_{p,e}^n}(\overline{R}_1,\overline{R}_2)$ and for any $x\in \operatorname{L}(\overline{R}_1)$, there is a representative $b_x$ of $(\Pr_n\circ \operatorname{L})(g)(x+\mathfrak{m}_1^n)$ such that $$\nu_2(\operatorname{L}(g)(x)-b_x)>M(\operatorname{L}(\overline{R}_1)).$$ \end{enumerate} \end{fact} Next we introduce two categories of valued hyperfields. Let $\mathcal H_{p,e}^n$ be a category consisting of the following data : \begin{itemize} \item $\operatorname{Ob}(\mathcal H_{p,e}^n)$ is the family of discrete normalized valued hyperfield of length $n$ and mixed characteristic $(0,p)$ having perfect residue fields, and for $n>e$ in addition, having ramification indices $e$; and \item For $n\le e$, $\operatorname{Mor}_{\mathcal H_{p,e}^n}(H_1,H_2):=\mathrm{Iso}(H_1,H_2)$ and for $n>e$, $\operatorname{Mor}_{\mathcal H_{p,e}^n}(H_1,H_2):=\mathrm{Iso}_{{\mathbb Z}}(H_1,H_2)$, where $\mathrm{Iso}_{{\mathbb Z}}(H_1,H_2)=\mathrm{Hom}_{\BZ}(H_1,H_2)\cap \mathrm{Iso}(H_1,H_2)$ for $H_1$ and $H_2$ in $\operatorname{Ob}(\mathcal H_{p,e}^n)$. \end{itemize} Let $\widehat \mathcal H_{p,e}^n$ be a category consisting of the following data : \begin{itemize} \item $\operatorname{Ob}(\widehat \mathcal H_{p,e}^n)=\operatorname{Ob}(\mathcal H_{p,e}^n)$; and \item $\operatorname{Mor}_{\widehat \mathcal H_{p,e}^n}(H_1,H_2):=\mathrm{Iso}(H_1,H_2)$ for $H_1$ and $H_2$ in $\operatorname{Ob}(\widehat \mathcal H_{p,e}^n)$. \end{itemize} By Remark/Definition \ref{rem/def:residuefield_vhf}, we have functors $\operatorname{Res}_{p,e}^n:\mathcal H_{p,e}^n\rightarrow {\mathcal R}_{p,e}^1$ and $\widehat{\res}:\widehat{\H}_{p,e}^n\rightarrow {\mathcal R}_{p,e}^1$. Next we introduce lifting functors for two categories $\mathcal H_{p,e}^n$ and $\widehat{\H}_{p,e}^n$, which are analogous to the lifting functors of ${\mathcal R}_{p,e}^n$ in \cite[Definition 4.2]{LL}. \begin{definition}\label{n-liftable} Fix a prime number $p$ and a positive integer $e$. We say that the category ${\mathcal C}_{p,e}$ is \emph {$n$-$\mathcal H$-liftable} if there is a functor $\operatorname{L}^{\mathcal H}:\mathcal H_{p,e}^n \rightarrow {\mathcal C}_{p,e}$ which satisfies the following: \begin{itemize} \item $(\mathcal H_n\circ \operatorname{L}^{\mathcal H})(\overline{R})\cong \overline{R}$ for each $\overline{R}$ in $\operatorname{Ob}({\mathcal R}_{p,e}^n)$. \item $\Pr_1 \circ\operatorname{L}^{\mathcal H}$ is equivalent to $\operatorname{Res}_{p,e}^n$. \item $\operatorname{L}^{\mathcal H}\circ \mathcal H_n$ is equivalent to $\operatorname{Id}_{{\mathcal C}_{p,e}}$. \end{itemize} We say that $\operatorname{L}^{\mathcal H}$ is an \emph{$n$-th $\mathcal H$-lifting functor} of ${\mathcal C}_{p,e}$. \end{definition} Now we see relationships between the categories ${\mathcal C}_{p,e}$, ${\mathcal R}_{p,e}^n$, $\mathcal H_{p,e}^n$, and $\widehat{\H}_{p,e}^n$. Define a functor $\widehat{\U}: {\mathcal R}_{p,e}^n\rightarrow \widehat{\H}_{p,e}^n$ as follows: \begin{itemize} \item For $\overline{R}\in \operatorname{Ob}({\mathcal R}_{p,e}^n)$, $\widehat{\U}(\overline{R}):=(\operatorname{U}\circ T_n\circ \operatorname{L})(\overline{R} )$ after rescaling $\theta_{(\operatorname{U}\circ T_n\circ \operatorname{L})(\overline{R} )}=1/e$, where $\operatorname{L}$ is the $n$-th lifting in Fact \ref{fact:lifting_DVR_valuedfield}; and \item For $f\in \operatorname{Mor}_{{\mathcal R}_{p,e}^n}(\overline{R}_1,\overline{R}_2)$, $\widehat{\U}(f):=(\operatorname{U}\circ \operatorname{T}_n)(f)$, \end{itemize} and define a functor $\widehat{\To}:\widehat{\H}_{p,e}^n\rightarrow {\mathcal R}_{p,e}^n$ as follows: \begin{itemize} \item For $H\in \operatorname{Ob}(\widehat{\H}_{p,e}^n)$, $\widehat{\To}(H):=\overline{R}$ where $\overline{R}$ is a truncated DVR in $\operatorname{Tr}(H)$. \item For $g\in \operatorname{Mor}_{\widehat{\H}_{p,e}^n}(H_1,H_2)$, $\widehat{\To}(g)$ is a morphism in $ \operatorname{Mor}_{{\mathcal R}_{p,e}^n}(\widehat{\To}(H_1),\widehat{\To}(H_1))$ such that $\operatorname{Tr}(f)=(1,\widehat{\To}(g),\eta)$. \end{itemize} By Remark \ref{rem:triple_to_vhf} and Proposition \ref{prop:DVR_enough}, we have the following result. \begin{theorem}\label{thm:equiv_DVR_hvf} Let $p$ be a prime number and $e$ be a positive integer. Fix $n>e$. \begin{enumerate} \item $\widehat{\U}\circ \widehat{\To}$ is equivalent to $\operatorname{Id}_{\widehat{\H}_{p,e}^n}$. \item $\widehat{\To}\circ \widehat{\U}$ is equivalent to $\operatorname{Id}_{{\mathcal R}_{p,e}^n}$ \end{enumerate} Therefore, two categories $\widehat{\H}_{p,e}^n$ and ${\mathcal R}_{p,e}^n$ are equivalent. \end{theorem} \begin{remark}\cite[Proposition 4.9]{LL}\label{prop:n_atleast_e} Let $R_1/\operatorname{W}(k)$ and $R_2/\operatorname{W}(k)$ be totally ramified extensions of degree $e$. Then $R_{1,e}$ is isomorphic to $R_{2,e}$ as $\operatorname{W}(k)$-algebras. So, the assumption that $n>e$ is natural in Theorem \ref{thm:equiv_DVR_hvf}. \end{remark} \noindent Each $R\in \operatorname{Ob}({\mathcal C}_{p,e})$ gives a discrete valued hyperfield $\mathcal H(R):=\mathcal H_n(K)$ in $\operatorname{Ob}(\mathcal H_{p,e}^n)$ where $K$ is the fraction field of $R$ after rescaling $\theta_{\mathcal H_n(R)}=1/e$. Also each $f\in \operatorname{Mor}_{{\mathcal C}_{p,e}}(R_1,R_2)$ induces a morphism $\mathcal H(f):\mathcal H(R_1)\rightarrow \mathcal H(R_2)$. So we have a functor $\mathcal H:{\mathcal C}_{p,e}\rightarrow \mathcal H_{p,e}^n$. For $n>l_{p,e}$, we define a functor $\operatorname{L}^{\mathcal H}:\mathcal H_{p,e}^n\rightarrow {\mathcal C}_{p,e}$. Since $\mathcal H_{p,e}^n$ is a subcategory of $\widehat{\H}_{p,e}^n$, we have a functor $\operatorname{To}:\mathcal H_{p,e}^n\rightarrow {\mathcal R}_{p,e}^n$ by restricting $\widehat{\To}$ to $\mathcal H_{p,e}^n$. For each $H\in \operatorname{Ob}(\mathcal H_{p,e}^n)$, there is a unique( up to isomorphic) complete valued field $K$ such that $H\cong \mathcal H_n(K)$, where $K$ is the fraction field of $(\operatorname{L}\circ\operatorname{To})(H)$ by Corollary \ref{cor:criteria_isom}. For $H_1,H_2\in \operatorname{Ob}(\mathcal H_{p,e}^n)$ and $f\in \operatorname{Mor}_{\mathcal H_{p,e}^n}(H_1,H_2)$, we have a morphism $\operatorname{L}^{\mathcal H}(f)\in \operatorname{Mor}_{{\mathcal C}_{p,e}}(R_1,R_2)$ where $R_1=(\operatorname{L}\circ \operatorname{To})(H_1)$ and $R_2=(\operatorname{L}\circ \operatorname{To})(H_2)$ by Theorem \ref{thm:mainhomlifting}. By Proposition \ref{prop:funtoriality}, $\operatorname{L}^{\mathcal H}$ forms a functor. By Remark \ref{rem:vf_rigid_hom}, Theorem \ref{thm:mainhomlifting}, and Proposition \ref{prop:funtoriality}, we have the following result which is analogous to Fact \ref{fact:lifting_DVR_valuedfield} in the case of valued hyperfields. \begin{theorem}\label{thm:lifting_vhf_valuedfield} Let $p$ be a prime number and $e$ be a positive integer. Fix $n>l_{p,e}$. Then ${\mathcal H}_{p,e}^n$ is $n$-$\mathcal H$-liftable and there is an $n$-th $\mathcal H$-lifting functor $\operatorname{L}^{\mathcal H}$ of ${\mathcal H}_{p,e}^n$ satisfying the following: \begin{enumerate} \item There is an isometric isomorphism, which is over $p$, between $(\mathcal H\circ \operatorname{L}^{\mathcal H})(H)$ and $H$ for each $H$ in $\operatorname{Ob}({\mathcal H}_{p,e}^n)$. \item $\Pr_1\circ\operatorname{L}^{\mathcal H}$ is equivalent to $\operatorname{Res}_{p,e}^n$. \item $\operatorname{L}^{\mathcal H}\circ \mathcal H$ is equivalent to $\operatorname{Id}_{{\mathcal C}_{p,e}}$. \end{enumerate} Moreover, there is a unique $n$-th lifting functor $\operatorname{L}^{\mathcal H}$ satisfying \begin{enumerate} \item[(4)] For each $g\in \operatorname{Mor}_{\mathcal H_{p,e}^n}(H_1,H_2)$ and for any $x\in \operatorname{L}^{\mathcal H}(H_1)$, there is a representative $b_x$ of $(\mathcal H_n\circ \operatorname{L})(g)([x])$ such that $$\nu_2(\operatorname{L}^{\mathcal H}(g)(x)-b_x)>M(\operatorname{L}^{\mathcal H}(H_1)).$$ \end{enumerate} If we take $\operatorname{L}$ satisfying (1)-(4) and $\operatorname{L}^{\mathcal H}$ satisfying (1)-(4), then $\operatorname{L}^{\mathcal H}=\operatorname{L}\circ \operatorname{To}$. \end{theorem} \noindent Moreover, by Theorem \ref{thm:tame_vf_vhf}, we have the following theorem. \begin{theorem}\label{thm:tame_equiv_vhf_valuedfields} Let $p$ be a prime number and $e$ be a positive integer. Suppose $p$ does not divide $e$ so that $l_{p,e}=e$. For any $n>e$, two categories $\mathcal H_{p,e}^n$ and ${\mathcal C}_{p,e}$ are equivalent \end{theorem} \noindent In summary, for $n> l_{p,e}$, we have the following diagram between categories of ${\mathcal C}_{p,e}$, ${\mathcal R}_{p,e}^n$, $\mathcal H_{p,e}^n$, and $\widehat{\H}_{p,e}^n$: $$ \begin{tikzcd} \mathcal H_{p,e}^n \arrow[d, hook] \arrow[dr, "\operatorname{To}" description] \arrow[r, shift left, "\operatorname{L}^{\mathcal H}"] \arrow[r, shift right, leftarrow, "\mathcal H"'] & {\mathcal C}_{p,e} \arrow[d, shift left, "\Pr_n"] \arrow[d, shift right, leftarrow, "\operatorname{L}"'] \\ \widehat{\H}_{p,e}^n \arrow[r, shift left, leftarrow, "\widehat{\U}"] \arrow[r, shift right, "\widehat{\To}"'] & {\mathcal R}_{p,e}^n \end{tikzcd} $$ , and we can take $\operatorname{L}^{\mathcal H}=\operatorname{L}\circ \operatorname{To}$. \begin{question}\label{question:equiv_vhf_valuedfields} Let $p$ be a prime number and $e$ be a positive integer. Fix $n>l_{p,e}$. We know that two categories $\widehat{\H}_{p,e}^n$ and ${\mathcal R}_{p,e}^n$ are equivalent, and $\operatorname{L}^{\mathcal H}\circ \mathcal H$ is equivalent to $\operatorname{Id}_{{\mathcal C}_{p,e}}$. And for the tame case, that is, $p\not|e$, $\mathcal H_{p,e}^n$ and ${\mathcal C}_{p,e}$ are equivalent. Is $\mathcal H_n\circ \operatorname{L}^{\mathcal H}$ equivalent to $\operatorname{Id}_{\mathcal H_{p,e}^n}$ if $p|e$? \end{question} \section{Relative completeness via hyper fields}\label{section:AKE_hyperfields} In this section, we aim to prove an AKE-type relative completeness theorem in terms of valued hyperfields for finitely ramified valued fields. We first recall basic facts on coarsenings of valuations. \begin{rem/def}\label{rem/def:coarse_valuation}\cite{PR} Let $(K,\nu, k, \Gamma)$ be valued field. Let $\Gamma^{\circ}$ be a convex subgroup of $\Gamma$ and $\dot{\nu}:K\setminus\{0\}\longrightarrow \Gamma/\Gamma^{\circ}$ be a map sending $x(\neq 0)\in K$ to $\nu(x)+\Gamma^{\circ}\in \Gamma/\Gamma^{\circ}$. The map $\dot{\nu}$ is a valuation, called {\em a coarse valuation} of $\nu$ with respect to $\Gamma^{\circ}$. The residue field $K^{\circ}$, called {\em the core field} of $(K,\nu)$ with respect to $\Gamma^{\circ}$, of $(K,\dot{\nu})$ forms a valued field equipped with a valuation $\nu^{\circ}$ induced from $\nu$ and the value groups $\Gamma^{\circ}$. More precisely, the valuation $\nu^{\circ}$ is defined as follows: Let $\pr_{\dot{\nu}} : R_{\dot{\nu}}\longrightarrow K^{\circ}$ be the canonical projection map and let $x\in R_{\dot{\nu}}$. If $x^{\circ}:=\pr_{\dot{\nu}}(x)\in K^{\circ}\setminus\{0\}$, then $\nu^{\circ}(x^{\circ}):=\nu(x)$. And $x^{\circ}=0\in K^{\circ}$ if and only if $\nu(x)>\Gamma^{\circ}$, that is, $\nu(x)>\gamma$ for all $\gamma\in \Gamma^{\circ}$. If $K$ is of characteristic $0$ and $\Gamma^{\circ}$ is non-trivial, then $(K,\dot{\nu})$ is always of equal characteristic $(0,0)$. \end{rem/def} \begin{fact}\label{fact:coarse_valuation} Let $(K,\nu,\Gamma)$ be valued field. Let $\dot{\nu}$ be the coarse valuation and $K^{\circ}$ be the core field with respect to $\Gamma^{\circ}$ for a non-trivial convex subgroup $\Gamma^{\circ}$ of $\Gamma$. \begin{enumerate} \item Let $R_{\nu}$, $R_{\dot{\nu}}$, and $R_{\nu^{\circ}}$ be the valuation rings of $(K,\nu)$, $(K,\dot{\nu})$, and $(K^{\circ},\nu^{\circ})$ respectively. Then $(\pr_{\dot{\nu}})^{-1}(R_{\nu^{\circ}})=R_{\nu}$. \item If $(K,\nu)$ is finitely ramified, then $(K^{\circ},\nu^{\circ})$ is finitely ramified, and $K$ and $K^{\circ}$ have the same ramification index. \item If $(K,\nu)$ is finitely ramified and $\aleph_1$-saturated, and $\Gamma^{\circ}$ is the smallest non-trivial convex subgroup, then $(K^{\circ},\nu^{\circ})$ is complete. \end{enumerate} \end{fact} \begin{definition}\label{def:restricted_hf} Let $(K,\nu,\Gamma)$ be valued field. Let $\Gamma^{\circ}\subset \Gamma$ be a non-trivial convex subgroup and $\gamma\in (\Gamma^{\circ})_{>0}$. Let $0_{\Gamma^{\circ}}^{\gamma}:=\{\alpha\in \mathcal H_{\gamma}(K)|\ \nu_{\mathcal H}(\alpha)>\Gamma^{\circ} \}$. If $\gamma$ is obvious, we write $0_{\Gamma^{\circ}}^{\gamma}=0_{\Gamma^{\circ}}$. Define $\mathcal H_{\gamma}^{\times}(K,\Gamma^{\circ}):=\{\alpha\in\mathcal H_{\gamma}(K): \nu_{\mathcal H}(\alpha)\in \Gamma^{\circ}\}$ and $\mathcal H_{\gamma}(K,\Gamma^{\circ}):=\mathcal H_{\gamma}^{\times}(K,\Gamma^{\circ})\cup \{ 0_{\Gamma^{\circ}}^{{\gamma}} \}$. \end{definition} \begin{remark} Let $\Gamma^{\circ}$ be a non-trivial convex subgroup of $\Gamma$. Then $\mathcal H_{\gamma}(K,\Gamma^{\circ})$ forms a valued hyperfield. \end{remark} \begin{proof} Let $(\mathcal H_{\gamma}(K),+_{\mathcal H},\cdot_{\mathcal H},\nu_{\mathcal H})$ be the valued $\gamma$-hyperfield of $K$. First, $\mathcal H_{\gamma}^{\times}(K,\Gamma^{\circ})$ is a multiplicative subgroup of $(\mathcal H_{\gamma}^{\times}(K),\cdot_{\mathcal H})$ because $\Gamma^{\circ}$ is a subgroup of $\Gamma$. Note that for $\alpha,\beta\in \mathcal H_{\gamma}^{\times}(K)$, if there is $x\in \alpha+_{\mathcal H}\beta$ such that $\nu(x)>\Gamma^{\circ}$, then $0_{\Gamma^{\circ}}\subset \alpha+_{\mathcal H}\beta$($\dagger$). For $\alpha=[a]$ and $\beta=[b]$, $\bigcup \alpha+_{\mathcal H}\beta=(a+b)+\mathfrak{m}^{\gamma+\min\{\nu(a),\nu(b)\}}$ by Lemma \ref{lem:basic_on_hyperfield}(2). Suppose $x\in \bigcup \alpha+_{\mathcal H}\beta$ such that $\nu(x)>\Gamma^{\circ}$. If $\nu(a+b)<\gamma+\min\{\nu(a),\nu(b)\}$, then $\nu(x)=\nu(a+b)<\gamma+\min\{\nu(a),\nu(b)\}\in\Gamma^{\circ}$ because $\gamma,\nu(a),\nu(b)\in \Gamma^{\circ}$. So we have that $\nu(a+b)\ge \gamma+\min\{\nu(a),\nu(b)\}$ and $\bigcup\alpha+_{\mathcal H}\beta=\mathfrak{m}^{\gamma+\min\{\nu(a),\nu(b)}\supset 0_{\Gamma^{\circ}}$. We define a multivalued operation $+_{\mathcal H}^{\Gamma^{\circ}}$ on $\mathcal H_{\gamma}(K,\Gamma^{\circ})$ as follows: For $\alpha,\beta\in \mathcal H_{\gamma}^{\times}(K)$, \begin{itemize} \item $\alpha+_{\mathcal H}^{\Gamma^{\circ}}0_{\Gamma^{\circ}}=\alpha=0_{\Gamma^{\circ}}+_{\mathcal H}^{\Gamma^{\circ}}\alpha$. \item $\alpha+_{\mathcal H}^{\Gamma^{\circ}}\beta=\begin{cases} \alpha+_{\mathcal H}\beta& \mbox{if } \forall x\in \bigcup \alpha+_{\mathcal H}\beta,\ \nu(x)\in \Gamma^{\circ}\\ (\alpha+_{\mathcal H}\beta)\cap(\mathcal H_n)^{\times}(K)\cup\{0_{\Gamma^{\circ}}\}& \mbox{if } \exists x\in \bigcup \alpha+_{\mathcal H}\beta,\ \nu(x)>\Gamma^{\circ} \end{cases}$. \end{itemize} By ($\dagger$) and covexity of $\Gamma^{\circ}$, $+_{\mathcal H}^{\Gamma^{\circ}}$ is well-defined. Define a map $\nu_{\mathcal H}^{\Gamma^{\circ}}$ on $\mathcal H_{\gamma}(K,\Gamma^{\circ})$ as follows: For $\alpha\in \mathcal H_{\gamma}(K,\Gamma^{\circ})$, $\nu_{\mathcal H}^{\Gamma^{\circ}}(\alpha)=\nu_{\mathcal H}(\alpha)$ if $\alpha\neq 0_{\Gamma^{\circ}}$ and $\nu_{\mathcal H}^{\Gamma^{\circ}}(\alpha)=\infty$ if $\alpha=0_{\Gamma^{\circ}}$. Then $(\mathcal H_{\gamma}(K,\Gamma^{\circ}),+_{\mathcal H}^{\Gamma^{\circ}}, \cdot_{\mathcal H}, \nu_{\mathcal H}^{\Gamma^{\circ}})$ forms a valued hyperfield. \end{proof} \begin{lemma}\label{lem:coarse_valuation} Let $(K,\nu,\Gamma)$ be valued field. Let $\dot{\nu}$ be the coarse valuation and $K^{\circ}$ be the core field with respect to $\Gamma^{\circ}$ for a non-trivial convex subgroup $\Gamma^{\circ}$ of $\Gamma$. For each $\gamma\in \Gamma^{\circ}_{>0}$, $\mathcal H_{\gamma}(K,\Gamma^{\circ})$ and $\mathcal H_{\gamma}(K^{\circ})$ are isomorphic. \end{lemma} \begin{proof} Consider a map $f: \mathcal H_{\gamma}(K,\Gamma^{\circ})\rightarrow \mathcal H_{\gamma}(K^{\circ})$ defined as follows: For $\alpha\in \mathcal H_{\gamma}(K,\Gamma^{\circ})$, $f(\alpha)=[a^{\circ}]$ if $\alpha=[a]$, and $f(\alpha)=0$ if $\alpha=0_{\Gamma^{\circ}}$. The map $f$ is well-defined. Suppose $[a]=[b]\in\mathcal H_{\gamma}(K,\Gamma^{\circ})$. Then $a/b\in (1+\mathfrak{m}^{\gamma})$ and $\nu(a)=\nu(b)\in \Gamma^{\circ}$. So, we have that $(a/b)^{\circ}=a^{\circ}/b^{\circ}\in (1+(\mathfrak{m}^{\circ})^{\gamma})$, and it implies $[a^{\circ}]=[b^{\circ}]$. And it is clear that for all $x\in K$ with $\nu(x)>\Gamma^{\circ}$, $x^{\circ}=0$. It is routinely to check that $f$ is a homomorphism. \end{proof} \noindent We recall the following facts before proving a relative completeness theorem. \begin{fact}\label{fact:KS_iso_theorem}{(Keisler-Shelah Isomorphism Theorem)} Let ${\mathcal M}$ and ${\mathcal N}$ be two first order structures. If ${\mathcal M}\equiv {\mathcal N}$, then there is a ultrafilter ${\mathcal U}$ on an infinite set $I$ such that $${\mathcal M}^{{\mathcal U}}\cong {\mathcal N}^{{\mathcal U}},$$ where ${\mathcal M}^{{\mathcal U}}$ and ${\mathcal N}^{{\mathcal U}}$ are the ultrapowers of ${\mathcal M}$ and ${\mathcal N}$ with respect to ${\mathcal U}$. \end{fact} \begin{fact}\cite[Lemma 1.5]{Be}\label{fact:def_resiue_ring} Let $(K,\nu)$ be a finitely ramified henselian valued field of mixed characteristic $(0,p)$. Then the valuation ring $R(K)$ of $(K,\nu)$ is definable by the formula $$\phi_q(x)\equiv \exists y\ y^{q}=1+px^{q}$$ for some $q>0$ such that $p\not|q$ and $q>e_{\nu}(p)$. \end{fact} \begin{theorem}\label{thm:AKE_hyperfield} Let $(K_1,\nu_1,k_1,\Gamma_1)$ and $(K_2,\nu_2,k_2,\Gamma_1)$ be finitely ramified henselian valued fields of mixed characteristic $(0,p)$. Suppose $k_1$ and $k_2$ are perfect fields. Let $n> e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$. The following are equivalent: \begin{enumerate} \item $K_1\equiv K_2$. \item $R_{n}(K_1)\equiv R_{n}(K_2)$ and $\Gamma_1\equiv \Gamma_2$. \item $\mathcal H_{n}(K_1)\equiv_{\mathcal H_{n}(\{p\})} \mathcal H_{n}(K_2)$. \end{enumerate} \end{theorem} \begin{proof} For $(1)\Leftrightarrow (2)$, see Fact \ref{fact:fined_AKE_Rn}. It is clear that $(1)$ implies $(3)$. We show $(3)$ implies $(1)$. Suppose $(3)$ holds. By Fact \ref{fact:KS_iso_theorem}, we may assume that $\mathcal H_{n}(K_1)\cong_{\mathcal H_{n}({\mathbb Z})} \mathcal H_{n}(K_2)$. By Remark \ref{rem:vhf_valuegroup}, $\Gamma_1\cong \Gamma_2$ and we may assume that $\Gamma_1=\Gamma_2=\Gamma$. By taking ultrapowers of $K_1$ and $K_2$ with respect to a nonprincipal ultrafilter on ${\mathbb N}$, we may assume that $K_1$ and $K_2$ are $\aleph_1$-saturated. Let $\dot{\nu}_1$ and $\dot{\nu}_2$ be the coarse valuation of $\nu_1$ and $\nu_2$ with respect to the smallest non-trivial convex subgroup $\Gamma^{\circ}$. We have two valued fields $(K_1,\dot{\nu}_1)$ and $(K_2,\dot{\nu}_2)$ of equal characteristic $(0,0)$ with residue fields $K_1^{\circ}$ and $K_2^{\circ}$ respectively. By Fact \ref{fact:coarse_valuation}(2) and (3). $(K_1^{\circ},\nu^{\circ}_1)$ and $(K_2^{\circ},\nu^{\circ}_2)$ are complete discrete valued fields of mixed characteristic $(0,p)$. Note that the value groups of $(K_1,\dot{\nu}_1)$ and $(K_2,\dot{\nu}_2)$ are $\Gamma/\Gamma^{\circ}$. Since $\mathcal H_n(K_1)\cong_{\mathcal H_n({\mathbb Z})}\mathcal H_n(K_2)$, we have that $\mathcal H_n(K_1^{\circ})\cong_{\mathcal H_n({\mathbb Z})} \mathcal H_n(K_1^{\circ})$ by Lemma \ref{lem:coarse_valuation}. Since $e_{\nu^{\circ}_i}(p)=e_{\nu_i}(p)$ for $i=1,2$, we have the $K^{\circ}_1\cong K^{\circ}_2$ by Corollary \ref{cor:criteria_isom}. By Fact \ref{fact:AKE}, we have that $(K_1,\dot{\nu}_1)\equiv (K_2,\dot{\nu}_2)$. To show that $(K_1,\nu_1)\equiv(K_2,\nu_2)$, it is enough to show that the valuation rings $R_{\nu}(K_1)$ of $(K_1,\nu_1)$ and $R_{\nu}(K_2)$ of $(K_2,\nu_2)$ are definable in $(K_1,\dot{\nu}_1)$ and $(K_2,\dot{\nu}_2)$ by the same formula. Recall the following result on a definability of a residue ring. Take $l>0$ large enough so that $q:=p^l+1>\max\{e_{\nu_1}(p),e_{\nu_2}(p)\}$. By Fact \ref{fact:def_resiue_ring}, $\phi_q(x)$ defines the residue rings $R_{\nu^{\circ}}(K_1)$ and $R_{\nu^{\circ}}(K_2)$ of $(K^{\circ}_1,\nu^{\circ}_1)$ and $(K^{\circ}_2,\nu^{\circ}_2)$. By Fact \ref{fact:coarse_valuation}(2), the valuation rings $R_{\nu}(K_1)$ and $R_{\nu}(K_2)$ are definable by the same formula in $(K_1,\dot{\nu}_1)$ and $(K_2,\dot{\nu}_1)$ so that $(K_1,\nu_1)\equiv (K_2,\nu_2)$. \end{proof} \noindent By Theorem \ref{thm:tame_vf_vhf} and the proof of $(3)\Rightarrow (1)$ of Theorem \ref{thm:AKE_hyperfield}, we have the following result. \begin{corollary}\label{cor:AKE_hyperfield_tame} Let $(K_1,\nu_1,k_1,\Gamma_1)$ and $(K_2,\nu_2,k_2,\Gamma_1)$ be finitely tamely ramified henselian valued fields of mixed characteristic $(0,p)$. Suppose $k_1$ and $k_2$ are perfect fields. The following are equivalent: \begin{enumerate} \item $K_1\equiv K_2$. \item $\mathcal H_1(K_1)\equiv_{\mathcal H_1(\{p\})} \mathcal H_1(K_2)$. \end{enumerate} \end{corollary} \noindent For local fields of mixed characteristic, they are elementary equivalent if and only if they are isomorphic. So we have the following corollary. \begin{corollary} Let $K_1$ and $K_2$ be local fields of mixed characteristic. Let $n>e_{\nu_2}(p)(1+e_{\nu_1}^2(p))$. The followings are equivalent: \begin{enumerate} \item $\mathcal H_{n}(K_1)\equiv_{\mathcal H_{n}(\{p\})} \mathcal H_{n}(K_2)$. \item $\mathcal H_{n}(K_1)\cong_{\mathcal H_{n}(\{p\})} \mathcal H_{n}(K_2)$. \end{enumerate} \end{corollary}

a421ca1cbc7f2c822cce5ed4aeb38f4786f0cf39

\section{Introduction}\label{sec:introduction} Suppose that you move to a new city and are interested in exploring the local music scene. Typically, you might pick up the arts section of the local newspaper or go online to find a community notice board. Either way you would likely come to a long listing of music events where each event description would provide a small amount of contextual information: the names of the artists, the name and location of the venue, the date and start time of the event, the price of the tickets, and perhaps a few genre labels or a sentence fragment that reflects the kind of music you would expert to hear at the event. While this ``public list of events'' model has been successful at getting fans to music events for many decades, we can use modern recommender systems to make music event discovery more efficient and effective. For example, companies like BandsInTown\footnote{https://www.bandsintown.com} and SongKick\footnote{https://www.songkick.com/} help users \emph{track} artists so that that the user can be notified when a favorite artist will be playing nearby. They also recommend upcoming events with artists who are similar to one or more of the artists that the user has selected to track. These services have been successful in growing both the number of users and in the number of artists and events covered by their service. For example, BandsInTown claims to have 38 million users and lists events for over 430,000 artists\footnote{According to https://en.wikipedia.org/wiki/Bandsintown on March 28, 2018.}. Event listings are added by aggregating information of ticket sellers (e.g., Ticketmaster\footnote{https://www.ticketmaster.com/}, TicketFly\footnote{https://www.ticketfly.com/}) and by artist managers and booking agents who have the ability to directly upload tour dates for their touring artists to these services. While this coverage is impressive, a large percentage of the events found in local newspapers are not listed on these commercial music event recommendation services. Many talented artists play at small venues (e.g., neighborhood pubs, coffee shops, and DIY shows) and are often not represented by (pro-active, tech-savvy) managers. Yet many music fans enjoy the intimacy of a small venue and a personal connection with local artists and may have a hard time discovering these events. As such, our goal is to develop a locally-focused music event recommendation system to help foster music discovery within a local music community. Here we define \emph{local} as all music events within a small geographic region (e.g., 10 square miles). This includes national and regional touring acts who may pass through town but it also includes non-touring artists (e.g., a high school punk band, a barber shop quartet, a jazz trio from the nearby music conservatory, or a neighborhood hip hop collective.) What makes this problem technically challenging is that a large percentage of our local artists have a small \emph{digital footprint} or no digital footprint at all. That is, we may not be able to find these artists on sites that typically provide useful music information \cite{turnbull2008five} (e.g., Spotify\footnote{https://developer.spotify.com/}, Last.fm\footnote{https://www.last.fm/api}, AllMusic\footnote{https://www.allmusic.com/}). Similarly, we often do not have music recordings from these artists so we will not be able to make use of content-based methods for automatic tagging \cite{turnbull2008semantic} or acoustic similarity\cite{mcfee2012learning}. Rather, we will rely the small amount of contextual information that can be scraped from the event listings in the local newspaper or community notice board. We will first introduce the concept of a \emph{Music Event Graph} as a 4-partite graph that connects genre tags to popular artists to event artists to events. We then use latent semantic analysis (LSA) \cite{deerwester1990indexing} to embedding tags and artists into a latent feature space. We show that LSA is particularly advantageous when considering new or not well-known (long-tail) artists who have small digital footprints. This approach also allows us to independently control the \emph{popularity bias}\cite{turnbull2008five} of our event recommendation algorithm so that events with popular artists are no more or less likely to be recommended than events featuring more obscure local artists. \section{Related Work}\label{sec:related-work} We have been unable to find previous research on the specific task of music event recommendation though there is a significant amount of work on both music recommendation \cite{celma2010, schedl2017} (i.e., recommending songs and artists) and event recommendation \cite{macedo15,dooms11, minkov2010} (i.e., events posted on social networking sites.) It both cases, it is common to explore content-based (i.e., the substance of an item), collaborative filtering-based (e.g., usage patterns from a large group of users), and hybrid approaches. We consider our approach to be a hybrid approach since we make use of both social tags (content) and artist similarity (collaborative filtering\footnote{While the details of the Last.fm algorithm for computing artist similarity remain a corporate trade secret, it would be reasonable to expect that these scores are computed using some form of collaborative filtering based on the large quantities of user listening histories that they collect \cite{barrington2009}.}) scores from Last.fm. As with many successful recommender systems, we make use of matrix factorization to embed data into a low dimensional space \cite{koren2009}. In particular, we use Latent Semantic Analysis (LSA) \cite{deerwester1990indexing} which is a common approach used in both text information retrieval \cite{manning2008} and music information retrieval systems (e.g., \cite{levy2007semantic, laurier2009, oramas2015semantic}). LSA is relatively easy to implement\footnote{For example, see http://scikit-learn.org/stable/modules/generated/ sklearn.decomposition.TruncatedSVD.html}, can improve recommendation accuracy, provides a compact representation of the data, works well with sparse input data, and can help alleviate problems caused by synonymy and polysemy \cite{manning2008}. We should note that other embedding techniques, such as probabilistic LSA \cite{hofmann1999probabilistic} and latent Dirichlet allocation (LDA) \cite{blei2003latent}, could also be used as an alternative to LSA. \section{Event Recommendation} When developing an event recommendation system, we will consider an interactive experience with three steps: \begin{enumerate} \item \textbf{User selects genre tags}: Ask the user to select one or more tags from a list of board genres (``rock'', ``hip hop'', ``reggae'') based on the most common genres of the artists who are playing at upcoming local events. \item \textbf{User selects preferred popular artists}: Ask the user to select one or more artists from a list of recognizable mainstream artists (The Beatles, Jay-Z, Bob Marley) based on the selected genres and related to the artists who are playing an upcoming event. \item \textbf{Display of recommended event list}: Show recommended events (with justification) to the user based on the the selected genre tag and popular artist preferences. \end{enumerate} This is a common \emph{onboarding} process for both commercial music event services (e.g., BandsInTown) and music streaming services (e.g., Apple Music) since it quickly gives recommender systems a small but sufficient amount of music preference information for new users. After onboarding, a user can drill down into specific artists or events, as well as listen to related music, explore a map of venues, etc. In this section, we describe the concept of a Music Event Graph and show how we can use it to efficiently recommend local music events based on the music preference information that is collected during user onboarding. \subsection{Music Event Graphs} When considering event recommendation, there are two phases that we need to consider: offline computation of relevance information for all upcoming events and real-time personalized event recommendation. We will use a Music Event Graph to help us structure our event recommendation system. The music event graph is a k-partite graph with $k=4$ levels. Our four levels represent common genre tags, popular artists, event artists, and events as is shown in Figure \ref{fig:event-graph-creation} \begin{figure} \centering \includegraphics[width=3.3in]{CreatingTheEventGraph.png} \caption{Construction a Music Event Graph: First we collect \emph{events} and the \emph{event artists} who will be performing at these events (right nodes and edges). Then, we select \emph{genre tags} and related \emph{popular artists} (left nodes and edges). Finally, we connect event artists to popular artists based on our artist similarity calculation (middle edges).} \label{fig:event-graph-creation} \end{figure} To construct the graph, we follow the following steps: \begin{enumerate} \item Collect a set of upcoming local \emph{events} \item Construct the set of \emph{event artists} from all of the local events \item Find the most frequently used \emph{genre tags} (e.g., ``rock'', ``jazz'', ``hip hop'') associated with the event artists. \item Using the genre tags, create a set of \emph{popular artists} by selecting the most well-known artists that are strongly associated with each genre. \item For each event artist, find the most similar artists from the set of popular artists. \end{enumerate} In Section \ref{sec:artist-similarity}, we will describe how we use harvested tags and artist similarity information to compute similarity between pairs of artists, as well as between artists and tags. These similarities are represented as real-valued weights, and as such, the event graph contains weighted edges. Based on the interactive design described above, we can efficiently recommend events using a Music Event Graph. The user selects one or more preferred genres and then a set of relevant popular artists. Next our algorithm selects the event artists and their related events that are connected to the user's selected genres and popular artists. This graph traversal algorithm is depicted in Figure \ref{fig:event-graph-recommendation}. \begin{figure} \centering \includegraphics[width=3.3in]{EventGraphRecommendation.png} \caption{Construction the Music Event Graph: a user selects genre tags $\{t_1,t_3\}$. She is then shown popular artists $\{pa_1, pa_2, pa_5, pa_6\}$ and selects $\{pa_2, pa_6\}$. We then use the graph to strongly recommend event $e_4$ with artists $ea_3$ and $ea_5$ based on multiple connections. We would also recommend events $e_1$ and $e_2$ based on their connections through $ea_1$ and $ea_3$, respectively. } \label{fig:event-graph-recommendation} \end{figure} We note that our algorithm uses weighted edges to compute a user-specific relevance scores for each event as we move from left to right in the graph structure. In addition, we can use the graph structure to provide \emph{recommendation transparency} \cite{sinha2002role} by keeping track of the paths that are used to get from the user genre and popular artist selections to the recommend event artists and events. \section{Artist Similarity and Tag Affinity}\label{sec:artist-similarity} At the core of the event recommendation system, we use Latent Semantic Analysis (LSA) when calculating artist similarity and artist-tag affinity. That is, we use truncated single value decomposition (SVD) to transform a large, sparse data matrix of artist similarity and tag information into a lower dimensional matrix such that each artist and tag is embedded into a dense, k-dimensional \emph{latent} feature space. Note that $k$ is a hyperparameter that is set based on empirical evaluation. We can then calculate artist-artist or artist-tag similarity using the cosine distance between pairs of vectors in this latent space. Before we describe LSA, we will start with some useful notation for our problem setup: \begin{itemize} \item [$\mathcal{A}$]: set of artists. \item [$\mathcal{T}$]: set of tags. Tags are any free text token that can be used to describe music. This may include genres, emotions, instruments, usages, etc. \item [$\mathcal{T}^G$]: a small subset of \emph{genre} tags (e.g., ``rock'', ``country'', ``blues'') that are frequently used to categorize music. \item [$\mathcal{A}^P$]: set of \emph{popular} artists where each artist in the set is none of the most recognizable artists associated with at least one of the genre tags $\mathcal{T}^G$. \item [$\mathcal{E}$]: set of local music events \item [$\mathcal{A}^{\mathcal{E}}$]: set of \emph{event} artists where each artist has one or more upcoming events in $\mathcal{E}$ \item [$\mathcal{F}$]: set of features where $\mathcal{F}= \mathcal{A} \cup \mathcal{T}$. That is, we will describe each artist $a$ as a (sparse) feature vector of artist similarity and tag affinity values in $\mathbb{R}^{|\mathcal{A}|+|\mathcal{T}|}$ \item [$X$]: (sparse) raw data matrix. The dimension of $X \in \mathbb{R}^{|\mathcal{A}| \times |\mathcal{F}|}$ where the $x_{i,j} \in [0,1]$ represents the affinity between the $i$-th artist (a row) and $j$-th feature (column). A value of 0 represents either no affinity or unknown affinity. Note that all artists are self-similar so that $x_{i,i} = 1$. In terms of practical implementation, we can construct $X$ by stacking our $|\mathcal{A}| \times |\mathcal{A}|$ artist similarity matrix next to our$ |\mathcal{A}| \times |\mathcal{T}|$ artist-tag affinity matrix. \end{itemize} LSA uses the truncated SVD algorithm to decompose the raw data matrix $X$ as follows: \begin{equation} X \approx X_k = U_k \Sigma_k V_k^{T} \end{equation} such that the matrix $X_k$ is a rank-$k$ approximation of $X$, $U_k$ is an $|\mathcal{A}| \times k$ matrix, $\Sigma_k$ is a diagonal $k \times k$ matrix of singular values, and $V_k^{T}$ is a $k \times |\mathcal{F}|$ matrix. We will then project each artist and tag in a $k$ dimensional latent feature space: \begin{equation} X_{SVD} = \Sigma_k V_k^{T} \end{equation} where $X_{SVD} \in \mathbb{R}^{k \times |\mathcal{F}|}$ or equivalently $\mathbb{R}^{k \times (|\mathcal{A}|+|\mathcal{T}|)}$ by construction. That is, the first $|\mathcal{A}|$ columns of $X_{SVD}$ represent artists and the last $|\mathcal{T}|$ columns represent tags all embedded into the same $k$ dimensional space. We can also embed a new artist with raw feature vector $\mathbf{x} \in \mathbb{R}^{1 \times |\mathcal{F}|}$ by computing \begin{equation} \mathbf{x}_{SVD} = \mathbf{x} V_k \Sigma_k^{-1}. \end{equation} so that $\mathbf{x}_{SVD} \in \mathbb{R}^{1 \times k}$ is projected in the same latent feature space. Finally, we can compute artist-artist, artist-tag, or tag-tag similarity in the embedded space by comparing their respective (column) vectors in $X_{SVD}$. For example, if we have two latent feature vectors $p$ and $q$, we can compute their cosine similarity: \begin{equation} \cos(p, q) = \frac{p \cdot q}{||p||~||q||} \end{equation} where $p$ and $q$ are $k$-dimensional vectors and $||x|| = \sqrt[]{\sum_{i=1}^{k} x_i^2}$ is the l2-norm of a vector $x$. One nice property of cosine similarity, is that it tends to remove popularity bias. That is, we normalize the feature vectors by their length (l2-norm) such that each artist (and tag) vector is the same length. Without length normalization, popular artists which tend to have a bigger digital footprint (resulting in a denser raw feature vector with a bigger l2-norm) tend to produce larger similarity scores on average than if we did not normalize by length. \section{Event and Artist Data}\label{sec:data} The data for our experiments is constructed by scraping local events from both TicketFly\footnote{https://www.ticketfly.com scraped February 15, 2018.} and the web-based public event calendar from a local newspaper\footnote{Details omitted during anonymous review process.}. We collected a total of $|\mathcal{E}| = 96$ events with 66 events from TicketFly, 36 events from the local newspaper, and 6 overlapping events between both websites. These events produced a set $|\mathcal{A}^{\mathcal{E}}| = 154$ event artists. We are also able to download short biographies of almost all of the event artists for events obtained from Ticketfly. The local newspaper only provides us with 1 to 3 genre tags for about half of the events we obtained from their site. We then used the Last.fm API \footnote{https://www.last.fm/api} to collect music information (popularity, biography text, artist similarity scores, and tags affinity scores) for each of our event artists. We then use snowball sampling on the similar artists and obtain this same Last.fm music information. We continue sampling these non-event artists until we have a set of 10,000 artists (i.e., $|\mathcal{A}| = 10,000$.) We define our set of tags $\mathcal{T}$ as the 1585 tags which are associated with 20 or more artists. Our set of genre tags $\mathcal{T}^{\mathcal{G}}$ are the top 20 tags which are most frequently associated with our event artists $\mathcal{A}^{\mathcal{E}}$. These include tags like ``rock'', ``jazz'', and ``reggae''. However, we manually prune tags which are obviously not genres like ``seen live'' and ``favorites''. Finally, for each artist, we concatenate all available biographies (Last.fm, TicketFly, local newspaper) and attempt to find each of our tags in the combined biography text. If a tag is found, we label the artist with that tag. This is especially important since otherwise, many of our event artists would not be labeled with any tags. In the end, we have 977,270 artist similarities and 456,867 artist-tag affinities. \section{Exploring Artist Similarity in the Long Tail} The core of our local event recommendation algorithm is our artist similarity calculation based on Latent Semantic Analysis (LSA). In this section, we show that most local event artists are relatively obscure \emph{long-tail} artists and that they tend to have small digital footprints. We also explore the relationship between digital footprint size and the accuracy of our artist similarity calculation. \subsection{Long-tail Event Artists} \begin{figure} \centering \includegraphics[width=3.3in]{LongTail.png} \caption{(Top) Plot of Last.fm Listener Count vs. popularity rank (divided by total number of artists) of 10,000 artists with most. (Bottom) Histogram of popularity for 154 event artists placed into 10 deciles of overall artist popularity. Note that most event artists reside in the long-tail of this popularity distribution.} \label{fig:long-tail} \end{figure} In the top plot of Figure \ref{fig:long-tail}, we rank all 10,000 of our artists by their Last.fm listener counts. This shows a typical long-tail (power-law) distribution where a small number of popular artists in the short-head (left) receive much more attention than the vast majority of other artists in the long tail (right) \cite{celma2010,anderson2004long}. For example, 16.3\% of the most popular artists represent 80\% of the listener counts. In the bottom plot, we show a histogram of the event artists' Last.fm listener counts broken down into deciles. We note that a disproportionate number of local event artists reside in the long-tale of this popularity distribution. In particular, 99 of the 154 event artists (64.2\%) are in the lowest three deciles of the ranking. \subsection{The Digital Footprint of Event Artists } \begin{figure} \centering \includegraphics[width=3.3in]{FootprintDistribution.png} \caption{Cumulative distributions of digital footprint size (i.e., the number of nonzero artist similarity and tag affinity scores for each artists) for event artists and all artists. } \label{fig:footprint-dist} \end{figure} As we discussed in the Introduction, obscure artists tend to have small digital footprints. To show this, we will consider the digital footprint of an artist to be the number of artist similarities plus the number of tag affinities for that artist. Equivalently, it is the number of nonzero values in the row of our raw data matrix $X$ that is associated with the artist. We note that digital footprint size is correlated with popularity rank ($r = -0.56$) such that popular artists tend to have a larger digital footprint. In Figure \ref{fig:footprint-dist}, we plot the empirical cumulative distribution for both event artists and all artists as a function of the digital footprint size. We see that about 27.2\% of the event artists have 15 or fewer digital footprints whereas only 2.8\% of all artists have so few digital footprints. This suggests that it will be important for us to design an artist similarity algorithm that works well in this \emph{small digital footprint} setting. \subsection{Artist Similarity with Latent Semantic Analysis} \begin{figure} \centering \includegraphics[width=3.3in]{ReducedDigitalFootprint.png} \caption{Plot of artist similarity ranking performance as a function of the (artificially reduced) digital footprint size for various LSA embeddings with ranks of 32, 64, 128, and 256 dimensions. The \emph{Raw} approach represents computing cosine distances without first applying LSA to the raw artist data vectors. } \label{fig:reduced-footprint} \end{figure} In Section \ref{sec:artist-similarity}, we introduced LSA as a algorithm for computing artist similarity. However, as we observed in the previous subsection, we are particularly interested in the case where an artist is represented by a small number of artist similarities and tag affinities (i.e., a small digital footprint.) To explore this, we will artificially reduce the digital footprint of artists to a fixed sized and see how well LSA is able to accurately compute artist similarity. To do this, we randomly split our data set of $\mathcal{A} = 10,000$ artists into a training set with $\mathcal{A}_{train} = 9,000$ artists and a test set of $\mathcal{A}_{test} = 1,000$ artists. Note that this involves removing 1000 rows \emph{and} 1000 columns from our raw data matrix $X$ since artists are also features. The training data will be used to to calculate our matrix decompositions $\mathcal{A}_{train} \approx U_k \Sigma_k V_k^{T}$ for a given embedding dimension $k$. Before projecting the $\mathcal{A}_{test}$ into the latent feature space, we limit the digital footprint size of each artists by randomly selecting artist similarity and tag affinity features to zero out. We can then project $\mathcal{A}_{test}$ into the latent feature space and calculate the cosine distance between each pair of test set artists. Finally, we can calculate the Area Under the ROC Curve (AUC) \cite{manning2008} for each artist where the original artist similarities serve as the ground truth. Figure \ref{fig:reduced-footprint} shows a plot of artificially reduced digital footprint size verses average AUC over the 1,000 test set artists for various LSA embedding dimensions. We also plot the curve for when we compute cosine distances between the \emph{raw} test artist vectors without projecting into a latent feature space. Here we note that LSA shows a improvement over raw cosine distance in small footprint setting of between 1 and 16 nonzero features. Once the digital footprint is larger than 128 nonzero features, the raw cosine approach slightly out performs LSA-based approach. However, the compactness of representing each artist with 32 or 64 floating point numbers may be advantageous in terms of storage size and computation time when we consider a much larger set of artists and tags. As such, we will use 64-dimensional LSA embeddings for the remaining experiments in this paper. \section{Exploring Event Recommendation} To explore the performance of event recommendation using event graphs and LSA-based artist similarity, we conducted a small user study with a short 2-phase survey. We recruited 51 participants who were very familiar with the local music scene and attend live events in the area on a weekly basis. In the first phase of our survey, we asked participants to select between 1 and 3 genres from a set of 20 common genres. For each selected genre, the test subject was then asked to select between 1 and 3 artists from a set of 16 popular artists that were representative of the genre (i.e., having a high cosine similarity score between the 64-dimensional latent feature vectors of the genre and the artist.) In the second stage, participants were shown a list of the 154 event artists in our data set. They were asked to select all artists that they would like to see at a live event in the local area and were required to select 5 or more event artists. To evaluate our system, we use each test subject's selected genres or popular artists from phase 1 of the survey to rank order the 154 event artists using one of the approaches described below. In all cases, we embedded artists and tags into a 64-dimensional latent feature space using LSA with the data set that is described in Section \ref{sec:data}. We then calculate the area under the ROC curve (AUC) for each user where ground truth relevance is determined from phase 2 of the survey. Each test subject provides multiple genre and multiple popular artist preferences. We explore a number of ways to combine these preferences to produce one ranking of the event artists for each test subject. We consider \emph{early fusion} and \emph{late fusion} steps for a number of approaches. In early fusion, we start with a set of latent feature vectors where each vector is associated with one of the users genre or artist preferences. We consider three approaches: \begin{itemize} \item \textbf{average} the latent feature vectors into one vector \item \textbf{cluster} the latent feature vectors and use the $k$ centroid vectors \item \textbf{none} use all of the latent feature vectors \end{itemize} When clustering, we use the $k$-means clustering algorithm\footnote{http://scikit-learn.org/stable/modules/generated/ sklearn.cluster.KMeans.html} with the number of clusters ($k$) equal to the rounded natural log of the number of user preferences. For \emph{late fusion}, we must output one ranking of the event artists for each user. We consider three approaches \begin{itemize} \item \textbf{average cosine} ranks event artists by the average of the cosine similarity scores between the event artist vector and each vector in the set of user preference vectors. \item \textbf{average rank} creates one ranking of event artists for each user preference vector, calculates the average rank for each event artist over this set of rankings, and then ranks them by this average rank. \item \textbf{interleave} creates a set of rankings of the event artists for each user preference vector, and then constructs a final ranking by alternating between these ranking lists and picking top remaining artists that have not already been added to the final ranking. \end{itemize} \begin{table}[] \centering \caption{Event artist recommendation performance. The mean and standard deviation of AUC for our 8 expert test subjects when considering popular artist preferences, genre preferences, and both preferences together. See text for details on the seven approaches and the two baselines. } \vspace{3mm} \label{tab:event-artist-recommend} \begin{tabular}{|l|c|c|c|} \hline \multicolumn{1}{|c|}{Approach} & \multicolumn{3}{|c|}{ User Preferences } \\ \hline \multicolumn{1}{|c|}{Early / Late Fusion} & Artists & Genres & Both \\ \hline none / avg. cosine & \textbf{.79} (.09) & .69 (.16) & .74 (.12) \\ none / avg. rank & .75 (.11) & .66 (.15) & .76 (.09) \\ none / interleave & \textbf{.79} (.11) & .69 (.15) & .71 (.15) \\ average / cosine & \textbf{.79} (.09) & .69 (.16) & .74 (.12) \\ cluster / avg. cosine & .78 (.09) & .69 (.18) & .74 (.11) \\ cluster / avg. rank & .74 (.13) & .66 (.20) & .68 (.17) \\ cluster / interleave & .78 (.10) & .69 (.16) & .75 (.09) \\ \hline \multicolumn{1}{|c|}{Baseline} & \multicolumn{3}{|c|}{~} \\ \hline random & \multicolumn{3}{|c|}{.50 (.12)}\\ popularity & \multicolumn{3}{|c|}{.53 (-)}\\ \hline \end{tabular} \end{table} Table \ref{tab:event-artist-recommend} shows average AUCs (and standard deviations) for our seven early/late fusion approaches when we use each user's popular artist preferences, genre tag preferences, and both sets of preferences together. We also include a popularity baseline that ranks all event artists by their Last.fm listener count as well as a random shuffle baseline. We observe that artist preferences alone result in the best performance and a number of our proposed early/late fusion approaches produce similar results. We should also mention that we collected survey data from individuals who attended local shows on a less frequent (monthly) basis. The results for these test subjects was significantly lower (average AUC of 0.61) and more variable (AUC standard deviation of 0.15) for our best performing approach (Genre Preferences / None / Interleave.) Having done error analysis on many of these less regular attendees, we often found that they selected a very eclectic set of event artists which did not match their preferences. As such, it would have been difficult for any recommender system to make accurate recommendations for many of these test subjects. This suggests that test subjects need to have a high level of familiarity with the local music community in order to provide useful ground truth for our experiment. \section{Discussion} In this paper, we explored the understudied task of local music event recommendation. This is an exciting task for the research community because it involves many interesting problems: long-tail recommendation, the new user \& new artist cold start problems, multiple types of music information (artist similarity, tags), and user preference modeling. It is also an interesting problem outside of the academic research community since music event recommender systems can be used to help grow and support the local arts community. By promoting the work of talented local musicians, such systems can help fans discover new artists and help musicians reach new audiences. These audiences in turn attend more events which help sustain concert venues, music festivals, and other (local) businesses who benefit from direct ticket sales and other forms of indirect support (e.g., food, drinks, merchandise.) While we were able to evaluate our system using a survey of local music experts, a more natural way to evaluate music event recommendation would be to build an interactive application thats collects user feedback over a longer period of time. We plan to develop such an app in the coming months and hope that it will be useful for expanding on the research that is presented in this paper.

10c1d4f0c9e2f2676bdc637fe722462eed43e5f6

\section{Introduction} During recent years, multicopter Unmanned Aerial Vehicles (UAVs) are becoming increasingly popular in both civil and military fields \cite{doyle2013avian,tauro2015large} including aerial photography, plant protection, package delivery and other fields. Limited by the battery technology, the flight time (hovering endurance) of multicopters is still too short for most applications. Since the performance and efficiency of a multicopter directly depend on the propulsion system, the design optimization for multicopter propulsion systems is urgently needed to increase the flight time. The design optimization problem studied in this paper is to find the optimal combination of propulsion system components to satisfy the given hovering endurance requirement, and the obtained propulsion system should have smaller weight and higher efficiency as possible. A typical multicopter propulsion system usually consists of four basic components including the propeller, BrushLess Direct-Current (BLDC) motor, Electronic Speed Control (ESC) and Lithium Polymer (LiPo) battery \cite{quan2017introduction}. Traditional methods to determine a propulsion system are usually based on the experience and trial-and-error experiments. Considering that there are thousands of component products on the market, it is a costly and time-consuming work for the traditional design methods. Meanwhile, in the whole design process of a multicopter system, the propulsion system need to be repeatedly modified according to the actual controlled system until all the performance requirements and safety requirements are satisfied. According to \cite{oktay2013simultaneous,oktay2016simultaneous}, a more efficient way is to simultaneously design the body system (including the propulsion system) and the control system subject to the optimal objective and additional constraints. Therefore, the automatic design and optimization technologies for propulsion systems will be beneficial for reducing the prototyping needs for the whole multicopter system, and minimizing development and manufacturing cost. For such reasons, this paper proposes a simple, practical and automatic design optimization method to help designers quickly find the optimal propulsion system according to the given design requirements. In our previous work \cite{Shi2017}, based on the mathematical modeling methods for the components of propulsion systems, a practical method is proposed to estimate the flight performance of multicopters according to the given propulsion system parameters. In fact, the study in this paper is the reverse process of our previous work, namely estimating the optimal propulsion system parameters according to the given design requirements. This problem is more complicated and difficult because the number of design requirements is much less than the number (more than 15) of the propulsion system parameters. There are many studies on the mathematical modeling \cite{Harrington2011,Ramana2013,Mccrink2015}, the efficiency analysis \cite{Lawrence2005,Stepaniak2009}, and the performance estimation \cite{Shi2017,Bershadsky2016a} of multicopter propulsion systems. To our best knowledge, there are few studies on the design optimization of multicopter propulsion systems. Most of them adopt numerical methods (fixed-wing aircraft \cite{Lundstrom2009} and multicopters \cite{Bershadsky2016a}) to search and traverse all the possible propulsion system combinations in the database based on the proposed cost functions. In \cite{Magnussen2014,Magnussen2015}, the multicopter optimization problem is described as a mixed integer linear program, and solved with the Cplex optimizer. However, these numerical methods have following problems: i) a large and well-covered product database is required for a better optimization effect; ii) the calculation speed is slow when there are large numbers of products in the database because the amount of product combinations is huge (the algorithm complexity is $O(n^{4})$, where $n$ is the number of the database products). In order to solve the above problems, this paper proposes an analytical method to estimate the optimal parameters of the propulsion system components. First, the modeling methods for each component of the propulsion system are studied respectively to describe the problem with mathematical expressions. Secondly, the whole problem is simplified and decoupled into several small problems. By solving these sub-problems, the optimal parameters of each component can be obtained respectively. Finally, based on the obtained parameters, selection algorithms are proposed to determine the optimal combination of the propeller, the motor, the ESC and the battery products from their corresponding databases. The contributions of this paper are as follows: i) an analytical method to solve the design optimization problem of multicopter propulsion systems is proposed for the first time; ii) the conclusion obtained through the theoretical analysis has a guiding significance for the multicopter design; iii) compared to the numerical traversal methods, the proposed method reduces the algorithm complexity from $O(n^{4})$ to $O(n)$, which is faster and more efficient for practical applications. The paper is organized as follows. \textit{Section\,\ref{Sec-2}} gives a comprehensive analysis of the design optimization problem to divide it into twelve sub-problems. In \textit{Section\,\ref{Sec-3}}, the modeling methods for each component of the propulsion system are studied the describe the sub-problems with mathematical expressions. In \textit{Section\,\ref{Sec-4}}, the sub-problems are solved respectively to obtain the optimal components of the desired propulsion system. In \textit{Section\,\ref{Sec-6}}, statistical analyses and experiments are performed to verify the proposed method. In the end, \textit{Section\,\ref{Sec-7}} presents the conclusions. \section{Problem Formulation} \label{Sec-2} \subsection{Design Requirements of Propulsion Systems} \label{Sec2-Des} The role of a propulsion system is to continuously generate the desired thrust within the desired time of endurance for a multicopter. The design requirements for a propulsion system are usually described by the following parameters: i) the number of the propulsion units $n_{\text{p}}$; ii) the hovering thrust of a single propeller $T_{\text{hover}}$ (unit: N) under the hovering mode when the multicopter stays fixed in the air; iii) the maximum thrust of a single propeller $T_{\text{max}}$ (unit: N) under the full-throttle mode when the autopilot gives the maximum throttle signal; iv) the nominal flight altitude $h_{\text{hover}}$ (unit: m); iv) the flight time $t_{\text{hover}}$ (unit: min) under the hovering mode. This paper only focuses on studying the design optimization of propulsion systems with assuming $T_{\text{hover}}$ and $T_{\text{max}}$ are known parameters. Although the propulsion system parameters $T_{\text{hover}}$ and $T_{\text{max}}$ are usually not directly available, according to our previous research \cite{Shi2017}, $T_{\text{hover}},T_{\text{max}}$ can be obtained by giving the aerodynamic coefficients, airframe parameters and the kinematic performance requirements. For example, for common multicopters, the hovering thrust $T_{\text{hover}}$ can be obtained by the total weight of the multicopter $G_{\text{total}}$ (unit: N) as \begin{equation} T_{\text{hover}}=\frac{G_{\text{total}}}{n_{\text{p}}}.\label{eq:T0} \end{equation} The kinematic performance of a multicopter is directly determined by the thrust ratio $\gamma\in\left(0,1\right)$ which is defined as \begin{equation} \gamma\triangleq\frac{T_{\text{hover}}}{T_{\text{max}}}.\label{Eq31T0TpmaxThr} \end{equation} where thrust ratio $\gamma$ describes the remaining thrust for the acceleration movement, which further determines the maximum forward speed and the wind resistance ability of a multicopter. Therefore, designers should estimate the desired $\gamma$ (usually $\gamma=0.5$ is selected for common multicopters) according to the kinematic performance requirements of the multicopter. Then, the desired full-throttle thrust $T_{\text{max}}$ of the propulsion system can be obtained through Eqs.\,(\ref{eq:T0})(\ref{Eq31T0TpmaxThr}). \subsection{Component Parameters} The ultimate goal of the design optimization problem is to select the optimal products from four component databases with component parameters listed in Table\,\ref{tab1}, where $\Theta_{\text{p}}$, $\Theta_{\text{m}}$, $\Theta_{\text{e}}$ and $\Theta_{\text{b}}$ represent the parameter sets for propellers, motors, ESCs and batteries. In order to ensure commonality of the method, all component parameters in Table\,\ref{tab1} are the basic parameters that can be easily found in the product description pages. The detailed introduction of each parameter in Table\,\ref{tab1} can also be found in \cite[pp. 31-46]{quan2017introduction}. \begin{table}[ptbh] \caption{Propulsion system parameters} \label{tab1}\centerin \begin{tabular}{|c|>{\raggedright}p{0.38\textwidth}|} \hline Items & Parameters \tabularnewline \hline Propeller & $\Theta_{\text{p}}\triangleq$\{Diameter $D_{\text{p}}$ (m), Pitch Angle $\varphi_{\text{p}}$ (rad), Blade Number $B_{\text{p}}$\}\tabularnewline \hline Motor & $\Theta_{\text{m}}\triangleq$\{Nominal Maximum Voltage $U_{\text{mMax}}$ (V), Nominal Maximum Current $I_{\text{mMax}}$ (A), KV Value $K_{\text{V}}$ (RPM/V), No-load Current $I_{\text{m0}}$ (A), Resistance $R_{\text{m}}$ ($\Omega$)\}\tabularnewline \hline ESC & $\Theta_{\text{e}}\triangleq$\{Nominal Maximum Voltage $U_{\text{eMax}}$ (V), Nominal Maximum Current $I_{\text{eMax}}$ (A), Resistance $R_{\text{e}}$ ($\Omega$)\}; \tabularnewline \hline Battery & $\Theta_{\text{b}}\triangleq$\{Nominal Voltage $U_{\text{b}}$ (V), Maximum Discharge Rate $K_{\text{b}}$ (A), Capacity $C_{\text{b}}$ (mAh), Resistance $R_{\text{b}}$ ($\Omega$)\}; \tabularnewline \hline \end{tabular} \end{table} The propeller pitch angle $\varphi_{\text{p}}$ (unit: rad) in Table\,\ref{tab1} is defined according to the propeller diameter ${D_{\text{{p}}}}$ (unit: m) and the propeller pitch ${H_{\text{{p}}}}$ (unit: m) as \begin{equation} \varphi_{\text{p}}\triangleq{\arctan\frac{{H_{\text{{p}}}}}{{\pi{D_{\text{{p}}}}}}}\label{Eq00-pitch} \end{equation} where ${H_{\text{{p}}}}$ and ${D_{\text{{p}}}}$ are usually contained in the propeller model name. The most commonly used unit for the battery voltage $U_{\text{b}}$, as well as the motor voltage $U_{\text{mMax}}$ and the ESC voltage $U_{\text{eMax}}$, is ``S\textquotedblright , which denotes the number of battery cells in series. For LiPo batteries, the voltage changes from 4.2\,V to 3.7\,V as the battery capacity decreases from full to empty, and the average voltage 4.0\,V is adopted for the unit conversion. For example, $U_{\text{b}}=12\text{\,S}=48\text{\,V}.$ \subsection{Optimization Constraints} \label{subsec:OptiConstr} \subsubsection{Requirement Constraints} According to \cite{Shi2017}, $t_{\text{hover}}$ and $T_{\text{max}}$ can be estimated by parameters $n_{\text{p}},T_{\text{hover}}$,$\Theta_{\text{p}},\Theta_{\text{m}},\Theta_{\text{e}},\Theta_{\text{b}}$. Therefore, two equality constraints can be obtained according to the design requirements in \textit{Section\,\ref{Sec2-Des}}. \subsubsection{Safety Constraints} The electric components should work within their allowed operating conditions to prevent from being burnt out. Therefore, a series of inequality constraints can be obtained for electric components including the battery, the ESC and the motor of the propulsion system. The detailed inequality expressions will be introduced in the later sections. \subsubsection{Product Statistical Constraints} In order to make sure that the obtained solution is meaningful and practical, the product statistical features should be considered. Otherwise, it is very possible that there is no product in reality matching with the obtained component parameters. The product features can be described by equality constraints based on the statistical models of the products in the database. In practice, products with different material, operating principle and processing technology may have different product features, for example, LiPo batteries and Ni-MH batteries. Therefore, different statistical models should be obtained for different types of products, and designers should select the required product type to the following optimization process to improve the precision and practicability of the obtained result. \subsection{Optimization Problem} In practice, there are two methods to increase the flight time of a multicopter: i) decrease the total weight of the multicopter to allocate more free weight for the battery capacity; ii) increase the efficiency of the multicopter propulsion system to decrease the required battery capacity. For a typical multicopter, according to the weight statistical model in \cite{Bershadsky2016a}, the propulsion system weight is the main source (usually more than 70\%) of the multicopter weight, and the weight of the battery is further the main source (usually more than 60\%) of the propulsion system weight. In fact, for a propulsion system, the maximum efficiency usually means the minimum battery weight (capacity). Therefore, the above two methods essentially share the same optimization objective, namely minimizing the weight of the propulsion system. Assuming that the total weight of the propulsion system is defined as $G_{\text{sys}}$ (unit: N), the optimization objective of the design optimization can be described as \begin{equation} \min_{\Theta_{\text{p}},\Theta_{\text{m}},\Theta_{\text{e}},\Theta_{\text{b}}}G_{\text{sys}}.\label{eq:OptiEqua} \end{equation} Along with the constraints in \textsl{Section\,\ref{subsec:OptiConstr}}, the optimal solutions for the parameters of each component can be obtained from Eq.\,(\ref{eq:OptiEqua}). \subsection{Problem Decomposition} \label{subsec:ProbDeco} According to \cite{Shi2017}, the detailed mathematical expressions for Eq.\,(\ref{eq:OptiEqua}) are very complex because there are 15 parameters listed in Table\,\ref{tab1} that need to be optimized through solving complex nonlinear equations. As a result, decomposition and simplification are required to solve this problem. \subsubsection{Weight Decomposition} The total weight of the propulsion system $G_{\text{sys}}$ is determined by the weight of each component as \begin{equation} G_{\text{sys}}=n_{\text{p}}\left(G_{\text{p}}+G_{\text{m}}+G_{\text{e}}\right)+G_{\text{b}}\label{eq:Gsys} \end{equation} where $G_{\text{p}},G_{\text{m}},G_{\text{e}},G_{\text{b}}$ (unit: N) denote the weight of the propeller, the motor, the ESC and the battery respectively. Therefore, based on the idea of the greedy algorithm, the optimization objective of minimizing $G_{\text{sys}}$ can be decomposed into four sub-problems of minimizing the weight of each component as \begin{equation} \min G_{\text{sys}}\Rightarrow\min G_{\text{p}},\min G_{\text{m}},\min G_{\text{e}},\min G_{\text{b}}.\label{eq:WeightDecom} \end{equation} As mentioned above, the battery weight $G_{\text{b}}$ is the most important factor for $G_{\text{sys}}$, and $G_{\text{b}}$ directly depends on the battery capacity $C_{\text{b}}$. Since, according to the analysis in \cite{quan2017introduction}, the battery capacity $C_{\text{b}}$ depends on the efficiency of each component, the optimization objective of minimizing $G_{\text{b}}$ can also be decomposed into four sub-problems of maximizing the efficiency of each component as \begin{equation} \min G_{\text{b}}\Rightarrow\max\eta_{\text{p}},\max\eta_{\text{m}},\max\eta_{\text{e}},\max\eta_{\text{b}}.\label{eq:EffDecom} \end{equation} where $\eta_{\text{p}},\eta_{\text{m}},\eta_{\text{e}},\eta_{\text{b}}$ denote the efficiency of the propeller, the motor, the ESC and the battery respectively. \begin{figure}[tbh] \centering \includegraphics[width=0.4\textwidth]{Fig04} \caption{Optimization objective decomposition diagram.} \label{Fig02Decom} \end{figure} Thus, by solving the eight sub-problems in Eqs.\,(\ref{eq:WeightDecom})(\ref{eq:EffDecom}), the optimal solutions for the parameters in Table\,\ref{tab1} can be obtained with the results presented in Fig.\,\ref{Fig02Decom}, where the obtained optimal solutions are marked with a subscript ``Opt''. Then, the optimal parameter sets for the propeller, the motor, the ESC and the battery can be determined and represented by $\Theta_{\text{pOpt}}$,$\Theta_{\text{mOpt}}$, $\Theta_{\text{eOpt}}$, and $\Theta_{\text{bOpt}}$ respectively. \subsubsection{Product Selection Decomposition} The ultimate goal of the design optimization of the propulsion system is to determine the optimal propeller, motor, ESC and propeller products from their corresponding databases according to the obtained optimal parameter sets $\Theta_{\text{pOpt}},\Theta_{\text{mOpt}},\Theta_{\text{eOpt}},\Theta_{\text{bOpt}}$. This problem can also be divided into four sub-problems. Through solving the four sub-problems, the parameter sets of the obtained products are represented by $\Theta_{\text{pOpt}}^{\ast}$, $\Theta_{\text{pOpt}}^{\ast}$, $\Theta_{\text{pOpt}}^{\ast}$ and $\Theta_{\text{pOpt}}^{\ast}$. \subsection{Solving Procedures} \label{Sec2-End} Through the above decomposition procedures, the whole design optimization problem can finally be simplified, decoupled and divided into twelve sub-problems. However, since there are argument-dependent relationships among the sub-problems, the solving sequence should be well-arranged. For instance, the propeller parameters $B_{\text{p}},\varphi_{\text{p}}$ are required for the motor optimization, and the optimal propeller diameter $D_{\text{p}}$ depends on the obtained motor parameters. As a result, the propeller and motor should be treated as a whole during the solving procedures. In this paper, the twelve sub-problems will be solved separately by twelve steps with the solving sequence shown in Fig.\,\ref{Fig02-0}. Fig.\,\ref{Fig02-0} also presents the inputs, outputs, and parameter dependency relationships of each sub-problem. The detailed solving methods of each step will be introduced in \textit{Section\,\ref{Sec-4}} with twelve subsections. \begin{figure}[tbh] \centering \includegraphics[width=0.4\textwidth]{Fig02-0} \caption{Solving procedures of the propulsion system optimization problem} \label{Fig02-0} \end{figure} \section{Propulsion System Modeling} \label{Sec-3} The whole propulsion system can be modeled by the equivalent circuit \cite{Shi2017} as shown in Fig.\,\ref{Fig02}, with which the optimization sub-problems can be described by mathematical expressions. \begin{figure}[tbh] \centering \includegraphics[width=0.45\textwidth]{Fig02} \caption{Equivalent circuit model of the whole propulsion system} \label{Fig02} \end{figure} \subsection{Propeller Modeling} \subsubsection{Propeller Aerodynamic Model} According to \cite{merchant2006propeller}, the thrust force $T$ (unit: N) and torque $M$ (unit: N$\cdot$m) of fixed-pitch propellers can be obtained through equations \begin{equation} \left\{ \begin{array}{c} T=C_{\text{T}}\rho\left(\frac{N}{60}\right)^{2}D_{\text{p}}^{4}\\ M=C_{\text{M}}\rho\left(\frac{N}{60}\right)^{2}D_{\text{p}}^{5} \end{array}\right.\label{Eq02PropTorque} \end{equation} where $N$ (unit: RPM) is the propeller revolving speed, $\rho$ is the local air density (unit: kg/m$^{3}$), $C_{\text{T}}$ is the propeller thrust coefficient, $C_{\text{M}}$ is the propeller torque coefficient, and $D_{\text{p}}$ (unit: m, from ${\Theta_{\text{p}}}$) is the propeller diameter. The air density $\rho$ is determined by the local temperature $T_{\text{t}}$ (unit: $^{\circ}$C) and the air pressure which is further determined by the altitude $h_{\text{hover}}$ (unit: m). According to the international standard atmosphere model \cite{cavcar2000international} \begin{equation} \rho=f_{\rho}\left(h_{\text{hover}}\right)\triangleq\frac{273}{(273+T_{\text{t}})}(1-0.0065\frac{h_{\text{hover}}}{273+T_{\text{t}}})^{5.2561}\rho_{\text{0}}\label{Eq03AirDen} \end{equation} where the standard air density $\rho_{\text{0}}=1.293$kg/m$^{3}$ ($^{\circ}$C, 273K). The propeller coefficients $C_{\text{T}}$ and $C_{\text{M}}$ can be modeled by using the blade element theory as presented in \cite{Shi2017}\cite{Mccrink2015}. A simplified form is introduced here as \begin{equation} \left\{ \begin{array}{lll} C_{\text{T}} & =f_{C_{\text{T}}}\left(B_{\text{p}},D_{\text{p}},\varphi_{\text{p}}\right) & \triangleq k_{\text{t0}}B_{\text{p}}\varphi_{\text{p}}\\ C_{\text{M}} & =f_{C_{\text{M}}}\left(B_{\text{p}},D_{\text{p}},\varphi_{\text{p}}\right) & \triangleq k_{\text{m0}}B_{\text{p}}^{2}\left(k_{\text{m1}}+k_{\text{m2}}\varphi_{\text{p}}^{2}\right) \end{array}\right.\label{Eq04Cm} \end{equation} where $k_{\text{t0}},k_{\text{m0}},k_{\text{m1}},k_{\text{m2}}$ are constant parameters determined by the shapes and aerodynamic characteristics of the propeller blades, and they can be obtained through the propeller model in \cite{Shi2017} as \begin{equation} \begin{array}{llcl} k_{\text{t0}} & \triangleq\frac{0.25{\pi^{3}}\lambda{\zeta^{2}}{K_{0}}\varepsilon}{{\pi A+{K_{0}}}}, & k_{\text{m0}} & \triangleq\frac{1}{{8A}}{\pi^{2}}\lambda{\zeta^{2}}\\ k_{\text{m1}} & \triangleq C_{{\rm {fd}}}, & k_{\text{m2}} & \triangleq\frac{\pi AK_{0}^{2}\varepsilon}{e{{\left({\pi A+{K_{0}}}\right)}^{2}}} \end{array}\label{eq:CTCMParamete} \end{equation} where the detailed definitions of the internal parameters of Eq.\,(\ref{eq:CTCMParamete}) can be found in \cite{Shi2017}. Note that $k_{\text{t0}},k_{\text{m0}},k_{\text{m1}},k_{\text{m2}}$ may slightly vary with the difference of types, material, and technology of propellers. Based on the propeller data from T-MOTOR website \cite{TMotor2017}, general parameters $k_{\text{t0}},k_{\text{m0}},k_{\text{m1}},k_{\text{m2}}$ for the carbon fiber propellers are given by \begin{equation} \begin{array}{ll} k_{\text{t0}}=0.323, & k_{\text{m0}}=0.0432\\ k_{\text{m1}}=0.01, & k_{\text{m2}}=0.9 \end{array}.\label{Eq04KCM} \end{equation} \subsubsection{Propeller Efficiency Objective Function} Similar to the lift-drag ratio for airfoils, a widely used efficiency index to describe the efficiency of propellers is $\eta_{\text{T/M}}$, which is defined as the ratio between the thrust coefficient $C_{\text{T}}$ and the torque coefficient $C_{\text{M}}$ as \begin{equation} \eta_{\text{T/M}}\triangleq\frac{C_{\text{T}}}{C_{\text{M}}}=\frac{k_{\text{t0}}\varphi_{\text{p}}}{k_{\text{m0}}B_{\text{p}}\left(k_{\text{m1}}+k_{\text{m2}}\varphi_{\text{p}}^{2}\right)}\label{Eq05EffTM} \end{equation} where $\eta_{\text{T/M}}$ only depends on the aerodynamic design of the blade shape, which is convenient for manufacturers to improve the aerodynamic efficiency. Moreover, a higher $\eta_{\text{T/M}}$ means a smaller torque for generating the same thrust. Since $\eta_{\text{T/M}}$ is adopted by most of the manufacturers, this paper will use $\eta_{\text{T/M}}$ as the propeller efficiency objective function to obtain the optimal $\varphi_{\text{pOpt}}$ and $B_{\text{pOpt}}$. \subsubsection{Propeller Weight Objective Function} Through analyzing the propeller products on the market, the propeller weight $G_{\text{p}}$ can be described by a statistical model that depends on the diameter $D_{\text{p}}$ and the blade number $B_{\text{p}}$ as \begin{equation} G_{\text{p}}=f_{G_{\text{p}}}\left(B_{\text{p}},D_{\text{p}}\right)\label{eq:Gp} \end{equation} where $f_{G_{\text{p}}}\left(\cdot\right)$ is an increasing function of $D_{\text{p}}$ and $B_{\text{p}}$. Therefore, the minimum propeller weight $G_{\text{p}}$ requires that both $D_{\text{p}}$ and $B_{\text{p}}$ should be chosen as small as possible, which is described as \begin{equation} \min G_{\text{p}}\Rightarrow\min B_{\text{p}},\min D_{\text{p}}.\label{eq:minGp} \end{equation} \subsection{Motor Modeling} \subsubsection{Motor Circuit Model } The equivalent circuit of a BLDC motor has been presented in Fig.\,\ref{Fig02}, where $U_{\text{m}}$ (unit: V) is the motor equivalent voltage and $I_{\text{m}}$ (unit: A) is the motor equivalent current. According to \cite{Shi2017,chapman2005electric}, $U_{\text{m}}$ and ${I_{\text{{m}}}}$ can be obtained through \begin{equation} \left\{ \begin{array}{lll} {I_{\text{{m}}}} & = & \frac{\pi{M{K_{\text{{V}}}}{U_{\text{{m0}}}}}}{{30({U_{\text{{m0}}}}-{I_{\text{{m0}}}}{R_{\text{{m}}}})}}+{I_{\text{{m0}}}}\\ {U_{\text{{m}}}} & = & {I_{\text{{m}}}R_{\text{{m}}}}+\frac{{{U_{\text{{m0}}}}-{I_{\text{{m0}}}}{R_{\text{{m}}}}}}{{{K_{\text{{V}}}}{U_{\text{{m0}}}}}}N \end{array}\right.\label{Eq05Um} \end{equation} where $M$ (unit: N$\cdot$m) is the output torque of the motor which equals to the propeller toque in Eq.\,(\ref{Eq02PropTorque}), $N$ is the motor rotating speed which equals to the propeller rotating speed in Eq.\,(\ref{Eq02PropTorque}). The current and voltage measured under no-load (no propeller) tests are called the no-load current $I_{\text{{m0}}}$ (unit: A) and the no-load voltage $U_{\text{{m0}}}$ (unit: V), where ${U_{\text{{m0}}}}$ is a constant value defined by manufacturers (usually ${U_{\text{{m0}}}=10}$\,V). Note that, the nominal motor resistance $R_{\text{m0}}$ on the product description is usually not accurate enough, so it is recommended to obtain the actual resistance $R_{\text{m}}$ according to the test data of the full-throttle current $I_{\text{{m}}}^{*}$ and speed $N^{*}$, where a correction expression is derived from Eq.\,(\ref{Eq05Um}) as $R_{\text{m}}$$\approx\left(U_{\text{{b}}}-N^{*}/K_{\text{V}}\right)/I_{\text{{m}}}^{*}\approx2\sim3R_{\text{m0}}$ according to our experimental results. \subsubsection{Motor Constraints} The actual operation voltage of the motor depends on the battery voltage $U_{\text{b}}$ instead of the motor nominal maximum voltage (NMV) $U_{\text{mMax}}$. The motor NMV $U_{\text{mMax}}$ defines the range of the battery voltage $U_{\text{b}}$ that can ensure the motor work safely, which is described as $U_{\text{b}}\leq U_{\text{mMax}}$. According to \cite{Shi2017}, to prevent the motor from burnout, the motor equivalent voltage $U$$_{\text{m}}$ and current $I_{\text{m}}$ satisfy the constraints that \begin{equation} \begin{array}{l} U_{\text{m}}\leq U_{\text{m\ensuremath{\sigma_{\text{max}}}}}=U_{\text{b}}\leq U_{\text{mMax}}\\ I_{\text{m}}\leq I_{\text{m\ensuremath{\sigma_{\text{max}}}}}\leq I_{\text{mMax}} \end{array}\label{eq:MotConstraint} \end{equation} where $U_{\text{m\ensuremath{\sigma_{\text{max}}}}}$ and $I_{\text{m\ensuremath{\sigma_{\text{max}}}}}$ are the motor voltage and current under the full-throttle mode ($\sigma_{\text{max}}=1$). By letting the motor work under the maximum limit condition as \begin{equation} U_{\text{m}}={U_{\text{{mMax}}}}\text{, }I_{\text{m}}=I_{\text{mMax}}.\label{eq:MotorConstraint} \end{equation} Then, the maximum rotating speed ${N}_{\max}$ (unit: RPM) and torque ${M}_{\max}$ (unit: N$\cdot$m) of the motor can be obtained by combining the motor model in Eq.\,(\ref{Eq05Um}) \begin{equation} \left\{ \begin{array}{l} {N}_{\max}=f_{{N}_{\max}}\left(\Theta_{\text{m}}\right)\triangleq\frac{{\left({U_{\text{{mMax}}}-{R_{\text{m}}}{I_{\text{mMax}}}}\right){K_{\text{V}}}{U_{\text{m0}}}}}{\left({{U_{\text{m0}}}-{I_{\text{m0}}}{R_{\text{m}}}}\right)}\\ {M}_{\max}=f_{{M}_{\max}}\left(\Theta_{\text{m}}\right)\triangleq\frac{{30\left({{I_{\text{mMax}}}-{I_{\text{m0}}}}\right)({U_{\text{m0}}}-{I_{\text{m0}}}{R_{\text{m}}})}}{\pi{K}_{\text{V}}{U_{\text{m0}}}} \end{array}\right..\label{Eq18Mmax} \end{equation} Moreover, according to the propeller model in Eq.\,(\ref{Eq02PropTorque}), the theoretical maximum thrust ${T_{\text{pMax}}}$ (unit: N) can be obtained as \begin{equation} T_{\text{pMax}}=\frac{{{C}_{\text{T}}}}{{{C}_{\text{M}}}}\frac{M_{\max}^{4/5}{{\rho}^{1/5}}{{C}_{\text{M}}^{1/5}{N}_{\max}^{2/5}}}{{{60}^{2/5}}}.\label{Eq19TpMax1} \end{equation} To satisfy the maneuverability requirement, the theoretical maximum thrust range $\left[0,T_{\text{pMax}}\right]$ of the propulsion system should cover the required thrust range $\left[0,T_{\text{max}}\right]$, which means the following constraint should also be satisfied \begin{equation} T_{\text{pMax}}\geq T_{\text{max}}.\label{eq:TmaxTpMaxEq} \end{equation} \begin{figure}[tbh] \centering \includegraphics[width=0.4\textwidth]{Fig05-1} \caption{Statistical relationship between the motor maximum input power $U_{\text{{mMax}}}\cdot{I}_{\text{mMax}}$ and propeller maximum thrust $T_{\text{pMax}}$. The testing data come from the motor products website \cite{TMotor2017}.} \label{Fig05ThustWattVolt-1} \end{figure} According to the statistical results in Fig.\,\ref{Fig05ThustWattVolt-1}, there is an equality constraint between the maximum input power ${U_{\text{{mMax}}}{I}_{\text{mMax}}}$ and the theoretical maximum thrust ${T_{\text{pMax}}}$ for motor products as \begin{equation} \frac{{T_{\text{pMax}}}}{{U_{\text{{mMax}}}{I}_{\text{mMax}}}}\approx G_{\text{WConst}}\label{Eq37GWconst} \end{equation} where $G_{\text{WConst}}$ (unit: N/W) is a constant coefficient that reflects the technological process and product quality of products. According to the curve fitting result in Fig.\,\ref{Fig05ThustWattVolt-1}, the coefficient for the tested motors is $G_{\text{WConst}}\approx0.0624$. \subsubsection{Motor Efficiency Objective Function } The motor power efficiency $\eta_{\text{m}}$ is defined as \begin{equation} \eta_{\text{m}}\triangleq\frac{P_{\text{p}}}{P_{\text{m}}}=\frac{M\frac{2\pi N}{60}}{U_{\text{m}}I_{\text{m}}}. \end{equation} According to Eq.\,(\ref{Eq05Um}), the propeller torque $M$ and rotating speed $N$ can be described by $U_{\text{m}}\text{ and }I_{\text{m}}$, which yields that \begin{equation} \begin{array}{c} \eta_{\text{m}}=\left(1-\frac{I_{\text{m}}}{{U_{\text{{m}}}}}{R_{\text{{m}}}}\right)\left(1-\frac{1}{I_{\text{m}}}I_{\text{m0}}\right)\end{array}.\label{eq:EqEffmOtor} \end{equation} It can be observed from Eq.\,(\ref{eq:EqEffmOtor}) that the motor efficiency $\eta_{\text{m}}$ has negative correlations with $R_{\text{{m}}}$ and $I_{\text{m0}}$. Therefore, $R_{\text{{m}}}$ and $I_{\text{m0}}$ should be chosen as small as possible for the maximum motor efficiency, which is described as \begin{equation} \max\eta_{\text{m}}\Rightarrow\min R_{\text{{m}}},\min I_{\text{m0}}.\label{eq:EffM} \end{equation} \subsubsection{Weight Optimization Objective Function} Through analyzing the motor products on the market, the motor weight $G_{\text{m}}$ can be described by a statistical model depending on the motor nominal maximum voltage (NMV) $U_{\text{{mMax}}}$, the Nominal Maximum Current (NMC) $I_{\text{mMax}}$, and the KV value $K_{\text{V}}$ as \begin{equation} G_{\text{m}}=f_{G_{\text{m}}}\left(U_{\text{mMax}},I_{\text{mMax}},K_{\text{V}}\right).\label{eq:GmOrgin} \end{equation} According to Eqs.\,(\ref{Eq19TpMax1})(\ref{Eq37GWconst}), $I_{\text{mMax}}$ and $K_{\text{V}}$ can be described by $U_{\text{mMax}}$ and $T_{\text{pMax}}$. Therefore, Eq.\,(\ref{eq:GmOrgin}) can be rewritten into the following form \begin{equation} G_{\text{m}}=f_{G_{\text{m}}}^{\prime}\left(U_{\text{mMax}},T_{\text{pMax}}\right).\label{eq:Gm} \end{equation} Thus, by combining the constraints in Eqs.\,(\ref{eq:MotConstraint})(\ref{eq:TmaxTpMaxEq}), the motor weight optimization problem can be written into \begin{equation} \begin{array}{c} \underset{U_{\text{mMax}},T_{\text{pMax}}}{\min}f_{G_{\text{m}}}^{\prime}\left(U_{\text{mMax}},T_{\text{pMax}}\right)\\ \text{s.t. }U_{\text{b}}\leq U_{\text{mMax}},T_{\text{max}}\leq T_{\text{pMax}} \end{array}.\label{eq:minGm} \end{equation} In practice, the motor weight $G_{\text{m}}$ has a positive correlation with $U_{\text{mMax}}\text{ and }T_{\text{pMax}}$. Therefore, the minimum motor weight $G_{\text{m}}$ requires that the $U_{\text{mMax}}$ and $T_{\text{pMax}}$ should both be chosen as small as possible, which is described as \begin{equation} \min G_{\text{m}}\Rightarrow\min U_{\text{mMax}},\min T_{\text{pMax}}.\label{eq:minGp-1} \end{equation} Thus, solving Eq.\,(\ref{eq:minGm}) gives that \begin{equation} U_{\text{b}}=U_{\text{mMax}},T_{\text{max}}=T_{\text{pMax}}\label{eq:UbTmax} \end{equation} where $U_{\text{mMax}}$ should be chosen as small as possible. \subsection{ESC Modeling} \subsubsection{ESC Circuit Model} After receiving the throttle signal $\sigma{\in}\left[0,1\right]$ from the flight controller, the ESC converts the direct-current power of the battery to the PWM-modulated voltage $\sigma U_{\text{{e}}}$ for the BLDC motor without speed feedback. Then, the motor rotating speed is determined by both the motor and propeller models. Since the propeller model is nonlinear, the motor speed is not in proportion to the input throttle signal. According to the ESC equivalent circuit in Fig.\,\ref{Fig02}, the ESC current $I_{\text{e}}$ (unit: A) and voltage $U_{\text{e}}$ (unit: V) are given by \begin{equation} \begin{array}{cl} \sigma U_{\text{{e}}} & =U_{\text{{m}}}+I_{\text{{m}}}R_{\text{e}}\\ I_{\text{{e}}} & ={\sigma I_{\text{{m}}}} \end{array}.\label{Eq06Ie} \end{equation} \subsubsection{ESC Efficiency Objective Function} According to Eq.\,(\ref{Eq06Ie}), the power efficiency of the ESC can be obtained as \begin{equation} \begin{array}{lll} \eta_{\text{e}} & \triangleq & \frac{U_{\text{m}}I_{\text{m}}}{U_{\text{e}}I_{\text{e}}}=\frac{1}{1+\frac{I_{\text{m}}}{U_{\text{m}}}R_{\text{e}}}\end{array} \end{equation} which shows that the ESC efficiency $\eta_{\text{e}}$ increases as the resistance $R_{\text{e}}$ decreases, which yields that \begin{equation} \max\eta_{\text{e}}\Rightarrow\min R_{\text{e}}.\label{eq:EffE} \end{equation} \subsubsection{ESC Weight Objective Function} Through analyzing the ESC products on the market, the ESC weight $G_{\text{e}}$ can be described by a statistical model depending on the ESC nominal maximum voltage (NMV) $U_{\text{{eMax}}}$, the Nominal Maximum Current (NMC) $I_{\text{eMax}}$ as \begin{equation} G_{\text{e}}=f_{G_{\text{e}}}\left(U_{\text{eMax}},I_{\text{eMax}}\right)\label{eq:GeOrgin} \end{equation} where $f_{G_{\text{e}}}\left(\cdot\right)$ is in positive correlation with $U_{\text{eMax}}\text{ and }I_{\text{eMax}}$. Therefore, to minimize $G_{\text{e}}$, the ESC parameters $U_{\text{eMax}},I_{\text{eMax}}$ should be chosen as small as possible, which is described as \begin{equation} \min G_{\text{e}}\Rightarrow\min U_{\text{eMax}},\min I_{\text{eMax}}.\label{eq:GeMin} \end{equation} \subsection{Battery Modeling} \subsubsection{Battery Circuit Model} The battery is used to provide energy to drive the motor through ESC. The most commonly-used type battery is the LiPo battery because of the superior performance and low price. According to Fig.\,\ref{Fig02}, the battery model is given by \begin{equation} U_{\text{b}}=U_{\text{\ensuremath{\text{e}}}}+{I}_{\text{b}}{R}_{\text{b}}\label{Eq08Ue} \end{equation} where ${U}_{\text{b}}$ (unit: V) is the nominal battery voltage, and ${I}_{\text{b}}$ (unit: A) is the output current. Assuming that the number of the propulsion unit on a multicopter is $n_{\text{p}}$, the battery current is given by \begin{equation} {I_{\text{{b}}}=n_{\text{p}}I_{\text{{e}}}+I_{\text{other}}}\label{Eq09Ob} \end{equation} where $I_{\text{other}}$ (unit: A) is the current from other devices on the multicopter such as the flight controller and the camera. Usually, according to \cite{quan2017introduction}, it can be assumed that $I_{\text{other}}\approx0.5\text{\,A}$ if there is only a flight controller on the multicopter. According to \cite{Shi2017}, the battery discharge time $t_{\text{discharge}}$ (unit: min) is determined by the battery capacity $C_{\text{b}}$ and the discharge current $I_{\text{b}}$ \begin{equation} t_{\text{discharge}}\approx\frac{0.85C_{\text{b}}}{I_{\text{b}}}\cdot\frac{60}{1000}\label{eq:tdischarge} \end{equation} where the coefficient 0.85 denotes a 15\% remaining capacity to avoid over discharge. Note that the endurance computation equations in Eqs.\,(\ref{Eq08Ue})-(\ref{eq:tdischarge}) are simplified to reduce the computation time. They can be replaced by more accurate and nonlinear methods as presented in \cite{donateo2017design,donateo2017new} to increase the precision of the battery optimization result. \subsubsection{Battery Constraints} The Maximum Discharge Rate (MDR) $K_{\text{b}}$ (unit: mA/mAh or marked with symbol ``C'') of the battery is defined as \begin{equation} K_{\text{b}}=\frac{1000I_{\text{bMax}}}{C_{\text{b}}} \end{equation} where $I_{\text{bMax}}$ (unit: A) is the maximum discharge current that the battery can withstand. Since the battery should be able to work safely under the full-throttle mode of the motor, the maximum discharge current $I_{\text{bMax}}$ should satisfy \begin{equation} I_{\text{bMax}}\geq n_{\text{p}}I_{\text{e\ensuremath{\sigma_{\text{max}}}}}+I_{\text{other}}=n_{\text{p}}I_{\text{{mMax}}}+I_{\text{other}}\label{eq:Kbbb} \end{equation} which yields that \begin{equation} K_{\text{b}}\geq\frac{1000\left(n_{\text{p}}I_{\text{{mMax}}}+I_{\text{other}}\right)}{C_{\text{b}}}.\label{eq:KbConstraint} \end{equation} \subsubsection{Battery Efficiency Objective Function} According to Eqs.\,(\ref{Eq08Ue})(\ref{Eq09Ob}), the battery efficiency $\eta_{\text{b}}$ can be written into \begin{equation} \begin{array}{cc} \eta_{\text{b}} & \triangleq\frac{n_{\text{p}}U_{\text{e}}I_{\text{e}}}{U_{\text{b}}I_{\text{b}}}=\left(1-\frac{I_{\text{b}}}{U_{\text{b}}}R_{\text{b}}\right)\left(1-\frac{I_{\text{other}}}{I_{\text{b}}}\right)\end{array}\label{Eq09BEFF} \end{equation} which shows that the battery efficiency $\eta_{\text{b}}$ increases as the resistance $R_{\text{b}}$ decreases, which is described as \begin{equation} \max\eta_{\text{b}}\Rightarrow\min R_{\text{b}}.\label{eq:Effb} \end{equation} \subsubsection{Battery Weight Objective Function} According to the definition of the battery power density $\rho_{\text{b}}$ (unit: Wh/kg), the battery weight can be described as \begin{equation} G_{\text{b}}=\frac{C_{\text{b}}U_{\text{b}}}{1000g\rho_{\text{b}}}\label{eq:Gb} \end{equation} where $g=9.8\text{\,m/s\ensuremath{^{2}}}$ is the acceleration of gravity. Limited by the battery technology, the power density $\rho_{\text{b}}$ for a specific battery type is statistically close to a constant value. For instance, $\rho_{\text{b}}\text{\ensuremath{\approx}}140\text{\,Wh/kg}$ for LiPo batteries. Moreover, according to the statistical results, the battery weight is positively correlated with $K_{\text{b}}$. Therefore, to minimize $G_{\text{b}}$, the battery parameters $C_{\text{b}}$, $U_{\text{b}}$ and $K_{\text{b}}$should be chosen as small as possible, which is described as \begin{equation} \min G_{\text{b}}\Rightarrow\min U_{\text{b}},\min K_{\text{b}},\min C_{\text{b}}. \end{equation} \section{Design Optimization} \label{Sec-4} \subsection{Step 1: Propeller Efficiency Optimization} \label{Sec-DesOpProp} \subsubsection{Optimal Blade Number $B_{\text{pOpt}}$} According to Eq.\,(\ref{Eq05EffTM}), if only the blade number $B_{\text{p}}$ is considered, the propeller thrust efficiency $\eta_{\text{T/M}}$ can be simplified into the following form \begin{equation} \eta_{\text{T/M}}\propto\frac{1}{B_{\text{p}}}\label{Eqw14-2Bp} \end{equation} where the symbol ``$\propto$'' means ``in proportion to''. Eq.\,(\ref{Eqw14-2Bp}) indicates that $\eta_{\text{T/M}}$ monotonically decreases as ${B_{\text{{p}}}}$ increases. Therefore, to maximize $\eta_{\text{T/M}}$, the blade number ${B_{\text{{p}}}}$ should be chosen as small as possible. Moreover, according to the weight optimization principle in Eq.\,(\ref{eq:minGp}), the blade number should also be chosen as small as possible to minimize the propeller weight $G_{\text{p}}$. Considering that the blade number should satisfy the constraint that ${B_{\text{{p}}}\geq2}$, the optimal blade number ${B_{\text{{pOpt}}}}$ should be chosen as \begin{equation} {B_{\text{{pOpt}}}=2}.\label{Eq14BpOpt} \end{equation} \subsubsection{Optimal Pitch Angle $\varphi_{\text{pOpt}}$} According to Eq.\,(\ref{Eq05EffTM}), if only the pitch angle $\varphi_{\text{p}}$ is considered, $\eta_{\text{T/M}}$ can be simplified into the form as \begin{equation} \eta_{\text{T/M}}\propto\frac{1}{k_{\text{m1}}\varphi_{\text{p}}^{-1}+k_{\text{m2}}\varphi_{\text{p}}}.\label{EqfH14EffFHD} \end{equation} It is easy to obtain from Eq.\,(\ref{EqfH14EffFHD}) that $\eta_{\text{T/M}}$ first increases then decreases as $\varphi_{\text{p}}$ increases. Therefore, the optimal pitch angle $\varphi_{\text{pOpt}}$ should be able to maximize $\eta_{\text{T/M}}$, which yields that \begin{equation} \varphi_{\text{pOpt}}=\sqrt{\frac{k_{\text{m1}}}{k_{\text{m2}}}}.\label{Eq15DHOpt0} \end{equation} Noteworthy, if it happens in some cases that there are few propeller products in database with pitch angles close to the obtained $\varphi_{\text{pOpt}}$, then the mean pitch angle $\overline{\varphi}_{\text{p}}$ can be chosen as the optimal pitch angle $\varphi_{\text{pOpt}}$=$\overline{\varphi}_{\text{p}}$ to ensure the method can find a proper propeller product. \subsection{Step 2: Motor Weight Optimization} \label{subsec:motorWeight} \subsubsection{Optimal Motor NMV $U_{\text{mMaxOpt}}$} \begin{figure}[tbh] \centering \includegraphics[width=0.4\textwidth]{Fig05} \caption{Statistical relationship between the motor NMV $U_{\text{{mMax}}}$ and the propeller maximum thrust $T_{\text{pMax}}$. The testing data come from the motor products website \cite{TMotor2017}.} \label{Fig05ThustWattVolt} \end{figure} According to statistical results in Fig.\,\ref{Fig05ThustWattVolt}, the relationship between ${U_{\text{{mMax}}}}$ and ${T_{\text{pMax}}}$ can be described as a piecewise function \begin{equation} f{_{{U_{\text{{mMax}}}}}}\left({T_{\text{pMax}}}\right){\triangleq}\left\{ \begin{array}{cc} U_{1}, & {0}<{T_{\text{pMax}}}\leq{T}_{1}\\ U_{2}, & {T}_{1}<{T_{\text{pMax}}}\leq{T}_{2}\\ \cdots & \cdots \end{array}\right.\label{Eq36UmMaxFunc} \end{equation} where $f{_{{U_{\text{{mMax}}}}}}\left(\cdot\right)$ is obtained with the principle of minimizing $U_{\text{{mMax}}}$ for a given $T_{\text{pMax}}$. Thus, by substituting $T_{\text{pMax}}=T_{\text{max}}$ into Eq.\,(\ref{Eq36UmMaxFunc}), the optimal motor NMV $U_{\text{{mMaxOpt}}}$ can be obtained as \begin{equation} {U_{\text{{mMaxOpt}}}=f{_{{U_{\text{{mMax}}}}}}\left({T_{\text{max}}}\right)}.\label{Eq36UmMax} \end{equation} \subsubsection{Optimal NMC $I_{\text{{mMaxOpt}}}$} According to Eq.\,(\ref{Eq37GWconst}), there is a function relationship between $T_{\text{pMax}}$, $I_{\text{{mMax}}}$ and $U_{\text{{mMax}}}$. Therefore, by substituting $U_{\text{{mMax}}}=U_{\text{{mMaxOpt}}}$ and $T_{\text{pMax}}=T_{\text{max}}$ into Eq.\,(\ref{Eq37GWconst}), the optimal motor NMC $I_{\text{{mMaxOpt}}}$ can be obtained as \begin{equation} {I_{\text{{mMaxOpt}}}=\frac{{T_{\text{max}}}}{G_{\text{WConst}}U_{\text{{mMaxOpt}}}}}.\label{Eq37ImMaX1} \end{equation} \subsubsection{Optimal KV Value ${K}_{\text{VOpt}}$} Since the motor no-load current ${I_{\text{m0}}}$ and resistance $R_{\text{m}}$ are both very small in practice, it is reasonable to assume that \begin{equation} {I_{\text{m0}}}\approx0\text{ and }{R_{\text{m}}\approx0}.\label{Eq33ImRm} \end{equation} Substituting Eq.\,(\ref{Eq33ImRm}) into Eq.\,(\ref{Eq18Mmax}) gives \begin{equation} \begin{array}{l} M_{\max}\approx\frac{30{{I}_{\text{mMax}}}}{\pi{{K}_{\text{V}}}}\\ N_{\max}\approx{U_{\text{{mMax}}}{K}_{\text{V}}} \end{array}.\label{Eq34MmNm} \end{equation} Thus, Eq.\,(\ref{Eq19TpMax1}) can be simplified into the following form \begin{equation} {T_{\text{pMax}}\approx{k}_{\text{tm}}}\cdot{{\left(\frac{{{I}_{\text{mMax}}^{2}U_{\text{{mMax}}}}}{{{K}_{\text{V}}}}\right)}^{2/5}}\label{EQ35IUK} \end{equation} where ${{k}_{\text{tm}}}$ is defined as \begin{equation} k_{\text{tm}}=f_{k_{\text{tm}}}\left(\text{\ensuremath{B_{\text{p}},\varphi_{\text{p}},h_{\text{hover}}}}\right)\triangleq\sqrt[5]{k_{\text{c}}\frac{255\rho C_{\text{T}}^{5}}{\pi^{4}C_{\text{M}}^{4}}}\label{Eq36Km} \end{equation} where $\rho=f_{\rho}\left(h_{\text{hover}}\right),$ $C_{\text{T}}=f_{C_{\text{T}}}\left(B_{\text{p}},D_{\text{p}},\varphi_{\text{p}}\right)$ and $C_{\text{M}}=f_{C_{\text{M}}}\left(B_{\text{p}},D_{\text{p}},\varphi_{\text{p}}\right)$ are defined in Eqs.\,(\ref{Eq03AirDen})(\ref{Eq04Cm}), and $k_{\text{c}}$ is a constant correction coefficient to compensate for the neglected factors including $I_{\text{m0}}$, $R_{\text{m}}$. According to the statistical analysis, the correction coefficient can be set as $k_{\text{c}}$$\approx0.82$$.$ Finally, by substituting $U_{\text{{mMaxOpt}}},$ $I_{\text{mMaxOpt}}$, $T_{\text{max}}$, $B_{\text{pOpt}},$ $\varphi_{\text{pOpt}}$ and $h_{\text{hover}}$ into Eqs.\,(\ref{EQ35IUK})(\ref{Eq36Km}), the expression for the optimal KV value ${{K}_{\text{VOpt}}}$ is obtained as \begin{equation} {K}_{\text{VOpt}}=f_{{k}_{\text{tm}}}^{5/2}\left(\text{\ensuremath{B_{\text{pOpt}},\varphi_{\text{pOpt}},h_{\text{hover}}}}\right)\frac{{{I}_{\text{mMaxOpt}}^{2}U_{\text{{mMaxOpt}}}}}{{T_{\text{max}}^{5/2}}}.\label{Eq37KVOpt} \end{equation} \subsection{Step 3: Motor Efficiency Optimization} \subsubsection{Optimal Motor Resistance ${R_{\text{{m}Opt}}}$ and No-load Current $I_{\text{m0Opt}}$} In practice, $R_{\text{{m}}}$ and $I_{\text{m0}}$ should satisfy the constraint that \begin{equation} R_{\text{{m}}}>0\text{ and }I_{\text{m0}}>0.\label{eq:mRE} \end{equation} By combining Eq.\,(\ref{eq:EffM}) and Eq.\,(\ref{eq:mRE}), the optimal motor resistance ${R_{\text{{m}Opt}}}$ and no-load current $I_{\text{m0Opt}}$ are marked with \begin{equation} {R_{\text{{m}Opt}}=0}\text{ and }I_{\text{m0Opt}}=0 \end{equation} which denote that the motor resistance and no-load current should be chosen as close to zero as possible. Note that, in addition to parameters $R_{\text{{m}}}$,$I_{\text{m0}}$, the motor efficiency $\eta_{\text{m}}$ in Eq. (\ref{eq:EqEffmOtor}) is also determined by the motor working state ($U_{\text{m}}$,$I_{\text{m}}$) which further depends on factors including the throttle, rotating speed, propeller parameters, other motor parameters, air density, etc. According to our experimental results, a larger motor with a larger propeller will have a higher efficiency for generating the same thrust. Therefore, only maximizing the motor efficiency $\eta_{\text{m}}$ to solve all the motor parameters may not obtain the desired result because the size and weight will be very large. Thus, this paper finds the optimal propulsion system by considering both the efficiency and the weight to ensure the practicability of the obtained results. \subsection{Step 4: Optimal Motor Product} \label{SecOptMotSel} Although the optimal motor parameters $\Theta_{\text{mOpt}}\triangleq\{U_{\text{mMaxOpt}}$, $I_{\text{mMaxOpt}}$, $K_{\text{VOpt}}$, $R_{\text{mOpt}}$, $I_{\text{m0Opt}}\}$ have been obtained through the above procedures, it is still difficult to determine a corresponding product from the database. For example, $U_{\text{mMax}}$ of motor products are usually given by discrete form like 20A, 30A, 40A, while the obtained solutions are usually given with continuous form like $I_{\text{mMaxOpt}}=33.5$\,A. To solve this problem, a method is proposed to determine the optimal motor product according to $\Theta_{\text{mOpt}}$. For simplicity, the parameter set of the obtained motor product is represented by $\Theta_{\text{mOpt}}^{\ast}\triangleq\{U_{\text{mMaxOpt}}^{\ast}$, $I_{\text{mMaxOpt}}^{\ast}$, $K_{\text{VOpt}}^{\ast}$, $R_{\text{mOpt}}^{\ast}$, $I_{\text{m0Opt}}^{\ast}\}$. There are two selection principles for the optimal motor product: (i) The product should be selected by comparing with every parameter of $\Theta_{\text{mOpt}}$ in a proper sequence. Through the statistical analysis of the motor products on the market, a comparison sequence is given by considering the influence of each parameter on the motor weight as \begin{equation} U_{\text{mMaxOpt}}^{*}\text{, }K_{\text{VOpt}}^{*}\text{, }I_{\text{mMaxOpt}}^{*}\text{, }R_{\text{{mOpt}}}^{*}\text{, }{I_{\text{m0Opt}}^{*}}.\label{eq:secPrin} \end{equation} (ii) When comparing one parameter, on the premise of ensuring safety requirements, the product should be chosen equal to or close to the corresponding parameter in $\Theta_{\text{mOpt}}$. For example, if $I_{\text{mMaxOpt}}=33.5$\,A and the available current options are 20A, 30A, 40A, then it should be chosen that $I_{\text{mMaxOpt}}^{\ast}=40$\,A for some safety margin. The safety constraints for the selection of motor products are given by \begin{equation} \begin{array}{l} U_{\text{mMaxOpt}}^{\ast}\geq U_{\text{mMaxOpt}},\,I_{\text{mMaxOpt}}^{\ast}\geq I_{\text{mMaxOpt}}\end{array}.\label{EqMotorSeq} \end{equation} \subsection{Step 5: Propeller Weight Optimization} \subsubsection{Optimal Diameter $D_{\text{pOpt}}$} According to Eq.\,(\ref{eq:TmaxTpMaxEq}), the following constraint equation should be satisfied \begin{equation} T_{\text{pMax}}=C_{\text{T}}\rho\left(\frac{N_{\text{max}}}{60}\right)^{2}D_{\text{p}}^{4}\geq T_{\text{max}} \end{equation} which yields that \begin{equation} D_{\text{p}}\geq\sqrt[4]{\frac{60^{2}T_{\text{max}}}{C_{\text{T}}\rho N_{\text{max}}^{2}}}.\label{eq:MinDia} \end{equation} According to the optimization objective in Eq.\,(\ref{eq:minGp}), the optimal diameter should be chosen as the minimum diameter under constraint in Eq.\,(\ref{eq:MinDia}). Therefore, the optimal diameter $D_{\text{pOpt}}$ can be obtained by combining Eqs.\,(\ref{Eq03AirDen})(\ref{Eq04Cm})(\ref{Eq18Mmax}) with the parameters $h_{\text{hover}}$, $D_{\text{pOpt}}$, $B_{\text{pOpt}}$ and $\Theta_{\text{mOpt}}^{*}$ as \begin{equation} \begin{array}{cl} D_{\text{pOpt}} & =\sqrt[4]{\frac{60^{2}T_{\text{max}}}{\rho C_{\text{T}}N_{\text{max}}^{2}}}=\sqrt[4]{\frac{60^{2}T_{\text{pMax}}}{\rho C_{\text{T}}N_{\text{max}}^{2}}}=\sqrt[5]{{\frac{{M}_{\max}}{{\rho{C_{\text{M}}}{{\left({\frac{{N}_{\max}}{{60}}}\right)}^{2}}}}}}\\ & =\sqrt[5]{{\frac{3600f_{{M}_{\max}}\left(\Theta_{\text{mOpt}}^{*}\right)}{f_{\rho}\left(h_{\text{hover}}\right)f_{C_{\text{M}}}\left(B_{\text{pOpt}},D_{\text{pOpt}}\right)f_{N_{\text{max}}}^{2}\left(\Theta_{\text{mOpt}}^{*}\right)}}}. \end{array}\label{Eq19DpOpt} \end{equation} Since the propeller pitch $H_{\text{{p}}}$ is more convenient to select a propeller product, according to the definition of the pitch angle $\varphi_{\text{p}}$ in Eq.\,(\ref{Eq00-pitch}), the optimal propeller pitch $H_{\text{{pOpt}}}$ is given by \[ H_{\text{{pOpt}}}=\pi\cdot D_{\text{pOpt}}\cdot\tan\varphi_{\text{pOpt}}. \] \subsection{Step 6: Optimal Propeller Product} \label{SecOpProp} With the obtained parameters $\Theta_{\text{pOpt}}=\left\{ B_{\text{pOpt}},\varphi_{\text{p}},D_{\text{pOpt}}\right\} $, the optimal propeller product can be determined by searching the propeller product database. The parameter set of the obtained optimal propeller product is represented by $\Theta_{\text{pOpt}}^{\ast}=\left\{ B_{\text{pOpt}}^{*},\varphi_{\text{p}}^{*},D_{\text{pOpt}}^{*}\right\} $. Similar to the selection principles in \textit{Section\,\ref{SecOptMotSel}},\textit{ }the optimal propeller product should be determined from the propeller database by comparing the parameters in the sequence $B_{\text{pOpt}}^{*}$,$\varphi_{\text{p}}^{*}$,$D_{\text{pOpt}}^{*}$, and each parameter in $\Theta_{\text{pOpt}}^{\ast}$ should be chosen equal or close to the corresponding parameter in $\Theta_{\text{pOpt}}$ with satisfying the safety constraints \begin{equation} \begin{array}{c} B_{\text{pOpt}}^{*}=B_{\text{pOpt}},\,D_{\text{pOpt}}^{*}\leq D_{\text{pOpt}}\end{array}.\label{eq:DpOpt} \end{equation} \subsection{Step 7: ESC Weight Optimization} \subsubsection{Optimal ESC NMV $U_{\text{eMaxOpt}}$ and NMI $I_{\text{eMaxOpt}}$} Since the ESC and the motor are connected in series, their voltage and current should match with each other to ensure proper operations. For the safety, the ESC NMV $U_{\text{eMax}}$ and NMI $I_{\text{eMax}}$ should be able to support the maximum voltage and current from the motor, which means \begin{equation} \begin{array}{c} U_{\text{eMax}}\geq U_{\text{mMaxOpt}},\,I_{\text{eMax}}\geq I_{\text{mMaxOpt}}\end{array}.\label{eq:GeCons} \end{equation} By combining Eq.\,(\ref{eq:GeMin}) and Eq.\,(\ref{eq:GeCons}), the optimal $U_{\text{eMaxOpt}}$ and $I_{\text{eMaxOpt}}$ are given by \begin{equation} \begin{array}{c} U_{\text{eMaxOpt}}=U_{\text{mMaxOpt}},\,I_{\text{eMaxOpt}}=I_{\text{mMaxOpt}}\end{array}.\label{eq:OptESC} \end{equation} \subsection{Step 8: ESC Efficiency Optimization} \subsubsection{Optimal ESC Resistance $R_{\text{eOpt}}$} According to Eq.\,(\ref{eq:EffE}), the optimal ESC resistance $R_{\text{eOpt}}$ should be chosen as small as possible for the maximum ESC efficiency $\eta_{\text{e}}$. Since $R_{\text{e}}>0$ in practice, the optimal ESC resistance $R_{\text{eOpt}}$ is marked with \begin{equation} R_{\text{eOpt}}=0 \end{equation} which denotes that the ESC resistance should be chosen as close to zero as possible. \subsection{Step 9: Optimal ESC Product} With the obtained parameters $U_{\text{eMaxOpt}},I_{\text{eMaxOpt}},R_{\text{eOpt}}$, the optimal ESC product can be determined by searching the ESC product database. The parameter set of the obtained optimal ESC product is represented by $\Theta_{\text{eOpt}}^{\ast}=\left\{ U_{\text{eMaxOpt}}^{*},I_{\text{eMaxOpt}}^{*},R_{\text{eOpt}}^{*}\right\} $. Similar to the selection principles in \textit{Section\,\ref{SecOptMotSel}}, the optimal ESC product should be determined from the ESC database by comparing the parameters in the sequence $U_{\text{eMaxOpt}}^{*},I_{\text{eMaxOpt}}^{*},R_{\text{eOpt}}^{*}$, and each parameter in $\Theta_{\text{eOpt}}^{\ast}$ should be chosen equal or close to the corresponding parameter in $\Theta_{\text{eOpt}}$ with satisfying the constraints \begin{equation} \begin{array}{c} U_{\text{eMaxOpt}}^{\ast}\geq U_{\text{eMaxOpt}},\,I_{\text{eMaxOpt}}^{\ast}\geq I_{\text{eMaxOpt}}\end{array}. \end{equation} \subsection{Step 10: Battery Efficiency Optimization} \subsubsection{Optimal Battery Resistance ${R}_{\text{bOpt}}$} According to Eq.\,(\ref{eq:Effb}), $R_{\text{b}}$ should be chosen as small as possible for the maximum battery efficiency $\eta_{\text{b}}$. considering that $R_{\text{b}}>0$ in practice, the optimal battery resistance ${R}_{\text{bOpt}}$ is marked with \begin{equation} {R}_{\text{bOpt}}=0 \end{equation} which denotes that the battery resistance should be chosen as close to zero as possible. \subsection{Step 11: Battery Weight Optimization} \subsubsection{Optimal Battery Nominal Voltage $U_{\text{bOpt}}$} As analyzed in \textit{Section\,\ref{subsec:motorWeight}}, the actual working voltage of the motor is determined by the battery voltage, and the constraint in Eq.\,(\ref{eq:UbTmax}) should be satisfied to make sure the motor has the minimum weight. Therefore, after the optimal motor NMV $U_{\text{mMaxOpt}}$ is determined, the optimal battery voltage $U_{\text{bOpt}}$ is also determined as \begin{equation} U_{\text{bOpt}}=U_{\text{mMaxOpt}}.\label{eq:UbOpt} \end{equation} \subsubsection{Optimal Battery Capacity $C_{\text{bOpt}}$ } After the above procedures, the propeller, motor, ESC and battery of the propulsion system both have the maximum efficiency, which means the battery has the minimum current $I_{\text{b0}}$ under the hovering mode with propeller thrust $T_{\text{hover}}$. According to \cite{Shi2017}, with knowing the parameters $\Theta_{\text{pOpt}}^{*},\Theta_{\text{mOpt}}^{*},\Theta_{\text{eOpt}}^{*},U_{\text{bOpt}},T_{\text{hover}},n_{\text{p}},h_{\text{hover}}$, the battery current $I_{\text{b0}}$ can be estimated through the propulsion system equivalent circuit in Fig.\,\ref{Fig02}. By substituting the desired propeller thrust $T_{\text{hover}}$ into the propeller, motor, ESC and battery models in Eqs.\,(\ref{Eq02PropTorque})(\ref{Eq05Um})(\ref{Eq06Ie})(\ref{Eq08Ue}) successively, the battery discharge current $I_{\text{b0}}$ can be obtained. Then, the optimal battery capacity $C_{\text{bOpt}}$ can be obtained by substituting $t_{\text{discharge}}=t_{\text{hover}}$ into Eq.\,(\ref{eq:tdischarge}), which yields that \begin{equation} C_{\text{bOpt}}=\frac{t_{\text{hover}}I_{\text{b0}}}{0.85}\cdot\frac{1000}{60}\label{EqCbOpt} \end{equation} \subsubsection{Optimal \textmd{MDR $K_{\text{bOpt}}$}} The battery MDR $K_{\text{b}}$ should be chosen as small as possible within the constraint in Eq.\,(\ref{eq:KbConstraint}). Therefore, by substituting the obtained $U_{\text{mMaxOpt}}$ and $C_{\text{bOpt}}$ into Eq.\,(\ref{eq:KbConstraint}), the optimal battery MDR K$_{\text{bOpt}}$ is given by \begin{equation} K_{\text{b}}=\frac{1000\left(n_{\text{p}}I_{\text{{mMaxOpt}}}+I_{\text{other}}\right)}{C_{\text{bOpt}}}.\label{eq:KBop} \end{equation} \subsection{Step 12: Optimal Battery Product} With the obtained parameters $\Theta_{\text{bOpt}}=\left\{ U_{\text{bOpt}},K_{\text{bOpt}},C_{\text{bOpt}},R_{\text{bOpt}}\right\} $, the optimal battery product can be determined by searching the battery product database. The parameter set of the obtained optimal battery product is represented by $\Theta_{\text{bOpt}}^{*}=\left\{ U_{\text{bOpt}}^{*},K_{\text{bOpt}}^{*},C_{\text{bOpt}}^{*},R_{\text{bOpt}}^{*}\right\} $. Similar to the selection principles in \textit{Section\,\ref{SecOptMotSel}}, the optimal battery product should be determined from the battery database by comparing the parameters in the sequence $U_{\text{bOpt}}^{*},K_{\text{bOpt}}^{*},C_{\text{bOpt}}^{*},R_{\text{bOpt}}^{*}$, and each parameter in $\Theta_{\text{bOpt}}^{\ast}$ should be chosen equal or close to the corresponding parameter in $\Theta_{\text{bOpt}}$ with satisfying the constraints \begin{equation} \begin{array}{l} U_{\text{bOpt}}^{\ast}=U_{\text{bOpt}},\,K_{\text{bOpt}}^{\ast}\geq K_{\text{bOpt}}\end{array}. \end{equation} Noteworthy, the battery voltage $U_{\text{bOpt}}^{\ast}$ should satisfy the constraint that $U_{\text{bOpt}}^{\ast}=U_{\text{bOpt}}=U_{\text{mMaxOpt}}$ to ensure that the motor can work under the desired voltage. In practice, designers have to build a battery pack to satisfy the above design requirements by connecting small battery cells in series or parallel. According to \cite[pp. 46]{quan2017introduction}, by combining battery cells in series, a higher voltage can be obtained, with capacity unchanged. On the other hand, by combining battery cells in parallel, larger capacity and discharge current can be obtained, with voltage unchanged. \section{Experiments and Verification} \label{Sec-6} \subsection{Statistical Model Verification} \label{Sec-6-1}Comprehensive statistical analyses for the products of propellers, motors, ESCs and batteries on the market are performed to verify the weight statistical functions $f_{G_{\text{p}}}\left(\cdot\right)$, $f_{G_{\text{m}}}\left(\cdot\right)$, $f_{G_{\text{e}}}\left(\cdot\right)$ and $f_{G_{\text{b}}}\left(\cdot\right)$ in Eqs.\,(\ref{eq:Gp})(\ref{eq:GmOrgin})(\ref{eq:GeOrgin})(\ref{eq:Gb}) respectively. Some typical results are presented in Fig.\,\ref{Fig09-0}, where the products come from four most well-known manufacturers (APC, T-MOTOR, Hobbywing, Gens ACE). Fig.\,\ref{Fig09-0} shows the relationship between the weight and the parameters of each component. The statistical results are consistent with the analysis results in \textit{Section\,}\ref{Sec-4}. \begin{figure}[tbh] \centering \includegraphics[width=0.45\textwidth]{Fig09-0} \caption{Weight statistical results for the propulsion system components.} \label{Fig09-0} \end{figure} \subsection{Optimization Method Verification} \label{SecPropDiam-1} According to Eq.\,(\ref{Eq15DHOpt0}), the optimal pitch angle for T-MOTOR series propellers (parameters are listed in Eq.\,(\ref{Eq04KCM})) is obtained as $\varphi_{\text{pOpt}}=0.1054$. For the comparative validation, a series of multicopter propellers from the website \cite{TMotor2017} are listed in Table\,\ref{Table2TProp}. It can be observed from Table\,\ref{Table2TProp} that the statistical diameter/pitch ratio result is $D_{\text{p}}/H_{\text{p}}\approx3$ ($\varphi_{\text{p}}\approx0.1065$), which is in good agreement with the theoretical result $\varphi_{\text{pOpt}}=0.1054$. The comparison between the calculation value and the statistical value shows that the pitch angle optimization method is effective to find the optimal pitch angle adopted by manufacturers. Meanwhile, the propeller test data from the UIUC website \cite{UIUC2017} also show that the obtained pitch angle $\varphi_{\text{pOpt}}$ can guarantee high efficient in increasing the thrust and decreasing the torque. \begin{table}[th] \caption{Diameter/pitch ratio of carbon fiber propellers from the website \cite{TMotor2017}.} \label{Table2TProp} \centerin \begin{tabular}{|c|c|c|c|c|} \hline Propeller & P12x4 & P14x4.8 & P15x5 & P16x5.4\tabularnewline \hline $D_{\text{p}}/H_{\text{p}}$ & 3 & 2.92 & 3 & 2.963\tabularnewline \hline \hline Propeller & P18x6.1 & $\cdots$ & G26x8.5 & G28x9.2\tabularnewline \hline $D_{\text{p}}/H_{\text{p}}$ & 2.951 & $\cdots$ & 3.06 & 3.04\tabularnewline \hline \hline Propeller & G30x10 & G32x11 & G34x11.5 & G40x13.1\tabularnewline \hline $D_{\text{p}}/H_{\text{p}}$ & 3 & 2.91 & 2.96 & 3.05\tabularnewline \hline \end{tabular} \end{table} The motor T-MOTOR U11 KV90 is adopted as an example to verify the proposed optimization method. The calibrated motor parameters of U11 KV90 are listed below \begin{equation} \begin{array}{c} K_{\text{V}}=90\text{RPM/V},\text{ }U_{\text{mMax}}=48\text{V},\text{ }I_{\text{mMax}}=36\text{A},\\ U_{\text{m0}}=10\text{V},\text{ }I_{\text{m0}}=0.7\text{A},\text{ }R_{\text{m}}=0.3\Omega \end{array}\label{eq:U11Param-1} \end{equation} and the corresponding experiment results from \cite{TMotor2017} are listed in Table\,\ref{Tab4PropData}. \begin{table}[ptb] \caption{Full-throttle test data of U11 KV90 with 12S Li-Po battery (48V)} \label{Tab4PropData}\centerin \begin{tabular}{|c|>{\centering}p{0.04\textwidth}|>{\centering}p{0.05\textwidth}|>{\centering}p{0.035\textwidth}|c|>{\centering}p{0.05\textwidth}|>{\centering}p{0.04\textwidth}|} \hline Prop. & Current (A) & Power (W) & Thrust (N) & RPM & Torque (N$\cdot$m) & Tempe. ($^{\circ}$C) \tabularnewline \hline 27x8.8CF & 24.6 & 1180.8 & 81.4 & 3782 & 2.623 & 58.5\tabularnewline \hline 28x9.2CF & 28.3 & 1358.4 & 91.3 & 3696 & 3.068 & 66.5\tabularnewline \hline 29x9.5CF & 31.9 & 1531.2 & 98.8 & 3602 & 3.41 & 78.5\tabularnewline \hline 30x10.5CF & 36.3 & 1742.4 & 106.8 & 3503 & 3.846 & HOT$!$\tabularnewline \hline \end{tabular} \end{table} From the perspective of theoretical calculation, the optimal diameter of motor U11 KV90 can be obtained by substituting the motor parameters in Eq.\,(\ref{eq:U11Param-1}) into Eq.\,(\ref{Eq19DpOpt}), where the obtained result is $D_{\text{pOpt}}=29.7$\,inches. Therefore, according to the constraint in Eq.\,(\ref{eq:DpOpt}), the propeller diameter should be chosen as ${D_{\text{pOpt}}^{\ast}}=29$\,inches because the motor will overheat if the propeller diameter is larger than $D_{\text{pOpt}}$. It can be observed from the experimental results in the last two rows of Table\,\ref{Tab4PropData} that the motor temperature becomes overheated when the propeller diameter changes from 29\ inches to 30\ inches. Therefore, the optimal propeller diameter obtained from experiments should be 29\,inches, which agrees with the theoretical optimal solution ${D_{\text{pOpt}}^{\ast}}=29$\,inches. \subsection{Design Optimization Example} As an example, assume that the given design task is to select an optimal propulsion system for a multicopter ($n_{\text{p}}=4$) whose total weight $G_{\text{total}}=196\text{\,N}$ (20\ kg) and flight time is $t_{\text{hover}}=17$\,min (flight altitude $h_{\text{hover}}=50\,\text{m}$). The component databases are composed of the ESC, BLDC motor, caber fiber propeller products from T-MOTOR website \cite{TMotor2017}, and the LiPo battery products from GENS ACE website \cite{ACE2017}. The key calculation results of the design procedures are listed as follows. i) The thrust requirements for the propulsion system are obtained from Eqs.\,(\ref{eq:T0})(\ref{Eq31T0TpmaxThr}) as $T_{\text{hover}}=49$\,N and $T_{\text{max}}=98$\,N, where $\gamma=0.5$ is adopted here. ii) The optimal propeller efficiency parameters are obtained from Eqs.\,(\ref{Eq14BpOpt})(\ref{Eq15DHOpt0}) as $B_{\text{pOpt}}=2$, $\varphi_{\text{pOpt}}=0.1065$\,rad. Then, with statistical models in Fig.\,\ref{Fig05ThustWattVolt-1} and Fig.\,\ref{Fig05ThustWattVolt}, the optimal motor parameters are obtained from Eqs.\,(\ref{Eq36UmMax})(\ref{Eq37ImMaX1})(\ref{Eq37KVOpt}) as $U_{\text{mMaxOpt}}=48$\,V, ${{I}_{\text{mMaxOpt}}}=34$\,A, ${{K}_{\text{VOpt}}}=91$\,RPM/V. Therefore, by searching products from the T-MOTOR website according to principles in Eqs.\,(\ref{eq:secPrin})(\ref{EqMotorSeq}), the optimal motor is determined as U11 KV90. iii) With the blade parameters in Eq.\,(\ref{Eq04KCM}), the optimal propeller diameter can be obtained from Eq.\,(\ref{Eq19DpOpt}) as $D_{\text{pOpt}}=0.7468\text{\,m}$, and the optimal propeller product is selected as 29x9.5CF 2-blade. In the same way, the ESC parameters can also be obtained from Eq.\,(\ref{eq:OptESC}) as $U_{\text{eMaxOpt}}=48$\,V and ${{I}_{\text{eMaxOpt}}}=34$\,A, and the optimal ESC product is selected as FLAME 60A HV. iv) For the battery, the optimal parameters are obtained from Eqs.\,(\ref{eq:UbOpt})(\ref{EqCbOpt})(\ref{eq:KBop}) as $U_{\text{bOpt}}=48$\,V, $C_{\text{bOpt}}=16000$\,mAh and $K_{\text{bOpt}}=10$\,C. The optimal battery selected from ACE website is TATTU LiPo 6S 15C 16000mAh $\times$ $2$. The obtained result has been verified by several multicopter designers. Experiments show that a quadcopter with the designed propulsion system can efficiently meet the desired design requirements. If a larger motor (like T-MOTOR U13) is selected, then a smaller propeller has to be chosen for generating the same full-throttle thrust $T_{\text{pMax}}$, which reduces the motor efficiency because the optimal operating condition cannot be reached. As a result, the obtained propulsion system is heavier than the optimal result according to our experiments. If a smaller motor is selected, then a larger propeller has to be chosen, which results in exceeding the safety current of the motor. Therefore, the obtained propulsion system is optimal within the given database. \subsection{Method Application} If the same optimization problem is solved by brute force searching methods to traverse all combinations and evaluate the performance of each combination, the time consumption will be far longer than the proposed method. For example, it takes about 100ms for our evaluation method in \cite{Shi2017} to calculate the performance of each propulsion combination. Assuming that the numbers of products in the propeller, motor, ESC and battery databases all equal to $n$, then it will take about $T\left(n\right)\approx C_{n}^{4}=O(n^{4})$ to traverse all products in the databases. By comparison, the computation amount of the proposed method to traverse 4 component databases is 4$n$. Since there are 15 parameters in Table\,\ref{tab1}, it is easy to verify that the total computation amount of the proposed method is approximate to $T\left(n\right)\approx15\cdot4n=O(n)$, which is much faster than the brute force searching methods. The optimization algorithm proposed in this paper is adopted as a sub-function in our online toolbox (URL: \uline{\url{www.flyeval.com/recalc.html}}) to estimate the optimal propulsion system with giving the multicopter total weight. The program is fast enough to be finished within 30\ ms by using a web server with low configuration (single-core CPU and 1GB of RAM). The feedback results from the users show that the optimization results are effective and practical for the multicopter design. \section{Conclusions} \label{Sec-7} In this paper, the precise modeling methods for the propeller, ESC, motor and battery are studied respectively to solve the optimization problem for the propulsion system of multicopters. Then, the key parameters of each component are estimated through mathematical derivations to make sure that the obtained propulsion system has the maximum efficiency. Experiments and feedback of the website demonstrate the effectiveness of the proposed method. The propulsion system is the most important part of a multicopter, and its design optimization method will be conducive to the fast, optimal and automatic design of the whole multicopter system or other types of aircraft systems. The theoretical analysis can be further used to directly maximize the endurance of all kinds of UAVs, which is interesting for future research. \bibliographystyle{IEEEtran}

9e6b6992e7ab7d88af8d1ad03c250a42e9b35cfc

\section{Introduction} The discovery of gravitational wave signals~\cite{Abbott:2016blz,Abbott:2016nmj,Abbott:2017oio,TheLIGOScientific:2017qsa} associated with the coalescence of two gravitationally interacting compact bodies (either black holes or neutron stars) and the ongoing analysis of the signals has demonstrated the importance of having accurate mathematical descriptions of the underlying dynamics. Hence, updating such models with useful information coming from different (analytic, semi-analytic or numeric) approximation methods remains an active research area. Spurring this on further is the promise of a wide range of complicated low-frequency gravitational wave sources visible by the space based interferometer LISA \cite{Klein:2015hvg,Caprini:2015zlo,Tamanini:2016zlh,Bartolo:2016ami,Babak:2017tow} All existing methods have indeed a limited range of applicability. For example, when the two-body dynamics occurs in a weak-field and slow motion regime the key method is the post-Newtonian (PN) expansion; when, instead, the field is weak but the motion is no longer slow one can apply the post-Minkowskian (PM) approximation; finally, when the mass-ratio of the two bodies is very small the general relativistic perturbation theory on the field of the large mass---referred to as the gravitational self-force approach (GSF in short)---can be conveniently used. In addition to these more analytic approaches, there is numerical relativity (NR) where one directly solves the full Einstein equations numerically without making any fundamental approximations. While NR offers the only direct approach to viewing the merger and ringdown phases of a binary coalescence, computational costs exclude the early inspiral and situations where the binary has a small mass ratio. These methods have been developed independently from each other, allowing fruitful crosschecking of results. This paper is concerned with phenomena associated with extreme mass-ratio inspirals, a source for LISA which is most naturally described using GSF. Of particular interest has been the identification of gauge invariant physical effects of the conservative self-force, for example the well known periastron advance \cite{Barack:2010ny, vandeMeent:2016hel} and redshift invariants \cite{Detweiler:2008ft,Barack:2011ed,Sago:2008id,Shah:2010bi,Shah:2012gu,vandeMeent:2015lxa,Kavanagh:2016idg,Kavanagh:2015lva,Bini:2016dvs}; for a recent review of this topic see \cite{Barack:2018yvs}. These invariants rely on the delicate regularization techniques for dealing with the singular nature of the point like source. Therefore, calculating and comparing such invariants with either other independent self-force codes, post-Newtonian methods or numerical relativity simulations, one can confirm difficult calculations and validate new codes. The development of conservative gauge invariants has probed ever higher derivatives of the metric perturbations, requiring more careful regularization. We summarise the current knowledge of these invariants in Table~\ref{tab:GaugeInvariants}. The aim of the present work is to expand the range of gauge invariants to include knowledge of the GSF corrections to the accumulated precession angle of the spin vector of a test gyroscope per radian of orbital motion, commonly referred to as the spin precession invariant. The gyroscope (carrying a small mass $m_1$ and a small spin $S_1$) moves along a circular geodesic orbit in Kerr spacetime (with mass $m_2$, spin $S_2$), generalizing previous results for a nonrotating black hole \cite{Dolan:2013roa,Bini:2014ica}. We see this also as a basis for generalizing the more difficult eccentric orbit calculation in Schwarzschild spacetime \cite{Akcay:2016dku,Kavanagh:2017wot} to Kerr spacetime as suggested in \cite{Akcay:2017azq}. We will use the notation $a_1=S_1/m_1$ and $a_2=S_2/m_2$ for the spin-to-mass ratio of the two bodies and associated dimensionless spin variables $\chi_1=S_1/m_1^2$ and $\chi_2=S_2/m_2^2$. Other standard notations are $M=m_1+m_2$ for the total mass of the system and \begin{equation} q=\frac{m_1}{m_2}\ll 1 \,,\qquad \nu=\frac{m_1 m_2}{(m_1+m_2)^2}\ll1 \,, \end{equation} for the ordinary and symmetric mass-ratios, respectively. See Table~\ref{tab:notation} for an overview of our notational conventions. Unless differently specified we will use units so that $c=G=1$. \begin{table}[htt] \caption{Overview of calculations of the various gauge invariants in the literature and a sample of references. } \begin{ruledtabular} \begin{tabular}{lcc} & \textbf{Schwarzschild} & \textbf{Kerr} \\ \hline Redshift & \checkmark \cite{Detweiler:2008ft,Sago:2008id,Shah:2010bi,Barack:2011ed,Bini:2013zaa,Bini:2013rfa, Bini:2014nfa,Kavanagh:2015lva,Hopper:2015icj,Johnson-McDaniel:2015vva,Bini:2015bla,Bini:2015bfb,Bini:2016qtx,Bini:2018zde} & \checkmark \cite{Shah:2012gu,Shah:2013uya,vandeMeent:2015lxa,Bini:2015xua,Kavanagh:2016idg,Bini:2016dvs} \\ Spin precession & \checkmark \cite{Dolan:2013roa,Bini:2014ica,Bini:2015mza,Kavanagh:2015lva,Shah:2015nva,Akcay:2016dku,Kavanagh:2017wot,Bini:2018aps} & This work \\ Quadrupolar Tidal & \checkmark \cite{Dolan:2014pja,Bini:2014zxa,Bini:2014ica,Kavanagh:2015lva,Shah:2015nva,Bini:2018kov} & \checkmark \cite{Bini:2018dki}\\ Octupolar Tidal & \checkmark \cite{Bini:2014zxa,Nolan:2015vpa} & \end{tabular} \end{ruledtabular} \label{tab:GaugeInvariants} \end{table} \begin{table}[htt] \caption{List of notations related to mass and spin used in this paper. Care is required since notation often varies between GSF and PN literature.} \begin{ruledtabular} \begin{tabular}{ll} $m_1$ & mass of small body \\ $m_2$ & mass of Kerr BH \\ $M$ & $m_1+m_2$ \\ $q$ & small mass ratio \\ $\nu$ & symmetric mass ratio \\ $S_i$ & spin magnitude of body $i$ \\ $a_i$ & $S_i/m_i$ \\ $a$ & $a_2$ \\ $\hat{a}$ & $a/m_2$ \\ $\chi_i$ & $a_i/m_i$ \end{tabular} \end{ruledtabular} \label{tab:notation} \end{table} \section{Kerr metric and perturbation} The (unperturbed) Kerr line element written in standard Boyer-Lindquist coordinates $(t,r,\theta,\phi)$ reads \begin{eqnarray} \label{kerrmet} ds_{(0)}^{2}&=& g^{(0)}_{\alpha\beta} dx^\alpha dx^\beta \nonumber\\ &=&-\left(1-\frac{2m_2r}{\Sigma} \right) dt^2-\frac{4am_2r \sin^2\theta}{\Sigma}dtd\phi\nonumber\\ &+&\frac{\Sigma}{\Delta}dr^2 + \Sigma d\theta^2 \nonumber\\ &+&\left( r^2+a^2+\frac{2m_2ra^2\sin^2\theta}{\Sigma} \right)\sin^2\theta d\phi^2\,, \end{eqnarray} where \begin{equation} \Delta= r^2+a^2-2m_2r\,,\qquad \Sigma=r^2+a^2\cos^2\theta\,. \end{equation} Let us consider the perturbation induced by a test gyroscope moving along a circular equatorial orbit at~$r=r_0$. The perturbed regularized metric will be denoted by $g^{\rm R}_{\alpha\beta}= g^{(0)}_{\alpha\beta}+q h^{\rm R}_{\alpha\beta}+\mathcal{O}(q^2)$, with corresponding line element \begin{equation} \label{kerrmet_pert} ds^2=(g^{(0)}_{\alpha\beta}+q h_{\alpha\beta}+\mc{O}(q^2)) dx^\alpha dx^\beta \,, \end{equation} and is assumed to keep a helical symmetry, with associated Killing vector $k=\partial_t +\Omega \partial_\phi$. Because of the helical symmetry, the metric perturbation depend only on $\bar \phi=\phi-\Omega t$, $r$ and $\theta$, i.e., $h_{\mu\nu}=h_{\mu\nu}(\bar \phi , r, \theta)$. The gyroscope world line (in both the unperturbed and perturbed cases) has its unit timelike tangent vector aligned with $k^\alpha$, i.e., \begin{equation} u^\alpha= u^tk^\alpha\,, \end{equation} where $u^t$ is a normalization factor (such that $u^\alpha u_\alpha=-1$). In the (unperturbed) Kerr case the orbital frequency is given by \begin{equation} \label{3.4_0} m_2\Omega^{(0)} = \frac{u^{3/2}}{1+{\hat a}u^{3/2}}\,, \end{equation} and \begin{equation} u^{t\,(0)}= \frac{1+\hat a u^{3/2}}{\sqrt{1-3u+2\hat a u^{3/2}}}\,, \end{equation} with $u=m_2/r$ the dimensionless inverse radius of the orbit. In the perturbed situation the frequency becomes \begin{align} \label{3.4} m_2\Omega &=&m_2\Omega^{(0)}\left(1-q\frac{1+{\hat a}u^{3/2}}{4u^2} m_2[\partial_r h_{kk}^{\rm R}]_1+\mc{O}(q^2)\right) \,, \end{align} where we have denoted as \begin{equation} \label{3.5} h_{kk}^{\rm R}=h_{\alpha\beta}^{\rm R}k^\alpha k^\beta\,, \end{equation} the double contraction of $h^{\rm R}$ with the Killing vector $k$, and the subscript 1 stands for the evaluation at the position of the particle 1. In the perturbed metric the geodesic condition also implies $[\partial_{\bar \phi} h_{kk}^{\rm R}]_1=0$ \cite{Detweiler:2008ft}. \subsection{Gyroscope precession} We are interested in computing the spin precession invariant $\psi(y)$, measuring the accumulated precession angle of the spin vector of a test gyroscope per radian of orbital motion defined in \cite{Dolan:2013roa}. That is, the ratio of the \lq\lq geodetic" spin precession frequency to the orbital frequency \begin{equation} \label{5.1} \psi(y)\equiv \frac{\Omega_{\rm prec}}{\Omega}, \end{equation} where $\psi$ is written as a function of the gauge-invariant dimensionless frequency parameter \begin{equation} \label{3.6} y=(m_2\Omega)^{2/3}\,. \end{equation} We thus compute the precession frequency $\Omega_{\rm prec}$ of the small-mass body 1 carrying a small-spin orbiting the large-mass spinning body 2, to linear order in the mass-ratio. The precession frequency, both in the background and in the perturbed spacetime, is defined by (see, e.g., Ref. \cite{Bini:2014ica}) \begin{equation} \label{1.12} \Omega_{\rm prec}=\Omega-|\nabla k|\,, \end{equation} where \begin{equation} |\nabla k|^2=\frac12 [K_{\mu\nu}K^{\mu\nu}]_1\,, \end{equation} with \begin{equation} \label{1.4} K_{\mu\nu}=\nabla_\mu^{\rm R} k_\nu=-\nabla_\nu^{\rm R} k_\mu=\frac12 (\partial_\mu k_\nu-\partial_\nu k_\mu)\,. \end{equation} In terms of the gauge-invariant dimensionless frequency parameter \eqref{3.6}, Eq. \eqref{3.4} implies \begin{equation} u = \frac{y}{(1-{\hat a}y^{3/2})^{2/3}}\left(1+q\frac{m_2[\partial_r h_{kk}^{\rm R}]_1}{6y^2(1-{\hat a}y^{3/2})^{2/3}}+\mc{O}(q^2)\right)\,. \end{equation} We then have \begin{equation} \label{3.9} m_2 |\nabla k|= m_2 |\nabla k|^{(0)}\, (1+q\, \delta(y)+\mathcal{O}(q^2))\,, \end{equation} where \begin{equation} m_2 |\nabla k|^{(0)}=y^{3/2}\left[\frac{1+{\hat a}y^{3/2}-3y\left(1-{\hat a}y^{3/2}\right)^{1/3}}{1-{\hat a}y^{3/2}}\right]^{1/2}\,, \end{equation} and \begin{eqnarray} \delta(y)&=& \frac{\hat a}{2u^{1/2}}(1+{\hat a}u^{3/2})m_2\partial_r h_{kk}^{\rm R}\nonumber\\ &&+\frac1{2u^{1/2}} \left(\partial_r h_{\phi k}^{\rm R}-\partial_{\phi} h_{r k}^{\rm R}\right)\nonumber\\ && -\frac{u}2 \frac{(1-2{\hat a}u^{3/2}+{\hat a}^2u^2)^2}{(1-3u+2{\hat a}u^{3/2})(1-2u+{\hat a}^2u^2)} h_{kk}^{\rm R}\nonumber\\ && -u^{3/2}\frac{1-2{\hat a}u^{3/2}+{\hat a}^2u^2}{(1-2u+{\hat a}^2u^2)(1+{\hat a}u^{3/2})}\frac{1}{m_2}h_{t\phi}^{\rm R} \nonumber\\ && -\frac{u^2}2 \frac{1-u+2{\hat a}u^{3/2}(1-2u)+2{\hat a}^2u^3}{(1-2u+{\hat a}^2u^2)(1+{\hat a}u^{3/2})^2}\frac{1}{m_2^2}h_{\phi\phi}^{\rm R}\nonumber\\ && -\frac12 (1-2u+{\hat a}^2u^2)h_{rr}^{\rm R}\,,\label{eq:delta} \end{eqnarray} with $u = {y}/{(1-{\hat a}y^{3/2})^{2/3}}$ in this (and in any) $\mathcal{O}(q)$ quantity. To linear order in $q$ (keeping $y$ as fixed), inserting \eqref{3.9} into \eqref{5.1}, $\psi(y)$ reads \begin{equation} \label{5.2} \psi (y)=1-\frac{|\nabla k|^{(0)}}{y^{3/2}}[1+q\, \delta(y)+\mathcal{O}(q^2)]\,, \end{equation} where \begin{align} \frac{|\nabla k|^{(0)}}{y^{3/2}}=\left[\frac{1+{\hat a}y^{3/2}-3y\left(1-{\hat a}y^{3/2}\right)^{1/3}}{1-{\hat a}y^{3/2}}\right]^{1/2}\,, \end{align} so that the GSF piece $\delta \psi (y)$ in $\psi (y)$ (such that $\psi(y)=\psi_{(0)}(y)+q\, \delta \psi (y)+\mathcal{O}(q^2)$) is related to $\delta(y)$ via \begin{equation} \label{5.3} \delta \psi (y)=-\frac{|\nabla k|^{(0)}}{y^{3/2}}\, \delta(y)\,. \end{equation} The well known (unperturbed) Kerr result is then recovered, namely \begin{eqnarray} m_2\Omega_{\rm prec}^{(0)}&=& m_2\Omega^{(0)}-m_2|\nabla k|^{(0)}\\ &=& y^{3/2}\left[1-\sqrt{\frac{1+\hat a y^{3/2}}{1-\hat a y^{3/2}}-3y(1-\hat a y^{3/2})^{-2/3}} \right]\,,\nonumber \end{eqnarray} together with the corresponding Schwarzschild limit ($\hat a\rightarrow0$) \begin{eqnarray} m_2\Omega_{\rm prec,\, schw}^{(0)}&=& y^{3/2}\left[1-\sqrt{1-3y} \right]\,. \end{eqnarray} Our goal for the remainder of the paper is to evaluate \eqref{5.3} using \eqref{eq:delta}. \section{Methods} \begin{table} \caption{Overview of the different implementations used in this work. The two analytic methods I and II differ for the field mode decomposition in terms of either scalar spherical harmonics $Y_{lm}$ or spin-weighted spheroidal harmonics ${}_s S_{lm}$ and their respective derivatives. The third approach III, instead, uses purely a scalar harmonic projection. } \begin{ruledtabular} \begin{tabular}{cccccc} \# & Weyl scalar & Gauge & Mode decomposition& Method & Ref \\ \hline I & $\psi_0$ & ORG& ${}_s S_{lm},\partial_\theta{}_s S_{lm} $& analytic & \cite{Kavanagh:2016idg} \\ II & $\psi_0$ & ORG& $Y_{lm},\partial_\theta Y_{lm}$ & analytic & \cite{Bini:2016dvs} \\ III & $\psi_4$ & ORG& $Y_{lm} $& numeric & \cite{vandeMeent:2015lxa} \end{tabular} \end{ruledtabular} \label{tab:1} \end{table} \subsection{Radiation gauge metric reconstruction} In this work we will follow the Chrzanowski-Cohen-Kegeles (CCK) procedure for obtaining metric perturbations in a radiation gauge \cite{Cohen:1974cm,Chrzanowski:1975wv,Kegeles:1979an,Wald:1978vm,Lousto:2002em,Ori:2002uv,Keidl:2010pm,Shah:2012gu,vandeMeent:2015lxa}. Once a solution for the perturbed Weyl scalar $\psi_0$ (or $\psi_4$) has been obtained by solving the $s=2$ (or $s=-2$) Teukolsky equation, one then construct the Hertz potential $\hat{\Psi}_{0/4}$, in terms of which one finally compute the components of the perturbed metric $h_{\alpha\beta}^{\mathrm{rec}}$ by applying a suitable differential operator. We will use the outgoing radiation gauge (ORG) (see Table \ref{tab:1}), such that the metric perturbation $h_{\alpha\beta}^{\mathrm{rec}}$ satisfies the conditions \begin{equation} n^\alpha h^{\mathrm{rec}}_{\alpha\beta} = 0,\qquad {h^{\mathrm{rec}}}_{\alpha}^{\phantom{\alpha}\alpha} =0, \end{equation} where $n^\alpha$ is the ingoing principal null vector. When sources are present, radiation gauge solutions to the Einstein equations feature singularities away from the source region \cite{Ori:2002uv,Barack:2001ph}. If the source is a point particle, a string-like (gauge) singularity will extend from the particle to infinity and/or the background horizon \cite{Pound:2013faa,Keidl:2006wk}. Alternatively, one can construct a solution obtained by gluing together the regular halves from two `half-string' solutions. The result is a metric perturbation with a gauge discontinuity on a hypersurface containing the point source's worldline. We work with this solution. The gauge discontinuity splits the spacetime in two disjoint regions: an `exterior' region that extends to infinity (labelled ``$+$''), and an `interior' region that includes the background horizon (labelled ``$-$''). All the necessary steps to perform this kind of computations are now well established in the literature (see, e.g., Refs. \cite{Keidl:2010pm,Shah:2012gu}). In this work we implement the CCK procedure using three separate codes, two using analytic methods resulting in high order post-Newtonian expansions of the metric perturbation and thus precession invariant, and one numerical code giving high accuracy data over a finite set of radii and Kerr spin values. We highlight the variations of these methods we follow and provide references for more details of their techniques in Table~\ref{tab:1}. An aspect in which our methods differ is the basis of angular harmonics in which our fields are represented. Method I keeps the natural basis of spin-weighted spheroidal harmonics for the representation of the Weyl scalars, resulting in an expression for the metric perturbation and spin precession invariant in a combination of spin-weighted spheroidal harmonics and their angular derivatives. In method II the spin-weighted spheroidal harmonics are expanded in scalar spherical harmonics, resulting in expressions which are a combination of scalar spherical harmonics and their derivatives. In both methods I and II, as a result of the CCK procedure the coefficients in the harmonic expansion can still depend on the angular variables. In method III \emph{all} angular dependence is projected onto an expansion in scalar spherical harmonics. In all methods we use the solutions of the radial Teukolsky equation due to Mano, Suzuki and Takasugi~(MST) \cite{Mano:1996gn,Mano:1996vt} satisfying the correct boundary conditions at the horizon and at spatial infinity. In methods I and II these are expanded as an asymptotic series in $u$ for certain low values of the harmonic $l$ value, and supplement the MST series with a PN type ansatz for all higher values of $l$, obtaining the spin-precesion invariant as a PN expression. Method III instead evaluates the MST solutions numerically following \cite{Fujita:2004rb,Fujita:2009us,Throwethesis}. \subsection{Regularization} The quantity $\delta\psi$ is defined in terms of the Detweiler-Whiting regular field \cite{Detweiler:2002mi}, $h_{\mu\nu}^{\rm R}$. In practice this is obtained as the difference, \begin{equation} h_{\mu\nu}^{\rm R} = h_{\mu\nu}^{\rm ret} -h_{\mu\nu}^{\rm S}, \end{equation} between the retarded field $h_{\mu\nu}^{\rm ret}$ and the Detweiler-Whiting singular field $h_{\mu\nu}^{\rm S}$. Since both $h_{\mu\nu}^{\rm ret}$ and $h_{\mu\nu}^{\rm S}$ are singular on the particle world line, this subtraction cannot be performed there, but requires the introduction of a suitable regulator. We here use a variant of the so called $l$-mode regularization of \cite{Barack:2001gx}. This calls for extending Eqs. \eqref{eq:delta} and \eqref{5.3} to field equations by choosing an extension of the four-velocity $u$ to a field. Eqs. \eqref{eq:delta} and \eqref{5.3} can then be applied separately to $h_{\mu\nu}^{\rm ret}$ and $h_{\mu\nu}^{\rm S}$, obtaining the fields $\delta\psi^{\rm ret}$ and $\delta\psi^{\rm S}$. The spherical harmonic modes of these fields are then finite, and the necessary subtraction can be done a the level of these modes \begin{equation}\label{eq:modesum} \delta\psi^{\rm R} = \sum_{l=0}^{\infty} (\psi^{\rm ret}_l -\psi^{\rm S}_l). \end{equation} Conventional $l$-mode regularization procedures continue to calculate $\psi^{\rm S}$ locally near the worldline with chosen gauge and extension of the 4-velocity, yielding an expression for the large $l$ behaviour of $\psi^{\rm S}_l$, \begin{equation} \label{psi_sing} \psi^{\rm S}_l = \pm L A_\psi + B_\psi + L^{-1} C_\psi + \mathcal{O}(L^{-2}), \end{equation} with $L=2l+1$, and $\pm$ sign depended on the direction from which the worldline is approached. Consequently, the mode-sum \eqref{eq:modesum} can be evaluate as \begin{equation}\label{eq:modesum2} \delta\psi^{\rm R} = \sum_{l=0}^{\infty} (\psi^{\rm ret}_l \mp L A_\psi- B_\psi -L^{-1} C_\psi) -D_\psi, \end{equation} with \begin{equation}\label{eq:Dterm} D_\psi = \sum_{l=0}^{\infty} (\psi^{\rm S}_l \mp L A_\psi- B_\psi -L^{-1} C_\psi). \end{equation} We follow a slightly different approach first applied in Refs. \cite{Shah:2010bi} and \cite{Bini:2013zaa}. Since $h_{\mu\nu}^{\rm R}$ is a smooth vacuum perturbation the large $l$ behaviour of $\delta\psi^{\rm R}_l$ is expected the be $\mathcal{O}(\exp(-c/L))$, and consequently we can read off the coefficients $A_\psi$, $B_\psi$, and $C_\psi$ from the large $l$ behaviour of $\psi^{\rm ret}_l$, which we can determine either numerically of in a PN expansion. However, it is fundamentally impossible to determine the ``D-term'' \eqref{eq:Dterm} from the retarded field alone. In general it will depend on the chosen gauge, extension, and type of harmonic expansion (e.g. scalar, spin-weighted, mixed, ...). For the GSF it is known to vanish for a large class of (regular) gauges and extensions \cite{Heffernan:2012vj}. However, it is known to take non-zero values in the radiation gauge used in this work. In particular, the D-term will be different in the interior and exterior solutions. In \cite{Pound:2013faa} it was shown that for the GSF, the corrections to the D-term relative to the Lorenz gauge cancel when one takes the average of the interior and exterior solutions. This argument extends a much wider class of quantities (at least for suitably chosen extensions) including $\delta\psi$ \cite{gaugecompletion}. Consequently, if $D_\psi$ vanishes in the Lorenz gauge, we can calculate $\delta\psi^{\rm R}$ through \begin{equation} \delta\psi^{\rm R}=\sum_{l=0}^\infty \left[\frac12 (\delta\psi_l^+ + \delta\psi_l^-)-B_\psi -C_\psi/L\right]. \end{equation} In this work, we \emph{conjecture} that $D_\psi$ vanishes in the Lorenz gauge for the chosen extensions and harmonic decompositions. In part, this conjecture will be motivated post-facto by the agreement of our results with standard PN results up to 4PN order. In methods I and II we find that the expressions for $B_\psi$ and $C_\psi$ agree (despite differences in harmonic decomposition). In particular we find that $C_\psi=0$ and $B_\psi$ is given by \begin{align} B_\psi=\frac{|\nabla k|^{(0)}}{y^{3/2}}B\,, \end{align} with \begin{equation} B=\sum_{n=0}^N {\hat a}^n B^{a^n}(y)\,, \end{equation} and \begin{eqnarray} B^{a^0}(y)&=& \frac{1}{2}y-\frac{1}{4}y^2-\frac{63}{128}y^3-\frac{995}{1024}y^4-\frac{63223}{32768}y^5\nonumber\\ &&-\frac{126849}{32768}y^6-\frac{16567767}{2097152}y^7-\frac{555080733}{33554432}y^8\nonumber\\ &&-\frac{77104836855}{2147483648}y^9 \,,\nonumber\\ B^{a^1}(y)&=& -\frac{1}{2}y^{3/2}+\frac{5}{6}y^{5/2}+\frac{277}{384}y^{7/2}+\frac{1385}{1024}y^{9/2}\nonumber\\ &&+\frac{272245}{98304}y^{11/2}+\frac{1253839}{196608}y^{13/2}+\frac{34614543}{2097152}y^{15/2}\nonumber\\ && +\frac{1563825339}{33554432}y^{17/2}+\frac{297179922135}{2147483648}y^{19/2} \,,\nonumber\\ B^{a^2}(y)&=& -\frac{1}{4}y^3+\frac{43}{144}y^4-\frac{335}{9216}y^5-\frac{5953}{4096}y^6\nonumber\\ &&-\frac{2547251}{294912}y^7-\frac{343117}{9216}y^8-\frac{4673348817}{33554432}y^9 \,,\nonumber\\ B^{a^3}(y)&=& -\frac{1}{16}y^{9/2}+\frac{103757}{82944}y^{11/2}+\frac{2002033}{331776}y^{13/2}\nonumber\\ &&+\frac{1566715}{65536}y^{15/2}+\frac{891749345}{10616832}y^{17/2}\nonumber\\ &&+\frac{759009677191}{2717908992}y^{19/2} \,,\nonumber\\ B^{a^4}(y)&=& -\frac{39}{128}y^5-\frac{1783}{1024}y^6-\frac{158107}{31104}y^7\nonumber\\ &&-\frac{13809563}{995328}y^8-\frac{76343163}{2097152}y^9 \,\nonumber\\ B^{a^5}(y)&=&\frac{15}{128} y^{11/2}-\frac{299}{1024} y^{13/2}-\frac{1903 }{512}y^{15/2}\nonumber\\ &&-\frac{103026043 }{5971968}y^{17/2} \,\nonumber\\ B^{a^6}(y)&=&\frac{765 }{1024}y^7+\frac{39125 }{12288}y^8+\frac{34971 }{4096}y^9 \,\nonumber\\ B^{a^7}(y)&=&-\frac{175}{1024} y^{15/2}+\frac{14395 }{12288}y^{17/2} \,\nonumber\\ B^{a^8}(y)&=&-\frac{38535 }{32768}y^9\,. \end{eqnarray} and higher powers of $\hat{a}$ appear at higher PN orders. \subsection{Completion} Since the operator $\hat{\Psi}_{0/4}$ is not injective, its inverse is fundamentally ambiguous up to an element of the kernel of $\hat{\Psi}_{0/4}$. Wald showed that the only (global) vacuum solutions of the linearized Einstein equation in this kernel are perturbations to the mass and angular momentum of the background Kerr spacetime and pure gauge solutions. Hence the full metric perturbation can be written \begin{equation} \label{eq:comph} h_{\alpha\beta}^\pm = h_{\alpha\beta}^{\mathrm{rec},\pm} + (\partial_{m_2} g^{(0)}_{\alpha\beta})\delta M^\pm+(\partial_{S_2} g^{(0)}_{\alpha\beta})\delta J^\pm + \nabla_{(\alpha}\xi_{\beta)}^\pm, \end{equation} with $\delta M^\pm$ and $\delta J^\pm$ numbers and $\xi_{\beta}^\pm$ gauge vector fields. It was shown in \cite{Merlin:2016boc,vandeMeent:2017fqk} that for a particle source on a bound geodesic the amplitudes of the mass and angular momentum perturbations are given by \begin{align} \delta M^{-} &=0\,, & \delta J^{-} &=0\,, \nonumber\\ \delta M^{+} &= m_2 E\,, & \delta J^{+} &= m_2 L\,, \end{align} where $\hat{E}$ and $\hat{L}$ are the specific energy and angular momentum of the particle, \begin{eqnarray} \hat{E} &=& \frac{1-2u+\hat a u^{3/2}}{\sqrt{1-3u+2\hat a u^{3/2}}}\,,\nonumber\\ \frac{\hat{L}}{m_2} &=& \frac{1-2\hat a u^{3/2} +\hat a^2 u^2}{\sqrt{u} \sqrt{1-3u+2\hat a u^{3/2}}}\,. \end{eqnarray} If $\delta\psi$ were a proper gauge invariant quantity, then we could simply ignore the gauge vectors $\xi^{\alpha}_\pm$. However, $\delta\psi$ (like the orbital frequency) is only a quasi-invariant in the sense of \cite{gaugecompletion}, meaning that it is only invariant under gauge vectors that are bounded in time. We thus have to (partially) fix the gauge contribution to the metric as well. We start this process by noting that the other contributions to the metric perturbation in \eqref{eq:comph} are all bounded in time. Consequently, by restricting our attention to gauges in which $h_{\alpha\beta}$ is bounded in time, we only have to consider gauge vectors that produce bounded metric perturbations. The most general such gauge vector \cite{gaugecompletion,Shah:2015nva} is \begin{equation} \xi_\pm=m_2 t\left(A^t_\pm\partial_t+A^\phi_\pm\partial_\phi\right) + o(t). \end{equation} Consequently, to uniquely fix the value of $\delta\psi$, we only need to fix the values of $A^t_\pm$ and $A^\phi_\pm$. The $A_{+}^{t/\phi}$ can be fixed by requiring the full metric perturbation $h_{\alpha\beta}^{+}$ to be asymptotically Minkowski, yielding $A_{+}^{t/\phi}=0$. The interior values $A_{-}^{t/\phi}$ can further be fixed by requiring the continuity of suitably chosen quasi-invariant fields constructed from the metric perturbation \cite{gaugecompletion}. For circular equatorial orbits in Kerr this procedure yields, \begin{align} \label{VdMleq01} A_{-}^{t}&= \frac{[(u^{1/2}\hat a -2)\hat a u^{3/2}-1] u}{ (1-3 u+2\hat a u^{3/2})^{1/2}(\hat a u^{3/2}+1)} \,,\\ A_{-}^{\phi}&= \frac{(u^{1/2}\hat a-2) u^{5/2}}{(1-3 u+2\hat a u^{3/2})^{1/2}(\hat a u^{3/2}+1)} \,. \end{align} With this the final expression for the GSF contribution to the spin precession invariant becomes \begin{equation} \delta\psi = \delta\psi^\mathrm{rec} + \frac12\left(\hat{E} \delta\psi^M + \frac{\hat{L}}{m_2} \delta\psi^J + A^t_{-} \delta\psi^t + A^\phi_{-} \delta\psi^\phi\right)\,, \end{equation} with \begin{align} \delta\psi^M &=\frac{u(1+\hat{a}u^{1/2} - \hat{a}u^{3/2} +\hat{a}^2 u^2 )}{\sqrt{1-3u+2\hat{a}u^{3/2}}}, \\ \delta\psi^J &=-\frac{u^{3/2}(1 - u + \hat{a}u^{3/2} )}{\sqrt{1-3u+2\hat{a}u^{3/2}}}, \\ \delta\psi^t &=-\frac{u^{3/2}(1-\hat{a}u^{1/2})(1 + \hat{a}u^{3/2})}{\sqrt{1-3u+2\hat{a}u^{3/2}}}, \\ \delta\psi^\phi &=-\frac{1+\hat{a} u^{3/2}}{u^{3/2}\sqrt{1-3u+2\hat{a}u^{3/2}}}\nonumber\\ &\qquad\times(1 - 4 u + 3\hat{a}u^{3/2}-\hat{a}u^{5/2}+\hat{a}^2u^{3} ). \end{align} \section{Results} \subsection{Spin-Exact results to 8PN} \noindent Omitting the intermediate results of the radiative and completion parts, the output of method I of Table~\ref{tab:1} is the spin-precession invariant written as a PN series in $y$ and $\log y$ with no restriction on the spin of the black hole $\hat{a}$. The results take the form \begin{align} \delta \psi &= c_{1.5}y^{1.5}+ c_2 y^2 + c_{2.5} y^{2.5} + c_3 y^3 + c_{3.5} y^{3.5} + c_{4} y^{4} \nonumber\\ &\quad + c_{4.5} y^{4.5} + (c_{5} + c^{\rm ln}_{5} \log y) y^{5} + c_{5.5} y^{5.5}\nonumber\\ &\quad+ (c_{6} + c^{\rm ln}_{6} \log y) y^{6} + (c_{6.5} + c_{6.5}^{\rm ln} \log y) y^{6.5}\nonumber\\ &\quad + (c_{7} + c^{\rm ln}_{7} \log y) y^{7} + (c_{7.5} + c_{7.5}^{\rm ln} \log y) y^{7.5}\nonumber\\ &\quad + (c_{8} + c^{\rm ln}_{8} \log y + c^{\rm ln^2}_{8} \log^2 y) y^{8} +\mc{O}(y^{8.5}) . \label{Eq:deltapsi8pn} \end{align} The coefficients in this expansion are given by \begin{eqnarray} c_{1.5} &=& \hat{a}, \quad c_{2} = 1, \quad c_{2.5} =0, \quad c_{3} = -3, \nonumber\\ c_{3.5} &=& \tfrac{16}{3} \hat{a}, \quad c_{4} =- \tfrac{15}{2} - 3 \hat{a}^2, \nonumber \\ c_{4.5} &=& \big[ \tfrac{233}{6} - \tfrac{41}{32} \pi^2 \big]\hat{a}+ \hat{a}^3, \nonumber \\ c_{5} &=& - \tfrac{6277}{30} - 16 \gamma + \tfrac{20471}{1024} \pi^2 - \tfrac{496}{15} \log(2) - \tfrac{163}{9} \hat{a}^2, \nonumber \\ c_{5}^{\rm ln} &=& -8, \end{eqnarray} \begin{widetext} \begin{align} c_{5.5} &= \big[-\tfrac{89}{15} + \tfrac{248}{5} \gamma + \tfrac{1585}{1024} \pi^2 + \tfrac{504}{5} \log(2)\big]\hat{a} + 4 \hat{a}^3, \nonumber \\ c_{5.5}^{\rm ln} &= \tfrac{124}{5}\hat{a}, \nonumber \\ c_{6} &= -\tfrac{87055}{28}-\tfrac{52}{5} \gamma +\tfrac{653629}{2048} \pi ^2+\tfrac{3772 }{105}\log (2)-\tfrac{729 }{14}\log (3)-\big[\tfrac{667}{18}+\tfrac{11023}{3072} \pi ^2\big] \hat{a}^2, \nonumber \\ c_{6}^{\rm ln} &=-\tfrac{26}{5}, \nonumber \\ c_{6.5} &= -\tfrac{26536 }{1575} \pi+ \big[-\tfrac{6767891}{700}-\tfrac{3736 }{35} \gamma+\tfrac{1528679}{1536} \pi ^2-\tfrac{102232}{315} \log (2)+\tfrac{729}{7} \log (3)\big]\hat{a}+\big[\tfrac{6121}{162}+\tfrac{21}{1024} \pi ^2\big] \hat{a}^3, \nonumber \\ c_{6.5}^{\rm ln} &=-\tfrac{1868 }{35}\hat{a}, \nonumber \\ c_{7} &= -\tfrac{149628163}{18900}+\tfrac{7628}{21} \gamma +\tfrac{297761947}{393216} \pi ^2-\tfrac{1407987}{524288} \pi ^4+\tfrac{4556}{21} \log (2)+\tfrac{12879 }{35}\log (3)+\tfrac{1284 }{25} \pi \hat{a} \nonumber \\ &\quad+ \big[-\tfrac{969713}{225}-\tfrac{152 }{3} \gamma+\tfrac{4906229}{12288} \pi ^2-\tfrac{1528 }{15}\log (2)\big] \hat{a}^2-\tfrac{20 }{3}\hat{a}^4, \nonumber \\ c_{7}^{\rm ln} &=\tfrac{3814}{21}-\tfrac{76 }{3}\hat{a}^2, \nonumber \\ c_{7.5} &= -\tfrac{113411 }{22050} \pi+ \big[-\tfrac{3715435931}{28350}-\tfrac{995212}{2835} \gamma +\tfrac{48197747581}{3538944} \pi ^2-\tfrac{7009733}{524288} \pi ^4+\tfrac{540788}{2835} \log (2)-\tfrac{4617}{7} \log (3)+\tfrac{16}{5} \psi ^{\{0,1\}}(\hat{a}) \nonumber \\ &\quad-\tfrac{16}{5} \psi ^{\{0,2\}}(\hat{a})\big]\hat{a}+ \big[\tfrac{148627}{450}+\tfrac{136 }{5} \gamma -\tfrac{164339}{12288} \pi ^2+\tfrac{392 }{5}\log (2)-24 \log (\kappa)-\tfrac{12}{5} \psi ^{\{0,1\}}(\hat{a})-\tfrac{48}{5} \psi ^{\{0,2\}}(\hat{a})\big]\hat{a}^3+\tfrac{2 }{5}\hat{a}^5, \nonumber \\ c_{7.5}^{\rm ln} &= -\tfrac{497606 }{2835}\hat{a}+\tfrac{8 }{5}\hat{a}^3, \nonumber \\ c_{8} &= \tfrac{403109158099}{9922500}-\tfrac{74909462}{70875} \gamma +\tfrac{3424}{25} \gamma ^2+\tfrac{164673979457}{353894400} \pi ^2-\tfrac{160934764317}{335544320} \pi ^4+\tfrac{340681718}{1819125} \log (2)+\tfrac{869696 }{1575} \gamma \log (2) \nonumber \\ &\quad+\tfrac{58208 }{105}\log ^2(2)-\tfrac{199989}{352} \log (3)-\tfrac{9765625}{28512} \log (5)-\tfrac{1344}{5} \zeta (3)-\tfrac{3207503 }{33075} \pi \hat{a}\nonumber\\ &\quad+ \big[-\tfrac{40220568253}{132300}-\tfrac{4996 }{9} \gamma+\tfrac{2015707491}{65536} \pi ^2-\tfrac{724004}{945} \log (2) -\tfrac{4617}{14} \log (3)-\tfrac{32}{5} \log (\kappa )-\tfrac{32}{15} \psi ^{\{0,1\}}(\hat{a})-\tfrac{16}{15} \psi ^{\{0,2\}}(\hat{a})\big]\hat{a}^2\nonumber \\ &\quad+ \big[-\tfrac{1075453}{24300}-\tfrac{16 }{5} \gamma+\tfrac{679}{1024} \pi ^2-\tfrac{16 }{5}\log (2)-\tfrac{16}{5} \log (\kappa ) +\tfrac{8}{5} \psi ^{\{0,1\}}(\hat{a})-\tfrac{16}{5} \psi ^{\{0,2\}}(\hat{a})\big]\hat{a}^4 , \nonumber \\ c_{8}^{\rm ln} &= -\tfrac{37454731}{70875}+\tfrac{3424 }{25} \gamma-\tfrac{12634 }{45}\hat{a}^2-\tfrac{16 }{5}\hat{a}^4+\tfrac{434848}{1575} \log (2), \nonumber \\ c_{8}^{\rm ln^{2}} &=\tfrac{856}{25}, \nonumber \end{align} \end{widetext} where $\gamma$ is Euler's constant, $\zeta(n)$ is the Riemann zeta function, $\psi^{\{n,k\}} (\hat{a}) \equiv \psi^{(n)}(\tfrac{i k \hat{a}}{\kappa}) + \psi^{(n)}(\tfrac{-i k \hat{a}}{\kappa}) = 2 \Re[\psi^{(n)}(\tfrac{i k \hat{a}}{\kappa})]$ and $\psi^{(n)}(z) = \frac{d^{n+1}}{dz^{n+1}}\ln \Gamma(z)$ is the polygamma function. These are partially confirmed by the output of method~II of Table \ref{tab:1}, which provides the same expansion, with at each PN order a Taylor expansion in small~$\hat{a}$. \subsection{The PN expectation} In a two-body system $(m_1,S_1)$ and $(m_2,S_2)$, the precession frequency of the body 1, $\Omega_{\rm prec}=\Omega_1$, can be computed following Ref. \cite{Blanchet:2012at} (see Eq. (4.9c) there) in terms of the dimensionless binding energy ${\mathcal E}\equiv (E_{\rm system}-M)/\mu $ and angular momentum ${\mathcal L}\equiv L_{\rm system}/(M\mu)$ of the system, as \begin{equation} \label{Omega1def} \Omega_1 = \mu\frac{\partial({\mathcal E}-M\Omega {\mathcal L})}{\partial S_1}\,, \end{equation} where ${\mathcal E}$ and ${\mathcal L}$ are considered as functions of $(m_1,m_2,\Omega, S_1,S_2)$. Introducing the dimensionless frequency variable $x=(M\Omega)^{2/3}$, Eq. \eqref{Omega1def} implies \begin{equation} m_2 \Omega_1=m_2 M \nu \frac{\partial ({\mathcal E}(x)-x^{3/2}{\mathcal L}(x))}{\partial S_1}\,, \end{equation} where the invariant expressions for ${\mathcal E}(x)$ and ${\mathcal L}(x)$ follow straightforwardly from Eqs. (5.2) and (5.4) of Ref. \cite{Levi:2015uxa} (concerning the lastly known next-to-next-to-leading order in spin terms; for lower order terms see, e.g., Ref. \cite{Blanchet:2012at}). We list below the resulting expressions for the quantity $\widetilde {E}(x)={\mathcal E}(x)-x^{3/2}{\mathcal L}(x)$, that is \begin{equation} \widetilde E(x)=\widetilde {E}^{\rm O}(x)+\widetilde {E}^{\rm S}(x)+\widetilde {E}^{\rm SS}(x)+\widetilde {E}^{\rm SSS}(x)+\widetilde {E}^{\rm SSSS}(x)\,, \end{equation} with \begin{align} &\widetilde{E}^{\rm O}(x)=\nonumber\\ &\quad-\frac32 x +\left(-\frac{9}{8}-\frac{1}{8}\nu\right) x^2 +\left(\frac{19}{16}\nu-\frac{27}{16}-\frac{1}{48}\nu^2\right)x^3 \nonumber\\ &\quad+\left(\frac{6889}{384}\nu-\frac{405}{128}-\frac{31}{64}\nu^2-\frac{7}{3456}\nu^3-\frac{41}{64}\nu\pi^2\right)x^4\nonumber\\ & \quad+\bigg(-\frac{1701}{256}+\frac{451}{384}\nu^2\pi^2+\frac{43}{1152}\nu^3-\frac{24689}{3840}\nu \nonumber \\ &\quad -\frac{71207}{2304}\nu^2+\frac{11}{20736}\nu^4+\frac{64}{5}\nu\gamma+\frac{128}{5}\nu\ln(2)\nonumber\\ &\quad +\frac{32}{5}\nu\ln(x)+\frac{1291}{1024}\nu\pi^2\bigg)x^5+\mc{O}(x^6) \,,\nonumber\\ &\widetilde{E}^{\rm S}(x)= \left[\left(-\Delta-\frac{1}{2}\nu+1\right)\chi_1 +\left(1-\frac{1}{2}\nu+\Delta\right)\chi_2\right]x^{5/2}\nonumber\\ & \quad+\bigg[\left(\frac{3}{2}+\frac{31}{48}\nu\Delta-\frac{3}{2}\Delta+\frac{1}{24}\nu^2-\frac{121}{48}\nu\right)\chi_1 \nonumber \\ &\quad +\left(\frac{3}{2}\Delta+\frac{3}{2}+\frac{1}{24}\nu^2-\frac{121}{48}\nu-\frac{31}{48}\nu\Delta\right)\chi_2\bigg]x^{7/2}\nonumber\\ &\quad+\left[\bigg(-\frac{373}{32}\nu-\frac{27}{8}\Delta+\frac{211}{32}\nu\Delta+\frac{43}{12}\nu^2-\frac{7}{48}\nu^2\Delta+\frac{1}{48}\nu^3 \right.\nonumber \\ &\quad+\frac{27}{8}\bigg)\chi_1 +\bigg(\frac{1}{48}\nu^3-\frac{373}{32}\nu+\frac{27}{8}+\frac{27}{8}\Delta+\frac{7}{48}\nu^2\Delta \nonumber \\ &\quad\left.-\frac{211}{32}\nu\Delta+\frac{43}{12}\nu^2\bigg)\chi_2\right]x^{9/2}+\mc{O}(x^{11/2}) \,,\nonumber\\ &\widetilde{E}^{\rm SS}(x)= \bigg[ \left(\frac14 \Delta-\frac14+\frac12 \nu\right)\chi_1^2+\bigg(-\frac14 \Delta-\frac14+\frac12 \nu\bigg)\chi_2^2 \nonumber \\ &\quad-\chi_1\chi_2\nu \bigg] x^3 +\bigg[ \left(\frac{13}{24}\Delta-\frac{7}{12}\nu^2-\frac{29}{24}\nu\Delta-\frac{13}{24}+\frac{55}{24}\nu\right) \nonumber \\ &\quad\times\chi_1^2+\left(-\frac16 \nu^2-\frac12 \nu\right)\chi _2\chi_1\nonumber\\ &\quad\left. +\left(-\frac{13}{24}\Delta+\frac{55}{24}\nu+\frac{29}{24}\nu\Delta-\frac{7}{12}\nu^2-\frac{13}{24}\right)\chi_2^2\right]x^4\nonumber\\ & \quad+\bigg[\bigg(\frac{3095}{288}\nu+\frac{607}{288}\nu^2\Delta+\frac{59}{144}\nu^3+\frac{67}{32}\Delta -\frac{3017}{288}\nu^2\nonumber \\ &\quad-\frac{67}{32}-\frac{1889}{288}\nu\Delta\bigg)\chi_1^2+\left(\frac{53}{72}\nu^2-\frac{15}{8}+\frac{143}{24}\nu\right)\nu\chi_2\chi_1\nonumber\\ & \quad+\bigg(\frac{3095}{288}\nu-\frac{3017}{288}\nu^2-\frac{607}{288}\nu^2\Delta+\frac{59}{144}\nu^3-\frac{67}{32}-\frac{67}{32}\Delta \nonumber \\ &\quad+\frac{1889}{288}\nu\Delta\bigg)\chi_2^2\bigg]x^5 +\mc{O}(x^6) \,,\nonumber\\ &\widetilde{E}^{\rm SSS}(x)= \left[\left(\frac{1}{2}\nu-\frac{1}{4}+\frac{1}{4}\Delta\right)\chi_1^3 +\left(-\frac{1}{4}+\frac{1}{4}\Delta-\frac{1}{2}\nu\right)\chi_2\chi_1^2\right.\nonumber\\ &\left. +\left(-\frac{1}{4}-\frac{1}{2}\nu-\frac{1}{4}\Delta\right)\chi_2^2\chi_1 +\left(-\frac{1}{4}-\frac{1}{4}\Delta+\frac{1}{2}\nu\right)\chi_2^3\right]\nu x^{9/2}+\mc{O}(x^{11/2}) \,,\nonumber\\ &\widetilde{E}^{\rm SSSS}(x)= \mc{O}(x^6) \,. \end{align} In the previous expressions we have replaced the spin variables $S_{1,2}$ by their dimensionless counterparts $\chi_{1,2}=S_{1,2}/m_{1,2}^2$. In order to compare with the GSF expression derived above, we compute the spin precession invariant $\psi=(m_2 \Omega_1)/y^{3/2}$, where the variable $y$ is related to $x$ by $x=(1+q)^{2/3}y$. Linearizing in $q$ we find \begin{eqnarray} \psi(y)&=& \frac{3}{2}y+\frac{9}{8}y^2+\frac{27}{16}y^3\nonumber\\ &+&\left(-y^{3/2}-\frac12 y^{5/2}-\frac{15}{8} y^{7/2}\right)\chi_2-\frac12 y^3\chi_2^2\nonumber\\ &+&\nu \delta\psi(y) + \mc{O}(\nu^2,y^4)\,, \end{eqnarray} with \begin{align} \delta\psi(y) &= y^2-3 y^3 +\left(y^{3/2}+\frac{16}{3}y^{7/2}\right)\chi_2\\ &\quad+ \left(-y^{3/2}+\frac32 y^{5/2}+\frac{9}{8} y^{7/2}-2y^3\chi_2\right)\chi_1+ \mc{O}(y^4)\,.\nonumber \end{align} Here the zeroth-order in $\nu$ contribution to $\delta \psi$ coincides with the Kerr value (see, e.g., Eq. (70) of Ref. \cite{Iyer:1993qa}); the $\mathcal{O}(\nu)$ Schwarzschild contribution to $\delta \psi$ coincides with previous results \cite{Bini:2014ica,Dolan:2013roa}; the first terms linear in spin in $\delta \psi$ agree with our first-order GSF result \eqref{Eq:deltapsi8pn}. \subsection{Numerical results} \begin{figure}[t] \includegraphics[width=\columnwidth]{numdataplot.pdf} \caption{Numerical results for $\delta\psi$ for various values of the spin $a$. The vertical dashed lines show the location of the lightring for that value of the spin, where $\delta\psi$ diverges.} \label{fig:numdataplot} \end{figure} \begin{figure}[t] \includegraphics[width=\columnwidth]{weakfieldcomp.pdf} \caption{Comparison of the numerical and PN results for $\delta\psi$ in the weak field regime for $\hat{a} = 0.9$. The plotted lines show the residual after subtracting the $n$th order PN approximation. The slopes of each line is compatible with an $(n+1/2)th$ order residual, as one would expect. The blue and red vertical dashed lines show the location of the ISCO and the lightring respectively. The shade region gives an estimate on the numerical error in the calculation of $\delta\psi_{\mathrm{num}}$} \label{fig:weakfieldcomp} \end{figure} \begin{figure}[t!] \includegraphics[width=\columnwidth]{strongfieldcomp.pdf} \caption{Comparison of the numerical and PN results for $\delta\psi$ in the strong field regime for $\hat{a} = 0.9$. Above $y\approx 0.23$ the PN residuals no longer consistently improve with higher PN order, demonstrating the asymptotic nature of the PN expansion. The relatively good performance of the $6PN$ approximant seems mostly coincidental due to a zero crossing. The blue and red vertical dashed lines show the location of the ISCO and the lightring respectively.} \label{fig:strongfieldcomp} \end{figure} Method III obtains high accuracy numerical results for $\delta\psi$ without any post-Newtonian assumptions. Fig.~\ref{fig:numdataplot} shows the results for a variety of spins. One obvious feature is that as the (unstable) circular orbits approach the lightring $\delta\psi$ diverges. This behaviour is well-known in the analogous case of the redshift invariant \cite{Barack:2011ed,Akcay:2012ea,Bini:2015xua}, and was studied in the case of $\delta \psi$ around Schwarzschild in \cite{Bini:2014zxa}, which concluded that the light-ring divergence of $\delta \psi$ is proportional to $E^2$, where $E$ is the orbital energy. The data here is also compatible with a divergence~$\propto E^2$. The full numerical results are available from the \emph{black hole perturbation toolkit} website \cite{BHPToolkit}. Fig.~\ref{fig:weakfieldcomp} shows a comparison of the numerical results with the obtained PN results in the weak field regime. Shown are the residuals after subtracting successive orders in the PN expansion. We see a consistent improvement in the weak field, providing a strong verification of both the analytical PN results and the numerical results. Fig.~\ref{fig:strongfieldcomp} shows the same plot but with a focus on the strong field regime. Here the picture is very different. Around $y\approx 0.24$ we observe a locus where all PN approximants do about equally well (with the 6 and 6.5 PN terms as notable exceptions). Above this there is no noticeable improvement from going to higher PN orders. \section{Discussion and outlook} In this paper we have, for the first time, calculated the GSF corrections to the spin precession invariant along circular equatorial geodesic orbits in a perturbed Kerr spacetime, generalizing previous results limited to the case of a perturbed Schwarzschild spacetime. This calculation has been done with a variety of methods and techniques providing ample cross-validation. Comparison with existing PN results using the first law of binary mechanics \cite{Blanchet:2012at}, provides a strong validation of the used radiation gauge GSF techniques employed here, while also validating the previous PN results. Cross validation between the different GSF calculations, which vary in the level of rigor in their derivation, validates some of the underlying assumptions. In particular, a subtle importance is the agreement we find between the methods despite the differences in harmonic projections. State of the art radiation gauge self-force codes project from spin-weighted spheroidal harmonics to scalar spherical harmonics to meet up with rigorously defined regularization techniques, which has a large negative impact on the computational costs. In this work we have shown agreement between such a projected numerical code, and an unprojected analytical code without needing and additional correction terms. Investigating if such agreements between projections hold in more generic orbital configurations or for gauge dependent quantities (such as the self-force itself) would be of great importance in developing more efficient numerical codes for realistic self-force models. An important application of the results in this paper will be to inform effective-one-body (EOB) theory \cite{Buonanno:1998gg,Damour:2001tu}. As shown in \cite{Bini:2014zxa}, the spin precession can be used to determine contributions to the effective-one-body Hamiltonian for spinning black holes relating to the secondary spin. This transcription will be left to future work. This work, focusing on circular equatorial orbits, is a first step in determining the spin precession around Kerr black holes. The formalism for extending this work to eccentric equatorial orbits has already been laid out \cite{Akcay:2017azq} and should provide a basis for generalizing to generically inclined orbits. This should provide additional avenues of cross-validating difficult GSF calculations and informing EOB. \section*{Acknowledgments} DB thanks ICRANet and the italian INFN for partial support and IHES for warm hospitality at various stages during the development of the present project. MvdM was supported by European Union's Horizon 2020 research and innovation programme under grant agreement No 705229. The numerical results in this paper were obtained using the IRIDIS High Performance Computing Facility at the University of Southampton.

ede076f0273a3db85bc42cd244142c9c17412703

\section{Introduction} The academic community has recently taken significant interest in the better understanding of the locomotion of microorganisms \citep{Lauga2009, Lauga2016} and the design of biomimetic devices \citep{Nelson2010}. One popular type of synthetic swimmers are those able to self-propel through autophoresis -- a type of design that escapes the technical challenges associated with manufacturing motile body parts on a small scale. This class of devices makes use of self-generated gradients in temperature (thermophoresis), electric potential (electrophoresis) or concentration of some chemical species (diffusiophoresis) in order to induce motion. While currently at the centre of active research, the physical ideas behind autophoresis are by no means novel and a comprehensive overview of the theory of phoretic transport can be found in the classical article by \citet{Anderson89}. Recent technological advances have facilitated the manufacturing of diffusiophoretic swimmers in a variety of shapes and chemical properties, starting with pioneering experimental work done by \citet{paxton2004} and followed by a suite of detailed experiments \citep[][]{Howse2007,Ebbens2011,ebbens2012,ebbens2014}. There has also been much interest recently in the study of thermophoretic particles \citep[][]{jiang2010,golestanian2012,bickel2013} as well as in the understanding of Marangoni self-propelled droplets \citep[][]{thutupalli2011,schmitt2013,izri2014}. In the present paper, we focus on the mechanism of diffusiophoresis due to concentrations of non-ionic species. Many recent papers investigate ways of exploiting geometric asymmetry \citep[][]{shklyaev2014, Lauga15}, asymmetry of surface properties \citep[][]{golestanian2005, Golestanian07} or a combination of the two \citep[][]{popescu2010} in order to generate a local imbalance of concentration in an otherwise uniform medium and induce motion. In this paper we quantify the mechanisms through which diffusiophoretic particles may achieve artificial chemotaxis, i.e.~the directed motion along an external chemical gradient, both at the individual and the non-interacting population level. It is known that an asymmetric swimmer placed in a uniform background gradient will undergo active reorientation and experience a torque that seeks to align its axis of symmetry with the direction of the gradient \citep{bickel2014,Saha2014}. Although the mechanisms through which diffusiophoretic swimmers achieve chemotaxis are qualitatively different from the techniques used by living organisms \citep[see][]{Berg75}, the prospect of achieving the same functionality is highly desirable for both biomedical and technological applications \citep{Nelson2010, popescu2011}. In this work we first quantify the propulsion and reorientation mechanisms associated with the canonical problem of a spherical axisymmetric swimmer placed in a uniform background gradient of solute concentration. We approach this problem using the classical continuum framework of diffusiophoresis \citep{Golestanian07,julicher2009,sabass2012} as opposed to the osmotic framework proposed by Brady and coworkers \citep{cordova2008,brady2011,cordova2013}. To the best of our knowledge, we use the same setup as \citet{Saha2014}, but we generalise and correct their results. Specifically, we derive a fully analytical law for the instantaneous propulsion and orientation of a phoretic swimmer with general axisymmetric surface properties, in the limit of zero P\'{e}clet number for both substrate and product. We compute the solution for a weakly-reactive swimmer as an expansion in small Damk\"{o}hler number, including the first-order effects from the substrate which had been neglected by \citet{Saha2014}. The advantage of having a complete analytical law is that we are able to apply our results to real-life designs of phoretic swimmers, such as the Janus sphere, which could not be achieved using the truncated results previously published. We next develop a continuum model for the long-time behaviour of a population of non-interacting phoretic swimmers. Most practical applications rely on the synchronised behaviour of a large number of phoretic swimmers \citep{Wang2013,palacci2013,palacci2014} and macroscopic models are essential in understanding how these micro-swimmers would disperse under the competing action of active propulsion and thermal fluctuations. Our approach is inspired by work done on gyrotactic micro-organisms which successfully explained the formation of bioconvection patterns in cells such as \textit{Chalmydomonas nivalis}. Specifically, we use the framework of generalised Taylor dispersion theory developed by \citet{Frank91, Frank93} and apply it to the artificial chemotaxis of phoretic swimmers. We finally compare the predictions of our continuum model with numerical simulations of large samples of swimmers and obtain excellent agreement. Notably, our continuum model demonstrates that the interplay between active propulsion and reorientation leads to the anisotropic diffusion of phoretic swimmers, a piece of physical understanding overlooked by previous studies. \section{Instantaneous behaviour: Artificial chemotaxis of an individual particle} \label{section-2} In this first part of the paper, we derive the instantaneous law controlling the linear and angular transport of a phoretically-active sphere in a solute (or ``substrate'') gradient. \subsection{Setup: Chemical problem} A spherical particle of radius $R$ is immersed in a uniform gradient of substrate concentration and induces a disturbance to the equilibrium concentration of substrate not only due to the volume that it displaces, but also due to the chemical reaction that it facilitates at its surface. Before introducing the particle, the medium is characterised by a known linear background concentration of substrate, $s_b(\rrb') = \rrb' \bcdot \bnabla s^{\infty}$, where $\rrb'$ is the position vector relative to some location of vanishing substrate concentration. The presence of the device leads to a perturbed concentration of substrate, $s$, which must be calculated. Due to a catalytic coating, we assume that the swimmer promotes the conversion of substrate molecules $S$ into product molecules $P$, after an intermediary stage of binding to the enzyme molecules $E$. The classical model for this type of chemical behaviour is referred to as Michaelis-Menten kinetics, and starts from the reaction formula \begin{equation} \ce{{E+S} <=>[k_1][k_{-1}] ES ->[k_2] {E+P}}, \end{equation} where $k_1, k_{-1}$ and $k_2$ are reaction rates \citep[for a review, see][]{nelson2008}. When the supply of substrate is much larger than the supply of enzyme, the reaction is limited by the availability of enzymatic sites, and the system classically reaches a steady non-equilibrium state. The rate of conversion of substrate molecules S into product molecules P is given by the Michaelis-Menten rule \begin{equation} \kappa(s) = \frac{\kappa_1 \kappa_2 s}{\kappa_2 + \kappa_1 s}, \label{eq-MM} \end{equation} where $s$ is the substrate concentration, $\kappa_1 = k_1k_2/(k_{-1}+k_2)$ and $\kappa_2=k_2$. Note that $\kappa(s)$ is the rate of reaction per enzyme molecule, so the total flux at the surface of the swimmer will be the reaction rate $\kappa(s)$ multiplied by a measure of surface activity $\sigma(\theta,\phi)$, which quantifies the abundance of enzyme molecules at a particular point on the surface of the particle, parameterised by the spherical coordinate angles $(\theta,\phi)$ in a body-fixed frame of reference. \begin{figure} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig1a.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig1b.pdf} \end{minipage} \caption{Sketch of the problem. (a): Axisymmetric active particle with axis of symmetry along $z$. The surface properties (surface activity $\sigma$, substrate mobility $\mu_s$, and product mobility $\mu_p$) are functions of the polar angle $\theta$ alone. (b): Active particle placed in a uniform substrate gradient, $\bnabla s^{\infty}$, which lies in the $yz$-plane. The presence of the background gradient, together with the catalytic coating of the swimmer, induce local tangential chemical gradients on the surface of the sphere (light grey arrows) which give rise to an apparent fluid slip velocity (white arrows) leading to translation (linear velocity $\VVb$) and rotation (angular velocity $\omb$) of the particle.} \label{fig1} \end{figure} We next assume that the spherical swimmer has axisymmetric surface properties, and define the $z$-axis to be its axis of symmetry (see figure~\ref{fig1}a). We choose the origin at the centre of the sphere, and define our spherical coordinate system such that the polar angle $\theta$ is measured relative to the axis of symmetry of the swimmer. In this coordinate system, any surface property of the swimmer can be expressed as series of Legendre polynomials of the polar angle $P_l(\cos\theta)$. In particular, the axisymmetric surface activity $\sigma(\theta)$ can be written as \begin{equation} \sigma(\theta) = \sum_{l=0}^{\infty} \sigma_l P_l(\cos\theta). \end{equation} We note that the typical scale of substrate concentration surrounding a swimmer centred at position $\rrb_c$ is given by the background concentration of substrate at that position, i.e. $s \sim s_b(\rrb_c)$. It is this scale that determines the magnitude of the reaction rate on the surface of the swimmer. Using this scale, the nonlinear reaction rate from \eqref{eq-MM} has two distinguished limits in which analytical progress can be made. In the case when $s_b(\rrb_c) \gg \kappa_2/\kappa_1$, the reaction saturates at a constant rate $\kappa_2$. This is the limit studied by \citet{golestanian2005,Golestanian07,popescu2010} and \citet{Lauga15}, amongst others. The swimmer could still perform chemotaxis in this regime but solely due to the substrate since the product problem has become independent of the substrate dynamics. If the product is released at a constant rate $\kappa_2 \sigma(\theta)$ from the surface of the swimmer, it can only lead to translation along the axis of symmetry of the swimmer. Therefore, there is no chemotactic effect from the product in this regime. In contrast, when $s_b(\rrb_c) \ll \kappa_2/\kappa_1$, the reaction rate becomes proportional to the substrate concentration, \begin{equation} \kappa(s) \sim \kappa_1 s, \label{eq-linear-regime} \end{equation} which means that the release of product molecules at the surface of the swimmer also depends on the direction of the background gradient in substrate concentration. In this limit, both substrate and product can contribute to chemotaxis in a non-trivial way. In what follows we shall focus on the linear regime \eqref{eq-linear-regime} which provides a mathematically tractable problem yet one that leads to interesting chemotactic effects. Traditional diffusiophoresis (whereby a passive colloid is propelled due to a background gradient in the concentration of a chemical species) will also be captured in our model as the special case where the surface activity $\sigma$ is zero, and the presence of chemical reactions will add further complexity to the chemotactic behaviour of the swimmer. In the limit of zero P\'{e}clet number for both substrate and product, we can neglect the advection of molecules and solve a diffusion-reaction problem with the appropriate boundary conditions. In the bulk of the fluid, we must solve the steady diffusion equation for each component \begin{equation} D_s \nabla^2 s = 0, \quad D_p \nabla^2 p = 0, \label{eq-setup-diff} \end{equation} where $s$ and $p$ are the volume concentrations of substrate and product molecules, while $D_s$ and $D_p$ are their respective diffusivities. The diffusive flux normal to the particle surface is given by the consumption and production of substrate and product molecules, respectively. Given the definition of the surface activity, $\sigma$, we thus write these boundary conditions as \begin{equation} \left. -D_s \frac{\p s}{\p n} \right|_{|\rrb| = R} = \left.-\kappa_1 \sigma(\theta) s\right|_{|\rrb| = R}, \quad \left. -D_p \frac{\p p}{\p n} \right|_{|\rrb| = R} = \left.\kappa_1 \sigma(\theta)s\right|_{|\rrb| = R}. \label{eq-chem-BC} \end{equation} Finally, we impose that the perturbation due to the presence of the particle decays at infinity, such that we recover the background concentrations far away from the particle. Therefore, as $|\rrb| \rightarrow \infty$, we impose that \begin{equation} s \rightarrow s_b(\rrb_c) + \rrb \bcdot \bnabla s^{\infty}, \quad p \ttz, \label{eq-chem-inf} \end{equation} where $\rrb_c$ is the instantaneous position of the centre of the swimmer relative to the reference location where $s_b (\rrb') = 0$, and $\rrb = \rrb' - \rrb_c$ is the position relative to the centre of the swimmer. \subsection{Setup: Hydrodynamics} \label{sec-hydro} The physical basis for the self-propulsion of phoretic swimmers lies in the interactions between the chemical molecules present in the fluid and the surface of the particle (see sketch in figure~\ref{fig1}b with notation). Following the classical framework \citep{Anderson89}, interactions are restricted to a thin diffuse layer around the particle where the chemical molecules obey a Boltzmann distribution. The energy of the particle is given by a total interaction potential $\Phi$ which includes effects such as van der Waals forces and excluded volume effects. We assume that the size of the boundary layer is much smaller than the inverse of the local curvature everywhere on the surface of the particle, so that we can use a locally flat approximation. If there is a gradient in the concentration of the chemical species outside the diffuse boundary layer, it will give rise generically to an osmotic flow near the surface of the swimmer. This flow will lead to an apparent slip velocity at the edge of the diffuse boundary layer proportional to the tangential gradient in concentration of the chemical species. The constant of proportionality is called the diffusio-osmotic mobility, $\mu$, and it is a property of the surface interactions. A detailed description of diffusiophoresis can be found in \citet{Anderson89}, accompanied by the classical integral expression for the mobility $\mu$ in terms of the total interaction potential $\Phi$. For the purpose of the present paper, we may ignore the details of the interaction potential and simply work with the mobility as a given surface property. Because our swimmer is axisymmetric, we can also express the diffusio-osmotic mobility as a series of Legendre polynomials, as we did for the surface activity. In general there are two relevant mobilities, one for the substrate molecules ($\mu_s$) and the other for the product molecules ($\mu_p$), both of which we decompose as \begin{equation} \mu_s(\theta) = \sum_{l=0}^{\infty} \mu_{sl} P_l(\cos\theta), \quad \mu_p(\theta) = \sum_{l=0}^{\infty} \mu_{pl} P_l(\cos\theta). \end{equation} In the case of our spherical particle, the slip velocity due to either one of the chemical species at a point $(\theta,\phi)$ on the surface of the swimmer can then be written as \begin{equation} \vvb^\alpha_{slip} (\theta,\phi) = \mu_\alpha(\theta) (\IIb-\hat{\rrb}\hat{\rrb})\bcdot\bnabla \alpha |_{r=R}(\theta,\phi), \label{eq-slip-velocity} \end{equation} where $\mu_\alpha$ is the diffusio-osmotic mobility of the surface with respect to the chemical species in question, $(\IIb-\hat{\rrb}\hat{\rrb})\bcdot\bnabla$ is the tangential gradient operator on the surface of the sphere, and $\alpha$ is the concentration of the chemical species of interest (substrate or product). The final step in the setup of the problem is to translate the local slip velocity into the equations of motion of the particle. Within the low Reynolds number regime, inertia can be neglected and therefore it is relevant to talk about the instantaneous linear and angular velocities of the particle. A classical calculation using the reciprocal theorem on a force-free, torque-free sphere \citep[see][]{Stone96} establishes that the linear velocity is given by \begin{equation} \VVb^\alpha = - \frac{1}{4\pi} \int_0^{2\pi} \int_0^{\pi} \vvb^\alpha_{slip}(\theta,\phi)\sin\theta\mathrm{d}\theta \mathrm{d}\phi, \label{eq-linear-velocity} \end{equation} and the angular velocity by \begin{equation} \omb^\alpha = - \frac{3}{8\pi R} \int_0^{2\pi} \int_0^{\pi} \hat{\rrb} \wedge \vvb^\alpha_{slip}(\theta,\phi)\sin\theta\mathrm{d}\theta \mathrm{d}\phi. \label{eq-angular-velocity} \end{equation} Using the short-hand $\langle \dots \rangle$ to denote the surface average, we can write these as \begin{equation} \VVb^\alpha = - \langle\vvb^\alpha_{slip}\rangle, \quad \omb^\alpha = -\frac{3}{2R} \langle \hat{\rrb} \wedge \vvb^\alpha_{slip}\rangle. \label{eq-both-velocities} \end{equation} Due to the linearity of Stokes flow and of the chemical problem, the contributions from substrate (s) and product (p) can be calculated separately and added up at the end to give the total linear and angular velocities as \begin{equation} \VVb = \VVb^s + \VVb^p, \quad \omb = \omb^s + \omb^p. \end{equation} In the simple case of a particle with uniform surface mobility and no surface activity, the only variations in concentration are due to the linear background gradient. If the interaction potential between the surface of the particle and the chemical species were attractive (resp.~repulsive), then this would gives rise to a negative (resp.~positive) mobility \citep[see][]{Anderson89}. Thus, the osmotic flows induced near the surface would go against (or up) the tangential chemical gradient. According to \eqref{eq-linear-velocity} this means that the particle in this case would move in the direction of increasing (resp.~decreasing) chemical concentration, as expected from a colloid which displays attractive (resp.~repulsive) interactions with the chemical species. This is the physical basis for the chemotaxis of phoretic swimmers, and the following subsections are concerned with formulating a precise mathematical description of the chemotactic mechanism. \subsection{Small Damk\"{o}hler number approximation} In its current form, the boundary condition \eqref{eq-chem-BC} at the surface of the sphere does not admit a closed-form solution for the substrate. In order to make further progress, we must make a final simplifying assumption about the relative importance of reaction and diffusion for the substrate molecules. This balance is quantified by the Damk\"{o}hler number, $\Dam$, which we define as \begin{equation} \Dam = \frac{\kappa_1 \Sigma R}{D_s} \sim \frac{|\kappa_1 \sigma(\theta) s|}{|D_s \p s/\p n|}, \end{equation} where $\kappa_1$ is the volumetric rate or reaction (in units $\mbox{L}^3\mbox{T}^{-1}$) and $\Sigma$ is the typical scale of surface activity (in units $\mbox{L}^{-2}$). In the limit of small Damk\"{o}hler number, $\Dam \ll 1$, the particle behaves as a passive sphere at leading order ($\Dam^0$) but, at first order ($\Dam^1$), it produces small disturbances in the substrate and product distributions due to the chemical reactivity of its surface. To make this statement more rigorous, we proceed with non-dimensionalising the equations of the chemical problem. For the chemical concentrations we use the scale imposed by the given background concentration of substrate, such that our dimensionless variables are $\tilde{s}=s/s_b(\rrb_c)$ and $\tilde{p}=p/s_b(\rrb_c)$. We scale lengths by the radius of the swimmer, $\tilde{r}=r/R$, and the surface activity by its typical scale, $\tilde{\sigma} = \sigma/\Sigma$. Under these scalings, the dimensionless versions of equations \eqref{eq-setup-diff}-\eqref{eq-chem-inf} are \begin{equation} \nabla^2 \tilde{s} = 0, \quad \nabla^2 \tilde{p} = 0, \end{equation} in the bulk of the fluid, with boundary conditions on the surface of the swimmer \begin{equation} \left.\frac{\p \tilde{s}}{\p n}\right|_{\tilde{r}=1} = \Dam \tilde{\sigma}(\theta) \tilde{s}|_{\tilde{r}=1}, \quad \left.\frac{\p \tilde{p}}{\p n}\right|_{\tilde{r}=1} = -\delta\Dam\tilde{\sigma}(\theta) \tilde{s}|_{\tilde{r}=1}, \label{eq-nondim-surf} \end{equation} and away from the swimmer \begin{equation} \tilde{s} \rightarrow 1 + \tilde{\rrb} \bcdot \bnabla \tilde{s}^{\infty}, \quad \tilde{p} \ttz ~~ \text{as} ~ \tilde{r}\tti. \label{eq-nondim-inf} \end{equation} In addition to the Damk\"{o}hler number, a second dimensionless parameter appears, $\delta = D_s/D_p$, which is the ratio of the two diffusivities. Due to existence of a background substrate profile, the nondimensional substrate distribution around the swimmer will include an $\mathcal{O}(1)$ term, such that $\tilde{s}=\tilde{s}_0 + \Dam \tilde{s}_1 + \mathcal{O}(\Dam^2)$. This can be seen from the boundary conditions \eqref{eq-nondim-surf} and \eqref{eq-nondim-inf}. However, since there was no product in the ambient medium to begin with, the scale of the product distribution is determined by the chemical reactions at the surface of the swimmer. In the small Damk\"{o}hler number approximation the swimmer is only weakly reactive, which means that the leading-order product distribution appears only at $\mathcal{O}(\Dam)$, such that $\tilde{p}=\Dam \tilde{p}_1 + \mathcal{O}(\Dam^2)$. Due to linearity of the Stokes equations, the contributions from substrate and product add up to give the total linear and angular velocities of the swimmer. Therefore, if we want to include all terms of equal importance for the motion of the swimmer, we must calculate the substrate and product concentrations to the same level of accuracy in our Damk\"{o}hler number expansion. In the current paper we work out the solution up to $\mathcal{O}(\Dam)$, which brings the first non-trivial contributions to the final answer and also captures the expected phenomenology. Returning to dimensional variables, we will approximate the substrate concentration to be $s \simeq s_0 + s_1$, where $s_0$ is the solution of the leading-order problem \begin{equation} \nabla^2 s_0 = 0, \quad \left.\frac{\p s_0}{\p n}\right|_{r=R} = 0, \quad s_0 \rightarrow s_b(\rrb_c) + \rrb \bcdot \bnabla s^{\infty} ~ \text{as} ~ r\tti, \label{eq-substrate-leading} \end{equation} while $s_1$ is the solution of the first-order problem \begin{equation} \nabla^2 s_1 = 0, \quad \left.\frac{\p s_1}{\p n}\right|_{r=R} = \frac{\kappa_1 \sigma(\theta)}{D_s}\left.s_0\right|_{r=R}, \quad s_1 \ttz ~ \text{as} ~ r\tti. \label{eq-substrate-first} \end{equation} Similarly, we approximate the product concentration to be $p \simeq p_1$, where $p_1$ is the solution of the first-order problem \begin{equation} \nabla^2 p_1 = 0, \quad \left.\frac{\p p_1}{\p n}\right|_{r=R} = -\frac{\kappa_1 \sigma(\theta)}{D_p}\left.s_0\right|_{r=R}, \quad p_1 \ttz ~ \text{as} ~ r\tti. \label{eq-product-first} \end{equation} Notice that the first-order problems for the substrate and product are the same, up to a factor of $-D_s/D_p$, so we only need to solve for one of them. \newpage \subsection{Leading-order effects} \label{subsec-leading} Recall that our $z$-axis is the axis of symmetry of the swimmer. Without loss of generality, we may then choose our $y$-axis such that the substrate gradient lies in the $yz$-plane, that is $\bnabla s^\infty = \nabla s_y^\infty \hat{\yyb} + \nabla s_z^\infty \hat{\zzb}$. The leading-order substrate concentration can be written as the sum of the background concentration plus a perturbation which satisfies Laplace's equation and decays at infinity. The general solution to Laplace's equation can be written classically as a sum of spherical harmonics $Y_l^m(\theta,\phi) \propto e^{im\phi}P_l^m(\cos\theta)$, where $P_l^m$ is an associated Legendre polynomial, multiplied by an appropriate radial component. For our perturbation to decay at infinity this radial component has to be a multiple of $r^{-(l+1)}$, so we can write \begin{equation} s_0(r,\theta,\phi) = s_b (\rrb_c) + \mathbf{r} \bcdot \bnabla s^\infty + \sum_{l=0}^{\infty} \sum_{m=-l}^{l} f_l^m r^{-(l+1)}Y_l^m(\theta,\phi). \end{equation} This expression can be simplified further if we take into account that the zero-flux boundary condition from \eqref{eq-substrate-leading} can only excite the spherical harmonics of order $m=0$ and $m=\pm 1$, which are already present in the solution due to the form of the concentration at infinity. Due to our choice of $y$-axis, the $m=\pm 1$ components must only lead to dependence on $\sin\phi$ and not $\cos\phi$, so the substrate concentration has at most the form \begin{equation} s_0(r,\theta,\phi) = s_b (\rrb_c) + \mathbf{r} \bcdot \bnabla s^\infty + \sum_{l=0}^{\infty} \left(\frac{R}{r}\right)^{l+1}\left[a_l P_l(\cos\theta) + b_l P_l^1(\cos\theta)\sin\phi\right]. \end{equation} Imposing the zero-flux boundary condition from \eqref{eq-substrate-leading}, we find the coefficients \begin{equation} a_1 = \frac{\nabla s_z^\infty R}{2}, \quad b_1 = -\frac{\nabla s_y^\infty R}{2}, \end{equation} and $a_l, b_l = 0$ for $l\neq 1$, which lead to the following substrate distribution \begin{equation} s_0 = s_b (\rrb_c) + \rrb \bcdot \bnabla s^\infty + \frac{R^3}{2r^2}\left(\nabla s_z^\infty \cos\theta + \nabla s_y^\infty\sin\theta\sin\phi\right). \end{equation} Notably, this expression can be simplified into \begin{equation} s_0 = s_b (\rrb_c) + \rrb \bcdot \bnabla s^\infty \left( 1 + \frac{1}{2} \left(\frac{R}{r}\right)^3\right). \label{eq-substrate-concentration} \end{equation} In order to compute the slip velocity, $\vvb^{s,0}_{slip}$, due to the leading-order substrate profile, we have to evaluate \begin{equation} (\IIb-\hat{\rrb}\hat{\rrb}) \bcdot \bnabla s_0 = \frac{1}{r} \frac{\p s_0}{\p \theta} \skew3\hat{\thb} + \frac{1}{r \sin\theta} \frac{\p s_0}{\p \phi} \skew3\hat{\phb} \end{equation} on the surface of the swimmer. Using \eqref{eq-slip-velocity}, we obtain \begin{equation} \vvb^{s,0}_{slip} = \frac{3\mu_s(\theta)}{2} \left[(- \nabla s_z^\infty \sin\theta + \nabla s_y^\infty \cos\theta\sin\phi) \skew3\hat{\thb} + \nabla s_y^\infty \cos\phi \ \skew3\hat{\phb} \right]. \label{eq1-solute} \end{equation} Then, by averaging this slip velocity over the surface of the spherical swimmer, we determine the contribution to linear velocity, $\VVb^{s}_0$, coming from the leading-order substrate problem as \begin{equation} \VVb^{s}_0 = -\left(\mu_{s0}+\frac{1}{10}\mu_{s2}\right) \bnabla s^\infty + \frac{3}{10}\mu_{s2} \ \skew3\hat{\zzb}\skew3\hat{\zzb} \bcdot \bnabla s^\infty. \label{eq-passive-trans} \end{equation} The calculations leading to this result are detailed in the first subsection of appendix \ref{appA}. Although this result is valid for a general surface mobility, $\mu_s(\theta)$, only two Legendre modes of the mobility contribute to the final results, namely those adjacent to the forcing coming from the linear gradient. Furthermore, our result is in agreement with the linear velocity calculated by Anderson for a passive sphere placed in a non-uniform field \citep[equation (37a) from][]{Anderson89}. In our case that field is the concentration of substrate molecules, and the sphere is passive, at leading order, due to our small Damk\"{o}hler number approximation. Similarly, one has to compute the average of $\hat{\rrb} \wedge \vvb^{s,0}_{slip}$ over the surface of the spherical swimmer in order to determine the angular velocity contribution, $\omb^{s}_0$, resulting from the leading-order substrate problem \begin{equation} \omb^{s}_0 = -\frac{3\mu_{s1}}{4R} \ \skew3\hat{\zzb} \wedge \bnabla s^\infty. \label{eq-passive-ang} \end{equation} The calculations leading to this are detailed in the second subsection of appendix \ref{appA}. This result agrees with the angular velocity calculated by Anderson for a passive sphere placed in a non-uniform field \citep[equation (37b) from][]{Anderson89}. Note again that this result is valid for a general surface mobility $\mu_s(\theta)$, but only one mode contributes to the reorientation of a passive swimmer. This mode corresponds to a linear dependence of mobility on the cosine of the polar angle, with opposite and equal in magnitude mobilities at the two poles. By symmetry we would expect this type of swimmer to align with the chemical gradient, but we can also predict the direction of rotation using a simple physical argument. Following the negative and positive chemotactic response of swimmers with uniform positive and negative mobility, respectively, as discussed at the end of section \S\ref{sec-hydro}, we would expect the pole of negative mobility to pull towards the higher concentration of substrate and the pole of positive mobility to pull towards the lower concentration. This will then make the swimmer rotate until its axis of symmetry is aligned with the substrate gradient and the pole of negative mobility faces higher concentration, which is precisely the dynamics captured by \eqref{eq-passive-ang}. Since the two poles are equally strong, the tug-of-war cannot be won by either opponent and this mode has zero linear velocity, in agreement with \eqref{eq-passive-trans}. \subsection{First-order effects} \label{subsec-first} We now turn our attention to finding the corrections to first order in Damk\"{o}hler number for the substrate problem. In order to satisfy Laplace's equation and the decay at infinity, the first-order substrate concentration must have the general form \begin{equation} s_1(r,\theta,\phi) = \sum_{l=0}^{\infty} \sum_{m=-l}^{l} F_l^m r^{-(l+1)}Y_l^m(\theta,\phi), \end{equation} but the normal-flux boundary condition from \eqref{eq-substrate-first} can only excite modes with either no $\phi$-dependence or $\sin\phi$-dependence which appear in the leading-order substrate concentration from \eqref{eq-substrate-concentration}. This means that the first-order substrate concentration can be simplified at most to the sum \begin{equation} s_1(r,\theta,\phi) = \sum_{l=0}^{\infty} \left(\frac{R}{r}\right)^{l+1} \left[A_l P_l(\cos\theta) - B_l \sin\theta P'_l(\cos\theta) \sin\phi \right], \label{eq-s-1-general-form} \end{equation} where we have used the fact that $P_l^1(\cos\theta) = -\sin\theta P'_l(\cos\theta)$. The coefficients $A_l$ and $B_l$ can be determined from the normal-flux boundary condition in \eqref{eq-substrate-first}. These calculations are detailed in the third subsection of appendix \ref{appA}, and lead to the following expressions \begin{equation} A_l = -\frac{\kappa_1 R}{(l+1)D_s} \left[s_b (\rrb_c) \sigma_l + \frac{3}{2} \nabla s_z^\infty R \left(\frac{l+1}{2l+3} \sigma_{l+1} + \frac{l}{2l-1}\sigma_{l-1}\right)\right], \end{equation} \begin{equation} B_l = -\frac{3\kappa_1 R^2 \nabla s_y^\infty}{2(l+1)D_s} \left(\frac{\sigma_{l+1}}{2l+3} - \frac{\sigma_{l-1}}{2l-1} \right). \end{equation} Since the first-order product problem \eqref{eq-product-first} is the same as the first-order substrate problem \eqref{eq-substrate-first}, except for a factor of $-D_s/D_p$, we can use our results for $s_1$ and state that the first-order product concentration $p_1$ is given by \begin{equation} p_1 = \sum_{l=0}^{\infty} \left(\frac{R}{r}\right)^{l+1} \left[C_l P_l(\cos\theta) - D_l \sin\theta P'_l(\cos\theta) \sin\phi \right], \end{equation} where the coefficients $C_l$ and $D_l$ can be written as \begin{eqnarray} C_l &=& \frac{\kappa_1 R}{(l+1)D_p} \left[s_b (\rrb_c) \sigma_l + \frac{3}{2} \nabla s_z^\infty R \left(\frac{l+1}{2l+3} \sigma_{l+1} + \frac{l}{2l-1}\sigma_{l-1}\right)\right],\\ D_l &=& \frac{3\kappa_1 R^2 \nabla s_y^\infty}{2(l+1)D_p} \left(\frac{\sigma_{l+1}}{2l+3} - \frac{\sigma_{l-1}}{2l-1} \right). \end{eqnarray} We next need to calculate the contribution of the first-order substrate concentration to the slip velocity. Using \eqref{eq-slip-velocity} and \eqref{eq-s-1-general-form}, we find that \begin{multline} \vvb^{s,1}_{\text{slip}} = \frac{\mu_s(\theta)}{R} \sum_{l=0}^{\infty} \left(- A_l\sin\theta P'_l(\cos\theta) -B_l \cos\theta P'_l(\cos\theta) \sin\phi\right. \\ + \left.B_l \sin^2\theta P''_l(\cos\theta)\sin\phi\right)\skew3\hat{\thb} + \frac{\mu_s(\theta)}{R} \sum_{l=0}^{\infty} \left(-B_l P'_l(\cos\theta)\cos\phi\right) \skew3\hat{\phb}. \end{multline} The last two subsections in appendix \ref{appA} present the detailed calculations involved in averaging $\vvb^{s,1}_{\text{slip}}$ and $\hat{\rrb} \wedge \vvb^{s,1}_{\text{slip}}$ over the surface of the spherical swimmer, in order to calculate the contributions of the first-order substrate problem to the linear and angular velocities of the swimmer. The final results are that the first-order substrate problem contributes \begin{multline} \VVb^{s}_1 = -\frac{\kappa_1 s_b (\rrb_c)}{D_s} \sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\sigma_l \left(\frac{\mu_{s,l+1}}{2l+3}-\frac{\mu_{s,l-1}}{2l-1}\right) \skew3\hat{\zzb} \\ - \frac{3 \kappa_1 R}{4 D_s}\sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\left(\frac{\sigma_{l+1}}{2l+3} - \frac{\sigma_{l-1}}{2l-1} \right)\left(\frac{l+1}{2l-1}\mu_{s,l-1}+\frac{l}{2l+3}\mu_{s,l+1}\right) \bnabla s^\infty \\ + \frac{3 \kappa_1 R}{4 D_s}\sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\left(\frac{3(l+1)\sigma_{l+1}\mu_{s,l-1}}{(2l+3)(2l-1)}-\frac{(l+2)\sigma_{l+1}\mu_{s,l+1}}{(2l+3)^2}\right. \\ +\left. \frac{(l-1)\sigma_{l-1}\mu_{s,l-1}}{(2l-1)^2} - \frac{3l\sigma_{l-1}\mu_{s,l+1}}{(2l-1)(2l+3)}\right) \skew3\hat{\zzb}\skew3\hat{\zzb}\bcdot\bnabla s^\infty. \end{multline} to the linear velocity of the swimmer, and \begin{equation} \omb^{s}_1= -\frac{9\kappa_1}{8 D_s} \sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\mu_{sl} \left(\frac{\sigma_{l+1}}{2l+3}-\frac{\sigma_{l-1}}{2l-1}\right) ~ \skew3\hat{\zzb} \wedge \bnabla s^\infty \end{equation} to its angular velocity. The contributions from the first-order product problem will be the same as the ones above, with an additional factor of $-D_s/D_p$ which carries through from the solution of the chemical problem, and a change of subscript on the mobility coefficients $\mu_{sl} \mapsto \mu_{pl}$. \subsection{Summary of results} In summary, we may write the instantaneous linear and angular velocities of the phoretic swimmer, when its centre is at position $\rrb_c$, in the compact form \begin{eqnarray} \VVb &=& U(\rrb_c) \skew3\hat{\zzb}+\alpha \bnabla s^\infty + \beta \ \skew3\hat{\zzb}\skew3\hat{\zzb}\bcdot\bnabla s^\infty,\label{summary-linear}\\ \omb &=& \Phi \ \skew3\hat{\zzb} \wedge \bnabla s^\infty \label{summary-angular} . \end{eqnarray} Due to the linearity of the Stokes equations, the total linear and angular velocities of the phoretic swimmer are simply the sum of contributions from the substrate and the product, at leading and first order. Using the results from the previous subsections, we deduce that the coefficients appearing in \eqref{summary-linear} and \eqref{summary-angular} are given by \begin{equation} U(\rrb_c) = \kappa_1 s_b (\rrb_c) \sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\sigma_l \left[\frac{1}{2l+3}\left(\frac{\mu_{p,l+1}}{D_p}-\frac{\mu_{s,l+1}}{D_s}\right)-\frac{1}{2l-1}\left(\frac{\mu_{p,l-1}}{D_p}-\frac{\mu_{s,l-1}}{D_s}\right)\right], \label{eqn-U} \end{equation} \begin{multline} \alpha = -\left(\mu_{s0}+\frac{1}{10}\mu_{s2}\right) + \frac{3\kappa_1 R}{4}\sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\left(\frac{\sigma_{l+1}}{2l+3} - \frac{\sigma_{l-1}}{2l-1} \right) \\ \times\left[\frac{l+1}{2l-1}\left(\frac{\mu_{p,l-1}}{D_p}-\frac{\mu_{s,l-1}}{D_s}\right)\right. +\left.\frac{l}{2l+3}\left(\frac{\mu_{p,l+1}}{D_p}-\frac{\mu_{s,l+1}}{D_s}\right)\right], \label{eqn-alpha-nu} \end{multline} \begin{multline} \beta = \frac{3\mu_{s2}}{10} - \frac{3\kappa_1 R}{4}\sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right) \left[\left(\frac{3(l+1)\sigma_{l+1}}{(2l+3)(2l-1)} + \frac{(l-1)\sigma_{l-1}}{(2l-1)^2}\right) \left(\frac{\mu_{p,l-1}}{D_p}-\frac{\mu_{s,l-1}}{D_s}\right) \right. \\ \left. -\left(\frac{(l+2)\sigma_{l+1}}{(2l+3)^2} + \frac{3l\sigma_{l-1}}{(2l-1)(2l+3)}\right)\left(\frac{\mu_{p,l+1}}{D_p}-\frac{\mu_{s,l+1}}{D_s}\right)\right], \label{eqn-beta-mu} \end{multline} \begin{equation} \Phi = -\frac{3\mu_{s1}}{4R} + \frac{9\kappa_1}{8} \sum_{l=1}^{\infty} \left(\frac{l}{2l+1}\right)\left(\frac{\mu_{pl}}{D_p}-\frac{\mu_{sl}}{D_s}\right) \left(\frac{\sigma_{l+1}}{2l+3}-\frac{\sigma_{l-1}}{2l-1}\right).\label{eqn-Omega} \end{equation} To the best of our knowledge, the setup and modelling assumptions used in deriving these results follow those used by \citet{Saha2014}. Our final equations, (\ref{eqn-U})-(\ref{eqn-Omega}), represent an extension of the instantaneous laws of motion published in \citet{Saha2014}, to a phoretic swimmer with the most general axisymmetric surface activity $\sigma(\theta)$ and surface mobilities $\mu_s(\theta)$, $\mu_p(\theta)$. It is also important to note that our solution includes the first-order contributions from the substrate, an effect which had been neglected by \citet{Saha2014}. \subsection{Janus spherical swimmer} \begin{figure} \centering \includegraphics[width=.5\textwidth]{fig2.pdf} \caption{Schematic representation of a Janus spherical swimmer with surface properties $\{\sigma^U,\mu_s^U,\mu_p^U\}$ on the upper hemisphere and $\{\sigma^L,\mu_s^L,\mu_p^L \}$ on the lower hemisphere. The linear and angular velocities of this phoretic swimmer are given by equations \eqref{eq-Janus-1}-\eqref{eq-Janus-4}.} \label{fig:Janus} \end{figure} Because our extended laws of motion allow for arbitrary axisymmetric surface properties, we can apply our results to one of the most common designs of phoretic swimmers, namely that of a Janus sphere. This type of swimmer, depicted in figure~\ref{fig:Janus}, has uniform surface properties on each of its two hemispheres such that the surface activity is given by \begin{equation}\label{Jsigma} \sigma(\theta) = \left\{ \begin{array}{ll} \sigma^U, & 0 \leq \theta < \pi/2, \\[2pt] \sigma^L, & \pi/2 < \theta \leq \pi, \end{array} \right. \end{equation} where the superscript denotes the lower ($L$) and upper ($U$) halves of the swimmer. We also assume similar spatial distributions of the two mobilities $\mu_s(\theta)$ and $\mu_p(\theta)$ (see full notation in figure~\ref{fig:Janus}). With the help of the integral identity \eqref{eq-half-integral}, we can project \eqref{Jsigma} onto the space of Legendre polynomials, and we find that \begin{equation} \sigma_0 = \frac{1}{2}(\sigma^U+\sigma^L), \end{equation} but that all other even modes vanish \begin{equation} \sigma_{2k} = 0, \quad k \geq 1. \end{equation} For the odd modes, we have \begin{equation} \sigma_{2k+1}=(-1)^{k}(\sigma^U-\sigma^L)\frac{(4k+3)(2k+1)!!}{(4k+2)(2k+2)!!}, \end{equation} and equivalent identities for the mobility coefficients $\mu_{sl}$ and $\mu_{pl}$. Upon substituting these expressions into our results for the linear and angular velocity of a general axisymmetric swimmer, we find that the parameters characterising the motion of a Janus sphere in a uniform chemical gradient, (\ref{eqn-U})-(\ref{eqn-Omega}), are given by \begin{eqnarray} U_{\text{Janus}}&=&\frac{\kappa_1 s_b}{8}(\sigma^L-\sigma^U)\left(\frac{\mu_p^L+\mu_p^U}{D_p}-\frac{\mu_s^L+\mu_s^U}{D_s} \right), \label{eq-Janus-1}\\ \alpha_{\text{Janus}} &=& -\frac{1}{2}(\mu_s^U+\mu_s^L)-\frac{\kappa_1 R}{8}(\sigma^U+\sigma^L)\left(\frac{\mu_p^L+\mu_p^U}{D_p}-\frac{\mu_s^L+\mu_s^U}{D_s} \right)\nonumber \\ &~& -\frac{3\gamma\kappa_1 R}{4}(\sigma^U-\sigma^L)\left(\frac{\mu_p^U-\mu_p^L}{D_p}-\frac{\mu_s^U-\mu_s^L}{D_s} \right), \quad\quad\\ \beta_{\text{Janus}} &=& -\frac{3\gamma\kappa_1 R}{4}(\sigma^U-\sigma^L)\left(\frac{\mu_p^U-\mu_p^L}{D_p}-\frac{\mu_s^U-\mu_s^L}{D_s} \right),\\ \Phi_{\text{Janus}}&=&\frac{9}{16R}(\mu_s^L-\mu_s^U)+\frac{9\kappa_1}{64}(\sigma^L+\sigma^U)\left(\frac{\mu_p^L-\mu_p^U}{D_p}-\frac{\mu_s^L-\mu_s^U}{D_s} \right), \label{eq-Janus-4} \end{eqnarray} with swimming kinematics as in \eqref{summary-linear} and \eqref{summary-angular}. Note that the coefficients $\alpha_{\text{Janus}}$ and $\beta_{\text{Janus}}$ are more easily calculated from equation \eqref{eq-needed-for-janus} than from equations \eqref{eqn-alpha-nu}-\eqref{eqn-beta-mu}, and that they involve a numerical factor \begin{equation} \gamma = \sum_{k=1}^\infty k\left(\frac{(2k-1)!!}{(2k-1)(2k+2)!!}\right)^2 \approx 0.0336. \end{equation} Our result for $U_{\text{Janus}}$ agrees with \citet{Golestanian07} if we equate their concentration-independent surface activity $\alpha(\theta)$ to the component of our surface activity which is independent of the background gradient, i.e.~$\kappa_1 s_b \sigma(\theta)$. Instead of a single chemical mobility to diffusivity ratio $\mu/D$, we also have an effective value, $(\mu/D)_{\text{eff}} = \mu_p/D_p - \mu_s/D_s$, due to the interplay of the two chemical species in our problem. The simple expressions in \eqref{eq-Janus-1}-\eqref{eq-Janus-4} provide an estimate for the linear and angular velocity of the swimmer based on information about its surface properties alone, and could thus be used in the design and fabrication of phoretic Janus particles. Note that, at leading order, the only non-zero parameters are $\alpha_{\text{Janus}}$ and $\Phi_{\text{Janus}}$. Thus, the dominant behaviour of a spherical Janus swimmer consists in a constant linear velocity along the direction of the chemical gradient, with the orientation of the swimmer being irrelevant. In order to fabricate an efficient chemotactic swimmer of this type, one must simply ensure that $\mu_s^U+\mu_s^L$ has the desired sign for positive or negative chemotaxis. The simplest realisation of this is a swimmer with uniform mobility. However, if it happens that $\mu_s^U+\mu_s^L=0$ then the chemotactic behaviour of the Janus sphere will depend on first-order effects coming from the chemical reactions. At the end of this subsection we discuss the different possible strategies for chemotaxis. For a swimmer that responds to both substrate and product molecules it is possible, in theory, to choose the surface properties $\{\sigma^L,\sigma^U,\mu_s^L,\mu_s^U,\mu_p^L,\mu_p^U \}$ in order to obtain any set of prescribed values for $\{ U_{\text{Janus}},\alpha_{\text{Janus}},\beta_{\text{Janus}},\Phi_{\text{Janus}}\}$, although this may be hard to achieve in practice. Surprisingly, the same cannot be said for a swimmer that responds only to product molecules and has $\mu_s =0$ even though there are four degrees of freedom and four target parameters. This is due to the specific combinations in which the four degrees of freedom appear in the expressions for $\{ U_{\text{Janus}},\alpha_{\text{Janus}},\beta_{\text{Janus}},\Phi_{\text{Janus}}\}$. We find that a Janus sphere with $\mu_s =0$ is subject to the following dependence relation between the four parameters describing its motion \begin{equation} \alpha_{\text{Janus}}=\frac{16\gamma R^2 U_{\text{Janus}}\Phi_{\text{Janus}}}{3s_b \beta_{\text{Janus}}}+\beta_{\text{Janus}}. \end{equation} This represents a constraint on the range of behaviour that can be obtained with a swimmer that has zero mobility with respect to the substrate molecules. For a swimmer that responds to the product molecules, and possibly the substrate molecules as well, we can distinguish different strategies for chemotaxis depending on the quantitative details of the problem. If the chemical gradient is sufficiently strong, the linear velocity of the swimmer will be dominated by the chemotactic `sedimentation' terms, $\alpha \bnabla s^\infty + \beta \ \skew3\hat{\zzb}\skew3\hat{\zzb}\bcdot\bnabla s^\infty$. In this case it is better to opt for a simpler swimmer with uniform surface properties, which ensures that $\beta_{\text{Janus}}=0$ and the swimmer undergoes direct chemotactic sedimentation along the gradient, with the sign of $\alpha_{\text{Janus}}$ determining the sense of chemotaxis. If, on the other hand, the chemical gradient is sufficiently weak, then the linear velocity of the swimmer will be dominated by propulsion along its axis of symmetry, i.e.~$U \skew3\hat{\zzb}$. In this case, an effective chemotactic swimmer must be able to align with the chemical gradient and then propel along its axis, which corresponds to non-zero values for both $ U_{\text{Janus}}$ and $\Phi_{\text{Janus}}$. In this situation, it is crucial that the two hemispheres of the Janus sphere have different surface activities as well as different surface mobilities. If $ U_{\text{Janus}}$ and $\Phi_{\text{Janus}}$ have the same sign then the swimmer will perform positive chemotaxis, whereas if they have opposite signs the swimmer will perform negative chemotaxis. For intermediate values of the chemical gradient, the optimal strategy for chemotaxis will consist of a combination of the two mechanisms. \section{Long-time behaviour: Artificial chemotaxis of a population of non-interacting particles} \label{section-3} \subsection{Motivation} Having established the physical law which describes the instantaneous movement of an individual swimmer, we now want to develop an understanding of the long-time behaviour of a phoretic swimmer, or equivalently that of a large number of non-interacting swimmers. Although we neglect the interaction between particles, we focus on understanding the effects of another essential ingredient - stochasticity. Thermal noise plays an important role on the scale of phoretic swimmers, which are typically a few microns in diameter. The orientation of the swimmers will fluctuate as a result of thermal energy in the surrounding fluid, leading to diffusive behaviour characterised by a rotational diffusivity $D_r$ given by the Einstein-Smoluchowski relation \citep{Einstein1905,Smolu1906} \begin{equation} D_r = \frac{k_B T}{8\pi\eta R^3}, \end{equation} where $k_B$ is Boltzmann's constant, $T$ is the temperature, and $\eta$ is the dynamic viscosity of the fluid. The physical basis for a macroscopic continuum model lies in the competition between the stochastic effect of rotational diffusion and the deterministic processes by which the swimmer propels and aligns with the external chemical gradient. If we consider the system on a timescale much larger than the characteristic timescale for rotational diffusion, such that we are essentially averaging over fluctuations in the swimmer orientation, we would expect to see the spatial distribution of swimmers evolve in time according to an effective advection-diffusion equation. The purpose of this section is to quantify this process rigorously. The first modelling approach we considered was based on the classical continuum model for gyrotactic swimming micro-organisms proposed by \citet{Pedley90}. This is a simple and intuitive model, but its simplicity comes at a cost. The authors use a phenomenological definition of the diffusivity tensor which requires the introduction of a direction correlation time, also called relaxation time by some authors \citep[see][]{Batchelor76}. This quantity is not calculated explicitly in their paper, and would have to be estimated from experiments or numerical simulations. Furthermore, the authors assume that the direction correlation time is isotropic, an assumption probably not suitable for a swimmer that has a preferential orientation. Therefore, we turned instead our attention to the theory of generalised Taylor dispersion (GTD) that Frankel and Brenner used for the study of orientable Brownian particles \citep{Frank91, Frank93} and which has been successfully applied to gyrotactic swimming micro-organisms in more recent years \citep[][]{Hill02, Frank03, Bees12}. By addressing the coupling between dynamics in the orientational space and in the physical space, GTD theory provides a rigorous approach to modelling collections of orientable particles. Fundamentally, generalised Taylor dispersion theory uses a moment expansion for the probability density of swimmers, in a similar way to \citet{golestanian2012,pohl2014,bickel2014}. What is novel in the present paper is that the chemical reactions considered in Section 2 lead to a more complex instantaneous behaviour than previously investigated, which brings a further degree of freedom to our continuum model (what we later call the ``indirect chemotactic index") and new phenomenology along with it. Furthermore, the diffusivity tensor is isotropic in all of the above references, because none of them possess the necessary combination of active reorientation and orientation-dependent velocity which promotes the emergence of anisotropic diffusion, as is the case with the phoretic swimmers in our paper. \subsection{Setup} We consider a uniform substrate gradient, as in the derivation of the instantaneous behaviour, and no background flow. Since the orientation of the swimmer changes in time, it no longer makes sense to define one of the principal Cartesian axes along the axis of symmetry of the swimmer. Instead, we take the positive $z$-axis to be aligned with the fixed chemical gradient such that $\bnabla s^\infty = |\bnabla s^\infty| \kkb$. We denote the position of the swimmer by $\xxb$ and its axis of symmetry using the unit vector $\nnb = (\sin\theta \cos\phi, \sin\theta\sin\phi, \cos\theta)$, where $\theta$ and $\phi$ are the usual polar and azimuthal angles from spherical coordinates. In this notation, the linear velocity of the phoretic swimmer in \eqref{summary-linear} can be written as \begin{equation} \VVb(\xxb,\nnb) = U(\xxb)\big ( \nnb + \nu(\xxb) \kkb + \mu(\xxb) \nnb\nnb\bcdot\kkb \big ), \label{linear-velocity} \end{equation} where we have introduced two dimensionless parameters \begin{equation} \nu(\xxb) \equiv \frac{\alpha |\bnabla s^\infty|}{U(\xxb)} \quad \text{and} \quad \mu(\xxb) \equiv \frac{\beta |\bnabla s^\infty|}{U(\xxb)}\cdot \end{equation} We call $\nu$ and $\mu$ the direct and indirect chemotactic indices, respectively, since they measure the relative effectiveness of direct chemotactic sedimentation, $\alpha \bnabla s^\infty$, and indirect chemotactic sedimentation, $\beta \nnb\nnb\bcdot\bnabla s^\infty$, against active propulsion, $U(\xxb)\nnb$. Note that the relative magnitude $\nu/\mu$ of direct to indirect chemotactic sedimentation determines the angle at which the phoretic swimmer sediments relative to the chemical gradient. This is similar to the Stokes flow sedimentation of an elongated object, such as a rod, under the action of gravity. In Stokes flow the sedimentation angle is given by the geometry (and the drag coefficients) of the elongated body, while in our case it is a function of the chemical properties of the spherical swimmer. Lastly, we need to consider the mechanism by which the swimmer aligns with the chemical gradient. The orientation of the swimmer evolves in time according to \begin{equation} \dot{\nnb} = \omb \wedge \nnb, \end{equation} and using \eqref{summary-angular} we arrive at the following reorientation law \begin{equation} \dot{\nnb} = \Omega [\kkb - (\kkb\bcdot\nnb) \ \nnb], \label{reorient-law} \end{equation} where $\Omega = \Phi |\bnabla s^\infty|$. This prompts us to define a third dimensionless parameter \begin{equation}\label{RPe} \lambda\equiv \frac{\Omega}{D_r}, \end{equation} which is a rotational P\'{e}clet number measuring the relative importance of active reorientation to rotational diffusion. \subsection{Generalised Taylor dispersion theory} Although generalised Taylor dispersion theory is now a classical tool, it is useful to clearly introduce it and define the differential operator involved in it. We start from the conservation equation for $P(\nnb,\xxb,t)$, the probability density function of finding a swimmer with orientation $\nnb$ at position $\xxb$ at time $t$, which is satisfies the conservation equation \begin{equation} \frac{\p P}{\p t} + \bnabla_r \bcdot \big (\VVb(\xxb,\nnb) P - D_t \bnabla_r P \big ) + \bnabla_n \bcdot \big (\dot{\nnb} P - D_r \bnabla_n P \big ) = 0, \label{eq-full-cons} \end{equation} where the differential operators for physical space and orientational space are \begin{equation} \bnabla_r \equiv \boldsymbol{i} \frac{\p}{\p x_1} + \boldsymbol{j} \frac{\p}{\p x_2} + \boldsymbol{k} \frac{\p}{\p x_3}, \quad \bnabla_n \equiv \skew3\hat{\thb} \frac{\p}{\p \theta} + \skew3\hat{\phb} \frac{1}{\sin\theta} \frac{\p}{\p \phi}, \end{equation} respectively. The law in \eqref{eq-full-cons} says that the local rate of change in the distribution of swimmers is due to the flux of swimmers in physical space (advection and translational diffusion) and the flux of swimmers in orientational space (advection and rotational diffusion). Since we are interested in the spatial distribution of the particles, we define the particle density \begin{equation} \rho(\xxb,t) \equiv \int_{S_2} P(\nnb,\xxb,t)~\mathrm{d}^2\nnb. \end{equation} Upon careful consideration of the moments of $P(\nnb,\xxb,t)$, Frankel and Brenner showed that at the macroscopic level, and on timescales $t \gg D_r^{-1}$, a population of non-interacting orientable particles obeys the advection-diffusion equation \begin{equation} \frac{\p \rho}{\p t} + \bnabla_{\rrb}\bcdot\big (\VVb_s \rho - \boldsymbol{D} \bcdot \bnabla_{\rrb}\rho\big ) = 0 \label{eq:gtdmacro}, \end{equation} where $\VVb_s$ is the mean swimming velocity and $\boldsymbol{D}$ is the diffusivity tensor. The mean swimming velocity is given by \begin{equation} \VVb_s (\xxb) = \int_{S_2}f(\nnb) \VVb(\xxb,\nnb)~ \mathrm{d}^2\nnb, \label{eq-GTDT-Vs-defn} \end{equation} whereas the diffusivity tensor is the sum of contributions from passive translational diffusion ($\boldsymbol{D}_{\text{pass}}$) and active phoretic motion ($\boldsymbol{D}_{\text{act}}$) as \begin{equation} \boldsymbol{D}(\xxb) = \boldsymbol{D}_{\text{pass}} + \boldsymbol{D}_{\text{act}}(\xxb). \end{equation} The passive part of the diffusivity tensor, $\boldsymbol{D}_{\text{pass}}$, is simply the average of the translational diffusivity tensor, $\boldsymbol{D}_t(\nnb)$, over the space of orientations \citep[see][]{Frank93}. For a spherical swimmer it is reasonable to assume that Brownian noise acts isotropically, and so $\boldsymbol{D}_{\text{pass}} = D_t \boldsymbol{I}$ with $D_t$ being the translational diffusion coefficient. Recently it has been shown by \citet{Agudo2018} that, at the nano scale, spatial variations in the translational diffusion coefficient of enzymes can compete with phoretic effects, and it may well be possible that the catalytic coating of phoretic swimmers could lead to enchanced diffusion coefficients as the swimmer moves through the chemical gradient. However, due to a lack of documentation on the enhanced diffusion capabilities of a microscopic phoretic swimmer, and due to the limitations of generalised Taylor dispersion theory, to consider spatial variations of $D_t$ would go beyond the scope of this paper. With the assumption that $\boldsymbol{D}_{\text{pass}}$ is constant and independent of the phoretic mechanisms for motion, we will focus from here onwards on quantifying the active diffusivity tensor given by \begin{equation} \boldsymbol{D}_{\text{act}}(\xxb) = \int_{S_2} [\bbb(\xxb,\nnb) \VVb(\xxb,\nnb)]^{\text{sym}} \mathrm{d}^2\nnb, \label{eq-GTDT-D-defn} \end{equation} where $[\dots ]^{\text{sym}}$ denotes the symmetric part of the tensor. The expressions for $\VVb_s (\xxb)$ and $\boldsymbol{D}_{\text{act}}(\xxb)$ contain two new fields that have to be computed: $f(\nnb)$ is the long-time steady distribution of the particle orientation, and $\bbb(\xxb,\nnb)$ represents the long-time relative displacement of a particle given that its instantaneous orientation is $\nnb$, compared to its expected position over all possible orientations. For a formal definition of these quantities we need introduce the linear operator \begin{equation} \Lin (\star) = \bnabla_{\nnb} \bcdot \left[ \dot{\nnb} (\star) - D_r \bnabla_{\nnb} (\star) \right], \end{equation} which describes the flux of a quantity in orientational space. Then $f$ and $\bbb$ are the solutions to the following equations and boundary conditions \begin{eqnarray} \Lin f &=& 0, \quad \int_{S_2}f(\nnb) \mathrm{d}^2\nnb = 1, \label{eq:f}\\ \Lin \bbb &=& f(\nnb) (\VVb(\xxb,\nnb) - \VVb_s(\xxb)), \quad \int_{S_2} \bbb ~\mathrm{d}^2\nnb = 0 \label{eq:b} \end{eqnarray} For a full derivation of these results see \citet{Frank93} or, for a more succinct account, \citet{Hill02}. To provide some physical intuition for these two equations, we observe that equation (\ref{eq:f}) corresponds to solving $\p f/\p t =0$ and therefore finding the steady-state distribution of orientations subject to a global normalisation condition. Similarly, equation (\ref{eq:b}) can be thought of as solving $\p \bbb/\p t =f(\nnb) (\VVb(\xxb,\nnb) - \VVb_s(\xxb))$ which means that the rate of change of the relative displacement is due to the relative velocity, weighted by the probability of finding a swimmer with that specific orientation. The obvious boundary condition to impose here is that, averaged over all possible orientations, the relative displacement must vanish. One important difference between our application of GTD theory to phoretic swimmers compared to previous applications to orientable Brownian particles and gyrotactic micro-swimmers is that the swimming velocity $\VVb(\xxb,\nnb)$ of phoretic swimmers depends on position as well. The classical theory still applies provided that the velocity is slowly-varying in space, such that the process of averaging over particle orientations can be carried out on timescales much larger than the diffusive timescale, $D_r^{-1}$, but much smaller than the timescale of travel to a region of significantly different swimming speed. So there must exist a timescale $\tau$, satisfying \begin{equation} D_r^{-1} \ll \tau \ll |\nabla U|^{-1}, \end{equation} on which we can average \eqref{eq-full-cons} over the space of swimmer orientations. In particular, we must have \begin{equation} |\nabla U| \ll D_r. \end{equation} In that case, the advection-diffusion equation \eqref{eq:gtdmacro} will be valid on timescales greater than or equal to $\tau$, and will involve a mean swimming velocity, $\VVb_s (\xxb)$, and a diffusivity tensor, $\boldsymbol{D}(\xxb)$, which are slowly-varying in space. Under this approximation, quantities such as $\nu(\xxb), \mu(\xxb), \VVb(\xxb,\nnb)$ and $\bbb(\xxb,\nnb)$ are also slowly-varying in space, and we will drop the explicit mention of spatial dependence from our notation. \subsection{Distribution of swimmer orientations} Since the orientation of the swimmers fluctuates over time from thermal noise, the only fixed meaningful direction in our problem is that of the chemical gradient. Therefore, quantities such as $f(\nnb)$ and $\bbb(\nnb)$ will inherit rotational symmetry around the direction of the chemical gradient (the $\kkb$ axis). For such an axisymmetric problem, the linear operator simplifies to \begin{equation} \Lin (\star) = -\frac{\Omega}{\sin\theta}\frac{\p}{\p\theta}\left(\sin^2\theta(\star)\right) - \frac{D_r}{\sin\theta}\frac{\p}{\p\theta}\left(\sin\theta\frac{\p(\star)}{\p\theta}\right). \end{equation} Then \eqref{eq:f} has a trivial first integral \begin{equation} \lambda\sin^2\theta f + \sin\theta \frac{\p f}{\p\theta}= 0 , \end{equation} where the integrating constant can be set to zero assuming that $f$ is well-behaved at the poles. The solution to this, subject to the appropriate normalisation condition, is the same as that found by \citet{Pedley90}. The long-time steady distribution of the swimmer orientation is \begin{equation} f(\nnb) = \hat{f}(\cos\theta) = \frac{\lambda e^{\lambda \cos\theta}}{4\upi \sinh\lambda}, \label{eq-equilibrium-orientation} \end{equation} where $\lambda$ is the rotational P\'{e}clet number, \eqref{RPe}. \subsection{Mean swimming velocity} From equations \eqref{linear-velocity} and \eqref{eq-GTDT-Vs-defn}, the mean swimming velocity is obtained as \begin{equation} \VVb_s = U \left( \int_{S_2}f(\nnb)\nnb \ \mathrm{d}^2\nnb + \nu \int_{S_2}f(\nnb)\kkb \ \mathrm{d}^2\nnb + \mu \int_{S_2} f(\nnb)\nnb\nnb\bcdot\kkb \ \mathrm{d}^2\nnb \right). \end{equation} We use the normalisation of $f(\nnb)$ to simplify the second integral, and average over the azimuthal angle in the other two integrals in order to get \begin{equation} \VVb_s = U \left( \int_0^\pi \hat{f}(\cos\theta)\cos\theta \sin\theta \mathrm{d}\theta + \nu + \mu \int_0^\pi \hat{f}(\cos\theta)\cos^2\theta \sin\theta \mathrm{d}\theta \right) \kkb. \end{equation} Using our previous expression for $\hat{f}(\cos\theta)$, we obtain the following analytical expression for the mean swimming velocity \begin{equation} \VVb_s = U \left[\frac{-1 + \lambda \coth\lambda}{\lambda} + \nu + \mu \left(\frac{2+\lambda^2-2\lambda\coth\lambda}{\lambda^2}\right)\right] \kkb. \end{equation} Note that the linear velocity of the swimmer relative to the mean swimming velocity is a function of $\lambda$ and $\mu$ only since \begin{equation} \VVb(\nnb) - \VVb_s = U \left[ \nnb + \mu \nnb\nnb\bcdot\kkb - \tilde{u}(\lambda,\mu) \kkb\right] \label{eq:relative_velocity}, \end{equation} where we have defined for convenience \begin{equation} \tilde{u}(\lambda,\mu) \equiv \frac{-1 + \lambda \coth\lambda}{\lambda} + \mu \left(\frac{2+\lambda^2-2\lambda\coth\lambda}{\lambda^2}\right). \end{equation} As a result, both $\bbb(\nnb)$ and the diffusivity tensor $\boldsymbol{D}_{\text{act}}$ will then only depend on the value of $\lambda$ and $\mu$. Physically, the direct chemotactic index $\nu$ corresponds to a component of velocity that is constant in time which imparts a constant drift to the phoretic swimmer without affecting the way in which the swimmer diffuses through space. \subsection{Active diffusivity tensor} The first step in calculating the active diffusivity tensor is to solve for the vector field $\bbb(\nnb)$. The expression for the relative swimming velocity from \eqref{eq:relative_velocity} together with the definition of $\bbb(\nnb)$ from equation (\ref{eq:b}) suggest that the only dependence on the azimuthal angle that can be excited in $\bbb(\nnb)$ is a $\cos\phi$ dependence in its $x$-component and a $\sin\phi$ dependence in its $y$-component. Therefore, we look to solve for $\bbb(\nnb)$ in the form \begin{equation} \bbb(\nnb) = (b_\perp(\theta)\cos\phi,b_\perp(\theta)\sin\phi,b_\parallel(\theta)), \end{equation} where the function $b_\perp(\theta)$ is the same in the $x$ and $y$-components due to rotational symmetry around the direction of the chemical gradient. We then decompose \begin{eqnarray} b_\perp(\theta) &=& \frac{U}{D_r}\sum_{l=1}^{\infty} - b^\perp_l P_l^1(\cos\theta) = \frac{U}{D_r}\sum_{l=1}^{\infty} b^\perp_l \sin\theta P_l'(\cos\theta),\\ b_\parallel(\theta) &=& \frac{U}{D_r}\sum_{l=0}^{\infty} b^\parallel_l P_l(\cos\theta), \end{eqnarray} and substitute into equation (\ref{eq:b}) to obtain, after some simplification, the equalities \begin{eqnarray} \sum_{l=1}^{\infty} b^\perp_l \left[\left(l(l+1)-3\lambda\cos\theta\right) P_l' + \lambda\sin^2\theta P_l''\right] &=& (1+\mu\cos\theta) \hat{f}(\cos\theta), \label{eq:bperp}\\ \sum_{l=1}^{\infty} b^\parallel_l \left[\left(l(l+1)-2\lambda\cos\theta\right) P_l + \lambda\sin^2\theta P_l'\right] &=& \left(\cos\theta(1+\mu\cos\theta)-\tilde{u}\right) \hat{f}(\cos\theta) \label{eq:bparl}.\quad \end{eqnarray} Note that $b^\parallel_0 = 0$ results from imposing $\int_{S_2} \bbb~d^2\nnb = 0$. The most important step in the derivation is choosing the appropriate inner product for these two equations, so that the integrals on the left-hand side may be evaluated exactly. In equation (\ref{eq:bperp}) we make the substitution $\xi = \cos\theta$ and take $\int \dots (1-\xi^2) P_m(\xi) d\xi$ of both sides. The left-hand side can then be evaluated using the identities \eqref{eq-1-u2-Pk'-Pl}, \eqref{eq-(1-u2)uPk_Pl'} and \eqref{eq-(1-u2)2_Pk_Pl''}. We obtain a set of linear equations for the coefficients $b_l^\perp$ as \begin{equation} P_{ml}b_l^\perp = \int_{-1}^{+1} (1+\mu\xi) \hat{f}(\xi) (1-\xi^2) P_m(\xi) \mathrm{d}\xi, \quad (m\geq 0) \label{eq:pentadiagonal} \end{equation} where the pentadiagonal matrix $\boldsymbol{P}$ has components \begin{multline} P_{ml} = \frac{2\lambda(m+3)(m+2)(m+1)^2}{(2m+5)(2m+3)(2m+1)} \delta_{l,m+2} \\ + \frac{2(m+2)^2(m+1)^2}{(2m+3)(2m+1)} \delta_{l,m+1} \qquad \qquad \qquad \qquad \\ - \frac{2\lambda m(m+1)(2m^2+2m-1)}{(2m+3)(2m+1)(2m-1)} \delta_{lm} \\ \qquad \qquad \qquad + \frac{2m^2(m-1)^2}{(2m+1)(2m-1)} \delta_{l,m-1} \\ + \frac{2\lambda m^2(m-1)(m-2)}{(2m+1)(2m-1)(2m-3)} \delta_{l,m-2}. \end{multline} In equation (\ref{eq:bparl}) we make the same substitution and take $\int \dots P_m(\xi) d\xi$ of both sides. Using identities \eqref{eq-Legendre-ortho-cond}, \eqref{eq-uPkPl-integral} and \eqref{eq-1-u2-Pk'-Pl} we obtain another set of linear equations, this time for the coefficients $b_l^\parallel$ \begin{equation} T_{ml}b_l^\parallel = \int_{-1}^{+1} \left[\xi(1+\mu\xi) - \tilde{u}\right] \hat{f}(\xi) P_m(\xi) \mathrm{d}\xi, \quad (m\geq 0) \label{eq:tridiagonal} \end{equation} where the tridiagonal matrix $\boldsymbol{T}$ has components \begin{multline} \qquad \qquad \qquad T_{ml} = \frac{2\lambda m(m+1)}{(2m+1)(2m+3)} \delta_{l,m+1} \\ + \frac{2m(m+1)}{2m+1} \delta_{lm}\\ - \frac{2\lambda m(m+1)}{(2m-1)(2m+1)} \delta_{l,m-1}. \qquad \qquad \label{eq:tridiagonal-components} \end{multline} Equations \eqref{eq:pentadiagonal} and \eqref{eq:tridiagonal} represent infinite systems of linear equations for the coefficients $b_{m}^\perp$ and $b_{m}^\parallel$ which cannot be inverted analytically. Progress can be made by truncating the series for $b_\perp(\theta)$ and $b_\parallel(\theta)$ to $N$ terms and inverting the resulting $N$-by-$N$ systems numerically. Rather fortunately, only the first two coefficients are needed to compute the active diffusivity tensor, as we shall now see. If we substitute our expressions for $\VVb(\nnb)$ and $\bbb(\nnb)$ into the definition of the active diffusivity tensor, \eqref{eq-GTDT-D-defn}, and average over the azimuthal angle, we find that all the off-diagonal terms vanish. Furthermore, we have $(\boldsymbol{D}_{\text{act}})_{11} = (\boldsymbol{D}_{\text{act}})_{22}$ due to rotational symmetry about the $\kkb$ axis, which is the direction of the chemical gradient. We denote $D_\parallel\equiv (\boldsymbol{D}_{\text{act}})_{33}$ to be the active diffusivity parallel to the chemical gradient, and $D_\perp\equiv (\boldsymbol{D}_{\text{act}})_{11} = (\boldsymbol{D}_{\text{act}})_{22}$ to be the active diffusivity in the plane normal to it. From \eqref{eq-GTDT-D-defn} these quantities are \begin{eqnarray} D_\perp &=& ~\upi \int_0^\pi b_\perp(\theta) \ U\sin\theta(1+\mu\cos\theta) \sin\theta \mathrm{d}\theta, \\ D_\parallel &=& 2\upi \int_0^\pi b_\parallel(\theta) \ U\cos\theta(1+\mu\cos\theta) \sin\theta \mathrm{d}\theta. \end{eqnarray} Using identities \eqref{eq-A-firstfew}, \eqref{eq-uPkPl-integral}, \eqref{eq-1-u2-Pk'-Pl} we can simplify these expression to finally obtain \begin{eqnarray} D_\perp &=& \frac{4\pi U^2}{D_r}\left(\frac{b_1^\perp}{3}+\frac{\mu b_2^\perp}{5}\right), \label{eq:Dperp_result}\\ D_\parallel &=& \frac{4\pi U^2}{D_r}\left(\frac{b_1^\parallel}{3}+\frac{2\mu b_2^\parallel}{15}\right). \label{eq:Dparl_result} \end{eqnarray} Since only the first two coefficients $b_{m}^\perp$ and $b_{m}^\parallel$ ($m=1,2$) are needed in the calculation of the diffusivity tensor, the truncated systems converge very rapidly and we find that it is usually sufficient to truncate to $N$ to order $\mathcal{O}(\lambda,\mu)$ or $N=2$, whichever is larger. \subsection{Numerical simulations} \begin{figure} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3a.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3b.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3c.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3d.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3e.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig3f.pdf} \end{minipage} \caption{Comparison of the distribution of swimmer orientations obtained from numerical simulations with the predictions of the continuum model. Left: histograms of final orientations for 10,000 swimmers whose initial orientations were chosen from a uniform distribution, after being evolved in time for 500 time steps of size $\Delta t = 10^{-2} D_r^{-1}$. The red solid line corresponds to the analytical expression from \eqref{eq-equilibrium-orientation} for the steady-state distribution of swimmer orientations. Right: sample of 1,000 swimmer orientations chosen from the steady-state distribution. Each swimmer orientation corresponds to a unit vector which is represented here as a blue dot on the unit sphere. From top to bottom the rotational P\'{e}clet number is $\lambda = 0.1, 3, 10$.} \label{fig:orientation} \end{figure} In order to validate the continuum model we performed numerical simulations of phoretic swimmers using a simple Euler-Maruyama scheme. To capture the correct dynamics we use a time step much smaller than the timescale $D_r^{-1}$ for rotational diffusion, and calculate all relevant macroscopic quantities as averages over a time period larger than $D_r^{-1}$. At each time step, the position and orientation of the swimmers are updated with the deterministic contribution from the linear and angular velocities of the swimmer, as in the classical Euler method. In addition to this, the orientation of the swimmer experiences a stochastic contribution due to thermal noise. During the $n$th time step the axis of symmetry of the swimmer is deflected in a random direction by an angle $\Delta \theta_n = 2 \sqrt{D_r \Delta t} \Delta W_n$, where $\Delta W_n$ are independent random variables taken from the standard normal distribution \citep[][]{saragosti12}. The only parameters that enter the simulation are the phenomenological constants $\mu$ and $\lambda$ which describe deterministic translation and reorientation, while the stochastic reorientation timescale $D_r^{-1}$ is normalised to one unit of time in our simulations. Since the direct chemotactic index $\nu$ contributes to a constant linear velocity of the particle, we can take it to be zero in our simulations, which means that we are effectively moving to a frame of reference that is sedimenting with the particle, translating at a constant speed $\alpha |\bnabla s^\infty|$ in the direction of the chemical gradient. We only consider a non-zero value for $\nu$ in figure \ref{fig:diffusivities} when we normalise the active diffusivity coefficients (derived from theory only, no simulations) using the mean swimming velocity. In this case we take $\nu$ to be an arbitrary non-zero value in order to avoid the singularity at $\lambda, \mu = 0$ where the mean swimming velocity would also go to zero otherwise. To validate our numerical method we first compared the steady-state distribution of orientations for our simulated swimmers against the analytical expression in \eqref{eq-equilibrium-orientation}, which is a well-established result in the literature \citep[][]{Pedley90,Bees12}. This validation is illustrated in figure~\ref{fig:orientation} which shows perfect agreement between numerical simulations and theory. \begin{figure} \centering \includegraphics[width=\textwidth]{fig4.pdf} \caption{Time evolution of 200 independent trajectories, all starting from the origin at $t=0$. Each snapshot is taken after a time interval of $3D_r^{-1}$, and the chemical gradient is along the horizontal axis. The large dots are placed at regular distances of $3V_s D_r^{-1}$ from the origin which confirms that the cloud of phoretic swimmers drifts along the direction of the chemical gradient at the mean swimming velocity ($V_s $), as predicted by the continuum model. This simulation corresponds to parameters $\lambda=5$ and $\mu=3$, and the time step used was $\Delta t = 10^{-2} D_r^{-1}$.} \label{fig:cloud} \end{figure} We then investigated qualitatively the behaviour of a cloud of swimmers initially starting from the origin. These results are depicted in figure~\ref{fig:cloud}. The cloud of swimmers is seen to diffuse anisotropically and to drift along the direction of the chemical gradient at the mean swimming velocity, $V_s$, as predicted by the continuum model. \begin{figure} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig5a.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig5b.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig5c.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig5d.pdf} \end{minipage} \caption{Comparison of the values obtained from numerical simulations against those predicted by generalised Taylor dispersion theory for the dependence of component $D_\perp$ of the active diffusivity tensor on $\lambda$. The simulations were run with 100 samples of 1,000 swimmers using a time step $\Delta t = 10^{-2} D_r^{-1}$ over a period $2D_r^{-1}$. Each sample was used to calculate one value of the diffusivity $D_\perp$, and then averaged to obtain one data point. The error bars represent one standard deviation amongst the 100 values obtained for the diffusivity. The solid lines represent the predictions of the continuum model from equations \eqref{eq:pentadiagonal}-\eqref{eq:tridiagonal-components} and \eqref{eq:Dperp_result}-\eqref{eq:Dparl_result}. In all plots the diffusivity is non-dimensionalised by $U^2D_r^{-1}$. The four graphs correspond to parameter values $\mu =0$, $\mu =1$, $\mu =5$ and $\mu =10$.} \label{fig:comparison} \end{figure} The final comparison focuses on the values of the active diffusivity tensor computed from stochastic simulations against those given by generalised Taylor dispersion theory. We plot in figure~\ref{fig:comparison} four comparative graphs for the dependence of $D_\perp$ on $\lambda$ at different values of the indirect chemotactic index $\mu$ (results for $D_\parallel$, not shown are very similar). In all cases, the agreement between the simulations and the theory is excellent and the data points obtained numerically are always within one standard deviation of the theoretical prediction. \begin{figure} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig6a.pdf} \end{minipage} \begin{minipage}[b]{0.5\textwidth} \centering \includegraphics[width=\textwidth]{fig6b.pdf} \end{minipage} \caption{Iso-values of the diffusivities $D_\perp$ and $D_\parallel$ computed using generalised Taylor dispersion theory in the $(\lambda,\mu)$ plane. Both diffusivities are normalised using the mean swimming velocity and the timescale of rotational diffusion, i.e.~$D \mapsto D/(V_s^2D_r^{-1})$. These contour plots are obtained with a value of $\nu=1.5$ for the direct chemotactic index, but there are no qualitative changes as $\nu$ is varied (not shown). We observe consistent decay of both diffusivities as $\lambda \tti$ at a fixed value of $\mu$, as well as a saddle point in diffusivity for $\lambda,\mu=\mathcal{O}(1)$.} \label{fig:diffusivities} \end{figure} We can also use the continuum model to understand how the active diffusivity tensor changes as we vary the values of the rotational P\'{e}clet number, $\lambda$, and the indirect chemotactic index, $\mu$. We plot in figure~\ref{fig:diffusivities} iso-values the two components of the active diffusivity tensor normalised by $V_s^2D_r^{-1}$, with $V_s$ being the magnitude of the mean swimming velocity. The active diffusivity decays as the rotational P\'{e}clet number $\lambda \tti$ for any fixed value of $\mu$, as expected, because the active reorientation of the particle enhances directed motion along the chemical gradient and suppresses diffusion. We also observe a local maximum in diffusion at $\lambda,\mu =0$ and a saddle point for $\lambda,\mu$ of $\mathcal{O}(1)$. The surprising finding is that diffusivity increases as $\mu \tti$ for a fixed value of $\lambda$, meaning that indirect chemotactic sedimentation tends to augment diffusion relative to the mean drift velocity of the swimmer. Finally, we investigate how the anisotropy of the diffusion depends on the parameters of the problem; specifically we measure the ratio $D_\perp/D_\parallel$ of the diffusivity in the plane normal to the chemical gradient to the diffusivity parallel to the gradient. Iso-values of this ratio in the $(\lambda,\mu)$ plane are shown in figure~\ref{fig:ratio}. We find that the contour $D_\perp/D_\parallel = 1$ passes through the origin, in agreement with our expectation to recover isotropic diffusion in the limit where both chemotactic alignment and indirect chemotactic sedimentation are weak. We observe that the ratio $D_\perp/D_\parallel$ falls below unity in the region of the $(\lambda,\mu)$ plane above this contour and rises above it in the region to the right of the contour. This indicates that chemotactic alignment, whose strength is given by $\lambda$, favours the reduction of diffusion along the chemical gradient. On the other hand, the indirect chemotactic sedimentation quantified by $\mu$ favours the reduction of diffusion in the plane normal to the chemical gradient. \begin{figure} \centering \includegraphics[width=.55\textwidth]{fig7.pdf} \caption{The ratio $D_\perp/D_\parallel$ computed using generalised Taylor dispersion theory in the $(\lambda,\mu)$ plane. As expected, the contour $D_\perp/D_\parallel=1$ passes through the origin, because the system reverts to isotropic diffusion in the limit $\lambda,\mu \ttz$. Increasing $\mu$ above this contour leads to a ratio $D_\perp/D_\parallel < 1$, meaning that indirect chemotactic sedimentation enhances diffusion along the chemical gradient relative to diffusion in the normal plane. On the other hand, increasing $\lambda$ to the right of this contour leads to a ratio $D_\perp/D_\parallel >1$, meaning that chemotactic alignment suppresses diffusion along the chemical gradient relative to diffusion in the normal plane. } \label{fig:ratio} \end{figure} From figure~\ref{fig:ratio} we also note that $D_\perp/D_\parallel > 1$ for the parameters $\lambda=5$ and $\mu=3$ which were used in our analysis of the time evolution of a cloud of swimmers. This is apparent in figure~\ref{fig:cloud} where the cloud of swimmers, initially at the origin, grows anisotropically over time, with more spreading being observed in the vertical than in the horizontal direction, the latter being the direction of the chemical gradient. \subsection{Asymptotics} For small $\lambda$ and $\mu$ we can invert the 2-by-2 systems obtained from the truncation of the linear systems of equations \eqref{eq:pentadiagonal} and \eqref{eq:tridiagonal} exactly, in order to obtain the approximate analytical expressions for $D_\perp$ and $D_\parallel$ that are given in Appendix \ref{appC}. We then take the limit $\lambda,\mu \ttz$ in these expressions and find that \begin{equation} D_\perp \rightarrow \frac{U^2}{6D_r}, \quad D_\parallel \rightarrow \frac{U^2}{6D_r} \end{equation} which is precisely what we would expect in the purely diffusive limit where both chemotactic alignment and indirect chemotactic sedimentation are weak. We recover isotropy of the active diffusivity tensor and the correct factor of $1/6$ for the dispersal of active swimmers in three dimensions \citep[][]{Berg75}. Furthermore, we discover that the linear terms vanish and that up to quadratic order \begin{eqnarray} D_\perp &\sim & \frac{U^2}{D_r}\left[\frac{1}{6} - \frac{\lambda^2}{40} - \frac{\mu^2}{90} + \mathrm{h.o.t.}\right],\\ D_\parallel &\sim & \frac{U^2}{D_r}\left[\frac{1}{6} - \frac{7\lambda^2}{135} -\frac{\lambda\mu}{9} - \frac{2\mu^2}{135} + \mathrm{h.o.t.}\right]. \end{eqnarray} This explains our previous observation that the diffusivity has a local maximum at $\lambda=0, \ \mu =0$, which can be seen in figure~\ref{fig:diffusivities}. Thus, for small values of $\lambda$ and $\mu$ both mechanisms of chemotactic alignment and indirect chemotactic sedimentation have the effect of suppressing diffusion relative to the mean drift velocity of the swimmers. \section{Conclusions} The idea of using autophoresis to design particles capable of independent propulsion and reorientation in the presence of chemical stimuli has gained much attention in recent years, not the least because it eliminates the need for active steering or fine control of external (e.g.~electric or magnetic) fields. In this paper we derive a general law for the instantaneous behaviour of a spherical axisymmetric swimmer placed in a uniform chemical gradient, which extends and corrects results published in \citet{Saha2014}. We also use our framework to calculate the linear and angular velocity of a Janus sphere, which has great relevance for experimental studies. The main contribution of the present paper is to obtain a fully analytical solution for a non-trivial transport problem involving a chemically active phoretic swimmer placed in a uniform chemical gradient, a canonical setup which had yet to be solved in a general form and presented in a pedagogical manner. Furthermore, our systematic analysis of the different sources contributing to chemotaxis could help inform the design of phoretic swimmers in future experiments, as we discussed in the subsection on Janus spheres. We reinforce the rationale for our first modelling assumption by noting that the regime in which the surface reaction rate is linear in the substrate concentration is the most interesting and relevant case to artificial chemotaxis, because the product problem is sensitive to the substrate gradient in this limit. Our second modelling assumption of small Damk\"{o}hler number could be relaxed, but in this case the substrate problem would no longer have a closed analytical solution and we would have to resort to numerical methods for solving a truncated system. The instantaneous model could be further extended through the inclusion of advective effects, which have been highlighted by several studies \citep{khair2013, michelin2013c, Lauga14}, or by considering the possible ionic effects recently pointed out by \citet{brown2014}. In the second half of the paper we present the calculations involved in applying generalised Taylor dispersion theory \citep{Frank91, Frank93} to the artificial chemotaxis of phoretic swimmers, and we obtain very good agreement between the continuum model and numerical simulations. We observe non-trivial variations of the active diffusivity in the two-dimensional parameter space of rotational P\'{e}clet number and indirect chemotactic index, which creates the possibility of novel pattern formation in systems where these parameters vary with position. The indirect chemotactic index is an added degree of complexity that is new to our continuum model compared to previous studies since it is a direct consequence of the complex chemical dynamics considered in our derivation of the instantaneous behaviour. The limitations of the present continuum model must however be acknowledged. The main difference between the problem of autophoretic swimmers and that of gyrotactic micro-organisms is that the swimming velocity of the former depends on position, whereas for the latter it is constant. This imposes a further condition on the applicability of the model, namely that the swimming velocity is sufficiently slowly varying in space, which may not be the case in certain practical applications. Furthermore, our model is three-dimensional, whereas many experimental projects are concerned with two-dimensional distributions of autophoretic swimmers above a flat plate. The model in this paper describes the long-time behaviour of a single phoretic swimmer, and therefore it is valid only for a population of non-interacting autophoretic swimmers. This makes our continuum model unable of explaining the complex collective behaviour observed in dense populations of interacting phoretic swimmers \citep{palacci2013, Saha2014}. Nevertheless, it represents an appropriate starting point for investigating the emergent behaviour of dilute suspensions of chemically active colloidal particles, such as the recent work done on phoretic swimmers ``riding'' active density waves \citep[see][]{geiseler2016}. In the phoretic swimming literature, many studies assume that active swimming leads to an enhanced, but still isotropic, spatial diffusion coefficient, whereas active reorientation leads to a modified rotational diffusion coefficient, and that the two mechanisms are independent of each other. We clearly show in this paper the importance of having an anisotropic diffusivity tensor which encapsulates the effect of both mechanisms, and which represents the correct way to think about the dispersion of autophoretic swimmers. On a phenomenological level, this represents an important contribution of our continuum model to the general understanding of the long-time behaviour of phoretic swimmers placed in a chemical gradient. \smallskip We thank the anonymous referees for their useful comments on the early version of our manuscript. This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement 682754 to EL). This work was also funded by a Summer Research Studentship from Trinity College and a George and Lilian Schiff Studentship (MTC).

c43c2f1d8c0a09f896e749cc8d00a5b26e60e3e2

\section{Introduction} \label{sec:Intro} One of the most important features of QCD relevant to the low-energy hadron physics is the chiral symmetry and its spontanious breaking. The spontanious symmetry breaking generates mass differences between chiral partners as well as the mixing among different chiral representations. It is interesting to study the role of the chiral symmetry breaking to determine the properties and structures of baryons such as amount of the masses of baryons generated by the chiral symmetry breaking and the chiral partner of ground state nucleon. In a hadronic model for light nucleons based on the parity doublet structure~\cite{DeTar,Jido1,Jido2,Gallas:2009qp,Gallas} the chiral partner of $N(939)$. One of the important feature of the model is the existence of the chiral invariant mass denoted by $m_0$ which is not originated from the spontaneous chiral symmetry breaking. In other word, the mass splitting between $N(939)$ and its chiral partner $N(1535)$ denoted by $m_0$, which implies that the masses of $N(939)$ and its chiral partner tend to $m_0$ when the chiral symmetry is restored and their mass splitting is given by spontanious chiral symmetry braking. This parity doublet structure is extended to include hyperons and/or excited nucleons such as $N(1440)$ and $N(1650)$ (see, e.g. Refs.~\cite{Nemoto1,Chen:2008qv,% Dmitrasinovic:2009vp,Dmitrasinovic:2009vy,Chen:2009sf,Chen:2010ba,Chen:2011rh, Dmitrasinovic1,Nishihara:2015fka,Olbrich:2015gln,Olbrich:2017fsd}). Particularly, in Ref.~\cite{Olbrich:2015gln}, two $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ representations under the chiral $\mbox{SU}(2)_L \otimes \mbox{SU}(2)_R$ are introduced to study $N(939)$, $N(1440)$, $N(1535)$ and $N(1650)$. Two types of solutions were found: In one type, the chiral partner of $N(939)$ is a mixture of $N(1535)$ and $N(1650)$ and two chiral invariant masses are almost $200$\,MeV. In another type, $N(939)$ and $N(1535)$ are chiral partners to each other and two chiral invariant masses are about $1000$\,MeV. On the other hand, in Ref.~\cite{Nishihara:2015fka}, a model is constructed in the three-flavor framework to further introduce the chiral $[(\bf{3},\bf{6})\oplus(\bf{6},\bf{3})]$ representations under the chiral $\mbox{SU}(3)_L \otimes \mbox{SU}(3)_R$ symmetry, which correspond to the chiral $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representations under the chiral $\mbox{SU}(2)_L \otimes \mbox{SU}(2)_R$ symmetry. It was shown that the $N(939)$ is dominated by the $[(\bf{3},\bf{6})\oplus(\bf{6},\bf{3})]$ to reproduce the axial charge of nucleon as pointed in Ref.~\cite{Dmitrasinovic:2009vp}, and that the chiral invariant mass of the $N(939)$ is about $800$\,MeV. The large value of the chiral invariant mass of $N(939)$ seems consistent with the results by the lattice QCD analysis in Ref.~\cite{Aarts:2015mma,Aarts:2017rrl}, which shows that is almost constant even if temperature is increased. In this paper, we make a general analysis using a two-flavor parity doublet model including chiral $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ and $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representations under the chiral $\mbox{SU}(2)_L \otimes \mbox{SU}(2)_R$ symmetry, with including derivative interactions to the pion fields. We will show that there exist five groups of solutions distinguishable by chiral inavarinat masses and mixing rates of nucleons. In a group of solutions, the chiral partner of $N(939)$ is $N(1440)$ having small chiral invariant mass of about 100MeV and $N(939)$ is dominated by $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representation. In another group, on the other hand, $N(939)$ belongs dominantly $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ representation having a large chiral invariant mass, and chiral partner of $N(939)$ is a mixture of $N(1535)$ and $N(1650)$. Futhermore, we give predictions of off-diagonal elements of axial-charge matrix, which could be checked in future lattice analysis. We also show changes of nucleon masses when the vacuum expectation value of $\sigma$, which is an order parameter of the spnoaneous chiral symmetry breaking, is changed. This paper is organized as follows: In section \ref{sec:EPDM}, we construct an extended model with parity doublet structure. Section~\ref{sec:CIM} is a main part, where we show the numerical results of fitting on the chiral invariant masses and chiral partner structure. In sections \ref{sec:axial charge} and \ref{sec:masses}, we study off-diagonal components of axial-charges matrix and change of nucleons mass as predictions. Finally we will give a brief summary and discussions in section~\ref{sec:summary}. \section{An Extended parity doublet model} \label{sec:EPDM} In this section we introduce four baryon fields with parity doublet structure and construct a Lagrangian for baryons and scalar and pseudoscalar mesons based on the ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ chiral symmetry. \subsection{Model construction} The chiral representations of quarks under ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ are written as \begin{equation} q_L\sim({\bf 2},{\bf 1})\hspace{5mm},\hspace{5mm} q_R\sim({\bf 1},{\bf 2}) \ , \end{equation} where the $\bf{2}$ and $\bf{1}$ in above bracket express doublet and singlet, respectively. Since baryons are expressed as direct products of three quarks, we have following possibilities for the chiral representations of baryons: \begin{align} q\otimes q\otimes q \sim & [({\bf 2},{\bf 1})\oplus({\bf 1},{\bf 2})]^3 \nonumber \\ \sim& 5[({\bf 2},{\bf 1})\oplus({\bf 1},{\bf 2})] \oplus3[({\bf 3},{\bf 2})\oplus({\bf 2},{\bf 3})] \notag \\ & \oplus[({\bf 4},{\bf 1})\oplus({\bf 1},{\bf 4})] \ . \end{align} After the chiral symmetry is spontaneously broken down to the flavor symmetry, nucleons appear from the representations of $(\bf{2},\bf{1})\oplus(\bf{1},\bf{2})$ and $(\bf{3},\bf{2})\oplus(\bf{2},\bf{3})$. In this paper we introduce two baryon fields corresponding to these two representations: \begin{eqnarray} &\psi_{1l}\sim({\bf 2},{\bf 1}),\hspace{10mm}\psi_{1r}\sim({\bf 1},{\bf 2})& \nonumber \\ &\psi_{2l}\sim({\bf 1},{\bf 2}),\hspace{10mm}\psi_{2r}\sim({\bf 2},{\bf 1})& \nonumber \\ &\eta_{1l}\sim({\bf 2},{\bf 3}),\hspace{10mm}\eta_{1r}\sim({\bf 3},{\bf 2})& \nonumber \\ &\eta_{2l}\sim({\bf 3},{\bf 2}),\hspace{10mm}\eta_{2r}\sim({\bf 2},{\bf 3})& \label{assignment} \end{eqnarray} Here the subscripts $l$ and $r$ express the chirality: \begin{eqnarray} &\gamma_5\psi_{il}=-\psi_{il},\hspace{10mm}\gamma_5\psi_{ir}=+\psi_{ir}& \nonumber \\ &\gamma_5\eta_{il}=-\eta_{il},\hspace{10mm}\gamma_5\eta_{ir}=+\eta_{ir}& \end{eqnarray} for $i=1,2$. For clarifying the representations under the chiral symmetry, we explicitly write the superscripts of the baryon fields as \begin{align} &(\psi_{1l})^a , (\psi_{1r})^\alpha , (\psi_{2l})^\alpha , (\psi_{2r})^a \nonumber \\ &\eta_{1l}^{(a,\alpha\beta)},\eta_{1r}^{(ab,\alpha)},\eta_{2l}^{(ab,\alpha)},\eta_{2r}^{(a,\alpha\beta)} \end{align} where $a,b=1,2$ are for ${\rm SU(2)_L}$ and $\alpha , \beta=1,2$ for ${\rm SU(2)_R}$. Note that the superscripts $ab$ and $\alpha \beta$ of $\eta$ fields are symmetrized to express ${\bf 3}$ representation: e.g. $\eta_{1l}^{(a,\alpha\beta)} = \eta_{1l}^{(a,\beta\alpha)}$. The transformation properties under the parity and the charge conjugation are defined as \begin{align} &\Psi_{1l,1r}\ \mathop{\to}_{P}\ \gamma_0\Psi_{1r,1l}\hspace{5mm},\hspace{5mm}\Psi_{2l,2r} \ \mathop{\to}_P\ -\gamma_0\Psi_{2r,2l}\hspace{5mm},\hspace{5mm} \\ &\Psi_{1l,1r}\ \mathop{\to}_{C}\ C(\bar{\Psi}_{1r,1l})^T\ ,\hspace{5mm}\Psi_{2l,2r}\ \mathop{\to}_{C}\ -C(\bar{\Psi}_{2r,2l})^T\ ,\hspace{5mm} \end{align} where $C=i\gamma^2\gamma^0$ and $\Psi=\psi,\eta$. The covariant derivatives for the fields are expressed as \begin{eqnarray} D_{\mu}\psi_{1r,2l}&=&(\partial_{\mu}-i\mathcal{R}_{\mu})\psi_{1r,2l} \ , \nonumber \\ D_{\mu}\psi_{1l,2r}&=&(\partial_{\mu}-i\mathcal{L}_{\mu})\psi_{1l,2r} \ , \end{eqnarray} and \begin{eqnarray} (D_{\mu}\eta_{1l,2r})^{(a,\alpha\beta)}&=&\partial_{\mu}\eta_{1l,2r}^{(a,\alpha\beta)}-i(\mathcal{L}_{\mu})^a_b\eta_{1l,2r}^{(b,\alpha\beta)} \nonumber \\ &&-i[(\mathcal{R}_{\mu})^{\alpha}_{\rho}\delta_{\sigma}^{\beta}+\delta_{\rho}^{\alpha}(\mathcal{R}_{\mu})^{\beta}_{\sigma}]\eta_{1l,2r}^{(a,\rho\sigma)} \ , \nonumber \\ (D_{\mu}\eta_{1r,2l})^{(ab,\alpha)}&=&\partial_{\mu}\eta_{1r,2l}^{(ab,\alpha)}-i(\mathcal{R}_{\mu})^{\alpha}_{\beta}\eta_{1r,2l}^{(ab,\beta)} \nonumber \\ &&-i[(\mathcal{L}_{\mu})^{a}_{c}\delta_{d}^{b}+\delta_{c}^{a}(\mathcal{L}_{\mu})^{b}_{d}]\eta_{1r,2l}^{(cd,a)} \ , \end{eqnarray} where $\mathcal{L}_\mu$ and $\mathcal{R}_\mu$ are the external gauge fields introduced by gauging the chiral ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ symmetry. Next we introduce a 2$\times$2 matrix field $M$ expressing a nonet of scalar and pseudoscalar mesons made of a quark and an antiquark. The representation under ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ of the $M$ is \begin{equation} M=\frac{\sigma}{2}+i{\vec \pi} \cdot {\vec T}\sim({\bf 2},{\bf 2}). \end{equation} The transformation properties under the parity and the charge congugation are given by \begin{align} M\ \mathop{\to}_P\ M^{\dag} \ , \quad M\ \mathop{\to}_{C}\ M^{T} \ . \end{align} The covariant derivative for $M$ is expressed as \begin{equation} (D_{\mu}M)^a_\alpha=\partial_{\mu}M^a_{\alpha}-i(\mathcal{L}_{\mu})^a_bM^b_\alpha+iM_\beta^a(\mathcal{R}_{\mu})_{\alpha}^\beta \end{equation} Using the fields introduced above we construct a Lagrangian invariant under the chiral ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ symmetry. \begin{widetext} Let us first consider terms including only $\psi_1$ and $\psi_2$ and their Yukawa interaction to $M$ field. In the present analysis, we include interaction terms with one $M$ field. Then, possible terms are expressed as \begin{eqnarray} \mathcal{L}^{(1)}=&&\bar{\psi}_{1l}i\Slash{D}\psi_{1l}+\bar{\psi}_{1r}i\Slash{D}\psi_{1r} \nonumber +\bar{\psi}_{2l}i\Slash{D}\psi_{2l}+\bar{\psi}_{2r}i\Slash{D}\psi_{2r} \nonumber \\ &&-g_1(\bar{\psi}_{1l}M\psi_{1r}+\bar{\psi}_{1r}M^{\dag}\psi_{1l}) -g_2(\bar{\psi}_{2r}M\psi_{2l}+\bar{\psi}_{2l}M^{\dag}\psi_{2r}) \nonumber \\ &&-m_0^{(1)}(\bar{\psi}_{1l}\psi_{2r}-\bar{\psi}_{1r}\psi_{2l}-\bar{\psi}_{2l}\psi_{1r}+\bar{\psi}_{2r}\psi_{1l}) \ . \end{eqnarray} A part including $\eta_1$ and $\eta_2$ with $M$ is \begin{eqnarray} \mathcal{L}^{(2)}=&&(\bar{\eta}_{1l})_{(a,\alpha\beta)}i\Slash{D}(\eta_{1l})^{(a,\alpha\beta)} +(\bar{\eta}_{1r})_{(ab,\alpha)}i\Slash{D}(\eta_{1r})^{(ab,\alpha)} +(\bar{\eta}_{2l})_{(ab,\alpha)}i\Slash{D}(\eta_{2l})^{(ab,\alpha)} +(\bar{\eta}_{2r})_{(a,\alpha\beta)}i\Slash{D}(\eta_{2r})^{(a,\alpha\beta)} \nonumber \\ &&-g_3[(\bar{\eta}_{1r})_{(ab,\alpha)}(M)^a_{\beta}(\eta_{1l})^{(b,\alpha\beta)} +(\bar{\eta}_{1l})_{(a,\alpha\beta)}(M^{\dag})^{\alpha}_{b}(\eta_{1r})^{(ab,\beta)}] \nonumber \\ &&-g_4[(\bar{\eta}_{2l})_{(ab,\alpha)}(M)^a_{\beta}(\eta_{2r})^{(b,\alpha\beta)} +(\bar{\eta}_{2r})_{(a,\alpha\beta)}(M^{\dag})^{\alpha}_{b}(\eta_{2l})^{(ab,\beta)}] \nonumber \\ &&-m_0^{(2)}[(\bar{\eta}_{1l})_{(a,\alpha\beta)}(\eta_{2r})^{(a,\alpha\beta)} -(\bar{\eta}_{1r})_{(ab,\alpha)}(\eta_{2l})^{(ab,\alpha)} +(\bar{\eta}_{2r})_{(a,\alpha\beta)}(\eta_{1l})^{(a,\alpha\beta)} -(\bar{\eta}_{2l})_{(ab,\alpha)}(\eta_{1r})^{(ab,\alpha)}] \ . \end{eqnarray} Yukawa interaction terms connecting $\psi$ fields to $\eta$ fields are expressed as \begin{eqnarray} \mathcal{L}^{(3)}=&&-y_1[\epsilon_{bc}(\bar{\psi}_{1r})_{\alpha}(M)^b_{\beta}(\eta_{1l})^{(c,\alpha\beta)} +\epsilon_{\beta\sigma}(\bar{\psi}_{1l})_a(M^{\dag})_b^{\beta}(\eta_{1r})^{(ab,\sigma)} +\epsilon^{bc}(\bar{\eta}_{1l})_{(c,\alpha\beta)}(M^{\dag})^{\beta}_b(\psi_{1r})^{\alpha} +\epsilon^{\beta\sigma}(\bar{\eta}_{1r})_{ab,\sigma}(M)^b_{\beta}(\psi_{1l})^a] \nonumber \\ &&-y_2[\epsilon_{\beta\sigma}(\bar{\psi}_{2r})_a(M^{\dag})_b^{\beta}(\eta_{2l})^{(ab,\sigma)} +\epsilon_{bc}(\bar{\psi}_{2l})_{\alpha}(M)^b_{\beta}(\eta_{2r})^{(c,\alpha\beta)} +\epsilon^{\beta\sigma}(\bar{\eta}_{2l})_{(ab,\sigma)}(M)^b_{\beta}(\psi_{2r})^a +\epsilon^{bc}(\bar{\eta}_{2r})_{(c,\alpha\beta)}(M^{\dag})^{\beta}_b(\psi_{2l})^{\alpha}] \ . \nonumber \\ \end{eqnarray} In addition to the non-derivative interactions shown above, we need to include derivative interactions. Possible interaction terms including one derivative are given by \begin{eqnarray} \mathcal{L}^{(4)}=\frac{i}{f_\pi}(&&-a_1[\bar{\psi}_{1l}\Slash{D}M\psi_{2l}-\bar{\psi}_{1r}\Slash{D}M^{\dag}\psi_{2r} +\bar{\psi}_{2l}\Slash{D}M^{\dag}\psi_{1l}-\bar{\psi}_{2r}\Slash{D}M\psi_{1r}] \nonumber \\ &&-a_2[\bar{\eta}_{1r (ab,\alpha)}(\Slash{D}M)^a_{\beta}\eta_{2r}^{(b,\alpha\beta)} -\bar{\eta}_{1l (a,\alpha\beta)}(\Slash{D}M^{\dag})^{\alpha}_{b}\eta_{2r}^{(ab,\beta)} +\bar{\eta}_{2r(a,\alpha\beta)}(\Slash{D}M^{\dag})^{\alpha}_{b}\eta_{1r}^{(ab,\beta)} -\bar{\eta}_{2l(ab,\alpha)}(\Slash{D}M)^a_{\beta}\eta_{1l}^{(b,\alpha\beta)}] \nonumber \\ &&-a_3[\epsilon_{ab}\bar{\psi}_{1r \alpha}(\Slash{D}M)^a_{\beta}\eta_{2r}^{(b,\alpha \beta)} -\epsilon_{\alpha \beta}\bar{\psi}_{1l a}(\Slash{D}M^{\dag})^{\alpha}_b\eta_{2l}^{(ab,\beta} +\epsilon^{ab}\bar{\eta}_{2r (b,\alpha \beta)}(\Slash{D}M^{\dag})^{\beta}_a\psi_{1r}^\alpha -\epsilon^{\alpha \beta}\bar{\eta}_{2l (ab,\beta)}(\Slash{D}M)^b_\alpha\psi_{1l}^a] \nonumber \\ &&-a_4[\epsilon_{\alpha \beta}\bar{\psi}_{2r a}(\Slash{D}M^{\dag})^{\alpha}_b\eta_{1r}^{(ab,\beta)} -\epsilon_{ab}\bar{\psi}_{2l \alpha}(\Slash{D}M)^{a}_{\beta}\eta_{1l}^{(\alpha \beta,b)} +\epsilon^{\alpha \beta}\bar{\eta}_{1r (ab,\beta)}(\Slash{D}M)^{b}_{\alpha}\psi_{2r}^a -\epsilon^{ab}\bar{\eta}_{1l (\alpha \beta,b)}(\Slash{D}M^{\dag})^{\beta}_a\psi_{2l}^\alpha]) \nonumber \\ \end{eqnarray} \end{widetext} Combining the above terms together, the Lagrangian in the present analysis is given by \begin{equation} \mathcal{L}=\mathcal{L}^{(1)}+\mathcal{L}^{(2)}+\mathcal{L}^{(3)}+\mathcal{L}^{(4)}+\mathcal{L}_{\rm meson} \ , \end{equation} where the mesonic part $\mathcal{L}_{\rm meson}$ is written as \begin{equation} \mathcal{L}_{\rm meson}={\rm Tr}[D_{\mu}M\cdot D^{\mu}M^{\dag}]-V(M) \end{equation} where $V(M)$ is a meson potential term. In this paper, we do not specify the form of the potential, but we assume that this potential provides the vacuum expectation value (VEV) of $M$ as $\langle M \rangle = \mbox{diag}( f_\pi/2 , f_\pi/2)$, where $f_\pi$ is the pion decay constant. \subsection{Mass matrix} We have constructed the Lagrangian by requiring the chiral ${\rm SU(2)_L}$$\otimes$${\rm SU(2)_R}$ invariance. To study the properties of nucleons, we decompose baryons in the chiral representation to irreducible representations of the flavor symmetry as \begin{align} & \psi_{1l,2r}=N^{(1)}_{1l,2r}\ ,\hspace{5mm}\psi_{1r,2l}=N^{(1)}_{1r,2l}\ , \\ & \eta_{1l,2r}^{(a,\alpha\beta)}=\Delta_{1l,2r}^{a \alpha \beta}+\frac{1}{\sqrt{6}}(\epsilon^{\alpha a}\delta^{\beta}_{k} +\epsilon^{\beta a}\delta^{\alpha}_{k})(N^{(2)}_{1l,2r})^k \ , \\ & \eta_{1r,2l}^{(ab,\alpha)}=\Delta_{1r,2l}^{ab\alpha}+\frac{1}{\sqrt{6}}(\epsilon^{a\alpha }\delta^{b}_{k} +\epsilon^{b\alpha }\delta^{a}_{k})(N^{(2)}_{1r,2l})^k \ . \end{align} In the following, for convenience we redefine the nucleon fields as: \begin{equation} N^{\prime(i)}_1 =N_1^{(i)} \ , \quad N^{\prime(i)}_2 = \gamma_5 N_2^{(i)} \ , \quad ( i = 1,2) \ . \label{redef} \end{equation} The mass terms for the redefined fields are expressed as \begin{equation} -\bar{N'}M'_NN' \end{equation} with \begin{equation} N'^T\equiv(N'^{(1)}_1\hspace{5mm}N'^{(2)}_1\hspace{5mm}N'^{(1)}_2\hspace{5mm}N'^{(2)}_2) \end{equation} and \begin{equation} M'_N= \begin{pmatrix} \frac{g_1}{2}f_\pi &-\frac{3y_1}{2\sqrt{6}}f_\pi &m^{(1)}_0 & 0\\ -\frac{3y_1}{2\sqrt{6}}f_\pi & \frac{g_3}{4}f_\pi & 0 & m^{(2)}_0\\ m^{(1)}_0 & 0 & -\frac{g_2}{2}f_\pi & \frac{3y_2}{2\sqrt{6}}f_\pi\\ 0 & m^{(2)}_0 & \frac{3y_2}{2\sqrt{6}}f_\pi & -\frac{g_4}{4}f_\pi \end{pmatrix}. \end{equation} The mass eigenstates denoted by \begin{equation} N'^T_{\rm phys}\equiv(N'^{(1)}_+\hspace{5mm}N'^{(2)}_+\hspace{5mm}N'^{(1)}_-\hspace{5mm}N'^{(2)}_-)_{\rm phys} \end{equation} are obtained by diagonalizing the above mass matrix $M'_N$. We note that parities of all fields in $N'$ are even due to the redefinition given in Eq.~(\ref{redef}). Two eigenvalues of the mass matrix $M'_N$ are negative in our analysis, and we regard the parities of these state as negative. \subsection{One pion interactions and axial charges} The interaction terms of nucleons to one pion are given from the Lagrangian as \begin{equation} \bar{N'}C'_{\pi NN}i\gamma_5\pi N' + \bar{N'}C'_{\partial \pi NN}\gamma^{\mu}\partial_{\mu}\pi \gamma_5N' \ , \end{equation} where $\pi =\vec{\pi}\cdot\vec{\tau}$ and \begin{equation} C'_{\pi NN}= \begin{pmatrix} -\frac{g_1}{2} &-\frac{y_1}{2\sqrt{6}} &0 & 0\\ -\frac{y_1}{2\sqrt{6}} & \frac{5g_3}{12} & 0 & 0\\ 0 & 0 & -\frac{g_2}{2} & -\frac{y_2}{2\sqrt{6}}\\ 0 & 0 & -\frac{y_2}{2\sqrt{6}} & \frac{5g_4}{12} \end{pmatrix} \ , \end{equation} \begin{equation} C'_{\partial \pi NN}= \begin{pmatrix} 0 & 0 &-\frac{a_1}{2f_\pi} &-\frac{a_3}{2\sqrt{6}f_\pi} \\ 0 & 0 &-\frac{a_4}{2\sqrt{6}f_\pi} &\frac{5a_2}{12f_\pi} \\ -\frac{a_1}{2f_\pi} & -\frac{a_4}{2\sqrt{6}f_\pi} &0 &0\\ -\frac{a_3}{2\sqrt{6}f_\pi} & \frac{5a_2}{12f_\pi}& 0& 0\\ \end{pmatrix} \ . \end{equation} Axial-vector charge matrix is determined as \begin{equation} \bar{N'}G'_A\gamma^{\mu}\mathcal{A}_{\mu}\gamma_5N' \ , \end{equation} where \begin{equation} G'_A= \begin{pmatrix} -1 & 0 &-\frac{a_1\sigma_0}{f_\pi} &-\frac{a_3\sigma_0}{\sqrt{6}f_\pi} \\ 0 & \frac{5}{3} &-\frac{a_4\sigma_0}{\sqrt{6}f_\pi} &\frac{5a_2\sigma_0}{6f_\pi} \\ -\frac{a_1\sigma_0}{f_\pi} & -\frac{a_4\sigma_0}{\sqrt{6}f_\pi} &1 &0\\ -\frac{a_3\sigma_0}{\sqrt{6}f_\pi} & \frac{5a_2\sigma_0}{6f_\pi}& 0& -\frac{5}{3}\\ \end{pmatrix} \ . \end{equation} In the present analysis, we identify the mass eigenstates as \begin{equation} N^T_{\rm phys}\equiv(N(939)\,,\,N(1440)\,,\,N(1535)\hspace{5mm}N(1650))_{\rm phys}. \end{equation} \section{Chiral invariant masses and partner structure} \label{sec:CIM} In this section we determine the values of model parameters and study the mixing structure of relevant baryons. As we said in the previous section, we set the VEV of $\sigma$ to be the pion decay constant: \begin{equation} \sigma_0=f_{\pi}=92.4{\rm MeV} \end{equation} Beside this parameter, there are twelve parameters in this model: \begin{equation} m_0^{(1)},m_0^{(2)},g_1,g_2,g_3,g_4,y_1,y_2,a_1,a_2,a_3,a_4 \end{equation} We list values of relevant physical quantities determined from experiments and lattice analyses in Table.\ref{expvalue}. Among them, we use the following ten physical values as inputs: nucleon masses: \begin{eqnarray} &&m_{N(939)}=939{\rm MeV} \ , \nonumber \\ &&m_{N(1440)}=1430{\rm MeV} \ , \nonumber \\ &&m_{N(1535)}=1535{\rm MeV} \ , \nonumber \\ &&m_{N(1650)}=1650{\rm MeV} \ , \label{mass values} \end{eqnarray} partial decay widths:~\footnote{ In calculating the decay widths, we use the pion mass as $m_{\pi}=137{\rm MeV}$. } \begin{align} &\Gamma(N(1440)\to N(939)+\pi)=228{\rm MeV}\ , \notag\\ &\Gamma(N(1535)\to N(939)+\pi)=68{\rm MeV}\ , \notag\\ &\Gamma(N(1650)\to N(939)+\pi)=84{\rm MeV}\ , \notag\\ &\Gamma(N(1650)\to N(1440)+\pi)=22{\rm MeV} \ , \label{width values} \end{align} and axial charges: \begin{align} &g_A(N(939))=1.272 \ , \notag\\ &g_A(N(1650)=0.55 \ . \label{gA values} \end{align} \begin{table} \begin{tabular}{|l|c|c|c|c|} \hline & $P$ & Mass & Width[$\Gamma_{N^*\to N\pi}$] & axial charge \\ \hline \hline N(939) & $+$ & 939.0$\pm$1.3 & - & 1.272$\pm$0.002 \\ N(1440) & $+$ & 1430$\pm$20 & 228$\pm$74 & -\\ N(1535) & $-$ & 1535$\pm$10 & 68$\pm$19 & $\mathcal{O}(0.1)$[lat]\\ \multirow{2}{*}{N(1650)} & \multirow{2}{*}{$-$} & \multirow{2}{*}{1655$\pm$15} & 84$\pm23$[to N(939)] & \multirow{2}{*}{0.55[lat]}\\ & & & 22$\pm$15[to N(1440)] & \\ \hline \end{tabular} \caption{Experimental values of masses and partial decay widths of baryons listed in Ref.~\cite{PDGC}. The column indicated by $ P = \pm$ shows the parity of the nucleon. Unit of masses and widths is MeV. The error of $m_{N(939)}$ expresses the mass difference between the proton and neutron. [lat] indicates that the value is obtained by the lattice analysis in Ref.~\cite{Takahashi:2008fy}. } \label{expvalue} \end{table} In addition to the above inputs, we use the following range of $g_A(N(1535))$ to restrict the parameters: \begin{equation} -0.25\leq g_A(N(1535)) \leq 0.25. \end{equation} Furthermore, we restrict the parameters by requiring all the components of axial-charge matrix on the physical base are no larger than 5. In this analysis, we first fix the chiral invariant masses $m_0^{(1)}$ and $m_0^{(2)}$ to certain values of every 5\,MeV from 0\,MeV to 1500\,MeV, and determine other ten parameters from the values shown in Eqs.~(\ref{mass values})-(\ref{gA values}). Here, $m_0^{(1)}$ and $m_0^{(2)}$ are chiral invariant masses of $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ and $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representations, respectively. We find that solutions are categorized into five groups as shown in Fig.\ref{chiralinvariantmass}. \begin{figure} \begin{center} \includegraphics[clip,width=8.0cm]{cimm1-cimm2_2.pdf} \end{center} \caption{Chiral invariant masses for which we find solutions to reproduce the values in Eqs.~(\ref{mass values})-(\ref{gA values}). } \label{chiralinvariantmass} \end{figure} In Group 1 indicated by purple $+$ symbols, both chiral invariant masses are less than $100$\,MeV. In the range where $m_0^{(1)}$ is between $200$\,MeV and $900$\,MeV, three lumps exist: Group 2 indicated by blue $\blacksquare$ symbols; Group 3 by light green $\times$ symbols; and Group 4 light blue $\times\hspace{-7.5pt}+$ symbols. They are characterized by \begin{align} m_0^{(1)}\sim m_0^{(2)} \ [{\rm Group 2}] \ , \notag\\ m_0^{(1)}\geq m_0^{(2)}\ [{\rm Group 3}] \ , \notag\\ m_0^{(1)}<m_0^{(2)} \ [{\rm Group 4}] \ , \notag\\ m_0^{(1)}\geq m_0^{(2)} \ [{\rm Group 5}] \ . \end{align} In Group 5 indicated by yellow $\square$, the chiral invariant mass $m_0^{(1)}$ takes large value of about 1000\,MeV. In Fig.~\ref{mixingrate}, we show the mixing structure of nucleons: $N(939)$, $N(1440)$, $N(1535)$ and $N(1650)$ for Group 1 to Group 5. Here the horizontal axis shows the value of axial-charge of $N(1535)$ and the vertical axis shows the percentages of $\psi_1$ indicated by magenta $\triangledown$ symbols, $\eta_1$ by brown $\circ$ symbols, $\psi_2$ by green $\diamond$ symbols and $\eta_2$ by navy $\vartriangle$ symbols. In Table~\ref{nucleonstate}, we summarize features of mixing rates for each group. The first row in Fig.~\ref{mixingrate} shows that the dominant component of $N(939)$ is $\eta_1$ indicated by brown $\circ$ belonging to $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representation. We note that we cannot find any solutions for $g_A(N(1535))\gtrsim -0.1$ in the Group 1. One can easily see that $N(1440)$ is dominated by $\eta_2$ (navy $\vartriangle$), $N(1535)$ by $\psi_1$(magenta $\triangledown$) belonging to $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ representation and $N(1650)$ bv $\eta_2$ (green $\diamond$). Since $\eta_1$ and $\eta_2$ are chiral partners to each other, we conclude that $N(1440)$ dominated by $\eta_2$ is the chiral partner to $N(939)$ dominated by $\eta_1$. We would like to stress that this partner structure can be realized when the chiral invariant masses of $N(939)$ and $N(1440)$ are small. In Group 2 (the second row of Fig.~\ref{mixingrate}), $\eta_1$ (brown $\circ$) belonging to $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representation is a dominant component in the $N(939)$ and $\psi_1$ (magenta $\triangledown$) almost occupies $N(1535)$, similarly to Group 1. A difference between Group 1 and Group 2 appears in the rate of $\psi_2$ (green $\diamond$) in $N(1440)$. In Group 1, the mixing rate of $\psi_2$ in $N(1440)$ is smaller than $0.1$ as can be seen in the first row of Fig.~\ref{mixingrate}. On the other hand, the rate of $\psi_2$ is larger than $0.2$ and $\psi_2$ is included in $N(1440)$ dominantly as shown in the second row of Fig.~\ref{mixingrate}. Here the rate of $\eta_2$ component (navy $\vartriangle$) included in $N(1650)$ is high, but $N(1440)$ and $N(1535)$ include a certain amount of the $\eta_2$ component. So, it is difficult to identify the chiral partner of $N(939)$ in Group 2. \begin{table} \begin{tabular}{|c|c|c|c|c|c|} \hline Group & Group 1 & Group 2 & Group 3 & Group 4 & Group 5 \\ color & purple & blue & light green & sky & yellow \\ \hline \hline $N(939)$ & $\eta_1>0.8$ & $\eta_1>0.35$ & $\eta_1<0.45$ & $\eta_1<0.35$ & $\eta_1<0.35$ \\ \hline \multirow{3}{*}{$N(1440)$} & $\psi_1<0.01$ & $\psi_1<0.025$ & $\psi_1>0.25$ & $\psi_1<0.75$ & $\psi_1<0.6$ \\ & $\psi_2<0.1$ & $\psi_2>0.2$ & $\psi_2>0.15$ & $\psi_2<0.2$ & $\psi_2<0.35$ \\ \hline $N(1650)$ & $\eta_2<0.1$ & $\eta_2>0.1$ & $\eta_2>0.1$ & $\eta_2<0.3$ & $\eta_2\leq0.1$ \\ \hline \end{tabular} \caption{Features of mixing rates. For example, in the column for Group 1, $\eta_1 > 0.8$ in the row of $N(939)$ implies that the percentage of $\eta_1$ component in $N(939)$ is always larger than $0.8$.} \label{nucleonstate} \end{table} In Group 3, Group 4 and Group 5, $N(939)$ is composed of $\psi_1$(magenta $\triangledown$) or $\eta_2$(navy $\vartriangle$) dominantly and negative parity nucleons, $N(1535)$ and $N(1650)$ have $\psi_2$ (green $\diamond$)) or $\eta_2$ (brown $\circ$) mainly, as can be seen in the third, fourth and fifth rows in Fig.~\ref{mixingrate}. This indicates that the chiral partner of $N(939)$ is a mixture of two negative parity nucleons in these groups, differently from Group 1 and Group 2. Table~\ref{nucleonstate} shows that Group 5 is distinguished from Group 3 by the mixing rate of $\eta_2$ component in $N(1650)$: The rate is larager than $0.1$ in Group 3, while it is no greater than $0.1$ in Group 5. On the other hand, it is difficult to distinguish Group 4 with Group 5 and Group 3 with Group 4 by mixing rates. In the present work, we use the values of chiral invariant masses in addition to the mixing rates to separate these Groups: The chiral invariant masses satisfy $m_0^{(1)}<m_0^{(2)}$ in Group 4; Group 3 is characterized by $m_0^{(1)}\geq m_0^{(2)}$ and $\eta_2>0.1$ in $N(1650)$, while Group 5 by $m_0^{(1)}\geq m_0^{(2)}$ and $\eta_2\leq0.1$ in $N(1650)$. \begin{figure}[H] \begin{center} \includegraphics[clip,width=8.5cm]{pointtype1.pdf} \\ \includegraphics[clip,width=8.9cm]{Group1-gA1535_2.pdf} \\ \hspace{1mm} \\ \includegraphics[clip,width=8.9cm]{Group2-gA1535_2.pdf} \\ \hspace{1mm} \\ \includegraphics[clip,width=8.9cm]{Group3-gA1535_3.pdf} \\ \hspace{1mm} \\ \includegraphics[clip,width=8.9cm]{Group4-gA1535_1.pdf} \\ \hspace{1mm} \\ \includegraphics[clip,width=8.9cm]{Group5-gA1535_1.pdf} \end{center} \caption{Mixing rates of nucleons. The first to fifth rows shows Group 1 to Group 5.} \label{mixingrate} \end{figure} In Ref.~\cite{Olbrich:2015gln}, two $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ representations are used to study $N(939)$, $N(1440)$, $N(1535)$ and $N(1650)$. The authors found two types of solutions. In one type (Type-A), One is that the chiral partner of $N(939)$ is a mixture of $N(1535)$ and $N(1650)$ in vacuum and both chiral invariant masses are almost $200$\,MeV. In another type (Type-B) $N(939)$ and $N(1535)$ are chiral partners to each other and both chiral invariant masses are about $1000$\,MeV. In the present analysis, we find solutions for which $N(939)$ is dominated by $[(\bf{1},\bf{2})\oplus(\bf{2},\bf{1})]$ representation in Groups 3, 4 and 5. Comparing mixing structure and the chiral invariant masses, we think that the Type-A is consistent with Group 3 and Type-B with Group 4 or 5. In Ref.~\cite{Nishihara:2015fka}, $[(\bf{3},\bf{\bar{3}})\oplus(\bf{\bar{3}},\bf{3})]$ and $[(\bf{3},\bf{6})\oplus(\bf{6},\bf{3})]$ representations under the chiral $U(3)_{\rm L}\otimes U(3)_{\rm R}$ symmetry are introduced to study six nucleons including $N(939)$, $N(1440)$, $N(1535)$ and $N(1650)$. This analysis indicates that $N(939)$ have $[(\bf{3},\bf{6})\oplus(\bf{6},\bf{3})]$ dominantly and the chiral invariant mass is 500MeV$\sim$800MeV. This is consistent with Group 2 in our analysis. \begin{table}[htb] \begin{tabular}{|c|c|c|c|c|c|c|} \hline Nucleon & rep & Group1 & Group 2 & Group 3 & Group 4 & Group 5 \\ \hline \multirow{4}{*}{$N(939)$}& $\psi_1$ & 0.113 & 0.21 & \cellcolor{yellow}0.518 & \cellcolor{yellow}0.677 &\cellcolor{yellow}0.768 \\ & $\eta_1$ & \cellcolor{yellow}0.856 & \cellcolor{yellow}0.735 & 0.098 & 0.032 &0.001 \\ & $\psi_2$ & 0.002 & 0.052 & 0.133 & 0.163 &0.227 \\ & $\eta_2$ & 0.029 & 0.002 & 0.251 & 0.128 &0.004 \\ \hline \multirow {4}{*}{$N(1440)$}& $\psi_1$ & 0.002 & 0.003 & \cellcolor{yellow}0.355 & 0.092 &0.002 \\ & $\eta_1$ & 0.029 & 0.043 & 0.021 & 0.223 &0.041 \\ & $\psi_2$ & 0.033 & \cellcolor{yellow}0.627 & \cellcolor{yellow}0.368 & 0.01 &0.001 \\ & $\eta_2$ & \cellcolor{yellow}0.936 & 0.327 & 0.256 & \cellcolor{yellow}0.675 &\cellcolor{yellow}0.956 \\ \hline \multirow{4}{*}{$N(1535)$}& $\psi_1$ & \cellcolor{yellow}0.819 & \cellcolor{yellow}0.641 & 0.126 & 0.012 &0.137 \\ & $\eta_1$ & 0.109 & 0.222 & \cellcolor{yellow}0.801 & 0.212 &0.322 \\ & $\psi_2$ & 0.068 & 0.007 & 0.057 & \cellcolor{yellow}0.615 &\cellcolor{yellow}0.528 \\ & $\eta_2$ & 0.004 & 0.130 & 0.015 & 0.161 &0.013 \\ \hline \multirow {4}{*}{$N(1650)$}& $\psi_1$ & 0.066 & 0.146 & 0.001 & 0.218 &0.093 \\ & $\eta_1$ & 0.006 & 0.0 & 0.08 & \cellcolor{yellow}0.533 &\cellcolor{yellow}0.636 \\ & $\psi_2$ & \cellcolor{yellow}0.897 & 0.314 & \cellcolor{yellow}0.442 & 0.212 &0.244 \\ & $\eta_2$ & 0.031 & \cellcolor{yellow}0.540 & \cellcolor{yellow}0.477 & 0.037 &0.027 \\ \hline \end{tabular} \caption{Typical values of mixing rates of nucleons. } \label{tablemixingrate} \end{table} \section{Axial charges} \label{sec:axial charge} In the previous section, we discussed the mixing structure of nucleons together with their chiral invariant masses. In this section, we study axial charges in detail. We define transition axial charge as off-diagonal elements of following axial-charge matrix on physical base. \begin{equation} \bar{N'}_{\rm phys}G'_{A {\rm phys}}\gamma^{\mu}\mathcal{A}_{\mu}\gamma_5N'_{\rm phys} \end{equation} where \begin{equation} G'_{A {\rm phys}}= \begin{pmatrix} g_A(N_1) & g_A(N_1N_2) & g_A(N_1N_3) &g_A(N_1N_4) \\ g_A(N_1N_2) & g_A(N_2) & g_A(N_2N_3) &g_A(N_2N_4) \\ g_A(N_1N_3) & g_A(N_2N_3) & g_A(N_3) &g_A(N_3N_4)\\ g_A(N_1N_4) & g_A(N_2N_4) & g_A(N_3N_4) &g_A(N_4)\\ \end{pmatrix} \end{equation} and \begin{equation} (N_1,N_2,N_3,N_4)\equiv(N'^{(1)}_+\hspace{5mm}N'^{(2)}_+\hspace{5mm}N'^{(1)}_-\hspace{5mm}N'^{(2)}_-)_{\rm phys} \end{equation} \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{gA1440_1.pdf} \end{center} \caption{The values of $g_A(N_2)$ and $g_A(N_3)$:points of all Groups. This relation is determined by $\sum_{i=1}^4g_A(N_i)=0$.} \label{gAN1440} \end{figure} In the present model, the following relation the diagonal axial charges is satiefied: \begin{equation} \sum_{i=1}^4g_A(N_i)=0 \end{equation} Now we use axial charge:$g_A(N_1)=g_A(N(939))=1.272$ and $g_A(N_4)=g_A(N(1650))=0.55$ as input. So this relation is \begin{equation} g_A(N_2)+g_A(N_3)=-1.822 . \end{equation} We plot this relation in Fig.\ref{gAN1440}. This plot shows that, when the axial charge of $N1535)$ is in the range consistent with the lattice analysis, the axial charge of $N(1440)$ is negative. \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{gAN2N3-gAN2N4_2.pdf} \end{center} \caption{Predicted values of $g_A(N_2N_3)$ and $g_A(N_2N_4)$.} \label{gAN2N3-gAN2N4} \end{figure} In Fig.~\ref{gAN2N3-gAN2N4} we plot predicted values of $g_A(N_2N_3)$ and $g_A(N_2N_4)$. This shows that $|g_A(N_2N_3)|$ of Group 1 and Group 2 is always larger than $2$ and $0.5$, respectively. On the other hand, $|g_A(N_2N_4)|$ of Group 1 is smaller than 2 and that of Group 2 is above 1. We see that Group 1 is able to be distinguished from other groups. \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{gAN1N3-gAN3N4_2.pdf} \end{center} \caption{Predicted values of $g_A(N_3N_4)$ and $g_A(N_1N_3)$. } \label{gAN3N4-gAN1N3} \end{figure} Predicted values of $g_A(N_3N_4)$ and $g_A(N_1N_3)$ are plotted in Fig.~\ref{gAN3N4-gAN1N3}. We note that $|g_A(N_1N_3)|$ belonging to Group 1 and Group 2 are larger than $1$, and that $|g_A(N_3N_4)|$ of Group 2 and Group 5 lies between $1$ and $3$ and that of Group 1 is above $1$. In particular, $|g_A(N_3N_4)|\sim2$ in Group 5. \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{gAN1N4-gAN1N2_2.pdf} \end{center} \caption{Predicted values of $g_A(N_1N_2)$ and $g_A(N_1N_4)$: There are no points of Group 3 in the vicinity of $|g_A(N_1N_4)|\sim 0$. Values of Group 1, 2, 4 and 5 are mixed there. } \label{gAN1N2-gAN1N4} \end{figure} We plot the values of $g_A(N_1N_2)$ and $g_A(N_1N_4)$ in Fig.~\ref{gAN1N2-gAN1N4}. From Fig.\ref{gAN1N2-gAN1N4}, we can find that $|g_A(N_1N_2)|$ belonging to Group 1 (purple $+$ symbols) and Group 2 (blue $\blacksquare$ symbols) are no larger than $1$, while that belonging to Group 3 (light green $\times$ symbols) and Group 4 light blue $\times\hspace{-7.5pt}+$ symbols are below $1.5$. $|g_A(N_1N_4)|$ belonging to Group 3 (light green $\times$ symbols) are larger than $0.5$. We summarize typical predicted values of transition axial-charges and the range of transition axial-charges in Table.~\ref{tableaxialcharge}. \begin{table} \scalebox{0.65}[1.0] { \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline group & ($m_0^{(1)}$ , $m_0^{(2)}$) & $g_A(N_1N_2)$ & $g_A(N_1N_3)$ & $g_A(N_1N_4)$ & $g_A(N_2N_3)$ & $g_A(N_2N_4)$ & $g_A(N_3N_4)$ \\ \hline 1 & (5 , 100) & 0.689 & 1.707 & -0.159 & -4.512 & -0.197 & -4.303 \\ 1 & (40 , 95) & -0.505 & -1.353 & 0.417 & -2.859 & 1.310 & 0.804 \\ :1: & - & $|g_A | < 1$ & $1<|g_A |<2$ & $|g_A | < 1$ & $|g_A |>2.5$ & $|g_A |<2$ & $|g_A|>0.5$ \\ \hline 2 & (510 , 540) & 0.449 & -2.224 & 0.247 & -4.644 & 2.614 & -1.979 \\ 2 & (645 , 715)& -0.434 & 1.868 & -1.132 & -1.039 & 2.204 & -0.921 \\ 2 & (790 , 855)& 0.252 & -1.912 & -1.306 & -1.294 & -1.614 & 1.238 \\ :2: & - & $|g_A|<1$ & $|g_A|>1$ & $|g_A|<2.5$ & $|g_A|>0.5$ &$|g_A|>1$ & $|g_A|>0.5$ \\ \hline 3 & (400 , 175) & 0.906 & -1.197 & -1.697 & 0.935 & -1.394 & 1.561 \\ 3 & (695 , 330) & 0.206 & 1.549 & 0.888 & 2.825 & 2.394 & 4.95 \\ 3 & (940 , 620) & 0.203 & 1.963 & 1.597 & 1.197 & 1.324 & -0.242 \\ :3: & - &$|g_A|<1.5$ & $|g_A|<2.5$ & $|g_A|<4$ & $|g_A|<3$ &$|g_A|<2.5$ & $g_A>-1$ \\ \hline 4 & (365 , 685) & 1.328 & -1.694 & 0.175 & -2.219 & 2.028 & -1.419 \\ 4 & (650 , 945) & -0.099 & -0.149 & 1.959 & -0.534 & -1.386 & 0.795 \\ 4 & (1070 , 1150) & -0.755 & -0.303 & 2.136 & -0.293 & -0.159 & 1.714 \\ :4: & - &$|g_A|<1.5$ & $|g_A|<2$ &$g_A<2.5$ & $|g_A|<3.5$ & $|g_A|<3$ & $g_A<2$ \\ \hline 5 & (1070 , 50) & -0.514 & -1.350 & -1.454 & 0.075 & -0.548 & -1.415 \\ 5 & (1085 , 500)& 0.713 & 1.210 & -1.599 & -0.058 & 0.107 & 1.492 \\ 5 & (1210 , 1050)& -0.295 & 0.175 & 1.788 & -0.245 & 0.116 & 1.867 \\ :5: & - &$|g_A|<2$ & $|g_A|<2.5$ &$|g_A|<2.5$ & $|g_A|<3$ & $|g_A|<2$ &$1<|g_A|<3$ \\ \hline \end{tabular} } \caption{ Typical predicted values of axial charges. Rows indicated by :$j$: ($j=1,\ldots,5$) show that the absolute value of axial charge lie in the range shown in the rows.} \label{tableaxialcharge} \end{table} \section{change of nucleon masses} \label{sec:masses} In this section we study the change of nucleon masses when the VEV of $\sigma$ changed . We plot the dependences of nucleon masses on the value of the VEV for Groups 1-4 in Fig.~\ref{mass_1} and those for Group 5 in Fig.~\ref{mass_2} for some choices of the chiral invariant masses, $m_0^{(1)}$ and $m_0^{(2)}$. \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{numass-sgm_1.pdf} \end{center} \caption{ Dependences of nucleon masses on the VEV of $\sigma$ for Groups 1-4.Values of chiral invariant masses used are $(m_0^{(1)},m_0^{(2)})=(40{\rm MeV},95{\rm MeV})$ (Group 1), $(m_0^{(1)},m_0^{(2)})=(645{\rm MeV},715{\rm MeV})$ (Group 2), $(m_0^{(1)},m_0^{(2)})=(400{\rm MeV},175{\rm MeV})$ (Group 3) and $(m_0^{(1)},m_0^{(2)})=(365{\rm MeV},685{\rm MeV})$ (Group 4). } \label{mass_1} \end{figure} Figure~\ref{mass_1} shows that nucleons masses are decreased as $\sigma_0$ is decreased in Group 1, 2, 3 and 4, if chiral invariant masses are smaller than mass of $N(939)$. We note that, for some parameter choices in Group 4, mass of ground state nucleon is increased as $\sigma_0$ is decreased. \begin{figure}[H] \begin{center} \includegraphics[clip,width=9.0cm]{massdegenerate_g5_1.pdf} \end{center} \caption{Dependences of nucleon masses on the VEV of $\sigma$ in Group 5 for $(m_0^{(1)}, m_0^{(2)})=(1070\mbox{MeV},50\mbox{MeV})$ (left figure) and $(1210\mbox{MeV},1050\mbox{MeV})$ (right figure). } \label{mass_2} \end{figure} In the case of Group 5 shown in Fig.~\ref{mass_2}, the value of $m_0^{(1)}$ is about $1000$\,MeV, while $m_0^{(2)}$ takes values in wide range. Left panel of FIg.~\ref{mass_2} [$(m_0^{(1)}, m_0^{(2)})=(1070\mbox{MeV},50\mbox{MeV})$] shows that the mass of the ground state is stable for $\sigma_0 > 60$\,MeV and it decreased towards $m_0^{(2)}$ as $\sigma_0$ is decreased from $60$\,MeV. On the other hand, the right panel [$(1210\mbox{MeV},1050\mbox{MeV})$] shows that all the masses are stable against the change or $\sigma_0$. Since $\sigma_0$ is an order parameter of chiral symmetry, Figs.~\ref{mass_1} and \ref{mass_2} show that nucleon masses are degenerated to chiral invariant masses when thechiral symmetry is restored in e.g., high temperature and/or density. \section{summary and discussions} \label{sec:summary} We introduced two types of nucleons belonging to $[(\bf{2},\bf{1})\oplus(\bf{1},\bf{2})]$ and $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ representations of the chiral ${\rm SU(2)_L}\otimes{\rm SU(2)_R}$ group together with their parity partners. We constructed an effective chiral Lagrangian based on the parity doublet structure. We fitted model parameters to the masses, decay widths and axial-charges of $N(939)$,$N(1440)$,$N(1535)$ and $N(1650)$. Our results show that there are five groups of solutions which are separated by chiral inavariant masses and mixing structure of nucleons. In Group 1, both the chiral invariant masses, $m_0^{(1)}$ for $[(\bf{2},\bf{1})\oplus(\bf{1},\bf{2})]$ and $m_0^{(2)}$ for $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$, are small as seen from Fig.~\ref{chiralinvariantmass}. In this group, the ground state $N(939)$ is dominated by $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ component and its chiral partner is $N(1440)$. In Group 5, on the other hand, the dominant component of $N(939)$ is $[(\bf{2},\bf{1})\oplus(\bf{1},\bf{2})]$, whose chiral invariant mass is about $1000$\,MeV, and the chiral partner of $N(939)$ is a mixture of negative parity nucleons. We gave predictions off-diagonal elements of axial-charge matrix called transition axial-charges, which shows that some groups can be excluded when some of them are determined by e.g., lattice analysis in future. We also study the change of nucleon masses when the VEV of $\sigma$, $\sigma_0$, is changed. In Groups 1-4, all nucleons masses are decreased with decreasing $\sigma_0$ if two chiral invariant masses are smaller than mass of $N(939)$ as shown in Fig.~\ref{mass_1}. In Group 5, on the other hand, the behavior depends on the value of $m_0^{(2)}$: For small $m_0^{(2)}$ the mass of $N(939)$ is stable for $\sigma_0 > 60$\,MeV and it decreases toward $m_0^{(2)}$, while for large $m_0^{(2)}$ it is stable for all $\sigma_0$. This seems consistent with the lattice analysis in Ref.~\cite{Aarts:2015mma,Aarts:2017rrl}, which shows that, with increasing temperature, the mass of the positive parity nucleon mass is stable, while that of the negative parity nucleon mass decreased. The chiral representation of $[(\bf{2},\bf{3})\oplus(\bf{3},\bf{2})]$ includes $\Delta$ baryon in addition to nucleons. For studying $\Delta$, we need to include $[(\bf{4},\bf{1})\oplus(\bf{1},\bf{4})]$ representations which do not include nucleons. It is interesting to study $\Delta$ baryons by constructing a model including $[(\bf{4},\bf{1})\oplus(\bf{1},\bf{4})]$ representations. [Study of $\Delta$ baryon based on the parity doublet structure are done in e.g., Refs.~\cite{Jido:1999hd,Jido2,Nagata,Takeda:2017mrm,Bicudo:2016eeu}.] We can extend the model to three flavor case based on the SU(3)$_{\rm L}\otimes$SU(3)$_{\rm R}$ symmetry to study hyperons as done in Refs.~\cite{Nemoto1,Chen:2008qv,Chen:2009sf,% Chen:2010ba,Chen:2011rh,Dmitrasinovic1,Nishihara:2015fka,Olbrich:2015gln,% Olbrich:2017fsd}. The parity doublet structure can also be extended to the baryons including heavy quarks as done in Refs.~\cite{Ma:2015lba,Ma:2015cfa,Ma:2017nik,Kawakami:2018olq}. It is interesting to construct nuclear matter in the present model and the density dependences of effective masses of nucleons as done in Refs.~\cite{Hatsuda:1988mv,% Zschiesche:2006zj,Dexheimer:2007tn,Dexheimer:2008cv,Sasaki:2010bp,Gallas:2011qp,% Steinheimer:2011ea,Dexheimer:2012eu,Benic:2015pia,Motohiro:2015taa,Takeda:2017mrm,% Suenaga:2017wbb,Takeda:2018ldi,Suenaga:2018kta,Shin:2018axs}. We leave this for future work. \acknowledgments This work was supported partly by JSPS KAKENHI Grant Number JP16K05345.

d6bb12a2f76058fb2244cb8a54563461acf11a4b

\section{Introduction} Hard X-ray (HXR) emission in solar flares has been studied for decades. In most flares HXR emission is dominated by footpoint sources, but more and more observed flares show a more complex evolution \citep{Krucker2008A&ARv}. In addition to the footpoint sources, a loop-top source is reported in several flare observations (e.g. \citealp{Masuda1994Natur, Krucker2007ApJ, Su2013NatPh}). The mechanism of HXR sources is widely believed to be non-thermal thick target bremsstrahlung. According to the CSHKP flare model, part of the energy released by magnetic reconnection above the flare loop is used to accelerate particles \citep{Carmichael1964NASSP, Sturrock1966Natur, Hirayama1974SoPh, Kopp1976SoPh}. These accelerated electrons are guided by the magnetic field, and move from loop top to footpoints along the flare loops. Footpoint sources are generated when these energetic electrons collide with ambient ions in chromospheric footpoints and produce X-ray photons via bremsstrahlung. Unlike foot-point sources, the loop-top source is usually not considered as a thick target. A coronal loop with a length of $10^9 \ \rm cm$ and a number density of $10^{10} \ \rm cm^{-3}$ is collisionally thin to electrons above $8 \ \rm keV$ \citep{Krucker2008A&ARv}, hence it seems difficult for the loop to stop energetic electrons with an energy of several ten $\ \rm keV$ efficiently enough to produce a loop-top HXR source. However, the efficiency of HXR photon generation increases rapidly when loop apex turbulence is taken into account. Turbulence is suggested to be an efficient trap for high energy electrons in the apex of flare loops \citep{Fang2016ApJ}. If turbulence exists in the apex, energetic electrons can effectively travel a much longer path before they leave the apex and the probability that the electrons collide or interact with ambient ions is effectively increased. Furthermore, more high energy electrons can participate in the bremsstrahlung to increase the intensity of loop-top HXR emission if the turbulence works as an efficient particle trap and accelerator. The efficiency of inverse Compton scattering (ICS) is improved by turbulence as well, since the probability that the electrons interact with soft X-ray (SXR) photons is also increased. Early on ICS was suggested, in addition to bremsstrahlung \citep{Korchak1967SvA}, as an alternative mechanism for HXR emission. The role of ICS can be more important than that of bremsstrahlung in HXR emission under special conditions, such as low plasma density, a hard spectrum of energetic electrons, and an anisotropic distribution of electrons \citep{Korchak1971SoPh, Chen2012ApJ}. \citet{Kontar2014ApJ} suggested the spectrum of loop-top electrons to be harder than expected in studying electron transport in flare loops, and they suggested transport is affected by turbulent pitch-angle scattering. A hard loop-top electron spectrum can imply that ICS may play an important role in the HXR emission in the loop top. The existence of turbulence in the apex of flare loops has been proved by multiple observations (e.g. \citealp{Antonucci1982SoPh, KontarPhysRevLett2017}). However, the origin of this turbulence needs further study. \citet{Fang2016ApJ} proposed a new probable turbulence source to interpret the generation of loop-top HXR emission. This new scenario suggests that Kelvin-Helmholtz instabilities (KHI) play an important role in the generation of turbulence in the apex. According to the CSHKP flare model, high speed evaporation flows are driven when high energy electrons deposit their energy into the chromospheric footpoints. The new scenario suggests that turbulence is produced via KHI when the flows go into the loop apex and interact with each other. High speed evaporation flows with a speed of $200 - 500 \ \rm km/s$ are frequently reported \citep{Feldman1980ApJ, Antonucci1982SoPh, Shimizu1994ApJ, Milligan2009ApJ, Nitta2012SoPh, Tian2014ApJ, Tian2015ApJ}. \citet{Nitta2012SoPh} investigated 13 flare events, 3 of which were found to have high speed flows with a velocity about $500 \ \rm km/s$. Evaporating upflows with speeds higher than $500 \ \rm km/s$ seem relatively rare and only a few cases are reported (e.g. \citealp{Bentley1994ApJ, Tomczak1997A&A, Liu2006ApJ}). While the evaporation flows have a speed of several hundred kilometers per second, the speed of the flows is likely comparable to the local Alfv\'en speed. Therefore, it is reasonable to consider the contribution of KHI. \citet{Fang2016ApJ} proved that KHI can be triggered and turbulence can be produced in the apex of flare loops. In their observational study, \citet{Antonucci1982SoPh} also found a link between high speed evaporation flows and turbulence. The turbulence in hot plasma increases and can reach a maximum turbulent velocity of $122 \ \rm km/s$ when a high speed upflow with a speed of $300 - 400 \ \rm km/s$ has been observed, while the mean turbulent velocities is of the order of $75 \ \rm km/s$ in the preflare active region. We suggest that this increase of turbulent velocity is likely caused by KHI. When the energetic particles deposit their energy into chromospheric footpoints, the energy is unlikely to be equally distributed to two footpoints. \citet{Aschwanden1999ApJ} found that the HXR emissions of two footpoints tend to be asymmetric in their survey, which indicates the asymmetry of energy deposition. The asymmetry of energy deposition is suggested to be determined by the asymmetry of the magnetic field. Figures 1 and 4 in \citet{Krucker2008A&ARv} also indicate a result of asymmetric energy deposition, in which one footpoint is closer to the loop-top HXR or SXR sources than the other. The asymmetry of energy deposition may influence the parameters of evaporation flows and the evolution of the flare loop. In our numerical study of KHI triggering, asymmetric energy deposition strategies are adopted to make our simulations comparable with the observations. We aim to continue the study of \citet{Fang2016ApJ} on the trigger of KHI and the generation of turbulence in flare loops. The set-up of our simulation is introduced in section 2. Energy is deposited into chromospheric footpoints of a coronal loop that drives evaporation flows. The interaction of the flows, trigger of KHI, and generation of turbulence are investigated in section 3. The thermodynamics and radiative evolution of the flare loop and energy cascade process of turbulence are studied with a high spatial resolution case. In addition, EUV and SXR images of the flare loop are synthesized in this section, and then the role of bremstrahlung versus the inverse Compton process in the generation of loop-top HXR sources is investigated. A parameter survey is performed to investigate important factors that influence the trigger of KHI and the generation of turbulence in section 4. We summarize our work in section 5. \section{Numerical set-up} We numerically solved the magnetohydrodynamics (MHD) equations with the MPI-parallelized Adaptive Mesh Refinement Versatile Advection Code \textit{MPI-AMRVAC} \citep{Keppens2012JCoPh, Porth2014ApJS, Xia2018ApJS}. The governing equations are written as follows: \begin{eqnarray} \frac{\partial \rho}{\partial t} + \nabla \cdot (\rho \bm{v}) & = & 0, \\ \frac{\partial \rho \bm{v}}{\partial t} + \nabla \cdot (\rho \bm{v} \bm{v} + p_{\rm{tot}} \bm{I} - \frac{\bm{BB}}{\mu_0}) & = & \rho \bm{g}, \\ \frac{\partial e}{\partial t} + \nabla \cdot (e \bm{v} + p_{\rm{tot}} \bm{v} - \frac{\bm{BB}}{\mu_0} \cdot \bm{v}) & = & \rho \bm{g} \cdot \bm{v} + \nabla \cdot (\bm{\kappa} \cdot \nabla T) \nonumber \\ & & - Q + H, \label{q-en} \\ \frac{\partial \bm{B}}{\partial t} + \nabla \cdot (\bm{v B} - \bm{B v}) & = & 0, \end{eqnarray} where $\bm{I}$ is the unit tensor and $\bm{\kappa}$ is the thermal conductivity tensor, and $\rho$, $\bm{v}$, $T,$ and $\bm{B}$ are density, velocity, temperature, and magnetic field, respectively. The total energy density is given by \begin{equation} e = \frac{p}{\gamma-1} + \frac{\rho v^2}{2} + \frac{B^2}{2 \mu_0}, \end{equation} and the total pressure is given by \begin{equation} p_{\rm{tot}} = p + \frac{B^2}{2 \mu_0}. \end{equation} The terms $Q$ and $H$ in the energy equation (\ref{q-en}) indicate optically thin radiative cooling and parametrized heating, respectively. We solved the MHD equations on a Cartesian box with a domain of $-40\ \textrm{Mm} \leq x \leq 40\ \textrm{Mm}$ and $0\ \textrm{Mm} \leq y \leq 50\ \textrm{Mm}$, but vector quantities are 2.5D (i.e. they have three components). We adopted an abundance ratio of $\rm{He/H = 0.1}$ in the calculations of density and pressure, where the plasma is assumed to be fully ionized. The gravitational acceleration is given by $\bm{g} = g_0 R_{\rm{s}}^2 / (R_{\rm{s}} + y)^2 \bm{\hat{y}}$, where $g_0 = -274\ \rm m\ s^{-2}$ and $R_{\rm{s}}$ is the solar radius. The thermal conductivity tensor is expressed as $\bm{\kappa} = \kappa_{\parallel} \bm{\hat{b}} \bm{\hat{b}}$ where $\bm{\hat{b}}=\bm{B}/B$, which indicates that only thermal conduction along magnetic field lines was considered. The parallel thermal conductivity is given by $\kappa_{\parallel} = 8 \times 10^{-7} T^{5/2} \ \rm{erg\ cm^{-1}\ s^{-1}\ K^{-7/2}}$. We adopted a maximum heat flux $F_{\rm{sat}} = 5 \phi \rho c_{\rm{s}}^3$ to include the saturation effect of thermal conduction \citep{Cowie1977ApJ}, where $\phi = 1$ and $c_s$ denotes the local acoustic speed. The radiative cooling function is expressed as $Q = n_{\rm{e}}^2 \Lambda (T)$, where $n_{\rm{e}}$ is the number density of electrons and $\Lambda (T)$ is a cooling curve supplied by \citet{Colgan2008ApJ}. This curve provides the radiative loss of hot and optically thin plasma; the emission from plasma with a temperature lower than $10,000 \ \rm K$ was not considered. The initial conditions and boundary conditions are as follows. The lower boundary is located in the photosphere and the upper boundary is located in the corona. The initial distribution of temperature consists of two parts: the VAL-C temperature profile \citep{Vernazza1981ApJS} is employed in the region between the lower boundary and $h_{\rm{tra}} =2.543 \ \rm Mm$, and the temperature distribution in the region above $h_{\rm{tra}}$ can be written as \begin{equation} T(y) = [3.5 F_{\rm{c}} (y - h_{\rm{tra}}) / \kappa + T_{\rm{tra}}^{7/2}]^{2/7}, \end{equation} where $F_{\rm{c}} = 2 \times 10^5 \ \rm erg \ cm^{-2} \ s^{-1}$, $\kappa = 8 \times 10^{-7} T^{5/2} \ \rm {erg\ cm^{-1}\ s^{-1}\ K^{-7/2}}$, $T_{\rm{tra}} = 0.447 \ \rm MK$. There is a ghost zone below the lower boundary ($y = 0$) and the thickness of the ghost zone is $2 \ \rm Mm$. We obtained the temperature profile in the ghost zone from a linear extrapolation of the temperature profile above the lower boundary. The plasma density at the bottom of the ghost zone was set to be $n_{\rm{e}} = 7.1 \times 10^{14} \ \rm cm^{-3}$. The initial plasma density in the ghost zone and the simulation region was derived from hydrostatic equilibrium $\partial p / \partial h = \rho g$. As a result, the number density at $y = 0$ is $n_{\rm{e}} = 2.15 \times 10^{14} \ \rm cm^{-3}$. Each boundary contains two layer of ghost cells. For the left and right boundary, $\rho$, $e$, $v_y$, and $B_y$ use a symmetric boundary, while $v_x$, $v_z$, $B_x$, and $B_z$ employ an anti-symmetric boundary. For the upper and bottom boundaries, the velocity components also employ an anti-symmetric boundary. In the lower boundary ghost cells, $\rho$ and $e$ and $\bm{B}$ were fixed at their initial values. For ghost cells of the upper boundary, $\rho$ and $e$ were calculated according to the gravity stratification, where the value of temperature was obtained from extrapolation. The value $\bm{B}$ in the upper boundary was extrapolated assuming zero normal gradient. We imposed zero velocities at the bottom and upper boundary surface. We used the same boundary conditions for all of the simulations. We used adaptive mesh refinement to refine the mesh and increase the spatial resolution based on the spatial smoothness of density and magnetic field. The initial simulation box contains $128 \times 80$ cells in a domain of $\rm 80\ Mm \times 50\ Mm$, and the maximum refinement level is set to five or six. A high spatial resolution of $19.5\ \rm km$ is achieved inside and near the flare loop, which ensured that we could simulate the triggering of KHI and evolution of turbulence. A background heating is employed to heat the corona, and is given by \begin{equation} H_0 = c_0 \exp (- \frac{y}{\lambda_0}), \end{equation} where $c_0 = 10^{-4} \ \rm erg\ cm^{-3} \ s^{-1}$ and $\lambda_0 = 80 \ \rm Mm$. We adopted a force-free magnetic arcade as an initial magnetic configuration, \begin{eqnarray} B_x & = & -B_0 \cos (\frac{\pi x}{L_0}) \sin \theta_0 \exp (- \frac{\pi y \sin \theta_0}{L_0}), \\ B_y & = & B_0 \sin (\frac{\pi x}{L_0}) \exp (- \frac{\pi y \sin \theta_0}{L_0}), \\ B_z & = & -B_0 \cos (\frac{\pi x}{L_0}) \cos \theta_0 \exp (- \frac{\pi y \sin \theta_0}{L_0}), \end{eqnarray} where $\theta_0 = 30^{\circ}$ is the angle between the apex of the magnetic loops and the neutral line, $L_0 = 80 \ \rm Mm$ is the horizontal size of the simulation box, and $B_0 = 80 \ \rm G$ is the strength of magnetic field at the bottom. The resulting magnetic field strength at the loop apex is about $50 \ \rm G$. We performed relaxation with this initial condition and background heating to obtain a stable plasma environment during which the precise thermodynamic balance gets adjusted. When the relaxation is finished, the maximum velocity in the simulation box is smaller than $4\ \rm km~s^{-1}$. The initial and the relaxed-state profiles of density and temperature at $x = 0$ are shown in Figure \ref{rhoT_profile}. \begin{figure}[htbp] \begin{center} \includegraphics[width=0.8\linewidth]{fig1.pdf} \caption{Height profiles of temperature and density before relaxation (dashed lines) and after relaxation (solid lines).} \label{rhoT_profile} \end{center} \end{figure} After the relaxation, we reset the time to zero and deposited flare energy into the chromospheric footpoints of a loop to trigger evaporation flows. According to the CSHKP flare model, high energy electron flows generated in the reconnection above the loop top move along the magnetic field, heat the lower atmosphere, and trigger the evaporation flows. The observation from Ramaty High Energy Solar Spectroscopic Imager (RHESSI) suggests that the high energy electron flows have a double power law spectrum \citep{Holman2003ApJ}. The high energy particles are found to deposit most of their energy in the upper chromosphere owing to Coulomb collisions \citep{Allred2005ApJ}. This effect is incorporated and represented by an energy source $H_1$ added in the upper chromosphere in the flare loops where the initial magnetic loops fulfil the following inequality: \begin{eqnarray} && A_z(24\ \textrm{Mm},0) < A_z(x,y) < A_z(23\ \textrm{Mm}, 0), \label{q-aset}\\ && A_z(x,y) = \frac{B_0 L_0}{\pi} \cos(\frac{\pi x}{L_0}) \exp(- \frac{\pi y \sin \theta_0}{L_0}), \end{eqnarray} where $A_z$ is a component of the magnetic vector potential $\bm{A} = (A_x, A_y, A_z)$ of the force-free magnetic arcade, as we have $\nabla \times \bm{A} = \bm{B}$. The $A_z$ has a maximum value at the point $(x=0, y=0)$ and decreases with distance along any ray from this point. The $A_z(24\ \textrm{Mm},0)$ and $A_z(23\ \textrm{Mm}, 0)$ indicate two magnetic field lines in the $x-y$ plane. We note that the flare energy source $H_1$ is added only in the region given by Equation (\ref{q-aset}). The expression of $H_1$ is written as \begin{equation} H_1 = c_1 \frac{1}{\sqrt{\pi \lambda_l^2}} \exp [\frac{-(y-y_c)^2}{\lambda_l^2}] f(t) g(x), \label{q-H1} \end{equation} where $c_1$ denotes the total energy flux transported by energetic particles, $\lambda_l^2 = 0.05 \ \rm Mm^2$, and the height $y_c = 1.75 \ \rm Mm$ is located in the upper chromosphere. The function $f(t),$ which describes the temporal evolution of energy flux, is written as follows: \begin{equation} f(t) = \frac{1}{\sqrt{\pi \lambda_t^2}} \exp [\frac{-(t-2 \lambda_t)^2}{\lambda_t^2}]. \label{q-ft} \end{equation} Obviously the energy flux has a temporal Gaussian distribution. The parameter $\lambda_t$ denotes the time scale of heating, and the maximum energy flux is achieved at $t = 2 \lambda_t$. The function $g(x),$ which indicates the asymmetry of energy deposition at both footpoints, is given by \begin{equation} g(x) = \begin{cases} asym / (1 + asym), & \quad x < 0 \\ 1 / (1 + asym). & \quad x > 0 \end{cases} \label{q-as} .\end{equation} The parameter $asym$ denotes the ratios of energy deposited at the left footpoint ($x < 0$) to that at the right footpoint ($x > 0$). We set the ratio to 0.8 in most of our simulations, since asymmetric energy deposition is more likely to happen \citep{Aschwanden1999ApJ, Fang2016ApJ}. Since the spatial distribution of the flare energy flux is static instead of dynamic, this method can not adapt to the change of density profile along the loop. The value of energy distributed per unit mass increases rapidly when a significant part of plasma leave this energy deposit area. As a result, the produced evaporation flows tend to have low densities and high speeds. However, this did not influence our research too much because the trigger of KHI is determined by Alfv\'en Mach number of the flows rather than the speed of the flows. In our numerical study, we tended to produce evaporation flows with Alfv\'en Mach numbers comparable to the observational flows, rather than produce evaporation flows that have speeds comparable to the observational flows. \section{Case study} In this section, we study a high resolution case to investigate the temporal evolution of the flare loop, the radiative evolution of the flare loop, and the energy cascade process of turbulence. In this case, we adopted a spatial resolution of $19.5\ \rm km$ within and near the flare loop. The value of $c_1$ in Equation (\ref{q-H1}) is set to $ 1.288 \times 10^{13} \ \rm erg~ cm^{-2}$, which ensures that about $1 \times 10^{29} \ \rm erg$ of energy is deposited into the chromospheric footpoints during the simulation. The amount of energy equals the estimated total power of non-thermal electrons ($> 20 \ \rm keV$) in an M2.2 flare on 2003 June 10 \citep{Milligan2006ApJ}. We used a timescale $\lambda_t = 60 \ \rm s$ in Equation (\ref{q-ft}) to deposit $90\%$ of energy in 4 min and set the parameter $asym$ in Equation (\ref{q-as}) to 0.8 to achieve asymmetric footpoint heating. Temporal evolution of number density, speed, and temperature in this case are shown in Figure \ref{rhovT}. As a result of the sudden heating, the pressure in the chromospheric footpoints increases rapidly and two high speed evaporation flows are launched. Two hot, dense, and fast flows can be found at $t=43 \ \rm s$ in the top panels of Figure \ref{rhovT}. The densities of the evaporation flows are about $2 \times 10^{10} \ \rm cm^{-3}$ and the speeds are about $600 \ \rm km/s$. The Alfv\'en Mach number of the upflows seems to be comparable to that in the flare event reported by \citet{Tian2014ApJ}, where the electron density is of the order of $10^{11} \ \rm cm^{-3}$ and the blueshift is about $~ 260 \ \rm km/s$. Even though the speed of the flows reaches as high as $600 \ \rm km/s$ in this case, no KHI is triggered near the boundary of the loop. The main reason is that the local Alfv\'en speed is also very high, as the Alfv\'en speed is about $600 \ \rm km/s$ in the inner loop and is higher than $4,000 \ \rm km/s$ outside the loop. When the dense flows meet each other near $t=86 \ \rm s$, a shock tube problem with two slow shocks is produced in the apex. The shock tube tends to expand along the magnetic field lines in the beginning, but this tendency is prevented quickly owing to the increase of number density and velocity in the evaporation flows that continue to impinge the bounding shocks. The increase of number density and velocity in the upstream drives the slow shock surfaces to move towards the middle, and the thermal pressure in the region between the two shock surfaces becomes higher and higher because of the plasma inflow. As a result, the high pressure forces the loop to expand vertically, as shown in panels (g)-(i) in Figure \ref{rhovT}. Because of the expansion, the strength of the local magnetic field becomes smaller and the Alfv\'en speed becomes lower in the loop apex. As another result of expansion, the evaporation flows shear with the dense plasma in the apex, and the KHI can be triggered since now both the conditions of high shear velocity and low Alfv\'en speed are satisfied. Thereafter, turbulence is produced by KHI, as shown in panels (j)-(l) of Figure \ref{rhovT}. Temperature fluctuations appear in the loop apex when turbulence has been produced. This temperature feature was pointed out in observational studies by \citet{Jakimiec1998A&A}. \begin{figure*}[htbp] \begin{center} \includegraphics[width=\textwidth]{fig2.pdf} \caption{Temporal evolution of number density (left column), velocity $v = \sqrt{v_x^2+v_y^2}$ (middle column), and temperature (right column) images at $t = 43, 86, 129,$ and $172\ \rm s$. The units of number density, velocity, and temperature are $10^9 \ \rm cm^{-3}$, $100 \ \rm km/s,$ and $\rm MK,$ respectively.} \label{rhovT} \end{center} \end{figure*} We provide the synthesized AIA $131 \ \rm \AA$ emission and thermal SXR emission in $4-10 \ \rm keV$ for comparison with observations. The intensity of the AIA $131 \ \rm \AA$ line is given by \begin{equation} I_{131} = G(T) n_e^2, \end{equation} where $G(T)$ is the contribution function provided by the CHIANTI atomic database \citep{DelZanna2015A&A}. The method to synthesize SXR emission is provided by \citet{Pinto2015A&A} and \citet{Fang2016ApJ}. According to their method, the thermal X-ray photon flux at the photon energy $h \nu$ is \begin{equation} I (h \nu, T) = I_0 \frac{EM}{h \nu \sqrt{k_{\textrm{b}} T}} g_{\textrm{ff}} (h \nu, T) \textrm{exp} (- \frac{h \nu}{k_{\textrm{b}} T}), \end{equation} where $g_{\textrm{ff}}$ is the Gaunt factor for free-free bremsstrahlung emission, which is written as\begin{equation} g_{\textrm{ff}} (h \nu, T) = \begin{cases} 1 , & h \nu \leq k_{\textrm{b}} T \\ (\frac{k_{\textrm{b}} T}{h \nu})^{0.4}, & h \nu > k_{\textrm{b}} T. \end{cases} \end{equation} The coefficient $I_0 = 1.07 \times 10^{-42} $ if the photon flux $I (h \nu, T)$ is measured at a distance of $1 \ \rm AU$ and is expressed in units of $\rm cm^{-2} \ s^{-1} \ keV^{-1}$. For the emission measure of a finite volume of plasma $\textrm{EM} = n_e^2 V$, we used the volume of a cell $V = 19.5 \times 19.5 \times 19.5 \ \rm km^3$. We divided the energy range $4 < h \nu < 10 \ \rm keV$ equally into 60 pieces, and then estimated and added up the photon flux density in each piece to calculate the total SXR flux density of the given range. The total SXR photon flux density can be described by $I_{\textrm{SXR}} = \sum I (h \nu, T) \Delta h \nu$. Integration of this emission flux along the light of sight was not performed, since our simulation data is 2D. Therefore, only the relative intensity of emission is meaningful. The evolution of SXR emission is demonstrated in Figure \ref{emission}. At $t = 86 \ \rm s$, high intensity emission only appears near the chromospheric footpoints, even though the evaporation flows have reached the apex. This is because the footpoints have higher plasma density and temperature ($\sim 17 \ \rm MK$). Later, the apex also shows high emission intensity at $t = 172 \ \rm s$, as the temperature of plasma increases to above $20 \ \rm MK$ via flow interaction. Thereafter, SXR emission disappears near the footpoints at $t = 258 \ \rm s$, as the heating is finished and the local temperature decreases to about $10 \ \rm MK$. Meanwhile, SXR emission at the apex is still strong as the temperature of plasma is still higher than $20 \ \rm MK$. The flare loop has a so-called fat and bright body accompanied by a so-called slim and less bright SXR emission down the legs. This shape is similar to the spider-like structure often observed in SXR images of flares \citep{Zhitnik2003MNRAS, Zhitnik2006SoSyR}. This spider-like shape is maintained until the end of the simulation ($t \approx 9 \ \rm min$). The total SXR flux of the flare heated loop at each time is calculated and plotted in panel (a) of Figure \ref{SXR}. The flux decreases to about $2/3$ of the maximum value when the simulation is finished at $558 \ \rm s$. Periodic signals are clearly demonstrated in the decreasing phase of the profiles. To investigate the periodic feature of the signals, wavelet analysis is performed and the results are demonstrated in panel (b). The original data is detrended before the wavelet analysis is performed. The method we perform in the detrending is given by \begin{equation} s_d (t) = \frac{s_o (t)}{s_b (t)} - 1, \end{equation} where $s_d$ is the detrended data, $s_o$ is the original data and $s_b$ is the smooth estimate of $s_o$. The original data is smoothed with Gaussian filter. We used Morlet wavelet in the analysis and set the wave number to be $\omega_0 = 6$. The data before $t = 43 \ \rm s$ were not taken into account in the wavelet analysis, while the SXR flux in this interval is nearly 0. The results of wavelet analysis show that the signal has a clear single periodicity in the SXR decreasing phase, and the period is about $25 \ \rm s$. The oscillation has a long duration, which is more than five periods. To investigate the nature of the oscillations, we employed running difference analyses. The analyses show that temperature and density near the centre of the apex (i.e. in the turbulent region) also vary with a period of about $25 \ \rm s$. The results of the running difference analysis of temperature are demonstrated in Figure \ref{Te_diff}. The total SXR flux is temporally minimal at $t = 288.1 \ \rm s$ and is located at a peak at $t = 301.0 \ \rm s$. This figure clearly shows that the temperature near the centre of the apex decreases with time when the total SXR flux decreases with time (from $t = 283.8 \ \rm s$ to $t = 288.1 \ \rm s$), and the temperature increases with time when the total SXR flux increases with time (from $t = 288.1 \ \rm s$ to $t = 301.0 \ \rm s$). The variation of density is in phase with the variation of temperature as well. This indicates that the oscillations in the total SXR flux are compressional signals. Oscillations in flares with periods of $\sim$ 1-60 s were interpreted as standing fast sausage modes \citep{Tian2016ApJ}. However, the oscillations in our simulation are difficult to interpret by this standard standing fast sausage wave model, which suggests that the phase speed of the waves is given by \begin{equation} C_p = 2 L / P\,, \end{equation} where $L$ is the loop length and $P$ is the period, and the phase speed should be smaller than the external Alfv\'en speed \citep{Aschwanden2004psci}. For the oscillations in our simulation, we find $2 L / P \approx \rm160 \ Mm / 25\ s = 6400 \ km/s$. This value is much larger than the Alfv\'en speed inner the flare loop (lower than $1000 \ \rm km/s$) and close to the maximum Alfv\'en speed outside the flare loop; the outside Alfv\'en speed ranges from $4000 \ \rm km/s$ to $8000 \ \rm km/s$. One probable interpretation is that the oscillations are standing fast sausage modes, where the waves are reflected before they arrive at the opposing footpoints. Instead, the waves are reflected near the apex because of the rapid change in loop radius. As a result, the wavelength is much shorter than $2 L$. We assume that the wavelength is $2 \times 30 \ \rm Mm = 60 \ \rm Mm$, while the size of the apex is about $20 - 30 \ \rm Mm$. In this assumption, the phase speed is $2400 \ \rm km/s$, which is lower than the Alfv\'en speed outside the apex and is a reasonable value of fast sausage wave phase speed. Besides the $25 \ \rm s$ oscillations near the centre of the apex, another type of oscillations is found in the running difference analyses. These oscillations appear near the boundaries of the flare loop. As shown in Figure \ref{Te_diff}, negative $\Delta T$ propagates from right to left near the upper boundary, while positive $\Delta T$ propagates from right to left near the lower boundary. The positive or negative $\Delta T$ near the boundary is caused by the variation of the loop boundaries. The oscillation can be seen in animated views of the density variation, where the loop boundaries show a fairly coherent swaying motion. This is reflected in the temperature difference views shown in Figure \ref{Te_diff} where these coherent motions give clear correlated patterns in the temperature at the top and bottom edges of the loop (near its apex). These sideways displacements are characteristic for more kink-type perturbations. \begin{figure*}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig3.pdf} \caption{Temporal evolution of synthesized AIA $131\ \rm \AA$ emission ($\rm DN \ cm^{-5} \ s^{-1} \ pix^{-1}$) and SXR emission ($\rm photon \ cm^{-2} \ s^{-1}$).} \label{emission} \end{center} \end{figure*} \begin{figure}[htbp] \begin{center} \includegraphics[width=1.0\linewidth]{fig4.pdf} \caption{ (a) Temporal evolution of the total flux of SXR emission (solid line) and the smooth estimate of the SXR emission (dashed line). (b) Wavelet analyses of the SXR emission. The black contour in panel (b) shows 95\% confidence level. The cross-hatched regions represent cone of influence regions, where the edge effect is important. Three dotted lines indicate the times $t = 288.1, 292.4$ and $301.0 \ \rm s,$ respectively.} \label{SXR} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[width=1.0\linewidth]{fig5.pdf} \caption{Running difference analysis of temperature. (a): $T (t = 288.1~ \textrm{s}) - T (t = 283.8~ \textrm{s})$; (b): $T(t = 292.4~ \textrm{s}) - T (t = 288.1~ \textrm{s})$; and (c): $T (t = 301.0~ \textrm{s}) - T (t = 296.7~ \textrm{s})$. The phases of oscillations at the times $t = 288.1, 292.4$ and $301.0 \ \rm s$ refer to Figure \ref{SXR}.} \label{Te_diff} \end{center} \end{figure} When the density of the plasma and photon density are given, we can compare the role of ICS with that of bremsstrahlung in HXR emission. The equation (5) and (7) in appendix A of \citet{Krucker2008A&ARv} are expressions to estimated volume emissivity of ICS and bremsstrahlung, respectively. Here, we estimate the relative contribution of ICS versus bremsstrahlung with these two equations. For HXR photons of energy $\epsilon$, this ratio is written as \begin{equation} R (\epsilon) = \frac{3 \pi}{2 \alpha} \frac{n_{ph}}{n_i} (2\delta-1) Q(\delta) \left( \frac{\epsilon}{4 \epsilon_i} \right)^{(1-\delta)/2} \left( \frac{\epsilon}{m_e c^2} \right)^{\delta-1/2}, \end{equation} where $n_{ph}$ is the number density of mono-energetic photons that have energy $\epsilon_i$ (and are upscattered by ICS), $n_i$ is the ambient coronal number density, and $\delta$ is the spectral index of energetic electrons. Furthermore, $\alpha$ is the fine-structure constant, and $m_e c^2$ is the rest energy of an electron. The dimensionless function $Q(\delta)$ is given by \begin{equation} Q (\delta) = \frac{2 (11 + 4 \delta + \delta^2)}{(1+\delta)(3+\delta)^2(5+\delta)} \,. \end{equation} The contribution of the inverse Compton process by scattering 2 eV photospheric photons, to 20 keV HXR photons, has been evaluated by \citet{Krucker2008A&ARv}. The authors suggest that the inverse Compton process may play a more important role than bremsstrahlung when the ambient coronal density is lower than $10^9 \ \rm cm^{-3}$. In our simulations, our plasma density is only $10^8 \ \rm cm^{-3}$ before the apex is filled with evaporation plasma, therefore their result indicates that this inverse Compton process may determine loop-top HXR emission only before evaporation flows reach the apex. The role of sustained loop-top turbulence, in higher density environments created by KHI interactions in which SXR photons are abundant, was not accounted for, however. Since \citet{Fang2016ApJ} suggested that high energy electrons scatter SXR photons to HXR photons, and thereby contribute to the loop-top HXR source, we investigate this effect here. Firstly, we need to estimate the number density of SXR photons. According to our SXR synthesized results shown in Figure \ref{emission}, we can assume that the contribution of a single computational cell of volume $19.5 \times 19.5 \times 19.5 \ \rm km^3$ to the SXR flux measured at a distance of $R = \rm 1\ AU = 1.5 \ \times 10^{13} \ cm$ is of order $f_m = 5 \times 10^{-5} \rm \ photon \ cm^{-2} \ s^{-1}$. The volume of the loop-top SXR source can clearly be approximated with a sphere of radius 10 Mm, making the total emitting volume $V_l = (10\ \textrm{Mm})^3 4 \pi / 3 $ and the area of the loop-top surface $S_l = 4 \pi (10\ \rm Mm)^2$. The total SXR flux is then estimated as $f_{t} = (V_l / V) f_m 4 \pi R^2$ and the SXR photon density is accordingly estimated as $n_{ph} = f_t / (S_l c) \approx 200 \ \rm cm^{-3}$, where $c$ is the light speed. To simplify the calculation, all SXR photons are assumed to have energy of 4 keV. The energetic electrons are assumed to have a hard spectrum $\delta = 2$, and the ion number density is obtained directly from our simulations and is of order $n_i = 10^{10} \ \rm cm^{-3}$. According to Equation (22), the relative contributions of scattering SXR photons to HXR photons and bremsstrahlung for the emission of 20 keV HXR photons is then $R\, (20\ \textrm{keV}) \approx 10^{-8}$. This implies that the contribution of the inverse Compton process to scattering SXR photons to HXR photons is very small for the generation of a loop-top HXR source. Still, we emphasize that the formulae used in this work do not account for the extra turbulence we find as an inevitable consequence of KHI. We now address these turbulence properties in what follows. Turbulence plays an important role in the new scenario presented by \citet{Fang2016ApJ}, as it is suggested to be an efficient accelerator for electrons and can trap high energy electrons in the loop top. Therefore, turbulence produced by KHI in the apex should be analysed. Figure \ref{PSD} shows the power spectral density (PSD) of velocity and magnetic field of the turbulence at $t=215 \ \rm s$, and the distribution of magnetic field and the flow field of velocity are also provided. A 2D region (green box in Figure \ref{PSD}) is selected to perform fast Fourier transform, and then the 2D results are organized to obtain 1D spectra. The 1D spectra are functions of wave number $k = \sqrt{k_x^2 + k_y^2}$, where $k_x$ and $k_y$ is the wave number in $x$ and $y$ direction respectively. Island structures and vortices can be found in the plots of magnetic field strength and velocity flow field, respectively. The spectral indexes are close to $-5/3$ in the spatial scale range of $10^{-6} < k < 10^{-5} \ \rm m^{-1} $. The cascade process of energy is clearly illustrated in Figure \ref{psd_evl}, which shows their temporal evolution. \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig6.pdf} \caption{(a) Magnetic field strength at $t = 215 \ \rm s$; (b) spectrum of magnetic field inside the green box; (c) flow field of velocity at $t = 215 \ \rm s$; and (d) spectrum of velocity inside the green box. White lines in (a) denote magnetic field lines.} \label{PSD} \end{center} \end{figure} To catch the energy cascade process, a larger region is selected to analyse the spectra (orange box in Figure \ref{psd_evl}a). Figure \ref{psd_evl}b provides the PSD of velocity at $t = 86, 129$ and $172 \ \rm s$. The value of PSD at the maximum wave number $k \approx 2 \times 10^{-4} \ \rm m^{-1}$ are set to one to compare the profiles. We find that the spectra coincide with each other very well in the range $k > 10^{-5} \ \rm m^{-1}$, which indicates that the energy cascade process stops at the scale $k \approx 10^{-5} \ \rm m^{-1}$. This may explain why the indexes of the spectra departs from $-5/3$ in the scale $k > 10^{-5} \ \rm m^{-1}$ in Figure \ref{PSD}. The energy cascades from large scale to small scale, as seen by comparing the three spectra: an enhancement can be found in scale $6 \times 10^{-6} > k > 2 \times 10^{-6} \ \rm m^{-1}$ at $t = 129 \ \rm s,$ while compared with time $t = 86 \ \rm s$; spectra at $172 \ \rm s$ show an enhancement in $10^{-5} > k > 6 \times 10^{-6} \ \rm m^{-1}$ compared with $t = 129 \ \rm s$. The plasma states at $t = 86, 129,$ and $172 \ \rm s$ are shown on the full domain size in Figure \ref{rhovT}. \begin{figure}[htbp] \begin{center} \includegraphics[width=1.0\linewidth]{fig7.pdf} \caption{(a) Out of plane current component $j_z$ at $t = 172 \ \rm s$; (b) temporal evolution of spectra. The region inside the orange box in panel (a) is selected to calculate the spectra.} \label{psd_evl} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig8.pdf} \caption{(a) Temporal evolution of total kinetic energy (black dashed line), thermal energy (red solid line), and magnetic energy (blue dotted line) for hot plasma ($\rm T \geq 5 \ MK$); (b) evolution of background heating flux (red solid line), and radiative cooling (blue dashed line) for hot plasma.} \label{energy} \end{center} \end{figure} We are also interested in the energy transformation and heat-loss evolution of the flare loop. Therefore, the temporal evolutions of total kinetic energy, thermal energy and magnetic energy are calculated. We define the boundaries of the flare loop with the help of plasma temperature. The region where $T > 5 \ \rm MK$ is assumed to belong to the loop, since the plasma outside the loop has a temperature $T < 2 \ \rm MK$ and the temperature of plasma inside the loop is higher than $10 \ \rm MK$. The results are demonstrated in panel (a) of Figure \ref{energy}. The ratio of thermal pressure to magnetic pressure $\beta$ has a wide range in the loop apex. The value of $\beta$ ranges from 1 to 10 in the dark region in panel (e) of Figure \ref{emission}, while $\beta$ is as high as several tens in the bright region in this panel. In contrast, $\beta$ is about $5 \ \times 10^{-4}$ in the apex and has an order of 1 in the chromospheric footpoints when energy has not been deposited. Before the evaporation flows collide with each other at $t \approx 75 \ \rm s$, the kinetic energy flow has the same value as the thermal energy flow. After the flows meet each other in the apex, compression causes kinetic energy to transform quickly into thermal energy. The evidence can be found in Figure \ref{rhovT}, where the plasma in the apex is rapidly heated to about $ 20 \ \rm MK$ from about $10 \ \rm MK$ when the flows collide with each other. The plasma in the loop remains hot for a long time after most of the energy is deposited in the footpoints, since the radiative cooling is less efficient compared with the heating (Figure \ref{energy}) in the hot loop. The temperature of the loop is about $20 \ \rm MK$ when the simulation is finished at $t = 558 \ \rm s$, and weak turbulence can still be found in the apex. The plasma temperature in our simulation is comparable to that in the flare observations. \citet{Nitta2001ApJ} investigated the loop-top temperatures in 36 flare events, and the temperatures in their study are in the range $18 - 23.5 \ \rm MK$. \section{Parameter surveys} In this section, we investigate what may influence the trigger of KHI and the generation of turbulence. The driving source of the evaporation flows, i.e. the chromospheric energy deposition process, is expected to have great influence on the dynamics of the flare loop. Therefore, we explore parameters of the energy deposition that are considered to be influential: the amount of energy, timescale of deposition, and asymmetry of deposition. The influence of the energy deposition timescale is investigated with two cases in this paper. Both cases have a resolution of $39 \ \rm km$. Two new cases listed in this work are referred to as case 2 and case 3, respectively, while the case in section 3 is referred to as case 1. Case 2 has the same setting as case 1 except for the resolution. Case 3 has $\lambda_t = 90 \ \rm s$ to let $90 \%$ of the energy be deposited in about 6 minutes, while $\lambda_t = 60 \ \rm s$ in case 2. Hence, the amounts of energy are the same but case 2 has a shorter timescale and stronger instantaneous energy flux than case 3. The temporal evolution of the velocity of case 2 is shown in Figure \ref{dtime}a-\ref{dtime}c, and that of case 3 is in Figure \ref{dtime}d-\ref{dtime}f. In case 2, turbulence is produced owing to KHI and vortices can be found in Figure \ref{dtime}c. In contrast, no turbulence is produced and no shear flows can be found in case 3. To study the difference between case 2 and case 3, the parameters of the evaporation flows are analysed. The temporal evolution of density and velocity at the axis of the right loop at a height of $h = 10 \ \rm Mm$ is shown in Figure \ref{rhov}. Figure \ref{rhov} indicates a positive correlation between density or velocity of the flows and the energy flux of footpoint heating. The energy flux is higher in case 2 compared with case 3, since the same amount of energy is deposited in a shorter time. Moreover, the density or velocity of the flow increase faster before a maximum value is achieved in case 2. The flows collide near the apex at $t \approx 70 \ \rm s$ in case 2 and at $t \approx 80 \ \rm s$ in case 3. The maximum speed of the flows is about $700 \ \rm km/s$ in case 2 and about $600 \ \rm km/s$ in case 3 at the measured point. As demonstrated in Figure \ref{rhovT}, the flows collide with each other in the apex and two slow shocks are generated before the KHI is triggered. From the Rankine-Hugoniot equations we know that the propagating velocity of a shock is determined by the velocity and density and pressure in the upstream and downstream. For case 2, the shock surface near the axis propagates towards the downstream since the velocity and density in the upstream increase quickly. At the same time, the shock fronts near the loop boundaries propagate upstream to maintain balances of pressure and mass transport because the velocity and density of flows near the boundary are much smaller than those near the axis. As a result, shear flows are generated in the loop apex and KHI can be triggered (Figure \ref{dtime}b). In contrast, the density and velocity of upstream in case 3 increase slowly, and therefore the shock front near the axis propagates upstream. Because the shock fronts near the boundaries propagate towards upstream as well, no shear flows can be found in the apex and KHI cannot be triggered. In addition, stronger energy flux leads to larger Alfv\'en Mach numbers. The Alfv\'en Mach number at the axis of the right loop at a height of $h = 10 \ \rm Mm$ at $t=151 \ \rm s$ is about $1.5$ in case 2 and about $0.9$ in case 3. This means that KHI is easier triggered in case 2 even if shear flows exist in the apex in both cases. For the same reason, the more energy is deposited the easier KHI is triggered. \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig9.pdf} \caption{Evolution of velocity for case 2 in which $\lambda_t = 60 \ \rm s$ (a, b, c) and case 3 in which $\lambda_t = 90 \ \rm s$ (d, e, f).} \label{dtime} \end{center} \end{figure} \begin{figure}[htbp] \begin{center} \includegraphics[width=1.0\linewidth]{fig10.pdf} \caption{Temporal evolution at the centre of the right loop leg at the height of $h = 10 \ \rm Mm$ of density (black lines) and velocity (red lines) for case 2 with higher energy flux and case 3 (dashed) with lower energy flux.} \label{rhov} \end{center} \end{figure} A new case (case 4) with a resolution of $39 \ \rm km$ is simulated and compared with case 2 to study the influence of the amount of deposited energy. The amount of energy deposited into the footpoints in case 4 is set to $1/3$ of that in case 2 with the same timescale of energy deposition. Figure \ref{denergy}a and \ref{denergy}b show the distribution of density at $t = 151 \ \rm s$ of case 2 and case 4, respectively. In case 2, the density and velocity of the evaporation flows increase quickly, since the energy flux deposited into the footpoints increases quickly. As a result, the flows push both shocks to the middle and high speed flows can go into the loop apex. Therefore, KHI can be triggered and turbulence can be produced. In contrast, evaporation flows fail to go into the apex in case 4, since the density and velocity of the flows increase slowly. Consequently, the flows fail to produce turbulence in this case. \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig11.pdf} \caption{(a) Number density at $t = 151 \ \rm s$ for case 2 with more energy deposited at the footpoints; (b) density at $t = 151 \ \rm s$ for case 4 with less energy deposited. The unit of number density is $10^{9} \ \rm cm^{-3}$.} \label{denergy} \end{center} \end{figure} In the observation of solar flares, the HXR emission often exhibits asymmetric features \citep{Aschwanden1999ApJ}. The asymmetry of HXR emission is likely caused by the asymmetric energy deposition at two footpoints. Therefore, it is meaningful to study the influence of asymmetry of energy deposition on the KHI and turbulence. Three new cases (case 5, 6, and 7) are simulated and compared with case 2 to investigate the influence in our work. In these four cases, the amounts of energy deposited in the loop and the timescales of energy deposition are the same, but the ratios of energy deposited at the left footpoint to the right foot-point are different. The ratios in cases 5, 6, and 7 are set to $0$, $0.4$ and $1.0,$ respectively, while that in case 2 is $0.8$. The corresponding simulation results of the out of plane current component $j_z$ at $t = 194 \ \rm s$ are demonstrated in Figure \ref{asym}. The resolution of the cases is $39 \ \rm km$. The vortical structures in each panel provide the information about the location where the KHI is triggered and turbulence is produced. Figure \ref{asym} denotes that the location where turbulence is produced is determined by the asymmetry of energy deposition. KHI is always triggered near the location where the interaction of the flows happen. Since the velocity of evaporation flow has a positive correlation with the energy flux deposited in the footpoint, an asymmetry of energy deposition leads to an asymmetry of flow speed. Consequently, KHI tends to be triggered near the apex when the ratio is close to $1$ and tends to be triggered away from the apex when the ratio is close to $0$, as shown in Figure \ref{asym}. We can still observe asymmetric vortex breaking when the heating is symmetric owing to numerical symmetry breaking and inaccuracy as shown in panel (d). \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig12.pdf} \caption{Out of plane current component $j_z$ at $t = 194 \ \rm s$ for four cases with different energy deposition asymmetries. The asymmetric parameters $asym$ in panels (a), (b), (c), and (d) are 0, 0.4, 0.8 and 1, respectively.} \label{asym} \end{center} \end{figure} When the turbulence is produced away from the apex because of asymmetric footpoint heating, the vortical structures tend to move back and forth near the apex (Figure \ref{motion}). The motion of the vortical structures seems to have a period of about $300 \ \rm s$. The periodic motion is probably a signal of standing slow-mode wave. The length of the loop is about $80 \ \rm Mm$ and the sound speed of $20 \ \rm MK$ plasma is about $500 \ \rm km/s$. The period of a standing slow wave should be about $320 \ \rm s$, which is close to the period of the motion in our simulations. \citet{Fang2015ApJ} and \citet{Mandal2016ApJ} have demonstrated that footpoint heating can generate slow waves and these reflect back and forth in the loop. \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig13.pdf} \caption{Temporal evolution of out of plane current component $j_z$ for case 6 where $asym = 0.4$. Turbulence is in the left of apex at $ t = 215 \ \rm s$, and then appear in the right of apex at $t = 366 \ \rm s$. Finally, turbulence can be found in the left of the apex at $t = 516 \ \rm s$.} \label{motion} \end{center} \end{figure} A case (case 8) in which two evaporation flows shear with each other is also studied. This case has a high resolution of $19.5 \ \rm km$, the same as the resolution of case 1 in section 3. In this new case, energy is deposited at various magnetic field lines to ensure that the evaporation flows shear each other rather than collide with each other. All other cases have the same set of magnetic field lines that are affected at both left and right footpoint, as given by equation~(\ref{q-aset}). In Figure \ref{shear}, the results of this new case are compared to that of case 1 where the flows collide with each other. Vortical structures can be found in both cases, which denotes that KHI can be triggered in both cases. However, more vortices can be found in the situation that the flows collide with each other. The collision of the evaporation flows leads to an expansion of the flare loop, a decrease of magnetic field strength and a decrease of local Alfv\'en speed. Therefore, KHI is more easily triggered in the situation in which the flows collide with each other. The interface of two evaporation flows is wider in the shearing case (case 8), which leads to KHI turbulence appearing on a long part of the loop. In contrast, two evaporation flows make contact with each other only in the apex in the collision case (case), which leads to a more local site of KHI and turbulence development. \begin{figure}[htbp] \begin{center} \includegraphics[width=\linewidth]{fig14.pdf} \caption{Trigger of KHI for two ways the evaporation flows interact with each other. Red denotes the plasma from left footpoint, while blue denotes plasma from the right footpoint. Panels (a) and (c): flows from two footpoints collide with each other; (b)\&(d): flows shear with each other. } \label{shear} \end{center} \end{figure} \section{Summary} We numerically study a new scenario proposed by \citet{Fang2016ApJ} concerning the origin of the observed turbulence in flare loops and generation of loop-top HXR emission. According to this new scenario, KHI can be triggered when chromospheric evaporation flows from two footpoints that meet and interact with each other in the flare loop apex, and turbulence can be produced by the KHI. The turbulence can act as a trapping region and an efficient accelerator to provide energetic electrons. The accelerated electrons can scatter SXR photons to HXR photons via the inverse Compton process to generate the loop-top HXR source, and add to the HXR bremsstrahlung, although the plasma densities are too high according to parametrized emission models \citet{Krucker2008A&ARv}. We focus on the trigger of KHI and the generation of turbulence. In our numerical study, energy is deposited into the chromospheric footpoints of a coronal loop and evaporation flows are produced owing to the sudden heating. The interaction of the flows, trigger of KHI, and generation of turbulence in the apex are investigated. The temporal evolution of a flare loop, radiative evolution of the flare loop, and energy cascade process of turbulence are studied with a high spatial resolution case. Thereafter, a parameter survey is performed to investigate what may influence the trigger of KHI and the generation of turbulence. The main results are summarized in the following: (1) KHI can be triggered when the evaporation flows interact with each other and turbulence can be produced by the KHI in our simulations. (2) Thermal SXR emission is synthesized and SXR sources are found in the results. The SXR emission appear in the footpoint at first, and then appear in the apex. The footpoint sources disappear when the footpoint heating is finished, but the loop-top source is maintained for several minutes. (3) The spectral indexes of velocity and magnetic field are close to $-5/3$ in the turbulent region. The energy cascade process in the turbulence is clearly demonstrated in the spectra. (4) When the evaporation flows are produced, the kinetic energy and thermal energy are very close. Kinetic energy is quickly transformed into thermal energy due to compression after the flows collide with each other at $t \approx 75 \ \rm s$ and the loop becomes very hot ($T > 20 \ \rm MK$). (5) The loop is still very hot ($T > 20 \ \rm MK$) when the simulation is finished at $t \approx 10 \ \rm minutes$ owing to inefficient radiative cooling. (6) The KHI is easily triggered when the density and velocity of the evaporation flows increase rapidly. The KHI turbulence is more easily produced when the amount of energy deposited into the footpoints is higher or the timescale of energy deposition is shorter. (7) The location of the turbulence is determined by the ratio of energy deposited at two footpoints. The site of KHI and turbulence development moves away from loop apex when the energy deposited into one footpoint is much more than that deposited into the other footpoint. (8) The conditions in which evaporation flows collide with other and the flows shear with each other are studied, and KHI turbulence can be produced in both conditions. (9) Abundant waves are produced in the flare loops, including compressional modulations in the turbulent zone with a period of about $25 \ \rm s$ and standing slow mode wave with a period of about $300 \ \rm s$. In our simulations, the upward velocity of the flows can reach $700 \ \rm km/s$ (see Figure \ref{rhov}). Such a high velocity is not necessary to trigger KHI. Whether or not KHI can be triggered is determined by the Alfv\'en Mach number, which depends on both velocity and density. In the observations, it does not seem easy for evaporation flows to achieve a velocity of $700 \ \rm km/s$, but the density of plasma tends to be higher. For example, the electron density of the observed evaporation flows is found to be of the order of $10^{11} \ \rm cm^{-3}$ in \citet{Tian2014ApJ}, which is several times what we found in our simulations. Meanwhile, the observed evaporation flows are reported to have a blueshift of $\sim 260 \ \rm km/s$. The Alfv\'en Mach number of the observed evaporation flows is comparable to that in our simulations if the strengths of magnetic field are of the same order. Therefore, we suggest that KHI is not difficult to trigger in solar flares. Future work is required to study the same scenario in 3D setting to get a clear insight into orientational line of sight effects. A flare loop that contains a reconnection site needs to be adopted to incorporate the KHI effects in a real flare scenario. A more dynamic energy deposit method that distributes flare energy base on the density profile along the loop should be adopted. Test particles need to be added into the simulation to study the effect and physics of particle acceleration and trapping in KHI turbulence. The emission fluxes of HXR generated via ICS and bremsstrahlung can be estimated with the distribution of the test particles. Such calculations can quantify the hardness ($\delta$) of the electron spectrum. The inverse Compton mechanism may contribute to loop-top HXR emission before evaporation flows go into the apex. The contribution is then from scattering photospheric lower energy photons to HXR photons rather than from scattering SXR photons to HXR photons. Bremsstrahlung dominates the HXR emission when the apex is filled with evaporation plasma, according to parametrized models that do not incorporate the turbulence effects. Future work must quantify the roles of the inverse Compton and bremsstrahlung mechanisms in the generation of loop-top HXR sources by comparing the HXR emissions produced via these mechanisms. \begin{acknowledgements} This work is supported by the Chinese scholarship council (CSC) and by project GOA/2015-014 (2014-2018 KU Leuven). We would like to thank Tom Van Doorsselaere, Kirit Makwana, Hui Tian, and Zheming Guo for stimulating discussions. The simulations were conducted on the VSC (Flemish Supercomputer Center funded by Hercules foundation and Flemish government). \end{acknowledgements} \bibliographystyle{aa}

a7ea4b2c80bb8a513d8e3e829d3151e8f0aff4b7

\section{Introduction} In recent years, deep neural networks \cite{LeCun2015} have re-emerged as a powerful tool in many nonlinear regression and classification applications, especially in the visual \cite{Voulodimos2018} and audio \cite{Gavat2015} processing domains. The general form of a neural network can be expressed as a sequence of linear and nonlinear mappings. In particular, we consider the following $N$-layer feed-forward network formulation: \begin{equation} y_{i+1} = f_i(W_iy_i + b_i), i = 1,..., N \label{eqn:nn} \end{equation} where $f_i$ is a nonlinear function mapping vectors of length $m_i$ to vectors of lengths $n_{i+1}$. They can be vectorized activation functions (ReLu, sigmoid, tanh, etc.), but can also be other functions such as pooling, softmax and even linear functions such as the identity function. Each $W_i$ is a matrix of length $m_i\times n_i$ and $y_i$ is a vector of length $n_i$, although for easier visualization and interpretation, they can be represented as (or reshaped into) a multidimensional tensor with the same number of elements. The vectors $b_i$ are of length $m_i$. The vectors $y_1$ and $y_N$ are the input and the output vectors of the entire network respectively and thus the network in Eq. (\ref{eqn:nn}) describes a nonlinear mapping between input and output. The functions $f_i$ are assumed to be fixed and given. The goal in training a network for classification and regression is that given $K$ pairs of vectors $(x_k, z_k)$, find the set of weights $W_i$ and biases $b_i$ such that $\sum_k d(\tilde{z}_k, z_k)$ is minimized, where $\tilde{z}_k$ is the output from the neural network when $x_k$ is fed as input to the neural network and $d(\cdot,\cdot)$ is an error or loss metric. \section{Model reduction} There are 2 ways in which model reduction is beneficial in deep learning. In the first case, after the model weights are obtained via training, model reduction can be used to reduce the resources needed for inference. In this case, the fixed weight matrices are implemented via approximations such as SVD \cite{Xue2013}. In the second case, which this paper is focused on, we want to reduce the complexity of the deep network that is used during training. One main factor in determining the size and complexity of the network in Eq. \ref{eqn:nn} is the number of trainable weights $n_W$. For the commonly implemented case where every weight can be set independently, this is equal to as the sum of the number of elements in all the matrices $W_i$ and vectors $b_i$. When $n_W$ is large, this can slow down training by requiring more iterations and more computations per iteration. The goal of this paper is to present a general framework for reducing $n_W$ that encompasses many model reduction techniques. The most well known model reduction technique is the introduction of the convolutional layer in convolutional neural networks. By making the weights shift-invariant across the image, the number of independently trainable weights is drastically reduced. This technique works well in problem domains that embody this shift-invariance, and this include approaches that works well in the frequency transform domain, whos transforms are circulant (or block circulant with circulant blocks in the case where the data occurs naturally in 2D). Other recent techniques include replacing the dense weight matrices with circulant matrices \cite{Ding2017}. \section{Model reduction via product-of-sums decompositions} The main idea in this paper is to reduce the full matrix $W_i$ and vectors $b_i$ with a parametrized version that has fewer parameters than the number of entries. In particular, we consider the following decomposition of a matrix\footnote{Even though we focused on decompositions of matrices, this decomposition is also applicable to higher order tensors. In higher order tensors, the sum is well-defined as addition elementwise (as long as they have the same shape) and the product operation can be defined in various ways, e.g. as a tensor product, a $k$-mode tensor product \cite{Kolda2009}, as a t-product \cite{Kilmer2011} or a generalized t-product \cite{Kernfeld2015}.} as a product of sums: \begin{equation}\label{eqn:decomp} W = \prod_{j=1}^p\sum_{k=1}^{s_i} g_{jk}(a_{jk})M_{jk} \end{equation} where $a_{jk}, j = 1,\cdots, p, k = 1,\cdots, s_j$, are the the {\em trainable parameters}, also denoted as the {\em variable parameters}. The total number of trainable parameters $n_W$ in Eq. \ref{eqn:decomp} is $\sum_j s_j $. The {\em fixed parameters} are the matrices $M_{jk}$ and the nonlinear functions $g_{jk}$. The size of the matrices $M_{jk}$ are such that the addition and product operations are valid, i.e. $M_{jk}$ and $M_{jk'}$ have the same dimension and the product of $M_{jk}$ and $M_{j+1,k}$ is a valid operation\footnote{We use the convention that $\prod_i M_i = M_1M_2M_3\cdots$, i.e. increasing indices are to the right.}. If the optimal weight matrix $W_{opt}$ has an (approximate) decomposition with a small number of parameters $a_{ij}$ (i.e. $\sum_i s_i$ is small), then this would reduce the model significantly. If a good approximation to $W_{opt}$ is known (either via transfer learning or training a small number of initial epochs), then a low rank decomposition can be computed to reduce the number of parameters. This approach was used in \cite{Sainath2013} for speech processing. Decompositions satisfying specific sparsity patterns of $M_i$ are explored in \cite{Wu2011}. We will restrict ourselves to the case $g_{jk}(x) = x$ in the sequel since the case of nonlinear $g_{jk}$ can be easily addressed when $g_{jk}$ are differentiable. It is clear that convolutional layers \cite{chua-yang88,LeCun2015}, KFC \cite{Grosse2016}, circulant matrices, Toeplitz matrices, Handel matrices, etc. \cite{Zhao2017}, all can be decomposed as Eq. (\ref{eqn:decomp}). The difficult part is to find the fixed parameters and the decompositions that would work best while keeping the $n_W$ small. The proposed framework offers more flexibility in choosing these parameters and the number of them. For instance, to approximate an $n\times n$ matrices using circulant matrices or the Adaptive Fastfood Transform \cite{Yang2015}, the number of parameters are fixed at $n$ and $3n$ respectively, whereas in the current approach, the number of trainable parameters can be varied arbitrarily, as we will illustrate in subsequent examples. Note that this decomposition can be easily implemented in current deep learning architectures. The weight matrix is a linear with respect to the trainable parameters, and thus its partial derivative with respect to the error will be related by the same linear operator. For example, for $W$ decomposed as a sum of matrices $M_k$, and a layer is described as $ y = f(\sum_{k}a_k M_k x + b)$, then $\frac{\partial y}{\partial a_k} = \nabla f( \sum_{k}a_k M_k x + b)M_kx$. Product of matrices are equivalant to multiple layers in a neural network where the nonlinearity is the identity function and its partial derivative for $y = f((\prod_i\alpha_i M_i) x+b)$ is equal to $\frac{\partial y}{\partial a_j} = \nabla f((\prod_i \alpha_i M_i) x + b) \prod_{i\neq j} \alpha_i (\prod_i M_i) x$. There are several benefits in reducing the number of trainable parameters. The first is that the gradient and Hessian (or approximate Hessian in quasi-Newton methods) are smaller objects and this can speed up training. The second is that the number of trainable prameters can be viewed as knobs adjusting the strengths of a small number of operators described by $M_{jk}$ and $b_i$. For a single fixed matrix $W_i$ that have been obtained after training, there exists hard coded approaches using physics \cite{Lineaat8084} or in silicon (using ROM or FPGA). If each of the $M_{jk}$ and $b_i$ are hard coded using these approaches then we can build a hardware-based classifier than can be trained using much fewer variable memory. Another benefit is that there is less chance of overfitting and thus reducing the need of a dropout layer. On the other hand, there is a tradeoff in performance as the model reduction reduces the degree of freedom. \section{Decomposition as a sum of matrices: $p=1$} In this section we consider the simplest case where the matrix $W$ is decomposed as a linear combination of matrices with the linear coefficients being the trainable parameters, i.e. the decomposition can be reduced to: \begin{equation} \label{eqn:lindecomp} W = \sum_{i=1}^s a_i M_i \end{equation} What should $s$ be, i.e. how many trainable parameters are needed? If $M_1$ is close to the optimal weight matrix $W_{opt}$, then we can trivially choose $s=1$. In practice, the optimal matrix $W_{opt}$ is unknown and the goal is to find a small set of matrices $M_i$ such that the $W_{opt}$ for the problem at hand is likely in the linear subspace of these matrices. Eq. (\ref{eqn:lindecomp}) can be rewritten as follows in order to be easily implemented in numpy and TensorFlow using the \verb|tensordot| operator: \begin{equation} \label{eqn:lindecomp_matrix} W = [M_1| M_2 | \cdots | M_u] (A\otimes I) \end{equation} where $A$ is the column vector $[a_1, \cdots a_u]^T$. In dimensionality reduction via random projection\cite{Dasgupta2000}, the input row vector $x$ of order $n$ is multiplied (on the right) by a matrix $W'$ of order $n\times m$ where $m << n$ in order to form a much smaller row vector of order $m$ which is then used as input to a neural network with a full dense layer with weight matrix $B$ of order $m\times k$ where $k$ is the number of features in the input layer. This is equivalent to decomposing the full matrix $W$ as $W'B$ where all entries of $B$ are the trainable parameters. This can be rewritten in the form of Eq. (\ref{eqn:lindecomp}) as $W = W'B = \sum_{ij}b_{ij}M_{ij}$ where $b_{ij}$ is the $ij$-th entry of $B$, $E_{ij}$ is a $0-1$ matrix of order $m\times k$ whose $ij$-th element is 1 and $0$ elsewhere and $M_{ij} = W'E_{ij}$. Next we will illustrate this decomposition with an example application. \subsection{Example 1: linear regression on MNIST dataset} We consider a single layer neural network to classify the MNIST handwritten digits dataset \cite{LeCun2010}. As is well known, the accuracy on the testing set is approximately 92\%. The input are 28 by 28 grayscale images and the output is a one-shot encoding of the classified digit which is decoded using softmax. Thus the weight matrix $W$ is a 784 by 10 matrix and the bias vector $b$ is a 10 by 1 vector resulting in 7850 trainable parameters (a model which we denote as the reference model\footnote{In the subsequent examples, we will pick as the reference model a classifier architecture in the literature where the weight matrices $W_i$ are not decomposed as Eq. (\ref{eqn:decomp}) and each entry is a trainable parameter.}). We consider 3 ways to reduce the number of model parameters of $W$ as a sum of matrices. In all these cases, we construct a fixed matrix $W'$ of size $784\times m$. The trainable parameters are the entries of a matrix $B$ of order $m\times 10$ and a vector $b$ of length $10$ and $W = W'B$ resulting in a total of $10(m+1)$ trainable parameters. In the first case, $W'$ consists of the map from the image to the $m$ lowest 2-d DCT (Discrete Cosine Transform) coefficients. This is a reasonable choice since image compression algorithms such as JPEG \cite{pennebaker:jpeg1993} leverage the insight that the lowest DCT coefficients contain most of the information relevant to the human observer. In the second case, $W'$ is random matrix with entries taken independently from a standard Gaussian normal distribution as is typically used in random projection methods to reduce dimensionality \cite{Dasgupta2000}\footnote{In the sequel, all random matrices used are generated this way.}. In the third case, $W'$ consists of $m$ columns of a random permutation matrix. This corresponds to a random subsampling of $m$ pixels of the input image to be used for classification. The results are shown in Fig. \ref{fig:MNIST-linear}. We see that with about 10\% of the total number of model parameters of the reference model, we can achieve similar performance when the truncated 2-d DCT matrix $W'$ is used. In fact, using about 22\% of coefficients we obtain a slightly better error rate on the testing data than when all the coefficients are used, indicating either overfitting due to the higher frequency coefficients containing minutiae and fine details or slower convergence to the optimal weights when the full dense matrix $W$ is trained. The performance for the random Gaussian matrix is lower, but again a subset of parameters (about a third) is sufficient to achieve similar performance as the full matrix. Finally, the performance of the reduced model is the worst for random subsampling requiring samples from more than half the image before similar classification performance is achieved. \begin{figure}[htbp] \centering \begin{subfigure}[b]{0.40\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_1.pdf}} \caption{MNIST data.}\label{fig:MNIST-linear} \end{subfigure} \quad \begin{subfigure}[b]{0.40\textwidth} \centerline{\includegraphics[width=\textwidth]{fashion_mnist_figure_1.pdf}} \caption{Fashion MNIST data.}\label{fig:fashionMNIST-linear} \end{subfigure} \caption{Model reduction of the linear model, averaged over 10 trials. The reference model has 7850 trainable parameters. } \end{figure} We also perform the same experiment on Fashion MNIST data \cite{xiao_fashion-mnist:_2017} with similar behavior (Fig. \ref{fig:fashionMNIST-linear}). This illustrates that for both the MNIST and Fashion MNIST classification tasks, about 10\% of the DCT coefficients of the image is sufficient to achieve the linear classifier performance. \section{Decomposition into products of multiple sums: $p > 1$} In the prior example we replace a dense matrix where each element is a trainable parameter with a linear combination of $m$ fixed matricess where the linear coefficients are the trainable parameters. In thise section, we look at the case of multiplying several such sums to form the dense matrix. \subsection{Linear classifier for MNIST} We perform an experiment using $10$ products, i.e. $p = 10$ with $s_i = K$ for all $i$ and picking the entries of matrices $M_{jk}$ from a normal distribution. The number of trainable parameters is $10(K+1)$. If $W$ has order $n\times m$, we choose $M_{jk}$ to be a square matrix of order $n\times n$ for $j=1,\cdots, p-1$ and $M_{jk}$ to be the same order as $W$ for $j=p$. The results are shown in Fig. \ref{fig:MNIST-prodsum}, which includes the data from Fig. \ref{fig:MNIST-linear}. We see that the performance is superior to a single product $(p=1)$ and approaches that of the DCT. Notice that due to the Fast Cosine Transform, the DCT can be decomposed into a product of sparse matrices. It is remarkable that product of random matrices can approach the performance of the truncated DCT which leverages the inherent knowledge of spectral decomposition and the observation that visual tasks can be performed using only a compressed version of the image (corresponding to using only the low frequency DCT components). We expect this product of sums architecture to perform well in non visual (or auditory) classification tasks. \subsection{Example 2: MNIST and Fashion MNIST using LeNet} We consider the LeNet \cite{LeCun1988} architecture, but using ReLu for the activation function instead and using the parameters as described in the Tensorflow tutorial "Deep MNIST for Experts". It consists of 2 convolutional layers, each followed by a maxpooling layer, and followed by a fully connected dense layer, a dropout layer and another fully connected dense layer. We obtain with this reference model an accuracy on the testing dataset of approximately $99.34\%$ and $92.5\%$ for MNIST and Fashion MNIST respectively. The total number of trainable parameters in this reference model is 3274634, with most of them in the 2 fully connected dense layers. In this section we look at various attempts of reducing the number of trainable parameters, while trading off test accuracy. \subsubsection{Convolutional feature extraction layers} The first convolutional layer is a $5\times5$ filter generating $32$ features corresponding to $(5\times 5 + 1)\times 32 = 832$ parameters, whereas the second convolutional layer is a $5\times 5$ filters taking $32$ features and mapping them to $64$ features resulting in $(5\times 5\times 32+1)\times 64 = 51264$ parameters, thus the $2$ convolutional layers has a total of $52096$ trainable parameters. We replace each convolutional layer with a product of weighted sums of random matrices for various values of $s_j$ and $p$ and matrices $M_{jk}$. As shown in Fig. \ref{fig:MNIST-cnn}, similar performance as the reference model can be obtained with far fewer trainable parameters in the convolutional layers on the MNIST dataset. On the Fashion MNIST dataset, an accuracy to within about 1\% with less than 1\% of the trainable parameters in the convolutional layers. \begin{figure}[htbp] \centering \begin{subfigure}[b]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_prodsum.pdf}} \caption{Model reduction in a linear model classifying MNIST data using a product of 10 sums of $K$ matrices for various $K$. The reference model has a total of 7850 trainable parameters.}\label{fig:MNIST-prodsum} \end{subfigure} \quad \begin{subfigure}[b]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_2.pdf}} \caption{Model reduction of the 2 convolutional layers in classifying MNIST and Fashion MNIST data. The reference model's convolutional layers has a total of 52096 trainable parameters.}\label{fig:MNIST-cnn} \end{subfigure} \caption{} \end{figure} Since the convolutional layers encompass only a small fraction of all trainable parameters, reducing the number of parameters in the convolutional layers may not seem like a useful approach in simplifying the network. However, this decomposition effectively decomposed the convolution into a weighted sum of a relatively small number of fixed {\em basic} or {\em core} filters and each such filter can be hardcoded into an optical filter and the processing can be performed in parallel and at the speed of light \cite{Weaver1966}. This is important as the implementation of the convolutional layers in a trained classifier takes up most of the time on a traditional computer. \subsubsection{Dense analysis layers} By replacing both the convolutional layers and the fully connected layers (that have the most trainable parameters) with various ways to decompose the weight matrix as Eq. (\ref{eqn:decomp}). In particular, we consider various combinations of $p$ and $s_i$ and also pick $M_{ik}$ to be both dense random matrices and rank-1 matrix of the form $u\otimes v$ where $u$ and $v$ are random vectors. The results are shown in Fig. \ref{fig:MNIST-dense}. We see that the accuracy for classifying MNIST has decreased, but with far fewer trainable parameters. We see that we get to within about $1\%$ of the accuracy of the reference LeNet architecture using only $0.1\%$ of the number of trainable parameters. In particular, using $200$ and $800$ fixed random matrices for the two convolutional layers, $100$ full rank plus $500$ rank-1 random matrices for the first dense layer and $1024$ random matrices for the second dense layers, we obtained a network with $3554$ trainable parameters and a test accuracy of $98.44\%$. The results for Fashion MNIST (Fig. \ref{fig:FMNIST-dense}) is a little worse, but still close to within $4\%$ with only $0.38\%$ of the number of parameters. \begin{figure}[htbp] \centering \begin{subfigure}[b]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_3.pdf}} \caption{MNIST data.}\label{fig:MNIST-dense} \end{subfigure} \begin{subfigure}[b]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_4.pdf}} \caption{Fashion MNIST data.}\label{fig:FMNIST-dense} \end{subfigure} \caption{Model reduction replacing both convolutional layers and full connected layers with products of sums. The reference model has a total of over $3\cdot 10^6$ trainable parameters.} \end{figure} Table \ref{tbl:comparison} compares this approach and the tradeoff in accuracy with Adaptive Fastfood Transform and KFC on MNIST. Since these prior approaches are special cases of Eq. (\ref{eqn:decomp}), and the number of trainable parameters is constrained by the special structure of the decomposition, the current framework provides a more flexible tradeoff between the number of trainable parameters and accuracy. \begin{table} \begin{center} \begin{tabular}{|l|r|r|} \hline {\bf Deep learning network architecture} & {\bf \# of trainable params} & {\bf Error} \\ \hline\hline Adaptive Fastfood 2048 \cite{Yang2015} & 52124 & 0.73\%\\ \hline Adaptive Fastfood 1024 \cite{Yang2015} & 38821 & 0.72\%\\ \hline KFC-Combined \cite{Grosse2016} & 52500 & 0.57\% \\ \hline KFC-II \cite{Grosse2016} & 27700 & 0.76\% \\ \hline ProfSumNet (Fig. \ref{fig:MNIST-dense}) & 3554 & 1.55\%\\ \hline \end{tabular} \end{center} \caption{Comparison with other methods of reducing trainable parameters on MNIST. Both Adaptive Fastfood Transform and KFC can be considered special cases of Eq. \ref{eqn:decomp} with $M_{jk}$ taking specific forms.} \label{tbl:comparison} \end{table} \subsection{Multi Layer Perceptron (MLP)} We consider a 3-layer neural network with dense connected layers as a reference model for MLP. The 3 layers have 1024, 512 and 10 neurons respectively with a ReLu activation function layer in between two layers. This network achieves 98.2\% accuracy on MNIST with 1333770 trainable parameters. We reduce the network by combining the decompositions described earlier. In particular, the first layer is reduced using the low frequency DCT coefficients, the second layer using a products of sums of random matrices and the third layer is either kept as a dense layer or decomposed as a product of sums. The performance is shown in Fig. \ref{fig:MNIST-mlp}. We see that we can achieve within 1.5\% of the performance of the reference model using only 4.3\% of the number of trainable parameters. \section{Deep learning for vision applications without convolutional layers} Even though convolutional layers appears crucial in modern visual recognition deep learning architectures, there are some issues with such architectures. First of all, the execution of the convolutional layer during classification is time consuming. Since convolution is equivalent to Hadamard product in transform space, this is exploited in \cite{Mathieu2014} to speed up the computation by transform the data to fourier space, apply the filtering via Hadamard product and transform the output back via the inverse fourier transform. Secondly, the filter size is fixed and relatively small, as the number of trainable parameters is proportional to the filter size. In this section we consider an approach of operating the nonlinearity entirely within the frequency transform domain. This approach addresses both the issues above. First of all, it allows the use of weight matrices, whose corresponding filter in the spatial domain is not restricted to a specific filter size. Secondly, the use of the Hadamard product speeds up the computation considerably, especially when the desired filter size is relatively large. Each filter can be considered has a Hadamard product with a diagonal matrix. We construct a neural network consisting of the following layers: \begin{enumerate} \item DCT layer: each image is converted to the 2D DCT via a matrix multiply \item Hadamard product layer 1 with 8 filters corresponding to 8 Hadamard products. This can be viewed as 8 diagonal matrix. We construct them as sum of 100 randomly generated 8-tuples of diagonal matrices. A bias vector is also added \item ReLu nonlinearity layer \item Reduction layer: a linear layer to map vectors of size $w\times h$ to vectors of size $w\times h/4$. \item Hadamard product layer 2 with 8 filters constructed as sum of 200 randomly generated 8-tuples of diagonal matrices. A bias vector is also added \item ReLu nonlinearity layer \item Dense layer 1: maps to 2048 features \item ReLu nonlinearity layer \item Dense layer 2: maps to 10 features \end{enumerate} Using these parameters, the Hadamard layers contains only $300$ trainable parameters, whereas the dense layers are similar (in terms of trainable parameters) to the reference network. We see in Fig. \ref{fig:MNIST-noconv} that we can get $98.2\%$ accuracy without the use of a convolutional layer and thus can be implemented much more quickly. A similar approach is discussed in \cite{Pratt2017}, but the FFT is used instead of DCT and the convolutional layers are matrices whose size is equal to length of the images $\times $ the number of features, thus fixing the number of trainable parameters to be relatively large and limiting the size of the images that can be processed. For instance, the approach in \cite{Pratt2017} reaches an error rate of $2.5\%$ for MNIST, whereas we obtain an error rate of $1.8\%$. \begin{figure}[htbp] \centering \begin{subfigure}[t]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_MLP.pdf}} \caption{Model reduction in classifying MNIST data using a 3-layer MLP. The reference model has a total of over $1.3 \cdot 10^6$ trainable parameters.}\label{fig:MNIST-mlp} \end{subfigure} \quad \begin{subfigure}[t]{0.45\textwidth} \centerline{\includegraphics[width=\textwidth]{figure_5.pdf}} \caption{Deep neural networks using filters in the DCT domain classifying MNIST data.}\label{fig:MNIST-noconv} \end{subfigure} \caption{} \end{figure} \section{Training weights in stages} The same weight matrix $W$ can have different decomposition with different number of trainable parameters. This suggests a training framework in stages where the same weight matrix is trained using successive decomposition with increasing number of trainable parameters. For instance, in Eq. (\ref{eqn:decomp}) suppose that in different stages, $p$ is constant, but the $s_i$'s are increasing in each stage. After training using the decomposition in one stage, the sums in the product can be approximated in the next stage efficiently by means of least squares. results in $W'$. This set of parameters can be used as initial conditions for training using the next stage's decomposition, etc. \section{Storage and implementation issues} When the matrix $W$ and $n_W = \sum_j s_j$ is relatively large, storing many large fixed matrices might overwhelm the storage requirements of the network. In the experiments above, this is addressed by picking $M_{jk}$ to be a small set of full rank matrices, in addition to lower rank matrices which can be stored more compactly, e.g. as sum of outer products. One appeal of using circulant matrices, Haar matrices, or low rank matrices for $M_i$ in existing literature is that they can be easily described. This typically translate to requiring less storage. In \cite{ailon2009} the Fast John-Lindenstrass Transform is used to implementation a random matrix multiplication that is fast and require less storage. An interesting question is how to choose a sequence of $M_i$ that are less "complex" and whose linear combination is rich enough to provide a good coverage. Perhaps a complexity-bounded Johnson-Lindenstrauss Theorem is needed. How should we choose $M_i$ that are less complex? One possibility is to find $M_i$ that has low Kolmogorov complexity and generate them on the fly using the smallest program necessary, but this approach would increase the computation time. Another approach is to have these matrices or their generators hard-coded in hardware which would result in faster execution as a side effect. \section{Degrees of freedom and complexity of problem.} Figs. (\ref{fig:MNIST-linear})-(\ref{fig:fashionMNIST-linear}) suggest that most of the information needed for classification of MNIST resp. Fashion MNIST resides in a small number of features. This is consistent with the image compression paradigm; images contain enough redundancies that can be removed without impacting its visual appearance and hence its recognition. The fact that in the JPEG compression standard, the high frequency DCT coefficients are less important, is consistent that a small number of DCT coefficients are sufficient to achieve high accuracy on MNIST and Fashion MNIST. Given a specific problem domain, an interesting question is what is a small set of $M_i$ that can cover the range of weight matrices while trading off accuracy. This analogy with compression goes further. Suppose we have multiple data sets besides MNIST and Fashion MNIST and train a certain CNN architecture (such as LeNet) on them, each resulting in a specific set of weight matrices. For simplicity, let us assume there is only one trained weight matrix $W_i$ in the network. The index $i$ is determined by the data set used to train it. Given a fixed $n_W$, what would be the set of products of sums of such matrices that is closest to these weight matrices. More specifically, focusing on a specific weight matrix in the CNN architecture, and let $W_i$ be the trained weight matrix (which is not unique and depends on initial conditions) for dataset $S_i$, define $d(W_i,\{M_{jk}\})$ as \[ d(W_i,\{M_{jk}\}) = \min_{a_{jk}} \left\|W_i-\prod_{j=1}^p\sum_{k=1}^s g_{jk}(a_{jk})M_{jk}\right\| \] where $\|\cdot\|$ is generally the Frobenius matrix norm. Then we want to find $M_{jk}$ such that $\mu(\{d(W_i,\{M_{jk}\})\}_i)$ is minimized, where $\mu$ can be the mean or the maximum. The experimental results shown in Figs. (\ref{fig:MNIST-cnn}-\ref{fig:FMNIST-dense}) indicate that the decrease in performance is larger for Fashion MNIST than MNIST suggesting that Fashion MNIST has more degree of freedom and requires more parameters to properly solve the classification problem. \section{Concluding remarks} We show that decomposition of a matrix as a product of sums can be useful in reducing the complexity of the model and can lead to faster implementations. The proposed approach allow for a flexible tradeoff between the number of trainable parameters and the accuracy of the network. On the other hand, the choice of such a decomposition is crucial and an efficient way to search for the best decomposition is important. The experiments also indicate that the number of trainable parameters (i.e. the degrees of freedom) in traditional deep networks is more than necessary, leading to overfitting and the need for drop out layers. Finally, product of sums of random matrices appear to be a promising architecture that is applicable to general problem domains, as contrasted with CNNs which are suitable mainly for image processing tasks. Future work include developing efficient methods for finding the best decomposition.

6d644e4ab339e1d5b3ea5551dffb62b64f0f30b9

\section{Introduction} Many classification problems in mathematics can be formally presented as pairs $(X,E)$, where $X$ is a Polish space and $E$ is an analytic equivalence relation on $X$. For example, the problem of classifying graph structures on domain $\omega$ up to isomorphism, or the problem of classifying self-adjoint operators on a separable Hilbert space up to unitary equivalence, are both instances of this formal setup. In order to compare the relative complexity of two such problems $(X,E),(Y,F),$ one often wants to know whether there exists a map $f\colon X\to Y$ such that: \begin{enumerate} \item $f$ is a {\bf reduction} from $E$ to $F$, i.e., $x E x'\iff f(x)F f(x')$, for all $x,x'\in X$; \item $f$ preserves some structural properties of $(X,E)$ and $(Y,F)$? \end{enumerate} In practice, and since otherwise the question trivializes,\footnote{By the axiom of choice, there is an ``abstract'' reduction from $E$ to $F$ if and only if $|X/E|\leq |Y/F|$.} the reduction $f$ is always assumed to be Borel, or at least Baire measurable. Besides this minimal definability requirement, one often wants $f$ to be sensitive to various other structural properties of $(X,E)$ and $(Y,F)$. For example, in the context of the topological Vaught conjecture it is useful to consider \emph{faithful reductions}, i.e., Borel reductions $f$ from $E$ to $F$ with the additionally property that the $F$-saturation of the image of any $E$-invariant Borel subset of $X$ is a Borel subset of $Y$ \cite{Friedman, Hjorth, Gao2, Gao3}. In the context of ergodic theory, when $X,Y$ additionally support probability measures $\mu,\nu$, and $E,F$ are orbit equivalence relations of measure preserving actions of countable groups, one often works with \emph{orbit equivalences}. These are Borel reductions from $E$ to $F$ which are bijective and measure preserving, i.e., they pull back $\nu$-null sets to $\mu$-null sets \cite{Gab,GabF,KechrisMiller}. In this paper we introduce and study \emph{$*$-reductions}. These are Baire measurable reductions which are category preserving.\footnote{For more information on category preserving maps between Polish spaces one may consult \cite[Appendix A]{Category}.} \begin{definition} Let $X,Y$ be Polish spaces and let $E,F$ be equivalence relations on $X$ and $Y$ respectively. A Baire measurable map $f\colon X\to Y$ is {\bf category preserving} if for every meager subset $M$ of $Y$, $f^{-1}(M)$ is a meager subset of $X$. A {\bf $*$-reduction} from $E$ to $F$ is a Baire measurable and category preserving reduction from $E$ to $F$. We write $E\leq_* F$ when such a $*$-reduction exists. \end{definition} As with Borel reductions, the relation $\leq_*$ induces a preordering among the various classification problems. This preordering reflects the relative complexity between two such problems from the point of view of $*$-reductions. Showing that some classification problem $*$-reduces to another often just amounts to finding a ``canonical way'' of coding $E$ into $F$. However, showing negative results predicates upon developing a basic obstruction theory for $*$-reductions. When it comes to simple Borel and Baire measurable reductions there are many well known descriptive set theoretic and dynamical obstructions \cite{Hjorth, Gao, GOP, Friedman}. Similar obstructions have been developed for orbit equivalences. For example Gaboriau's theory of cost \cite{Gab,GabF} implies that two free groups of different rank can never produce orbit equivalent equivalence relations via free and measure preserving actions on a standard measure space. The main goal of this paper is to develop certain obstructions for $*$-reductions by advancing further some of the techniques introduced in \cite{GOP}. We briefly describe here the main ideas; definitions and details can be found in Section \ref{S: Becker graph embeddings}. Given a Polish $G$-space $X$, let $E^G_X$ be the associated orbit equivalence relation on $X$. In \cite{GOP}, the authors introduce a digraph structure $\mathcal{B}(X/G)$ on the quotient $X/E^G_X$ based on Becker's notion of right-embeddings.\footnote{In \cite{Becker}, Becker studies left-embeddings. Here and in \cite{GOP}, with an eye on the applications, it is convenient to develop everything in terms of right-embeddings. Analogous results hold for left-embeddings.} The \emph{right-Becker graph} $\mathcal{B}(X/G)$ has the elements $[x]$ of $X/E^G_X$ as vertices, and an arrow $[x]\to[y]$ whenever there is a right-Cauchy sequence $(g_n)$ in $G$ so that $g_n y$ converges to $x$. The main structural result in \cite{GOP} (\cite[Proposition 2.8]{GOP}) states that a Baire measurable reduction $f$ from $E^G_X$ to $E^H_Y$ induces an injective digraph homomorphism from $\mathcal{B}(X_0/G)$ to $\mathcal{B}(Y/H)$, where $X_0$ is an invariant generic subset of $X$. Using this result and the fact that the Becker graph associated to the action of a CLI group $H$ is trivial,\footnote{A group is CLI if it admits a complete and left-invariant metric.} they introduce---in analogy to Hjorth's turbulence condition---a new dynamical obstruction for classifying $E^G_X$ by CLI-group actions. Here we show that under $*$-reductions the main structural result from \cite{GOP} can be strengthened to the following. \begin{theorem}\label{T: main} Suppose that $G,H$ are Polish groups, $X$ is a Polish $G$-space, $Y$ is a Polish $H$-space, and $f\colon X\to Y$ is $*$-reduction from $E^G_X$ to $E^H_Y$. Then there is a $G$-invariant dense $G_{\delta}$ subset $X_0$ of $X$ and an $H$-invariant non-meager Baire measurable subset $Y_0$ of $Y$ so that the induced map $[f]$ restricted on $X_0/G$ is an isomorphism from the digraph $\mathcal{B}(X_0/G)$ to the digraph $\mathcal{B}(Y_0/H)$. \end{theorem} This strengthening allows us to utilize ``higher dimensional'' properties of the Becker graphs $\mathcal{B}(X/G)$ and $\mathcal{B}(Y/H)$ as obstructions for $*$-reducing $E^G_X$ to $E^H_Y$. We say that the \emph{dimension} of the Becker graph $\mathcal{B}(X/G)$ is at least $n$, if the combinatorial $n$-cube embeds in $\mathcal{B}(X/G)$. We say that the \emph{generic dimension} of $\mathcal{B}(X/G)$ is at least $n$ if the combinatorial $n$-cube embeds in $\mathcal{B}(X_0/G)$, for every invariant comeager subset $X_0$ of $X$. Similarly, we say the \emph{locally generic dimension of $\mathcal{B}(X/G)$} is at least $n$ if the combinatorial $n$-cube embeds in $\mathcal{B}(X_0/G)$, for every invariant non-meager subset $X_0$ of $X$ with the Baire property. \begin{figure}[!ht] \centering \begin{tikzcd}[column sep=0.9em,row sep=0.8em] & & & & & & & & & & \bullet & \\ & & & & & & & \bullet & & \bullet \arrow[ur] & \bullet \arrow[u] \arrow[from=dr]& \bullet \arrow[ul] \\ & & & & \bullet & & \bullet\arrow[ur] & & \bullet \arrow[ul] & \bullet\arrow[u]\arrow[ur] & \bullet \arrow[ul, crossing over] \arrow[ur, crossing over]& \bullet \arrow[u]\\ \bullet & & & & \bullet\arrow[u] & & & \bullet \arrow[ul]\arrow[ur] & & & \bullet \arrow[ul]\arrow[u]\arrow[ur] & \\ n=0 & & & & n=1 & & & n=2 & & & n=3 & \\ \end{tikzcd} \caption{The combinatorial $n$-cube.}\label{Picture1} \end{figure} From this point of view, the anti-classification criterion in \cite{GOP} corresponds to the fact that CLI groups induce $0$-dimensional orbit equivalence relations and hence they cannot reduce---in a Baire measurable fashion---orbit equivalence relations which are generically at least $1$-dimensional. In Section \ref{S: Applications} we define for every $n>0$ a free action of $S_{\infty}$ on a Polish space of countable structures $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ which, as we show in Theorem \ref{thm:bklmain}, locally generically $(n-1)$-dimensional but not $n$-dimensional . Theorem~\ref{T: main} then implies that the associated classification problems $(\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}}),\simeq_{\mathrm{iso}})$, $n>0$, are incomparable under $*$-reductions. The structures in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ are \emph{labeled} versions of certain families of structures that were introduced and studied in~\cite{BKL} for their interesting behavior with respect to \emph{disjoint $n$-amalgamation}. In the process of proving Theorem \ref{thm:bklmain} we develop a general method which can be used for computing the generic dimension of other similar problems; see Remark \ref{Remark fin}. In particular, our method implies that $=^{+}$ is the orbit equivalence relation of an $\infty$-dimensional $S_{\infty}$-space and therefore incomparable under $*$-reductions to $(\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}}),\simeq_{\mathrm{iso}})$ for all $n>0$. \subsubsection*{Acknowledgments} We are grateful to Alexander Kechris, Martino Lupini, and Ronnie Chen for many helpful discussions and suggestions. \section{$*$-reductions and Becker graphs}\label{S: Becker graph embeddings} In what follows $X$ and $Y$ will be Polish spaces and $E$ and $F$ will be analytic equivalence relations on $X$ and $Y$ respectively. Often these equivalence relations will be \emph{orbit equivalence relations} associated with a Polish group action. Let $G$ be a Polish group. A {\bf Polish $G$-space} is a Polish space $X$ together with a continuous action $G\times X\to X$ on $X$. The associated {\bf orbit equivalence relation} $E^G_X$ on $X$ is defined by $x E^G_X y \iff [x]=[y]$, where $[x]$ denotes the orbit of $x\in X$ under the action. If $A$ is any subset of $X$, we will denote by $[A]$ the $G$-saturation of $A$: $\{x\mid \exists g\in G\, g x\in A\}$. Note that if $A$ is an analytic subset of $X$, then $[A]$ is analytic. There are some immediate noticeable differences between the usual definable reductions and $*$-reductions. For example, since the preimages of a comeager set under a category preserving map is comeager, an equivalence relation $E$ cannot $*$-reduce to an equivalence relation $F$ with a comeager $F$-class, unless $E$ has also a comeager $E$-class. As a consequence, and in contrast to the Borel reduction hierarchy \cite{Friedman}, countable graph isomorphism is not $\leq_*$-universal amongst orbit equivalence relations of $S_{\infty}$ actions. On the other hand, if $E$ and $F$ are orbit equivalence relations both having a comeager equivalence class, then any Baire measurable reduction from $E$ to $F$ can be modified to a $*$-reduction. This follows simply from the fact that all orbits of a Polish $G$-space are Borel. The problem becomes more interesting when both $E$ and $F$ have only meager equivalence classes, i.e., when generic properties do not concentrate on a single equivalence class. In general $*$-reductions reveal to be much more sensitive to the dynamical aspects of the classification problems under comparison compared to the usual definable reductions. For example, we have the following proposition. Recall that an equivalence relation $E$ on $X$ is {\bf generically ergodic} if every Baire measurable $E$-invariant subset of $X$ is either meager or comeager. \begin{proposition}\label{P: gen erg} Let $E$ and $F$ be analytic equivalence relations on Polish spaces $X$ and $Y$ respectively. If $E\leq_* F$ and $F$ is generically ergodic then $E$ is generically ergodic. \end{proposition} \begin{proof} Let $f$ be the $*$-reduction from $E$ to $F$ and let $A$ be a Baire measurable $E$-invariant subset of $X$. Let $B$ be any $G_{\delta}$ subset of $A$ so that $A\setminus B$ is meager. By restricting $B$ further if necessary we can assume that $f\mathord{\upharpoonright} B$ is continuous. Let $C$ be the $F$-saturation of the image of $B$ under $f$. Then $C$ is an invariant analytic subset of $Y$ and hence it is either meager or comeager. It follows that $A$ is meager or comeager as well. \end{proof} We restrict now our attention entirely to orbit equivalence relations. Recall that a sequence $(g_n)$ in $G$ is Cauchy with respect to some left invariant metric on $G$ if and only if it is Cauchy with respect to any left invariant metric on $G$; see \cite{Becker}. In this case we say that $(g_n)$ is {\bf left-Cauchy}. Similarly we define when $(g_n)$ is right-Cauchy. Let $X$ be a Polish $G$-space and let $x,x'\in X$. Following Becker \cite{Becker}, we say that {\bf $x$ left-embeds in $x'$} if there is a left-Cauchy sequence $(g_n)$ so that $(g_n x)$ converges to $x'$. Similarly we say that {\bf $x$ right-embeds in $x'$} if there is a right-Cauchy sequence $(g_n)$ so that $(g_n x')$ converges to $x$. We recall now some definitions and a result from \cite{GOP}. In view of our applications in Section \ref{S: Applications} we are going to develop here everything in terms of right embeddings. All results hold equally for left embeddings. The {\bf right Becker digraph} $\mathcal{B}(X/G)$ associated to Polish $G$-space is a graph on domain $X/G=\{[x]\mid x\in X\}$, whose arrows are precisely all pairs $([x],[x'])$ such that $x$ right-embeds in $x'$. The main anti-classification result developed in \cite{GOP} was a consequence of the following proposition. Recall that an {\bf $(E,F)$-homomorphism} is any map $f\colon X\to Y$ with \[x E x' \implies f(x) F f(x').\] An $(E,F)$-homomorphism $f$ induces a map $[f]\colon X/E\to Y/F$ between the quotients, sending $[x]$ to $[f(x)]$. This map is injective if and only if $f$ is a reduction. \begin{proposition}[\cite{GOP}, Proposition 2.8]\label{Prop} Let $G,H$ be Polish groups, let $X$ be a Polish $G$-space, and let let $Y$ be a Polish $H$-space. If $f\colon X\to Y$ is Baire measurable $(E^G_X,E^H_Y)$-homomorphism, then there is a $G$-invariant dense $G_{\delta}$ subset $X_0$ of $X$ so that the induced map $[f]\colon X_0/G\to Y/H$ is a graph homomorphism from the digraph $\mathcal{B}(X_0/G)$ to the digraph $\mathcal{B}(Y/H)$. \end{proposition} The above proposition was used in \cite{GOP} in order to show that certain equivalence relations do not reduce in a Baire measurable fashion to any orbit equivalence relation of a CLI group action. In particular, if $H$ is CLI then $\mathcal{B}(Y/H)$ contains only loops~\cite[Lemma 2.7]{GOP}. Hence if $\mathcal{B}(X/G)$ contains non-trivial edges in every invariant $G_\delta$ subset $X_0$ of $X$ then, by Proposition \ref{Prop}, every Baire measurable $(E^G_X,E^H_Y)$-homomorphism would fail to be a reduction. In the context of $*$-reductions, we can strengthen the conclusion of Proposition \ref{Prop} so that $[f]\mathord{\upharpoonright} (X_0/G)$ is an embedding of digraphs rather than an injective digraph homomorphism. This is the essence of Theorem \ref{T: main}. For the proof we will need the following minor strengthening of \cite[Lemma 2.5]{GOP}\footnote{This is essentially \cite[Lemma 3.17]{Hjorth} modified as in the beginning of the proof of \cite[Theorem 3.18]{Hjorth}.}. \begin{lemma}\label{L: orbit continuity} Suppose that $G,H$ are Polish groups, $X$ is a Polish $G$-space, and $Y$ is a Polish $H$-space. Let $C\subseteq X$ be a $G_{\delta}$ subset of $X$ such that for any $x\in C$ the set $\{g\in G \mid gx\in C \}$ is comeager in $G$. Let $f\colon C\to Y$ be a Baire measurable homomorphism from the equivalence relation $E^G_X\mathord{\upharpoonright} C$ to $E^H_Y$. Then there is a dense in $C$, $G_{\delta}$ subset $\widetilde{C}$ of $C$ so that: \begin{enumerate}[(a)] \item $f\mathord{\upharpoonright} \widetilde{C}$ is continuous; \item for any $x\in \widetilde{C}$ the set $\{g\in G \mid gx\in \widetilde{C} \}$ is comeager in $G$; \item for any $x_0\in \widetilde{C}$ and for any open neighborhood $W$ of the identity in $H$ there exists an open neighborhood $U$ of $x_0$ and an open neighborhood $V$ of the identity of $G$ such that for any $x\in U\cap \widetilde{C}$ and for a comeager set of $g\in V$, we have that $f(gx)\in W f(x)$ and $gx\in \widetilde{C}$. \end{enumerate} \end{lemma} \begin{proof} Let $V\subseteq_1 G$ stand for ``$V$ is an open neighborhood of identity in $G$'' and let $\forall^*x\in A$ stand for ``for a comeager collection of elements in $A$.'' The proof is exactly the same as in \cite[Lemma 2.5]{GOP} so we will omit the details. First one needs to show that for any fixed $W\subseteq_1 H$ we have that \[\forall x_0\in C \;\; \forall^* g_0\in G \;\; \; \exists V\subseteq_1 G \;\; \; \forall^{*} g_1 \in V \;\;\; f(g_1 g_0 x_0)\in W f(g_0 x_0).\] Then by an application of Kuratowski-Ulam theorem, and since by assumption we have that $\forall x\in C \; \;\forall^* g\in G \; \; gx\in C$, we get a dense $G_{\delta}$ subset $C_0$ of $C$ so that \[\forall x\in C_0 \;\; \; \exists V\subseteq_1 G \;\; \; \forall^{*} g \in V \;\;\; f(g x)\in W f(x).\] Then we use $C_0$ exactly as in the proof of \cite[Lemma 2.5]{GOP} to define $C_1$ and we finish by setting $\widetilde{C}:=\{x\in C_1 \mid \forall^* g\in G\; \; gx\in C_1\}$. \end{proof} We now proceed to the proof of Theorem \ref{T: main}. \begin{proof}[Proof of Theorem \ref{T: main}] Let $f\colon X\to Y$ be as in the statement of the theorem. As in \cite[Lemma 2.8]{GOP}, if $\widetilde{C}$ is the set provided by Lemma \ref{L: orbit continuity} with $C=X$, then setting $X'_0=\{x\in X\mid \forall^* g\in G \; gx\in \widetilde{C}\}$ we get the invariant dense $G_{\delta}$ set appearing in the statement of Proposition \ref{Prop}, i.e. we get that $[f]\colon X'_0/G\to Y/H$ is homomorphism from $\mathcal{B}(X'_0/G)$ to $\mathcal{B}(Y/H)$. For the convenience of the reader we include a brief sketch: assume that $x,x'$ are in $X'_0$ and $x$ right-embeds in $x'$. Since this is a $G$-invariant property we can assume that $x,x'$ are in $\widetilde{C}$. Let $(g_n)$ be a right-Cauchy sequence with $g_n x'\to x$. As in \cite[Lemma 2.3]{GOP} we can use properties (b) and (c) of Lemma \ref{L: orbit continuity} to slightly modify $(g_n)$ so that $g_nx$ is in $\widetilde{C}$ and so that there is a right-Cauchy sequence $(h_n)$ in $H$ with $f(g_n x)=h_n f(x)$, for all $n>0$. By property (a) of Lemma \ref{L: orbit continuity} we get that $h_n f(x)\to f(x')$. Since $f$ is by assumption a reduction we also have that $[f]$ is injective. In what follows, we will intersect $X'_0$ with another invariant dense $G_{\delta}$ set $X''_0$ with the property that for every $x,x'\in X''_0$, if $f(x)$ right-embeds in $f(x')$, then $x$ right-embeds in $x'$. The desired sets will then be: $X_0=X'_0\cap X''_0$, and $Y_0=[f(X_0)]$. \begin{claim} There exists $D\subseteq Y$ and $h\colon D\to X$ so that: \begin{enumerate} \item $D$ is a $G_{\delta}$ subset of $Y$ with $f^{-1}(D)$ dense $G_{\delta}$ in $X$; \item for every $ y\in D$ we have that $\{g\in H \mid g y \in D\}$ is a comeager subset of $H$; \item $h$ is a Baire measurable reduction from $E^H_Y\mathord{\upharpoonright} D$ to $E^G_X$, with $[f]\circ [h]=\mathrm{id}$ and $[h]\circ [f\mathord{\upharpoonright} f^{-1}(D)]=\mathrm{id}$. \end{enumerate} \end{claim} \begin{proof}[Proof of Claim.] Since $f$ is Baire measurable, we can find a dense $G_{\delta}$ subset $C_0$ of $X$ so that $f\mathord{\upharpoonright} C_0$ is continuous on $C_0$. Let $A=[f(C_0)]\subseteq Y$ be the $H$-saturation of the image of $C_0$ under $f$. Since $F=\{(y,x)\in A\times C_0 \mid y E^H_Y f(x)\}$ is an analytic set, by the Jankov--von Neumann selection theorem \cite[Theorem 18.1]{Kechris} there is Baire measurable\footnote{$\sigma(\Sigma^1_1)$-measurable, in particular.} map $h'\colon A \to C_0$ uniformizing $F$. Let $D_0$ be a $G_\delta$ subset of $A$ so that $B:=A\setminus D_0$ is meager. We claim that the set $D=\{y\in D_0 \mid \forall^* g\in H\, g y \in D_0\}$ and the map $h=h'\mathord{\upharpoonright} D$ are as required. To see this notice that property (3) follows immediately from the fact that $h'$ uniformizes $F$, and property (2) follows from \cite[Proposition 3.2.5(v)]{Gao}. For property (1), $D$ is $G_{\delta}$ since it is the Vaught transform of a $G_{\delta}$ set intersected with a $G_{\delta}$ set; see \cite[Proposition 3.2.7]{Gao}. Finally, notice that since $A$ is $H$-invariant we have that \[f^{-1}(D)= f^{-1}\bigg(\bigg\{y\in (Y\setminus B) \mid \forall^* g\in H\, g y \in (Y\setminus B)\bigg\}\bigg),\] and the set $\big\{y\in (Y\setminus B) \mid \forall^* g\in H\, g y \in (Y\setminus B)\big\}$ is comeager in $Y$ by Kuratowski-Ulam. Since $f$ preserves category and it is continuous on the dense $G_{\delta}$ subset $C_0$ of $X$, $f^{-1}(D)$ is a dense $G_{\delta}$ subset of $X$ as well. \end{proof} By Lemma \ref{L: orbit continuity} applied to $h\colon D\to X$ we can now find a dense $G_\delta$ subset $\widetilde{D}$ of $D$ satisfying the conclusions of Lemma \ref{L: orbit continuity}. By the proof of \cite[Proposition 2.8]{GOP}, as described in the first paragraph of this proof, for every $y,y'\in \widetilde{D}$, if $y$ right-embeds in $y'$ then $h(y)$ right-embeds in $h(y')$. Setting now $X''_0$ to be the Vaught transform $\{x\in X \mid \forall^*g\in G \; gx\in f^{-1}(\widetilde{D})\}$ of $f^{-1}(\widetilde{D})$ we get the desired sets: $X_0=X'_0\cap X''_0$ and $Y_0=[f(X_0)]$. To see this, let $x,x'\in X''_0$ so that $f(x)$ right-embeds in $f(x')$. By replacing $x,x'$ with a generic translate we can assume without the loss of generality that $x,x'\in f^{-1}(\widetilde{D})$. Since $f(x),f(x')\in \widetilde{D}$ we get that $h(f(x))$ right-embeds in $h(f(x'))$. Property (3) of the above claim implies that $x$ right-embeds in $x'$. Finally, since $[f(X_0)]=[f(X_0\cap\widetilde{C})]$ and $f\mathord{\upharpoonright}\widetilde{C}$ is continuous, we have that $Y_0$ is analytic and hence Baire measurable. Moreover, since $f^{-1}(Y_0)=X_0$, $f$ is category preserving, and $X_0$ is comeager in $X$, $Y_0$ cannot be meager in $Y$. \end{proof} \section{Higher dimensional obstructions}\label{S: Applications} For every $n\geq 0$, the {\bf combinatorial $n$-cube} is the poset category $\Delta^{n-1}$, whose set of objects is the powerset $\mathcal{P}(\{0,\ldots,n-1\})$ of $[n] = \{0,\ldots,n-1\}$ and whose arrows are precisely the inclusions $\sigma\subseteq\tau$ between subsets $\sigma,\tau$ of $\{0,\ldots,n-1\}$. For $n=0$, the combinatorial $n$-cube is the one-object poset $\Delta^{-1}=\mathcal{P}(\emptyset)=\{\emptyset\}$. \begin{figure}[!ht] \centering \begin{tikzcd}[column sep=1 em,row sep=0.9em] & \{0,1,2\} & \\ & & \\ \{0,1\} \arrow[uur] & \{0,2\} \arrow[uu] \arrow[from=ddr]& \{1,2\} \arrow[uul] \\ & & \\ \{0\}\arrow[uu]\arrow[uur] & \{1\} \arrow[uul, crossing over] \arrow[uur, crossing over]& \{2\} \arrow[uu]\\ & & \\ & \emptyset \arrow[uul]\arrow[uu]\arrow[uur] & \\ \end{tikzcd} \caption{Hasse diagram of the combinatorial $3$-cube.} \end{figure} Forgetting for a moment the composition law between the arrows we can view $\Delta^{n-1}$ as a digraph. Let $X$ be a Polish $G$-space and let $C$ be a non-empty $G$-invariant subset of $X$. The {\bf dimension of $C$} is the largest natural number $n\geq 0$ so that the combinatorial $n$-cube embeds in $\mathcal{B}(C/G)$, if such $n$ exists; and it is $\infty$ otherwise. Notice that for every $C$ as above the dimension of $C$ is at least $0$. We say that $X$ is {\bf generically $n$-dimensional} if $n$ is the largest element in $\{0,1,\ldots\}\cup\{\infty\}$ so that the dimension of every invariant comeager subset $X$ is at least $n$. Similarly we say that $X$ is {\bf locally generically $n$-dimensional} if $n$ is the largest element in $\{0,1,\ldots\}\cup\{\infty\}$ so that the dimension of every invariant non-meager subset of $X$ with the Baire property is at least $n$. If $\mathrm{dim}(X)$ denotes the dimension of $X$, $\mathrm{dim}^{*}_{\forall}(X)$ denotes the generic dimension of $X$, and $\mathrm{dim}^{*}_{\exists}(X)$ denotes the locally generic dimension of $X$, then we always have that $\mathrm{dim}(X)\geq \mathrm{dim}^{*}_{\forall}(X)\geq \mathrm{dim}^{*}_{\exists}(X)$. While the inequalities are strict in general, the last two quantities are equal whenever $X$ is generically ergodic. The obstruction, developed in \cite{GOP}, for classifying orbit equivalence relations by CLI group actions relied on the fact that the image of a combinatorial $1$-cube under an injective digraph homomorphism can never be a self-loop. Working with digraph embeddings---rather than just injective homorphisms---allows us to utilize combinatorial $n$-cubes as obstructions for classification under $*$-reductions, even when $n>1$. \begin{theorem}\label{T: main 2} Let $X$ be a Polish $G$-space and let $Y$ be a Polish $H$ space where $G$ and $H$ are Polish groups. Let also $n\in \{0,1,2,\ldots\}\cup\{\infty\}$. \begin{enumerate} \item If $X$ is generically $n$-dimensional and the dimension of $Y$ is less than $n$, then $E^G_X$ does not $*$-reduce to $E^H_Y$. \item If the dimension of $X$ is less than $n$ and $Y$ is locally generically $n$-dimensional, then $E^G_X$ does not $*$-reduce to $E^H_Y$. \end{enumerate} \end{theorem} \begin{proof} The proof is a direct consequence of Theorem \ref{T: main}. \end{proof} In the rest of this section we develop examples of Polish $G$-spaces $X$ which are $n$-dimensional and locally generically $n$-dimensional. We then use Theorem \ref{T: main 2} to deduce that these spaces are pairwise incomparable under $*$-reductions. Let $\mathcal{L}$ be a countable language. We consider the space $\mathcal{X}_{\mathcal{L}}$ of $\mathcal{L}$-structures with domain $\omega$. For any formula $\varphi(x)$ and any tuple $a$ from $\omega$, define \[[\varphi(a)] = \{A\in \mathcal{X}_{\mathcal{L}} \mid A\models \varphi(a)\}.\] Then a subbasis for the topology on $\mathcal{X}_{\mathcal{L}}$ is given by the sets of the form $[R(a)]$ and $[\lnot R(a)]$ for every $n$-ary relation symbol $R$ in $\mathcal{L}$ and every $n$-tuple $a$ from $\omega$, and $[f(a) = b]$ for every $n$-ary function symbol $f$ in $\mathcal{L}$, every $n$-tuple $a$ from $\omega$, and every element $b\in \omega$. That is, $\mathcal{X}_{\mathcal{L}}$ is a Polish space homeomorphic to a product of Cantor spaces $2^{(\omega^{n})}$ for each $n$-ary relation symbol in $\mathcal{L}$ and Baire spaces $\omega^{(\omega^n)}$ for each $n$-ary function symbol in $\mathcal{L}$. Consider now the Polish group $S_\infty$, of all bijections from $\omega$ to $\omega$, endowed with the pointwise convergence topology. There is a natural continuous action of $S_\infty$ on $\mathcal{X}_{\mathcal{L}}$; namely, if $g\in S_\infty$ and $A\in\mathcal{X}_{\mathcal{L}}$, then $gA$ is the unique $B\in\mathcal{X}_{\mathcal{L}}$ so that for every tuple $a=(a_0,\ldots,a_{n-1})$ in $\omega$ and every quantifier free formula $\varphi$ we have that \[B\models \varphi(a_0,\ldots,a_{n-1}) \iff A\models \varphi(g^{-1}(a_0),\ldots,g^{-1}(a_{n-1}))\] In other words, $g A = B$ if and only if $g$ is an isomorphism $A\to B$. The orbit equivalence relation on $\mathcal{X}_{\mathcal{L}}$ induced by the logic action is denoted $\simeq_{\mathrm{iso}}$. Recall that a sequence $(g_n)$ in any Polish group is left-Cauchy if it is Cauchy with respect to some left-invariant metric, and moreover, that this is equivalent with $(g_n)$ being Cauchy with respect to every left-invariant metric; see \cite{Becker}. In $S_{\infty}$ a compatible left-invariant metric is given by $d_l(g,h)=1/2^m$, where $m$ is the least natural number with $g(m)\neq h(m)$. Elements $\gamma$ of the left completion of $S_{\infty}$ can be identified with injections $\gamma\colon \omega\to\omega$ which are not necessarily surjective; see e.g. \cite{Gao paper}. Since for every left-invariant metric $d_l$, in any Polish group, the metric $d_r(g,h):=d_l(g^{-1},h^{-1})$ is right-invariant, we have that a sequence $(g_n)$ in $S_{\infty}$ is right-Cauchy if and only if $(g_n^{-1})$ is left-Cauchy. Similarly to the logic action, where $g A =B$ if and only if $g$ is an isomorphism from $A$ to $B$, the following proposition states that right-embeddings from $A$ to $B$ correspond to model-theoretic embeddings from $A$ to $B$. For left-embeddings the situation is a bit more complicated; see \cite{Becker}. \begin{proposition}\label{Proposition Emb} Let $A,B\in \mathcal{X}_{\mathcal{L}}$ and let $(g_n)$ be a right-Cauchy sequence in $S_{\infty}$. Let also $\gamma\colon \omega\to \omega$ be the injective map that is the limit of $(g^{-1}_n)$. The following are equivalent: \begin{enumerate} \item $(g_n B)$ converges to $A$; \item $\gamma$ is an embedding from $A$ to $B$. \end{enumerate} \end{proposition} \begin{proof} Let $\varphi$ be a quantifier free formula and let $a$ be a tuple in $\omega$. By definition of the logic action, $\varphi(a)$ holds in $g_n B$ for large $n$ if and only if $\varphi(g^{-1}_n a)$ holds in $B$ for large $n$. Since $(g^{-1}_n)$ is converging to $\gamma$, the later is equivalent to $B \models \varphi(\gamma a)$. The rest follows from the fact that $g_nB$ converges to $A$ if and only if for all $a$ and $\varphi$ as above and for all large enough $n$ we have that $A\models \varphi(a)\iff g_n B \models \varphi(a)$. \end{proof} When $\mathcal{K}$ is a class of $\mathcal{L}$-structures, we write $\mathrm{Mod}_{\omega}(\mathcal{K})$ for the subspace of $\mathcal{X}_{\mathcal{L}}$ consisting of structures in $\mathcal{K}$. Abusing notation, we also define \[[\varphi(a)] = \{A\in \mathrm{Mod}_{\omega}(\mathcal{K})\mid A\models \varphi(a)\}.\] When $\mathrm{Mod}_{\omega}(\mathcal{K})$ is an invariant subspace of $\mathcal{X}_{\mathcal{L}}$, the logic action descends to $\mathrm{Mod}_{\omega}(\mathcal{K})$. We are particularly interested in the case when $\mathrm{Mod}_{\omega}(\mathcal{K})$ is a $G_\delta$, and hence Polish, subspace of $\mathcal{X}_{\mathcal{L}}$. Note that $\mathrm{Mod}_{\omega}(\mathcal{K})$ is an invariant Borel subspace of $\mathcal{X}_{\mathcal{L}}$ whenever $\mathcal{K}$ is axiomatizable by a countable $\mathcal{L}_{\omega_1,\omega}$-theory, and it is often possible to check that $\mathrm{Mod}_{\omega}(\mathcal{K})$ is $G_\delta$ by looking at the form of this axiomatization. Given a countable language $\mathcal{L}$, we denote by $\widehat{\mathcal{L}}$ the language $\mathcal{L}\cup \{P_i\mid i\in \omega\}$, where the $P_i$ are unary relation symbols which do not appear in $\mathcal{L}$. Given a class $\mathcal{K}$ of $\mathcal{L}$-structures, we denote by $\widehat{\mathcal{K}}$ the class of $\widehat{\mathcal{L}}$-structures whose reduct to $\mathcal{L}$ is in $\mathcal{K}$ and such that no two elements satisfy exactly the same set of predicates $P_i$, i.e. the satisfy the following axiom: \[\forall x\forall y\, \left((x\neq y)\rightarrow \bigvee_{i\in \omega} (P_i(x)\not\leftrightarrow P_i(y))\right).\tag{A1}\] We call the structures in $\widehat{\mathcal{K}}$ \textbf{labeled} $\mathcal{K}$-structures. \begin{example}\label{Example:ctbl sets of reals} If $\mathcal{L}_\emptyset$ is the empty language and $\mathcal{S}$ is the class of all sets then every structure in $\mathrm{Mod}_{\omega}(\widehat{\mathcal{S}})$ is essentially a sequence of distinct reals (elements of $2^\omega$). Up to isomorphism, such a structure is essentially a countable set of reals. The orbit equivalence relation $\simeq_{\mathrm{iso}}$ on $\mathrm{Mod}_{\omega}(\widehat{\mathcal{S}})$ is often denoted $=^{+}$. For any class of $\mathcal{L}$-structures $\mathcal{K}$, $\mathrm{Mod}_{\omega}(\widehat{\mathcal{K}})$ and $\mathrm{Mod}_{\omega}(\mathcal{K})\times \mathrm{Mod}_{\omega}(\widehat{\mathcal{S}})$ are isomorphic as Polish $S_\infty$-spaces. Note that $\mathrm{Mod}_{\omega}(\widehat{\mathcal{S}})$ is an invariant $G_\delta$ subspace of $\mathcal{X}_{\widehat{\mathcal{L}_{\emptyset}}}$. Hence, when $\mathrm{Mod}_{\omega}(\mathcal{K})$ is an invariant $G_\delta$ subspace of $\mathcal{X}_{\mathcal{L}}$, then $\mathrm{Mod}_{\omega}(\widehat{\mathcal{K}})$ is also an invariant $G_\delta$ subspace of $\mathcal{X}_{\widehat{\mathcal{L}}}= \mathcal{X}_{\mathcal{L}}\times \mathcal{X}_{\widehat{\mathcal{L}_\emptyset}}$. \end{example} Besides the fact that they give rise to natural equivalence relations which generalize $=^{+}$ it is convenient to consider labeled $\mathcal{K}$-structures for two more reasons. First, the logic action on $\mathrm{Mod}_{\omega}(\widehat{\mathcal{K}})$ automatically has meager orbits. This is crucial if one wants to apply Theorem \ref{T: main 2} on $\mathrm{Mod}_{\omega}(\widehat{\mathcal{K}})$ in any meaningful way since the existence of a comeager orbit in any $G$-space $X$ implies that $X$ is generically $0$-dimensional. Second, if $A$ and $B$ are labeled $\mathcal{K}$-structures, then there is at most one embedding $A\to B$. As a consequence, every diagram of labeled $\mathcal{K}$-structures and embeddings is automatically commutative. This labeling trick will allows us to work at the level of the embeddability relation (i.e., the information contained within the Becker digraph) without having to keep track of the composition relation between embeddings. We leave it to future work to develop a functorial version of the results in this paper, perhaps within the framework of Polish groupoids \cite{Lupini}. The main examples of logic actions that we will consider below consist of \emph{labeled BKL$_n$-structures}. Fix $n\geq 1$ and let $\mathcal{L}_n$ be the language which contains a collection $\{s_i\mid i\in \omega \}$ of $n$-ary function symbols, and a collection $\{R_j \mid j\in \omega\}$ of $n$-ary relation symbols. A {\bf BKL$_n$-structure} is any $\mathcal{L}_n$-structure $A$ which satisfies the following list of axioms: \begin{enumerate}[(B1)] \item The $n$-ary relations $(R_j)_{j\in \omega}$ partition $A^n$. \item For any $n$-tuple $a = (a_0,\dots,a_{n-1})$ satisfying $R_j$, we have $s_i(a) = a_0$ for all functions $s_i$ with $i>j$. \item There is no substructure-independent set of size $(n+1)$: If $|B| = n+1$, then there is some $b\in B$ such that $b$ is in the substructure generated by $B\setminus \{b\}$. \item $A$ is locally finite: For any finite tuple $a$ in $A$, the substructure generated by $a$ is finite. \end{enumerate} We will denote by $\mathcal{B}_n$ and by $\mathcal{B}_n^{\mathrm{fin}}$ the classes of all BKL$_n$-structures and all finite BKL$_n$-structures, respectively. Note that the empty structure is a BKL$_n$-structure. The BKL$_n$ structures were introduced by Baldwin, Koerwein, and Laskowski in~\cite{BKL}, based on a similar construction by Laskowski and Shelah in~\cite{LS}. In~\cite{BKL}, the classes $\mathcal{B}_n$ and $\mathcal{B}_n^{\mathrm{fin}}$ were called $\hat{\mathbf{K}}^{n-1}$ and $\mathbf{K}_0^{n-1}$, respectively. Our goal for the rest of this section is to prove the following theorem. \begin{theorem}\label{thm:bklmain} For every $n\geq 1$ the $S_{\infty}$-space $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is $(n-1)$-dimensional and locally generically $(n-1)$-dimensional. \end{theorem} We will conclude with the proof of Theorem \ref{thm:bklmain} at the end of the section after we collect the necessary lemmas. First we record the following corollary which is an immediate consequence of Theorem \ref{T: main 2} and Theorem \ref{thm:bklmain}. \begin{corollary}\label{cor:main} There is no $*$-reduction from $(\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{m}}),\simeq_{\mathrm{iso}})$ to $(\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}}),\simeq_{\mathrm{iso}})$, when $m\neq n$. \end{corollary} Recall from the beginning of this section the poset category $\Delta^{n-1}=(\mathcal{P}(\{0,\ldots,n-1\}),\subseteq)$ where $[n]=\{0,1,\ldots,n-1\}$ is the terminal object. Consider also the full subcategory $\partial\Delta^{n-1}$ of $\Delta^{n-1}$ whose set of objects is $\Delta^{n-1}\setminus \{[n]\}$. We view each class of structures $\mathcal{K}$ as a category whose arrows are embeddings. An {\bf $n$-cube in $\mathcal{K}$} is a functor $\mathcal{A}$ from $\Delta^{n-1}$ to $\mathcal{K}$, i.e., a pair $\mathcal{A} = \big((A_{\sigma})_{\sigma},(f^{\sigma}_{\tau})_{\sigma\subseteq\tau} \big)$, where $(A_{\sigma})_{\sigma}$ is a collection of structures from $\mathcal{K}$, indexed by elements $\sigma$ of $\Delta^{n-1}$; together with embeddings $f^{\sigma}_{\tau}\colon A_{\sigma}\to A_{\tau}$, so that each $f^\sigma_\sigma$ is the identity map, and for every $\sigma\subseteq\tau\subseteq\rho$ we have that $f^{\tau}_{\rho} \circ f^{\sigma}_{\tau}=f^{\sigma}_{\rho}$. Similarly, a \textbf{partial $n$-cube} in $\mathcal{K}$ is a functor from $\partial\Delta^{n-1}$ to $\mathcal{K}$. An $n$-cube in $\mathcal{K}$ is \textbf{disjoint} if $f^\sigma_{\sigma\cup \tau}(A_\sigma) \cap f^\tau_{\sigma\cup \tau}(A_\tau) = f^{\sigma\cap \tau}_{\sigma\cup\tau}(A_{\sigma\cap\tau})$ for all $\sigma$ and $\tau$. Similarly, a partial $n$-cube is disjoint if the same condition holds whenever $\sigma\cup \tau\subsetneq [n]$. The class $\mathcal{K}$ has \textbf{disjoint $n$-amalgamation} if every disjoint partial $n$-cube can be extended to a disjoint $n$-cube. Let $\mathcal{A} = \big((A_\sigma),(f^\sigma_\tau)\big)$ and $\mathcal{B} = \big((A_\sigma),(f^\sigma_\tau)\big)$ be disjoint $n$-cubes. A \textbf{disjoint embedding} $\mathcal{A}\to \mathcal{B}$ is a family of embeddings $h_\sigma\colon A_\sigma\to B_\sigma$ such that for all $\sigma\subseteq \tau$, $h_\tau\circ f^{\sigma}_\tau = g^{\sigma}_\tau \circ h_\sigma$, and for all $\sigma$ and $\tau$, \[(h_{\sigma\cup \tau}\circ f^\sigma_{\sigma\cup \tau})(A_\sigma) \cap g^\tau_{\sigma\cup \tau}(B_{\sigma}) = (h_{\sigma\cup \tau}\circ f^{\sigma\cap \tau}_{\sigma\cup \tau})(A_{\sigma\cap \tau}).\] An $n$-cube is {\bf reducible} if there are $\sigma,\tau\in\Delta^{n-1}$ with $\sigma\not\subseteq\tau$ and some embedding $f\colon A_{\sigma}\to A_{\tau}$ so that $f^{\tau}_{\rho}\circ f=f^{\sigma}_{\rho}$ and $f\circ f^{\rho'}_{\sigma}=f^{\rho'}_{\tau}$, for all $\rho'\subseteq \sigma$ and $\tau \subseteq \rho$. Equivalently, an $n$-cube is reducible if and only if there are $\sigma,\tau\in \Delta^{n-1}$ with $\sigma\not\subseteq\tau$ such that the image of $A_{\sigma} $ in $A_{[n]}$ is contained in the image of $A_{\tau}$ in $A_{[n]}$. \begin{lemma}\label{lem:bklprops} \begin{enumerate} \item The class $\mathcal{B}_n^{\mathrm{fin}}$ has disjoint $k$-amalgamation for all $1\leq k\leq n$. \item If $\mathcal{A}$ is an $n$-cube in $\mathcal{B}_n$ such that $A_\emptyset$ is infinite, then $\mathcal{A}$ is reducible. \end{enumerate} \end{lemma} \begin{proof} \cite[Theorem 3.1.1]{BKL} essentially proves (1) but we repeat the short proof for completeness. Suppose $((A_\sigma)_{\sigma},(f^\sigma_\tau)_{\sigma\subseteq \tau})$ is a disjoint partial $k$-cube in $\mathcal{B}_n^{\mathrm{fin}}$. Let $A_{[k]}$ be the colimit of this diagram in the category of sets. Explicitly, $A_{[k]}$ is the disjoint union of the $A_\sigma$ for $\sigma\in \partial \Delta^{k-1}$, with $f^{\sigma}_\tau(a)$ and $f^{\sigma}_{\tau'}(a)$ identified for all $\sigma\subseteq \tau,\tau'$ and $a\in A_\sigma$. For all $\sigma\in \partial \Delta^{k-1}$, we define $f^\sigma_{[k]}$ to be the canonical inclusion $A_\sigma\to A_{[k]}$. Then the disjointness conditions are satisfied. It remains to make $A_{[k]}$ into a structure in $\mathcal{B}_n^{\mathrm{fin}}$ by defining the functions $s_i$ and relations $R_j$ on all $n$-tuples. So let $a = (a_0,\dots,a_{n-1})$ be an $n$-tuple in $A_{[k]}$. If $a$ is in the image of any $A_\sigma$ for $\sigma\in \partial \Delta^{k-1}$, there is a unique way to define the $s_i$ and $R_j$ so that $f^\sigma_{[k]}$ is an embedding. So we may assume that $a$ is not contained in the image of any $A_\sigma$. Enumerate $A_{[k]}$ as $c_0,\dots,c_N$, set $R_N(a)$, and define $s_i(a) = c_i$ for all $i\leq N$ and $s_i(a) = a_0$ for all $i>N$. Axioms (B1) and (B2) are satisfied by construction, and (B4) is trivially satisfied, since $A_{[k]}$ is finite. For (B3), suppose for contradiction that $B\subseteq A_{[k]}$ is substructure-independent, with $|B| = n+1$. Let $B = \{b_0,\dots,b_n\}$, and for all $0\leq i\leq n$, let $b^i = (b_0,\dots,b_{i-1},b_{i+1},\dots,b_n)$. If each $b^i$ is contained in the image of some $A_{[k]\setminus \{j\}}$, then since $k\leq n$, there are $i\neq i'$ and $j$ such that $b^i$ and $b^{i'}$ are both contained in the image of $A_{[k]\setminus \{j\}}$, and hence all of $B$ is contained in the image of $A_{[k]\setminus \{j\}}$. This contradicts the fact that $A_{[k]\setminus \{j\}}\in \mathcal{B}_n^{\mathrm{fin}}$. So there is some $i$ such that $b^i$ is an $n$-tuple not contained the image of any $A_\sigma$. Then $b_i = c_j$ for some $0\leq j\leq N$, and $s_j(b^i) = b_i$, so $B$ is not substructure-independent. For (2), assume towards contradiction that there is an irreducible $n$-cube $\mathcal{A}$ in $\mathcal{B}_n$ such that each structure $A_\sigma$ in $\mathcal{A}$ is infinite. Since $\mathcal{A}$ is not reducible, for each $0\leq i< n$, the image of $A_{\{i\}}$ in $A_{[n]}$ is not contained in the image of $A_{[n]\setminus \{i\}}$ in $A_{[n]}$. Pick $x_i$ in $f^{\{i\}}_{[n]}(A_{\{i\}})\setminus f^{[n]\setminus \{i\}}_{[n]}(A_{[n]\setminus \{i\}})$ for all $i$. Now by axiom (B4), $\langle x_0,\dots,x_{n-1}\rangle$ is finite, so we can pick some $x_n\in f^\emptyset_{[n]}(A_\emptyset)\setminus \langle x_0,\dots,x_{n-1}\rangle$. Then the set $\{x_0,\dots,x_{n}\}$ is substructure-independent. We have already seen that $x_n\notin \langle x_0,\dots,x_{n-1}\rangle$. And for all $0\leq i < n$, $\langle x_0,\dots,x_{i-1},x_{i+1},\dots,x_n\rangle \subseteq f^{[n]\setminus \{i\}} _{[n]}(A_{[n]\setminus \{i\}})$, so $x_i\notin \langle x_0,\dots,x_{i-1},x_{i+1},\dots,x_n\rangle$. This contradicts (B3). \end{proof} The statement of Lemma~\ref{lem:bklprops}(2) also holds for $\widehat{\mathcal{B}_n}$, by taking reducts to $\mathcal{L}_n$. As a consequence, there are no irreducible $n$-cubes in $\widehat{\mathcal{B}_n}$ all of whose structures lie in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$. Most of our remaining work is to prove that, in contrast, $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ has many irreducible $k$-cubes whenever $k<n$. We begin by observing that $\mathcal{B}^{\mathrm{fin}}_n$ is a Fra\"iss\'e\xspace class when $n\geq 2$ (this was also used in~\cite{BKL}). Then we will find an irreducible $k$ cube in $\mathcal{B}_n$ for $1\leq k<n$ such that every structure in the cube is isomorphic to the Fra\"iss\'e\xspace limit of $\mathcal{B}^{\mathrm{fin}}_n$. Finally, we will use the existence of this $k$-cube and a Baire category argument to find $k$-cubes in any comeager subset of $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$. \begin{lemma}\label{lem:bklspaces} \begin{enumerate} \item Let $A$ be a finite BKL$_n$-structure with domain $\{a_0,\dots,a_k\}$. Then there is a finite conjunction of atomic and negated atomic $\mathcal{L}_n$-sentences with parameters from $A$, denoted $\theta_A(a_0,\dots,a_{k})$, such that for any BKL$_n$-structure $B$, $B\models \theta_A(b_0,\dots,b_{k})$ if and only if the map $a_i\mapsto b_i$ is an embedding $A\to B$. \item When $n\geq 2$, $\mathcal{B}_n^{\mathrm{fin}}$ is a Fra\"iss\'e\xspace class. \end{enumerate} \end{lemma} \begin{proof} For any $n$-tuple $c$ from $A$, let $j_c$ be the unique natural number such that $A\models R_{j_c}(c)$, and let $d_c^i = s_i(c)$ for all $i\in \omega$. Then let $\theta_A(a_0,\dots,a_k)$ be the conjunction of the following formulas: \begin{itemize} \item $a_i\neq a_j$, for all $1\leq i < j \leq k$. \item $R_{j_c}(c)$, for each $n$-tuple $c$ from $A$. \item $s_i(c) = d^i_c$, for each $n$-tuple $c$ from $A$ and each $i\leq j_c$. \end{itemize} This finite piece of the diagram of $A$ is sufficient to determine the rest of its diagram. Indeed, by axioms (B1) and (B2), $R_{j_c}(c)$ implies $\lnot R_{j}(c)$ for all $j\neq j_c$ and $s_i(c) = c_0$ for all $i>j_c$. This establishes (1). For (2), it is clear from the definition that $\mathcal{B}_n^{\mathrm{fin}}$ has the hereditary property, and the amalgamation property is Lemma~\ref{lem:bklprops}(1) in the case $k=2$. The amalgamation property implies the joint embedding property, since $\mathcal{B}_n^{\mathrm{fin}}$ includes the empty structure. Finally, the fact that $\mathcal{B}_n^{\mathrm{fin}}$ is countable up to isomorphism follows from part (1), since there are only countably many $\mathcal{L}_n$ formulas. \end{proof} We denote by $\mathcal{B}^*_{n}$ the class of all structures in $\mathcal{B}_{n}$ which are isomorphic to the Fra\"iss\'e\xspace limit of $\mathcal{B}_{n}^{\mathrm{fin}}$. These are exactly the countably infinite BKL$_n$-structures which satisfy the following \textbf{extension axiom} for every pair of finite BKL$_n$-structures $A\subseteq B$ with domains $\{a_0,\dots,a_k\}$ and $\{a_0,\dots,a_k,a_{k+1},\dots, a_\ell\}$, respectively: \[\forall x_0,\dots,x_k\, (\theta_A(x_0,\dots,x_k)\rightarrow \exists x_{k+1},\dots,x_\ell \,\theta_B(x_0,\dots,x_\ell)).\] \begin{lemma}\label{L: in-out} Suppose $1\leq k<n$. Let $\mathcal{A} = \big((A_{\sigma}),(f^{\sigma}_{\tau}) \big)$ be a disjoint $k$-cube in $\mathcal{B}_{n}^{\mathrm{fin}}$. For every $\rho\in \Delta^{k-1}$ and every embedding $h\colon A_{\rho} \to B$ in $\mathcal{B}_{n}^{\mathrm{fin}}$, there is a disjoint $k$-cube $\mathcal{B} = \big((B_{\sigma}),(g^{\sigma}_{\tau}) \big)$ in $\mathcal{B}_{n}^{\mathrm{fin}}$ and a disjoint embedding $(h_\sigma)\colon \mathcal{A}\to \mathcal{B}$ such that $B_\rho = B$ and $h_\rho = h$, and further, for all $\tau$ such that $\rho\not\subseteq \tau$, we have $B_\tau = A_\tau$, $h_\tau\colon A_\tau\to B_\tau$ is the identity map, and $g^\sigma_\tau = f^\sigma_\tau$ for all $\sigma\subseteq \tau$. \end{lemma} \begin{proof} We define $\mathcal{B}$ and $(h_{\sigma})$ in stages, ensuring that the parts of $\mathcal{B}$ and $(h_\sigma)$ that we have defined so far are functorial and satisfy the relevant disjointness conditions. Namely, for each $\tau$, we must check: \begin{enumerate}[(a)] \item For all $\sigma\subseteq \sigma'\subseteq \tau$, $g^\tau_\tau$ is the identity, and $g^{\sigma'}_\tau\circ g^\sigma_{\sigma'} = g^\sigma_\tau$. \item For all $\sigma\subseteq \tau$, $h_\tau\circ f^\sigma_\tau = g^\sigma_\tau\circ h_\sigma$. \item For all $\sigma\cup \sigma' = \tau$, $g^\sigma_{\tau}(B_\sigma) \cap g^{\sigma'}_{\tau}(B_{\sigma'}) = g^{\sigma\cap \sigma'}_{\tau}(B_{\sigma\cap \sigma'})$. \item For all $\sigma\cup \sigma' = \tau$, $(h_{\tau}\circ f^\sigma_{\tau})(A_\sigma) \cap g^{\sigma'}_{\tau}(B_{\sigma'}) = (h_{\tau}\circ f^{\sigma\cap \sigma'}_{\tau})(A_{\sigma\cap \sigma'})$. \end{enumerate} For $\tau$ such that $\rho\not\subseteq\tau$, set $B_{\tau}=A_{\tau}$, let $h_{\tau}\colon A_{\tau}\to B_{\tau}$ be the identity map, and let $g^{\sigma}_{\tau}=f^{\sigma}_{\tau}$ for $\sigma\subseteq \tau$, as required in the statement of the lemma. Conditions (a), (b), (c), and (d) are satisfied for all such $\tau$; (a) by functoriality of $\mathcal{A}$, (b) trivially, and (c) and (d) because $\mathcal{A}$ is disjoint. For $\tau$ with $\rho\subseteq\tau$, we proceed by induction on the size $m=| \tau \setminus \rho|$. For $m=0$, i.e., for $\tau=\rho$, let $h_{\tau}\colon A_{\tau}\to B_{\tau}$ be $h\colon A_{\rho}\to B$. Let $g^\tau_\tau$ be the identity and $g^{\sigma}_{\tau}= h \circ f^{\sigma}_{\tau}$ for $\sigma\subsetneq \tau$. Condition (a) is trivially satisfied when $\sigma' = \tau$, and letting $\sigma\subseteq \sigma'\subsetneq \tau$, we have $\rho\not\subseteq \sigma'$, so $g^{\sigma'}_\tau\circ g^\sigma_{\sigma'} = h\circ f^{\sigma'}_\tau \circ f^{\sigma}_{\sigma'} = h\circ f^{\sigma}_\tau = g^\sigma_\tau$. Condition (b) is trivially satisfied when $\sigma = \tau$, so it suffices to consider $\sigma\subsetneq \tau$. Then $\rho\not\subseteq\sigma$, and we have $g^{\sigma}_{\tau}\circ h_{\sigma}=h_{\tau} \circ f^{\sigma}_{\tau}$, since $h_\sigma$ is the identity map. Condition (c) is trivial when $\sigma = \tau$ or $\sigma' = \tau$, so we may assume $\rho\not\subseteq\sigma$ and $\rho\not\subseteq \sigma'$. Then \[g^\sigma_\tau(B_\sigma) \cap g^{\sigma'}_\tau(B_{\sigma'}) = h(f^\sigma_\tau(A_\sigma)\cap f^{\sigma'}_\tau(A_{\sigma'})) = h(f^{\sigma\cap \sigma'}_\tau(A_{\sigma\cap \sigma'})) = g^{\sigma\cap \sigma'}_\tau(B_{\sigma\cap \sigma'}).\] Similarly, condition (d) is trivial when $\sigma' = \tau$, so we may assume $\rho\not\subseteq \sigma'$. Then \[(h_{\tau}\circ f^\sigma_{\tau})(A_\sigma) \cap g^{\sigma'}_{\tau}(B_{\sigma'}) = h(f^\sigma_\tau(A_\sigma)\cap f^{\sigma'}_\tau(A_{\sigma'})) = (h_{\tau}\circ f^{\sigma\cap \sigma'}_{\tau})(A_{\sigma\cap \sigma'}).\] Assume now that we have defined $B_{\tau}$, $g^\sigma_\tau$, and $h_{\tau}$ for all $\tau$ with $|\tau \setminus \rho|<m$ and all $\sigma\subseteq \tau$, such that conditions (a), (b), (c), and (d) are satisfied. We consider $\tau$ with $|\tau \setminus \rho|=m$. Let $\{a_0,\ldots,a_{m-1}\}$ be an enumeration of the set $\tau \setminus \rho$. First we define a disjoint partial $(m+1)$-cube, i.e.\ a functor $F\colon \partial\Delta^{m}\to \mathcal{B}^{\mathrm{fin}}_n$. Let \[ F(\sigma) = \begin{cases} B_{\rho\cup \{a_i\mid i\in \sigma,i\neq m\}} & m\in \sigma\\ A_{\rho\cup \{a_i\mid i\in \sigma\}} & m\notin \sigma. \end{cases} \] The image of the relation $\sigma\subseteq \sigma'$ under $F$ is defined as the unique choice amongst $f^{F(\sigma)}_{F(\sigma')}$, $g^{F(\sigma)}_{F(\sigma')}$, or $h_{F(\sigma')} \circ f^{F(\sigma)}_{F(\sigma')}$ which makes sense. Functoriality of $F$ follows from conditions (a) and (b) and functorality of $\mathcal{A}$, and disjointness of the $(m+1)$-cube follows from conditions (c) and (d) and disjointness of $\mathcal{A}$. Since $m+1\leq k+1\leq n$, $\mathcal{B}^{\mathrm{fin}}_n$ has disjoint $(m+1)$-amalgamation by Lemma~\ref{lem:bklprops}, so we can extend $F$ to a disjoint $(m+1)$-cube $G\colon \Delta^m\to \mathcal{B}^{\mathrm{fin}}_n$. Set $B_{\tau}=G([m+1])$, and let $h_{\tau}$ be the image of the relation $[m]\subseteq[m+1]$ under $G$. For any subset $\sigma$ of $\tau$, we define $g^{\sigma}_{\tau}$ to be the appropriate embedding in the image of $G$, if $\rho\subseteq \sigma$, or $h_{\tau}\circ f^{\sigma}_{\tau}$ if not. Now conditions (a) and (b) follow from functoriality of $G$ and induction, and conditions (c) and (d) follow from disjointness of $G$ and induction. \end{proof} \begin{lemma}\label{L: in-out2} For all $1\leq k < n$, there is an irreducible $k$-cube in $\mathcal{B}^*_n$. \end{lemma} \begin{proof} We build a sequence of disjoint $k$-cubes in $\mathcal{B}_n^{\mathrm{fin}}$, $\mathcal{A}^i = \big((A^i_\sigma),(f^{\sigma,i}_\tau)\big)$ for $i\in \omega$, together with disjoint embeddings $(h^i_\sigma) \colon \mathcal{A}^i\to \mathcal{A}^{i+1}$ for all $i$. Then for each $\sigma$, we will let $C_\sigma$ be the directed colimit of the $A^i_\sigma$, and for each $\sigma\subseteq \tau$, let $g^\sigma_\tau$ be the natural map $C_\sigma\to C_\tau$ induced by the maps $f^{\sigma,i}_\tau$ for all $i$. This defines a $k$-cube $\mathcal{C} = \big((C_\sigma),(g^\sigma_\tau)\big)$. The idea is to carry out the Fra\"iss\'e\xspace construction simultaneously for each $\sigma$, so that each $C_\sigma$ in $\mathcal{B}_n^*$. We begin with the empty $k$-cube $\mathcal{A}^0$, where $A_\sigma^0$ is the empty structure for all $\sigma$. Suppose we have defined $\mathcal{A}^i$, and we are given some $\rho$, some substructure $B\subseteq A^i_\rho$, and some embedding $e\colon B\to B'$ in $\mathcal{B}^{\mathrm{fin}}_n$. By disjoint $2$-amalgamation, there is a structure $A_\rho^{i+1}\in \mathcal{B}_n^{\mathrm{fin}}$ and embeddings $h^i_\rho\colon A_\rho^i\to A_\rho^{i+1}$ and $e'\colon B\to A_\rho^{i+1}$ such that $h^i_\rho\mathord{\upharpoonright} B = e'\circ e$ and $h^i_\rho(A_\rho)\cap e'(B) = h(B)$. By Lemma~\ref{L: in-out}, we can extend $A_\sigma^{i+1}$ to a disjoint $k$-cube $\mathcal{A}^{i+1}$ and extend $h^i_\rho$ to a disjoint embedding $(h^i_\sigma)\colon \mathcal{A}^i\to \mathcal{A}^{i+1}$. Taking care in the course of the construction to handle every embedding from a substructure of some $A^i_\rho$ in this way, we may ensure that each $C_\sigma$ is in $\mathcal{B}_n^*$. It remains to show that the $k$-cube $\mathcal{C}$ is irreducible. Suppose $\sigma\not\subseteq \tau$. We will show that $g^\sigma_{[k]}(C_\sigma)\not\subseteq g^\tau_{[k]}(C_\tau)$. Pick some $i$ such that the extension $h_\sigma^i\colon A_\sigma^i\to A_\sigma^{i+1}$ is determined by some proper extension of a substructure of $A_\sigma^i$. Since $A_\sigma^{i+1}$ is defined by disjoint $2$-amalgamation, there is some $x_{i+1}\in A_\sigma^{i+1}\setminus h_\sigma^i(A_\sigma^i)$, and further, by Lemma~\ref{L: in-out}, we know that $h^{i}_\tau\colon A_\tau^{i}\to A_\tau^{i+1}$ is the identity map, since $\sigma\not\subseteq \tau$. Let $y_{i+1} = f^{i+1,\sigma}_{[k]}(x)\in A^{i+1}_{[k]}$. Then $y_{i+1}\notin f^{i+1,\tau}_{[k]}(A^{i+1}_\tau)$. Indeed, if it were, then there would be some $z\in A^i_\tau$ such that $f^{i+1,\tau}_{[k]}(h^i_\tau(z)) = y_{i+1}$, since $h^i_\tau$ is the identity map. But then since $(h_\sigma)$ is a disjoint embedding, $x_{i+1}$ and $z$ would both be in the image of some element in $A^i_{\sigma\cap \tau}$, and in particular $x_{i+1}$ would be in the image of some element of $A^i_\sigma$, contradicting the choice of $x_{i+1}$. Now for $j>i+1$, define $x_{j+1} = h^j_\sigma(x_j)\in A^{j+1}_\sigma$ and $y_{j+1} = h^j_{[k]}(y_j)\in A^{j+1}_{[k]}$ by induction. Let $x$ and $y$ be the common images of the $x_j$ and $y_j$ in $C_\sigma$ and $C_{[k]}$. Then $y = g^\sigma_{[k]}(x)$. We will show by induction on $j\geq i+1$ that $y_j\notin f^{j,\tau}_{[k]}(A^j_\tau)$, from which it follows that $y\in g^\sigma_{[k]}(C_\sigma)\setminus g^\tau_{[k]}(C_\tau)$. We have already established the base case, when $j = i+1$. For the inductive step, if $y_{j+1} = h^j_{[k]}(y_j) \in f^{j+1,\tau}_{[k]}(A^{j+1}_\tau)$, then since $(h_\sigma)$ is a disjoint embedding, then $y_j\in f^{j,\tau}_{[k]}(A^j_\tau)$, contradicting the inductive hypothesis. \end{proof} It will be convenient to encode an entire $k$-cube in a single structure. Fix $1\leq k<n$, and consider the language \[\mathcal{L}^k_{n}=\mathcal{L}_{n}\cup\{D_{\sigma}\mid \sigma\in \Delta^{k-1}\},\] where the $D_{\sigma}$ are unary relation symbols. Let $\mathcal{D}^k_{n}$ be the class of all $\mathcal{L}^k_{n}$-structures $M$ such that: \begin{enumerate} \item $D^M_{[k]} = M$. \item Each $D^M_{\sigma}$ is an $\mathcal{L}_n$-substructure of $M\mathord{\upharpoonright} \mathcal{L}_n$. \item $D^M_\sigma\subseteq D^M_\tau$ when $\sigma\subseteq \tau$. \item $(D^M_{\sigma})_\sigma$ together with the inclusion maps is an irreducible $k$-cube in $\mathcal{B}^*_{n}$. \end{enumerate} \begin{lemma} \begin{enumerate} \item $\mathrm{Mod}_\omega(\mathcal{B}_{n})$ is an invariant $G_\delta$ subspace of $\mathcal{X}_{\mathcal{L}_n}$. \item A basis for the topology on $\mathrm{Mod}_\omega(\mathcal{B}_{n})$ is given by the collection of sets $[\theta_A(a_0,\dots,a_{k})]$, where $A$ is a finite BKL$_n$-structure with domain $\{a_0,\dots,a_{k}\}\subseteq \omega$. \item $\mathrm{Mod}_\omega(\mathcal{B}^{*}_{n})$ is an invariant \emph{dense} $G_{\delta}$ subspace of $\mathrm{Mod}_\omega(\mathcal{B}_{n})$. \item $\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ is an invariant \emph{dense} $G_{\delta}$ subspace of $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$. \item For $1\leq k<n$, $\mathrm{Mod}_\omega(\mathcal{D}^k_{n})$ is an invariant \emph{nonempty} $G_\delta$ subspace of $\mathcal{X}_{\mathcal{L}^k_{n}}$. \item For $1\leq k<n$, $\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$ is an invariant \emph{nonempty} $G_\delta$ subspace of $\mathcal{X}_{\widehat{\mathcal{L}^k_{n}}}$. \end{enumerate} \end{lemma} \begin{proof} For (1), we give an axiomatization of $\mathcal{B}_n$ in $\mathcal{L}_{\omega_1,\omega}$: \begin{itemize} \item $\forall x_0,\dots,x_{n-1} \bigvee_{j\geq 0} R_j(x_0,\dots,x_{n-1})$; \item $\forall x_0,\dots,x_{n-1} \bigwedge_{0\leq i<j} \lnot (R_i(x_0,\dots,x_{n-1})\land R_j(x_0,\dots, x_{n-1})$; \item $\forall x_0,\ldots,x_{n-1} \bigwedge_{0\leq i< j} (R_i(x_0,\dots,x_{n-1})\rightarrow s_j(x_0,\ldots,x_{n-1})=x_0)$; \item $\forall x_0,\ldots,x_{n} \vee_{0\leq k\leq n} \bigvee_{t(x_1,\dots,x_{k-1},x_{k+1},\dots,x_n)} t(x_1,\dots,x_{k-1},x_{k+1},\dots,x_n) = x_k$,\\ where $t$ ranges over all terms in the variables $x_0,\dots,x_n$, with $x_k$ omitted. \item $\forall x_0,\dots,x_{k}(\bigwedge_{1\leq i<j\leq k} x_i\neq x_j \rightarrow \bigvee_{A = \{a_0,\dots,a_\ell\}} \exists x_{k+1},\dots,x_\ell\, \theta_A(x_0,\dots,x_\ell))$,\\ where $A$ ranges over all isomorphism types of structures in $\mathcal{B}_n^{\mathrm{fin}}$ with a fixed enumeration $a_0,\dots,a_\ell$, where $k\leq \ell$. Note that this is only a countable disjunction, since each such $A$ corresponds to the finitary $\mathcal{L}_n$-formula $\theta_A$. \end{itemize} From the form of the axiomatization it follows that $\mathrm{Mod}_\omega(\mathcal{B}_{n})$ is $G_\delta$ in $\mathcal{X}_{\mathcal{L}_n}$. For (2), let $U$ be a nonempty open subset of $\mathrm{Mod}_\omega(\mathcal{B}_{n})$. We may assume that $U = [\varphi(a)]$, where $\varphi(x)$ is a finite conjunction of atomic and negated atomic $\mathcal{L}_n$-formulas. Let $A\in [\varphi(a)]$. By axiom (B4), $A$ is locally finite. Let $B$ be the substructure of $A$ generated by the tuple $a$, and fix an enumeration of $B$ by a tuple $b$. Then $M\in [\theta_B(b)] \subseteq [\varphi(a)]$. For (3), the fact that $\mathrm{Mod}_\omega(\mathcal{B}^{*}_{n})$ is an invariant $G_{\delta}$ subspace of $\mathrm{Mod}_\omega(\mathcal{B}_{n})$ is clear from the axiomatization by extension axioms. Density follows from (2), since every finite BKL$_n$-structure with domain a subset of $\omega$ is a substructure of a structure in $\mathrm{Mod}_\omega(\mathcal{B}^{*}_{n})$. Then (4) follows immediately. For (5), note that conditions (1)-(3) in the definition of $\mathcal{D}^k_n$ are closed, the condition that the $k$-cube is irreducible is open, and the condition that each $D^M_\sigma$ is in $\mathcal{B}^*_n$ is $G_\delta$, by relativizing the extension axioms for $\mathcal{B}^*_n$ to $D_\sigma$. To see that it is nonempty, let $((A_\sigma),(f^\sigma_\tau))$ be the irreducible $k$-cube in $\mathcal{B}^*_n$ constructed in Lemma~\ref{L: in-out2}. We may assume that $M = A_{[k]}$ has domain $\omega$ and define the relations $D^M_\sigma$ as $f^\sigma_{[k]}(A_\sigma)$. Then (6) follows immediately. \end{proof} If $M\in \mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$, then for all $\sigma\in \Delta^{k-1}$, we view $D^M_\sigma$ as a $\widehat{\mathcal{L}_n}$-substructure of $M\mathord{\upharpoonright} \widehat{\mathcal{L}_n}$. Notice that the domain of $D^M_{\sigma}$ is always an infinite subset of $\omega$, since $M\mathord{\upharpoonright} \widehat{\mathcal{L}_n}$ is in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n+1}})$. Let $e^M_\sigma\colon \omega \to \omega$ enumerate $D^M_\sigma$ in increasing order. Then there is a unique structure $N\in \mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ such that $e^M_\sigma\colon N\to M$ defines an $\widehat{L_n}$-isomorphism between $N$ and $D^M_\sigma$. The assignment $M\mapsto N$ defines a map: \[\phi_{\sigma}\colon\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})\to\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}}).\] \begin{lemma}\label{L: cont open} For all $1\leq k<n$, and for every $\sigma\in \Delta^{k-1}$, the map $\phi_{\sigma}\colon\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})\to\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ is continuous and open. \end{lemma} \begin{proof} For continuity, let $U$ be an open set in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$. We may assume $U$ has the form $[\psi(a)]$, where $\psi$ is a finite conjunction of atomic and negated atomic $\widehat{\mathcal{L}_n}$-formulas and $a = (a_1,\dots,a_n)$ is a tuple from $\omega$ with $a_1<\dots<a_n$. Let $M\in \phi_\sigma^{-1}(U)$. Then $M\models \psi(e^M_\sigma(a))$. Let $\chi$ be the conjunction of all formulas of the form $D_\sigma(b)$ and $\lnot D_\sigma(b)$ which are satisfied in $M$, for $b\leq e^M_\sigma(a_n)$, and let $V = [\psi(e^M_\sigma(a))\land \chi]$. Then $V$ is an open set in $\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$ and $M\in V\subseteq \phi_\sigma^{-1}(U)$. Indeed, if $M'\in V$, then $e^{M'}_\sigma$ agrees with $e^{M}_\sigma$ on the initial segment of $\omega$ up to $a_n$, so $M'\models \psi(e^{M'}_\sigma(a))$, and $\phi_\sigma(M')\in U$. For openness, let $U$ be an open set in $\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$. We may assume $U$ has the form $[\psi(a)]$, where $\psi$ is a finite conjunction of atomic and negated atomic $\widehat{\mathcal{L}^k_n}$-formulas and $a = (a_1,\dots,a_n)$ is a tuple from $\omega$ with $a_1<\dots<a_n$. Note also that $\psi$ mentions only finitely many of the labeling predicates, $P_{i_1},\dots,P_{i_\ell}$. Let $N\in \phi_\sigma(U)$, and let $M\in U$ with $\phi_\sigma(M) = N$. Let $b\in \omega$ be the largest element such that $e_\sigma(b)$ is in the tuple $a$ in $M$, and let $B = \langle 0,\dots,b\rangle$ be the substructure of $N$ generated by the initial segment of $\omega$ up to $b$. Enumerate $B$ as $\{b_1,\dots,b_m\}$. Recall that $\theta_B(b_1,\dots,b_m)$ determines the $\mathcal{L}_n$-isomorphism type of $B$. Let $\chi$ be the conjunction of all formulas of the form $P_{i_j}(b_{j'})$ and $\lnot P_{i_j}(b_{j'})$ which are satisfied in $N$, for $1\leq j\leq \ell$ and $1\leq j' \leq m$. And let $V = [\theta_B(b_1,\dots,b_m)\land \chi]$. Then $V$ is an open set in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ and $N\in V\subseteq \phi_\sigma(U)$. To see this, given $N'\in V$, we will find $M'\in U$ such that $\phi_\sigma(M') = N'$. We obtain $M'$ by modifying our original structure $M$ in stages. \begin{enumerate} \item Since $N'\models \theta_B(b_1,\dots,b_m)$, $M\models \theta_B(e^M_\sigma(b_1),\dots,e^M_\sigma(b_m))$. So $e^M_\sigma$ restricts to an $\mathcal{L}_n$-embedding $B\to D^M_\sigma$. Since $D^M_\sigma\mathord{\upharpoonright} \mathcal{L}_n$ and $N'\mathord{\upharpoonright}\mathcal{L}_n$ are both isomorphic to the Fra\"iss\'e\xspace limit of $\mathcal{B}^{\text{fin}}_n$, this embedding extends to an $\mathcal{L}_n$-isomorphism $f\colon N'\to D^M_\sigma$ which agrees with $e^M_\sigma$ on $B$. We will use this $f$ to re-enumerate $M$ so that $\phi_\sigma$ produces a structure which agrees with the target structure $N'$ in the reduct to the language $\mathcal{L}_n$. \item We view the composition $e^M_\sigma \circ f^{-1}\colon D^M_\sigma\to D^M_\sigma$ as a permutation of $D^M_\sigma\subseteq \omega$. Extending this permutation by the identity outside $D^M_\sigma$ defines a permutation $g\in S_\infty$. Let $M_0 = g(M)$ (by the logic action of $S_\infty$ on $\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$. Note that $g$ is the identity on the tuple $a$. Indeed, if $a_i\notin D^M_\sigma$, then $g(a_i) = a_i$ by definition. And if $a_i\in D^M_\sigma$, then there is some $b'$ such that $e^M_\sigma(b') = a_i$. Since $a_i<a_n$, $b'<b$, so $b'\in B$. But since $f$ agrees with $e^M_\sigma$ on $B$, $g(a_i) = e^M_\sigma(f^{-1}(a_i)) = a_i$. It follows that $M_0\in [\psi(a)] = U$. Further, since $g$ fixes $D^M_\sigma$ set-wise, we have $D^{M_0}_\sigma = D^M_\sigma$ as subsets of $\omega$, and $e^{M_0}_\sigma = e^M_\sigma$. But $g^{-1}\circ e^{M_0}_\sigma = f$ is an $\mathcal{L}_n$-isomorphism $N'\to D^M_\sigma$, so $e^{M_0}_\sigma$ is an $\mathcal{L}_n$-isomorphism $N'\to D^{M_0}_\sigma$. It follows that $\phi_\sigma(M_0)\mathord{\upharpoonright} \mathcal{L}_n = N'\mathord{\upharpoonright}\mathcal{L}_n$. \item It remains to improve this to equality of $\widehat{\mathcal{L}_n}$-structures by relabeling $M_0$ by the predicates $P_i$. Let $M_1 = M_0\mathord{\upharpoonright}\mathcal{L}^k_n$, dropping all the labels. We now define the $\widehat{\mathcal{L}_n}$-structure $M'$ expanding $M_0$. Define the labels on the elements of $D^{M_0}_\sigma$ so that $e^{M_0}_\sigma$ is an $\widehat{\mathcal{L}_n}$-isomorphism. Note that no two elements get exactly the same labels, since this is true in $N'$. Now assign labels arbitrarily to the remaining elements of $M_1$, only making sure that no two elements get exactly the same labels and that for each element $a_i$ of the tuple $a$, the new label of $a_i$ agrees with its old label on the predicates $P_{i_1},\dots,P_{i_\ell}$. In fact, this is already the case when $a_i$ is in $D^{M_0}_\sigma$, since $N'in V$ and the information about the predicates $P_{i_1},\dots,P_{i_\ell}$ is included in $\chi$. We have now arranged that $\phi_\sigma(M') = N'$. \item It remains to check that $M'\in U = [\psi(a)]$. This follows from the fact that $M_0\in [\psi(a)]$, and $M'$ agrees with $M_0$ on all atomic formulas in $\mathcal{L}^k_n$, as well as the values of all the predicates mentioned in $\psi$ on the tuple $a$. \qedhere \end{enumerate} \end{proof} We now have all the ingredients necessary to prove Theorem~\ref{thm:bklmain}. \begin{proof}[Proof of Theorem~\ref{thm:bklmain}] First we observe that having a $k$-cube in $\widehat{\mathcal{B}_{n}}$ with structures lying in some invariant subset $C$ of $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is the same as having an embedding of the combinatorial $k$-cube in the Becker graphs $\mathcal{B}(C/S_{\infty})$. This follows from Proposition \ref{Proposition Emb} and the fact that we are working with labeled structures; see the discussion after the definition of labeled $\mathcal{K}$-structures. By Lemma~\ref{lem:bklprops}(2) it follows that the dimension of $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is at most $(n-1)$. When $n=1$, it is immediate that $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is locally generically $(n-1)$ dimensional (this just means that every non-meager invariant set with the Baire property is nonempty). So in the remainder of the proof, we assume $n>1$. First, we show that $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is generically $(n-1)$-dimensional. Let $C\subseteq \mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ be a comeager set. Since $\mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ is dense $G_\delta$ in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$, the restriction $C^* = C\cap \mathrm{Mod}_\omega(\widehat{\mathcal{B}^{*}_{n}})$ is comeager in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$. For each $\sigma\in \Delta^{k-1}$, let $C^*_\sigma = \phi_\sigma^{-1}(C^*)\subseteq \mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$. Continuous open maps are category preserving, so each $C^*_\sigma$ is comeager in $\mathrm{Mod}_\omega(\widehat{\mathcal{D}^k_{n}})$. Since this space is nonempty and Polish, $\bigcap_{\sigma\in \Delta^{k-1}} C^*_\sigma$ is nonempty. Let $M$ be a structure in the intersection. We define a $k$-cube $\mathcal{A}$ by setting $A_\sigma = \phi_\sigma(M)$. In particular, $A_{[k]} = M\mathord{\upharpoonright} \widehat{\mathcal{L}_n}$. To define the connecting maps, let $f^\sigma_{[k]} = e^M_\sigma \colon A_\sigma\to A_{[k]}$. Now for any $\sigma\subseteq \tau\subseteq [k]$, the map $f^\sigma_{[k]}$ factors through the map $f^\tau_{[k]}$ by a unique map $f^\sigma_\tau$. By construction, $\mathcal{A}$ is irreducible and $A_\sigma\in C^*\subseteq C$. Finally, we claim that $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ is generically ergodic when $n>1$, so that the fact that it is generically $(n-1)$-dimensional immediately implies that it is locally generically $(n-1)$-dimensional. By~\cite[Proposition 6.1.9]{Gao}, it suffices to show that there is a structure in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ with a dense orbit. Consider the $\widehat{\mathcal{L}_n}$-theory which axiomatizes labeled BKL$_n$-structures (axioms (A1), (B1), (B2), (B3), and (B4)) and contains the additional axiom: \[\bigwedge_{A=\{a_0,\ldots,a_k\}} \bigwedge_{(T_0,F_0),\ldots,(T_k, F_k)} \exists x_0,\ldots,x_k \bigg(\theta_A(x_0,\ldots,x_k) \wedge \bigwedge_{i\leq k} \big(\bigwedge_{j\in T_i} P_j(x_i) \land \bigwedge_{j\in F_i} \lnot P_j(x_i)\big)\bigg),\] where $A = \{a_0,\dots,a_k\}$ varies over all isomorphism types of structures in $\mathcal{B}^{\mathrm{fin}}_{n}$, and for each $i$, $(T_i,F_i)$ is a pair of disjoint finite subsets of $\omega$. It is easy to see that there are structures in $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$ which satisfy this theory and that every such structure has a dense orbit under the $S_{\infty}$ action. \end{proof} \begin{remark}\label{Remark fin} Keeping track of the properties of $\mathcal{B}_n$ used in the proof of Theorem~\ref{thm:bklmain}, we see that for any class of structures $\mathcal{K}$ satisfying the following hypotheses, the $S_{\infty}$-space $\mathrm{Mod}_\omega(\widehat{\mathcal{K}})$ is locally generically $(\geq (n-1))$-dimensional: \begin{enumerate} \item $\mathcal{K}$ is closed under substructure. \item $\mathrm{Mod}_\omega(\mathcal{K})$ is a $G_\delta$-subspace of $\mathcal{X}_{\mathcal{L}}$. \item $\mathcal{K}^{\mathrm{fin}}$ is a Fra\"iss\'e\xspace class, whose Fra\"iss\'e\xspace limit is in $\mathcal{K}$. \item $\mathcal{K}^{\mathrm{fin}}$ has disjoint $k$-amalgamation for all $1\leq k\leq n$. \item $\mathcal{K}^{\mathrm{fin}}$ is \emph{separable}: this is the condition in Lemma~\ref{lem:bklspaces}(1). \end{enumerate} It follows, for example, that if $\mathcal{K}$ is the class $\mathcal{S}$ of all sets or $\mathcal{K}$ is the class of all graphs, then $\mathrm{Mod}_\omega(\widehat{\mathcal{K}})$ is locally generically $\infty$-dimensional. In both cases Theorem \ref{T: main 2} implies that $\simeq_{\mathrm{iso}}$ on $\mathrm{Mod}_\omega(\widehat{\mathcal{K}})$ is incomparable to $\simeq_{\mathrm{iso}}$ on $\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}})$, for all $n>0$. In particular, $=^{+}$ is incomparable to $(\mathrm{Mod}_\omega(\widehat{\mathcal{B}_{n}}),\simeq_{\mathrm{iso}})$, for all $n>0$; see Example \ref{Example:ctbl sets of reals}. \end{remark} \begin{bibdiv} \begin{biblist} \bib{Becker}{article} { title = {Polish group actions: Dichotomies and generalized elementary embeddings} author = {Becker, H.} date = {1998} journal = {Journal of the American Mathematical Society} volume = {11} pages = {397--449} } \bib{BKL}{article} { title = {Disjoint amalgamation in locally finite AEC}, author={Baldwin, John T.} author={Koerwien, Martin} author={Laskowski, Michael C.} volume={82}, number={1}, journal={J. Symbolic Logic}, publisher={Cambridge University Press}, year={2017}, pages={98--119} } \bib{Friedman}{article} { title = {A Borel reducibility theory for classes of countable structures} author = {Friedman, H.} author = {Stanley, L.} date = {1989} journal = {J. Symbolic Logic} volume = {54} number={3} pages = {894--914} } \bib{GabF}{article} { title = {Co\^ut des relations d'\'e quivalence et des groupes} author = {Gaboriau, D.} date = {2000} journal = {Invent. Math.} volume = {139} number={1} pages = {41--98} } \bib{Gao paper}{article} { title = {On Automorphism Groups of Countable Structures} author = {Gao, S.} date = {1998} journal = {The Journal of Symbolic Logic} volume = {63} pages = {891--896} } \bib{Gab}{book} { title = {On orbit equivalence of measure preserving actions} author = {Gaboriau, D.} series = {Rigidity in Dynamics and Geometry (Cambridge, 2000)} year = {2002} publisher = {Springer, Berlin} pages = {167--186} } \bib{Gao}{book} { title = {Invariant descriptive set theory} author = {Gao, S.} year = {2008} publisher = {Taylor \& Francis} } \bib{Gao2}{article} { title = {Some dichotomy theorems for isomorphism relations of countable models} author = {Gao, S.} date = {2001} journal = {J. Symbolic Logic} volume = {66} number={2} pages = {902-922} } \bib{Gao3}{article} { title = {Unitary group actions and Hilbertian Polish metric spaces} author = {Gao, S.} date = {2005} journal = {Logic and its applications, Contemp. Math.} volume = {380} pages = {53–72} } \bib{Hjorth}{book} { title = {Classification and orbit equivalence relations} author = {Hjorth, G.} series = {Mathematical Surveys and Monographs} volume = {75} year = {2000} publisher = {American Mathematical Society} } \bib{Kechris}{book} { title = {Classical descriptive set theory} author = {Kechris, S. A.} series = { Graduate Texts in Mathematics} year = {2011} publisher = {Springer, New York} } \bib{KechrisMiller}{book} { title = { Topics in orbit equivalence} author = {Kechris, S. A.} author = {Miller, B. D.} series = {Lect. Notes in Math.} year = {2004} volume ={1852} publisher = {Springer, Berlin} } \bib{GOP}{article} { title = {Games orbits play and obstructions to Borel reducibility} author = {Lupini, M.} author = {Panagiotopoulos, A.} date = {} journal = {Groups, Geometry, and Dynamics} } \bib{Lupini}{article} { title = {Polish groupoids and functorial complexity} author = {Lupini, M.} note = {With an appendix by Anush Tserunyan} date = {2017} number = {9} journal = {Trans. Amer. Math. Soc.} volume = {369} pages = {6683–6723} } \bib{LS}{article} { title = {On the existence of atomic models} author={Laskowski, M. C.} author={Shelah, S.} volume={58}, number={4}, journal={J. Symbolic Logic}, year={1993}, pages={1189--1194} } \bib{Category}{article} { author = {Melleray, J.} author = {Tsankov, T.} title = {Generic representations of abelian groups and extreme amenability} journal={Israel Journal of Mathematics} date={2013} volume={198} number={1} pages={129--167} } \end{biblist} \end{bibdiv} \end{document}

7e89678b63ac49d8e54cfbd522e10c430473d510

\section{Introduction} Language grounding, i.e., understanding how words and expressions are anchored in data, is one of the initial tasks that are essential for the conception of a data-to-text (D2T) system \cite{bib_connectinglanguage,nlg_datatotext}. This can be achieved through different means, such as using heuristics or machine learning algorithms on an available parallel corpora of text and data \cite{novikova2017e2e} to obtain a mapping between the expressions of interest and the underlying data \cite{bib_mousam}, getting experts to provide these mappings, or running surveys on writers or readers that provide enough data for the application of mapping algorithms \cite{bib_ramos2017empirical}. Performing language grounding allows ensuring that generated texts include words whose meaning is aligned with what writers understand or what readers would expect \cite{bib_connectinglanguage}, given the variation that is known to exist among writers and readers \cite{reiter2002should}. Moreover, when contradictory data appears in corpora or any other resource that is used to create the data-to-words mapping, creating models that remove inconsistencies can also be a challenging part of language grounding which can influence the development of a successful system \cite{bib_mousam}. This paper presents a resource for language grounding of geographical descriptors. The original purpose of this data collection is the creation of models of geographical descriptors whose meaning is modeled as graded or fuzzy \cite{fisher2000sorites,fisher2006approaches}, to be used for research on generation of geographical referring expressions, e.g., \cite{bib_ross_3,nlg_roadsafe,bib_rodrigo,bib_fuzzygre,bib_ramos2017empirical}. However, we believe it can be useful for other related research purposes as well. \section{The resource and its interest} The resource is composed of data from two different surveys. In both surveys subjects were asked to draw on a map (displayed under a Mercator projection) a polygon representing a given geographical descriptor, in the context of the geography of Galicia in Northwestern Spain (see Fig. \ref{fig:survey}). However, the surveys were run with different purposes, and the subject groups that participated in each survey and the list of descriptors provided were accordingly different. \begin{figure}[t] \centering \includegraphics[width=\columnwidth]{img/survey_shot} \caption{Snapshot of the version of the survey answered by the meteorologists (translated from Spanish).} \label{fig:survey} \end{figure} The first survey was run in order to obtain a high number of responses to be used as an evaluation testbed for modeling algorithms. It was answered by 15/16 year old students in a high school in Pontevedra (located in Western Galicia). 99 students provided answers for a list of 7 descriptors (including cardinal points, coast, inland, and a proper name). Figure \ref{fig:north_teens} shows a representation of the answers given by the students for ``Northern Galicia'' and a contour map that illustrates the percentages of overlapping answers. \begin{figure}[b] \centering \includegraphics[width=\columnwidth]{img/norte_teens} \caption{Representation of polygon drawings by students and associated contour plot showing the percentage of overlapping answers for ``Northern Galicia''.} \label{fig:north_teens} \end{figure} The second survey was addressed to meteorologists in the Galician Weather Agency \cite{meteogalicia}. Its purpose was to gather data to create fuzzy models that will be used in a future NLG system in the weather domain. Eight meteorologists completed the survey, which included a list of 24 descriptors. For instance, Figure \ref{fig:east_experts} shows a representation of the answers given by the meteorologists for ``Eastern Galicia'' and a contour map that illustrates the percentage of overlapping answers. Table \ref{table:desc_list} includes the complete list of descriptors for both groups of subjects. 20 out of the 24 descriptors are commonly used in the writing of weather forecasts by experts and include cardinal directions, proper names, and other kinds of references such as mountainous areas, parts of provinces, etc. The remaining four were added to study intersecting combinations of cardinal directions (e.g. exploring ways of combining ``north'' and ``west'' for obtaining a model that is similar to ``northwest''). \begin{figure}[t] \centering \includegraphics[width=\columnwidth]{img/east_experts} \caption{Representation of polygon drawings by experts and associated contour plot showing the percentage of overlapping answers for ``Eastern Galicia''.} \label{fig:east_experts} \end{figure} \begin{table}[b] \centering \includegraphics[width=\columnwidth]{img/descriptor_list} \caption{List of geographical descriptors in the resource.} \label{table:desc_list} \end{table} The data for the descriptors from the surveys is focused on a very specific geographical context. However, the conjunction of both data sets provides a very interesting resource for performing a variety of more general language grounding-oriented and natural language generation research tasks, such as: \begin{itemize} \item Testing algorithms that create geographical models. These models would aggregate the answers from different subjects for each descriptor. The differences among the subjects can be interpreted from a probabilistic or fuzzy perspective that allows a richer characterization of the resulting models. For instance, in Fig. \ref{fig:north_teens} the contour plots could be taken as the basis or support for the semantics of the expression ``Northern Galicia'', with a core region that is accepted by the majority, and a gradual decay as one moves to the outer periphery of the regions outlined. \item Analyzing differences between the expert and non-expert groups for the descriptors they have in common (as Table \ref{table:desc_list} shows, both groups share 6 descriptors). \item Studying how to combine models representing the semantics of different cardinal directions, such as ``south'' and ``east'' to obtain a representation of ``southeast''. \item Developing geographical referring expression generation algorithms based on the empirically created models. \end{itemize} \section{Qualitative analysis of the data sets} The two data sets were gathered for different purposes and only coincide in a few descriptors, so providing a direct comparison is not feasible. However, we can discuss general qualitative insights and a more detailed analysis of the descriptors that both surveys share in common. At a general level, we had hypothesized that experts would be much more consistent than students, given their professional training and the reduced number of meteorologists participating in the survey. Comparing the visualizations of both data sets we have observed that this is clearly the case; the polygons drawn by the experts are more concentrated and therefore there is a higher agreement among them. On top of these differences, some students provided unexpected drawings in terms of shape, size, or location of the polygon for several descriptors. \begin{figure}[b] \centering \includegraphics[width=\columnwidth]{img/north_south} \caption{Areas overlapping ``north'' and ``south'' for both subject groups (in blue).} \label{fig:north_south} \end{figure} If we focus on single descriptors, one interesting outcome is that some of the answers for ``Northern Galicia'' and ``Southern Galicia'' overlap for both subject groups. Thus, although `north' and `south' are natural antonyms, if we take into account the opinion of each group as a whole, there exists a small area where points can be considered as belonging to both descriptors at the same time (see Fig. \ref{fig:north_south}). In the case of ``west'' and ``east'', the drawings made by the experts were almost divergent and showed no overlapping between those two descriptors. Regarding ``Inland Galicia'', the unions of the answers for each group occupy approximately the same area with a similar shape, but there is a very high overlapping among the answers of the meteorologists. A similar situation is found for the remaining descriptor ``R\'ias Baixas'', where both groups encompass a similar area. In this case, the students' answers cover a more extensive region and the experts coincide within a more restricted area. \subsection{A further analysis: apparent issues} \label{sec:issues} As in any survey that involves a task-based collection of data, some of the answers provided by the subjects for the described data sets can be considered erroneous or misleading due to several reasons. Here we describe for each subject group some of the most relevant issues that any user of this resource should take into account. In the case of the students, we have identified minor drawing errors appearing in most of the descriptors, which in general shouldn't have a negative impact in the long term thanks to the high number of participants in the original survey. For some descriptors, however, there exist polygons drawn by subjects that clearly deviate from what could be considered a proper answer. The clearest example of this problem involves the `west' and `east' descriptors, which were confused by some of the students who drew them inversely (see Fig. \ref{fig:weast_east_problem}, around 10-15\% of the answers). \begin{figure}[b] \centering \includegraphics[width=\columnwidth]{img/west_east_studs} \caption{Contour maps of student answers for ``Western Galicia'' and ``Eastern Galicia''.} \label{fig:weast_east_problem} \end{figure} In our case, given their background, some of the students may have actually confused the meaning of + ``west'' and ``east''. However, the most plausible explanation is that, unlike in English and other languages, in Spanish both descriptors are phonetically similar (``este'' and ``oeste'') and can be easily mistaken for one another if read without attention. As for the expert group, a similar case is found for ``Northeastern Galicia'' (see Fig. \ref{fig:northeast_problem}), where some of the given answers (3/8) clearly correspond to ``Northwestern Galicia''. However, unlike the issue related to ``west'' and ``east'' found for the student group, this problem is not found reciprocally for the ``northwestern'' answers. \begin{figure}[t] \centering \includegraphics[width=\columnwidth]{img/northeast_issue} \caption{Representation of polygon drawings by experts and associated contour plots showing the percentage of overlapping answers for ``Northeastern Galicia''.} \label{fig:northeast_problem} \end{figure} \section{Resource materials} The resource is available at \cite{bib_resource} under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Both data sets are provided as SQLite databases which share the same table structure, and also in a compact JSON format. Polygon data is encoded in GeoJSON format \cite{geojson}. The data sets are well-documented in the repository's README, and several Python scripts are provided for data loading, using Shapely \cite{shapely}; and for visualization purposes, using Cartopy \cite{Cartopy}. \section{Concluding remarks} The data sets presented provide a means to perform different research tasks that can be useful from a natural language generation point of view. Among them, we can highlight the creation of models of geographical descriptors, comparing models between both subject groups, studying combinations of models of cardinal directions, and researching on geographical referring expression generation. Furthermore, insights about the semantics of geographical concepts could be inferred under a more thorough analysis. One of the inconveniences that our data sets present is the appearance of the issues described in Sec. \ref{sec:issues}. It could be necessary to filter some of the answers according to different criteria (e.g., deviation of the centroid location, deviation of size, etc.). For more applied cases, manually filtering can also be an option, but this would require a certain knowledge of the geography of Galicia. In any case, the squared-like shape of this region may allow researchers to become rapidly familiar with many of the descriptors listed in Table \ref{table:desc_list}. As future work, we believe it would be invaluable to perform similar data gathering tasks for other regions from different parts of the world. These should provide a variety of different shapes (both regular and irregular), so that it can be feasible to generalize (e.g., through data-driven approaches) the semantics of some of the more common descriptors, such as cardinal points, coastal areas, etc. The proposal of a shared task could help achieve this objective. \section*{Acknowledgments} This research was supported by the Spanish Ministry of Economy and Competitiveness (grants TIN2014-56633-C3-1-R and TIN2017-84796-C2-1-R) and the Galician Ministry of Education (grants GRC2014/030 and "accreditation 2016-2019, ED431G/08"). All grants were co-funded by the European Regional Development Fund (ERDF/FEDER program). A. Ramos-Soto is funded by the ``Conseller\'{i}a de Cultura, Educaci\'{o}n e Ordenaci\'{o}n Universitaria'' (under the Postdoctoral Fellowship accreditation ED481B 2017/030). J.M. Alonso is supported by RYC-2016-19802 (Ram\'on y Cajal contract). The authors would also like to thank Juan Taboada for providing the list of most frequently used geographical expressions by MeteoGalicia, and Jos\'e Manuel Ramos for organizing the survey at the high school IES Xunqueira I in Pontevedra, Spain.

da419832f8ffa016dcc2a04a10bf904016004beb

\section{Conclusion} \label{sec:conc} We have described the Uppsala submission to the CoNLL 2018 shared task, consisting of a segmenter that jointly extracts words and sentences from a raw text, a tagger that provides UPOS tags and morphological features, and a parser that builds a dependency tree given the words and tags of each sentence. For the parser we applied multi-treebank models both monolingually and multilingually, resulting in only 34 models for 82 treebanks as well as significant improvements in parsing accuracy especially for low-resource languages. We ranked 7th for the official LAS and MLAS scores, and first for the unofficial scores on word segmentation, UPOS tagging and morphological features. \section*{Acknowledgments} We are grateful to the shared task organizers and to Dan Zeman and Martin Potthast in particular, and we acknowledge the computational resources provided by CSC in Helsinki and Sigma2 in Oslo through NeIC-NLPL (www.nlpl.eu). Aaron Smith was supported by the Swedish Research Council. \section{Introduction} \label{sec:intro} The CoNLL 2018 Shared Task on Multilingual Parsing from Raw Text to Universal Dependencies \citep{conll2018} requires participants to build systems that take as input raw text, without any linguistic annotation, and output full labelled dependency trees for 82 test treebanks covering 46 different languages. Besides the labeled attachment score (LAS) used to evaluate systems in the 2017 edition of the Shared Task \citep{udst:overview}, this year's task introduces two new metrics: morphology-aware labeled attachment score (MLAS) and bi-lexical dependency score (BLEX). The Uppsala system focuses exclusively on LAS and MLAS, and consists of a three-step pipeline. The first step is a model for joint sentence and word segmentation which uses the BiRNN-CRF framework of \citet{shao17,shao18} to predict sentence and word boundaries in the raw input and simultaneously marks multiword tokens that need non-segmental analysis. The second component is a part-of-speech (POS) tagger based on \citet{bohnet18acl}, which employs a sentence-based character model and also predicts morphological features. The final stage is a greedy transition-based dependency parser that takes segmented words and their predicted POS tags as input and produces full dependency trees. While the segmenter and tagger models are trained on a single treebank, the parser uses multi-treebank learning to boost performance and reduce the number of models. After evaluation on the official test sets \citep{ud22data}, which was run on the TIRA server \citep{tira}, the Uppsala system ranked 7th of 27 systems with respect to LAS, with a macro-average F1 of 72.37, and 7th of 27 systems with respect to MLAS, with a macro-average F1 of 59.20. It also reached the highest average score for word segmentation (98.18), universal POS (UPOS) tagging (90.91), and morphological features (87.59). \paragraph{Corrigendum:} After the test phase was over, we discovered that we had used a non-permitted resource when developing the UPOS tagger for Thai PUD (see Section \ref{sec:tagger}). Setting our LAS, MLAS and UPOS scores to 0.00 for Thai PUD gives the corrected scores: LAS 72.31, MLAS 59.17, UPOS 90.50. This does not affect the ranking for any of the three scores, as confirmed by the shared task organizers. \section{Resources} \label{sec:resources} All three components of our system were trained principally on the training sets of Universal Dependencies v2.2 released to coincide with the shared task \citep{ud22data}. The tagger and parser also make use of the pre-trained word embeddings provided by the organisers, as well as Facebook word embeddings \cite{bojanowski2016enriching}, and both word and character embeddings trained on Wikipedia text\footnote{\url{https://dumps.wikimedia.org/backup-index-bydb.html}} with word2vec \citep{mikolov2013distributed}. For languages with no training data, we also used external resources in the form of Wikipedia text, parallel data from OPUS \citep{tiedemann2012OPUS}, the Moses statistical machine translation system \citep{Moses07}, and the Apertium morphological transducer for Breton.\footnote{\url{https://github.com/apertium/apertium-bre}} \section{Multi-Treebank Models} \label{sec:multilingual} One of our main goals was to leverage information across treebanks to improve performance and reduce the number of parsing models. We use two different types of models: \begin{enumerate}[topsep=2pt,itemsep=1pt] \item Single models, where we train one model per treebank (17 models applied to 18 treebanks, including special models for Breton KEB and Thai PUD). \item Multi-treebank models \begin{itemize}[topsep=0pt,itemsep=1pt] \item Monolingual models, based on multiple treebanks for one language (4 models, trained on 10 treebanks, applied to 11 treebanks). \item Multilingual models, based on treebanks from several (mostly) closely related languages (12 models, trained on 48 treebanks, applied to 52 treebanks; plus a special model for Naija NSC). \end{itemize} \end{enumerate} When a multi-treebank model is applied to a test set from a treebank with training data, we naturally use the treebank embedding of that treebank also for the test sentences. However, when parsing a test set with no corresponding training data, we have to use one of the other treebank embeddings. In the following, we refer to the treebank selected for this purpose as the \emph{proxy} treebank (or simply \emph{proxy}). In order to keep the training times and language balance in each model reasonable, we cap the number of sentences used from each treebank to 15,000, with a new random sample selected at each epoch. This only affects a small number of treebanks, since most training sets are smaller than 15,000 sentences. For all our multi-treebank models, we apply the treebank embeddings described in Section \ref{sec:parsing}. Where two or more treebanks in a multilingual model come from the same language, we use separate treebank embeddings for each of them. We have previously shown that multi-treebank models can boost LAS in many cases, especially for small treebanks, when applied monolingually \cite{stymne+18acl}, and applied it to low-resource languages \cite{uu-conll17}. In this paper, we add POS tags and pre-trained embeddings to that framework, and extend it to also cover multilingual parsing for languages with varying amounts of training data. Treebanks sharing a single model are grouped together in Table \ref{tab:results}. To decide which languages to combine in our multilingual models, we use two sources: knowledge about language families and language relatedness, and clusterings of treebank embeddings from training our parser with all available languages. We created clusterings by training single parser models with treebank embeddings for all treebanks with training data, capping the maximum number of sentences per treebank to 800. We then used Ward's method to perform a hierarchical cluster analysis. We found that the most stable clusters were for closely related languages. There was also a tendency for treebanks containing old languages (i.e., Ancient Greek, Gothic, Latin and Old Church Slavonic) to cluster together. One reason for these languages parsing well together could be that several of the 7 treebanks come from the same annotation projects, four from PROIEL, and two from Perseus, containing consistently annotated and at least partially parallel data, e.g., from the Bible. For the multi-treebank models, we performed preliminary experiments on development data investigating the effect of different groupings of languages. The main tendency we found was that it was better to use smaller groups of closely related languages rather than larger groups of slightly less related languages. For example, using multilingual models only for Galician-Portuguese and Spanish-Catalan was better than combining all Romance languages in a larger model, and combining Dutch-German-Afrikaans was better than also including English. A case where we use less related languages is for languages with very little training data (31 sentences or less), believing that it may be beneficial in this special case. We implemented this for Buryat, Uyghur and Kazakh, which are trained with Turkish, and Kurmanji, which is trained with Persian, even though these languages are not so closely related. For Armenian, which has only 50 training sentences, we could not find a close enough language, and instead train a single model on the available data. For the four languages that are not in a multilingual cluster but have more than one available treebank, we use monolingual multi-treebank models (English, French, Italian and Korean). For the nine treebanks that have no training data we use different strategies: \begin{itemize}[topsep=2pt,itemsep=1pt] \item For Japanese Modern, we apply the mono-treebank Japanese GSD model. \item For the four PUD treebanks, we apply the multi-treebank models trained using the other treebanks from that language, with the largest available treebank as proxy (except for Finnish, where we prefer Finnish TDT over FTB; cf.\ Section~\ref{sec:segmentation} and \citet{stymne+18acl}). \item For Faroese, we apply the model for the Scandinavian languages, which are closely related, with Norwegian Nynorsk as proxy (cf.\ Section~\ref{sec:tagger}). In addition, we map the Faroese characters \{\'{I}\'{y}\'{u}ð\}, which do not occur in the other Scandinavian languages, to \{Iyud\}. \item For Naija, an English-based creole, whose treebank according to the README file contains spoken language data, we train a special multilingual model on English EWT and the three small spoken treebanks for French, Norwegian, and Slovenian, and usd English EWT as proxy.\footnote{ We had found this combination to be useful in preliminary experiments where we tried to parse French Spoken without any French training data. } \item For Thai and Breton, we create multilingual models trained with word and POS embeddings only (i.e., no character models or treebank embeddings) on Chinese and Irish, respectively. These models make use of multilingual word embeddings provided with Facebook's MUSE multilingual embeddings,\footnote{\url{https://github.com/facebookresearch/MUSE}} as described in more detail below. \end{itemize} For all multi-treebank models, we choose the model from the epoch that has the best mean LAS score among the treebanks that have available development data. This means that treebanks without development data rely on a model that is good for other languages in the group. In the cases of the mono-treebank Armenian and Irish models, where there is no development data, we choose the model from the final training epoch. This also applies to the Breton model trained on Irish data. \paragraph{Thai--Chinese} For the Thai model trained on Chinese, we were able to map Facebook's monolingual embeddings for each language to English using MUSE, thus creating multilingual Thai-Chinese embeddings. We then trained a monolingual parser model using the mapped Chinese embeddings to initialize all word embeddings, and ensuring that these were not updated during training (unlike in the standard parser setup described in Section~\ref{sec:parsing}). At test time, we look up all OOV word types, which are the great majority, in the mapped Thai embeddings first, otherwise assign them to a learned OOV vector. Note that in this case, we had to increase the word embedding dimension in our parser to 300 to accomodate the larger Facebook embeddings. \paragraph{Breton--Irish} For Breton and Irish, the Facebook software does not come with the necessary resources to map these languages into English. Here we instead created a small dictionary by using all available parallel data from OPUS (Ubuntu, KDE and Gnome, a total of 350K text snippets), and training a statistical machine translation model using Moses \cite{Moses07}. From the lexical word-to-word correspondences created, we kept all cases where the translation probabilities in both directions were at least 0.4 and the words were not identical (in order to exclude a lot of English noise in the data), resulting in a word list of 6027 words. We then trained monolingual embeddings for Breton using word2vec on Wikipedia data, and mapped them directly to Irish using MUSE. A parser model was then trained, similarly to the Thai-Chinese case, using Irish embeddings as initialization, turning off updates to the word embeddings, and applying the mapped Breton embeddings at test time. \section{Dependency Parsing} \label{sec:parsing} We use a greedy transition-based parser \citep{nivre2008algorithms} based on the framework of \citet{kiperwasser16} where BiLSTMs \citep{hochreiter1997long,graves2008bilstms} learn representations of tokens in context, and are trained together with a multi-layer perceptron that predicts transitions and arc labels based on a few Bi\-LSTM vectors. allow the construction of non-projective dependency trees \cite{nivre09acl}. We also introduce a static-dynamic oracle to allow the parser to learn from non-optimal configurations at training time in order to recover better from mistakes at test time \citep{delhoneux17arc}. In our parser, the vector representation $x_i$ of a word type $w_i$ before it is passed to the BiLSTM feature extractors is given by: \vspace{-1.5mm} \begin{equation*} x_i = e(w_i) \circ e(p_i) \circ \text{BiLSTM}(ch_{1:m}) . \vspace{-1.5mm} \end{equation*} Here, $e(w_i)$ represents the word embedding and $e(p_i)$ the POS tag embedding \citep{chen14}; these are concatenated to a character-based vector, obtained by running a BiLSTM over the characters $ch_{1:m}$ of $w_i$. With the aim of training multi-treebank models, we additionally created a variant of the parser which adds a treebank embedding $e(tb_i)$ to input vectors in a spirit similar to the language embeddings of \citet{ammar16} and \citet{uu-conll17}: \vspace{-1.5mm} \begin{equation*} x_i = e(w_i) \circ e(p_i) \circ \text{BiLSTM}(ch_{1:m}) \circ e(tb_i). \vspace{-1.5mm} \end{equation*} an effective way to combine multiple monolingual heterogeneous treebanks \citep{stymne+18acl} and applied them to low-resource languages \citep{uu-conll17}. In this shared task, the treebank embedding model was used both monolingually, to combine several treebanks for a single language, and multilingually, mainly for closely related languages, both for languages with no or small treebanks, and for languages with medium and large treebanks, as described in Section~\ref{sec:multilingual}. During training, a word embedding for each word type in the training data is initialized using the pre-trained embeddings provided by the organizers where available. For the remaining languages, we use different strategies: \begin{itemize}[topsep=2pt,itemsep=1pt] \item For Afrikaans, Armenian, Buryat, Gothic, Kurmanji, North Sami, Serbian and Upper Sorbian, we carry out our own pre-training on the Wikipedia dumps of these languages, tokenising them with the baseline UDPipe models and running the implementation of word2vec in the Gensim Python library\footnote{\url{https://radimrehurek.com/gensim/}} with 30 iterations and a minimum count of 1. \item For Breton and Thai, we use specially-trained multilingual embeddings (see Section~\ref{sec:multilingual}). \item For Naija and Old French, we substitute English and French embeddings, respectively. \item For Faroese, we do not use pre-trained embeddings. While it is possible to train such embeddings on Wikipedia data, as there is no UD training data for Faroese we choose instead to rely on its similarity to other Scandinavian languages (see Section~\ref{sec:multilingual}). \end{itemize} Word types in the training data that are not found amongst the pre-trained embeddings are initialized randomly using Glorot initialization \cite{glorot2010understanding}, as are all POS tag and treebank embeddings. Character vectors are also initialized randomly, except for Chinese, Japanese and Korean, in which case we pre-train character vectors using word2vec on the Wikipedia dumps of these languages. At test time, we first look for out-of-vocabulary (OOV) words and characters (i.e., those that are not found in the treebank training data) amongst the pre-trained embeddings and otherwise assign them a trained OOV vector.\footnote{An alternative strategy is to have the parser store embeddings for all words that appear in either the training data or pre-trained embeddings, but this uses far more memory.} A variant of word dropout is applied to the word embeddings, as described in \citet{kiperwasser16b}, and we apply dropout also to the character vectors. plus first item on the buffer with its leftmost dependent). We train all models for 30 epochs with hyper-parameter settings shown in Table~\ref{tbl:hyper}. Note our unusually large character embedding sizes; we have previously found these to be effective, especially for morphologically rich languages \citep{smith18}. Our code is publicly available. We release the version used here as UUParser 2.3.\footnote{\url{https://github.com/UppsalaNLP/uuparser}} \begin{table}[t] \begin{center} \begin{scalebox}{0.85}{ \begin{tabular}{ l | c } \hline Character embedding dimension & 500 \\ Character BiLSTM layers & 1 \\ Character BiLSTM output dimension & 200 \\ \hline Word embedding dimension & 100 \\ POS embedding dimension & 20 \\ Treebank embedding dimension & 12 \\ Word BiLSTM layers & 2 \\ Word BiLSTM hidden/output dimension & 250 \\ \hline Hidden units in MLP & 100 \\ \hline Word dropout & 0.33 \\ $\alpha$ (for OOV vector training) & 0.25 \\ Character dropout & 0.33 \\ $p_{agg}$ (for exploration training) & 0.1 \\ \hline \end{tabular}} \end{scalebox} \caption{Hyper-parameter values for parsing.} \label{tbl:hyper} \end{center} \end{table} \paragraph*{Using Morphological Features} \indent Having a strong morphological analyzer, we were interested in finding out whether or not we can improve parsing accuracy using predicted morphological information. We conducted several experiments on the development sets for a subset of treebanks. However, no experiment gave us any improvement in terms of LAS and we decided not to use this technique for the shared task. What we tried was to create an embedding representing either the full set of morphological features or a subset of potentially useful features, for example case (which has been shown to be useful for parsing by \citet{kapociute13lithuanian} and \citet{eryigit08cl}), verb form and a few others. That embedding was concatenated to the word embedding at the input of the BiLSTM. We varied the embedding size (10, 20, 30, 40), tried different subsets of morphological features, and tried with and without using dropout on that embedding. We also tried creating an embedding of a concatenation of the universal POS tag and the Case feature and replace the POS embedding with this one. We are currently unsure why none of these experiments were successful and plan to investigate this in the future. It would be interesting to find out whether or not this information is captured somewhere else. A way to test this would be to use diagnostic classifiers on vector representations, as is done for example in \citet{hupkes2018visualisation} or in \citet{adi17fine}. \section{Results and Discussion} \label{sec:results} Table~\ref{tab:results} shows selected test results for the Upp\-sala system, including the two main metrics LAS and MLAS (plus a mono-treebank baseline for LAS),\footnote{Since our system does not predict lemmas, the third main metric BLEX is not very meaningful.} the sentence and word segmentation accuracy, and the accuracy of UPOS tagging and morphological features (UFEATS). To make the table more readable, we have added a simple color coding. Scores that are significantly higher/lower than the mean score of the 21 systems that successfully parsed all test sets are marked with two shades of green/red. The lighter shade marks differences that are outside the interval defined by the standard error of the mean ($\mu \pm \textsc{se}, \textsc{se} = \sigma / \sqrt{N}$) but within one standard deviation (std dev) from the mean. The darker shade marks differences that are more than one std dev above/below the mean ($\mu \pm \sigma$). Finally, scores that are no longer valid because of the Thai UPOS tagger are crossed out in yellow cells, and corrected scores are added where relevant. Looking first at the LAS scores, we see that our results are significantly above the mean for all aggregate sets of treebanks (\textsc{all}, \textsc{big}, \textsc{pud}, \textsc{small}, \textsc{low-resource}) with an especially strong result for the low-resource group (even after setting the Thai score to 0.00). If we look at specific languages, we do particularly well on low-resource languages like Breton, Buryat, Kazakh and Kurmanji, but also on languages like Arabic, Hebrew, Japanese and Chinese, where we benefit from having better word segmentation than most other systems. Our results are significantly worse than the mean only for Afrikaans AfriBooms, Old French SRCMF, Galician CTG, Latin PROIEL, and Portuguese Bosque. For Galician and Portuguese, this may be the effect of lower word segmentation and tagging accuracy. To find out whether our multi-treebank and multi-lingual models were in fact beneficial for parsing accuracy, we ran a post-evaluation experiment with one model per test set, each trained only on a single treebank. We refer to this as the mono-treebank baseline, and the LAS scores can be found in the second (uncolored) LAS column in Table~\ref{tab:results}. The results show that merging treebanks and languages did in fact improve parsing accuracy in a remarkably consistent fashion. For the 64 test sets that were parsed with a multi-treebank model, only four had a (marginally) higher score with the mono-treebank baseline model: Estonian EDT, Russian SynTagRus, Slovenian SSJ, and Turkish IMST. Looking at the aggregate sets, we see that, as expected, the pooling of resources helps most for \textsc{low-resource} (25.33 vs.\ 17.72) and \textsc{small} (63.60 vs.\ 60.06), but even for \textsc{big} there is some improvement (80.21 vs.\ 79.61). We find these results very encouraging, as they indicate that our treebank embedding method is a reliable method for pooling training data both within and across languages. It is also worth noting that this method is easy to use and does not require extra external resources used in most work on multilingual parsing, like multilingual word embeddings \cite{ammar16} or linguistic re-write rules \cite{aufrant+16} to achieve good results. Turning to the MLAS scores, we see a very similar picture, but our results are relatively speaking stronger also for \textsc{pud} and \textsc{small}. There are a few striking reversals, where we do significantly better than the mean for LAS but significantly worse for MLAS, including Buryat BDT, Hebrew HTB and Ukrainian IU. Buryat and Ukrainian are languages for which we use a multilingual model for parsing, but not for UPOS tagging and morphological features, so it may be due to sparse data for tags and morphology, since these languages have very little training data. This is supported by the observation that low-resource languages in general have a larger drop from LAS to MLAS than other languages. For sentence segmentation, the Uppsala system achieved the second best scores overall, and results are significantly above the mean for all aggregates except \textsc{small}, which perhaps indicates a sensitivity to data sparseness for the data-driven joint sentence and word segmenter (we see the same pattern for word segmentation). However, there is a much larger variance in the results than for the parsing scores, with altogether 23 treebanks having scores significantly below the mean . For word segmentation, we obtained the best results overall, strongly outperforming the mean for all groups except \textsc{small}. We know from previous work \citep{shao18} that our word segmenter performs well on more challenging languages like Arabic, Hebrew, Japanese, and Chinese (although we were beaten by the Stanford team for the former two and by the HIT-SCIR team for the latter two). By contrast, it sometimes falls below the mean for the easier languages, but typically only by a very small fraction (for example 99.99 vs.\ 100.00 for 3 treebanks). Finally, it is worth noting that the maximum-matching segmenter developed specifically for Thai achieved a score of 69.93, which was more than 5 points better than any other system. Our results for UPOS tagging indicate that this may be the strongest component of the system, although it is clearly helped by getting its input from a highly accurate word segmenter. The Upp\-sala system ranks first overall with scores more than one std dev above the mean for all aggregates. There is also much less variance than in the segmentation results, and scores are significantly below the mean only for five treebanks: Galician CTG, Gothic PROIEL, Hebrew HTB, Upper Sorbian UFAL, and Portuguese Bosque. For Galician and Upper Sorbian, the result can at least partly be explained by a lower-than-average word segmentation accuracy. The results for morphological features are similar to the ones for UPOS tagging, with the best overall score but with less substantial improvements over the mean. The four treebanks where scores are significantly below the mean are all languages with little or no training data: Upper Sorbian UFAL, Hungarian Szeged, Naija NSC and Ukrainian IU. All in all, the 2018 edition of the Uppsala parser can be characterized as a system that is strong on segmentation (especially word segmentation) and prediction of UPOS tags and morphological features, and where the dependency parsing component performs well in low-resource scenarios thanks to the use of multi-treebank models, both within and across languages. For what it is worth, we also seem to have the highest ranking single-parser transition-based system in a task that is otherwise dominated by graph-based models, in particular variants of the winning Stanford system from 2017 \citep{dozat-conll-2017}. \section{Extrinsic Parser Evaluation} \label{sec:epe} In addition to the official shared task evaluation, we also participated in the 2018 edition of the Extrinsic Parser Evaluation Initiative (EPE) \citep{epe18}, where parsers developed for the CoNLL 2018 shared task were evaluated with respect to their contribution to three downstream systems: biological event extraction, fine-grained opinion analysis, and negation resolution. The downstream systems are available for English only, and we participated with our English model trained on English EWT, English LinES and English GUM, using English EWT as the proxy. In the extrinsic evaluation, the Uppsala system ranked second for event extraction, first for opinion analysis, and 16th (out of 16 systems) for negation resolution. Our results for the first two tasks are better than expected, given that our system ranks in the middle with respect to intrinsic evaluation on English (9th for LAS, 6th for UPOS). By contrast, our performance is very low on the negation resolution task, which we suspect is due to the fact that our system only predicts universal part-of-speech tags (UPOS) and not the language specific PTB tags (XPOS), since the three systems that only predict UPOS are all ranked at the bottom of the list. \section{Sentence and Word Segmentation} \label{sec:segmentation} We employ the model of \citet{shao18} for joint sentence segmentation and word segmentation. Given the input character sequence, we model the prediction of word boundary tags as a sequence labelling problem using a BiRNN-CRF framework \citep{huang2015bidirectional,shao17}. This is complemented with an attention-based LSTM model \citep{bahdanau2014neural} for transducing non-segmental multiword tokens. To enable joint sentence segmentation, we add extra boundary tags as in \citet{uu-conll17}. We use the default parameter settings introduced by \citet{shao18} and train a segmentation model for all treebanks with at least 50 sentences of training data. For treebanks with less or no training data (except Thai discussed below), we substitute a model for another treebank/language: \begin{itemize}[topsep=2pt,itemsep=1pt] \item For Japanese Modern, Czech PUD, English PUD and Swedish PUD, we use the model trained on the largest treebank from the same language (Japanese GSD, Czech PDT, English EWT and Swedish Talbanken ). \item For Finnish PUD, we use Finnish TDT rather than the slightly larger Finnish FTB, because the latter does not contain raw text suitable for training a segmenter. \item For Naija NSC, we use English EWT. \item For other test sets with little or no training data, we select models based on the size of the intersection of the character sets measured on Wikipedia data (see Table \ref{tab:results} for details).\footnote{North Sami Giella was included in this group by mistake, as we underestimated the size of the treebank.} \end{itemize} \paragraph{Thai} Segmentation of Thai was a particularly difficult case: Thai uses a unique script, with no spaces between words, and there was no training data available. Spaces in Thai text can function as sentence boundaries, but are also used equivalently to commas in English . For Thai sentence segmentation, we exploited the fact that four other datasets are parallel, i.e., there is a one-to-one correspondence between sentences in Thai and in Czech PUD, English PUD, Finnish PUD and Swedish PUD.\footnote{This information was available in the README files distributed with the training data and available to all participants.} First, we split the Thai text by white space and treat the obtained character strings as potential sentences or sub-sentences. We then align them to the segmented sentences of the four parallel datasets using the Gale-Church algorithm \citep{gale1993program}. Finally, we compare the sentence boundaries obtained from different parallel datasets and adopt the ones that are shared within at least three parallel datasets. For word segmentation, we use a trie-based segmenter with a word list derived from the Facebook word embeddings.\footnote{\url{github.com/facebookresearch/fastText}} The segmenter retrieves words by greedy forward maximum matching \citep{wong96}. This method requires no training but gave us the highest word segmentation score of 69.93\% for Thai, compared to the baseline score of 8.56\%. \section{Tagging and Morphological Analysis} \label{sec:tagger} We use two separate instantiations of the tagger\footnote{\url{https://github.com/google/meta_tagger}} described in \citet{bohnet18acl} to predict UPOS tags and morphological features, respectively. The tagger uses a Meta-BiLSTM over the output of a sentence-based character model and a word model. There are two features that mainly distinguishes the tagger from previous work. The character BiLSTMs use the full context of the sentence in contrast to most other taggers which use words only as context for the character model. This character model is combined with the word model in the Meta-BiLSTM relatively late, after two layers of BiLSTMs. For both the word and character models, we use two layers of BiLSTMs with 300 LSTM cells per layer. We employ batches with 8000 words and 20000 characters. We keep all other hyper-parameters as defined in \newcite{bohnet18acl}. From the training schema described in the above paper, we deviate slightly in that we perform early stopping on the word, character and meta-model independently. We apply early stopping due to the performance of the development set (or training set when no development set is available) and stop when no improvement occurs in 1000 training steps. We use the same settings for UPOS tagging and morphological features. To deal with languages that have little or no training data, we adopt three different strategies: \begin{itemize}[topsep=2pt,itemsep=1pt] \item For the PUD treebanks (except Thai), Japanese Modern and Naija NSC, we use the same model substitutions as for segmentation (see Table \ref{tab:results}). \item For Faroese we used the model for Norwegian Nynorsk, as we believe this to be the most closely related language. \item For treebanks with small training sets we use only the provided training sets for training. Since these treebanks do not have development sets, we use the training sets for early stopping as well. \item For Breton and Thai, which have no training sets and no suitable substitution models, we use a bootstrapping approach to train taggers as described below. \end{itemize} \paragraph{Bootstrapping} We first annotate an unlabeled corpus using an external morphological analyzer. We then create a (fuzzy and context-independent) mapping from the morphological analysis to universal POS tags and features, which allows us to relabel the annotated corpus and train taggers using the same settings as for other languages . For Breton, we annotated about 60,000 sentences from Breton OfisPublik, which is part of OPUS,\footnote{\url{https://opus.nlpl.eu/OfisPublik.php}} using the Apertium morphological analyzer. The Apertium tags could be mapped to universal POS tags and a few morphological features like person, number and gender. For Thai, we annotated about 33,000 sentences from Wikipedia using PyThaiNLP\footnote{\url{https://github.com/PyThaiNLP/pythainlp/wiki/PyThaiNLP-1.4}} and mapped only to UPOS tags (no features). Unfortunately, we realized only after the test phase that PyThaiNLP was not a permitted resource , which invalidates our UPOS tagging scores for Thai, as well as the LAS and MLAS scores which depend on the tagger. Note, however, that the score for morphological features is not affected, as we did not predict features at all for Thai. The same goes for sentence and word segmentation, which do not depend on the tagger. \paragraph{Lemmas} Due to time constraints we chose not to focus on the BLEX metric in this shared task. In order to avoid zero scores, however, we simply copied a lowercased version of the raw token into the lemma column.

468da2158d12c3432709409dc93c6e93aa3cda92

\section{Introduction} Before presenting our results, let us discuss briefly the following simple problem. Assume that, on a suitable probability space, we have two sequences of random variables $\{X_n\}_{n\ge 1}$ and $\{Y_n\}_{n\ge 1}$ which fulfill the following conditions: \begin{itemize} \item[a)] for every $n\geq 1$ we have $\mathbb{E}(X_n)=\mathbb{E}(Y_n)=0$, \item[b)] for some $M>0$ and every $n\geq 1$ we assume that $\max(|X_n|,|Y_n|)\leq M$ almost surely, \item[c)] for every $n\geq 1$, $\mathbb{E}(X_n Y_n)=0$, \item[d)] $X_n+Y_n\xrightarrow[n\rightarrow\infty]{\text{a.s.}}~0.$ (a.s. is short for ``almost surely''.) \end{itemize} \begin{center} {\it Do we necessarily have that $X_n\xrightarrow[n\rightarrow\infty]{\text{a.s.}}~0$ and $Y_n\xrightarrow[n\rightarrow\infty]{\text{a.s.}}~0$?} \end{center} Using dominated convergence and the orthogonality assumption c), we get $\mathbb{E}(X_n^2)+\mathbb{E}(Y_n^2)\rightarrow 0$ which is not far from a positive answer to the previous question, up to extracting a subsequence. However, without additional structure assumptions, the answer is {\it totally negative}, as indicated by the following counterexample. The starting point is to consider an uniformly bounded sequence of $1$-periodic functions that are centered and converge to zero in $L^2([0,1])$ but not almost surely for the Lebesgue measure $\lambda$ on $[0,1]$. A standard way to build this object is to consider for every $p\ge 0$ and every integer $n \in [2^p,2^{p+1}-1]$, the indicator function $f_n$ of the set $\left[ (n-2^p)2^{-p}, (n-2^p+1)2^{-p}\right]$. The convergence to zero of the $L^2([0,1])$-norm is straightforward while one may observe that for every non dyadic $x \in [0,1]$ and every $A>0$ one can find $n\ge A$ such that $f_n(x)=1$, which disproves the almost sure convergence. We shall denote by $\{f_n\}_{n\ge 1}$ the resulting sequence of functions and by $\tilde{f}_n$ the centered functions $f_n-\int_0^1 f_n(x) dx$. Clearly the previous observations are also valid for $\tilde{f}_n$, since $\int_0^1 f_n(x) dx \rightarrow 0.$ Next, given $ (e_p(x),f_p(x))_{p\ge 0}:=\big(\sqrt{2}\cos(2 p \pi x),\sqrt{2}\sin(2 p \pi x) \big)_{p\ge 0} $ an orthonormal basis of $L^2([0,1])$, we can expand $\tilde{f}_n$ in the $L^2([0,1])$ as follows: $$ \tilde{f}_n=\sum_{p=1}^\infty \alpha_{p,n} e_p +\beta_{p,n} f_p \, . $$ Besides, set $\epsilon_n=\sum_{p=2}^\infty \alpha_{p,n}^2+\sum_{p=1}^\infty \beta_{p,n}^2$ and $g_n=\tilde{f}_n-\alpha_{1,n} e_1+\sqrt{\epsilon_n} e_1$ and it is clear that $g_n$ and $\tilde{f}_n$ enjoy the same aforementionned properties. Finally we set $h_n=-g_n+2\sqrt{\epsilon_n} e_1$. We claim that $\{g_n,h_n\}$ fulfill the assumptions a), b), c), d) on the probability space $([0,1],\mathcal{B}([0,1]),\lambda)$. Since $\tilde{f}_n$ and $e_1$ are centered we get that $g_n$ and $h_n$ are centered. Since $f_n$ is uniformly bounded as well as the sequences $\{\alpha_{1,n}\}_{n\ge 1}$ and $\{\epsilon_n\}_{n\ge 1}$ the same conclusion is valid for $\{g_n,h_n\}_{n\ge 1}$. Thus a) and b) hold. Besides, $g_n+h_n=2\sqrt{\epsilon_n} e_1$ tends to zero almost surely and d) is true. Finally, the orthogonality assumption c) also holds as illustrated by the next computations: \begin{eqnarray*} \int_0^1 g_n(x) h_n(x) dx &=& -\int_0^1 g_n^2(x) dx +2 \epsilon_n\\ &=& -2 \epsilon_n+2 \epsilon_n\\ &=&0. \end{eqnarray*} As a conclusion, the orthogonality assumption is not sufficiently powerful to \textit{decompose} the almost sure convergence on the orthogonal projections. This short note explores the previous problem in the specific framework of Gaussian Wiener chaoses and homogeneous sums in independent random variables, which are well known to display much more structure besides orthogonality. In this setting, we shall see that the previous question has a \textit{positive} answer. Let us begin with one of our main results: We assume that all the random objects in this note are defined on a common probability space $\big( \Omega, \mathscr{F}, \mathbb{P} \big)$. \begin{thm}\label{thm-Gauss} Fix an integer $q\geq 2$ and let $\{F_n, n\geq 1\}$ be a sequence in the sum of the first $q$ Gaussian Wiener chaoses $($associated to a given isonormal Gaussian process $W$$)$ such that as $n\rightarrow+\infty$, $F_n$ converges almost surely to some random variable, denoted by $F$. Then, $F$ belongs to the sum of the first $q$ Gaussian Wiener chaoses. Moreover, as $n\rightarrow+\infty$, \begin{align}\label{ASG} \text{$J_p(F_n)$ converges almost surely to $J_p(F)$} \end{align} for each $p\in\{0,\cdots,q\}$. Here $J_p(\bullet)$ denotes the projection operator onto the $p$th Gaussian Wiener chaos, see Section \ref{sec-Gauss} for more details. \end{thm} The concept of chaos that we consider dates back to N. Wiener's 1938 paper \cite{Wiener38}, in which Wiener first introduced the notion of \emph{multiple Wiener integral} calling it polynomial chaos. It was later modified in K. It\^o's work \cite{Ito51, Ito56} so that the multiple Wiener integrals of different orders are orthogonal to each other. The multiple Wiener integrals have proved to be a very useful tool in the investigation of the asymptotic behavior of partial sums of dependent random variables; see P. Major's book \cite{Major-book}. Recently, the multiple Wiener integrals that we call Wiener chaos here have gained considerable attention, as one can see along the expanding research line passing from D. Nualart and G. Peccati's fourth moment theorem \cite{FMT} to the birth of the so-called Malliavin-Stein approch \cite{NP09MS}; see the book \cite{bluebook} and \cite[Chapter 1]{GZthesis} for more details. The results in \cite{Major-book, bluebook, FMT} mostly concern the convergence in distribution, and related ideas have been used to investigate the total-variation convergence on Wiener chaoses, see for example \cite{NourdinPoly13}. Our Theorem \ref{thm-Gauss}, to the best of our knowledge, is the very first instance that focuses on the almost sure convergence on Wiener chaoses. Note that the almost sure convergence in \eqref{ASG} occurs along the whole sequence, {\it not just} along some subsequence. In fact, the almost sure convergence along some subsequence is quite straightforward: As a consequence of \emph{hypercontractivity property} of the \emph{Ornstein-Uhlenbeck semigroup}, the almost sure convergence of $F_n$ to $F$ implies the $L^2(\mathbb{P})$ convergence of $F_n$ to $F$, so that $F$ belongs to the sum of the first $q$ Gaussian Wiener chaoses; at the same time, we have $J_p(F_n)$ converges in $L^2(\mathbb{P})$ to $J_p(F)$ because of the orthogonality property, hence the desired almost sure convergence holds true along some subsequence; see Section \ref{sec-Gauss} for more explanation. Let us first sketch our strategy of showing \eqref{ASG}, which motivates us to investigate the almost sure convergence on homogeneous polynomials in independent random variables. By assumption, $F_n = \mathfrak{f}_n(W)\rightarrow F=\mathfrak{f}(W)$ almost surely, where $\mathfrak{f}_n, \mathfrak{f}$ are deterministic representations of $F_n$ and $F$ respectively. Then it also holds true that $\mathfrak{f}_n(W^t)\rightarrow \mathfrak{f}(W^t)$ almost surely for each $t>0$, where \begin{align}\label{Wt} W^t : = e^{-t}W + \sqrt{1 - e^{-2t}} \widehat{W} \quad\text{with $\widehat{W}$ an independent copy of $W$,} \end{align} is an isonormal Gaussian process that is exchangeable with $W$. That is, $\big( W^t, W \big)$ has the same law as $\big( W, W^t \big)$ and in particular, $W^t$ has the same law as $W$. Conditioning on $W$, $\mathfrak{f}_n(W^t)$ belongs to the sum of the first $q$ Gaussian Wiener chaoses associated to $\widehat{W}$ and converges to $\mathfrak{f}(W^t)$ almost surely, therefore we deduce from the hypercontractivity that \begin{center} conditioning on $W$, $\mathfrak{f}_n(W^t)$ converges in $L^2\big(\Omega, \sigma\{ \widehat{W} \}, \mathbb{P} \big)$ to $\mathfrak{f}(W^t)$. \end{center} It follows immediately by taking expectation with respect to $\widehat{W}$ that almost surely, \[ \mathbb{E}\Big[ \mathfrak{f}_n(W^t) \vert \sigma\{W\} \Big] \rightarrow \mathbb{E}\Big[ \mathfrak{f}(W^t) \vert \sigma\{W\} \Big] \,. \] By Mehler's formula \eqref{Mehler}, we rewrite the above convergence as follow: $P_t F_n \rightarrow P_t F$, where $(P_t, t\geq 0)$ is the Ornstein-Uhlenbeck semigroup associated to $W$. By the definition of this semigroup, one has \begin{align}\label{OUASG} P_t F_n = \sum_{k=0}^q e^{-kt} J_k(F_n) \quad\text{converges almost surely to } \quad P_t F = \sum_{k=0}^q e^{-kt} J_k(F) \end{align} for each $t > 0$. With many enough $t > 0$, we can deduce \eqref{ASG} from \eqref{OUASG}. Now let us summarize the main idea of the proof: We introduced extra randomness $\widehat{W}$ to construct an identical copy $W^t$ of $W$, then we carried out the conditioning argument that ``separates'' Wiener chaoses of different orders (see \eqref{OUASG}). The above construction \eqref{Wt} of exchangeable pairs has been used in \cite{NZ17} to provide an elementary proof of the quantitative fourth moment theorem in \cite{NP09MS}, whose ideas have been further extended in the Ph.D thesis \cite{GZthesis}. We remark here that the use of exchangeable pairs in aforementioned references is for proving the distributional convergence within the framework of Stein's method of probabilistic approximation \cite{ICM14, ChenPoly, Stein86}; while we use exchangeable\footnote{In fact, we just need the identical pairs to proceed our arguments in this note.} pairs to investigate the almost sure convergence, which is of some independent interest. As announced, following the above strategy (but using different constructions of exchangeable pairs), we are able to provide two more results concerning homogeneous sums. Let us introduce a few notation first. \medskip \noindent{\bf Notation.} Throughout this note, $\mathbb{N}: =\{1, 2, \ldots \}$, $\mathbb{X} = \{ X_i, i\in\mathbb{N} \}$ denotes a sequence of centered independent real random variables and $ \ell^2_0(\mathbb{N})^{\odot p}$, with $p\geq 1$, stands for the set of kernels $f\in \ell^2(\mathbb{N})^{\otimes p}$ that is symmetric in its arguments and vanishes on its diagonals, that is, $f(i_1, \ldots, i_p) =0$ whenever $i_j = i_k$ for some $j\neq k$. Given $f\in \ell^2_0(\mathbb{N})^{\odot p}$, we define the homogeneous sums (or homogeneous polynomials) with order $p$, based on the kernel $f$, by setting \begin{align}\label{QPF} Q_p\big(f; \mathbb{X} \big) = \sum_{i_1, \ldots, i_p\in\mathbb{N}} f(i_1, \ldots, i_p) X_{i_1} \cdots X_{i_p} \,, \end{align} whenever the above sum is well defined, for example when $f$ has finite support, or when $X_i$ has unit variance for each $i\in\mathbb{N}$. \begin{obs}\label{rem12} {\rm If $\mathbb{X}$ is a sequence of symmetric Rademacher random variables\footnote{That is, $X_i$ takes values in $\{-1, +1\}$ with equal probability for each $i\in\mathbb{N}$.}, the random variable defined in \eqref{QPF} is an element in the so-called Rademacher chaos of order $p$; if $\mathbb{X}$ is a sequence of centered standard Poisson random variables\footnote{That is, $X_i = P_i - 1$ with $\{ P_i, i\in\mathbb{N}\}$ a sequence of i.i.d. standard Poisson random variables.}, the random variable $Q_p\big(f; \mathbb{X} \big)$ in \eqref{QPF} is a particular example of the Poisson Wiener chaos of order $p$; see Section 1.3 in \cite{GZthesis}. } \end{obs} We are now in a position to state our results. \begin{thm} \label{HP-bdd} Fix an integer $q\geq 2$ and let $\mathbb{X}$ be a sequence of centered independent real random variables such that $\mathbb{E}[ X_i^2] =1$ for each $i\in\mathbb{N}$ and $ \sup\{\mathbb{E}[ \vert X_j\vert^{2+\delta}] : j\in\mathbb{N} \} < \infty$ for some $\delta>0$. Assume $\big(F_n, n\geq 1\big)$ is a sequence of random variables satisfying the following conditions: \begin{enumerate} \item[\rm (i)] For each $n\geq 1$, $F_n$ is of the form \begin{align}\label{Fnp} F_n = \mathbb{E}[F_n] + \sum_{p=1}^q Q_p\big(f_{p,n}; \mathbb{X} \big) \quad\text{with $f_{p,n}\in\ell^2_0(\mathbb{N})^{\odot p}$ for each $p\in\{1,\ldots, q\}$.} \end{align} \item[\rm (ii)] $F_n$ converges almost surely to the random variable $F$ of the following form: \begin{align}\label{Fpinf} F = \mathbb{E}[ F] + \sum_{p=1}^q Q_p(f_{p}; \mathbb{X}) \quad\text{with $f_{p}\in\ell^2_0(\mathbb{N})^{\odot p}$ for each $p\in\{1,\ldots, q\}$.} \end{align} \end{enumerate} Then, as $n\rightarrow+\infty$, $\mathbb{E}[F_n]$ converges to $\mathbb{E}[F]$; and for each $p\in\{1, \ldots, q\}$, $Q_p\big(f_{p,n}; \mathbb{X} \big) $ converges almost surely to $Q_p\big(f_{p}; \mathbb{X} \big) $. \end{thm} Note that Theorem \ref{HP-bdd} covers the examples from Remark \ref{rem12}: In particular when $\mathbb{X}$ is a sequence given therein, we do not need to assume that $F$ belongs to the sum of the first $q$ chaos, because this comes as a consequence of the hypercontractivity property (Lemma \ref{hyper}). \medskip In the following result, we remove the $(2+\delta)$-moment assumption with the price of imposing finite-support assumption. It is worth pointing out that we only require $X_k$'s have first moments. \begin{thm}[\bf Unbounded case]\label{HP-ubdd} Fix an integer $q \geq 2$ and consider a sequence of centered independent {\rm (non-deterministic)} random variables $\mathbb{X} = \{X_k, k\in\mathbb{N} \}$. Let $F_n$ be a sequence of random variables given by \[ F_n = \mathbb{E}[F_n] + \sum_{p=1}^q Q_p\big(f_{p,n}; \mathbb{X}\big) \quad\text{with $f_{p,n}\in\ell^2_0(\mathbb{N})^{\odot p}$ for each $p\in\{1,\ldots, q\}$.} \] We assume that $d > q$ is an integer such that the support of $f_{p,n}$ is contained in $\{1, \ldots, d\}^p$ for any $p=1,\ldots, q$, $n\geq 1$. If $F_n$ converges almost surely to $F$ with $F$ given by \[ F = \mathbb{E}[F] + \sum_{p=1}^q Q_p\big(f_{p}; \mathbb{X}\big) \quad\text{with $f_{p}\in\ell^2_0(\mathbb{N})^{\odot p}$ for each $p\in\{1,\ldots, q\}$.} \] Then, as $n\rightarrow\infty$, $\mathbb{E}[F_n]\rightarrow \mathbb{E}[F]$; and $\forall p\in\{1,\ldots, q\}$, $Q_p\big(f_{p,n}; \mathbb{X}\big)$ converges almost surely to $Q_p\big(f_{p}; \mathbb{X}\big)$. \end{thm} This paper naturally leads to the following open questions. \noindent{\bf Open question 1.} Let $\mathbb{Y}$ be a sequence of independent (nonsymmetric) Rademacher variables, that is, $\mathbb{P}\big( Y_k = 1) = p_k = 1 - \mathbb{P}\big( Y_k = -1)\in(0,1)$ for each $k\in\mathbb{N}$. Define $X_k = (Y_k +1 - 2p_k)/(2\sqrt{p_k - p_k^2} )$ for each $k\in\mathbb{N}$. Assume $F_n$ and $F$ have the form \eqref{Fnp} and \eqref{Fpinf} respectively such that $F_n$ converges almost surely to $F$. Does the almost surely convergence in Theorem \ref{HP-bdd} hold true? \medskip \noindent{\bf Open question 2.} Fix an integer $q\geq 2$ and let $(F_n, n\in\mathbb{N})$ be a sequence of random variables that belong to the sum of the first $q$ Poisson Wiener chaoses\footnote{Say, these Poisson Wiener chaoses are defined based a Poisson random measure over a $\sigma$-finite measure space; see the book \cite{LP-book} for more details.} such that $F_n$ converges almost sure to some $F$ in the sum of the first $q$ Poisson Wiener chaoses. \begin{center} {\it Does $J_p(F_n)$ converge almost surely to $J_p(F)$, as $n\rightarrow+\infty$?} \end{center} Here $J_p(\bullet)$ denotes the projection operator onto the $p$th Poisson Wiener chaos, and one can refer to the book \cite{LP-book} for any unexplained term. \medskip The rest of this article is organized as follows: Section \ref{sec-Gauss} is devoted to a detailed proof of Theorem \ref{thm-Gauss} and Section \ref{sec-HP} consists of two subsections that deal with Theorem \ref{HP-bdd} and Theorem \ref{HP-ubdd} respectively. \section{Almost sure convergence on the Gaussian Wiener chaoses}\label{sec-Gauss} Let $W:= \big\{ W(h) : h\in\mathfrak{H} \big\}$ be an isonormal Gaussian process over a real separable Hilbert space $\mathfrak{H}$, \emph{i.e.} $W$ is a centered Gaussian family such that $\mathbb{E}[ W(h) W(g) ] = \langle h, g \rangle_\mathfrak{H}$ for any $h, g\in\mathfrak{H}$. The resulting $L^2(\Omega, \sigma\{W\}, \mathbb{P})$ can be decomposed in an {\it orthogonal} manner as a direct sum of Gaussian Wiener chaoses: $ L^2(\Omega, \sigma\{W\}, \mathbb{P}) = \bigoplus_{p=0}^\infty \mathbb{C}^W_p $\, where $\mathbb{C}^W_0 = \mathbb{R}$ and for $p\geq 1$, the $p$th Gaussian Wiener chaos admits a complete orthonormal system given by \begin{align}\label{OBCP} \left\{ \prod_{i\in\mathbb{N}} \frac{H_{a_i}(W(e_i) )}{\sqrt{a_i!}} \,: \, a_i\in\mathbb{N}\cup\{0\} \,\, \text{such that} \,\, \sum_{i\in\mathbb{N}}a_i = p \right\} \,. \end{align} In this note, $H_p(x) := (-1)^p \exp(x^2/2) \frac{d^p}{dx^p}\exp(-x^2/2) $ is the $p$th Hermite polynomial\footnote{The first few Hermite polynomials are given by $H_0(x) =1$, $H_1(x) = x$, $H_2(x) = x^2 - 1$ and $H_{p+1}(x) = xH_p(x) - pH_{p-1}(x)$ for any $p\in\mathbb{N}$. } and $\{ e_i : i\in\mathbb{N}\}$ stands for an orthonormal basis of $\mathfrak{H}$. Recall that $J_p(\bullet)$ denotes the projection operator onto $\mathbb{C}^W_p$, and we can define the Ornstein-Uhlenbeck semigroup $(P_t, t\geq 0)$ by setting $ P_t F = \mathbb{E}[ F] + \sum_{p=1}^\infty e^{-pt} J_p(F)$, $\forall F\in L^2(\Omega, \sigma\{W\}, \mathbb{P})$. It also has the following nice representation that is of central importance to our approach: given $F\in L^2(\Omega, \sigma\{W\}, \mathbb{P})$, we first have $F = \mathfrak{f}(W)$ for some deterministic representation $\mathfrak{f}: \mathbb{R}^\mathfrak{H}\rightarrow \mathbb{R}$, then Mehler's formula reads as follows: \begin{align}\label{Mehler} P_tF= \mathbb{E}\big[ \mathfrak{f}(W^t) \vert \sigma\{ W\} \big] \end{align} where $W^t$ is another isonormal Gaussian process defined in \eqref{Wt}. As a consequence of the hypercontractivity property of the Ornstein-Uhlenbeck semigroup, we have the equivalence of all $L^r(\mathbb{P})$-norm $(1< r < +\infty)$ on a fixed Gaussian Wiener chaos. For any unexplained term, one can refer to \cite{bluebook, Nualart06}. \medskip Now we are ready to present the proof of Theorem \ref{thm-Gauss}. \begin{proof}[Proof of Theorem \ref{thm-Gauss}] As a consequence of the \emph{hypercontractivity} (see \emph{e.g.} \cite[Section 1.4.3]{Nualart06}), $\{F_n, n\geq 1\}$ is bounded in $L^m(\mathbb{P})$ for any $m>1$. Therefore, $\mathbb{E}[ \vert F\vert^m] < +\infty$ for any $m >1$ and \[ \mathbb{E}\big[ (F_n - F)^2 \big] = \mathbb{E}[ F_n^2] + \mathbb{E}[ F^2] - 2 \mathbb{E}[ F_n F] \xrightarrow{n\rightarrow+\infty} \mathbb{E}[ F^2] + \mathbb{E}[ F^2] - 2 \mathbb{E}[ F^2] =0 \, , \] where the limit follows from the almost sure convergence of $F_nF$ and its uniform integrability. So we can conclude that $F$ belongs to the sum of the first $q$ Gaussian Wiener chaoses. \noindent{\bf Observation:} Given $G = \mathfrak{g}(W)\in L^2(\mathbb{P})$ for some deterministic representation $\mathfrak{g}:\mathbb{R}^\mathfrak{H}\rightarrow \mathbb{R}$, then in view of \eqref{OBCP}, $G$ belongs to the sum of the first $q$ Gaussian Wiener chaoses (associated to $W$) if and only if $G$ is a polynomial in i.i.d. Gaussians $\{ W(e_i)\,:\, i\in\mathbb{N} \}$ with degree $\leq q$. It follows from this equivalence that conditioning on $W$, $\mathfrak{g}(W^t)$ belongs to the sum of the first $q$ Gaussian Wiener chaoses (associated to $\widehat{W}$), which can be seen from the simple formula $H_n(ax+by) = \sum_{k=0}^n {n\choose k} a^kb^{n-k}H_k(x) H_{n-k}(y)$ with $a,b\in\mathbb{R}$ such that $a^2 + b^2 =1$. \medskip With the above observation and assuming $F_n = \mathfrak{f}_n(W)$, $F = \mathfrak{f}(W)$ for some deterministic $\mathfrak{f}_n$ and $\mathfrak{f}$, it holds that $\mathfrak{f}_n(W^t)$ converges almost surely to $\mathfrak{f}(W^t)$. Therefore, conditioning on $W$, $\mathfrak{f}_n(W^t)$ belongs to the sum of the first $q$ Gaussian Wiener chaoses (associated to $\widehat{W}$) for each $n\in\mathbb{N}$ and converges almost surely to $\mathfrak{f}(W^t)$. Thus, due to the hypercontractivity, we have conditioning on $W$, the expectation of $\mathfrak{f}_n(W^t)$ with respect to $\widehat{W}$ converges to that of $\mathfrak{f}(W^t)$ with respect to $\widehat{W}$. That is, almost surely, $ P_tF_n = \mathbb{E}\big[ \mathfrak{f}_n(W^t)\, \vert \sigma\{W\} \big] \rightarrow \mathbb{E}\big[ \mathfrak{f}(W^t)\, \vert \sigma\{W\} \big] = P_tF $, as $n\rightarrow+\infty$. Note that the two equalities in the above display follow from Mehler's formula \eqref{Mehler}. Hence by definition, we have \[ \mathbb{E}[F_n] + \sum_{k=1}^q e^{-kt} J_k(F_n) \longrightarrow \mathbb{E}[F] + \sum_{k=1}^q e^{-kt} J_k(F) \, \] almost surely, as $n\rightarrow+\infty$. Applying the above argument for sufficiently many $t > 0$ gives the desired result. \qedhere \end{proof} As an interesting corollary, we show that, for sequences of random variables lying in a fixed finite sum of Gaussian Wiener chaoses, the almost sure convergence is robust to the application of the standard Malliavin calculus operators $L$ and $\Gamma$, where the {\it Ornstein-Uhlenbeck} generator $L$ is defined formally by $L = \sum_{p\geq 1} -p J_p$ and the \emph{carr\'e-du-champ} operator is defined by $\Gamma[F,G] = \frac{1}{2}\big( L[FG] - FL[G] - GL[F] \big)$ whenever the expressions make sense. For example, when $F, G$ live in a fixed finite sum of Gaussian Wiener chaoses, $L[F]$ and $\Gamma[F,G]$ are well defined. More precisely, we have the following result. \begin{cor}\label{corr0} For fixed integers $p, q\in\mathbb{N}$, we consider a sequence $F_n \in \bigoplus_{k\le p} \mathbb{C}_k^{W}$ which converges almost surely towards $F$, as $n\rightarrow+\infty$ Then, we have $ L[F_n] \xrightarrow[n\rightarrow\infty]{\text{a.s.}} L[F] \,.$ Consider another sequence $G_m \in \bigoplus_{k\le q} \mathbb{C}_k^{W}$ that converges almost surely towards $G$, as $m\rightarrow+\infty$. Then, we have $ \Gamma[F_n, G_n] \xrightarrow[n\rightarrow\infty]{\text{a.s.}} \Gamma[F, G]$. \end{cor} \begin{proof} First of all, Theorem \ref{thm-Gauss} tells us that $J_k(F_n)$ converges towards $J_k(F)$ almost surely for every $k\in\{ 0, \ldots, p\}$. Thus, we get \[ L[F_n]= \sum_{k=0}^p -k J_k(F_n) \xrightarrow[n\rightarrow\infty]{\text{a.s.}} L [F]. \]This achieves the proof of the first assertion and in the same way, $ L[G_n] \xrightarrow[n\rightarrow\infty]{\text{a.s.}} L[G] \,.$ Second assertion follows easily from the first one: Clearly $F_nG_n$ lives in the sum of the first $(p+q)$ Gaussian Wiener chaos in view of the multiplication formula (see \emph{e.g.} \cite{Nualart06}) and $F_nG_n$ converges to $FG$ almost surely, thus $L[F_nG_n]$ converges almost surely to $L[FG]$, implying $ \Gamma[F_n, G_n] \xrightarrow[n\rightarrow\infty]{\text{a.s.}} \Gamma[F, G]$. This completes our proof. \qedhere \end{proof} \begin{obs} {\rm As we have pointed out, Theorem \ref{HP-bdd} covers the case where $F_n$ belongs to a fixed sum of Rademacher chaoses (in the symmetric setting), so a similar result to the above Corollary \ref{corr0} can be formulated. } \end{obs} \section{Almost sure convergence on homogeneous sums}\label{sec-HP} This section is divided into two parts, presenting the proof of Theorems \ref{HP-bdd}, \ref{HP-ubdd} respectively. Having our general strategy in mind, we will first present our use of auxiliary randomness both in Section \ref{bddsec} and in Section \ref{ubddsec}. \subsection{The case of bounded $(2+\delta)$ moments}\label{bddsec} Recall that $\mathbb{X} = ( X_i, i\in\mathbb{N})$ is a sequence of centered independent random variables. Let $\widehat{\mathbb{X}} = \big( \widehat{X}_i, i\in\mathbb{N} \big)$ is an independent copy of $\mathbb{X}$ and let $\Theta = \big( \theta_i, i\in\mathbb{N} \big)$ be a sequence of i.i.d. standard exponential random variables such that $\mathbb{X},\widehat{\mathbb{X}}$ and $\Theta$ are independent. Given any $t > 0$, we define a new sequence $\mathbb{X}^t= ( X^t_i, i\in\mathbb{N})$ by setting $ X^t_i = X_i\cdot\mathbbm{I} _{\{\theta_i \geq t \}} + \widehat{X}_i\cdot \mathbbm{I} _{\{\theta_i < t \}}$ for each $i\in\mathbb{N}$. It is routine to verify that $\mathbb{X}^t$ has the same law as $\mathbb{X}$: Indeed, given any $y\in\mathbb{R}, i\in\mathbb{N}$, we have $ \mathbb{P}\big( X_i^t \leq y \big) = \mathbb{P}\big( X_i \leq y , \theta_i \geq t \big) + \mathbb{P}\big( \widehat{X}_i \leq y , \theta_i < t \big) = \mathbb{P}\big( X_i \leq y \big) $ and moreover it is clear that $\mathbb{X}^t$ is a sequence of independent random variables, thus $\mathbb{X}^t$ has the same law as $\mathbb{X}$. Another ingredient for our proof is the following {\it hypercontractivity property}. \begin{lemma}\label{hyper} Let $\Xi=(\xi_i, i\in\mathbb{N})$ be a sequence of real centered independent random variables such that $\mathbb{E}[ \xi_i^2] =1$ for each $i\in\mathbb{N}$ and $M:=\sup\big\{\mathbb{E}[ \vert \xi_j\vert^{2+\delta} ] \,:\, j\in\mathbb{N} \big\}$ is finite for some $\delta > 0$. Given any $f\in\ell^2_0(\mathbb{N})^{\odot d}$, one has \begin{align}\label{onehas} \big\| Q_d(f ; \Xi ) \big\| _{L^{2+\delta}(\mathbb{P})} \leq \left( 2\sqrt{1+\delta} \cdot M^{1/(2+\delta)} \right)^d \big\| Q_d(f ; \Xi ) \big\| _{L^2(\mathbb{P})} \,. \end{align} As a consequence: Given a fixed integer $q\geq 2$, $ F_n := \mathbb{E}[F_n] + \sum_{p=1}^q Q_p\big(f_{p,n};\Xi \big)$, with $f_{p,n}\in\ell^2_0(\mathbb{N})^{\odot p}$ for each $p\in\{1,\ldots, q\}$, if $\{F_n\}_{n\geq 1}$ is tight, then $\sup\big\{ \mathbb{E}[ \vert F_n\vert^{2+\delta} ] : n\in\mathbb{N} \big\}< +\infty$. \end{lemma} \begin{proof} Our lemma follows easily from Propositions 3.11, 3.12 and 3.16 in \cite{MOO10}: let us first truncate the kernel $f_n = f \cdot\mathbbm{I} _{\{1, \ldots, n\}^d}$ with any $n\geq d$, then $Q_d(f ; \Xi )$ is simply a multilinear polynomial in $\xi_1, \ldots, \xi_n$ so that the results in \cite{MOO10} imply $ \big\| Q_d(f_n ; \Xi ) \big\| _{L^{2+\delta}(\mathbb{P})} \leq \left( 2\sqrt{1+\delta} \cdot M^{1/(2+\delta)} \right)^d \big\| Q_d(f_n ; \Xi ) \big\| _{L^2(\mathbb{P})} $, thus by passing $n$ to infinity, we get \eqref{onehas}; and by Minkowski's inequality, we get $\| F_n \| _{L^{2+\delta}(\mathbb{P})} \leq \kappa\cdot \| F_n \| _{L^2(\mathbb{P})}$ with $\kappa:= \left( 2\sqrt{1+\delta} \cdot M^{1/(2+\delta)} \right)^q (q+1)$. Note that \begin{align*} \mathbb{E}\big[ F_n^2 ] \leq \mathbb{E}\big[ F_n^2 \cdot\mathbbm{I} _{\{ F_n^2 > \mathbb{E}[ F_n^2]/2 \}} \big] + \frac{1}{2} \mathbb{E}[ F_n^2] \leq \frac{1}{2} \mathbb{E}[ F_n^2]+ \| F_n \| _{L^{2+\delta}(\mathbb{P})}^{2} \cdot \mathbb{P}\big( F_n^2 > \mathbb{E}[ F_n^2]/2 \big)^{\delta/(2+\delta)} \end{align*} where the last inequality follows from the H\"older inequality. Therefore, \[ \mathbb{P}\big( F_n^2 > \mathbb{E}[ F_n^2]/2 \big) \geq \left(1/2\right)^{(2+\delta)/\delta} \left( \| F_n \| _{L^2(\mathbb{P})} \big/ \| F_n \| _{L^{2+\delta}(\mathbb{P})} \right)^{2(2+\delta)/\delta} \geq (2\kappa^2)^{-(2+\delta)/\delta} \,\,~\text{for each $n\in\mathbb{N}$} \] while due to tightness, one can find $K > 0$ large enough such that $\mathbb{P}\big( F_n^2 > K \big) < (2\kappa^2)^{-(2+\delta)/\delta}$ for each $n\in\mathbb{N}$, implying that $ \mathbb{P}\big( F_n^2 > \mathbb{E}[ F_n^2]/2 \big) > \mathbb{P}\big( F_n^2 > K \big)$. This gives us $\mathbb{E}[ F_n^2] \leq 2K $, $\forall n\in\mathbb{N}$. Hence, $\| F_n \| _{L^{2+\delta}(\mathbb{P})} \leq \kappa \cdot \sqrt{2K} < +\infty$ for each $n\in\mathbb{N}$. \end{proof} \begin{proof}[Proof of Theorem \ref{HP-bdd}] Without losing any generality, we can assume that $F_n = \mathfrak{f}_n(\mathbb{X})$ and $F = \mathfrak{f}(\mathbb{X})$ for some deterministic mappings $\mathfrak{f}_n, \mathfrak{f}$ from $\mathbb{R}^\mathbb{N}$ to $\mathbb{R}$. By our construction of $\mathbb{X}^t$, it follows that $ \mathfrak{f}_n(\mathbb{X}^t) = \mathbb{E}[ F_n] + \sum_{p=1}^q Q_p\big(f_{p,n}; \mathbb{X}^t\big) \longrightarrow \mathfrak{f}(\mathbb{X}^t) = \mathbb{E}[ F] + \sum_{p=1}^q Q_p\big(f_{p}; \mathbb{X}^t\big)$ almost surely. Now fixing a generic realization of $\mathbb{X}$, $\mathfrak{f}_n(\mathbb{X}^t)$ and $\mathfrak{f}(\mathbb{X}^t)$ are polynomials (with bounded degrees) in \[ \Xi:= \left\{ \widehat{X}_1, \frac{\mathbbm{I} _{\{ \theta_1 < t \}} -(1- e^{-t})}{ \sqrt{ e^{-t}(1-e^{-t}) } } , \widehat{X}_2, \frac{\mathbbm{I} _{\{ \theta_2 < t \}} -(1- e^{-t})}{ \sqrt{ e^{-t}(1-e^{-t}) } } , \widehat{X}_3, \frac{\mathbbm{I} _{\{ \theta_3 < t \}} -(1- e^{-t})}{ \sqrt{ e^{-t}(1-e^{-t}) } } \,, \ldots \, \right\} \] and $\mathfrak{f}_n(\mathbb{X}^t)$ converges to $\mathfrak{f}(\mathbb{X}^t)$ almost surely with respect to the randomness $\Theta, \widehat{\mathbb{X}}$, as $n\rightarrow+\infty$. Note that $\Xi$ defined above satisfies the assumptions in Lemma \ref{hyper}. Thus, by fixing a generic realization of $\mathbb{X}$, $\mathfrak{f}_n(\mathbb{X}^t)$ is uniformly bounded in $L^{2+\delta}(\mathbb{P})$ (with respect to the randomness $\Theta, \widehat{\mathbb{X}}$). Therefore, we have $ \mathbb{E}\big[ \mathfrak{f}_n( \mathbb{X}^t) \vert \mathbb{X} \big] \longrightarrow \mathbb{E}\big[ \mathfrak{f}( \mathbb{X}^t) \vert \mathbb{X} \big] $, as $n\rightarrow+\infty$. The above conditional expectations can be easily computed as follows: \begin{align*} \mathbb{E}\big[ \mathfrak{f}_n( \mathbb{X}^t) \vert \mathbb{X} \big]& = \mathbb{E}[F_n] + \sum_{p=1}^q \mathbb{E}\big[ Q_p\big(f_{p,n}; \mathbb{X}^t\big) \vert \mathbb{X} \big] = \mathbb{E}[F_n] + \sum_{p=1}^q \sum_{i_1, \ldots, i_p\in\mathbb{N}}f_{p,n}(i_1, \ldots, i_p) \mathbb{E}\big[ X^t_{i_1}\cdots X^t_{i_p} \vert \mathbb{X} \big] \\ &= \mathbb{E}[F_n] + \sum_{p=1}^q \sum_{i_1, \ldots, i_p\in\mathbb{N}}f_{p,n}(i_1, \ldots, i_p) e^{-pt} X_{i_1}\cdots X_{i_p} = \mathbb{E}[ F_n] + \sum_{p=1}^q e^{-pt} Q_p\big(f_{p,n}; \mathbb{X}\big) \,; \end{align*} in the same way, we have $ \mathbb{E}\big[ \mathfrak{f}( \mathbb{X}^t) \vert \mathbb{X} \big] = \mathbb{E}[ F] + \sum_{p=1}^q e^{-pt} Q_p\big(f_{p}; \mathbb{X}\big)$. Hence the desired result follows from the same ending argument as in previous section. \qedhere \end{proof} \subsection{The unbounded case}\label{ubddsec} Recall from the statement of Theorem \ref{HP-ubdd} that $\mathbb{X}$ is a sequence of centered independent random variables and the kernels $(f_{p,n}, n\geq 1)$ have their supports uniformly contained in $\{1, \ldots, d\}^p$, for each $p\in\{1, \ldots, q\}$. It follows that $F_n$ and $F$ only depend on the first $d$ coordinates $X_1, \ldots, X_d$. Now given any $m \geq d$, in what follows, we construct random variables $X^{(I)}_1, \ldots, X^{(I)}_m$ such that $\big(X^{(I)}_1, \ldots, X^{(I)}_m\big)$ is equal in law to $(X_1, \ldots, X_m)$: Let $\widehat{\mathbb{X}}$ be an independent copy of $\mathbb{X}$ and let $I:=I_m$ be a uniform random variable on $[m]:=\{1, \ldots, m\}$ such that $\mathbb{X}$, $\widehat{\mathbb{X}}$ and $I$ are independent, then for each $i, k\in[m]$, we define $ X_k^{(i)} = X_k \cdot \mathbbm{I} _{\{ i \neq k \}} + \widehat{X}_k \cdot \mathbbm{I} _{\{ i = k \}}$ and $X_k^{(I)} = X_k \cdot \mathbbm{I} _{\{ I \neq k \}} + \widehat{X}_k \cdot \mathbbm{I} _{\{ I = k \}}$. It is routine to verify that $\big(X^{(I)}_1, \ldots, X^{(I)}_m\big)$ and $(X_1, \ldots, X_m)$ are exchangeable pairs, which is well-known in the community of Stein's method, see for instance \cite{Stein86}. \begin{proof}[Proof of Theorem \ref{HP-ubdd}] Since $F_n, F$ only depend on the first $d$ coordinates, we can write $$Q_p\big(f_{p}; \{X_k, k\in[m] \} \big) = Q_p\big(f_{p}; \mathbb{X}\big) \quad\text{for any $m\geq d$.} $$ With this convention, we define $F^{(i)}$ to be the following sum $ \mathbb{E}[F]+ \sum_{p=1}^q Q_p\big(f_{p}; \{X^{(i)}_k, k\in[m] \} \big) $ and similarly we define $F_n^{(I)}$, $F^{(I)}$ and $F_n^{(i)}$ for each $n\geq 1$. It follows that $F_n^{(j)}$ converges almost surely to $F^{(j)}$ for each $j$, since $(X^{(j)}_k, k\in[m] )$ is clearly equal in law to $(X_k, k\in[m] )$. Now we fix $i\in[m]$, $F_n^{(i)}$ can be expressed as a sum of two parts: $F_n^{(i)} = \alpha_n(i) + \beta_n(i) \widehat{X}_i$, where $\alpha_n(i)$ and $\beta_n(i)$ are polynomials in $X_1, \ldots, X_{i-1}, X_{i+1}, \ldots, X_m$. In the same way, we can rewrite $F^{(i)} = \alpha(i) + \beta(i) \widehat{X}_i$, where $\alpha(i)$ and $\beta(i)$ are polynomials in $X_1, \ldots, X_{i-1}, X_{i+1}, \ldots, X_m$. (Note that $\beta_n(i) = \beta(i) =0$ for $i > d$.) Therefore, conditioning on $\{X_1, \ldots, X_m\}$, $ \alpha_n(i)\rightarrow \alpha(i) $ and $\beta_n(i)\rightarrow \beta(i)$, as $n\rightarrow+\infty$. Thus, $ \mathbb{E}\big[ F_n^{(i)} \vert \mathbb{X} \big] = \alpha_n(i)\rightarrow \alpha(i) = \mathbb{E}\big[ F^{(i)} \vert \mathbb{X} \big] $. Hence, almost surely \begin{align}\label{U1} \mathbb{E}\big[ F_n^{(I)} \vert \mathbb{X} \big] = \frac{1}{m} \sum_{i=1}^m \mathbb{E}\big[ F_n^{(i)} \vert \mathbb{X} \big] \rightarrow \frac{1}{m} \sum_{i=1}^m \mathbb{E}\big[ F^{(i)} \vert \mathbb{X} \big] = \mathbb{E}\big[ F^{(I)} \vert \mathbb{X} \big] \,. \end{align} Let us now compute the above conditional expectations: since \begin{align*} &\quad \mathbb{E}\Big[ Q_p\big(f_{p}; \{ X^{(I)}_k, k\in[m] \} \big) \big\vert \mathbb{X} \Big] - Q_p\big(f_{p}; \mathbb{X} \big) = \sum_{i=1}^m m^{-1} \mathbb{E}\Big[ Q_p\big(f_{p}; \{ X^{(i)}_k, k\in[m] \} \big) - Q_p\big(f_{p}; \mathbb{X} \big) \big\vert \mathbb{X} \Big] \\ & = - \sum_{i=1}^m m^{-1} \sum_{j_1, \ldots, j_p\leq d} f_p(j_1, \ldots, j_p) X_{j_1}\cdots X_{j_p} \cdot\mathbbm{I} _{( i\in\{ j_1, \ldots, j_p \} )} = - p m^{-1} Q_p\big(f_{p}; \mathbb{X} \big) \,, \end{align*} we have $ \mathbb{E}\Big[ Q_p \big(f_{p}; \{ X^{(I)}_k, k\in[m] \} \big) \vert \mathbb{X} \Big] = (1 - pm^{-1} ) Q_p(f_p; \mathbb{X} )$ so that \eqref{U1} implies \[ \mathbb{E}[F_n]+ \sum_{p=1}^q (1 - p m^{-1} ) Q_p\big(f_{p,n}; \mathbb{X} \big) \xrightarrow[a.s.]{n\rightarrow\infty} \mathbb{E}[F]+ \sum_{p=1}^q (1 - p m^{-1} ) Q_p\big(f_{p}; \mathbb{X} \big) . \] Hence the desired result follows from iterating the above process for many enough $m \geq d$. \end{proof}

d2ae5f1ad2949547a764069b0e8af5a7f0f435c9

\subsection{Copy-paste Data Augmentation with Random Placement is not Effective for Object Detection}\label{sec:random_copy-paste} In \cite{cut_paste}, data augmentation is performed by positioning segmented objects at random locations in new scenes. As mentioned previously, the strategy was shown to be effective for object \emph{instance} detection, as soon as an appropriate procedure is used for preventing the object detector to overfit blending artefacts---that is, the main difficulty is to prevent the detector to ``detect artefacts'' instead of detecting objects of interest. This is achieved by using various blending strategies to smooth object boundaries such as Poisson blending~\cite{perez2003poisson}, and by adding ``distractors'' - objects that do not belong to any of the dataset categories, but which are also synthetically pasted on random backgrounds. With distractors, artefacts occur both in positive and negative examples, for each of the categories, preventing the network to overfit them. According to~\cite{cut_paste}, this strategy can bring substantial improvements for the object instance detection/retrieval task, where modeling the fine-grain appearance of an object instance seems to be more important than modeling visual context as in the general category object detection task. Unfortunately, the augmentation strategy described above does not improve the results on the general object detection task and may even hurt the performance as we show in the experimental section. To justify the initial claim, we follow~\cite{cut_paste} as close as possible and conduct the following experiment on the PASCAL VOC12 dataset~\cite{pascal}. Using provided instance segmentation masks we extract objects from images and store them in a so-called instance-database. They are used to augment existing images in the training dataset by placing the instances at random locations. In order to reduce blending artifacts we use one of the following strategies: smoothing the edges using Gaussian or linear blur, applying Poisson blending~\cite{perez2003poisson} in the segmented region, blurring the whole image by simulating a slight camera motion or leaving the pasted object untouched. As distractors, we used objects from the COCO dataset~\cite{coco} belonging to categories not present in PASCAL VOC~\footnote{Note that external data from COCO was used only in this preliminary experiment and not in the experiments reported later in Section~\ref{sec:exp}.}. \begin{figure}[btp!] \begin{center} \includegraphics[width=0.99\linewidth,trim=90 115 230 30,clip]{images/context_inputs.pdf} \end{center} \caption{\textbf{Contextual images - examples of inputs to the context model}. A subimage bounded by a magenta box is used as an input to the context model after masking-out the object information inside a red box. The top row lists examples of positive samples encoding real objects surrounded by regular and predictable context. Positive training examples with ambiguous or uninformative context are given in the second row. The bottom row depicts negative examples enclosing background. This figure shows that contextual images could be ambiguous to classify correctly and the task of predicting the category given only the context is challenging.} \label{fig:context_input} \end{figure} For any combination of blending strategy, by using distractors or not, the naive data augmentation approach with random placement did not improve upon the baseline without data augmentation for the classical object detection task. A possible explanation may be that for instance object detection, the detector does not need to learn intra-class variability of object/scene representations and seems to concentrate only on appearance modeling of specific instances, which is not the case for category-level object detection. This experiment was the key motivation for proposing a context model, which we now present. \subsection{Explicit Context Modeling by CNN} The core idea behind the proposed method is that it is possible to some extent to guess the category of an object just by looking at its visual surroundings. That is precisely what we are modeling by a convolutional neural network, which takes contextual neighborhood of an object as input and is trained to predict the object's class. Here, we describe the training data and the learning procedure in more details. \textit{Contextual data generation.}\label{sec:context_inputs} In order to train the contextual model we use a dataset that comes with bounding box and object class annotations. Each ground-truth bounding box in the dataset is able to generate positive ``contextual images'' that are used as input to the system. As depicted in the Figure~\ref{fig:context_input}, a ``contextual image'' is a sub-image of an original training image, fully enclosing the selected bounding box, whose content is masked out. Such a contextual image only carries information about visual neighborhood that defines middle-range context and no explicit information about the deleted object. One box is able to generate multiple different context images, as illustrated in Figure~\ref{fig:contextual_image_variations}. Background ``contextual images'' are generated from bounding boxes that do not contain an object and are formally defined in~\cite{dvornik2018modeling}. To prevent distinguishing between positive and background images only by looking at the box shape and to force true visual context modeling, we estimate the shape distribution of positive boxes and sample the background ones from it. Precisely, we estimate the joint distribution of scale $s$ and aspect ratio $a$ with a two-dimensional histogram, as described in~\cite{dvornik2018modeling}, and we draw a pair $(s, a)$ from this distribution in order to construct a background box. Since in natural images there is more background boxes than the ones actually containing an object, we address the imbalance by sampling more background boxes, following sampling strategies in \cite{dvornik2018modeling,faster-rcnn}. \begin{figure}[t!] \begin{center} \includegraphics[width=1.09\linewidth,trim=110 150 65 10,clip]{images/contextual_imagte_variations.pdf} \end{center} \vspace{-0.5cm} \caption{\textbf{Different contextual images obtained from a single bounding box.} A single ground-truth bounding box (in blue) is able to generate a set of different context images (in green and orange) by varying the size of the initial box and the context neighborhood. While the orange contextual images may be recognized as a chair, the green ones make it more clear that the person was masked out. This motivates the need to evaluate several context images for one box during the context estimation phase.} \label{fig:contextual_image_variations} \end{figure} \textit{Model training.} Given the set of all contexts, gathered from all training data, we train a convolutional neural network to predict the presence of each object in the masked bounding box. The input to the network are the ``contextual images'' obtained during the data generation step. These contextual images are resized to $300 \times 300$ pixels, and the output of the network is a label in $\{0,1,...,C\}$, where $C$ is the number of object categories. The $0$-th class represents background and corresponds to a negative ``context image''. For such a multi-class image classification problem, we use the classical ResNet50 network~\cite{resnet} pre-trained on ImageNet, and change the last layer to be a softmax with $C+1$ activations (see experimental section for details). \begin{figure*}[btp!] \begin{center} \includegraphics[width=0.99\linewidth,trim=10 0 2 4,clip]{images/augmented_samples_full.pdf} \end{center} \caption{\textbf{Data augmentation for different types of annotations.} The first column contains samples from the training dataset with corresponding semantic/instance segmentation and bounding box annotations. Columns 2-4 present the result of applying context-driven augmentation to the initial sample with corresponding annotations.} \label{fig:augmented_samples} \end{figure*} \subsection{Context-driven Data Augmentation}\label{subsec:context_da} Once the context model is trained, we use it to provide locations where to paste objects. In this section, we elaborate on the context network inference and describe the precise procedure used for blending new objects into existing scenes. \textit{Selection of candidate locations for object placement.} A location for pasting an object is represented as a bounding box. For a single image, we sample 200 boxes at random from the shape distribution used in \ref{sec:context_inputs} and later select the successful placement candidates among them. These boxes are used to build corresponding contextual images, that we feed to the context model as input. As output, the model provides a set of scores in range between 0 and 1, representing the presence likelihood of each object category in a given bounding box, by considering its visual surrounding. The top scoring boxes are added to the final candidate set. Since the model takes into account not only the visual surroundings but a box's geometry too, we need to consider all possible boxes inside an image to maximize the recall. However this is too costly and using 200 candidates was found to provide good enough bounding boxes among the top scoring ones.\\ After analyzing the context model's output we made the following observation: if an object of category $c$ is present in an image it is a confident signal for the model to place another object of this class nearby. The model ignores this signal only if no box of appropriate shape was sampled in the object's neighborhood. This often happens when only 200 candidate locations are sampled; however, evaluating more locations would introduce a computational overhead. To fix this issue, we propose a simple heuristic, which consists of drawing boxes in the neighborhood of this object and adding them to the final candidate set. The added boxes have the same geometry (up to slight distortions) as the neighboring object's box. \textit{Candidate scoring process.} As noted before, we use the context model to score the boxes by using its softmax output. Since the process of generating a contextual image is not deterministic, predictions on two contextual images corresponding to the same box may differ substantially, as illustrated in Figure~\ref{fig:contextual_image_variations}. We alleviate this effect by sampling 3 contextual images for one location and average the predicted scores. After the estimation stage we retain the boxes where an object category has score greater than $0.7$; These boxes together with the candidates added at the previous step form the final candidate set that will be used for object placement. \textit{Blending objects in their environment.} Whenever a bounding box is selected by the previous procedure, we need to blend an object at the corresponding location. This step follows closely the findings of~\cite{cut_paste}. We consider different types of blending techniques (Gaussian or linear blur, simple copy-pasting with no post-processing, or generating blur on the whole image to imitate motion), and randomly choose one of them in order to introduce a larger diversity of blending artefacts. Figure~\ref{fig:blendings} presents the blending techniques mentioned above. We also do not consider Poisson blending in our approach, which was considerably slowing down the data generation procedure. Unlike~\cite{cut_paste} and unlike our preliminary experiment described in Section~\ref{sec:random_copy-paste}, we do not use distractors, which were found to be less important for our task than in~\cite{cut_paste}. As a consequence, we do not need to exploit external data to perform data augmentation. \textit{Updating image annotation.} Once an image is augmented by blending in a new object, we need to modify the annotation accordingly. In this work, we consider data augmention for both object detection and semantic segmentation, as illustrated in Figure~\ref{fig:augmented_samples}. Once a new object is placed in the scene, we generate a bounding box for object detection by drawing the tightest box around that object. In case where an initial object is too occluded by the blended one, i.e. the IoU between their boxes is higher than 0.8, we delete the bounding box of the original object from the annotations. For semantic segmentation, we start by considering augmentation on instance masks (Figure~\ref{fig:augmented_samples}, column 4) and then convert them to semantic masks (Figure~\ref{fig:augmented_samples}, column 3). If a new instance occludes more than $80\%$ of an object already present in the scene, we discard annotations for all pixels belonging to the latter instance. To obtain semantic segmentation masks from instance segmentations, each instance pixel is labeled with the corresponding objects class. \subsection{Dataset, Tools, and Metrics} \label{sec:datasets} \textit{Datasets.} In our experiments, we use the Pascal VOC'12~\cite{pascal} and COCO~\cite{coco} datasets. In the VOC'12 dataset, we only consider a subset that contains segmentation annotations. The training set contains $1\,464$ images and is dubbed \texttt{VOC12train-seg} later in the paper. Following standard practice, we use the test set of VOC'07 to evaluate the detection performance, which contains $4\,952$ images with the same 20 object categories as VOC'12. We call this image set \texttt{VOC07-test}. When evaluating segmentation performance, we use the validation set of the VOC'12 annotated with segmentation masks \texttt{VOC12val-seg} that contains $1\,449$ images.\\ The COCO dataset \cite{coco} is used for large-scale object detection experiments. It includes 80 object categories for detection and instance segmentation. For both tasks, there are 118K images for training that we denote as \texttt{COCO-train2017} and 5K for validation and testing denoted as \texttt{COCO-val2017}. \textit{Models.} To test our data-augmentation strategy we chose a single model capable of performing both object detection and semantic segmentation. BlitzNet \cite{blitznet} is an encoder-decoder architecture, which is able to solve either of the tasks, or both simultaneously if trained with box and segmentation annotations together. The open-source implementation is available online. If used to solve the detection task, BlitzNet achieves close to the state-of-the-art results ($79.1\%$ mAP) on \texttt{VOC07-test} when trained on the union of the full training and validation parts of VOC'07 and VOC'12, namely \texttt{VOC07-train+val} and \texttt{VOC12train+val} (see~\cite{blitznet}); this network is similar to the DSSD detector of \cite{dssd} that was also used in the Focal Loss paper \cite{focal_loss}. When used as a segmentor, BlitzNet resembles the classical U-Net architecture~\cite{ronneberger2015u} and also achieves results comparable to the state-of-the-art on VOC'12-test set ($75.5\%$ mIoU). The advantage of such class of models is that it is relatively fast (it may work in real time) and supports training with big batches of images without further modification.\\ To make the evaluation extensive, we also consider a different region-based class of detectors. For that purpose we employ an open-source implementation of Faster-RCNN \cite{jjfaster2rcnn} which uses ResNet50~\cite{resnet} architecture as a feature extractor. Finally, when tackling object detection and instance segmentation on COCO, we use Mask-RCNN~\cite{mask_rcnn} that solves both tasks simultaneously. For each region proposal the network outputs estimated class probabilities, regressed box offsets and a predicted instance mask. We run the official implementation of~\cite{Detectron2018} that uses ResNet50 as a backbone, followed by an FPN~\cite{lin2016feature} module. This setup corresponds to the current state of the art in object detection and instance segmentation. \textit{Evaluation metric.} In VOC'07, a bounding box is considered to be correct if its Intersection over Union (IoU) with a ground truth box is higher than 0.5. The metric for evaluating the quality of object detection and instance segmentation for one object class is the average precision (AP). Mean Average Precision (mAP) is used to report the overall performance on the dataset. Mean Intersection Over Union (mIoU) is used to measure performance on semantic segmentation. \subsection{Implementation Details}\label{sec:details} \textit{Training the context model.} After preparing the ``contextual images'' as described in \ref{sec:context_inputs}, we re-scale them to the standard size $300 \times 300$ and stack them in batches of size 32. We use ResNet50~\cite{resnet} with ImageNet initialization to train a contextual model in all our experiments. Since we have access only to the training set at any stage of the pipeline we define two strategies for training the context model. When the amount of positive samples is scarce, we train and apply the model on the same data. To prevent overfitting, we use early stopping. In order to determine when to stop the training procedure, we monitor both training error on our training set and validation error on the validation set. The moment when the loss curves start diverging noticeably is used as a stopping point. We call this training setting ``small-data regime''. When the size of the training set is moderate and we are in ``normal-data regime'', we split it in two parts ensuring that for each class, there is a similar number of positive examples in both splits. The context model is trained on one split and applied to another one. We train the model with ADAM optimizer \cite{adam} starting with learning rate $10^{-4}$ and decreasing it by the factor of 10 once during the learning phase. The number of steps depends on a dataset. We sample 3 times more background contextual images, as noted in Section~\ref{sec:context_inputs}. Visual examples of augmented images produced when using the context model are presented in Figure~\ref{fig:context_output}. Overall, training the context model is about 4-5 times faster than training the detector. \textit{Training detection and segmentation models.} In this work, the BlitzNet model takes images of size $300 \times 300$ as an input and produces a task-specific output. When used as a detector, the output is a set of candidate object boxes with classification scores and in case of segmentation it is an estimated semantic map of size $75\times75$; like our context model, it uses ResNet50 \cite{resnet} pre-trained on ImageNet as a backbone. The models are trained by following~\cite{blitznet}, with the ADAM optimizer \cite{adam} starting from learning rate $10^{-4}$ and decreasing it later during training by a factor 10 (see Sections~\ref{sec:det_voc}~and~\ref{sec:segmentation} for number of epochs used in each experiment). In addition to our data augmentation approach obtained by copy-pasting objects, all experiments also include classical data augmentation steps obtained by random-cropping, flips, and color transformations, following~\cite{blitznet}. For the Faster-RCNN detector training, we consider the classical model of \cite{faster-rcnn} with ResNet50 backbone and closely follow the instructions of~\cite{jjfaster2rcnn}. On the Pascal VOC12 dataset, training images are rescaled to have both sides between 600 and 1000 pixels before being passed to the network. The model is trained with the Momentum optimizer for 9 epochs in total. The starting learning rate is set to $10^{-2}$ and divided by 10 after first 8 epochs of training. When using Mask-RCNN~\cite{mask_rcnn}, the images are rescaled to have a maximum size of 1333 pixel on one side or a minimum one of 800 pixels. \red{Following the original implementation~\cite{Detectron2018}, training and evaluation is performed on 2 GPUs, where images are grouped in batches of size 16}. We set the starting learning rate to $2 \cdot 10^{-2}$ which is decreased by a factor of 10 twice later during training. For both Faster-RCNN and Mask-RCNN standard data augmentation includes only horizontal flipping. \textit{Selecting and blending objects.} Since we widely use object instances extracted from the training images in all our experiments, we create a database of objects cut out from the \texttt{VOC12train-seg} or \texttt{COCO-train} sets to quickly access them during training. For a given candidate box, an instance is considered as matching if after scaling it by a factor in $[0.5, 1.5]$ the re-scaled instance's bounding box fits inside the candidate's one and takes at least 80\% of its area. The scaling factor is kept close to 1 not to introduce scaling artefacts. When blending the objects into the new background, we follow~\cite{cut_paste} and use randomly one of the following methods: adding Gaussian or linear blur on the object boundaries, generating blur on the whole image by imitating motion, or just paste an image with no blending. By introducing new instances in a scene we may also introduce heavy occlusions of existing objects. The strategy for resolving this issue depends on the task and is clarified in Sections \ref{sec:det_voc} and \ref{sec:segmentation}. \subsection{Why is Random Placement not Working?}\label{sec:prelim} As we discovered in the Section~\ref{sec:random_copy-paste}, random copy-paste data augmentation does not bring improvement when used to augment object detection datasets. There are multiple possible reasons for observing this behavior, such as violation of context constraints imposed by the dataset, objects looking ``out of the scene'' due to different illumination conditions or simply artifacts introduced due to blending techniques. To investigate this phenomenon, we conduct a study, that aims to better understand (i) the importance of visual context for object detection, (ii) the role of illumination conditions and (iii) the impact of blending artefacts. For simplicity, we choose the first 5 categories of VOC'12, namely \textit{aeroplane, bike, bird, boat, bottle}, and train independent detectors per category. \textit{Baseline when no object is in context.} To confirm the negative influence of random placing, we consider one-category detection, where only objects of one selected class are annotated with bounding boxes and everything else is considered as background. Images that do not contain objects of the selected category become background images. After training 5 independent detectors as a baseline, we construct a similar experiment by learning on the same number of instances, but considering as positive examples only objects that have been synthetically placed in a random context. This is achieved by removing from the training data all the images that have an object from the category we want to model, and replacing it by an instance of this object placed on a background image. The main motivation for such study is to consider the extreme case where (i) no object is placed in the right context; (iii) all objects may suffer from rendering artefacts. As shown in Table~\ref{tab:preliminary}, the average precision degrades significantly by about $14\%$ compared to the baseline. As a conclusion, either visual context is indeed crucial for learning, or blending artefacts is also a critical issue. The purpose of the next experiment is to clarify this ambiguity. \begin{table}[btp!] \centering \renewcommand{\arraystretch}{1.4} \renewcommand{\tabcolsep}{0.5mm} \begin{tabular}{l| c c c c c |c} Method & aero & bike & bird & boat & bottle & average \\ \hline Base-DA & 58.8 & 64.3 & 48.8 & 47.8 & 33.9 & 48.7\\ Random-DA & 60.2 & 66.5 & 55.1 & 41.9 & 29.7 & 48.3\\ \hline Removing context & 44.0 & 46.8 & 42.0 & 20.9 & 15.5 & 33.9\\ \hline Enlarge + Reblend-DA & 60.1 & 63.4 & 51.6 & 48.0 & 34.8 & 51.6\\ \end{tabular} \caption{Ablation study on the first five categories of VOC'12. All models are learned independently. We compare classical data augmentation techniques (Base-DA), approaches obtained by copy-pasting objects, either randomly (Random-DA) or by preserving context (Enlarge+Reblend-DA). The line ``Removing context'' corresponds to the first experiment described in Section~\ref{sec:prelim}; Enlarge-Reblend corresponds to the second experiment.} \label{tab:preliminary} \end{table} \begin{figure*}[hbtp!] \begin{center} \includegraphics[width=0.99\linewidth,trim=20 70 30 0,clip]{images/context_outputs.pdf} \end{center} \caption{\textbf{Examples of instance placement with context model guidance.} The figure presents samples obtained by placing a matched examples into the box predicted by the context model. The top row shows generated images that are visually almost indistinguishable from the real ones. The middle row presents samples of good quality although with some visual artifacts. For the two leftmost examples, the context module proposed an appropriate object class, but the pasted instances do not look visually appealing. Sometimes, the scene does not look natural because of the segmentation artifacts as in the two middle images. The two rightmost examples show examples where the category seems to be in the right environment, but not perfectly placed. The bottom row presents some failure cases.} \label{fig:context_output} \end{figure*} \begin{figure*}[hbtp!] \begin{center} \includegraphics[width=0.99\linewidth,trim=60 110 120 60,clip]{images/enlarge.pdf} \end{center} \caption{\textbf{Illustration of artifacts arising from enlargement augmentation.} In the enlargement data augmentation, an instance is cut out of the image, up-scaled by a small factor and placed back at the same location. This approach leads to blending artefacts. Modified images are given in the top row. Zoomed parts of the images centered on blending artifacts are presented in the bottom line.} \label{fig:enlargement} \end{figure*} \textit{Impact of blending when the context is right.} In the previous experiment, we have shown that the lack of visual context and the presence of blending artefacts may explain the performance drop observed in the third row of Table~\ref{tab:preliminary}. Here, we propose a simple experiment showing that neither (iii) blending artefacts nor (ii) illumination difference are critical when objects are placed in the right context: the experiment consists of extracting each object instance from the dataset, up-scale it by a random factor slightly greater than one (in the interval $[1.2,1.5]$), and blend it back at the same location, such that it covers the original instance. To mimic the illumination change we apply a slight color transformation to the segmented object. As a result, the new dataset benefits slightly from data augmentation (thanks to object enlargement), but it also suffers from blending artefacts for \emph{all object instances}. As shown on the forth row of Table~\ref{tab:preliminary}, this approach improves over the baseline, which suggests that the lack of visual context is probably the key explaining the result observed before. The experiment also confirms that the presence of difference in illumination and blending artefacts is not critical for the object detection task. Visual examples of such artefacts are presented in Figure~\ref{fig:enlargement}. \begin{table*}[hbtp!] \centering \renewcommand{\arraystretch}{1.3} \renewcommand{\tabcolsep}{0.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l|c c c c c c c c c c c c c c c c c c c c |c|} method & aero & bike & bird & boat & bott. & bus & car & cat & chair & cow & table & dog & horse & mbike & pers. & plant & sheep & sofa & train & tv & avg.\\ \hline Base-DA & 58.8 & 64.3 & 48.8 & 47.8 & 33.9 & 66.5 & 69.7 & 68.0 & 40.4 & 59.0 & 61.0 & 56.2 & 72.1 & 64.2 & 66.7 & 36.6 & 54.5 & 53.0 & 73.4 & 63.6 & 58.0\\ Random-DA & 60.2 & 66.5 & 55.1 & 41.9 & 29.7 & 66.5 & 70.0 & 70.1 & 37.4 & 57.4 & 45.3 & 56.7 & 68.3 & 66.1 & 67.0 & 37.0 & 49.9 & 55.8 & 72.1 & 62.6 & 56.9\\ Enlarge-DA & 60.1 & 63.4 & 51.6 & 48.0 & 34.8 & 68.8 & 72.1 & 70.4 & 41.1 & 63.7 & 62.3 & 56.3 & 70.1 & 67.8 & 65.3 & 37.9 & 58.1 & 61.2 & 75.5 & 65.9 & 59.7\\ Context-DA & 68.9 & 73.1 & 62.5 & 57.6 & 38.9 & 72.5 & 74.8 & 77.2 & 42.9 & 69.7 & 59.5 & 63.9 & 76.1 & 70.2 & 69.2 & 43.9 & 58.3 & 59.7 & 77.2 & 64.8 & 64.0\\ \hline Impr. Cont. & \bf{10.1} & \bf{8.7} & \bf{13.7} & \bf{9.2} & 5.0 & 6.0 & 5.1 & \bf{9.2} & 2.5 & \bf{10.7} & 1.5 & \bf{7.5} & 4.0 & 6.0 & 2.5 & \bf{7.3} & 3.8 & 6.7 & 4.2 & 1.2 & 5.8 \\ \end{tabular} } \caption{Comparison of detection accuracy on \texttt{VOC07-test} for the single-category experiment. The models are trained independently for each category, by using the $1\,464$ images from \texttt{VOC12train-seg}. The first row represents the baseline experiment that uses standard data augmentation techniques. The second row uses in addition copy-pasting of objects with random placements. ``Enlarge-DA'' augmentation blends up-scaled instances back in their initial location, which is given in row 3. The forth row presents the results achieved by our context-driven approach and the last row presents the improvement it brings over the baseline. The numbers represent average precision per class in \%. Large improvements over the baseline (greater than $7\%$) are in bold. All numbers are averaged over 3 independent experiments. } \label{tab:single_cat} \end{table*} \begin{table*}[hbtp!] \centering \renewcommand{\arraystretch}{1.3} \renewcommand{\tabcolsep}{0.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l| c |c c c c c c c c c c c c c c c c c c c c | c |} model & CDA & aero & bike & bird & boat & bott. & bus & car & cat & chair & cow & table & dog & horse & mbike & pers. & plant & sheep & sofa & train & tv & avg.\\ \hline \multirow{2}{*}{BlitzNet300} & & 63.6 & 73.3 & 63.2 & 57.0 & 31.5 & 76.0 & 71.5 & 79.9 & 40.0 & 71.6 & 61.4 & 74.6 & 80.9 & 70.4 & 67.9 & 36.5 & 64.9 & 63.0 & 79.3 & 64.7 & 64.6\\ & \checkmark & 69.9 & 73.8 & 63.9 & 62.6 & 35.3 & 78.3 & 73.5 & 80.6 & 42.8 & 73.8 & 62.7 & 74.5 & 81.1 & 73.2 & 68.9 & 38.1 & 67.8 & 64.3 & 79.3 & 66.1 & \bf{66.5}\\ \hline \multirow{2}{*}{F-RCNN} & & 65.8 & 70.9 & 66.5 & 54.6 & 45.9 & 72.7 & 72.9 & 80.3 & 36.8 & 70.3 & 48.0 & 78.9 & 70.7 & 70.6 & 66.3 & 33.1 & 64.7 & 59.8 & 71.8 & 61.1 & 63.1\\ & \checkmark & 67.4 & 67.7 & 64.9 & 58.0 & 50.4 & 71.6 & 74.9 & 80.4 & 36.8 & 70.2 & 56.4 & 75.7 & 73.7 & 71.6 & 71.5 & 39.4 & 68.6 & 63.5 & 67.7 & 60.1 & \bf{64.5} \\ \end{tabular} } \caption{Comparison of detection accuracy on \texttt{VOC07-test} for the multiple-category experiment. The model is trained on all categories at the same time, by using the $1\,464$ images from \texttt{VOC12train-seg}. The first column specifies the detector used in the experiment, the second column notes if Context-driven Data Augmentation (CDA) was used. The numbers represent average precision per class in \%.} \label{tab:multiple} \end{table*} \subsection{Object Detection Augmentation on VOC PASCAL}\label{sec:det_voc} In this subsection, we are conducting experiments on object detection by augmenting the PASCAL VOC'12 dataset. In order to measure the impact of the proposed technique in a ``small data regime'', we pick the single-category detection scenario and also consider a more standard multi-category setting. We test single-shot region-based families of detectors---with BlitzNet and Faster-RCNN respectively---and observe improved performance in both cases. \subsubsection{Single-category Object Detection}\label{sec:single} In this section, we conduct an experiment to better understand the effect of the proposed data augmentation approach, dubbed ``Context-DA'' in the different tables, when compared to a baseline with random object placement ``Random-DA'', and when compared to standard data augmentation techniques called ``Base-DA''. The study is conducted in a single-category setting, where detectors are trained independently for each object category, resulting in a relatively small number of positive training examples per class. This allows us to evaluate the importance of context when few labeled samples are available and see if conclusions drawn for a category easily generalize to other ones. The baseline with random object placements on random backgrounds is conducted in a similar fashion as our context-driven approach, by following the strategy described in the previous section. For each category, we treat all images with no object from this category as background images, and consider a collection of cut instances as discussed in Section~\ref{sec:datasets}. During training, we augment a negative (background) image with probability 0.5 by pasting up to two instances on it, either at randomly selected locations (Random-DA), or using our context model in the selected bounding boxes with top scores (Context-DA). The instances are re-scaled by a random factor in $[0.5, 1.5]$ and blended into an image using a randomly selected blending method mentioned in Section~\ref{sec:datasets}. For all models, we train the object detection network for 6K iterations and decrease the learning rate after 2K and 4K iterations by a factor 10 each time. The context model was trained in ``small-data regime'' for 2K iterations and the learning rate was dropped once after 1.5K steps. The results for this experiment are presented in Table~\ref{tab:single_cat}. The conclusions are the following: random placement indeed hurts the performance on average. Only the category bird seems to benefit significantly from it, perhaps because birds tend to appear in various contexts in this dataset and some categories significantly suffer from random placement such as boat, table, and sheep. Importantly, the visual context model always improves upon the random placement one, on average by 7\%, and upon the baseline that uses only classical data augmentation, on average by 6\%. Interestingly, we identify categories for which visual context is crucial (aeroplane, bird, boat, bus, cat, cow, dog, plant), for which context-driven data augmentation brings more than 7\% improvement and some categories that display no significant gain or losses (chair, table, persons, tv), where the difference with the baseline is less noticeable (around 1-3\%). \begin{table*}[hbtp!] \centering \renewcommand{\arraystretch}{1.3} \renewcommand{\tabcolsep}{0.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l|c c c c c c c c c c c c c c c c c c c c | c |} method & aero & bike & bird & boat & bott. & bus & car & cat & chair & cow & table & dog & horse & mbike & pers. & plant & sheep & sofa & train & tv & avg.\\ \hline Base-DA & 79.0 & 43.7 & 65.8 & 57.9 & 53.8 & 83.8 & 77.9 & 76.7 & 19.2 & 56.6 & 46.6 & 67.6 & 59.0 & 73.1 & 77.9 & 46.8 & 69.4 & 37.8 & 73.7 & 70.3 & 63.3\\ Random-DA & 78.1 & 47.1 & 75.4 & 57.8 & 57.2 & 83.5 & 76.2 & 76.6 & 20.5 & 57.0 & 43.1 & 69.2 & 57.5 & 71.5 & 78.2 & 40.0 & 63.3 & 42.0 & 74.5 & 64.1 & 63.1 \\ Enlarge-DA & 77.2 & 45.4 & 67.9 & 57.9 & 61.0 & 84.1 & 78.8 & 76.3 & 20.3 & 58.4 & 46.9 & 67.5 & 60.5 & 73.9 & 78.1 & 45.2 & 71.1 & 38.8 & 73.6 & 71.1 & 64.1 \\ Context-DA & 81.7 & 46.4 & 73.4 & 60.7 & 59.4 & 85.3 & 78.8 & 79.1 & 20.6 & 60.0 & 48.0 & 68.1 & 62.2 & 75.3 & 78.8 & 47.6 & 71.6 & 39.9 & 73.6 & 70.3 & 65.4\\ \hline Impr. Cont. & \textbf{2.7} & \textbf{2.7} & \textbf{7.6} & \textbf{2.8} & \textbf{4.6} & 1.5 & 1.1 & 2.3 & 1.4 & \textbf{3.4} & 1.4 & 0.5 & \textbf{3.2} & 2.3 & 2.2 & 0.9 & 0.8 & 2.1 & -0.1 & 0 & 2.1 \\ \end{tabular} } \vspace*{0.1cm} \caption{Comparison of segmentation accuracy on \texttt{VOC12val-seg}. The model is trained on all 20 categories by using the $1\,464$ images from \texttt{VOC12train-seg}. Base-DA represents the baseline experiment that uses standard data augmentation techniques. Context-DA uses also our context-driven data augmentation. Random-DA is its context-agnostic analogue. Enlarge-DA corresponds to randomly enlarging an instance and blending it back. The last row presents absolute improvement over the baseline. The numbers represent IoU per class in \%. Categories enjoying an improvement higher than $2.5\%$ are in bold. All numbers are averaged over 3 independent experiments. } \label{tab:seg} \end{table*} \begin{table}[hbtp!] \centering \renewcommand{\arraystretch}{1.2} \renewcommand{\tabcolsep}{1.0mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l | c | c c c | c c c} { Model} & DA &{@0.5:0.95} & {@0.5} & @0.75 & S & M & L \\ \hline \multicolumn{8}{c}{Object Detection} \\ \hline BlitzNet300 & & 27.3 & 46.0 & 28.1 & 10.7 & 26.8 & 46.0 \\ BlitzNet300 & Rnd & 26.8 & 45.0 & 27.6 & 9.3 & 26.0 & 45.7 \\ BlitzNet300 & Cont & \bf{28.0} & \bf{46.7} & \bf{28.9} & 10.7 & \bf{27.8} & \bf{47.0} \\ \hline \red{Mask-RCNN} & & 38.6 & 59.7 & 42.0 & 22.1 & 41.5 & 50.6 \\ \red{Mask-RCNN} & Rnd & 36.9 & 57.3 & 39.7 & 20.5 & 39.6 & 48.0 \\ \red{Mask-RCNN} & Cont & \bf{39.1} & \bf{60.3} & \bf{42.3} & \bf{22.4} & \bf{42.2} & \bf{51.2} \\ \hline \multicolumn{8}{c}{Instance Segmentation} \\ \hline \red{Mask-RCNN} & & 34.5 & 56.5 & 36.3 & 15.7 & 37.1 & 52.1 \\ \red{Mask-RCNN} & Rnd & 33.6 & 55.2 & 35.8 & 14.8 & 35.5 & 50.0 \\ \red{Mask-RCNN} & Cont & \bf{34.8} & \bf{57.0} & \bf{36.5} & \bf{15.9} & \bf{37.6} & \bf{52.5} \\ \end{tabular} } \caption{Comparison of object detection and instance segmentation accuracy on \texttt{COCO-val2017} for the multiple-category experiment. The model is trained on all categories at the same time, by using the $118\,783$ images from \texttt{COCO-train2017}. The first column specifies a model used to solve a task, the second column notes if Context-driven (Cont) or random-placement (Rnd) Data Augmentation was used. For different IoU thresholds @0.5:0.95, @0.5 and @0.75) and for different object size (S, M, L), the numbers represent mAP in \%. Best results are in bold. } \label{tab:coco} \end{table} \subsubsection{Multiple-Categories Object Detection}\label{sec:multiple} In this section, we conduct the same experiment as in Section~\ref{sec:single}, but we train a single multiple-category object detector instead of independent ones per category. Network parameters are trained with more labeled data (on average 20 times more than for models learned in Table~\ref{tab:single_cat}). When training the context model, we follow the ``normal-data strategy'' described in Section \ref{sec:details} and train the model for 8K iterations, decreasing the learning rate after 6K steps. The results are presented in Table~\ref{tab:multiple} and show a modest average improvement of $2.1\%$ for a single shot and $1.4\%$ for a region-based detector on average over the corresponding baselines, which is relatively consistent across categories. This confirms that data augmentation is crucial when few labeled examples are available. \begin{table*}[!t] \centering \renewcommand{\arraystretch}{1.3} \renewcommand{\tabcolsep}{0.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l|c c c c c c c c c c c c c c c c c c c c |c|} Aug. type & aero & bike & bird & boat & bottle & bus & car & cat & chair & cow & table & dog & horse & mbike & pers. & plant & sheep & sofa & train & tv & avg.\\ \hline Inst seg & \bf{68.9} & \bf{73.1} & \bf{62.5} & \bf{57.6} & \bf{38.9} & \bf{72.5} & \bf{74.8} & \bf{77.2} & \bf{42.9} & \bf{69.7} & 59.5 & \bf{63.9} & 76.1 & 70.2 & \bf{69.2} & \bf{43.9} & 58.3 & \bf{59.7} & 77.2 & 64.8 & \bf{64.0} \\ Gt seg & 67.8 & 70.3 & 61.5 & 56.6 & 38.2 & 71.2 & 74.7 & 75.7 & 41.6 & 68.3 & 59.0 & 63.2 & 75.6 & 71.0 & 68.7 & 42.6 & 59.5 & 59.1 & \bf{78.4} & \bf{65.3} & 63.4 \\ Weak seg & 68.9 & 71.3 & 59.0 & 54.2 & 37.3 & 71.9 & 74.5 & 75.2 & 40.8 & 67.6 & \bf{59.8} & 62.8 & \bf{76.4} & \bf{71.3} & 68.4 & 43.8 & \bf{59.9} & 57.2 & 76.6 & 64.4 & 63.0 \\ \hline No & 58.8 & 64.3 & 48.8 & 47.8 & 33.9 & 66.5 & 69.7 & 68.0 & 40.4 & 59.0 & 61.0 & 56.2 & 72.1 & 64.2 & 66.7 & 36.6 & 54.5 & 53.0 & 73.4 & 63.6 & 58.0\\ \end{tabular} } \caption{Comparison of detection accuracy on \texttt{VOC07-test} for the single-category experiment. The models are trained independently on each category, by using the \texttt{VOC12train-seg}. The first column specifies the type of object mask used for augmentation: ground-truth instance segmentations (Inst. Seg.), ground-truth semantic segmentation (GT Seg.), or weakly-supervised semantic segmentations (Weak Seg.). Inst. seg. stands for the original instance segmentation ground truth masks. The numbers represent AP per class in \%. The best result for a category is in bold. All numbers are averaged over 3 independent experiments. } \label{tab:single_weak} \end{table*} \begin{table*}[!t] \centering \renewcommand{\arraystretch}{1.3} \renewcommand{\tabcolsep}{0.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l|c c c c c c c c c c c c c c c c c c c c |c|} Aug. type & aero & bike & bird & boat & bott. & bus & car & cat & chair & cow & table & dog & horse & mbike & pers. & plant & sheep & sofa & train & tv & avg.\\ \hline Inst. seg & \bf{69.9} & 73.8 & \bf{63.9} & \bf{62.6} & 35.3 & \bf{78.3} & 73.5 & 80.6 & \bf{42.8} & \bf{73.8} & 62.7 & \bf{74.5} & 81.1 & 73.2 & 68.9 & 38.1 & \bf{67.8} & 64.3 & 79.3 & \bf{66.1} & \bf{66.5}\\ GT Seg. & 68.7 & 74.5 & 60.1 & 60.0 & 34.9 & 75.4 & \bf{74.4} & \bf{81.7} & 41.1 & 72.4 & \bf{64.2} & 74.4 & \bf{81.3} & \bf{74.6} & \bf{69.6} & \bf{39.7} & 67.6 & 64.2 & \bf{80.4} & 65.5 & 66.2\\ Weak Seg. & 69.2 & \bf{75.2} & 63.2 & 59.8 & \bf{35.6} & 77.1 & 73.4 & 78.7 & 41.3 & 72.9 & 62.8 & 72.7 & 79.6 & 72.5 & 68.1 & 39.2 & 67.6 & \bf{66.1} & 79.5 & 64.2 & 65.9\\ \hline No & 63.6 & 73.3 & 63.2 & 57.0 & 31.5 & 76.0 & 71.5 & 79.9 & 40.0 & 71.6 & 61.4 & 74.6 & 80.9 & 70.4 & 67.9 & 36.5 & 64.9 & 63.0 & 79.3 & 64.7 & 64.6\\ \end{tabular} } \caption{Comparison of detection accuracy on \texttt{VOC07-test} for the multy-category experiment depending on the type of object masks used for augmentation. The models are trained on all categories together, by using the $1\,464$ images from \texttt{VOC12train-seg}. The first column specifies the type of object mask used for augmentation: ground-truth instance segmentations (Inst. Seg.), ground-truth semantic segmentation (GT Seg.), or weakly-supervised semantic segmentations (Weak Seg.). Inst. seg. stands for the original instance segmentation ground truth masks. The numbers represent AP per class in \%. The best result for a category is in bold. All numbers are averaged over 3 independent experiments. } \label{tab:multy_weak} \end{table*} \subsection{Object Detection and Instance Segmentation Augmentation on COCO}\label{exp:coco} In order to test our augmentation technique at large scale, we use in this section the COCO dataset~\cite{coco} whose training set size is by 2 orders of magnitude larger than \texttt{voc12train-seg}, and consider both object detection and instance segmentation tasks. \subsubsection{Object Detection with BlitzNet}\label{exp:coco_det} By design, the experiment is identical to the one presented in Section~\ref{sec:multiple}. However, for the COCO dataset we need to train a new context model. This is done by training for 350K iterations (decay at 250K) as described in Section~\ref{sec:details}. The non data-augmented baseline was trained according to \cite{blitznet}; when using our augmentation pipeline, we train the detector for 700K iterations and decrease the learning rate by a factor of 10 after 500K and 600K iterations. Table~\ref{tab:coco} shows that we are able to achieve a modest improvement of $0.7\%$, and that data augmentation still works and does not degrade the performance regardless the large amount of data available for training initially. \subsubsection{Detection and Segmentation with Mask-RCNN}\label{exp:coco_seg} \red{For this experiment, we use Mask-RCNN~\cite{mask_rcnn} that jointly solves object detection and instance segmentation. When training the baseline model, we closely follow original guidelines\footnote{\url{https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md}} and train the model with 2x schedule (for 180K iterations) to maximize the baseline's performance. Training the model with 1x schedule (for 90K iterations) results in underfitting, while training with x4 schedule (for 360K iterations), results in overfitting. In order to improve the performance of Mask-RCNN for both tasks, we train the model with x4 schedule and use the context-driven data augmentation. In order to reduce pasting artifacts negatively affecting Mask-RCNN, we decrease the augmentation probability during the training. More precisely, augmentation probability is set to 0.5 in the beginning of the training and then linearly decreased to 0 by the end of the training procedure. Training with constant augmentation probability did not improve the performance over the x2 baseline. On the other hand, gradually reducing augmentation probability results in less aggressive regularization and brings more benefits when training on a large dataset, such as COCO. As Table~\ref{tab:coco} shows, following this augmentation strategy results in a $0.5\%$ an $0.3\%$ mAP improvement for detection and segmentation respectively, when comparing to the most accurate baseline Mask-RCNN, trained with x2 schedule. Augmenting the training data with random placement strategy hurts the performance substantially, which highlights the importance of context for data augmentation.} \subsection{Semantic Segmentation Augmentation}\label{sec:segmentation} In this section, we demonstrate the benefits of the proposed data augmentation technique for the task of semantic segmentation by using the VOC'12 dataset. First, we set up the baseline by training the BlitzNet300~\cite{blitznet} architecture for semantic segmentation. Standard augmentation techniques such as flipping, cropping, color transformations and adding random noise were applied during the training, as described in the original paper. We use \texttt{voc12train-seg} subset for learning the model parameters. Following the training procedure described in Section~\ref{sec:details}, we train the model for 12K iterations starting from the learning rate of $10^{-4}$ and decreasing it twice by the factor of 10, after 7K and 10K steps respectively. Next, we perform data augmentation of the training set with the proposed context-driven strategy and train the same model for 15K iterations, dropping the learning rate at 8K and 12K steps. In order to blend new objects in and to augment the ground truth we follow routines described in Section~\ref{subsec:context_da}. We also carry out an experiment where new instances are placed at random locations, which represents a context-agnostic counterpart of our method. We summarize the results of all 3 experiments in Table~\ref{tab:seg}. As we can see from the table, performing copy-paste augmentation at random locations for semantic segmentation slightly degrades the model's performance by $0.2\%$. However when objects are placed in the right context, we achieve a boost of $2.1\%$ in mean intersection over union. These results resemble the case of object detection a lot and therefore highlight the role of context in scene understanding. We further analyze the categories that benefit from our data augmentation technique more than the others. If improvement for a class AP over the baseline is higher than $2.5\%$, Table~\ref{tab:seg} marks the result in bold. Again, we can notice correlation with the detection results from Section~\ref{sec:single} which demonstrates the importance of context for the categories that benefit from our augmentation strategy in both cases. \subsection{Reducing the need for pixel-wise object annotation}\label{exp:weak} Our data augmentation technique requires instance-level segmentations, which are not always available in realistic scenarios. In this section, we relax the annotation requirements for our approach and show that it is possible to use the method when only bounding boxes are available. \begin{figure}[btp!] \begin{center} \includegraphics[width=0.99\linewidth,trim=35 50 30 35,clip]{images/annotation_types.pdf} \end{center} \caption{\textbf{Possible types of instance-level annotation.} The left column presents an image annotated with object boxes. Column 2 shows semantic segmentation annotations with object boxes on top and approximate instance segmentations derived from it. The last column presents the original instance segmentation annotations.} \label{fig:annotation_types} \end{figure} \textbf{Semantic segmentation + bounding box annotations.} Instance segmentation masks provide annotations to each pixel in an image and specify (i) an instance a pixel belongs to and (ii) class of that instance. If these annotations are not available, one may approximate them with semantic segmentation and bounding boxes annotations. Figure~\ref{fig:annotation_types} illustrates possible annotation types and the difference between them. Semantic segmentation annotations are also pixel-wise, however they annotate each pixel only with the object category. Instance-specific information could be obtained from object bounding boxes, however this type of annotation is not pixel-wise and in some cases is not sufficient to assign each pixel to the correct instance. As Figure~\ref{fig:annotation_types} suggests, as long as a pixel in semantic map is covered by only one bounding box, it uniquely defines the object it belongs to (row 1); otherwise, if more than one box covers the pixel, it is not clear which object it comes from (row 2). When deriving approximate instance masks from semantic segmentation and bounding boxes (see Figure~\ref{fig:annotation_types}, column 2), we randomly order the boxes and assign pixels from a semantic map to the corresponding instances. Whenever a pixel could be assigned to multiple boxes we choose a box that comes first in the ordering. Once the procedure for obtaining object masks is established we are back to the initial setting and follow the proposed data augmentation routines described above. As could be seen in Tables~\ref{tab:single_weak}~and~\ref{tab:multy_weak} detection performance expiriences a slight drop of $0.6\%$ in single-category and $0.3\%$ in multi-category settings respectively, comparing to using instance segmentation masks. These results are promising and encourage us to explore less elaborate annotations for data augmentation. \textbf{Bounding box annotations only.} Since we have an established procedure for performing data augmentation with semantic segmentation and bounding boxes annotations, the next step to reducing pixel-wise annotation is to approximate segmentation masks. We employ weakly-supervised learning to estimate segmentations from bounding boxes and use the work of \cite{khoreva2017simple}. When trained on the \texttt{VOC12train} dataset, augmented with more training examples according to \cite{khoreva2017simple,gould2009decomposing}, it achieves $65.7\%$ mIoU on the \texttt{VOC12val-set}. Unfortunately, we have found that naively applying this solution for estimating segmentation masks and using them for augmentation results in worse performance. The reason for that was low quality of estimated masks. First, inaccurate object boundaries result in non-realistic instances and may introduce biases in the augmented dataset. But more importantly, confusion between classes may hampers the performance. For example, augmenting a category ``cow'' with examples of a ``sheep'' class may hurt the learning process. Hence, we need a model with a more discriminative classifier. To this end we propose the following modifications to the segmentation method: we change the architecture from DeepLab\_v1 \cite{chen2014semantic} to DeepLab\_v4 \cite{chen2018deeplab}, perform multi-scale inference and process the resulting masks with a conditional random field. The later helps to refine the object edges, which was found not necessary in the original work of \cite{chen2018deeplab}, when learning with full supervision. By training on the same data as the original method of \cite{khoreva2017simple} but with the proposed modifications we achieve $75.8\%$ mIoU, which is more than 10\% improvement to the initial pipeline. This accuracy seems to be sufficient to use automatically-estimated segmentation masks for augmentation purposes. When the semantic maps are estimated, we follow the augmentation routines of the previous section with only one difference; specifically, an instance is kept if the bounding box of its segmentation covers at least $40\%$ of its corresponding ground truth box. Otherwise, the object is not used for data augmentation. The results of applying this strategy to the single- and multy-category object detection are presented in Table~\ref{tab:single_weak}~and~\ref{tab:multy_weak}, respectively. Table~\ref{tab:single_weak} shows which categories are unable to provide high-quality masks, even though the quality seems to be sufficient to improve upon the non-augmented baseline. It is surprising that by using object boxes instead of segmentation masks we lose only $0.6\%$ of mAP in the multi-class scenario while still outperforming non-augmented training by $1.6\%$. These results show that the method is widely applicable even in the absence of segmentation annotations. \subsection{Studying the Importance of Context Modeling Quality for Scene Understanding}\label{exp:context_quality} First, we make an assumption that the quality of a context model is mainly influenced by the amount of data it has received for training. Hence, to study this relation, we mine a bigger dataset \texttt{VOC07-trainval+VOC12-trainval} which results in 16551 images. Then, we proceed by taking subsets of this dataset of increasing size and train the context model on them. Finally, we use the obtained context models to augment \texttt{VOC12-trainval} and train BlitzNet300 on it for detection and segmentation. Table \ref{tab:context_size} summarizes the object detection performance on \texttt{VOC07-test} and semantic segmentation performance on \texttt{VOC12val-seg}. In the current experiment, 10\% of the full set (1655 images) is roughly equal to the size of the \texttt{VOC12train-seg} (1464 images) initially used for training the context model. As we increase the data size used for context modeling, we can see how both detection and segmentation improve; however, this gain diminishes as the data size keeps growing. This probably mean that to improve scene understanding, the context model has to get visual context ``approximately right'' and further improvement is most likely limited by other factors such as unrealistic generated scenes and limited number of instances that are being copy-pasted. On the other hand, if the context model is trained with little data, as in the case of using only 5\% of the full set, our augmentation strategy tends to the random one and shows little improvement. \begin{table}[btp!] \centering \renewcommand{\arraystretch}{1.4} \renewcommand{\tabcolsep}{1.5mm} \resizebox{\linewidth}{!}{ \begin{tabular}{l |c| c c c c c c} $\%$ of data used & 0 & 5 & 10 & 25 & 50 & 75 & 100 \\ \hline Det. mAP & 64.6 & 65.3 & 66.1 & 66.4 & 66.7 & 66.9 & 66.9 \\ Seg. mIoU & 63.3 & 64.6 & 65.1 & 65.3 & 65.5 & 65.9 & 66.0 \\ \hline \end{tabular} } \caption{Object detection and semantic segmentation performance depending on amount of data used for building the context model. First row depicts the portion (in \%) of the \texttt{VOC0712trainval} used for training the context model. Second column corresponds to performance of baseline models. The second row gives the final detection mAP \% evaluated on \texttt{VOC07test}, while the third row lists segmentation mIou in \% on \texttt{VOC12val-seg}. For both tasks we used BlitzNet300 trained on augmented \texttt{VOC12train-seg}.} \label{tab:context_size} \end{table} \section{Introduction}\label{sec:introduction}} \input{intro.tex} \section{Related Work} \input{related.tex} \section{Approach} \input{approach.tex} \section{Experiments}\label{sec:exp} \input{exp.tex} \section{Conclusion} \input{ccl.tex} \section*{Acknowledgment} This work was supported by a grant from ANR (MACARON project under grant number ANR-14-CE23-0003-01), by the ERC grant number 714381 (SOLARIS project), the ERC advanced grant ALLEGRO and gifts from Amazon and Intel. \ifCLASSOPTIONcaptionsoff \newpage \fi \begin{IEEEbiography}[{\includegraphics[width=1.1in,height=1.55in,trim=0 40 0 0, clip,keepaspectratio]{./images/dvornik.jpg}}]{Nikita Dvornik} recieved the bachelor degree at the Moscow Institute of Physics and Technology~(MIPT) and master degree at INP Grenoble. He is currently working towards the PhD degree at INRIA Grenoble under supervision of Cordelia Schmid and Julien Mairal. His research interests include scene understanding tasks, such as object detection and semantic segmentaion, data augmentation, few-shot learning and learning general image representations under constraints. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1.1in,height=1.25in, clip,keepaspectratio]{./images/mairal.jpg}}]{Julien Mairal} (SM’16) received the Graduate degree from the Ecole Polytechnique, Palaiseau, France, in 2005, and the Ph.D. degree from Ecole Normale Superieure, Cachan, France, in 2010. He was ´ a Postdoctoral Researcher at the Statistics Department, UC Berkeley. In 2012, he joined Inria, Grenoble, France, where he is currently a Research Scientist. His research interests include machine learning, computer vision, mathematical optimization, and statistical image and signal processing. In 2016, he received a Starting Grant from the European Research Council and in 2017, he received the IEEE PAMI young research award. He was awarded the Cor Baayen prize in 2013, the IEEE PAMI young research award in 2017 and the test-of-time award at ICML 2019. \end{IEEEbiography} \begin{IEEEbiography}[{\includegraphics[width=1.4in,height=1.3in,trim=0 20 0 20, clip,keepaspectratio]{./images/cordelia.jpg}}]{Cordelia Schmid} holds a M.S. degree in Computer Science from the University of Karlsruhe and a Doctorate, also in Computer Science, from the Institut National Polytechnique de Grenoble (INPG). She is a reserach director at Inria Grenoble. She has been an editor-in-chief for IJCV (2013--2018), a program chair of IEEE CVPR 2005 and ECCV 2012 as well as a general chair of IEEE CVPR 2015. In 2006, 2014 and 2016, she was awarded the Longuet-Higgins prize for fundamental contributions in computer vision that have withstood the test of time. She is a fellow of IEEE. She was awarded an ERC advanced grant in 2013, the Humbolt research award in 2015 and the Inria \& French Academy of Science Grand Prix in 2016. She was elected to the German National Academy of Sciences, Leopoldina, in 2017. In 2018 she received th Koenderink prize for fundamental contributions in computer vision that have withstood the test of time. Starting 2018 she holds a joint appointment with Google research. \end{IEEEbiography} \bibliographystyle{IEEEtran}

13fcf939b1e8030517c7c6203760cc1fbdde4c17

\section{Introduction} Owing to its optimality under white noise \cite{brennan:2003}, MRC has become almost ubiquitous in multi-antenna systems used in the current wireless ecosystem. However, like other diversity-combining techniques, it suffers from performance losses in the presence of spatially correlated interference \cite{ganti:mhaenggi:comml:2009}. Therefore, it is important to characterize its exact performance under correlated interference, which is the main topic of this letter. Initial attempts towards characterizing the post-combining SIR, or equivalently the MRC outage/success probability, under spatially correlated interference include \cite{tanbourgi:2014} \cite{tanbourgi:dhillon:etel:trancom:2014}, which model the interference field as PPP and employ tools from stochastic geometry for the analysis. However, due to the structure of the problem, the exact analysis was limited to only the case of $N=2$. For the general case of $N>2$, only simple bounds were provided. Therefore, in this letter, our objective is to characterize, or tightly approximate, the Cumulative Distribution Function (CDF) of the interference-affected post-combining SIR for arbitrary $N$. To achieve this, we employ the mixture-based method \cite{arindam:comml:2017} of modeling the correlation in the interference powers experienced at multiple antennas. This method, at its core, employs a correlation framework constructed using \textit{mixture of random variables}, which not only offers a much more amenable analytical structure but also mimics the PPP characteristics quite well. The derived analytical expressions for MRC outage probability are shown to tightly approximate the exact PPP-based results obtained using Monte Carlo simulations. As a part of this contribution, we also provide a method for improving the accuracy of the mixture-based approximation method of \cite{arindam:comml:2017}. Originally, in \cite{arindam:comml:2017}, the mixture weights ($q$'s) were chosen so as to match the interference correlation of the mixture-based and PPP-based models. Different from \cite{arindam:comml:2017}, in this letter, we show that tuning $q$ instead to match the joint SIR statistics, i.e. joint complementary CDF (CCDF) of SIR, of the two models results in much tighter approximations. This tuning requires solving an $N^{\text{th}}$ degree polynomial equation in $q$, which can be easily accomplished using commonly available numerical packages. \vspace{-0.25cm} \section{System Model} We consider a Poisson dipole network \cite{Baccelli:Blaszczyszyn:Muhlethaler2006} wherein the transmitters (single antenna) are distributed as per a homogeneous PPP $\Phi \equiv \{x\} \subset \mathbb{R}^{2}$ of intensity $\lambda$. Each transmitter has, at a distance $d$ in an arbitrary direction, an $N$-antenna receiver capable of MRC processing. By Slivnyak's theorem \cite{mhaenggi:ganti:2009}, we then add a new reference transmitter into the network with its receiver placed at the origin $o$, which can be referred to as the ``typical" pair. In this letter, we focus on this pair to study the perfomance of the considered MRC. We denote the location of the typical receiver's antennas by $\{z_i\}_{i=1}^{N}\in \mathbb{R}^{2}$. The separation between these antennas is assumed to be negligible (of the order of wavelength) in comparison to other distances in the network (such as $d$). Therefore, for the purpose of path-loss calculations, we will assume that all antennas are co-located with the receiver itself, i.e. $\{z_i \equiv o\}$. That said, it should be noted that the small separation between the antennas may still allow them to experience significantly different fading characteristics. For the typical receiver, the nodes at $\Phi$ (not including the serving transmitter) act as interferers. These interferers are assumed to follow ALOHA protocol whereby each interferer transmits independently on the same time-frequency resource block as the typical link with probability $p$. Therefore, the set of active interferers $\hat{\Phi}$ that interfere with the typical receiver in any given time slot is simply a thinned PPP of intensity $\lambda p$ (parent PPP $\Phi$ thinned with probability $p$). For channel fluctuations, we assume i.i.d. Rayleigh fading across all the links. For path-loss, between $x$ and $i$-th antenna ($z_i$), we consider the standard function $\ell(x,z_i)=\frac{1}{\epsilon+\|x-z_i\|^{\alpha}}$, where $\alpha>2$ is the path-loss exponent and $\epsilon \rightarrow 0$. For the case of $z_i \equiv o$, we simply have $\ell(x,z_i)=\frac{1}{\epsilon+\|x\|^{\alpha}}$. Lastly, we assume that all nodes in the system transmit with $P$ units of power. Then, in this setting, the SIR at the $i$-th antenna due to a transmission from the reference transmitter is given by \vspace{-0.1cm} \begin{equation} \text{SIR}_i = \frac{Ph_{i}d^{-\alpha}}{\sum_{x\in\hat{\Phi}}^{}Ph_{xz_i}\|x\|^{-\alpha}}=\frac{h_{i}d^{-\alpha}}{I(z_i)}, \label{SIR} \end{equation} where, $I(z_i)$ is the interference power normalized by $P$, and $h_{xz_i}$, $h_i$ $\sim \exp(1)$ are the exponentially distributed fading gains of the link between $x$ and $z_i$ and between the reference transmitter and $i$-th antenna, respectively. We focus on interference-limited scenario in which thermal noise is negligible compared to interference. Since all antennas of the receiver of interest are collocated from the path-loss perspective and all links experience i.i.d. Rayleigh fading, it is easy to argue that the interference power experienced across different antennas is identically distributed, i.e., $\{I(z_i)\}$ is a sequence of identically distributed random variables. Further, these interference powers exhibit spatial correlation because of the common interference field. The spatial correlation coefficient $\zeta_{ij} = \text{Corr}[I(z_i), I(z_j)]$, $\forall i\neq j$, can be readily quantified using only the spatial aspect of \cite[(11)]{ganti:mhaenggi:comml:2009}, i.e. \begin{equation} \begin{split} \zeta_{ij} &=\lim\limits_{\epsilon \rightarrow 0}\frac{\int_{\mathbb{R}^2}^{}\ell(x,o)^2\dd x}{\mathbb{E}[h_{xz_i}^2]\int_{\mathbb{R}^2}^{}\ell(x,o)^2\dd x}\\ & = \lim\limits_{\epsilon \rightarrow 0}\frac{\int_{\mathbb{R}^2}^{}\frac{1}{(\epsilon+\|x\|^{\alpha})^{2}}\dd x}{\mathbb{E}[h_{xz_i}^2]\int_{\mathbb{R}^2}^{}\frac{1}{(\epsilon+\|x\|^{\alpha})^{2}}\dd x}=\lim\limits_{\epsilon \rightarrow 0}\frac{1}{\mathbb{E}[h_{xz_i}^2]} = \frac{1}{2}, \end{split} \label{zeta_same_point} \end{equation} where, the last step follows from $\mathbb{E}[h^n] = n!$ for $h\sim \exp(1)$. For MRC, we consider the interference-aware case, as also assumed in \cite{tanbourgi:2014} \cite{tanbourgi:dhillon:etel:trancom:2014}, where the receiver has perfect knowledge of the instantaneous interference powers and the transmitter-to-receiver link fading gains for every antenna. The combiner treats the interference as white noise and the MRC weights are taken to be proportional to the ratio of fading amplitude and interference power \cite{brennan:2003}. The post-MRC SIR, denoted by $\text{SIR}_\text{MRC}$, is therefore given by \begin{equation} \text{SIR}_\text{MRC} =\frac{h_{1} d^{-\alpha}}{I(z_1)} + \ldots + \frac{h_{N} d^{-\alpha}}{I(z_N)}. \label{sir_mrc_expansion} \end{equation} For the case of $N=2$, this SIR has been fully characterized in \cite{tanbourgi:2014}. However, for the general case, the standard approaches are not analytically tractable, which makes the analysis extremely difficult. In the next section, we present an alternate approach, wherein we employ the mixture-based method of interference modeling \cite{arindam:comml:2017}, to obtain the CDF of $\text{SIR}_\text{MRC}$ for arbitrary $N$. \vspace{-0.15cm} \section{Outage Probability of MRC} The mixture-based model in \cite{arindam:comml:2017} can be used to construct a set of arbitrarily (non-negatively) correlated interference powers $\{I(z_i)\}_{i=1}^N$. For the multi-antenna case of this letter, where $I(z_i)$ and $I(z_j)$ are equally correlated $\forall i\neq j$, the mixture-based framework of \cite{arindam:comml:2017} can be reduced to a more simpler form as presented in the following. \vspace{-0.4cm} \subsection{Mixture-based Construction of Interference Powers} The main idea is to represent the PPP-based interference random variables $\{I(z_i\equiv o)\}_{i=1}^N$ with their mixture-based equivalents having the same PPP-based distributional and correlational properties. To achieve this, we first generate a sequence of i.i.d. random variables $\{J_n\}_{n=0}^{N}$ that are distributed identically to $I(o)$ of (1). For that, we imagine an auxiliary setup wherein we take a set of independent homogeneous PPPs of interferers $\{{\Psi}_n\}_{n=0}^N$, of intensity $\lambda p$, on $\mathbb{R}^2$ and model $\{J_n\}$ to be the respective interference powers observed at the origin $o$ due to these PPPs, i.e. \begin{equation} J_{n} = \sum_{x\in{\Psi}_{n}}^{}h_{xo}\ell(x,o) \qquad \text{for}\ n\in\{0,\ldots,N\}. \label{mixture_rv_J} \end{equation} Note that $\{{\Psi}_n\}$, which are used to obtain $\{J_n\}$, belong only to the auxiliary system and in no way interfere with the actual PPP $\hat{\Phi}$ of the interferers in the original system of Section II. Next, using these $J_n$'s, we model $\{{I}(z_i)\}$ as the mixtures \begin{equation} \begin{split} I(z_i) = J_{A_i}, \qquad \text{for}\ i\in\{1,\ldots,N\}, \label{I_mixture} \end{split} \end{equation} where, $\{A_i\}_{i=1}^{N}$ are independent binary random variables whose probability mass functions are given by \begin{equation} p_{A_i}(a_i)= \begin{cases} q, & \text{if}\ a_i=0 \\ 1-q, & \text{if}\ a_i=i\\ 0, & \text{otherwise}. \end{cases} \end{equation} From the distribution preservation property of mixtures [5], we have the mixture-based $I(z_i)$ distributed identically to $J_0$ and $J_i$ and therefore to PPP-based $I(z_i)$ $\forall i$. Further, it can be easily verified that the mixture-based $\text{Corr}[{I}(z_i),{I}(z_j)]$ $=$ $q^2$, for all $i\neq j$. At this point, it appears natural to select $q^2 = 0.5$ so as to match the values of the correlation coefficient with (2), but, as discussed in the sequel, this does not necessarily offer the most accurate approximation. Next, we use this mixture-based framework to derive the MRC outage probability for an arbitrary $N$. \vspace{-0.3cm} \subsection{CDF of $\text{SIR}_\text{MRC}$ for arbitrary $N$} The main result of this paper is stated in the next Theorem. \noindent\textbf{Theorem 1} \textit{The outage probability of MRC, defined as $\text{P}_{\text{Out}} \stackrel{\Delta}{=}\mathbb{P}(\text{SIR}_\text{MRC} < T)$, derived using the mixture-based method, is given by \eqref{eq_no_big_Pout}, where $T$ is the given threshold, $C=\frac{2\pi^2\lambda p}{\alpha \sin(2\pi/\alpha)}$, $f_{V_k}(v_k) = \frac{2C}{\alpha}v_k^{\frac{2}{\alpha}-1}e^{-C v_k^{2/\alpha}}$, and $B_{m,j}(\cdot)$ is the Bell polynomial \cite{jhonson:bell}.} \textit{Proof:} From \eqref{sir_mrc_expansion}, we have \vspace{-0.4cm} \begin{figure*}[!t] \normalsize \begin{equation*} \begin{split} \text{P}_{\text{Out}} &=\sum_{n=0}^{N} {N \choose n} q^n (1-q)^{N-n}\hspace{-0.15cm}\int_{0}^{Td^{\alpha}}\hspace{-0.3cm}\int_{0}^{Td^{\alpha}\hspace{-0.1cm}-v_1}\hspace{-0.7cm}\cdots \int_{0}^{Td^{\alpha}\hspace{-0.1cm}-v_1\hspace{-0.05cm}-\ldots -v_{N-n-1}}\hspace{-0.1cm}\Biggl\{\hspace{-0.05cm}1\hspace{-0.05cm}-\hspace{-0.17cm}\sum_{m=0}^{n-1}\hspace{-0.075cm}(-1)^m\hspace{-0.05cm}\frac{\left(\hspace{-0.05cm}Td^{\alpha}\hspace{-0.15cm}-\hspace{-0.1cm}\sum_{k=1}^{N-n} \hspace{-0.1cm}v_k\hspace{-0.05cm}\right)^{\hspace{-0.1cm}m}}{m!}\hspace{-0.15cm}\sum_{j=1}^{m}\hspace{-0.05cm}(\hspace{-0.05cm}-1\hspace{-0.05cm})^je^{-C \left(\hspace{-0.05cm}Td^{\alpha}\hspace{-0.1cm}-\hspace{-0.05cm}\sum_{k=1}^{N-n} \hspace{-0.075cm}v_k\hspace{-0.05cm}\right)^{\frac{2}{\alpha}}}\hspace{-0.1cm}.\\ &\hspace{1.2cm}B_{m,j}\biggl\{\hspace{-0.05cm}\frac{2C\hspace{-0.05cm}\left(\hspace{-0.1cm}Td^{\alpha}\hspace{-0.15cm}-\hspace{-0.175cm}\sum_{k=1}^{N-n} \hspace{-0.1cm}v_k\hspace{-0.1cm}\right)^{\hspace{-0.1cm}\frac{2}{\alpha}\hspace{-0.025cm}-\hspace{-0.025cm}1\hspace{-0.5cm}}}{\alpha}, \ldots, \hspace{-0.075cm}\frac{\hspace{-0.05cm}2\hspace{-0.05cm}\left(\hspace{-0.05cm}2\hspace{-0.075cm}-\hspace{-0.075cm}\alpha\hspace{-0.05cm}\right)\hspace{-0.05cm}\cdots\hspace{-0.05cm} \left(\hspace{-0.05cm}2\hspace{-0.075cm}-\hspace{-0.075cm}\alpha(\hspace{-0.05cm}m\hspace{-0.075cm}+\hspace{-0.075cm}j\hspace{-0.05cm})\hspace{-0.05cm}\right)\hspace{-0.05cm}C\hspace{-0.05cm}\left(\hspace{-0.1cm}Td^{\alpha}\hspace{-0.15cm}-\hspace{-0.175cm}\sum_{k=1}^{N-n}\hspace{-0.1cm}v_k\hspace{-0.1cm}\right)^{\hspace{-0.1cm}\frac{2}{\alpha}\hspace{-0.05cm}-\hspace{-0.05cm}1\hspace{-0.05cm}-\hspace{-0.05cm}m\hspace{-0.05cm}+\hspace{-0.05cm}j\hspace{-0.8cm}}}{\alpha^{m-j+1}}\biggr\}\hspace{0.4cm}\Biggr\}\Biggl\{\prod_{k=1}^{N-n}\hspace{-0.075cm}f_{V_k}\hspace{-0.05cm}(v_k)\hspace{-0.05cm}\Biggr\}\dd v_{N-n}\ldots \dd v_1 \end{split} \tag{15} \label{eq_no_big_Pout} \end{equation*} \hrulefill \vspace*{-0.4cm} \end{figure*} \begin{equation*} \begin{split} \text{P}_{\text{Out}} &= \mathbb{P}\left(\frac{h_{1} d^{-\alpha}}{I(z_1)} + \ldots + \frac{h_{N} d^{-\alpha}}{I(z_N)} < T\right)\\ &\stackrel{(b)}{=}\mathbb{P}\left(\frac{h_1}{J_{A_1}} + \ldots + \frac{h_N}{J_{A_N}} < Td^{\alpha}\right)\\ &\stackrel{(c)}{=} \sum_{a_1\in \{0,1\}}^{} \sum_{a_2\in \{0,2\}}^{} \ldots \sum_{a_N \in \{0,N\}}^{} \left(\prod_{i=1}^{N}p_{A_i}(a_i)\right).\\ &\ \ \ \mathbb{P}\left(\frac{h_1}{J_{A_1}} + \ldots + \frac{h_N}{J_{A_N}} < Td^{\alpha}\Bigr\rvert A_1=a_1,\ldots,A_N=a_N\right) \end{split} \end{equation*} where, (b) follows from the mixture-based representations of $\{{I}(z_i)\}$ in \eqref{I_mixture}, and $(c)$ from the law of total probability. Next, by collecting the common terms, the above can be compactly written as: $\text{P}_{\text{Out}} =$ \begin{equation} \begin{split} \hspace{-0.1cm}\sum_{n=0}^{N} \hspace{-0.1cm}{N \choose n} q^n (1\hspace{-0.075cm}-\hspace{-0.075cm}q)^{N-n}\underbrace{\mathbb{P}\left(\frac{\sum_{i=1}^{n}h_i}{J_0} + \sum_{k=1}^{N-n}\frac{h_k}{J_k} < Td^{\alpha}\right)}_{=W_n} \end{split} \label{P_out} \end{equation} Here, we define $U_n=\frac{\sum_{i=1}^{n}h_i}{J_0}$ and $V_k=\frac{h_k}{J_k}$ and find their distributions. In this regard, we observe that $U_n d^{-\alpha}$ is equivalent to the random variable $\text{SIR}_\text{MRC}$ when all the $N$ antennas see the same interference power. This has been studied as the full-correlation case in \cite{tanbourgi:2014}, and therefore, from \cite[Lemma 1, Proposition 1]{tanbourgi:2014}, we have the CDF of $U_n$ given by \vspace{-0.4cm} \begin{equation} F_{U_n}(u) = 1-\sum_{m=0}^{n-1}(-1)^m\frac{u^m}{m!}\frac{\partial^m }{\partial u^m}\exp\left(-C u^{2/\alpha}\right), \label{Un_cdf} \end{equation} where $C=\frac{2\pi^2\lambda p}{\alpha \sin(2\pi/\alpha)}$. Next, note that $V_k$ is distributed identically to $U_1$; hence, from \eqref{Un_cdf}, we have the CDF and PDF of $V_k$, respectively, given by \vspace{-0.2cm} \begin{align} F_{V_k}(v_k)& = 1-\exp\left(-C v_k^{2/\alpha}\right),\ \ \text{and}\\ f_{V_k}(v_k)& = \frac{2C}{\alpha}v_k^{\frac{2}{\alpha}-1}\exp\left(-C v_k^{2/\alpha}\right). \label{Vk_cdf_pdf} \end{align} \vspace{-0.3cm} Using these distributions, $W_n$ can now be derived as \begin{equation} \begin{split} W_n &= \mathbb{P}\left(U_n + \sum_{k=1}^{N-n} V_k < Td^{\alpha}\right)\\ &=\mathbb{E}_{\{V_k\}}\left[\mathbb{P}\left(U< Td^{\alpha}-\sum_{k=1}^{N-n} V_k\right)\right]\\ &=\mathbb{E}_{\{V_k\}}\left[\left.1-\sum_{m=0}^{n-1}(-1)^m\frac{s^m}{m!}\frac{\partial^m }{\partial s^m}\exp\left(-C s^{2/\alpha}\right)\right\rvert_{s}\right], \end{split} \end{equation} where $\rvert_{s}$ means the function inside the expectation is evaluated at $s=Td^{\alpha}-\sum_{k=1}^{N-n} V_k$. Next, using Faa di Bruno's formula \cite{bruno:1857} and Bell polynomial $B_{m,j}(\cdot)$ \cite{jhonson:bell}, for the $m$-th derivative, we get \vspace{-0.4cm} \begin{equation} \begin{split} W_n&=\mathbb{E}_{\{V_k\}}\biggl[1-\sum_{m=0}^{n-1}(-1)^m\frac{s^m}{m!}\sum_{j=1}^{m}(-1)^je^{-C s^{\frac{2}{\alpha}}}.\\ &\qquad\qquad B_{m,j}\left\{\frac{\partial}{\partial s}C s^{\frac{2}{\alpha}}, \ldots, \frac{\partial^{m-j+1}}{\partial s^{m-j+1}}C s^{\frac{2}{\alpha}} \right\}\biggr\rvert_{s}\biggr], \end{split} \end{equation} Finally, averaging over the i.i.d. $\{V_k\}$ and substituting into \eqref{P_out}, we get the result. \vspace{-0.3cm} \subsection{Tuning of the Mixture Parameter ``$q$"}\label{mapping_ssec} We conclude the mixture-based model of this letter by choosing a value for the parameter $q$ that results in highly accurate mixture-based approximations. Here, we propose a new method of tuning (based on matching of joint statistics of the SIR) which offers a much tighter approximation than the tuning presented originally in \cite{arindam:comml:2017}. Recall that tuning in \cite{arindam:comml:2017} is based on interference correlation matching, which results in $q^2=0.5$ from \eqref{zeta_same_point}. The MRC outage probability involves (jointly) multiple correlated SIRs; and, as the accuracy of the mixture-based approximation depends on how closely it mimics the PPP-based characteristics, we propose to match the two models directly at the level of joint SIR statistics. That is, tune $q$ such that the joint CCDFs of SIR match. The difference between the two CCDFs, for a threshold $T$, is given by \begin{equation} \begin{split} &f(q) = \ e^{\frac{-B\Gamma\left(N+\frac{2}{\alpha}\right)}{(N-1)!\ \Gamma\left(1+\frac{2}{\alpha}\right)}}\\ &- \sum_{n=0}^{N}{N \choose n}q^n(1-q)^{N-n}\exp{-B(n^{\frac{2}{\alpha}}+N-n)}, \end{split} \label{eq_joint_ccdf_matching} \end{equation} where, $B=\frac{2\pi^2\lambda p d^2 T^{\frac{2}{\alpha}}}{\alpha \sin(2\pi/\alpha)}$. The first CCDF term is PPP-based that is available from \cite[Theorem 1]{haenggi:2012}. The second term is for the mixture-based case, which is given as: $\mathbb{P}\left(\text{SIR}_1>T, \ldots,\text{SIR}_N>T\right)$ = \begin{equation*} \begin{split} & \mathbb{P}\left(\cap_{i=1}^Nh_i>yd^{\alpha}I(z_i)\right)= \mathbb{E}_{\{J_{A_i}\}}\biggl[e^{-\sum_{i=1}^{N}yd^{\alpha}J_{A_i}}\biggr]\\ &\stackrel{(b)}{=} \hspace{-0.25cm}\sum_{a_1\in \{0,1\}}^{} \hspace{-0.25cm}\ldots\hspace{-0.2cm} \sum_{a_N \in \{0,N\}}^{} \hspace{-0.15cm}\left(\prod_{i=1}^{N}p_{A_i}(a_i)\right)\mathbb{E}_{\{J_{a_i}\}}\biggl[e^{-\sum_{i=1}^{N}yd^{\alpha}J_{a_i}}\biggr]\\ &\stackrel{(c)}{=}\sum_{n=0}^{N} {N \choose n} q^n (1\hspace{-0.1 cm}-\hspace{-0.1 cm}q)^{N\hspace{-0.05 cm}-\hspace{-0.05 cm}n}\mathbb{E}_{J_0}\hspace{-0.05 cm}\biggl[e^{-nyd^{\alpha}J_{0}}\biggr]\mathbb{E}_{J_1}\hspace{-0.05 cm}\biggl[e^{-yd^{\alpha}J_{1}}\biggr]^{\hspace{-0.05 cm}N-n}, \end{split} \end{equation*} where, (b) follows from the total probability law and (c) by collecting the common terms. Lastly, by taking the Laplace transform of interference \cite{mhaenggi:ganti:2009}, we get the expression in \eqref{eq_joint_ccdf_matching}. The tuned $q$ which matches the two CCDFs is simply the solution of the equation $f(q)=0$. From Abel-Ruffini's theorem \cite{jacobson:2009}, there is no general algebraic solution for polynomial equations with arbitrary coefficients for $N\geq 5$. Although this does not mean that some particular classes of higher degree polynomials may not have explicit form of algebraic solutions, such an exploration for the above equation is outside the scope of this study. Therefore, we simply use commonly available numerical packages from Mathematica or Matlab to find the appropriate value of $q\in[0,1]$ which satisfies the above equation. The above tuning, as shown in the next section, results in far better accuracy compared to simply setting $q^2=\zeta_{ij}=0.5$ as per \cite{arindam:comml:2017}. This is because even if the two models are matched at the interference correlation level, deviations may still appear at the SIR correlation level. To understand this, note that the SIR correlation can be expressed as \begin{equation} \text{Corr}[\text{SIR}_i,\text{SIR}_j] = \zeta^{\text{inv}}_{ij}\sqrt{\frac{\text{Var}[I(z_i)^{-1}]\text{Var}[I(z_j)^{-1}]}{\text{Var}[h_iI(z_i)^{-1}]\text{Var}[h_jI(z_j)^{-1}]}} \end{equation} where, $\zeta^{\text{inv}}_{ij}= \text{Corr}\left[ I(z_i)^{-1},I(z_j)^{-1}\right]$. From results obtained using Monte Carlo simulations, we find that for an interference correlation of $0.5$ (for $\alpha=4$ and $\lambda p=10^{-2}/\text{m}^2$), $\zeta^{\text{inv}}_{ij}\approx 0.76$, which is then brought down to the SIR correlation of $\approx 0.3$. However, for mixtures, it can be easily shown that if $\zeta_{ij}=0.5$, $\zeta^{\text{inv}}_{ij}$ is also $0.5$, which therefore will result in an SIR correlation of $< 0.3$. Clearly, this mismatch will affect the accuracy of the derived approximations, which motivated us to tune $q$ as per \eqref{eq_joint_ccdf_matching} instead. \vspace{-0.2cm} \section{Simulation Results}\label{simulation_sec} \begin{figure}[t] \centering \includegraphics[scale=0.46]{PvsT_N4.eps} \caption{$P_{\text{Out}}$ vs. $T$ for $N=4$, $\alpha=4$, and $d=10$m.} \label{Pout_vs_T_N4} \vspace{-0.5cm} \end{figure} \begin{figure}[t] \centering \includegraphics[scale=0.45]{Pout_vs_N.eps} \caption{$P_{\text{Out}}$ vs. $N$, $T=1$dB, $\alpha=4$, $d=10$m.} \label{Pout_vs_N} \vspace{-0.5cm} \end{figure} \begin{figure}[t] \centering \includegraphics[scale=0.47]{q_tuned.eps} \caption{Tuned $q^2$, $T=1$ dB, $\lambda p=10^{-4}/\text{m}^2$, $\alpha=4$, $d=10$m.} \label{fig_tuned_q_values} \vspace{-0.5cm} \end{figure} For simulations, we consider a square region $[-L,L]^2$ centered at the origin such that there are on an average 1000 interferers present, i.e. $L$ satisfies $(2L)^2\lambda p=10^3$. The receiver is then placed at the center with its reference transmitter at a distance $d=10$m. For path-loss, we take $\alpha=4$. Figure \ref{Pout_vs_T_N4} plots the outage probability of the $4$-antenna MRC receiver with respect to the threshold $T$. Clearly, the accuracy of the mixture-based approximations depends on the value of $q$. The tightest match with the simulation data is obtained when $q$ is tuned as per \eqref{eq_joint_ccdf_matching}. Setting $q^2=\zeta^{\text{inv}}_{ij}=0.76$ (which matches the two models at the SIR correlation level) gives better accuracy than $q^2=\zeta_{ij}=0.5$, however, it still slightly undershoots the plot. A very high value of $q^2$ ($=0.9$), on the other hand, causes overshooting. Similar behavior is observed in Fig. \ref{Pout_vs_N}, which plots $\text{P}_{\text{Out}}$ against the number of receiver antennas. We see that the tuned values of $q$ (plotted in Fig. \ref{fig_tuned_q_values}), provide the tightest approximations to the actual PPP-based data. For comparison purpose, we also plot the upper and lower bounds of the MRC outage probability that were proposed in \cite[Proposition 2]{tanbourgi:2014} for arbitrary $N$. Not surprisingly, the mixture-based approximations offer far accurate results than the PPP-based bounds. \vspace{-0.15cm} \section{Conclusions} In this letter, using the mixture-based method of modeling spatially correlated interference, we derived accurate expressions for the MRC outage probability for any arbitrary number of antennas, which was earlier not possible through standard PPP-based approach. In addition, different from \cite{arindam:comml:2017}, we present a new method of tuning the mixture parameter $q$ that is based on matching the joint CCDF of SIR. These tuned mixture-based approximations are shown to obtain far better accuracy than the tuning in \cite{arindam:comml:2017} and the previously known bounds in the literature. The mixture-based model in this letter, therefore, can be used to study many such related scenarios of correlated interference in multi-antenna systems. \vspace{-0.2cm} \bibliographystyle{IEEEtran}

2e14c3b3894c64d69bf42f7ea4b5ce237c75ad29

\section{Introduction} \label{sec:intro} \begin{figure}[t] \includegraphics[width=\linewidth]{figures/mirflickr_examples_2col} \caption{Comparison of candidate batches selected for the second annotation round by ITAL with the selection of TCAL \cite{demir2015tcal} on 4 exemplary queries from the MIRFLICKR \cite{mirflickr} dataset. The border colors correspond to different topics that could be associated with the query. Obviously, ITAL explores much more diverse relevant topics than TCAL.} \label{fig:mirflickr-examples} \end{figure} For content-based image retrieval (CBIR) \cite{niblack1993qbic,smeulders2000cbir}, it is, in general, not sufficient to just classify the query image or to identify the same object on different images. Since images encode complex semantic and stylistic information---sometimes more than text can express---a single query is usually insufficient to comprehend the search interest of the user. A common approach to overcome this issue is enabling the user to provide \textit{relevance feedback} by tagging some retrieval results as relevant or irrelevant \cite{cox2000pichunter,tong2001support,deselaers2008learning}. This way, however, the user will only be able to give feedback regarding images about which the retrieval system is already very confident. The effect of relevance feedback could hence be significantly improved when the user is not asked to provide feedback regarding the currently top-scoring results, but for those instances of the dataset that are most informative for the classifier to distinguish between relevant and irrelevant images. Finding such a set of most informative samples is the objective of \textit{batch-mode active learning (BMAL)} \cite{brinker2003incorporating,guo2008discriminative,giang2014batch,yang2015usdm}, which has recently been explored for CBIR \cite{demir2015tcal,cardoso2017rbmal}. However, the performance of existing approaches usually varies substantially between datasets, which is not only observable in our experiments, but also in comparative evaluations in the existing literature (\textit{e.g.}, \cite{guo2008discriminative}). In this work, we propose \textit{Information-Theoretic Active Learning (ITAL)}, a BMAL method for relevance feedback that does not suffer from this instability, but provides state-of-the-art performance across different datasets. This is demonstrated by a comparison of our approach with a variety of competitor methods on 5 image datasets of very different type and structure. Our method for selecting unlabeled samples for annotation by the user \begin{enumerate*}[label=\upshape(\itshape\alph*\upshape)] \item implicitly maintains both diversity and informativeness of the candidate images, \item employs an explicit model of the user behavior for dealing with the possibility of incorrect annotations and \textit{unnameable instances} \cite{Kaeding15_ALD}, \textit{i.e.}, images which the user cannot classify at all, \item takes the model output change caused by the expected user feedback into account, \item can easily be parallelized for processing large datasets, and \item works with as few as a single initial training sample. \end{enumerate*} The user model allows ITAL to compensate for unreliable users, who are likely to make mistakes or to refuse giving feedback. It acts as an implicit mechanism for controlling the trade-off between redundancy and diversity of the batch of samples selected for annotation. Because care has to be taken not only that all images in the batch of unlabeled samples selected for annotation are \textit{informative} individually, but that they are also \textit{diverse} compared to each other to avoid unnecessary redundant feedback. The majority of existing works on BMAL try to achieve this using a combination of several heuristics to simultaneously maximize the \textit{diversity} within the batch and the \textit{uncertainty} of the selected samples or their \textit{density} in the dataset \cite{brinker2003incorporating,zhu2008sud,giang2014batch,demir2015tcal,yang2015usdm,cardoso2017rbmal}. Our proposed ITAL method, in contrast, aims to maximize the mutual information (MI) between the expected user feedback and the relevance model. By taking the joint distribution of the predictive relevance of the samples in the batch into account, the MI criterion implicitly maintains diversity without the need for any heuristics or manually tuned linear combinations of different criteria. Instead, our method does not only take the structure of the data and the current relevance predictions into account, but also considers the expected impact that annotating the selected samples would have on the predicted relevance after updating the model. This integration of the expected model output change (EMOC) has successfully been used for one-sample-at-a-time active learning \cite{freytag2014EMOC}, but, to the best of our knowledge, not been applied to BMAL yet. However, computing the expected model output change requires relevance models that can be updated efficiently. In addition, both the relevance model and the active learning technique should be capable of working with as few training data as a single positive query example provided by the user. We achieve both by using a Gaussian process (GP) \cite{rasmussen2006gaussian} for classification, which can be fitted to a single training sample and can be updated using a closed-form solution without the need for iterative optimization. This is in contrast to many other works on active learning, which are based on logistic regression \cite{guo2007optimistic,li2013adaptive} or support vector machines (SVMs) \cite{tong2001support,brinker2003incorporating,giang2014batch,demir2015tcal} as classification technique. Moreover, SVMs require a fair amount of both positive and negative initial training data for learning a robust hyperplane. Thus, such an approach is not feasible for image retrieval. \Cref{fig:mirflickr-examples} illustrates the advantages of our approach: While existing methods often select images similar to the query, but with high uncertainty (\textit{e.g.}, only dogs for a dog query or birds for a bird query), ITAL additionally explores the different meanings of the query image. The query showing a bird in front of the sea could as well refer to images of the sea or to animals at the sea in general. The dog query, on the other hand, could refer to images showing two dogs, images of dogs in general, or images of white animals. Finally, the user providing the beach image as query could be interested in images of the coast, of creatures at the beach, or also just in images of people in action without necessarily being at the beach. All these various options are explored by ITAL, which actively asks the user for the feedback to resolve these ambiguities. We will briefly review related methods in the following section and explain our ITAL method in detail in \cref{sec:ital}. The experiments mentioned above are presented in \cref{sec:experiments} and \cref{sec:conclusions} concludes this paper. \section{Related Work} \label{sec:related-work} The use of active learning methods is, of course, not limited to information retrieval applications, but also evident in the scenario of manual annotation of large unlabeled datasets: One would prefer spending money and human effort on labeling the most useful samples instead of outliers. Thus, active learning has been extensively studied for several years across various application domains, including binary classification \cite{tong2001support,brinker2003incorporating,guo2008discriminative,freytag2013labeling,freytag2014EMOC}, multi-class classification \cite{guo2007optimistic,jain2009pknn,li2013adaptive,Kaeding15_ALD,yang2015usdm}, and regression \cite{guestrin2005near,krause2007nonmyopic,Kaeding18_ALR}. With regard to batch-mode active learning (BMAL), most existing methods employ some combination of the criteria uncertainty, diversity, and density: Brinker \cite{brinker2003incorporating} proposes to select samples close to the decision boundary, while enforcing diversity by minimizing the maximum cosine similarity of samples within the batch. Similarly, ``Sampling by Uncertainty and Density (SUD)'' \cite{zhu2008sud} selects samples maximizing the product of entropy and average cosine similarity to the nearest neighbors and ``Ranked Batch-mode Active Learning (RBMAL)'' \cite{cardoso2017rbmal} constructs a batch by successively adding samples with high uncertainty and low maximum similarity to any other already selected sample. ``Triple Criteria Active Learning (TCAL)'' \cite{demir2015tcal}, on the other hand, first selects a subset of uncertain samples near the decision boundary, divides them into $k$ clusters, and chooses that sample from each cluster that has the minimum average distance to all other samples in the same cluster. Following a more complex approach, ``Uncertainty Sampling with Diversity Maximization (USDM)'' \cite{yang2015usdm} finds a trade-off between the individual entropy of the samples in the batch and their diversity by formulating this optimization problem as a quadratic program, whose parameters to be determined are the ranking-scores of the unlabeled samples. In addition, two works use an information-theoretic approach and are, thus, particularly similar to our method: Guo \& Greiner \cite{guo2007optimistic} propose to maximize the mutual information between the selected sample and the remaining unlabeled instances, given the already labeled data. They reduce this objective to the minimization of conditional entropy of the predictive label distribution of the unlabeled samples, given the existing labels and a proxy-label for the selected instance. With regard to the latter, they make an optimistic guess assuming the label which would minimize mutual information. If this guess turns out to be wrong, they fall back to uncertainty sampling for the next iteration. Though the results obtained by this approach called MCMI[min]+MU are convincing, it is computationally demanding and not scalable to real-world scenarios, even though the authors already employed some assumptions to make it more tractable. In particular, they assume that the conditional entropy of a set of samples can be decomposed as a sum of the entropy of individual samples. However, this assumption ignores relationships between unlabeled samples and is hence not suitable for a batch-mode scenario. In our work, we employ different approximations and Gaussian processes to enable the use of mutual information for BMAL. Using Gaussian processes instead of logistic regression or SVMs also allows us to take the impact of user feedback on the model output into account, since updating a GP does not involve iterative algorithms. On the other hand, Li \& Guo \cite{li2013adaptive} employ mutual information as a measure for the information density of the unlabeled samples and combine it with the conditional entropy of their individual labels as uncertainty measure. Similar to our approach, they use a GP to estimate the mutual information, but then employ logistic regression for the actual classification. Furthermore, their method cannot be applied to a batch-mode scenario and does not scale to large datasets, so that they need to randomly sub-sample the unlabeled data. Our ITAL method, in contrast, forms a consistent framework, provides a batch-mode, considers the impact of annotations on the model output, and relies solely on the solid theoretical basis of mutual information to implicitly account for uncertainty, density, and diversity. \section{Information-Theoretic Active Learning} \label{sec:ital} We begin with a very general description of the idea behind our ITAL approach and then describe its individual components in more detail The implementations of ITAL and the competing methods described in \cref{subsec:competitors} are available as open source at \url{https://github.com/cvjena/ITAL/}. \subsection{Idea and Ideal Objective} \label{subsec:idea} Let $\mathfrak{U} = \{ x_1, \dotsc, x_m \}$ be a set of features of unlabeled samples and $\mathfrak{L} = \{ (x_{m+1}, y_{m+1}), \dotsc, (x_{m+\ell}, y_{m+\ell}) \}$ be a set of features of labeled samples $x_i \in \mathbb{R}^d$ and their labels $y_i \in \{-1,1\}$. The label $1$ is assigned to relevant and $-1$ to irrelevant samples. $\mathfrak{X} = \{x_1,\dotsc,x_m,x_{m+1},\dotsc,x_{m+\ell} \}$ denotes the set of all $n=m+\ell$ samples. In the scenario of content-based image retrieval, $\mathfrak{L}$ usually consists initially of the features of a single relevant sample: the query image provided by the user. However, queries consisting of multiple and even negative examples are possible as well. Intuitively, we want to ask the user for relevance feedback for a batch $u \subseteq \mathfrak{U}$ of $k = |u|$ unlabeled samples, whose feedback we expect to be most helpful for classifying the remaining unlabeled instances, \textit{i.e.}, assessing their relevance to the user. Note that these chosen samples are also often referred to as ``queries'' in the active learning literature. To avoid confusion caused by this conflicting terminology, we will refer to the query image as ``query'' and to the unlabeled samples chosen for annotation as ``candidates''. Ideally, the most informative batch $u$ of candidates can be found by maximizing the conditional mutual information $\mathfrak{I}(R,F \mid u)$ between the relevance $R$ of both labeled and unlabeled samples, which is a multivariate random variable over the space $\{-1,1\}^n$ of relevance labels, and the user feedback $F$, being a multivariate random variable over the space $\{-1,0,1\}^n$ of possible feedbacks. A feedback of $0$ represents the case that the user has not given any feedback for a certain candidate. This option is a special feature of our approach, which allows the user to omit candidates that cannot be labeled reliably. Since the size $n$ of the dataset can be huge, this problem is not solvable in practice. We will show later on how it can be approximated to become tractable. But for now, let us consider the ideal optimization objective: \begin{equation} \label{eq:objective} u = \argmax_{\hat{u} \subseteq \mathfrak{U}} \mathfrak{I}(R,F \mid \hat{u}) \;. \end{equation} Writing the mutual information (MI) in terms of entropy reveals the relationship of our approach to uncertainty sampling by maximizing the entropy $H(R \mid u)$ of the candidate batch \cite{li2013adaptive,zhu2008sud,yang2015usdm} or minimizing the conditional entropy $H(R \mid F, u)$ \cite{guo2007optimistic}: \begin{equation} \label{eq:mi-entropy} \mathfrak{I}(R,F \mid u) = H(R \mid u) - H(R \mid F, u) \;. \end{equation} In contrast to pure uncertainty maximization, we also take into account how the relevance model is expected to change after having obtained the feedback from the user: To select those samples whose annotation would reduce uncertainty the most, we maximize the difference between the uncertainty $H(R \mid u)$ according to the current relevance model and the uncertainty $H(R \mid F, u)$ after an update of the model with the expected user feedback. In contrast to existing works \cite{li2013adaptive,zhu2008sud,yang2015usdm}, we do not assume $H(R \mid u)$ to be equal to the sum of individual entropies of the samples, but use their joint distribution to compute the entropy. Thus, maximizing $H(R \mid u)$ is also a possible novel approach, which will be compared to maximization of MI in the experiments (cf.\ \cref{sec:experiments}). In more detail, the mutual information can be decomposed into the following components (a derivation is provided in \cref{app:mi-derivation}): \begin{multline} \label{eq:mi-integral} \mathfrak{I}(R,F \mid u) = \sum_{\substack{r \in \{-1,1\}^n \\ f \in \{-1,0,1\}^n}} \Biggl[ P(R=r \mid u) \cdot P(F=f \mid R=r, u) \\ \cdot \log\left( \frac{P(R=r \mid F=f, u)}{P(R=r \mid u)} \right) \Biggr] \,. \end{multline} \noindent The individual terms can be interpreted as follows: \begin{itemize} \item $P(R \mid u) = P(R)$ is the probability of a certain relevance configuration according to the current relevance model. \item $P(R \mid F, u)$ is the probability of a certain relevance configuration after updating the relevance model according to the user feedback. Thus, $\frac{P(R \mid F, u)}{P(R \mid u)}$ quantifies the model output change. Compared to other active learning techniques taking model output change into account, \textit{e.g.}, EMOC \cite{freytag2014EMOC}, we do not just consider the change of the predictive mean, but of the joint relevance probability and hence take all parameters of the distributions into account. \item $P(F \mid R, u)$ is the probability of observing a certain feedback, given that the true relevance of the samples is already known. One might assume that the feedback will always be equal to the true relevance. However, users are not perfect and tend to make mistakes or prefer to avoid difficult samples (so-called \textit{unnameable instances} \cite{Kaeding15_ALD}). Thus, this term corresponds to a \textit{user model} predicting the behavior of the user. \end{itemize} \noindent In the following subsections \ref{subsec:rel-model} and \ref{subsec:user-model}, we will first describe our relevance and user model, respectively. Thereafter, we introduce assumptions to approximate \cref{eq:objective} in the subsections \ref{subsec:approx-mi} and \ref{subsec:approx-greedy}, since finding an optimal set of candidates would require exponential computational effort. \subsection{Relevance Model} \label{subsec:rel-model} We fit a probabilistic regression to the training data $\mathfrak{L}$, \textit{i.e.}, the query images and the images annotated so far, using a Gaussian process \cite[chapter 2]{rasmussen2006gaussian} with an RBF kernel given by the kernel matrix $K \in \mathbb{R}^{n \times n}$ over the entire dataset $\mathfrak{X}$: \begin{equation} \label{eq:rbf-kernel} K_{ij} = \sigma_\mathrm{var}^2 \cdot \exp\left(-\frac{\|x_i - x_j\|^2}{2 \cdot \sigma_\mathrm{ls}^2}\right) + \sigma_\mathrm{noise}^2 \cdot \delta_{ij} \;, \end{equation} where $\delta_{ij}$ is the Kronecker delta function with $\delta_{ij} = 1 \leftrightarrow i=j$ and zero otherwise, and $\sigma_\mathrm{var}$, $\sigma_\mathrm{ls}$, and $\sigma_\mathrm{noise}$ are the hyper-parameters of the kernel. The computation of the kernel matrix can be performed off-line in advance and does hence not contribute to the run-time of our active learning method. However, if time and memory required for computing and storing the kernel are an issue, alternative kernels for efficient large-scale Gaussian process inference \cite{rodner2017large} can be used. The prediction of the Gaussian process for any finite set of $k$ samples consists of a multivariate normal distribution $\mathcal{N}(\mu, \Sigma)$ over continuous values $\hat{y} \in \mathbb{R}^k$, where $\mu \in \mathbb{R}^k$ is a vector of predictive means of the samples and $\Sigma \in \mathbb{R}^{k \times k}$ is their predictive joint covariance matrix. Let $p(\hat{y}) = \mathcal{N}(\hat{y} \mid \mu, \Sigma)$ denote the probability density function of such a distribution. We use this probabilistic label regression for binary classification by considering samples $x_i$ with $\hat{y}_i > 0$ as relevant. The probability of a given relevance configuration $r \in \{-1,1\}^n$ for the samples in $\mathfrak{X}$ is hence given by \begin{equation} \label{eq:rel-prob} P(R=r) = \int_{a_1}^{b_1} \dotsi \int_{a_n}^{b_n} p(y_1,\dotsc,y_n) \diff y_n \dotsm \diff y_1 \;, \end{equation} with \begin{equation} \setlength\arraycolsep{0pt} a_i = \left\{\begin{array}{rrrr} 0, & \quad r_i\, &=& \, 1, \\ -\infty, & \quad r_i\, &=& \, -1, \end{array}\right. \qquad b_i = \left\{\begin{array}{rrrr} \infty, & \quad r_i \, &=& \, 1, \\ 0, & \quad r_i \, &=& \, -1, \end{array}\right. \end{equation} for $i = 1, \dotsc, n$. This is a multivariate normal distribution function, which can be efficiently approximated using numerical methods \cite{genz1992mvndst}. The posterior probability $P(R \mid F, u)$ can be obtained in the same way after updating the GP with the expected feedback. For such an update, it is not necessary to re-fit the GP to the extended training data from scratch, which would involve an expensive inversion of the kernel matrix of the training data. Instead, efficient updates of the inverse of the kernel matrix \cite{luetz2013want} can be performed to obtain updated predictions at a low cost. This is an advantage of our GP-based approach compared with other methods relying on logistic regression (\textit{e.g.}, \cite{guo2007optimistic,li2013adaptive}), which requires expensive iterative optimization for updating the model. \subsection{User Model} \label{subsec:user-model} We employ a simple, but plausible user model for $P(F \mid R, u)$, which comes along with a slight simplification of the optimization objective in \cref{eq:objective}: First of all, we assume that if we already know the true relevance $r = [r_1, \dotsc, r_n]^\top$ of all samples, the feedback $f_i$ given by the user for an individual sample $x_i$ is conditionally independent from the feedback provided for the other samples. More formally: \begin{equation} \label{eq:user-model-independency} P(F = f \mid R = r, u) = \prod_{i=1}^{n} P(F_i = f_i \mid R_i = r_i, u) . \end{equation} Clearly, if a sample $x_i$ has not been included in the candidate batch $u$, the user cannot give feedback for that sample, \textit{i.e.}, $x_i \notin u \rightarrow f_i = 0$. Furthermore, we assume that the user will, on average, label a fraction $p_\mathrm{label}$ of the candidate samples. For each labeled sample, the user is assumed to provide an incorrect label with probability $p_\mathrm{mistake}$. In summary, this user model can be formalized as follows: \begingroup \setlength\arraycolsep{3pt} \begin{equation} \label{eq:user-model} P(F_i = f_i \mid R_i = r_i, u) = \left\{\begin{array}{crcl} 0 \,, & x_i \notin u &\wedge & f_i \neq 0, \\ 1 \,, & x_i \notin u &\wedge & f_i = 0, \\ 1 - p_\mathrm{label} \,, & x_i \in u &\wedge & f_i = 0, \\ p_\mathrm{label} \cdot p_\mathrm{mistake} \,, & x_i \in u &\wedge & f_i \neq r_i, \\ p_\mathrm{label} \cdot (1 - p_\mathrm{mistake}) \,, & x_i \in u &\wedge & f_i = r_i. \end{array}\right. \end{equation} \endgroup The fact that $\left(\exists_{i \in \{1,\dotsc,n\}}: x_i \notin u \wedge f_i \neq 0\right) \rightarrow P(F=f \mid R=r, u) = 0$ allows us to adjust the sum in \cref{eq:mi-integral} to run over only $3^k$ instead of $3^n$ possible feedback vectors, where $k \ll n$ is the batch size and independent from the size $n$ of the dataset. This is not an approximation, but an advantage of our user model, that decreases the complexity of the problem significantly. Modeling the user behavior can enable the active learning technique to find a trade-off between learning as fast as possible by asking for feedback for very diverse samples and improving confidence regarding existing knowledge by selecting not extremely diverse, but slightly redundant samples. The latter can be useful for difficult datasets or tasks, where the user is likely to make mistakes or to refuse to give feedback for a significant number of candidates. Nevertheless, the assumption of a perfect user, who labels all samples in the batch and never fails, is an interesting special case since it results in a simplification of the MI term from \cref{eq:mi-integral} and can reduce computation time drastically: \begin{multline} \label{eq:mi-perfectuser} (p_\mathrm{label} = 1 \wedge p_\mathrm{mistake} = 0) \rightarrow \\ \displaystyle \mathfrak{I}(R,F \mid u) = \displaystyle \sum_{r \in \{-1,1\}^n} \Biggl[ P(R=r \mid u) \displaystyle \cdot \log\left( \frac{P(R=r \mid F=r, u)}{P(R=r \mid u)} \right) \Biggr] \,. \end{multline} \subsection{Approximation of Mutual Information} \label{subsec:approx-mi} Even with the perfect user assumption, evaluating \cref{eq:mi-perfectuser} still involves a summation over $2^n$ possible relevance configurations, which does not scale to large datasets. To overcome this issue, we employ an approximation based on the assumption, that the probability of observing a certain relevance configuration depends only on the samples in the current candidate batch: \begin{equation} \label{eq:batch-condition-assumption} P(R = r \mid u = \{x_{i_1}, \dotsc, x_{i_k}\}) \;=\; P(R_{i_1} = r_{i_1}, \dotsc, R_{i_k} = r_{i_k}) \;. \end{equation} This means that we indeed condition $P(R \mid u)$ on the current batch $u$, though actually $P(R \mid u) = P(R)$ holds for the original problem formulation. This assumption allows us to restrict the sum in \cref{eq:mi-integral} to $2^k$ instead of $2^n$ possible relevance configurations, leading to the approximate mutual information \begin{multline} \label{eq:mi-approx} \tilde{\mathfrak{I}}(R,F \mid u) = 2^{n-k} \sum_{\substack{r \in \{-1,1\}^k \\ f \in \{-1,0,1\}^k}} \Biggl[ P(R_u = r) \cdot P(F_u = f \mid R_u = r) \\ \cdot \log\left( \frac{P(R_u = r \mid F_u = f)}{P(R_u = r)} \right) \Biggr] \end{multline} with $u = \{x_{i_1}, \dotsc, x_{i_k}\}$ and, by an abuse of notation, $R_u = [R_{i_1}, \dotsc, R_{i_k}]^\top$ (analogously for $F_u$). The number of involved summands now does not depend on the size of the dataset anymore, but only on the number $k$ of candidates chosen at each round for annotation. On the other hand, this assumption also restricts the estimation of the expected model output change, expressed by the term $P(R_u \mid F_u) / P(R_u)$, to the current batch, which is probably the most severe drawback of this approximation. However, estimating the model output change for the entire dataset would be too expensive and the experiments in \cref{subsec:results} show that our approach can still benefit from the expected model output change. Future work might explore the option of taking a tractable subset of context into account additionally. \subsection{Greedy Batch Construction} \label{subsec:approx-greedy} Although the computational effort required to calculate the approximate MI given in \cref{eq:mi-approx} is independent from the size of the dataset, finding the exact solution to the optimization problem from \cref{eq:objective} would still require computing the MI for all possible $2^m$ candidate batches $u \subseteq \mathfrak{U}$ of unlabeled samples. Since we do not want to confront the user with an unlimited number of candidates anyway, we set the batch size to a fixed number $k$, which leaves us with a number of $\binom{m}{k}$ possible candidate batches. Assessing them all would involve a polynomial number of subsets and is, thus, still too time-consuming in general. On first sight, one might think that this problem can be solved more efficiently using dynamic programming, but this is unfortunately not an option since the MI is not adequately separable. Thus, we follow a linear-time greedy approach to approximate the optimal batch by successively adding samples to the batch \cite{guestrin2005near}, taking their relationship to already selected samples into account: We first select the sample $x_{i_1}$ with maximum $\tilde{\mathfrak{I}}(R, F \mid u = \{x_{i_1}\})$. The second sample $x_{i_2}$ is chosen to maximize MI together with $x_{i_1}$. This continues until the batch contains $k$ samples. At each iteration, the unlabeled samples eligible for being added to the current batch can be treated completely independently from each other, allowing for straightforward parallelization. \section{Experiments} \label{sec:experiments} We demonstrate the performance of our ITAL approach on five image datasets of varying type and structure, described in \cref{subsec:datasets}, and compare it against several existing active learning techniques briefly explained in \cref{subsec:competitors}. The quantitative results in \cref{subsec:results} show that ITAL is the only method that can provide state-of-the-art performance across all datasets. Qualitative examples are shown in \cref{fig:mirflickr-examples} and failure cases can be found in \cref{app:failure-cases}. \subsection{Datasets} \label{subsec:datasets} All datasets used in our experiments consist of multiple classes and are divided into a training and a test set. We define image retrieval tasks for each dataset as follows: Pick a single random instance from the training set of a certain class as query image and consider all other images belonging to that class as relevant, while instances from other classes are irrelevant. Batch-mode active learning is performed for 10 successive rounds with a batch-size of $k = 4$ candidates per round and retrieval performance is evaluated after each round by means of average precision on the test set. This process is repeated multiple times with different random queries for each class and we report the mean average precision (mAP) over all repetitions. Note that our goal is not to achieve state-of-the-art performance in terms of classification accuracy, but with respect to the active learning objective, \textit{i.e.}, obtaining better performance after fewer feedback rounds. The smallest dataset used is the \textbf{Butterflies} dataset \cite{butterflies}, comprising 1,500 images of 5 different species of butterflies captured over a period of 100 years. We use the CNN features provided by the authors and, following their advice, reduce them to 50 dimensions using PCA. A random stratified subset of 20\% of the dataset is used as test set. Second, we use the \textbf{USPS} dataset \cite{USPS} consisting of 9,300 gray-scale images of handwritten digits, scanned from envelopes by the U.S.\ Postal Service. The number of images per class is very unevenly distributed. All images have a size of $16 \times 16$ pixels and are used without further feature extraction as 256-dimensional feature vectors. We use the canonical training-test split provided with the dataset. As a more real-word use-case, we perform evaluation on the \textbf{13 Natural Scenes} dataset \cite{natural_scenes} and the \textbf{MIRFLICKR-25K} dataset \cite{mirflickr}. The former consists of more than 3,400 images from 13 categories of natural scenes such as forests, streets, mountains, coasts, offices, or kitchens. The latter comprises 25,000 images, each assigned to a subset of 14 very general topics such as ``clouds'', ``tree'', ``people'', ``portrait'' etc. Thus, query images can belong to multiple categories and will be ambiguous. Asking the user the right questions is hence of great importance. There are also ``wide-sense annotations'' assigning images to categories if they could be related to a small degree. If a candidate image is annotated in this way, we consider it as unnameable during our simulation. For both datasets, we extracted image features from the first fully-connected layer of the VGG-16 convolutional neural network \cite{simonyan2014vgg}, pre-trained for classification on ImageNet, and reduce their dimensionality to 512 using PCA. Experiments in \cref{app:mirflickr-pca-dim} show that the relative performance of the different methods is not very sensitive w.r.t.\ the dimensionality of the features. 25\% of the natural scenes and 20\% of the MIRFLICKR dataset are used as test set. Finally, we derive further challenging image retrieval tasks from the \textbf{ImageNet Large Scale Visual Recognition Challenge (ILSVRC)} \cite{ILSVRC15}, which comprises more than 1,2 million images from 1,000 classes. Following Freytag et al.\ \cite{freytag2014EMOC}, we obtain binary classification tasks by randomly choosing a single positive and 19 negative classes. This is repeated 25 times and 10 random queries are chosen for each task, leading to a total of 250 image retrieval scenarios. We use the bag-of-words (BoW) features provided with ImageNet instead of CNN features, mainly for two reasons: First, the BoW features are public and hence facilitate reproduction. Second, most neural networks are pre-trained on ImageNet, which could bias the evaluation. The number of random repetitions per class for the natural scenes and the MIRFLICKR dataset has been set to 10 as well, while we use 25 queries per class for USPS and 50 for the butterflies dataset. The features of all datasets were scaled to be in $[0,1]$. \subsection{Competitor Methods} \label{subsec:competitors} We compare ITAL with a variety of baselines and competing methods, including \textbf{SUD} \cite{zhu2008sud}, \textbf{TCAL} \cite{demir2015tcal}, \textbf{RBMAL} \cite{cardoso2017rbmal}, and the method of Brinker \cite{brinker2003incorporating} referred to as ``\textbf{border\_div}'' in the following. All these native BMAL methods have been described in \cref{sec:related-work}. In addition, we evaluate the following successful one-by-one active learning techniques in the BMAL scenario by selecting the $k$ samples with the highest selection scores: Uncertainty sampling for SVM active learning by choosing samples close to the decision boundary (\textbf{border}) \cite{tong2001support}, uncertainty sampling for Gaussian processes (\textbf{unc}) \cite{kapoor2007active}, where uncertainty is defined as the ratio between absolute predictive mean and predictive standard deviation, and sample selection by maximizing the expected model output change (\textbf{EMOC}) \cite{freytag2014EMOC}. All methods have to compete against the baselines of \textbf{random} selection, selecting the \textbf{topscoring} samples with maximum predictive mean, resembling the standard retrieval scenario \cite{ayache2007evaluation}, and variance sampling (\textbf{var}) by maximizing the difference of the sum of variances and the sum of covariances in the batch. Finally, we also investigate maximizing the \textbf{joint entropy} $H(R_{i_1}, \dotsc, R_{i_k}) = - \sum_{r \in \{-1,1\}^k} P(R_u=r) \cdot \log(P(R_u=r))$ of candidate batch $u = \{x_{i_1},\dotsc,x_{i_k}\}$. Being a component of our ITAL method, this is also a novel approach, but lacks the model output change term and the user model (cf.\ \cref{eq:mi-entropy}). We also tried applying \textbf{USDM} \cite{yang2015usdm}, \textbf{MCMI[min]} \cite{guo2007optimistic} and \textbf{AdaptAL} \cite{li2013adaptive}, especially since the latter two also maximize a mutual information criterion. However, all these methods scale so badly to datasets of realistic size, that they could not be applied in practice. AdaptAL, for example, would require 14 hours for composing a single batch on MIRFLICKR, which is clearly intractable. Our ITAL method, in contrast, can handle this dataset with less than a minute per batch. These three competitors could, thus, only applied to USPS, MIRFLICKR, and ImageNet by randomly sub-sampling 1000 candidates to choose from, as suggested by Li et al.\ \cite{li2013adaptive}. This usually leads to a degradation of performance, as can be seen from the results reported in \cref{app:adaptal}. \subsection{Hyper-parameters} \label{subsec:hyper-params} The hyper-parameters of the RBF kernel, \textit{i.e.}, $\sigma_\mathrm{ls}$, $\sigma_\mathrm{var}$, and $\sigma_\mathrm{noise}$ (cf. \cref{subsec:rel-model}), potentially have a large impact on the performance of the active learning methods. However, the overall goal is to eventually obtain a classifier that performs as well as possible. Therefore, we determine the optimal kernel hyper-parameters for each dataset using tenfold cross-validation on the training set and alternating optimization to maximize mean average precision. This optimization aims only for good classification performance independent of the active learning method being used and the same hyper-parameters are used for all methods. With regard to the hyper-parameters of the user model used by ITAL, we employ the perfect user assumption for being comparable to competing methods that do not model the user. An experiment evaluating the effect of different user model parameters is presented in \cref{subsec:exp-user-model}. In case that other methods have further hyper-para\-me\-ters, we use the default values provided by their authors. \subsection{Results} \label{subsec:results} \Cref{fig:performance} depicts the average precision obtained on average after 10 feedback rounds using the different BMAL methods. On the Butterflies dataset, ITAL obtains perfect performance after the least number of feedback rounds. TCAL and border\_div perform similar to ITAL on USPS, but ITAL learns faster at the beginning, which is important in interactive image retrieval scenarios. While sampling candidates based on batch entropy behaves almost identical to ITAL on Butterflies, USPS, and MIRFLICKR, it is slightly superior on the Natural Scenes dataset, but fails to improve after more than 4 rounds of feedback on the ImageNet benchmark, where ITAL is clearly superior to all competitor methods. This indicates that taking the effect of the expected user feedback on the model output change into account is of great benefit for datasets as diverse as ImageNet. \begin{figure}[t] \includegraphics[width=\linewidth]{figures/italia_performance} \caption{Comparison of retrieval performance after different numbers of feedback rounds for various active learning methods. The thick, orange line corresponds to our proposed method. Figure is best viewed in color.} \label{fig:performance} \end{figure} Since it is often desirable to compare different methods by means of a single value, we report the area under the learning curves (AULC) in \cref{table:auc}, divided by the number of feedback rounds so that the best possible value is always 1.0. In all cases, our method is among the top performers, achieving the best of all results in 3 out of 5 cases. The improvement over the second-best method on 13 Natural Scenes and ImageNet is significant on a level of \textless1\% and on a level of 7\% on USPS, according to Student's paired t-test. \begin{table}[t] \caption{Area under the learning curves from \cref{fig:performance}. The numbers in parentheses indicate the position in the ranking of all methods. The best value in each column is set in bold face, while the second-best and third-best values are underlined.} \label{table:auc} \begin{scriptsize} \setlength\tabcolsep{1pt} \begin{tabularx}{\linewidth}{lXccXccXccXccXccXr} \toprule && \multicolumn{14}{c}{Area under Learning Curve (AULC)} && \\ \cmidrule{3-16} Method && \multicolumn{2}{c}{Butterflies} && \multicolumn{2}{c}{USPS} && \multicolumn{2}{c}{Nat.\ Scenes} && \multicolumn{2}{c}{MIRFLICKR} && \multicolumn{2}{c}{ImageNet} && Avg. Rank \\ \midrule random && 0.7316 & (8) && 0.5416 & (8) && 0.5687 & (8) && 0.4099 & (9) && 0.1494 & (9) && 8.4 \\ topscoring && 0.5991 & (11) && 0.5289 & (9) && 0.5419 & (9) && 0.4358 & (7) && 0.1708 & (7) && 8.6 \\ var && 0.6800 & (9) && 0.5550 & (7) && 0.5831 & (5) && 0.3957 & (10) && 0.1383 & (11) && 8.4 \\ border && 0.7434 & (6) && 0.6393 & (5) && 0.5775 & (7) && 0.4559 & (6) && 0.1743 & (4) && 5.6 \\ border\_div && 0.7456 & (5) && \underline{0.6465} & (3) && \underline{0.6031} & (3) && \textbf{0.4795} & (1) && \underline{0.1791} & (3) && \underline{3.0} \\ unc && 0.7373 & (7) && 0.6391 & (6) && 0.5793 & (6) && 0.4585 & (5) && 0.1725 & (5) && 5.8 \\ EMOC && \underline{0.7561} & (2) && 0.4723 & (10) && 0.4654 & (11) && 0.4357 & (8) && 0.1483 & (10) && 8.2 \\ SUD && 0.3887 & (12) && 0.3903 & (12) && 0.3766 & (12) && 0.3883 & (11) && 0.1626 & (8) && 11.0 \\ TCAL && \textbf{0.7720} & (1) && 0.6459 & (4) && 0.6016 & (4) && 0.4688 & (4) && 0.1708 & (6) && 3.8 \\ RBMAL && 0.6023 & (10) && 0.4457 & (11) && 0.5046 & (10) && 0.3732 & (12) && 0.1356 & (12) && 11.0 \\ \midrule[.03em] entropy (ours) && \underline{0.7512} & (3) && \underline{0.6484} & (2) && \textbf{0.6547} & (1) && \underline{0.4703} & (3) && \underline{0.1793} & (2) && \underline{2.2} \\ ITAL (ours) && 0.7511 & (4) && \textbf{0.6522} & (1) && \underline{0.6233} & (2) && \underline{0.4731} & (2) && \textbf{0.1841} & (1) && \textbf{2.0} \\ \bottomrule \end{tabularx} \end{scriptsize} \end{table} The performance of the competing methods, on the other hand, varies significantly across datasets. ITAL, in contrast, is not affected by this issue and provides state-of-the-art performance independent from the characteristics of the data. To make this more visible, we construct a ranking of the tested methods for each dataset and report the average rank in \cref{table:auc} as well. ITAL achieves the best average rank and can thus be considered most universally applicable. This is of high importance because, in an active learning scenario, labeled data is usually not available before performing the active learning. Thus, adaptation of the AL method or selection of a suitable one depending on the dataset is difficult. A widely applicable method such as ITAL is hence very desirable. \subsection{Effect of the User Model} \label{subsec:exp-user-model} To evaluate the effect of the user model integrated into ITAL, we simulated several types of users behaviors on the 13 Natural Scenes dataset: \begin{enumerate*}[label=\itshape\alph*\upshape)] \item\label{item:aggressive-user} an aggressive user annotating all images but assigning a wrong label in 50\% of the cases, \item\label{item:conservative-user} a conservative user who always provides correct labels but only annotates 25\% of the images on average, and \item\label{item:realistic-user} a blend of both, labeling 50\% of the candidate images on average and having a 25\% chance of making an incorrect annotation. \end{enumerate*} The same active learning methods as in the previous sections are applied and the parameters $p_\mathrm{label}$ and $p_\mathrm{mistake}$ of ITAL are set accordingly. The results presented in \cref{app:imperfect-users} show that the user model helps ITAL to make faster improvements than with the perfect user model. A possible reason for this effect is that the parameters of the user model control an implicit trade-off between diversity and redundancy used by ITAL: For an imperfect user, selecting samples for annotation more redundantly can help to reduce the impact of wrong annotations. However, ITAL performs reasonably well even with the perfect user assumption. This could hence be used to speed-up ITAL noticeably with only a minor loss of performance. \section{Conclusions} \label{sec:conclusions} We have proposed information-theoretic active learning (ITAL), a novel batch-mode active learning technique for binary classification, and applied it successfully to image retrieval with relevance feedback. Based on the idea of finding a subset of unlabeled samples that maximizes the mutual information between the relevance model and the expected user feedback, we propose suitable models and approximations to make this NP-hard problem tractable in practice. ITAL does not need to rely on manually tuned combinations of different heuristics, as many other works on batch-mode active learning do, but implicitly trades off uncertainty against diversity by taking the joint relevance distribution of the instances in the dataset into account. Our method also features an explicit user model that enables it to deal with unnameable instances and the possibility of incorrect annotations. This has been demonstrated to be beneficial in the case of unreliable users. We evaluated our method on five image datasets and found that it provides state-of-the-art performance across datasets, while many competitors perform well on certain datasets only. Moreover, ITAL outperforms existing techniques on the ImageNet dataset, which we attribute to its ability of taking the effect of the expected user feedback on the model output change into account. \section*{Acknowledgements} This work was supported by the German Research Foundation as part of the priority programme ``Volunteered Geographic Information: Interpretation, Visualisation and Social Computing'' (SPP 1894, contract number DE 735/11-1). \bibliographystyle{splncs04} \section{Performance of MCMI[min] and AdaptAL} \label{app:adaptal} Since MCMI[min] \cite{guo2007optimistic} and AdaptAL \cite{li2013adaptive} also maximize a mutual information criterion and are, thus, similar to our method, we also tried to apply those methods to our benchmark datasets. Even though we replaced the expensive logistic regression with Gaussian process inference for being comparable to our method, they could only be applied to the Butterflies and 13 Natural Scenes dataset within reasonable time. For the remaining 3 datasets, we randomly sub-sampled 1000 candidates from the entire dataset, as suggested by \cite{li2013adaptive}. \begin{table} \caption{Comparison of ITAL with MCMI[min] and AdaptAL in terms of AULC.} \label{tbl:mi-comp} \setlength\tabcolsep{6pt} \begin{tabularx}{\linewidth}{Xccccc} \toprule Method & Butterflies & USPS & Nat.\ Scenes & MIRFLICKR & ImageNet \\ \midrule random & 0.7316 & 0.5416 & 0.5687 & 0.4099 & 0.1494 \\ MCMI[min] & 0.6846 & 0.5293 & 0.4554 & 0.4087 & 0.1413 \\ AdaptAL & \textbf{0.7716} & 0.6487 & 0.6424 & 0.4643 & 0.1746 \\ \midrule entropy (ours) & 0.7512 & 0.6484 & \textbf{0.6547} & 0.4703 & 0.1793 \\ ITAL (ours) & 0.7511 & \textbf{0.6522} & 0.6233 & \textbf{0.4731} & \textbf{0.1841} \\ \bottomrule \end{tabularx} \end{table} The results in \cref{tbl:mi-comp} show that MCMI[min] does not work well in a batch-mode scenario and performs worse than random. AdaptAL, on the other hand, is the top performer on the Butterflies dataset and the second-best method on Natural Scenes, directly behind our batch-entropy approach. These are the two datasets where it could be applied in reasonable time on the entire dataset. The sub-sampling that is necessary on the remaining three datasets, however, negatively impacts performance, especially on ImageNet. To the best of our knowledge, our method is the first one that makes an information-theoretic approach to batch-mode active learning applicable in realistic scenarios without sub-sampling the dataset. \newpage \section{Simulation of Imperfect Users} \label{app:imperfect-users} As described in section 4.5 of the paper, we have investigated the effect of three different extreme user behavior models on the performance of the tested BMAL methods. With regard to our approach, we have evaluated both ITAL with the user model parameters $p_\mathrm{label}$ and $p_\mathrm{mistake}$ set according to the simulated user and ITAL with the perfect user assumption, which is faster. We have selected batches of 4 images for annotation at each round. \begin{figure} \includegraphics[width=\linewidth]{figures/user-behaviors} \caption{Comparison of different user behavior models on Natural Scenes.} \label{fig:user-behaviors} \end{figure} As expected, all methods suffer from imperfect user feedback compared to a perfect user. While an adequate user model helps ITAL to learn faster during the first rounds, the difference is small enough to justify the use of the perfect user assumption even if it is not true in order to gain a significant speed-up. The case of overly aggressive but error-prone users obviously cannot be handled by the active learning method alone, but also requires adequate handling of such scenarios by the classifier. \newpage \section{Sensitivity of Results regarding Feature Dimensionality} \label{app:mirflickr-pca-dim} To assess to which extent the results presented in the paper are affected by certain transformations applied to the features, we experimented with different dimensionalities of the feature space on the MIRFLICKR dataset. To this end, we have applied PCA to the features extracted from the first fully-connected layer of VGG16, which comprise 4096 dimensions, and projected them onto spaces with 64, 128, 256, 512, and 1024 features. Experiments with all BMAL methods have been conducted on those features for 10 rounds of user feedback and the area under the learning curve (AULC) for the various dimensionalities is reported in \cref{fig:mirflickr-pca-dim}. \begin{figure} \centering \includegraphics[width=.75\linewidth]{figures/mirflickr_pca} \caption{Area under Learning Curve (AULC) of various BMAL methods on MIRFLICKR with varying feature dimensionality.} \label{fig:mirflickr-pca-dim} \end{figure} The results show that the relative performance of the different methods compared to each other is largely insensitive to the number of features. The performance of ITAL is stable up to as few as 128 dimensions, while some other methods such as TCAL and EMOC already degrade after reducing the number of features to less than 256. When using 1024 features, ITAL is even able to catch up to border\_div, which is the best performing method on this particular dataset. However, we have used 512 features for our experiments in the paper due to the increased computational cost incurred by higher-dimensional feature spaces. \newpage \section{Examples for Failure Cases} \label{app:failure-cases} \begin{figure}[h] \centering \includegraphics[width=\linewidth]{figures/mirflickr_failures} \caption{Four queries from MIRFLICKR where ITAL performed worst.} \label{fig:failure-cases} \end{figure} To analyze the possible shortcomings of our method, we have picked four queries from the MIRFLICKR dataset where ITAL had the worst AULC score. These are depicted in \cref{fig:failure-cases}, along with the candidate images selected for annotation over 4 rounds of feedback and the top results retrieved by the relevance model after each round. The first query could be interpreted in multiple ways: The user could be searching for images of people, of babies, or of adults with babies. All these options are covered by the candidate images selected by ITAL. Only one of those image shows a baby alone, which is the actual search objective in this example. That image, however, has not been annotated confidently as showing a baby in the MIRFLICKR dataset, so that it remains unnameable here. The second query shows a swarm of birds on a power pole, but the simulated user actually searches for birds. The features used in our experiment are apparently not sufficient to capture the semantics of this image well enough for recognizing that it is about birds. Thus, the selected candidates do not contain any image of a bird in a different scene and the classifier cannot abstract away from power poles. The ``night'' query, on the other hand, is again an example of erroneous annotations in the dataset: Several images of night scenes have been selected as candidates, but have been annotated either as unnameable or even as irrelevant. Finally, the last query image shows a river and the candidates are actually quite suitable to identify whether the user is more interested in mountain scenes, water scenes, river scenes, or natural scenes in general. However, either the features or the small number of annotated images seem to be insufficient in this case for distinguishing between rivers and other bodies of water. \clearpage \section{Derivation of Eq. (3)} \label{app:mi-derivation} Plugging in the definitions of entropy and conditional entropy into the definition of mutual information given in eq. (2) leads to the following: \begin{multline*} \mathfrak{I}(R,F \mid u) = - \Biggl[ \sum_{r \in \{-1,1\}^n} P(R=r \mid u) \cdot \log P(R=r \mid u) \Biggr] + \\ \Biggl[ \sum_{\substack{r \in \{-1,1\}^n \\ f \in \{-1,0,1\}^n}} P(F=f \mid u) \cdot P(R=r \mid F=f, u) \cdot \log P(R=r \mid F=f, u) \Biggr] \;. \end{multline*} \noindent Expressing $P(R=r \mid u)$ in the first sum as the marginalization \begin{equation*} P(R=r \mid u) = \sum_{f \in \{-1,0,1\}^n} P(F=f \mid u) \cdot P(R=r \mid F=f, u) \end{equation*} allows us to merge the two sums: \begin{multline*} \mathfrak{I}(R,F \mid u) = \sum_{\substack{r \in \{-1,1\}^n \\ f \in \{-1,0,1\}^n}} \Biggl[ P(F=f \mid u) \cdot P(R=r \mid F=f, u) \\ \cdot \log \left( \frac{P(R=r \mid F=f, u)}{P(R=r \mid u)} \right) \;. \Biggr] \end{multline*} \noindent Using Bayes' Theorem we can substitute \begin{equation*} P(F=f \mid u) \cdot P(R=r \mid F=f, u) = P(R=r \mid u) \cdot P(F=f \mid R=r, u) \;, \end{equation*} finally leading to eq. (3) from the main paper. \end{appendices}

d06c8152cf4f7eabee9ab51a88e1e665be69d006

\section{Introduction} Single-channel speech enhancement aims at attenuating the noise component of noisy speech to increase the intelligibility and perceived quality of the speech component \cite{loizou2013speech}. It is commonly used to improve the quality of mobile speech communication in noisy environments and enhance the speech signal before amplification in hearing aids and cochlear implants. More importantly, speech enhancement is widely applied as a front-end pre-processing stage to improve the performance of automatic speech recognition (ASR) \cite{hinton2012deep, jaitly2012application, sainath2011making, deng2013recent, yu2017recent} and speaker recognition under noisy conditions \cite{Li14overview, Li15robust}. With the advance of deep learning, deep neural network (DNN) based approaches have achieved great success in single-channel speech enhancement. The mask learning approach \cite{narayanan2013ideal, wang2014training, weninger2015speech} is proposed to estimate the ideal ratio mask or ideal binary mask based on noisy input features using a DNN. The mask is used to filter out the noise from the noisy speech and recover the clean speech. However, it has the presumption that the scale of the masked signal is the same as the clean target and the noise is strictly additive and removable by the masking procedure which is generally not true for real recorded stereo data. To deal with this problem, the feature-mapping approach \cite{xu2015regression, lu2013speech, maas2012recurrent, feng2014speech, weninger2014single, chen2017improving} is proposed to train a feature-mapping network that directly transforms the noisy features to enhanced ones. The feature-mapping network serves as a non-linear regression function trained to minimize the feature-mapping loss, i.e., the mean square error (MSE) between the enhanced features and the paralleled clean ones. The application of MSE estimator is based on the homoscedasticity and no auto-correlation assumption of the noise, i.e., the noise needs to have the same variance for each noisy feature and the noise needs to be uncorrelated between different noisy features \cite{freedman2009statistical}. This assumption is in general violated for real speech signal (a kind of time series data) under non-stationary unknown noise. Recently, adversarial training \cite{gan} has become a hot topic in deep learning with its great success in estimating generative models. It was first applied to image generation \cite{radford2015unsupervised, denton2015deep}, image-to-image translation \cite{isola2017imagetoimage, zhu2017unpaired} and representation learning \cite{chen2016infogan}. In speech area, it has been applied to speech enhancement \cite{pascual2017segan, donahue2017exploring, mimura2017cross, meng2018cycle}, voice conversion \cite{kaneko2017parallel, hsu2017voice}, acoustic model adaptation \cite{sun2017unsupervised, meng2017unsupervised, meng2018adversarial}, noise-robust \cite{grl_shinohara, grl_serdyuk} and speaker-invariant \cite{saon2017english, meng2018speaker} ASR using gradient reversal layer (GRL) \cite{ganin2015unsupervised}. In these works, adversarial training is used to learn a feature or an intermediate representation in DNN that is invariant to the shift among different domains (e.g., environments, speakers, image styles, etc.). In other words, a generator network is trained to map data from different domains to the features with similar distributions via adversarial learning. Inspired by this, we advance the feature-mapping approach with adversarial learning to further diminish the discrepancy between the distributions of the clean features and the enhanced features generated by the feature-mapping network given non-stationary and auto-correlated noise at the input. We call this method adversarial feature-mapping (AFM) for speech enhancement. In AFM, an additional discriminator network is introduced to distinguish the enhanced features from the real clean ones. The feature-mapping network and the discriminator network are jointly trained to minimize the feature-mapping loss and simultaneously mini-maximize the discrimination loss with adversarial multi-task learning. With AFM, the feature-mapping network can generate pseudo-clean features that the discriminator can hardly tell whether they are real clean features or not. To achieve better performance on ASR task, senone-aware adversarial feature-mapping (SA-AFM) is proposed in which an acoustic model network is introduced and is jointly trained with the feature-mapping and discriminator networks to optimize the senone classification loss in addition to the feature-mapping and discrimination losses. Note that AFM is different from \cite{donahue2017exploring} in that: (1) In AFM, the inputs to the discriminator are enhanced and clean features while in \cite{donahue2017exploring} the inputs to the discriminator are the concatenation of enhanced and noisy features and the concatenation of clean and noisy features. (2) The primary task of AFM is feature-mapping, i.e., to minimize the $L_2$ distance (MSE) between enhanced and clean features and it is advanced with adversarial learning to further reduce the discrepancy between the distributions of the enhanced and clean features while in \cite{donahue2017exploring} the primary task is to generate enhanced features that are similar to clean features with generative adversarial network (GAN) and it is regularized with the minimization of $L_1$ distance between noisy and enhanced features. (3) AFM performs adversarial multi-task training using GRL method as in \cite{ganin2015unsupervised} while \cite{donahue2017exploring} conducts conditional GAN iterative optimization as in \cite{gan}. (4) In this paper, AFM uses long short-term memory (LSTM)-recurrent neural networks (RNNs) \cite{sak2014long, meng2017deep, erdogan2016multi} to generate the enhanced features and a feed-forward DNN as the discriminator while \cite{donahue2017exploring} uses convolutional neural networks for both. We perform ASR experiments with features enhanced by AFM on CHiME-3 dataset \cite{barker2015third}. Evaluated on a clean acoustic model, AFM achieves 16.95\% and 5.27\% relative word error rate (WER) improvements respectively over the noisy features and feature-mapping baseline and the SA-AFM achieves 9.85\% relative WER improvement over the multi-conditional acoustic model. \section{Adversarial Feature-Mapping Speech Enhancement} \label{sec:cse} With feature-mapping approach for speech enhancement, we are given a sequence of noisy speech features $X=\{x_1, \ldots, x_T\}$ and a sequence of clean speech features $Y=\{y_1, \ldots, y_T\}$ as the training data. $X$ and $Y$ are \emph{parallel} to each other, i.e., each pair of $x_i$ and $y_i$ is frame-by-frame synchronized. The goal of speech enhancement is to learn a non-linear feature-mapping network $F$ with parameters $\theta_f$ that transforms $X$ to a sequence of enhanced features $\hat{Y}=\{\hat{y}_1, \ldots, \hat{y}_T\}$ such that the distribution of $\hat{Y}$ is as close to that of $Y$ as possible: \begin{align} & \hat{y_i} = F(x_i), \quad i = 1, \ldots, T \\ & P_{\hat{Y}}(\hat{y}) \rightarrow P_Y(y). \label{eqn:fm} \end{align} To achieve that, we minimize the noisy-to-clean feature-mapping loss $\mathcal{L}_{F}(\theta_f)$, which is commonly defined as the MSE between $\hat{Y}$ and $Y$ as follows: \begin{align} \mathcal{L}_{F}(\theta_f) &= \frac{1}{T}\sum_{i=1}^T (\hat{y}_i -y_i)^2 = \frac{1}{T}\sum_{i=1}^T \left[F(x_i) - y_i\right]^2. \label{eqn:loss_f} \end{align} However, the MSE that feature-mapping approach minimizes is based on the homoscedasticity and no auto-correlation assumption of the noise, i.e., the noise has the same variance for each noisy feature and the noise is uncorrelated between different noisy features. This assumption is in general invalid for real speech signal (time series data) under non-stationary unknown noise. To address this problem, we further advance the feature-mapping network with an additional discriminator network and perform adversarial multi-task training to further reduce the discrepancy between the distribution of enhanced features and the clean ones given non-stationary and auto-correlated noise is at the input. \begin{figure}[htpb!] \centering \includegraphics[width=0.65\columnwidth]{afm.png} \vspace{-0.0cm} \caption{\small The framework of AFM for speech enhancement. } \label{fig:afm} \end{figure} As shown in Fig. \ref{fig:afm}, the discriminator network $D$ with parameters $\theta_d$ takes enhanced features $\hat{Y}$ and clean features $Y$ as the input and outputs the posterior probability that an input feature belongs to the clean set, i.e., \begin{align} P(y_i \in \mathbb{C})& = D(y_i) \\ P(\hat{y}_i \in \mathbb{E})& = 1 - D(\hat{y}_i) \label{eqn:discriminator} \end{align} where $\mathbb{C}$ and $\mathbb{E}$ denote the sets of clean and enhanced features respectively. The discrimination losses $\mathcal{L}_{D}(\theta_f, \theta_d)$ for the $D$ is formulated below using cross-entropy: \begin{align} & \mathcal{L}_{D}(\theta_f, \theta_d) = \frac{1}{T}\sum_{i = 1}^{T} \left[ \log P(y_i \in \mathbb{C}) + \log P(\hat{y}_i \in \mathbb{E}) \right] \nonumber \\ & = \frac{1}{T}\sum_{i = 1}^{T} \log D(y_i) + \log \left[1 - D(F(x_i))\right]. \label{eqn:loss_d} \end{align} To make the distribution of the enhanced features $\hat{Y}$ similar to that of the clean ones $Y$, we perform adversarial training of $F$ and $D$, i.e, we minimize $\mathcal{L}_{D}(\theta_f, \theta_d)$ with respect to $\theta_d$ and maximize $\mathcal{L}_{D}(\theta_f, \theta_d)$ with respect to $\theta_f$. This minimax competition will first increase the generation capability of $F$ and the discrimination capability of $D$ and will eventually converge to the point where the $F$ generates extremely confusing enhanced features that $D$ is unable to distinguish whether it is a clean feature or not. The total loss of AFM $\mathcal{L}_{AFM}(\theta_f, \theta_d)$ is formulated as the weighted sum of the feature-mapping loss and the discrimination loss below: \begin{align} &\mathcal{L}_{\text{AFM}}(\theta_f, \theta_d) = \mathcal{L}_{F}(\theta_f) - \lambda \mathcal{L}_{D}(\theta_f, \theta_d) \label{eqn:loss_afm} \end{align} where $\lambda > 0$ is the gradient reversal coefficient that controls the trade-off between the feature-mapping loss and the discrimination loss in Eq. \eqref{eqn:loss_f} and Eq. \eqref{eqn:loss_d} respectively. $F$ and $D$ are jointly trained to optimize the total loss through adversarial multi-task learning as follows: \begin{align} \hat{\theta}_f = \argmin_{\theta_f} \mathcal{L}_{\text{AFM}}(\theta_f, \hat{\theta}_d) \label{eqn:min_f} \\ \hat{\theta}_d = \argmax_{\theta_d} \mathcal{L}_{\text{AFM}}(\hat{\theta}_f, \theta_d) \label{eqn:max_d} \end{align} where $\hat{\theta}_f$ and $\hat{\theta}_d$ are optimal parameters for $F$ and $D$ respectively and are updated as follows via back propagation through time (BPTT) with stochastic gradient descent (SGD): \begin{align} & \theta_f \leftarrow \theta_f - \mu \left[ \frac{\partial \mathcal{L}_{\text{F}}(\theta_f)}{\partial \theta_f} - \lambda \frac{\partial \mathcal{L}_{\text{D}}(\theta_f, \theta_d)}{\partial \theta_f} \right] \label{eqn:grad_f} \\ & \theta_d \leftarrow \theta_d - \mu \frac{\partial \mathcal{L}_{\text{D}}(\theta_f, \theta_d)}{\partial \theta_d} \label{eqn:grad_d} \end{align} where $\mu$ is the learning rate. Note that the negative coefficient $-\lambda$ in Eq. \eqref{eqn:grad_f} induces reversed gradient that maximizes $\mathcal{L}_{D}(\theta_f, \theta_d)$ in Eq. \eqref{eqn:loss_d} and makes the enhanced features similar to the real clean ones. Without the reversal gradient, SGD would make the enhanced features different from the clean ones in order to minimize Eq. \eqref{eqn:loss_d}. For easy implementation, gradient reversal layer is introduced in \cite{ganin2015unsupervised}, which acts as an identity transform in the forward propagation and multiplies the gradient by $-\lambda$ during the backward propagation. During testing, only the optimized feature-mapping network $F$ is used to generate the enhanced features given the noisy test features. \section{Senone-Aware Adversarial Feature-Mapping Enhancement} \label{sec:saafm} For AFM speech enhancement, we only need parallel clean and noisy speech for training and we do not need any information about the content of the speech, i.e., the transcription. With the goal of improving the intelligibility and perceived quality of the speech, AFM can be widely used in a broad range of applications including ASR, mobile communication, hearing aids, cochlear implants, etc. However, for the most important ASR task, AFM does not necessarily lead to the best WER performance because its feature-mapping and discrimination objectives are not directly related to the speech units (i.e., word, phoneme, senone, etc.) classification. In fact, with AFM, some decision boundaries among speech units may be distorted in searching for an optimal separation between speech and noise. To compensate for this mismatch, we incorporate a DNN acoustic model into the AFM framework and propose the senone-aware adversarial feature-mapping (SA-AFM), in which the acoustic model network $M$, feature-mapping network $F$ and the discriminator network $D$ are trained to jointly optimize the primary task of feature-mapping, secondary task of the third task of clean/enhanced data discrimination and the third task of senone classification in an adversarial fashion. The transcription of the parallel clean and noisy training utterances is required for SA-AFM speech enhancement. Specifically, as shown in Fig. \ref{fig:saafm}, the acoustic model network $M$ with parameters $\theta_m$ takes in the enhanced features $\hat{Y}$ as the input and predicts the senone posteriors $P(q|\hat{y}_i; \theta_y), q\in \mathcal{Q}$ as follows: \begin{align} M(\hat{y}_i) = P(q | \hat{y}_i; \theta_m)， \label{eqn:senone_classify_1} \end{align} after the integration with feature-mapping network $F$, we have \begin{align} M(F(x_i)) = P(q | x_i; \theta_f, \theta_m). \label{eqn:senone_classify_2} \end{align} \begin{figure}[htpb!] \centering \includegraphics[width=0.8\columnwidth]{saafm.png} \vspace{-0.0cm} \caption{\small The framework of SA-AFM for speech enhancement. } \label{fig:saafm} \vspace{-20pt} \end{figure} We want to make the enhanced features $\hat{Y}$ senone-discriminative by minimizing the cross-entropy loss between the predicted senone posteriors and the senone labels as follows: \begin{align} \mathcal{L}_{\text{M}}(\theta_m, \theta_f) & = - \frac{1}{T}\sum_{i = 1}^{T} \log P(s_i | x_i;\theta_f, \theta_m) \nonumber \\ & = -\frac{1}{T}\sum_{i = 1}^{T} \log M(F(x_i)) \label{eqn:loss_m} \end{align} where $S$ is a sequence of senone labels $S=\{s_{1},\ldots, s_{T}\}$ aligned with the noisy data $X$ and enhanced data $\hat{Y}$. Simultaneously, we minimize feature-mapping loss $\mathcal{L}_{F}(\theta_f)$ defined in Eq. \eqref{eqn:loss_f} with respect to $F$ and perform adversarial training of $F$ and $D$, i.e, we minimize $\mathcal{L}_{D}(\theta_f, \theta_d)$ defined in Eq. \eqref{eqn:loss_d} with respect to $\theta_d$ and maximize $\mathcal{L}_{D}(\theta_f, \theta_d)$ with respect to $\theta_f$, to make the distribution of the enhanced features $\hat{Y}$ similar to that of the clean ones $Y$. The total loss of SA-AFM $\mathcal{L}_{\text{SA-AFM}}(\theta_m, \theta_f, \theta_d)$ is formulated as the weighted sum of $\mathcal{L}_{F}(\theta_f)$, $\mathcal{L}_{D}(\theta_f, \theta_d)$ and the senone classification loss $\mathcal{L}_{\text{M}}(\theta_m, \theta_f)$ as follows: \begin{align} \mathcal{L}_{\text{SA-AFM}}(\theta_f, \theta_d, \theta_m) = & \mathcal{L}_{F}(\theta_f)- \lambda_1 \mathcal{L}_{D}(\theta_f, \theta_d) \nonumber \\ & + \lambda_2 \mathcal{L}_{M}(\theta_f, \theta_m) \label{eqn:loss_saafm} \end{align} where $\lambda_1 > 0$ is the gradient reversal coefficient that controls the trade-off between $\mathcal{L}_{F}(\theta_f)$ and $\mathcal{L}_{D}(\theta_f, \theta_d)$, and $\lambda_2 > 0$ is the weight for $\mathcal{L}_{M}(\theta_f, \theta_m)$. $F$, $D$ and $M$ are jointly trained to optimize the total loss through adversarial multi-task learning as follows: \begin{align} (\hat{\theta}_f, \hat{\theta}_m) = \argmin_{\theta_f, \theta_m} \mathcal{L}_{\text{SA-AFM}}(\theta_f, \hat{\theta}_d, \theta_m) \label{eqn:saafm_fm} \\ \hat{\theta}_d = \argmax_{\theta_d} \mathcal{L}_{\text{SA-AFM}}(\hat{\theta}_f, \theta_d, \hat{\theta}_m) \label{eqn:saafm_d} \end{align} where $\hat{\theta}_f$, $\hat{\theta}_d$ and $\hat{\theta}_m$ are optimal parameters for $F$, $D$ and $M$ respectively and are updated as follows via BPTT with SGD as in Eq. \eqref{eqn:grad_f}, Eq. \eqref{eqn:grad_d} and Eq. \eqref{eqn:grad_m} below: \begin{align} \theta_m \leftarrow \theta_m - \mu \frac{\partial \mathcal{L}_{\text{M}}(\theta_f, \theta_m)}{\partial \theta_m}. \label{eqn:grad_m} \end{align} During decoding, only the optimized feature-mapping network $F$ and acoustic model network $M$ are used to take in the noisy test features and generate the acoustic scores. \section{Experiments} In the experiments, we train the feature-mapping network $F$ with the parallel clean and noisy training utterances in CHiME-3 dataset \cite{barker2015third} using different methods. The real far-field noisy speech from the 5th microphone channel in CHiME-3 development data set is used for testing. We use a pre-trained clean DNN acoustic model to evaluate the ASR WER performance of the test features enhanced by $F$. The standard WSJ 3-gram language model with 5K-word lexicon is used in our experiments. \subsection{Feedforward DNN Acoustic Model} \label{sec:dnn_am} To evaluate the ASR performance of the features enhanced by AFM, we first train a feedforward DNN-hidden Markov model (HMM) acoustic model using 8738 clean training utterances in CHiME-3 with cross-entropy criterion. The 29-dimensional log Mel filterbank (LFB) features together with 1st and 2nd order delta features (totally 87-dimensional) are extracted. Each feature frame is spliced together with 5 left and 5 right context frames to form a 957-dimensional feature. The spliced features are fed as the input of the feed-forward DNN after global mean and variance normalization. The DNN has 7 hidden layers with 2048 hidden units for each layer. The output layer of the DNN has 3012 output units corresponding to 3012 senone labels. Senone-level forced alignment of the clean data is generated using a Gaussian mixture model-HMM system. A WER of 29.44\% is achieved when evaluating the clean DNN acoustic model on the test data. \subsection{Adversarial Feature-Mapping Speech Enhancement} \label{sec:exp_afm} We use parallel data consisting of 8738 pairs of noisy and clean utterances in CHiME-3 as the training data. The 29-dimensional LFB features are extracted for the training data. For the noisy data, the 29-dimensional LFB features are appended with 1st and 2nd order delta features to form 87-dimensional feature vectors. $F$ is an LSTM-RNN with 2 hidden layers and 512 units for each hidden layer. A 256-dimensional projection layer is inserted on top of each hidden layer to reduce the number of parameters. $F$ has 87 input units and 29 output units. The features are globally mean and variance normalized before fed into $F$. The discriminator $D$ is a feedforward DNN with 2 hidden layers and 512 units in each hidden layer. $D$ has 29 input units and one output unit. We first train $F$ with 87-dimensional LFB features as the input and 29-dimensional LFB features as the target to minimize the feature-mapping loss $\mathcal{L}(\theta_f)$ in Eq. \eqref{eqn:loss_f}. This serves as the feature-mapping baseline. Evaluated on clean DNN acoustic model trained in Section \ref{sec:dnn_am}, the feature-mapping enhanced features achieve 25.81\% WER which is 12.33\% relative improvement over the noisy features. Then we jointly train $F$ and $D$ to optimize $\mathcal{L}_{\text{AFM}}(\theta_f, \theta_d)$ as in Eq. \eqref{eqn:loss_afm} using the same input features and targets. The gradient reversal coefficient $\lambda$ is fixed at $60$ and the learning rate is $5\times 10^{-7}$ with a momentum of $0.5$ in the experiments. As shown in Table \ref{table:wer_afm}, AFM enhanced features achieve 24.45\% WER which is 16.95 \% and 5.27\% relative improvements over the noisy features and feature-mapping baseline. \begin{table}[h] \centering \begin{tabular}[c]{c|c|c|c|c|c} \hline \hline Test Data & BUS & CAF & PED & STR & Avg.\\ \hline Noisy & 36.25 & 31.78 & 22.76 & 27.18 & 29.44\\ \hline FM & 31.35 & 28.64 & 19.80 & 23.61 & 25.81 \\ \hline AFM & 30.97 & 26.09 & 18.40 & 22.53 & 24.45 \\ \hline \hline \end{tabular} \caption{The ASR WER (\%) performance of real noisy dev set in CHiME-3 enhanced by different methods evaluated on a clean DNN acoustic model. FM represents feature-mapping.} \label{table:wer_afm} \vspace{-20pt} \end{table} \subsection{Senone-Aware Adversarial Feature-Mapping Speech Enhancement} The SA-AFM experiment is conducted on top of the AFM system described in Section \ref{sec:exp_afm}. In addition to the LSTM $F$ and feedforward DNN $D$, we train a multi-conditional LSTM acoustic model $M$ using both the 8738 clean and 8738 noisy training utterances in CHiME-3 dataset. The LSTM $M$ has 4 hidden layers with 1024 units in each layer. A 512-dimensional projection layer is inserted on top each hidden layer to reduce the number of parameters. The output layer has 3012 output units predicting senone posteriors. The senone-level forced alignment of the training data is generated using a GMM-HMM system. As shown in Table \ref{table:wer_saafm}, the multi-conditional acoustic model achieves 19.28\% WER on CHiME-3 simulated dev set. \begin{table}[h] \centering \begin{tabular}[c]{c|c|c|c|c|c} \hline \hline System & BUS & CAF & PED & STR & Avg. \\ \hline Multi-Condition & 18.44 & 23.37 & 16.81 & 18.50 & 19.28 \\ \hline SA-FM & 18.19 & 22.29 & 15.31 & 18.26 & 18.51 \\ \hline SA-AFM & 17.02 & 21.01 & 14.41 & 17.13 & 17.38 \\ \hline \hline \end{tabular} \caption{The ASR WER (\%) performance of simulated noisy dev set in CHiME-3 by using multi-conditional acoustic model and different enhancement methods.} \label{table:wer_saafm} \vspace{-10pt} \end{table} Then we perform senone-aware feature-mapping (SA-FM) by jointly training $F$ and $M$ to optimize the feature-mapping loss and the senone classification loss in which $M$ takes the enhanced LFB features generated by $F$ as the input to predict the senone posteriors. The SA-FM achieves 18.51\% WER on the same testing data. Finally, SA-AFM is performed as described in Section \ref{sec:saafm} and it achieves 17.38\% WER which is 9.85\% and 6.10\% relative improvements over the multi-conditional acoustic model and SA-FM baseline. \section{Conclusions} In this paper, we advance feature-mapping approach with adversarial learning by proposing AFM method for speech enhancement. In AFM, we have a feature-mapping network $F$ that transforms the noisy speech features to clean ones with parallel noisy and clean training data and a discriminator $D$ that distinguishes the enhanced features from the clean ones. $F$ and $D$ are jointly trained to minimize the feature-mapping loss (i.e., MSE) and simultaneously mini-maximize the discrimination loss. On top of feature-mapping, AFM pushes the distribution of the enhanced features further towards that of the clean features with adversarial multi-task learning.To achieve better performance on ASR task, SA-AFM is further proposed to optimize the senone classification loss in addition to the AFM losses. We perform ASR experiments with features enhanced by AFM on CHiME-3 dataset. AFM achieves 16.95\% and 5.27\% relative WER improvements over the noisy features and feature-mapping baseline when evaluated on a clean DNN acoustic model. Furthermore, the proposed SA-AFM achieves 9.85\% relative WER improvement over the multi-conditional acoustic model. As we show in \cite{ts_adapt}, teacher-student (T/S) learning \cite{ts_learning} is better for robust model adaptation without the need of transcription. We are now working on the combination of AFM with T/S learning to further improve the ASR model performance. \clearpage \bibliographystyle{IEEEtran}

44e1cd9739cec837d6ab531b9c892f520fe26b3d

\section{Introduction} M-dwarfs are quickly stepping into the forefront as some of the best candidates in modern searches for habitable, Earth-like exoplanets. This is due mainly to their small masses and luminosities, favoring close-in Goldilocks zones which translate to stronger and more frequent signals for many exoplanet detection schemes. The Goldilocks zone may not provide the whole picture for habitability, however, for many M-dwarfs exhibit extraordinary flaring events \citep{flares} which may bombard these exoplanets with ionizing radiation. As flares are primarily magnetic phenomenae, one of the fundamental questions for assessing their effects on exoplanet habitability is "How are they formed?" It has been clear for some time that the magnetic activity a star is capable of generating is closely tied to its rotation rate, with faster rotating stars being more active up to a saturation threshold at roughly $Ro=t_{rot}/t_{conv}=0.1$ \citep{rotact}. Rotation rate alone, however, cannot provide the entire picture. \subsection{The Tachocline Divide} Considering the activity of late type stars, a sharp transition can be seen at roughly M3.5 ($0.35 M_\odot$) \citep{tachodiv}. Earlier than M3.5, stars are dominantly inactive with only $10\%$ demonstrating significant markers for magnetism. Among later stars than M3.5, however, nearly $90\%$ display magnetic activity. Suggestively, stellar modeling tells us that it is later than M3.5 where main-squence stars become fully convective (FC). Among other things, becoming FC means losing the transition region between the convection zone (CZ) and the underlying radiative zone (RZ). Helioseismology tells us that within the Sun, this transition is a layer of substantial velocity shear and thus it has come to be called the tachocline \citep{soimdi}. The stably-stratified shearing flows of the solar tachocline are often considered to be fundamental in organizing the Sun's dynamo \citep{babcock}. We seek to understand here how their presence or absence may be contributing to the divide observed between early and late type M-dwarfs. \subsection{Past Dynamo Simulations} While the tachocline was thought to be critical to the solar dynamo, simulations have shown that a solar-like CZ can, if rotating rapidly enough, sustain globally organized and periodically cycling \it wreaths \rm of magnetism even in the absence of a tachocline. These wreaths can be statistically steady solutions \citep{brownsteady} or go through periodic cycles \citep{browncycle}, among a variety of other behaviors stemming from an intricate and nonlinear parameter space. In the realm of M-dwarfs, several simulations of FC stars, e.g. \citet{browning08} and \citet{yadav15}, have demonstrated very strong magnetism reaching mean toroidal field strengths in excess of 10kG. The simulations with particularly strong fields appear to damp away nearly all of the differential rotation achieved by their hydrodynamic precursors. In some cases, FC simulations have produced broad dipolar caps of magnetism powerful enough to partly suppress convection and create polar dark spots. \section{Framing the Problem} This work employs the open-source 3D MHD code Rayleigh \citep{rayleigh} to evolve the anelastic equations in rotating spherical shells. Rayleigh performs competitively in benchmarks relative to codes such as MagIC and ASH and demonstrates efficient parallelization up to $O(10^5)$ cores. Rayleigh is a pseudospectral code, employing both a physical grid and a basis of spherical harmonics and Chebyshev polynomials. Time stepping is achieved with a hybrid implicit-explicit approach, where the linear terms are advanced via a 2nd order Crank-Nicolson method and the nonlinear terms by 2nd order Adams-Bashforth. \subsection{The Anelastic Equations} The anelastic equations are a fully nonlinear form of the fluid equations from which sound waves have been filtered out. This provides an appropriate framework for exploring subsonic convection within stellar interiors, where fast-moving p-modes would otherwise throttle the maximum timestep. The thermodynamic variables are linearized against a one dimensional, time independent background state given by $\bar{\rho},\,\bar{P},\,\bar{T},$ and $\bar{S}$, with deviations from the background written without overbars. The exact form of the anelastic equations solved in Rayleigh is \begin{equation} \nabla\cdot(\bar{\rho}\mathbf{v})=0\;, \end{equation} \begin{equation} \bar{\rho}\frac{D\mathbf{v}}{Dt}=-\bar{\rho}\nabla\frac{P}{\bar{\rho}}-\frac{\bar{\rho}S}{c_p}\mathbf{g}+\nabla\cdot\mathcal{D}\;, \end{equation} \begin{eqnarray} \lefteqn{\bar{\rho}\bar{T}\frac{DS}{Dt}=\nabla\cdot[\kappa\bar{\rho}\bar{T}\nabla S]\;+}\nonumber \\ && 2\bar{\rho}\nu\times[e_{ij}e_{ij}-\frac{1}{3}(\nabla\cdot\mathbf{v})^2]+Q\;, \end{eqnarray} \begin{equation} \frac{\rho}{\bar{\rho}}=\frac{P}{\bar{P}}-\frac{T}{\bar{T}}=\frac{P}{\gamma\bar{P}}-\frac{S}{c_p}\;, \end{equation} where $Q$ is the volumetric heating function, $e_{ij}$ is the strain rate tensor, and $\mathcal{D}$ is the viscous stress tensor defined as \begin{equation} \mathcal{D}_{ij}=2\bar{\rho}\nu[e_{ij}-\frac{1}{3}(\nabla\cdot\mathbf{v})\delta_{ij}]\;. \end{equation} Due to the resolutions accessible to modern computing, the viscosity, conductivity, and resistivity we employ are not the molecular values, but rather eddy diffusivities. These values are inflated by many orders of magnitude as a parameterization of the turbulent mixing occurring at sub-grid scales. \subsection{Modeling an M-Dwarf} The calculations were performed within a radial hydrodynamic background state derived using the stellar evolution community code MESA \citep{MESA}. We consider a ZAMS star of 0.4 M$_\odot$ with solar metalicity and rotating at 2 $\Omega_\odot=828\,\mathrm{nHz}$. In the outermost layers of stars, the anelastic equations begin to break down as flows approach the sound speed and non-diffusive radiative transfer becomes important. As a result, we must restrict our computational domain to exclude this region. In our notation, models are either "H" hydrodynamic or "D" dynamo, followed by the frame rotation rate $\Omega_0$ in multiples of $\Omega_\odot$ and lastly by "t" if the computational domain includes the tachocline. A plot of the density stratification and entropy gradient is presented in Figure \ref{fig:reference}. All simulations (H2, H2t, D2, D2t) had CZs extending from $R_t= 0.44R_*$ to $R_o=0.97R_*$, where $R_*=2.588\times 10^{10}$ cm containing $N_r=192$ radial grid points, and spanning $N_\rho=5$ density scale heights. The tachocline models (H2t, D2t) contained an additional radial domain spanning the tachocline region and underlying stable layer, $R_i=0.35R_*$ to $R_t$ with $N_r=48$. An angular resolution of $N_\theta\times N_\phi = 512 \times 1024$ was chosen for all models. \begin{figure} \centering \includegraphics[width=1.0\linewidth]{reduced_reference.png} \caption{The density stratification (red) and background entropy gradient (blue) employed by the simulations. Due to the numerical noise in MESA's entropy profiles, a smoothing function was applied before taking the gradient which results in a more gradual transition to convective stability than indicated by the stellar model. $R_t$ is marked with a vertical dashed line, while $R_i$ and $R_o$ are at the endpoints of the profiles.} \label{fig:reference} \end{figure} As used previously, e.g. \citet{brownsteady}, we employ viscosity profiles for models H2 and D2 proportional to $\rho^{-0.5}$ where the viscosity at the top of the domain is chosen to be $\nu_0=6.65\times 10^{11} $ cm$^2 $s$^{-1}$. The conductivity $\kappa$ and resistivity $\eta$ are chosen to yield a Prandtl number P$_\mathrm{r}=\nu/\kappa=1/4$ and magnetic Prandtl number P$_\mathrm{rm}=\nu/\eta=4$ throughout the domain. These choices for diffusivity yield a Rayleigh number 1247 times the empirically determined critical point for this system. Models H2t and D2t have similar structure in the CZ, but with dramatic reductions of diffusive amplitudes in the tachocline to increase the viscous time scale and delay its eventual unravelling, as \begin{equation} a = a_t+\frac{a_0(\frac{\rho}{\rho_0})^{-0.5}}{1+\exp{(c(R_t-r)/(R_o-r_i))}}\;. \end{equation} Here we choose tachocline diffusivities $a_t=10^{-3}a_0$ and a transition steepness $c=200$. In all simulations, the mean $(l=0)$ entropy field sees a separate, much smaller conductivity $\kappa_0$ which serves to discourage thermal conduction as a means of energy transport in the bulk of the CZ, and consequently forces the convective motions to carry the full luminosity of the star. As is common practice for simulations such as these, the model was first evolved using purely hydrodynamics. After a steady state was achieved, magnetism was introduced as white-noise perturbations and allowed to self-consistently reshape the flows while growing to its mature amplitudes. \subsubsection{Boundary Conditions} The upper and lower boundary conditions are impenetrable and stress free, \begin{equation} v_r|_{\mathrm{bc}}=\frac{d}{dr}(v_\theta/r)|_{\mathrm{bc}}=\frac{d}{dr}(v_\phi/r)|_{\mathrm{bc}}=0\;. \end{equation} The lower boundary is thermally insulating, and the top boundary extracts the star's luminosity $L_*=9.478\times 10^{31}$ erg s$^{-1}$ through a fixed conductive gradient, with \begin{equation} \frac{dS}{dr}|_{\mathrm{bot}}=0,\; \frac{dS}{dr}|_{\mathrm{top}}=\mathrm{const}\;. \end{equation} With no conductive input, energy balance is instead maintained through the volumetric heating function $Q$ which is adapted from the $\epsilon_{nuc}$ and $\nabla\cdot\mathcal{F}_{rad}$ reported by MESA. Finally, the magnetic field matches onto an external potential field at both boundaries, as \begin{equation} B=\nabla\Phi,\; \nabla^2\Phi|_{R_i,R_o}=0\;. \end{equation} \section{Toroidal Fields} \begin{figure} \centering \includegraphics[width=0.85\linewidth]{Bphi_average_joint.png} \caption{(a) Time- and longitude-averaged $B_\phi$ for case D2, showing the strong wreaths with core averages of 10kG, and the high-latitude structures with average strengths of 2-4kG. (b) The same for case D2t, with a tachocline. In the mid-CZ, the lower hemisphere wreath is dominating and extending across the equator. A second pair of wreaths exist in the tachocline with the same average strengths and polarities as their mid-CZ counterparts at this time, though this parity is not always present.} \label{fig:toroidal} \end{figure} \begin{figure} \centering \includegraphics[width=0.85\linewidth]{omega_joint.png} \caption{(a) Time- and longitude-averaged $\Omega-\Omega_0$ for case H2t, showing a rotational contrast of 170 nHz from equator to pole in the CZ and a transition to solid body rotation in the RZ. (b) The same for case D2t, where the differential rotation has been all but eliminated in the CZ. A degree of radial shear persists in the tachocline, especially at higher latitudes, as well as in the equatorial near-surface layers.} \label{fig:difrot} \end{figure} \begin{figure*}[ht] \centering \includegraphics[width=0.85\linewidth]{Bphi_slices.png} \caption{Toroidal fields of model D2t seen in Mollweide projection at mid-depth (a) and in the tachocline (b). Large-scale structures of opposite polarity form in opposite hemispheres, but at the time shown, $t = 1328.5$ rotations, the southern wreath is dominating. The small-scale structures clearly evident in (a) are absent in (b).} \label{fig:tslice} \end{figure*} \begin{figure*}[ht] \centering \includegraphics[width=0.95\linewidth]{Bp_evolution_joint.png} \caption{(a$_\mathrm{s}$ and a$_\mathrm{m}$) Azimuthally averaged $B_\phi$ at both depths $r=0.945R_*$ (near surface) and $r=0.684R_*$ (mid-layer), for case D2 as varying in time and latitude. Frequent polarity reversals are evident with a cycle period varying between 100 and 150 rotations, with an outlier at 200 rotations. (b$_\mathrm{s}$ and b$_\mathrm{m}$) The same for case D2t, which underwent only two reversals separated by 220 rotations. Time averaging intervals for Figures \ref{fig:toroidal} and \ref{fig:poloidal} are shown for each model.} \label{fig:evolution} \end{figure*} Several differences emerge from the two simulations when considering the final configurations of their toroidal fields. As clearly visible in Figure \ref{fig:toroidal}a, model D2 possesses two wreaths of opposite polarity forming at a depth of roughly $0.76R_*$ and at latitudes of about $\pm 25^\circ$. From shell slices as shown for D2t in Figure \ref{fig:tslice}, we see that these wreaths contain substantial longitudinal modulation, most likely an imprint of the vigorous convection taking place in this central region of the CZ. Toroidal field strengths within the wreaths peak close to 20kG, whereas temporal and azimuthal averages hover around 10kG in the wreath cores. Additionally, we see intermittent coherent toroidal fields near the poles. These high-latitude fields are weaker than those in the wreaths, achieving peak strengths of 16kG and azimuthal averages of about 4kG. The inclusion of a tachocline in model D2t alters the character of the mid-CZ wreaths. While the field strengths and choppiness remain unaffected, we note from Figure \ref{fig:tslice} a tendency for one wreath to dominate over the other in magnitude. This in turn allows the dominant wreath to push closer to the equator and in some cases extend a few degrees across it. The same intermittent behavior at high latitudes as in case D2 is observed here, though with marginally weaker amplitudes. We find that the tachocline provides a reservoir for the toroidal field as can be seen in Figure \ref{fig:toroidal}b. The wreaths here are for the most part not produced locally, and are instead pumped into this region of reduced resistivity by overshooting convection. In solar models, the shear of the tachocline is thought to provide a mean field $\Omega$-effect for converting poloidal to toroidal field, but that is not occurring in this simulation. As evident in Figure \ref{fig:difrot}, the strong fields have damped the latitudinal contrast in rotation rate from 170 nHz ($21\%$) in H2t to just 30 nHz ($3.6\%$) in D2t. Since the CZ has been brought nearly to solid body rotation, the transition to RZ mandates less rotational shear. In a different region of parameter space, with weaker fields through slowed rotation or reduced P$_\mathrm{rm}$, for example, we might expect to find that some differential rotation remains in the CZ and thus the tachocline could contribute a more significant $\Omega$-effect for toroidal field generation. With less disruption by the turbulent convective motions, we find that the wreaths in the tachocline shown in Figure \ref{fig:tslice}b are nearly uniform in longitude and thus their peak field strengths are close to their average core strengths at 10kG. \subsection{Time Evolution} In examining the evolution of the toroidal fields over time, we find further differences in the magnetic behaviors of these two simulations. Figure \ref{fig:evolution} shows that both models undergo reversals in the polarities of their mean fields, though only case D2 exhibits any regularity in its cycling period during the time captured by the simulations here. D2 has its toroidal fields reverse in both hemispheres every 100 to 150 rotations, with a failed reversal in the southern hemisphere near rotation 1610 leading to a half-cycle of twice the usual length in both hemispheres and a period where both wreaths had the same polarity. With a tachocline, however, we observe only two reversals in the 1550 rotations captured so far in simulation D2t. These reversals are separated by roughly 220 rotations, longer than any half-cycle in the CZ-only simulation. The latter of the two reversals occurred just before the end of the run for case D2t, so it is not clear yet whether these reversals and thus the interval between were random occurrences or if they mark the onset of a cycling phase for this star's dynamo. \section{Poloidal Fields} In addition to its effects on the internal toroidal fields, inserting a tachocline at the base of our CZ has led to significantly stronger and more organized poloidal fields near the stellar surface. We consider the total spectral power of the radial field at the top of the domain $S=\sum_{l,m}f^2_{lm}$ where $f_{lm}$ is a spherical harmonic coefficient. With no stable layer, we find $S=5.68$ kG$^2$, while case D2t has a surface spectral power of 6.63 kG$^2$, an increase of $17\%$. \begin{figure} \centering \includegraphics[width=0.85\linewidth]{Br_average_joint.png} \caption{(a) Time- and longitude-averaged $B_r$ for case D2, showing high latitude caps of poloidal field at strengths in excess of 5kG along with weaker mid-latitude fields. (b) The same for case D2t, low and mid-latitude structures are similar to case D2 but with greater amplitude, and the polar caps show more asymmetry between the two hemispheres.} \label{fig:poloidal} \end{figure} Figure \ref{fig:pslice} shows that the character of the surface magnetism varies significantly with latitude. Near the equator, the radial fields trace the narrow convective downflow lanes. In some instances, the concentrated fields become intense enough to fully suppress their local convection. There is an analogy to be made with starspots here, but we must keep in mind that these simulations extend only to $0.97R_*$, and thus such structures would still need to extend through the more turbulent layers of the star before they could be observable. At high latitudes, in the shadow of the tangent cylinder, the radial field covers much more area and typically has stronger fields. The structure present at high-latitude is suggestive of polar caps, and indeed an azimuthal average as shown in Figure \ref{fig:poloidal} confirms their presence. Spherical harmonic decomposition reveals another role played by the tachocline: organizing surface poloidal fields into large-scale structures. Not only does case D2t have more power in its poloidal field at the surface, but the fraction which is axisymmetric $S_{AS}=\sum_{l}f^2_{l,m=0}$ is nearly twice as great at 0.306 as that of case D2 at 0.158. While neither model produced a dipole dominated field, with axisymmetric dipole fractions $ f_{1,0}^2/S$ of 0.004 and 0.026 for D2 and D2t, case D2t showed a strong preference for its axisymmetric quadrupole and octupole modes, which together contained $(f^2_{2,0}+f^2_{3,0})/S=0.193$ of the power in the radial field. By way of contrast, these modes contained only $0.054$ of the total spectral power in case D2. \begin{figure} \centering \includegraphics[width=1.0\linewidth]{Br_slice.png} \caption{Radial fields of model D2 shown in Mollweide projection near the upper boundary of the simulation at depth 0.967R$_*$. Near the equator, the fields trace downflow lanes and reach magnitudes of 2-6 kG, while high-latitude fields have greater filling factors and attain strengths on the order of 10kG.} \label{fig:pslice} \end{figure} \subsection{Spin-Down Implications} The low-order modes of the surface poloidal field are particularly important due to their interactions with the stellar winds and consequently with the spin-down histories of these stars. Since the radial decay of magnetic multipoles goes as $r^{-(l+1)}$, the effective lever-arm each mode could use to exert a torque shrinks very quickly with increasing $l$. Thus, magnetized wind spin-down analyses tend to focus on modes with $l=1,2,$ or 3 \citep{winds}. Applying this principle qualitatively to the field configurations achieved in our models, we see some hints that by enhancing the low order poloidal fields near the surface, a tachocline may cause a star to spin-down more rapidly. We must emphasize that our comparison is not between early and late M-dwarfs, but rather between an early M-dwarf and another model of an early M-dwarf whose tachocline has been replaced with an impenetrable boundary. \section{Conclusions} We have compared the dynamos operated and fields generated in simulations of two M2-like stars, differing initially only in that one has a computational domain including only the CZ, while the other also includes a portion of the underlying RZ. In doing so, we have arrived at three main conclusions regarding the magnetism of such stars and how the presence of a tachocline may modify it: \begin{enumerate} \item The CZs of early M-dwarfs are perfectly capable of generating and organizing strong toroidal fields with or without an underlying tachocline of shear. \item The tachocline can provide a reservoir for the fields produced in the bulk of the CZ, and coupling between this reservoir and the mid-CZ dynamo can slow the reversals of the global field. \item The tachocline helps to organize near-surface poloidal fields onto larger spatial scales, which may create a favorable condition for the host star to shed angular momentum through its magnetized wind. \end{enumerate} While few experiments of this type have been conducted for early M-dwarfs, the more extensively studied parameter spaces in the solar regime have proven to house a rich diversity of behavior. More work is currently underway to examine the local sensitivities of our model in parameter space, and thus to assess the robustness of these conclusions concerning the features of deep convective shells with underlying tachoclines. \section*{Acknowledgements} {We thank Ben Brown, Sacha Brun, and Brad Hindman for helpful advice in developing this work. We thank Nick Featherstone for his assistance with Rayleigh, as well as the Computational Infrastructure for Geodynamics (http://geodynamics.org) which is funded by the National Science Foundation. The calculations presented here were performed on the NASA Pleiades supercomputer. This work was supported by NASA grant NNX17AG22G.}

108890154ac7122db9e6b08b2bbfd82d3a5bf4ae

\section*{Abstract} {\bf Theories where a fermionic dark matter candidate interacts with the Standard Model through a vector mediator are often studied using minimal models, which are not necessarily anomaly-free. In fact, minimal anomaly-free simplified models are usually strongly constrained by either direct detection experiments or collider searches for dilepton resonances. In this paper, we study the phenomenology of models with a fermionic dark matter candidate that couples axially to a leptophobic vector mediator. Canceling anomalies in these models requires considerably enlarging their field content. For an example minimal scenario we show that the additional fields lead to a potentially much richer phenomenology than the one predicted by the original simplified model. In particular collider searches for pair-produced neutralinos and charginos can be more sensitive than traditional monojet searches in thermally motivated parts of the parameter space where the mediator is outside the reach of current searches. } \vspace{10pt} \noindent\rule{\textwidth}{1pt} \tableofcontents\thispagestyle{fancy} \noindent\rule{\textwidth}{1pt} \vspace{10pt} \section{Introduction} \label{sec:intro} Dark matter phenomenology has often been studied using so-called Simplified Models \cite{Buchmueller:2013dya,Busoni:2013lha,Alves:2013tqa,Busoni:2014sya,Malik:2014ggr,Abdallah:2015ter,Bauer:2016gys,Alves:2016cqf} which aim to consider only the minimal Lagrangian and particle content relevant for relic density calculations, direct and indirect detection, and collider searches. Such models can be understood as a \emph{bottom-up} approach to dark matter model building, to be contrasted with the \emph{top-down} approach focusing on plausible dark matter scenarios in UV-complete theories such as Supersymmetry. However, the requirement for minimality and the focus on phenomenology has also resulted in many of these models failing basic theoretical self-consistency requirements. Models with a new heavy gauge boson, notably, often obfuscate how the masses of the mediator and dark matter are generated in a gauge-invariant manner. Furthermore, models with new fermions that are charged under the Standard Model or under a new gauge group are often plagued by anomalies. Fixing this often requires the introduction of further interactions and fields that will lead to a much richer phenomenology than predicted by the original simplified model~\cite{Duerr:2014wra,Duerr:2016tmh,Jacques:2016dqz,Ismail:2016tod,Das:2017ski,Das:2017flq,Das:2017deo,Ellis:2017tkh,Bauer:2018egk,Ellis:2018xal,Caron:2018yzp}. It is worth stressing that these problems mean that the models can not be consistently quantised and only provide predictions at tree level, which makes it clear that they are not models of nature in any meaningful sense but rather just benchmarks for specific experimental signatures. Studies of the phenomenology of anomaly-free $U(1)$ extensions in various contexts can be found in \cite{Ma:2001kg,Barr:2005je,PhysRevD.74.115017,Abel:2008ai,Mambrini:2010dq,Basso:2010yz,Chun:2010ve,Mambrini:2011dw,Heeck:2014zfa,Kile:2014jea,Elahi:2015vzh,Okada:2016gsh,Patra:2016shz,Klasen:2016qux,Ibe:2016dir,Liu:2017lpo,Arcadi:2017jqd,Cline:2017lvv,Chen:2017usq,Baek:2017sew,Cui:2017juz,DeRomeri:2017oxa,Nanda:2017bmi}. Perfect illustrations of this consistency issue are the so-called gauge portal models, involving a fermionic dark matter candidate that is charged under a new dark $U(1)'$ gauge group associated with a massive gauge boson $Z'$. While generating a mass for $Z'$ only requires introducing a new Higgs field, cancelling the anomalies associated with the dark fermion can prove particularly cumbersome. Since these models typically introduce family-independent couplings for the leptons and quarks, they often correspond to $U(1)_B$ and $U(1)_L$ models of gauged baryon and lepton numbers, which were originally discussed in \cite{FileviezPerez:2010gw,FileviezPerez:2011pt,Duerr:2013dza} with the dark matter phenomenology studied in \cite{Duerr:2013lka,Perez:2014qfa,Perez:2015rza}. While simple anomaly-free gauge portal solutions based on this family of models have been discussed in~\cite{Ellis:2017tkh}, they involve either a large vector-like coupling between the dark matter and the $Z'$, or a sizable coupling between the $Z'$ and the Standard Model (SM) leptons. Both features are somewhat undesirable from a dark matter perspective, as they respectively imply large direct detection cross-sections or a clear dilepton resonance signal at the LHC for large parts of the thermally motivated values of the parameters. Given the current sensitivities for these types of signals~\cite{Aprile:2017iyp,ATLAS-CONF-2017-027}, it is well-motivated to also investigate anomaly-free models for which the $Z'$ is leptophobic and couples primarily axially to the DM. As outlined in~\cite{FileviezPerez:2010gw,Ellis:2017tkh} however, anomaly cancellation for these models is non-trivial and requires dramatically increasing the model's particle content. In particular, it is not clear that the mono-X searches which typically are benchmarked with these models at the LHC are the most sensitive once mediator searches are avoided due to the new signatures which are introduced. In this paper we derive a minimal field content and charge assignment which cancels all gauge anomalies for a model with a fermionic dark matter candidate a vector mediator from a new broken $U(1)_{Y'}$ gauge group. As outlined above, in order to avoid the current direct detection and collider constraints, we require that the mediator is strictly leptophobic, with a purely axial coupling to at least one fermionic interaction eigenstate. We then use this scenario as a benchmark example to illustrate that, in the region of parameter space outside the reach of the dijet searches, anomaly-free gauge portal models will often be more sensitive to searches for new heavy states with electroweak charges at the LHC than to monojet searches. Requiring theoretical consistency for simplified models thus uncovers a particularly intriguing interplay between dark matter searches and the wider BSM programme of the LHC which inspired the title of this work\footnote{Whac-a-mole is a classic arcade game in which new moles appear when previous ones have been pushed back into their holes.}: constructing a consistent model which avoids the strongest constraints currently set on simplified models predicts new signatures which can allow other searches to exclude the model. This also suggests the usefulness of the Simplified Model as a benchmark for monojet searches is rather questionable. This model was recently independently derived and studied in \cite{Caron:2018yzp} in the limit where all anomaly-cancelling fermions are decoupled from the dark matter and collider phenomenology, we will here instead focus on the phenomenological consequences of keeping the masses of these new fermions at the same scale as the dark matter fermion. The structure of the paper is as follows: in Section~\ref{sec:modelcontent} we derive and discuss the model, with a discussion of the mass mixing scenario we will employ in Section~\ref{sec:massmixing} and the effect of kinetic mixing of the $U(1)$s in Section~\ref{subsec:kinetic}. This is followed by a study of the dark matter phenomenology in Section~\ref{sec:dmobs} and a study of the LHC phenomenology for a selection of thermal and non-thermal benchmark parameter points in Section~{sec:lhcpheno}. We conclude in Section~\ref{sec:conclude}. \section{Model details} \label{sec:modelcontent} We start from a minimal gauge-portal model, with a new Dirac fermion $\chi$ that is charged under a new $U(1)_{Y'}$ gauge group associated with a massive gauge boson $Z'$. The corresponding new physics Lagrangian is \begin{equation} \mathcal{L} = -\frac{1}{4} B^{'\mu\nu}B'_{\mu\nu} - \frac{\epsilon}{4} B^{\mu\nu}B'_{\mu\nu} + \bar{\chi}(\slashed{\partial}-m)\chi - \bar{\chi}(g_V + g_A\gamma_5) \gamma^\mu \chi Z'_\mu \end{equation} where $B_{\mu\nu}$ and $B'_{\mu\nu}$ are the field strengths for the SM hypercharge and the new $U(1)_{Y'}$ group respectively. In the rest of our study, we set the kinetic mixing $\epsilon$ to zero, briefly discussing this choice in section~\ref{subsec:kinetic}. The relative values of the vector and axial couplings $g_{V,A}$ depend on the dark hypercharges of the left and right-handed components of $\chi$, that is, $\chi_L$ and $\chi_R$. This basic scenario has been widely used to derive constraints from relic density, direct detection, and collider searches for gauge portal models. When the coupling between $\chi$ and $Z'$ is axial, however, which is the configuration with the loosest direct detection constraints, this model has non-zero $U(1)_{Y'}$ anomalies. In what follows, we present an example model with an extended dark sector that allows to cancel these anomalies while keeping the DM-$Z'$ coupling mostly axial and the $Z'$ leptophobic. This corresponds to a specific implementation of the general gauged baryon and lepton-number motivated models derived in \cite{Duerr:2013dza,Duerr:2014wra}. We stress that, although our chosen model has a particularly large number of fields, it is among the most minimal models that can be built with our requirements. A more detailed discussion of the construction of these models is presented by the authors of~\cite{Ellis:2017tkh}. In the rest of our study, we will use \textsc{SARAH} \cite{Staub:2008uz,Staub:2009bi,Staub:2010jh,Staub:2012pb,Staub:2013tta,Staub:2015kfa} throughout in order to implement the model and derive relevant quantities. In order to have an anomaly-free gauge portal model that satisfies the existing experimental constraints, we need to introduce additional fields which transform non-trivially under $ U(1)_Y \times SU(2)_L$ \cite{FileviezPerez:2010gw,Ellis:2017tkh}. A minimal solution is to add $6$ new Weyl fermions in the following $U(1)_Y \times SU(2)_L \times SU(3)_C \times U(1)_{Y'}$ representations\footnote{An alternative solution with fewer new fields using both $SU(2)_L$ doublets and triplets was derived in \cite{Perez:2014qfa} and its dark matter phenomenology was studied in \cite{Ohmer:2015lxa}.}: \begin{itemize} \item $\chi_L \sim (0,\mathbf{1},\mathbf{1},Y_{\chi}')$ \item $\chi_R \sim (0,\mathbf{1},\mathbf{1},-Y_{\chi}')$ \item $\theta_L \sim (Y_\theta,\mathbf{1},\mathbf{1},Y_{\theta_L}')$ \item $\theta_R \sim (Y_\theta,\mathbf{1},\mathbf{1},Y_{\theta_R}')$ \item $\phi_L \sim (Y_\phi,\mathbf{2},\mathbf{1},Y_{\phi_L}')$ \item $\phi_R \sim (Y_\phi,\mathbf{2},\mathbf{1},Y_{\phi_R}')$ \end{itemize} Here we have already made $\theta$ and $\phi$ vectorlike under the Standard Model gauge group to avoid having to consider anomaly equations for products of gauge groups not involving $U(1)_{Y'}$. The relevant anomaly equations assuming $Y'_{l,e} = 0$ (which requires $Y'_{q} = Y'_{u,d}$ to keep the Standard Model Yukawa terms gauge invariant) are then: \begin{align} 9 Y_q'+Y_{\phi_L}'-Y_{\phi_R}' &= 0~~~SU(2)_L^2 \times U(1)_{Y'} \\ -18 Y_q' + 2 Y_{\phi_L}' Y_{\phi}^2 - 2 Y_{\phi_R}' Y_{\phi}^2 + Y_{\theta_L}' Y_{\theta}^2 - Y_{\theta_R}' Y_{\theta}^2 &= 0~~~U(1)_Y^2 \times U(1)_{Y'} \\ 2 Y_{\phi_L}'^2 Y_{\phi} - 2 Y_{\phi_R}'^2 Y_{\phi} + Y_{\theta_L}'^2 Y_{\theta} - Y_{\theta_R}'^2 Y_{\theta} &= 0~~~U(1)_Y \times U(1)_{Y'}^2 \\ 2 Y_{\chi}'^3 + Y_{\theta_L}'^3 - Y_{\theta_R}'^3 + 2 Y_{\phi_L}'^3 - 2 Y_{\phi_R}'^3 &= 0~~~U(1)_{Y'}^3 \\ 2 Y_{\chi}' + Y_{\theta_L}' - Y_{\theta_R}' + 2 Y_{\phi_L}' - 2 Y_{\phi_R}' &= 0~~~U(1)_{Y'} \end{align} In order to avoid having charged dark matter candidates we require $Y_\phi = m + \frac{1}{2}$ and $Y_\theta = n$ where $m,n \in \mathbb{Z}$, so that the lightest charged state can decay into the lightest neutral state by emitting a Standard Model particle. Restricting ourselves to configurations with $|Q| \in \{0,1\}$ for all new fields we find the solution \begin{align} Y_{\theta} &= \pm 1 \\ Y_{\phi} &= \pm \frac{1}{2} \\ Y_q' &= -\frac{2}{9} Y_{\phi_L}' \\ Y_{\chi}' &= -Y_{\phi_L}' \\ Y_{\theta_L}' &= -Y_{\phi_L}' \\ Y_{\theta_R}' &= Y_{\phi_L}' \\ Y_{\phi_R}' &= -Y_{\phi_L}' \end{align} For simplicity we take $Y_{\phi_L}' = 1$ and $Y_\theta, Y_\phi = 1, \frac{1}{2}$ for the rest of the paper\footnote{While writing up this paper a similar study appeared in \cite{Caron:2018yzp} which directly connected the model to $U(1)_B$ and hence chose to normalise such that $Y_q' = \frac{1}{3}$, however the models are otherwise identical. We will take a different approach to studying the phenomenology for the remainder of this paper.}. Due to the gauge invariance of the SM lepton Yukawa interaction term mentioned above, the Standard Model Higgs doublet $H$ must be a singlet under $U(1)_{Y'}$. Therefore in order to break $U(1)_{Y'}$ and give masses to all of the new fields we have to add a new complex scalar which is a Standard Model singlet: \begin{itemize} \item $\tilde{S} \sim (0,\mathbf{1},\mathbf{1},2)$ \end{itemize} This scalar gets a vev $v_{S}$: \begin{align} \tilde{S} &= \frac{1}{\sqrt{2}} \left(v_{S} + S + i a\right) \end{align} which generates a mass for the $Z'$ gauge boson of $U(1)_{Y'}$ \begin{equation} m_{Z'}^2 = 4 g_{Y'}^2 v_S^2. \end{equation} The allowed mass and interaction terms in the Lagrangian are then as follows: \begin{align} \mathcal{L} \supset &-\mu_H^2 H^\dagger H - \mu_{S}^2 \tilde{S}^2 -\lambda_H |H^\dagger H|^2 -\lambda_{H,S} H^\dagger H \tilde{S}^2 - \lambda_{S} \tilde{S}^4 \\ &- y_d H \bar{d} q - y_e H \bar{e} l - y_u \widetilde{H} \bar{u} q + h.c. \\ &- y_{\theta_R} \widetilde{H} \overline{\phi_L} \theta_R - y_{\chi_R} H \overline{\phi_L} \chi_R - y_{\chi_L} \widetilde{H} \overline{\chi_L} \phi_R - y_{\theta_L} H \overline{\theta_L} \phi_R + h.c. \\ &- y_{\chi} \tilde{S} \overline{\chi_L} \chi_R - y_{\theta} \tilde{S} \overline{\theta_L} \theta_R - y_{\phi} \tilde{S}^\dagger \overline{\phi_L} \phi_R - y_{\chi_L,M} \tilde{S} \overline{\chi_L} \chi_L^c - y_{\chi_R,M} \tilde{S}^\dagger \overline{\chi_R} \chi_R^c + h.c. \end{align} where $\widetilde{H} = i \sigma^2 H^*$. \subsection{From interactions to mass eigenstates} \label{sec:massmixing} In the broken phase we get the following Majorana mass matrix for the neutral states after expanding the $SU(2)$ doublets as $\phi_{L/R} = \begin{pmatrix} \phi_{L/R,1} \\ \phi_{L/R,2} \end{pmatrix}$: \begin{equation} \begin{pmatrix}\overline{\chi_L} & \overline{\phi_{R,2}^c} & \overline{\chi_R^c} & \overline{\phi_{L,2}} \end{pmatrix} \begin{pmatrix}\sqrt{2} v_{S} y_{\chi_L} & 0 & \frac{v_{S} y_{\chi}}{\sqrt{2}} & \frac{ v_H y_{\chi_L}}{\sqrt{2}} \\ 0 & 0 & \frac{ v_H y_{\chi_R}}{\sqrt{2}} & \frac{v_{S} y_{\phi}}{\sqrt{2}} \\ \frac{ v_{S} y_{\chi}}{\sqrt{2}} & \frac{ v_H y_{\chi_R}}{\sqrt{2}} & \sqrt{2} v_{S} y_{\chi_R} &0 \\ \frac{ v_H y_{\chi_L}}{\sqrt{2}} & \frac{ v_{S} y_{\phi} }{\sqrt{2}} & 0 & 0\end{pmatrix} \begin{pmatrix}\chi_L^c \\ \phi_{R,2} \\ \chi_R \\ \phi_{L,2}^c \end{pmatrix}. \label{eq:neutralmix}\end{equation} Turning off the Majorana terms $y_{\chi_L}$ and $y_{\chi_R}$ allows us to describe the propagating degrees of freedom as two Dirac fermions $\chi_{1/2}$ where $\chi_1$ will generically denote the lightest fermion a.k.a. the DM candidate, and that we will always take as mostly $\chi$ to avoid direct detection constraints. Keeping them turned on but small splits the two Dirac pairs into four Majorana fermions \cite{DeSimone:2010tf,Davoli:2017swj}, which will naturally suppress the diagonal vectorlike couplings of the Majorana dark matter fermion. \begin{figure}[!tbp] \centering \includegraphics[width=0.6\textwidth]{spectrum.png} \caption{ Two representative model spectra which give the correct relic density. The left spectrum shows a scenario with large mixing in both the neutral and charged fermion sectors, and is ruled out by direct detection constraints. The right spectrum is allowed by direct detection and dijet constraints and will be used as benchmark point \textbf{[3]} in Section~\ref{sec:lhcpheno}. \label{fig:spectrum}} \end{figure} The terms mixing the Standard Model singlets with the neutral components of the $SU(2)$ doublets, $y_{\chi_R}$ and $y_{\chi_L}$, also introduce couplings to the $Z$ for $\chi_1$ which again rule the model out through direct detection unless these terms are very small: $y_{\chi_R}, y_{\chi_L} v_H \ll 10^{-2} y_\chi v_S$ for $y_\phi = 1.1 y_\chi$. We keep track of this effect and note that this is a somewhat unfortunate consequence of this particular charge assignment as it requires some fine-tuning to achieve the stated goal of avoiding direct detection constraints. Other anomaly-cancelling solutions such as models involving a $SU(2)_L$ triplet with $Y=0$ could avoid this hurdle as the neutral fermions would not couple to the $Z$. We present an analysis of the direct detection sensitivity of non-zero $y_{\chi_R}, y_{\chi_L}$ in Section~\ref{sec:dmobs} to give an estimate of the degree of fine-tuning involved. We note that turning on the Majorana terms in the Lagrangian would also suppress the spin-independent direct detection cross section due to the absence of diagonal vectorlike interactions among the Majorana fermions, and the parameter space and phenomenology would also be more complex due to the presence of four Majorana fermions and a Dirac fermion in the spectrum in the limit where the doublet Yukawa is not much larger than that of the singlet. However this limit will share many qualitative features with the same limit in the Dirac case which we explore here\footnote{In this sense the Dirac scenario provides a good way to study the phenomenology with a smaller parameter space with the understanding that in particular the fine-tuning in the neutral fermion sector mass-mixing would be alleviated by turning on small Majorana terms.}. Although the $y_{\chi_R}$ and $y_{\chi_L}$ couplings should be small in order to avoid direct detection constraints, they should also not be identically zero in order to avoid having the neutral component of the doublet $\phi$ ---which always couples to the $Z$--- being stable, and thus a dark matter candidate. In what follows, we therefore always choose small non-zero values for these couplings in our benchmark models. Note that the existence of these couplings allows $\chi_1$ and $\chi_2$ to coannihilate if they are close enough in mass (so $y_\chi \sim y_\phi$) by ensuring that the two particles remain in equilibrium in the early Universe. We emphasize that studies that involve relic density constraints need to take this possibility into account. The mass matrix for the charged fermions is: \begin{equation} \begin{pmatrix} \overline{\theta_L} & \overline{\phi_{L,1}} \end{pmatrix} \begin{pmatrix} \frac{v_{S} y_{\theta}}{\sqrt{2}} & \frac{v_H y_{\theta_L}}{\sqrt{2}} \\ \frac{v_H y_{\theta_R}}{\sqrt{2}} & \frac{v_{S} y_{\phi}}{\sqrt{2}}\end{pmatrix} \begin{pmatrix} \theta_R \\ \phi_{R,1} \end{pmatrix} .\label{eq:chargedmix} \end{equation} We denote the two mass eigenstates by $\chi_{1/2}^\pm$. If the off-diagonal elements are small and $y_\theta \gg y_\phi$, $\chi_1^\pm$ is the charged component of the doublet $\phi$, $m_{\chi_1^\pm} \sim m_{\chi_2}$ and $\chi_1^\pm$ can also become a coannihilation partner of the dark matter. We will study this maximal coannihilation scenario in Section~\ref{sec:dmobs}. For completeness we also include the mixing in the scalar sector in the Appendix~\ref{sec:appendix}. This mixing has no incidence on the DM phenomenology as long as the contribution of the scalar portal interactions to the DM relic density remain negligible. We will therefore assume $\lambda_{H,S} = 0$ such that the mixing completely vanishes and $H$ is completely Standard Model-like, with the caveat that our results only apply as long as the approximation that any such effects are negligible holds. Including higher order corrections will still induce effective interactions between the dark matter and light quarks through the 125 GeV state, which will in general be constrained by the spin-independent direct detection cross section, however the $y H \overline{\chi} \chi $ interaction is only constrained to have $y \lesssim 0.1$ by Xenon1T \cite{Aprile:2017iyp} so this mixing is not as fine-tuned as that in the neutral fermion sector. An interesting question in a model with such a rich particle content is whether there is a decoupling limit for the new anomaly-cancelling fermion states. One possible decoupling direction would be to take the $v_S\rightarrow \infty$ limit while requiring $y_\chi$ to be much small than the other Yukawa couplings. We observe, however, that $v_S$ controls not only the fermion masses, but also the mass of the new gauge boson $Z'$ that mediates the dark matter annihilation rate. Notably, in the $m_{\chi_1} > m_{Z'}$ limit where the additional fermions are heavy, most of the dark matter annihilation occurs through the $Z'$, and $y_\chi$ is chosen to maximise this annihilation channel, the associated velocity-averaged cross-section in the $s$-wave verifies $$\langle\sigma v\rangle \lesssim \frac{g_{Y'}^2}{6 \,v_S^2}.$$ The observed relic density is achieved for $\langle\sigma v\rangle\approx 10^{-8}$~GeV, which implies that $v_S$ cannot be much larger than $10$~TeV for $g_{Y'}\approx 1$. Since, for gauge portal dark matter models, perturbative unitarity sets order one bounds on Yukawa couplings~\cite{Hedri:2014mua}, the new fermionic states cannot be made heavier than a few tens of TeV in this limit. In the 'on-shell' region $2m_{\chi_1} < m_{Z'}$ where collider searches are particularly sensitive, it is typically possible to make the anomaly cancelling fermions heavier than the $Z'$ since $m_{Z'} = 2 v_S,~ m_{\chi_2}/m_{\chi_1^\pm} \sim y_{\phi}/\sqrt{2} v_S$ which will heavily suppress their pair production cross section. The simplified model framework for gauge-portal dark matter therefore proves mostly consistent for LHC studies, however as we will show coannihilation effects are well-motivated from a relic density perspective in the on-shell region and does predict clear new LHC signatures. A potential worry for our model is that since $U(1)_{Y'} \sim U(1)_B$ at low energies, breaking $U(1)_{Y'}$ might induce proton decay or neutron-antineutron oscillations. The general form of the corresponding operator is: \begin{equation} \mathcal{O} \propto q^{3p} \ell^q \tilde{S}^r \end{equation} with $p,q,r \in \mathbb{N}$. Since $Y_q' = 2/9$ and $Y_{\tilde{S}}' = 2$, the minimal gauge invariant realization is given by: \begin{equation} \mathcal{O} \propto (q^9 v_S) \ell \end{equation} with $|\Delta B| = 3$ and mass dimension 16. As such we are not sensitive to proton decay bounds or searches for neutron-antineutron oscillations \cite{Babu:2013jba}, and any other effect is highly suppressed at $Q \sim \Lambda_{QCD}$ by 12 powers of the scale where $U(1)_{Y'}$ is broken. The general insensitivity of searches for baryon number violation to electroweak scales of $U(1)_B$ breaking when it is gauged has been previously pointed out in e.g. \cite{Carone:1995pu,Aranda:2000ma,FileviezPerez:2010gw}. \subsection{The effect of kinetic mixing} \label{subsec:kinetic} In general a kinetic mixing term between $U(1)_Y$ and $U(1)_{Y'}$ is allowed by all symmetries: \begin{equation} -\frac{\epsilon}{4} B^{\mu \nu} B'_{\mu \nu}. \end{equation} If we remain agnostic about the UV completion of the theory, we can in theory set $\epsilon = 0$ at any scale we wish. However a motivated completion scenario of this model would have the gauge groups unify at some high scale $\Lambda$ which would set $\epsilon = 0$ at that scale, and since the quarks are charged under both $U(1)_Y$ and $U(1)_{Y'}$ and do not have orthogonal charges, this term would nevertheless run to a non-zero value at the weak scale. The effect on electroweak precision observables for models similar to ours has been studied in \cite{Kahlhoefer:2015bea,Ellis:2018xal,Caron:2018yzp} which found the resulting constraints to be rather weak. However this mixing will also introduce couplings of the $Z'$ to leptons, as after rotating the kinetic mixing away the general form of the covariant derivative becomes: \begin{equation} D_\mu \eta = \left(\partial_\mu -i \sum_{x,y} Q^x_\eta g_{xy} V^y_\mu \right) \eta \label{eq:kinmixing} \end{equation} where the sum over $x,y = Y, Y'$ denotes the two $U(1)$s. When $\epsilon = 0$, $g_{Y Y'} = g_{Y' Y} = 0$. A non-zero kinetic mixing will therefore not affect the coupling structure of the dark matter to the $Z'$ but it could make our model dramatically sensitive to LHC searches for dilepton resonances. In order to show that the effect of the kinetic mixing remains small in most of our parameter space, we have estimated the running of $g_{Y}$ and $g_{YY'}$ using $g_{Y'} (100 \textrm{ GeV}) = 0.2, 1$ and $g_Y(m_Z) = 0.358$ when $\epsilon = 0$ at $\Lambda = 10$ TeV. The one-loop renormalisation group equations of these parameters, derived using \textsc{SARAH} \cite{Fonseca:2013bua} (assuming $g_{Y Y'}, g_{Y' Y} \ll g_{Y}, g_{Y'}$ and that all new fermions have masses $< m_{Z'}$), are given by: \begin{equation} \mu \frac{d g_{Y}}{d \mu} = \frac{53 g_Y^3}{160 \pi^2}, ~~~~ \mu \frac{d g_{Y'}}{d \mu} = \frac{53 g_{Y'}^3}{108 \pi^2} \end{equation} \begin{equation} \mu \frac{d g_{Y Y'}}{d \mu} = - \frac{ g_Y^2 g_{Y'}}{6 \sqrt{15} \pi^2}, ~~~~ \mu \frac{d g_{Y' Y}}{d \mu} = - \frac{ g_Y g_{Y'}^2}{6 \sqrt{15} \pi^2} \end{equation} The evolution of the mixing couplings as a function of the energy scale $\mu$ is shown in Figure~\ref{fig:running}. Calculating di-lepton constraints requires the branching ratio to leptons to be correctly evaluated taking into account $Z' \to \chi \chi$ decays, after which the model-independent cross section limits for narrow resonances in \cite{Aaboud:2017buh} can be recasted after taking into account the production cross section, and the branching fraction into $e^+ e^-$ and $\mu^+ \mu^-$. The ultimate sensitivity therefore depends on the value of $g_{Y'}$, so ultimately a combination of $g_{Y'}$ and $\Lambda$ will be constrained as demonstrated in Figure~\ref{fig:running}. The constraint gets weaker for larger values of $m_{Z'}$. According to~\cite{Caron:2018yzp} dilepton constraints become relevant again for thermally motivated values of $g_{Y'}$ for $m_{Z'} \sim \Lambda/100$, but this conclusion only applies when $g_{Y'}$ is fixed by the relic density assuming no coannihilation. In fact, in the coannihilation region, relic density requirements typically point to smaller values of $g_{Y'}$, that lead to a slower running of $g_{Y Y'}$. In this region, dilepton constraints will therefore be negligible even for smaller values of $m_{Z'}/\Lambda$. Since it is in general always possible to choose $\Lambda$ such that dilepton constraints are negligible we will not consider these when determining the benchmark points which we take as not constrained by dijet constraints in Section~\ref{sec:lhcpheno}, but a full study of the parameter space should of course specify the assumptions made on $\Lambda$. An interesting study we leave for future work would be to derive a model where $\text{Tr}(YY') = 0$ and the kinetic mixing would not run above the scale where all of the new fermions are dynamical, which would allow the gauge unification scale to be pushed arbitrarily high. \begin{figure}[!tbp] \centering \includegraphics[width=0.6\textwidth]{running.pdf} \caption{Running of the effective mixing between $U(1)_Y$ and $U(1)_{Y'}$ using the notation introduced in Equation~\ref{eq:kinmixing} for $\Lambda = 10$ TeV. The $g_{YY'}$ coupling determines the coupling of the mostly $U(1)_{Y'}$ mass eigenstate to leptons. \label{fig:running}} \end{figure} In the following section, we derive the constraints associated with the model described at the beginning of this section. We notably derive the relic density and direct detection constraints for well-motived choices of parameters, and derive and discuss the impact of the current $13$~TeV LHC searches for selected benchmark points. \section{Phenomenology} This section explores the constraints on our example model from relic density requirements (for benchmark parameter points we require the relic density to have the value measured by \textsc{Planck}~\cite{Ade:2015xua}), direct detection, and LHC searches. In order to evaluate these constraints we have implemented our model in \textsc{SARAH} and exported it to the \textsc{UFO} format \cite{Degrande:2011ua}. The mixing parameters as well as the masses of the different eigenstates are derived from the Yukawa couplings and the vevs using a separate Python code, and directly inputed into the parameter card of the model. We use \textsc{MG5\_aMC@NLO}~\cite{Alwall:2011uj} to numerically evaluate the widths of $\chi_2$ and $\chi_{1/2}^\pm$ \cite{Alwall:2014bza} which typically decay into a three-body final state in the coannihilation region. The widths of $Z'$ and $h_2$ are evaluated analytically at tree-level assuming vanishing mixing in the neutral and charged scalar sectors using the general forms given in \cite{Abdallah:2015ter}. Finally, we evaluate the $13$~TeV LHC constraints using \textsc{CheckMATE} 2 \cite{Drees:2013wra,Dercks:2016npn,deFavereau:2013fsa,Cacciari:2011ma,Cacciari:2005hq,Cacciari:2008gp,Read:2002hq,Lester:1999tx,Barr:2003rg,Cheng:2008hk,Tovey:2008ui,Polesello:2009rn,Matchev:2009ad}. \subsection{Dark Matter Observables} \label{sec:dmobs} We use \textsc{MadDM} 3.0 \cite{Alwall:2011uj,Backovic:2013dpa,Backovic:2015cra,Ambrogi:2018jqj} to calculate the relic density and the direct detection cross-section for our model. We begin by investigating the sensitivity to direct detection experiments induced by the mixing of the DM candidate $\chi_1$ with the neutral components of the new $SU(2)_L$ doublet $\phi$ as discussed above. We find the limits on the spin-independent cross-section with nucleons from Xenon1T \cite{Aprile:2017iyp} to always be the most constraining\footnote{The other experimental constraints we make use of are the spin-dependent proton interaction limits set by PICO-60 \cite{Amole:2017dex} and the spin-dependent neutron interaction limits set by LUX \cite{PhysRevLett.118.251302}.} and plot the resulting constraints as a function of $(y_{\chi_R}, y_{\chi_L})/y_\chi$ for $v_S = v_H$ in Figure~\ref{fig:xenon1t}. In general these couplings have to verify $y_{\chi_R}, y_{\chi_L} v_H \ll 10^{-2} y_\chi v_S$ for the induced coupling to the $Z$ not to generate a too large cross section when $y_\chi = 1.1 y_\phi$, however this constraint significantly relaxes if there is a hierarchy between $y_\chi$ and $y_\phi$ or $v_H$ and $v_S$. For diagonal Yukawa couplings which are roughly equal, the resulting upper bounds on the mixing between $\chi_1$ and $\chi_2$ still allow for prompt decays of the heavy dark sector particles in the coannihilation region as long as the mass difference is greater than 10 GeV. Another source of direct detection constraints come from recent constraints on the $\mathcal{O}_{13}$ momentum-dependent operator which appears when integrating out the $Z'$ set by PandaX-II \cite{Xia:2018qgs}. These constrain a combination of $m_{Z'}$ and $g_{Y'}$ and are stronger than dijet constraints for low $m_{Z'}$. We plot the constraint and the value for some selected parameter points in our model in Figure~\ref{fig:pandax}. To check that $\chi_2$ and $\chi_1^\pm$ can be interconverted with $\chi_1$ efficiently enough for coannihilation to occur we have followed the procedure in \cite{Ellis:2015vaa,ElHedri:2017nny}. Namely, we compute the rate $\Gamma$ associated with the $t$-channel $\chi_1\, q \,\leftrightarrow\,(\chi_2,\chi_1^\pm)\,q'$ scattering, that is the dominant $\chi_1$--$\chi_2,\chi_1^\pm$ conversion process in our model since the quarks are light. Neglecting the quark masses, the average rate can be expressed as a function of the scattering cross-section $\sigma_\text{scatt}$ as \begin{align} \langle \Gamma\rangle = \int_{E_{min}}^\infty \sigma_\text{scatt} \frac{2\times 4\pi}{(2\pi)^3}\frac{p_{q,rest}^2 dp_{q,rest}}{e^{p_{q,rest}/T} + 1} \end{align} where $p_{q,rest}$ is the momentum of the incoming quark in the rest frame of the heavy initial state fermion. The minimal energy of this incoming quark, $E_{min}$, is defined by \begin{align} E_{min} &= \text{max}\left\{0, \frac{m_f^2 - m_i^2}{2 m_i}\right\} \end{align} for initial and final states of masses $m_i$ and $m_f$ respectively. We notably find that for $m_{\chi_1} = 150$ GeV, $m_{\chi_2} = 165$ GeV ---a relative mass splitting of order $10$\%--- the $q \chi_1 \to q' \chi_2$ average rate is still many orders of magnitude larger than the Hubble rate at freeze-out for a singlet-doublet mixing approaching the direct detection limit. An example diagram of the process is shown in Figure~\ref{fig:annihilation}. We therefore consider dark matter masses as low as $150$~GeV to be safe when selecting parameter points later in the study, however we display results for $m_{\chi_1}$ below this value in the relic density plots in Figure~\ref{fig:relic}. \begin{figure}[!h] \centering \begin{subfigure}[b]{0.49\textwidth} \centering \includegraphics[width=\textwidth]{direct_detection.png} \caption{} \label{fig:xenon1t} \end{subfigure} \begin{subfigure}[b]{0.49\textwidth} \centering \includegraphics[width=\textwidth]{direct_detection_o13.png} \caption{} \label{fig:pandax} \end{subfigure} \caption{The impact of direct detection constraints on (a) the mixing between $\chi_{L,R}$ and the neutral components of the doublets $\phi_{L,R}$ where we have used the shorthand $y_Z = y_{\chi_R}, y_{\chi_L}$ in the legend to indicate that these arise from the coupling to the $Z$. We have set $v_S = v_H$ here, and (b) the mass of the $Z'$ and and $g_{Y'}$ arising from constraints on the $\mathcal{O}_{13}$ operator (here in the form used in \cite{Xia:2018qgs}, where the dimensionless coupling is defined as the squared dimensionless Wilson coefficient when setting the suppression scale to the electroweak vev). \label{fig:dirdet}} \end{figure} We now compute the DM relic density as a function of $m_{\chi_1}$ and $m_{Z'}$ for a few representative scenarios, showing the results in Figure~\ref{fig:relic}. On one hand we considered a non-coannihilating case, where there is a sizable mass hierarchy between $\chi_1$ and $\chi_2/\chi_1^\pm$ and the mixing between the two states does not contribute significantly to the final relic density. This scenario corresponds to that found in typical axial-vector Simplified Models with the addition of diagrams involving $S$, drawn in the top row of Figure~\ref{fig:annihilation}. On the other hand, we considered two coannihilating scenarios where the $\chi_1-\chi_2$ and $\chi_1-\chi_1^\pm$ relative mass splittings, defined by \begin{equation} \Delta m = \frac{m_{\chi_2/\chi_1^\pm} - m_{\chi_1}}{m_{\chi_1}} \end{equation} are fixed to $10\%$ and $2\%$, and the $\chi_1-\chi_2$ mixing, albeit small, is sufficient to ensure thermal and chemical equilibrium between the two particles in the early Universe for these values of $\Delta$ (example diagrams of the relevant processes are drawn in the bottom row of Figure~\ref{fig:annihilation}). We present our results for these three configurations in figure~\ref{fig:relic}. In the non-coannihilating scenario the relic density observed by \textsc{Planck}~\cite{Ade:2015xua} can generically be obtained for perturbative values of all couplings for DM masses around the TeV scale. Setting the DM relic density to its Planck value also allows to uniquely determine the coupling $g_{Y'}$ after $m_{\chi_1}$, $m_{Z'}$, and $m_S$ are fixed. This interesting feature however does not hold when $m_S$ and $\Delta m$ are small enough for coannihilation effects to be significant. \begin{figure}[h] \hspace*{-.8cm} \begin{center} \begin{tabular}{ccc} \begin{fmffile}{diag1} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\bar{\chi}_1$,label.side=left}{v1,i2} \fmf{dashes,label=$S$}{v1,v2} \fmf{photon,label=$Z'$,label.side=left}{v2,o1} \fmf{photon,label=$Z'$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag2} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\chi_1$,label.side=left}{v1,v2} \fmf{fermion,label=$\bar{\chi}_1$,label.side=left}{v2,i2} \fmf{photon,label=$Z'$,label.side=left}{v1,o1} \fmf{dashes,label=$S$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag3} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\chi_1$,label.side=left}{v1,v2} \fmf{fermion,label=$\bar{\chi}_1$,label.side=left}{v2,i2} \fmf{dashes,label=$S$,label.side=left}{v1,o1} \fmf{dashes,label=$S$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} \end{tabular} \end{center} \begin{center} \begin{tabular}{ccc} \begin{fmffile}{diag6} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,v2} \fmf{fermion,label=$\bar{\chi_2}$,label.side=left}{v2,i2} \fmf{photon,label=$Z$,label.side=left}{v1,o1} \fmf{photon,label=$Z$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag7} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_2$}{i1,v1} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,v2} \fmf{fermion,label=$\bar{\chi_2}$,label.side=left}{v2,i2} \fmf{photon,label=$Z$,label.side=left}{v1,o1} \fmf{photon,label=$Z$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag8} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1^+$}{i1,v1} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,v2} \fmf{fermion,label=$\chi_1^-$,label.side=left}{v2,i2} \fmf{photon,label=$W$,label.side=left}{v1,o1} \fmf{photon,label=$W$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} \end{tabular} \end{center} \begin{center} \begin{tabular}{ccc} \begin{fmffile}{diag9} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{photon,label=$Z$,label.side=left}{v1,v2} \fmf{fermion,label=$q$,label.side=right}{i2,v2} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,o1} \fmf{fermion,label=$q$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} \begin{fmffile}{diag4} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,v2} \fmf{photon,label=$Z$,label.side=left}{v2,i2} \fmf{photon,label=$Z$,label.side=left}{v1,o1} \fmf{fermion,label=$\chi_2$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag5} \begin{fmfgraph*}(100,100) \fmfleft{i1,i2} \fmfright{o1,o2} \fmf{fermion,label=$\chi_1$}{i1,v1} \fmf{fermion,label=$\chi_2$,label.side=left}{v1,v2} \fmf{photon,label=$W$,label.side=left}{v2,i2} \fmf{photon,label=$Z$,label.side=left}{v1,o1} \fmf{fermion,label=$\chi_1^\pm$,label.side=right}{v2,o2} \end{fmfgraph*} \end{fmffile} \end{tabular} \end{center} \caption{Top row: new diagrams for $\chi_1 \bar{\chi_1}$ annihilation involving $S$ absent in a Simplified Model where the $Z'$ gets its mass using the St\"uckelberg mechanism. Middle row: example diagrams which contribute to the co-annihilation cross section into weak bosons. Bottom row: example diagrams which allow the conversion processes $\chi_1 \psi \to \chi_2 \psi$ ans $\chi_1 \psi \to \chi_1^\pm \psi$ (where $\psi$ is any Standard Model particle) to occur, which are suppressed by one factor of the singlet-doublet mixing. The first process on this row strongly dominates the interconversion rates in the dark sector due to the low quark masses.\label{fig:annihilation} } \end{figure} Besides the $Z'$ and $S$ funnel regions, where resonance DM annihilation significantly loosens the relic density constraints, we note that the $m_{\chi_1}/2 \ll m_{Z'}$ region always overcloses in the non-coannihilating scenario. Conversely, in the $m_{\chi_1} \gtrsim 2 m_{Z'}$ region, the DM relic density decreases when the DM mass increases. This rather counter-intuitive behaviour is due to the fact that this mass is proportional to $v_S$ and $y_\chi$. For a fixed value of $m_{Z'}$ and $g_{Y'}$, with hence fixed $v_S$, increasing this mass automatically implies increasing the associated Yukawa coupling, which in turn leads to higher DM annihilation rates through $S$ interactions. We therefore expect perturbativity constraints to set upper limits on the masses of the particles in the dark sector. This is demonstrated in the bottom right plot in Figure~\ref{fig:relic} where we show the value of $y_\chi$ as a function of $m_{\chi_1}$ for fixed $g_{Y'}, v_S, m_{Z'}$ as a red line. In the coannihilating models with $\Delta m = 2,10$~\%, the shapes of the funnel and $m_{\chi_1} \gtrsim 2 m_{Z'}$ regions are mostly unchanged although the relic density constraints significantly relax at low $\Delta m$. Part of the $m_{\chi_1} \lesssim 2 m_{Z'}$ region, however, is now allowed and the relic density constraints in this area now translate into an upper bound on $m_{\chi_1}$ that does not depend on $m_{Z'}$. This behaviour is due to the fact that the new coannihilation processes that lower the DM relic density in this region are annihilations of dark sector fermions into SM gauge bosons and thus do not involve the $Z'$. These observations are confirmed by the one-dimensional profile of $\Omega h^2$ on the bottom-right plot of figure~\ref{fig:relic}. For the non-coannihilating scenario, the relic density is much larger than its Planck value and overall decreases when the DM mass increases. When coannihilation is significant, on the other hand, the relic density first increases with the DM mass before decreasing again after the $Z'$ funnel. Note also the appearance of the $W$ and $Z$ funnels for coannihilating scenarios where SM electroweak processes dominate at low $m_{\chi_1}$. \begin{figure}[!tbp] \centering \includegraphics[width=0.49\textwidth]{relic_nocoann.png} \includegraphics[width=0.49\textwidth]{relic_withcoann_10p.png} \\ \includegraphics[width=0.49\textwidth]{relic_withcoann_2p.png} \includegraphics[width=0.49\textwidth]{relic_1d.png} \caption{The relic density after freeze-out for $g_{Y'} = 0.4$ as a function of $m_{Z'}$ and $m_{\chi_1}$ for $m_S = 1500$ GeV, $y_\theta = 1.5 y_\chi$. Results are presented as a function of $m_{\chi_1}$ and $m_{Z'}$ for $y_\phi \gg y_\chi$ (no coannihilation, top left), $\Delta m = 10 \%$ (top right), $\Delta m = 2 \%$ (bottom left), and as a function of $m_{\chi_1}$ for $v_S = 1000$ and $g_{Y'} = 0.4, 1.0$ for these scenarios (bottom right). In the 2D plots the red dashed line shows the contour with the correct relic density for $g_{Y'} = 0.4$, whereas the black dashed line shows the same contour for $g_{Y'} = 1.0$ for comparison. \label{fig:relic}} \end{figure} Finally, note that indirect detection experiments will also place competitive constraints on the model, especially in the high-$m_{Z'}$ region (see \cite{Jacques:2016dqz} for a comparative study for a model similar to ours in the limit where all additional fermions are heavy enough to decouple). We will focus on the impact on direct detection, relic density, and LHC constraints from not decoupling the additional fermions in the model here, and leave a detailed study of indirect detection constraints to future work. \subsection{LHC Searches} \label{sec:lhcpheno} For models that evade the direct detection bounds on the $\chi_1$--$\chi_2$ mixing discussed in section~\ref{sec:dmobs}, the leading experimental constraints are set by colliders, and notably by the $13$~TeV LHC searches. Dijet resonance searches are particularly sensitive in the off-shell region where $m_{\chi_1} > m_{Z'}/2$ where the $Z'$ branching ratio to jets is always $100$~\% \cite{Bauer:2018egk,Ellis:2018xal,Caron:2018yzp}, but only down to $m_{Z'} \sim 450$~GeV, which is the minimal $Z'$ mass probed by $jj$ searches at the LHC, while lower values are probed by $jj$ + ISR searches at the LHC \cite{Aaboud:2017yvp,Aaboud:2018zba,ATLAS-CONF-2016-070,Aaboud:2018fzt}\footnote{We use the combination of all of these searches derived in \cite{Ellis:2018xal} with $g Y'_q = \frac{2}{9} g_{Y'}$ in our model to assess these constraints.}. Other particularly sensitive probes are direct searches for pair-produced $e^+ e^- \to \chi_1^\pm \overline{\chi_1^\pm}$ charged fermions decaying to a lighter stable neutral fermion $\chi_{1}$ by emitting a $W$ at LEP. In the MSSM, LEP results rule out charginos lighter than $92.4$ GeV in the case where the two charged and neutral fermions are pure doublet Higgsino components \cite{Egana-Ugrinovic:2018roi}. This limit can not be directly translated to our model but suggests that the $Z$ and $W$ funnels that can occur in the co-annihilating case are firmly ruled out. A third source of collider constraints are monojet searches for $p p \to \chi_1 \overline{\chi_1}\, j$ production where the additional jet comes from ISR. These searches will be competitive at $m_{Z'} < 450$ GeV and $m_{\chi_2/\chi_1^\pm} > 100$ GeV where the dijet constraints are weaker and there are no LEP constraints. When the additional fermions in the model are close in mass to $\chi_1$ there are additional $pp \to \chi_2 \overline{\chi_2} j$\dots processes that also will contribute significantly to the monojet cross section (since the additional decay products of the heavier fermions in the dark sector might be too soft to be picked up by the detector, or invisible in the case of $\chi_2 \to \nu \bar{\nu} \chi_1$) and have to be taken into account. An additional sensitive channel that has so far been neglected for gauge-portal models is direct searches for pair production of $\chi_2,\chi_1^\pm$ at the LHC. In the coannihilation region these particles will often decay to $\chi_1$ promptly by emitting a $Z$ or a $W$ and will therefore have SUSY-like signatures with leptons and jets plus missing transverse energy. The usefulness of searches for these signatures for $U(1)'$ extensions of the MSSM was recently studied in detail in \cite{Araz:2017wbp}. In general the pair-production cross sections for $\chi_2$ and $\chi_1^\pm$ do not depend on $g_{Y'}$ when this coupling is small since SM gauge interactions dominate, which generates a signal which can compete in sensitivity with monojet searches even when the couplings to the $Z'$ are very weak. Note that constraints from these searches can be expected to be most significant when these additional particles are not too heavy. This feature suggests a kind of complementarity between these searches and the relic density constraints since the latter can often be alleviated by bringing the masses of the dark fermions down to the $m_{\chi_2/\chi_1^\pm}\sim m_{\chi_1}$ coannihilation region. \begin{figure}[h] \hspace*{-.8cm} \begin{center} \begin{tabular}{ccc} \begin{fmffile}{diag11} \begin{fmfgraph*}(150,120) \fmfleft{i1,i2} \fmfright{o1,o2,o3,o4} \fmf{fermion,label=$q$,label.side=left,tension=2}{i1,v1} \fmf{fermion,label=$\bar{q}$,label.side=left,tension=2}{v1,i2} \fmf{photon,label=$Z'$,tension=2}{v1,v2} \fmf{fermion,label=$\chi_2$,label.side=right,tension=2}{v2,v3} \fmf{fermion,label=$\bar{\chi}_2$,label.side=right,tension=2}{v4,v2} \fmf{fermion,label=$\chi_1$,label.side=left}{v3,o2} \fmf{fermion,label=$\bar{\chi}_1$,label.side=left}{o3,v4} \fmf{photon,label=$Z/Z'$,label.side=left}{o1,v3} \fmf{dashes,label=$H$}{o4,v4} \end{fmfgraph*} \end{fmffile} & \begin{fmffile}{diag12} \begin{fmfgraph*}(150,120) \fmfleft{i1,i2} \fmfright{o1,o2,o3,o4} \fmf{fermion,label=$q$,label.side=left,tension=2}{i1,v1} \fmf{fermion,label=$\bar{q}$,label.side=left,tension=2}{v1,i2} \fmf{photon,label=$Z'$,tension=2}{v1,v2} \fmf{fermion,label=$\chi_1^+$,label.side=right,tension=2}{v2,v3} \fmf{fermion,label=$\chi_1^-$,label.side=right,tension=2}{v4,v2} \fmf{fermion,label=$\chi_1$,label.side=left}{v3,o2} \fmf{fermion,label=$\bar{\chi}_1$,label.side=left}{o3,v4} \fmf{photon,label=$W$,label.side=left}{o1,v3} \fmf{photon,label=$W$}{o4,v4} \end{fmfgraph*} \end{fmffile} \end{tabular} \end{center} \caption{ Example pair production processes which contribute to the constraints derived from LHC analyses. Note that the decays of the heavier fermions are three-body if the mass difference to $\chi_1$ is $\ll m_Z/m_{Z'}/m_W$. The branching ratios of the heavier fermions for our benchmark points are given in Tables~\ref{tab:points_dd}-\ref{tab:points_nonthermal}. \label{fig:lhcprocesses} } \end{figure} To demonstrate the importance of these electroweakino searches and compare them to the monojet searches we use \textsc{CheckMATE} 2.0.26 to scan all of the implemented 13 TeV searches for the following processes: \begin{itemize} \item $pp \to \chi_m \overline{\chi_n} j$ \item $pp \to \chi_m \overline{\chi_n}$ \end{itemize} where $\chi_{m/n}$ is summed over all neutral and charged fermions in the dark sector, and the jet is required to have a $p_T > 200$ GeV in order to fill the phase space where monojet searches are sensitive. Some example diagrams contributing to these processes are presented in Figure~\ref{fig:lhcprocesses}. In general the collider phenomenology is somewhat similar to the Dirac bino -- Higgsino limit of the MSSM, with the addition of the $Z'$ which allows for much more efficient s-channel pair production than through the $Z/W$ when $m_{\chi_1} \gg m_{Z}$. The electroweakino mono-$X$ phenomenology of the MSSM and NMSSM was recently studied in detail in \cite{Bernreuther:2018nat} and shares many features with our model. Due to the computational cost of performing a sensitivity analysis over all the phase space of our model we leave a parameter scan taking all constraints into account to future work and focus on a few benchmark points where we generate $5\times 10^5$ events per process with minimal phase space cuts. Since many processes often contribute to the same signal region it is necessary to include them all and add up the contributions to derive the strongest constraints possible. Depending on $m_{\chi_1}$ and $\Delta m$ the visible decay products can be either hard ---allowing to reconstruct the $W/Z$ masses--- or soft, and some three-body decays can be phase-space suppressed and soft compared to others. If the mass splitting is large but pairs of $\chi_2,\chi_1^\pm$ still have non-negligible production cross sections there can be new decay channels opening up, such as $\chi_2 \to h \chi_1$ which can be sensitive to unexpected searches for new signatures like opposite-sign (OS) tau pairs and missing energy. The lifetimes of $\chi_2$ and $\chi_1^\pm$ for the points we consider are shorter than for $b$ mesons in order to be able to consider these prompt. In general smaller mass splitting will further increase the lifetime of the heavier fermions. In this nearly degenerate region where long-lived searches can be expected to be highly sensitive to $\chi_1^\pm$, higher order mass corrections need to be taken into account properly. We leave the analysis of this region to future work, but note that this again suggests the model is sensitive to the wider search program at the LHC for well-motivated parameter choices and that the full mixing scenario needs to be taken into consideration to evaluate these. Here we choose six representative benchmark points that all approximately give the correct relic density within 20\% and are summarised in Tables~\ref{tab:points_dd},\ref{tab:points_nonthermal}. These points all escape current collider constraints and show how collider searches provide complementary information in the coannihilation region, however the points in Table~\ref{tab:points_nonthermal} are in tension with the latest PandaX-II constraints on the momentum-suppressed spin-independent operator for direct detection. We include them here as examples of possible collider signatures in the model when the direct detection constraints are lifted slightly. \begin{table}[] \centering \begin{tabular}{lccc} Benchmark point & \textbf{[1]} & \textbf{[2]} & \textbf{[3]} \\ \hline $g_{Y'}$ & 0.6 & 0.3 & 0.2 \\ $m_{Z'}$ & 3000 GeV & 1500 GeV & 1000 GeV \\ $m_{S}$ & 1500 GeV & 1500 GeV & 1500 GeV \\ $m_{\chi_1}$ & 200 GeV & 300 GeV & 200 GeV \\ $m_{\chi_2}$ & 210 GeV & 312 GeV & 210 GeV \\ $m_{\chi_1^\pm}$ & 210 GeV & 312 GeV & 210 GeV \\ $d_{\chi_2}$ ($\frac{|p|}{m} = 1$) & 3.8 $\times$ 10$^{-4}$ m &8.1 $\times$ 10$^{-5}$ m & 3.8 $\times$ 10$^{-4}$ m \\ $d_{\chi_1^\pm}$ ($\frac{|p|}{m} = 1$) & 1.4 $\times$ 10$^{-4}$ m & 3.0 $\times$ 10$^{-5}$ m & 1.4 $\times$ 10$^{-4}$ m \\ $\text{Br}(\chi_2 \to \chi_1 q \bar{q})$ & 63\% & 61.5\% & 63\% \\ $\text{Br}(\chi_2 \to \chi_1 b \bar{b})$ & 0\% & 4.4\% & 0\% \\ $\text{Br}(\chi_2 \to \chi_1 l^+ l^-)$ & 8.2\% & 7.9\% & 8.2\% \\ $\text{Br}(\chi_2 \to \chi_1 \tau^+ \tau^-)$ & 3.3\% & 3.4\% & 3.3\% \\ $\text{Br}(\chi_2 \to \chi_1 \nu \bar{\nu})$ & 23.9\% & 22.8\% & 23.9\% \\ $\text{Br}(\chi_1^\pm \to \chi_1 q \bar{q})$ & 67.0 \% & 66.8\% & 67.0\% \\ $\text{Br}(\chi_1^\pm \to \chi_1 l \bar{\nu_l})$ & 23.0\% & 22.9\% & 23.0\% \\ $\text{Br}(\chi_1^\pm \to \chi_1 \tau \bar{\nu_\tau})$ & 10.0\% & 10.3\% & 10.0\% \end{tabular} \caption{Summary of benchmark points for scan of LHC searches which can explain the relic density while avoiding all other constraints. In the branching ratios $q$ denotes light quarks and $l$ an electron or muon. The mixing in the neutral fermion sector is set to be small enough to avoid direct detection constraints, which will in general mean it's small enough to be phenomenologically negligible for all other observables except for the lifetimes of the heavier fermions. As an example of these lifetimes, the smallest 10 GeV splitting between $\chi_2$ and $\chi_1$ and $\Delta m = 5\%$ with a mixing within direct detection bounds of points \textbf{[1]}, \textbf{[3]} give a $\chi_2$ lifetime of the same order of magnitude as $b$-mesons ($\sim 1.5$ picosecond) and is therefore approaching the limit for where we can ignore searches for long-lived particles, especially for smaller masses where the boosts can be significant. \label{tab:points_dd} } \end{table} \begin{table}[] \centering \begin{tabular}{lcccc} Benchmark point & \textbf{[4]} & \textbf{[5]} & \textbf{[6]} \\ \hline $g_{Y'}$ & 0.2 & 0.5 & 0.8 \\ $m_{Z'}$ & 400 GeV & 380 GeV & 240 GeV \\ $m_{S}$ & 1500 GeV & 1500 GeV & 1000 GeV \\ $m_{\chi_1}$ & 150 GeV & 150 GeV & 60 GeV \\ $m_{\chi_2}$ & 165 GeV & 250 GeV & 600 GeV \\ $m_{\chi_1^\pm}$ & 165 GeV & 250 GeV & 600 GeV \\ $d_{\chi_2}$ ($\frac{|p|}{m} = 1$) & 7.7 $\times$ 10$^{-6}$ m &2.1 $\times$ 10$^{-9}$ m & 1.9 $\times$ 10$^{-11}$ m \\ $d_{\chi_1^\pm}$ ($\frac{|p|}{m} = 1$) & 3.0 $\times$ 10$^{-6}$ m & 3.6 $\times$ 10$^{-10}$ m & 4.3 $\times$ 10$^{-11}$ m \\ $\text{Br}(\chi_2 \to \chi_1 q \bar{q})$ & 59.8 \% & $\text{Br}(Z \to q \bar{q})$ & $\frac{\text{Br}(Z' \to q \bar{q})}{1.72} + \frac{\text{Br}(Z \to q \bar{q})}{4.76}+ \frac{\text{Br}(h\to q \bar{q})}{5.0}$ \\ $\text{Br}(\chi_2 \to \chi_1 b \bar{b})$ & 7.8 \% & $\text{Br}(Z \to b \bar{b})$ & $\frac{\text{Br}(Z' \to b \bar{b})}{1.72} + \frac{\text{Br}(Z \to b \bar{b})}{4.76}+ \frac{\text{Br}(h\to b \bar{b})}{5.0}$ \\ $\text{Br}(\chi_2 \to \chi_1 l^+ l^-)$ & 7.4 \% & $\text{Br}(Z \to l^+ l^-)$ & $\frac{\text{Br}(Z \to l^+ l^-)}{4.76}+ \frac{\text{Br}(h\to l^+ l^-)}{5.0}$ \\ $\text{Br}(\chi_2 \to \chi_1 \tau^+ \tau^-)$ & 3.4 \% & $\text{Br}(Z \to \tau^+ \tau^-)$ & $\frac{\text{Br}(Z \to \tau^+ \tau^-)}{4.76}+ \frac{\text{Br}(h\to \tau^+ \tau^-)}{5.0}$ \\ $\text{Br}(\chi_2 \to \chi_1 \nu \bar{\nu})$ & 21.6 \% & $\text{Br}(Z \to \nu \bar{\nu})$ & $\frac{\text{Br}(Z \to \nu \bar{\nu})}{4.76}^\dagger$ \\ $\text{Br}(\chi_1^\pm \to \chi_1 q \bar{q})$ & 67.0 \% & $\text{Br}(W \to q \bar{q})$ & $0.98 \times \text{Br}(W \to q \bar{q})^{\dagger \dagger}$ \\ $\text{Br}(\chi_1^\pm \to \chi_1 l \bar{\nu_l})$ & 22.5 \% & $\text{Br}(W \to l \bar{\nu_l})$ & $0.98 \times\text{Br}(W \to l \bar{\nu_l})^{\dagger \dagger}$ \\ $\text{Br}(\chi_1^\pm \to \chi_1 \tau \bar{\nu_\tau})$ & 10.5 \% & $\text{Br}(W \to \tau \bar{\nu_\tau})$ & $0.98 \times\text{Br}(W \to \tau \bar{\nu_\tau})^{\dagger \dagger}$ \end{tabular} \caption{Summary of non-thermal benchmark points for scan of LHC searches. For point \textbf{[6]} we write out the decay modes contributing to the final branching ratio for an experimental signature to demonstrate that larger mass hierarchies in the dark sector means these can no longer be described by (phase-space suppressed) $W/Z$ branching ratios but rather can include cascade decays into $Z'$ and $h$ with modified kinematics, introducing new signatures. \newline $^\dagger$: there is also an invisible $\text{Br}(Z' \to \chi_1 \overline{\chi_1})$ contribution which is experimentally indistinguishable. \newline $^{\dagger \dagger}$: the total branching ratios are slightly rescaled due to rare $\chi_1^\pm \to \chi_1 t\bar{b}, \chi_1 h W, \chi_1 Z W, \chi_1 Z' W$ three-body decays. \label{tab:points_nonthermal} } \end{table} We present the results using the $r$ value for the most sensitive signal regions defined by \textsc{CheckMATE}, \begin{equation} r = \frac{S - 1.64 \Delta S}{S95} \end{equation} where $S$ is the predicted number of signal events, $\Delta S$ is the uncertainty on the signal prediction, and $S95$ is the reported 95\% C.L. limit on the BSM cross section in the signal region. Here, we take $\Delta S$ to be solely the statistical uncertainty due to insufficient simulated events and do not consider any theoretical systematic uncertainties or $K$-factors. A point is therefore considered excluded at 95 \% C.L. if $r > 1$. Note that since we are only considering models that are allowed by dijet resonance bounds, searches giving $r> 1$ will be the leading collider probe for the corresponding benchmark point. \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[1]} \\ \hline ATLAS 36.1 fb$^{-1}$ jets + MET \cite{Aaboud:2017vwy} & $r = 0.08$, 2j-2100 \\ ATLAS 36.1 fb$^{-1}$ soft OS lepton pair + MET \cite{Aaboud:2017leg} & $r = 0.33$, SRI-MLL-10 \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.12$, EM10 \\ CMS 12.9 fb$^{-1}$ soft OS lepton pair + MET \cite{CMS-PAS-SUS-16-025} & $r = 0.061$, stop low MET low $p_{T,l_1}$ \\ CMS 35.9 fb$^{-1}$ soft OS lepton pair + MET \cite{Sirunyan:2018iwl} & $r = 0.049$, weakino low MET high $m_{ll}$ \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[1]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. The two last results where a similar analysis with less data is more sensitive can be explained by differences in the base selection criteria. All the signal regions shown here require at least $\slashed{E}_T \gtrsim 100~$GeV. SRI-MLL-10 is inclusive in lepton flavor and select lepton pairs with invariant mass $m_{\ell\ell}\in[1, 20]~$GeV. EM2 simply requires $\slashed{E}_T\in[300,350]~$GeV. Finally, the last region targets relatively compressed spectra and cuts on the leading lepton $p_T$.\label{tab:point1cons}} \end{table} \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[2]} \\ \hline ATLAS 36.1 fb$^{-1}$ soft OS lepton pair + MET \cite{Aaboud:2017leg} & $r = 0.66$, SRI-MLL-20 \\ ATLAS 14.8 fb$^{-1}$ isolated lepton + jets + MET \cite{ATLAS-CONF-2016-054}& $r = 0.12$, GG2J \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.45$, EM10 \\ CMS 12.9 fb$^{-1}$ soft OS lepton pair + MET \cite{CMS-PAS-SUS-16-025} & $r = 0.21$, stop low MET low $p_{T,l_1}$ \\ CMS 35.9 fb$^{-1}$ soft OS lepton pair + MET \cite{Sirunyan:2018iwl} & $r = 0.18$, stop medium MET high $p_{T,l_1}$ \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[2]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. The EM2 region is the same as the one shown in Table~\ref{tab:point1cons}. SR$_W^{\mathrm{3-body}}$-SF is originally designed for stops with $\Delta m(\widetilde{t}, \chi)\sim m_W$ and hence imposes a $b$-quark veto and cuts on a ``super-Razor'' variable $M_\Delta^R$, that reaches an endpoint near the stop-neutralino mass splitting. 3LI simply imposes a moderate missing $E_T$ cut and a tight requirement on the $p_T$ of the third lepton. Finally, SR A25 imposes a harder $\slashed{E}_T$ cut as well as a transverse mass cut, and tags $Z$ bosons.\label{tab:point2cons}} \end{table} \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[3]} \\ \hline ATLAS 36.1 fb$^{-1}$ $\gamma$ + MET \cite{Aaboud:2017dor} & $r = 0.010$, SRI1 \\ ATLAS 36.1 fb$^{-1}$ jets + MET \cite{Aaboud:2017vwy} & $r = 0.024$, 2j-1200 \\ ATLAS 36.1 fb$^{-1}$ soft OS lepton pair + MET \cite{Aaboud:2017leg} & $r = 1.3$, SRI-MLL-10 \\ ATLAS 14.8 fb$^{-1}$ iso lepton + jets + MET \cite{ATLAS-CONF-2016-054} & $r = 0.033$, GG2J \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.24$, EM2 \\ CMS 12.9 fb$^{-1}$ soft OS lepton pair + MET \cite{CMS-PAS-SUS-16-025} & $r = 0.17$, stop low MET low $p_{T,l_1}$ \\ CMS 35.9 fb$^{-1}$ soft OS lepton pair + MET \cite{Sirunyan:2018iwl} & $r = 0.15$, weakino high MET low $p_{T,l_1}$ \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[3]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. SRI1, EM2, and the stop low MET low $p_{T,\ell_1}$ regions are already described in Table~\ref{tab:point1cons}. 2j-1200 imposes a high MET cut , requires at least two hard jets with invariant mass larger than $1.2$~TeV. SRI-MLL-10 is inclusive in lepton flavour and select lepton pairs with invariant mass $m_{\ell\ell}\in[1, 10]~$GeV. GG2j is similar to the regions used in monojet searches but also requires leptons. Finally, the weakino high MET low $p_{T, \ell_1}$ region requires a same-flavour opposite sign lepton pair and cuts on its invariant mass.\label{tab:point3cons}} \end{table} \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[4]} \\ \hline ATLAS 36.1 fb$^{-1}$ $\gamma$ + MET \cite{Aaboud:2017dor} & $r = 0.021$, SRI1 \\ ATLAS 36.1 fb$^{-1}$ OS $\tau$ pair + MET \cite{Aaboud:2017nhr} & $r = 0.058$, SR2 highmass \\ ATLAS 36.1 fb$^{-1}$ soft OS lepton pair + MET \cite{Aaboud:2017leg} & $r = 1.2$, SRI-MLL-30 \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.12$, EM2 \\ CMS 12.9 fb$^{-1}$ soft OS lepton pair + MET \cite{CMS-PAS-SUS-16-025} & $r = 3.3$, stop low MET low $p_{T,l_1}$ \\ CMS 35.9 fb$^{-1}$ soft OS lepton pair + MET \cite{Sirunyan:2018iwl} & $r = 1.1$, stop low MET low $p_{T,l_1}$ \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[4]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. The two last results where a similar analysis with less data is more sensitive can be explained by differences in the base selection criteria. All the signal regions shown here require at least $\slashed{E}_T \gtrsim 100~$GeV. The SRI1 signal region selects events with no more than two jets and a hard photon. The SR2 highmass region is designed for events with two hard taus, with large invariant mass and $m_{T2}$. SRI-MLL-30 is inclusive in lepton flavor and select lepton pairs with invariant mass $m_{\ell\ell}\in[1, 30]~$GeV. EM2 simply requires $\slashed{E}_T\in[300,350]~$GeV. Finally, the last region targets relatively compressed spectra and cuts on the leading lepton $p_T$.\label{tab:point1cons}} \end{table} \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[5]} \\ \hline ATLAS 13.3 fb$^{-1}$ 2 OS leptons + 2 $b$s + MET \cite{ATLAS-CONF-2016-076} & $r = 0.10$, SR$_W^{\text{3-body}}$-SF \\ ATLAS 13.3 fb$^{-1}$ 2,3 leptons + MET \cite{ATLAS-CONF-2016-096} & $r = 0.15$, 3LI \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.046$, EM2 \\ CMS 35.9 fb$^{-1}$ 2,3 leptons + MET \cite{Sirunyan:2017lae} & $r = 0.14$, SR A25 \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[5]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. The EM2 region is the same as the one shown in Table~\ref{tab:point1cons}. SR$_W^{\mathrm{3-body}}$-SF is originally designed for stops with $\Delta m(\widetilde{t}, \chi)\sim m_W$ and hence imposes a $b$-quark veto and cuts on a ``super-Razor'' variable $M_\Delta^R$, that reaches an endpoint near the stop-neutralino mass splitting. 3LI simply imposes a moderate missing $E_T$ cut and a tight requirement on the $p_T$ of the third lepton. Finally, SR A25 imposes a harder $\slashed{E}_T$ cut as well as a transverse mass cut, and tags $Z$ bosons.\label{tab:point2cons}} \end{table} \begin{table}[] \begin{tabular}{ll} Analysis & Constraint on \textbf{[6]} \\ \hline ATLAS 36.1 fb$^{-1}$ OS $\tau$ pair + MET \cite{Aaboud:2017nhr} & $r = 0.31$, SR2 highmass \\ ATLAS 36.1 fb$^{-1}$ jets + MET \cite{Aaboud:2017vwy} & $r = 0.18$, 4j-1400 \\ ATLAS 36.1 fb$^{-1}$ $j$ + MET \cite{ATLAS-CONF-2017-060} & $r = 0.49$, EM2 \\ \end{tabular} \caption{Summary of constraints on benchmark point \textbf{[6]} from the most sensitive 13 TeV searches implemented in \textsc{CheckMATE} at the LHC. The SR2 highmass and EM2 regions have been described in Table~\ref{tab:point1cons}. The 4j-1400 region imposes a hard cut on missing $E_T$, requires at least four hard jets with invariant mass larger than $1.4$~TeV.\label{tab:point4cons}} \end{table} The results for the four selected parameter points are summarised in Tables~\ref{tab:point1cons}--\ref{tab:point4cons}. Points \textbf{[1]} and \textbf{[3]}, which achieve the correct relic density through coannihilation processes, are out of the reach of the monojet searches implemented in \textsc{CheckMATE} but is excluded by searches for opposite-sign (OS) soft lepton pairs and MET \cite{Aaboud:2017leg,CMS-PAS-SUS-16-025,Sirunyan:2018iwl}. Point \textbf{[2]} which does not have coannihilation is allowed by the LHC searches studied here but is also more sensitive to searches for multiple leptons and MET \cite{ATLAS-CONF-2016-096,Sirunyan:2017lae} than to monojet searches. Only point \textbf{[4]} with its large mass splitting between the DM and the other fermions is most sensitive to the monojet search, and even still shows almost comparable sensitivity to a search for opposite-sign $\tau$s + MET in the high $m_{\tau \tau}$ signal region ($m_{\tau \tau} > 110$ GeV) \cite{Aaboud:2017nhr}, which can be traced back to the 20\% branching fraction of $\chi_2 \to h \chi_1$. These results demonstrate that monojet searches often do not provide the only or most sensitive collider constraints on anomaly-free gauge portal models which do not couple to leptons in the part of parameter space where dijet constraints are not applicable. Viable coannihilation scenarios, for example, are often particularly sensitive to searches for compressed SUSY spectra. These scenarios, with their compressed topologies and relaxed relic density constraints are expected to gain importance as collider and direct detection constraints become even more stringent in the rest of the parameter space\footnote{Of course direct detection constraints can also eventually push the allowed mixing in the fermion sector to be too small for coannihilation to occur.}. Even in scenarios without coannihilation it is possible to find interesting new signatures due to the heavier fermions if they are not completely decoupled, which often is necessary to avoid nonperturbative Yukawas after fixing the other masses in the model. Thus, even seemingly simple scenarios such as portal models are in fact associated with a wide variety of collider searches and a full study of these models requires the scan of as many published analyses and signal regions as possible, which mirrors the situation in for example the pMSSM. Finally, we note that aside from SUSY searches, studying constraints from Standard Model measurements using \textsc{Contur} could also lead to meaningful constraints on this class of models~\cite{Butterworth:2016sqg}. Finally, let us point out that if we loosen the relic density constraint to also allow relic densities which are smaller than the measured one, we will in general require larger $g_{Y'}$, which suggests even more sensitivity to LHC searches (assuming the direct detection constraint on the momentum-suppressed spin-independent operator still is avoided). \section{Conclusions} \label{sec:conclude} Simplified dark matter models where a fermion annihilates into SM particles via a vector mediator are popular benchmark scenarios. In order to avoid very strong direct detection and dilepton resonance constraints in these models it is convenient to keep the coupling of the vector boson to the DM candidate axial and make it leptophobic. This choice of parameters, however, introduces gauge anomalies when the vector is the gauge boson of a new broken $U(1)$ gauge group. We have demonstrated that the phenomenological consequences of avoiding such anomalies by enlarging the field content of the model can be wide-reaching even for the simplest possible solution. The additional particles allow for much richer scenarios than in simplified models, due to the presence of new weakly charged fermions and the possibility of coannihilation. It follows that constraints on simplified dark matter models which assume leptophobia and an axial DM--$Z'$ coupling should be understood as applying only to a specific corner of the full parameter space of the minimal realistic model. We have in effect played a game of Whac-A-Mole with experimental constraints, avoiding some but in the process introduced several new ones. In particular it is humorous that constructing a leptophobic model to avoid dilepton resonance constraints predicts new signatures for searches for opposite-sign lepton pairs. Notably, regions of the parameter space of the anomaly free theory can be better constrained using LHC searches for electroweakinos rather than monojet searches. These results further highlight the value of the full width of the BSM program at the LHC from a dark matter perspective. Finally, we note that the anomaly-free model studied here is only one example of how to extend gauge portal simplified models to make them theoretically consistent. A wider study of how to cancel anomalies in $Z'$ models without running into direct detection or LHC dilepton resonance constraints and the typical additional constraints that one should expect would be particularly useful. Additionally, an in-depth study of some of the more exotic signatures associated with our scenario such as long-lived particles would be of prime importance to fully understand this class of models. Although simplified models proved incredibly useful tools for exploring new physics model at colliders and dark matter experiments, it is crucial to keep their limitations in mind and question their minimality. The example of anomaly cancellation illustrates that even enforcing basic consistency requirements on these models can considerably enlarge the associated phenomenology, with a wide palette of new signatures to explore.

88539675ffaedabac35af99cc8366e08621082b9

\section{Introduction and motivations\label{sec:intro}} As a class of the simplest hadrons, heavy quarkonium is usually viewed as the ``hydrogen atom" in the strong-interaction theory QCD. While the knowledge of the nonperturbative aspect in QCD is still quite limited, heavy quarkonium provides a unique opportunity to probe the quark confinement in QCD by exploring the physics at the scale around the nonperturbative and perturbative boundary. The intrinsic scales of the heavy quark mass $m_Q$ and their binding energy $m_Qv^2$ lie in the perturbative and nonperturbative regimes respectively, where $v$ is the relative velocity between the heavy quark pair in the rest frame of the quarkonium. Due to the smallness of the relative velocity $v$ (e.g. $v^2\simeq 0.3$ and $v^2\simeq 0.1$ for the charmonium and bottomonium), the relativistic QCD can be reorganized via the operator product expansion in the power counting of $v$. The effective theory was dubbed as non-relativistic QCD (NRQCD)~\cite{Bodwin:1994jh}. The reformulation of QCD provides a factorization conjecture for calculating the rates of the quarkonium production and decay. In the case of the quarkonium ${\cal H}$ production, the (differential) cross section at leading-order (LO) in the QCD strong coupling constant $\alpha_s$ can be schematically written as \begin{eqnarray} d\sigma({\cal H})&=&\sum_{n}{d\hat{\sigma}(n)\langle\mathcal{O}^{\cal H}(n)\rangle},\label{eq:xsfac} \end{eqnarray} where $n$ represents a Fock state, $d\hat{\sigma}(n)$ is a perturbatively calculable short-distance coefficient (SDC) with the heavy quark pair in the Fock state $n$ and $\langle\mathcal{O}^{\cal H}(n)\rangle$ is the vacuum expectation number of an operator $\mathcal{O}^{\cal H}(n)$. If the factorization formula Eq.(\ref{eq:xsfac}) holds, the nonperturbative long-distance matrix element (LDME) $\langle\mathcal{O}^{\cal H}(n)\rangle$ is independent of quarkonium production process as well as the production environment. The universal LDMEs, which are analogous to the parton-distribution functions (PDFs) in the perturbative QCD factorization, are to be determined from a subset of the experimental data and to predict all of the rest experimental measurements. They have the probability explanations at LO, while LDMEs depend on the renormalization scheme and they are not physical objects. The prediction power of Eq.(\ref{eq:xsfac}) heavily relies on the perturbative convergences of $v^2$ and $\alpha_s$ in $d\sigma({\cal H})$. The leading power counting of various Fock states up to $\mathcal{O}(v^7)$ for S-wave and P-wave quarkonia is listed in Table.~\ref{tab:powcnt} according to the NRQCD velocity scaling rule~\cite{Bodwin:1994jh}. The convergence in $v^2$ can be improved by including the relativistic corrections. However, the prices to pay are that one has to introduce more nonperturbative LDMEs that can not be determined from the first principle, and the good relations like heavy-quark spin symmetry holding at LO in $v$ will be violated too. \begin{table}[ht!] \begin{center} \begin{tabular}{{|c|}*{4}{c|}}\hline Power counting & $\eta_Q$ & $\psi,\Upsilon$ & $h_Q$ & $\chi_{QJ}$\\\hline $v^3$ & ${\bigl.^1\hspace{-1mm}S^{[1]}_0}$ & ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ & $-$ & $-$\\ $v^5$ & $-$ & $-$ & ${\bigl.^1\hspace{-1mm}P^{[1]}_1},{\bigl.^1\hspace{-1mm}S^{[8]}_0}$ & ${\bigl.^3\hspace{-1mm}P^{[1]}_J},{\bigl.^3\hspace{-1mm}S^{[8]}_1}$\\ $v^7$ & ${\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}S^{[8]}_1},{\bigl.^1\hspace{-1mm}P^{[8]}_1}$ & ${\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}S^{[8]}_1},{\bigl.^3\hspace{-1mm}P^{[8]}_J}$ & $-$ & $-$ \\\hline \end{tabular} \end{center} \caption{The leading power counting of various Fock states contributing to various quarkonium within NRQCD velocity scaling rule~\cite{Bodwin:1994jh}.\label{tab:powcnt}} \end{table} The most subtle part is the $\alpha_s$ stability in the SDCs $d\hat{\sigma}(n)$, which is the main point to be discussed in this paper. For the high-transverse-momentum ($P_T$) quarkonium production at a high-energy hadron collider, it was found that ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ receives a giant K factor from QCD corrections to its SDC a decade ago~\cite{Campbell:2007ws}, which was understood by the fact that due to the quantum number conservation, there is a factor $\alpha_s \frac{P_T^2}{4m_Q^2}$ enhancement at $\mathcal{O}(\alpha_s^4)$ (next-to-leading order, NLO) compared to $\mathcal{O}(\alpha_s^3)$ (LO). This enhancement spoils the perturbative convergence in $\alpha_s$, shedding light on another possible enhancement from $\mathcal{O}(\alpha_s^5)$ (next-to-next-to-leading order, NNLO) corrections, while the accomplishment of the full NNLO calculation is even lacking today. The sole reason is the partonic cross sections $\frac{d\hat{\sigma}}{dP_T^2}$, before convoluting PDFs, are asymptotically scaling as $\left(\frac{2m_Q}{P_T}\right)^4\frac{1}{P_T^4}$ (next-to-next-to-leading power, NNLP) , $\left(\frac{2m_Q}{P_T}\right)^2\frac{1}{P_T^4}$ (next-to-leading power, NLP) and $\frac{1}{P_T^4}$ (leading power, LP) at LO, NLO, NNLO respectively.~\footnote{Rigorously speaking, the associated production of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ with the same flavoured heavy quark pair contributes $\mathcal{O}(\alpha_s^4)$ and is LP in $P_T$. We guide the readers to the discussion on this part in section~\ref{sec:charmfrag}.} Therefore, even with a full NNLO calculation at $\mathcal{O}(\alpha_s^5)$, the accuracy for the LP part of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ hadroproduction is still at LO level, while the NLP piece is indeed NLO accurate. A NLO accuracy of the LP contribution can only be achieved with a next-to-NNLO calculation in $\alpha_s$ for the SDC. The situation is slightly better though still similar for the other Fock states listed in Table~\ref{tab:powcnt}. Like ${\bigl.^1\hspace{-1mm}S^{[1,8]}_0},{\bigl.^3\hspace{-1mm}P^{[1,8]}_J},{\bigl.^1\hspace{-1mm}P^{[1,8]}_1}$, the NLP (LP) parts of SDCs appear firstly at LO (NLO) in $\alpha_s$. On the other hand, because of the same quantum number as the gluon, ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ has the leading $P_T$ behaviour as the jet, which means the LP channel is already opened at LO $\mathcal{O}(\alpha_s^3)$. In Table~\ref{tab:accuracy}, we have collected the first $\alpha_s$ powers in order to achieve the LO and NLO QCD accuracies for various Fock states at both LP and NLP in $P_T$. \begin{table}[ht!] \begin{center} \begin{tabular}{{|c|}*{7}{c|}}\hline \multirow{2}{*}{Accuracy} & \multicolumn{2}{c|}{${\bigl.^3\hspace{-1mm}S^{[1]}_1}$} & \multicolumn{2}{c|}{${\bigl.^3\hspace{-1mm}S^{[8]}_1}$} & \multicolumn{2}{c|}{${\bigl.^1\hspace{-1mm}S^{[1,8]}_0},{\bigl.^3\hspace{-1mm}P^{[1,8]}_J},{\bigl.^1\hspace{-1mm}P^{[1,8]}_1}$}\\\cline{2-7} & {\rm LP} & {\rm NLP} & {\rm LP} & {\rm NLP} & {\rm LP} & {\rm NLP}\\\hline {\rm LO} & $\alpha_s^5$ & $\alpha_s^4$ & $\alpha_s^3$ & $\alpha_s^3$ & $\alpha_s^4$ & $\alpha_s^3$\\\hline {\rm NLO} & $\alpha_s^6$ & $\alpha_s^5$ & $\alpha_s^4$ & $\alpha_s^4$ & $\alpha_s^5$ & $\alpha_s^4$\\\hline \end{tabular} \end{center} \caption{The first $\alpha_s$ orders needed in the SDCs for both LP and NLP in $P_T$ of various Fock states in their hadroproduction in order to achieve the LO and NLO QCD accuracies.\label{tab:accuracy}} \end{table} Following this observation, the complete NLO result for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ production is possible to be reproduced by the tree-level matrix element alone at $\mathcal{O}(\alpha_s^4)$ after introducing an {\it ad hoc} infrared cutoff. A first attempt was given in Ref.~\cite{Artoisenet:2008fc} to introduce an invariant-mass cut on any final-final and initial-final massless parton pairs, which was called NLO$^\star$. It can successfully reproduce the high-$P_T$ NLO calculation for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ production.~\footnote{Besides the single ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ production, NLO$^\star$ cut was also applied to ${\bigl.^3\hspace{-1mm}S^{[1]}_1}+{\bigl.^3\hspace{-1mm}S^{[1]}_1}$ hadroproduction in Ref~\cite{Lansberg:2013qka}. NLO$^\star$ calculation is able to well reproduce the complete NLO result~\cite{Sun:2014gca} in the double charmonium/bottomonium production. Its good performance may rely on the fact that like the single ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ production, the LO SDC of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}+{\bigl.^3\hspace{-1mm}S^{[1]}_1}$ is also NNLP in $P_T$ in the large transverse momentum region.} The same infrared cut can be imposed in the phase-space integration of the $\mathcal{O}(\alpha_s^5)$ tree-level matrix element. Another giant K factor was observed compared to the NLO calculation at high $P_T$, which may question on the extractions of color-octet LDMEs based on NLO calculations~\cite{Butenschoen:2010rq,Ma:2010yw,Chao:2012iv,Gong:2012ug,Bodwin:2014gia}. In contrast, a suspicion in Ref.~\cite{Ma:2010jj} on the size of $\mathcal{O}(\alpha_s^5)$ was given from their $P_T$ scaling reanalysis of the NNLO$^\star$ curves. Instead of the $P_T$ power enhancement, the observed giant K factor $\frac{d\sigma^{\rm NNLO^\star}}{d\sigma^{\rm NLO}}$ is mainly attributed to the introduction of the infrared cutoff. Therefore, a reliable estimate of the size of $\mathcal{O}(\alpha_s^5)$ is still missing. It is necessary to clarify the situation before drawing a solid conclusion. The aim of this paper is to introduce an infrared-safe method to cure the problematic giant K factors appearing in the SDC calculations in particular for high-$P_T$ quarkonium production without performing complete higher-order calculations.~\footnote{In the processes of elementary particle production, a few proposals to cure the giant K factors, which are mainly from logarithmic terms in perturbative calculations, are present~\cite{Caravaglios:1998yr,Mangano:2001xp,Catani:2001cc,Krauss:2002up,Lonnblad:2001iq, Lavesson:2005xu,Lonnblad:2011xx,Hamilton:2010wh,Hoche:2010kg,Lavesson:2008ah,Lonnblad:2012ix, Gehrmann:2012yg,Hoeche:2012yf,Frederix:2012ps,Lonnblad:2012ng,Hamilton:2013fea,Rubin:2010xp}. Unfortunately, none of them is straightforwardly applicable to the power-enhanced contributions in quarkonium production.} In contrast to the NLO$^\star$ calculations, the new method will not introduce the logarithmic dependence from the infrared cutoff. The estimate of the missing higher orders is to use the conventional renormalization and factorization scale variations. It is complemented with the fragmentation function approach, which requires the analytical calculations of different single- and double-parton fragmentation functions for single and multiple quarkonium production. Another nontrivial task to use the fragmentation function approach is to solve the corresponding coupled evolution equations. It has been shown in Ref.~\cite{Ma:2014svb} that the fragmentation function approach without scale evolution can reproduce the spin-summed NLO cross sections of ${\bigl.^3\hspace{-1mm}S^{[1]}_1},{\bigl.^3\hspace{-1mm}S^{[8]}_1},{\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[8]}_J}$ at high $P_T$, which shows the necessity of taking into account both the single-parton (at LP) and the double-parton (at NLP) fragmentation contributions. The factorization theorem for the single-inclusive quarkonium production cross sections in terms of single- and double-parton fragmentation functions was first proven in Ref.~\cite{Kang:2011mg} under the assumption of perturbative QCD factorization. There are also other appealing reasons to introduce such a method. First of all, it can be used to stabilize the higher-order QCD corrections in quarkonium associated production processes, where most of them are absent of complete NLO calculations. The possible cancellations between S-wave and P-wave are guaranteed in our approach. For instance, in the double $J/\psi$ at the LHC, it requires a NNLO calculation to have the full cancellations between S-wave and P-wave Fock states. As we will see later in this paper, the good reproduction of the NLO results both in the spin-summed and spin-dependent cross sections for single quarkonium production at high $P_T$ can serve as a fast way to the future phenomenology studies. In practice, the phenomenology from a complicated calculation scales as an inverse power of the computation time. The outline of the remaining context is following. After introducing the remainders of P-wave counterterms in section ~\ref{sec:pCTs}, we will show that one can reproduce the NLO results for most of the Fock states (except ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$) with fairly simple cuts based on tree-level matrix elements in section \ref{sec:simpleNLO}. These simple cuts are not sufficient to remove large logarithms introduced by the phase space cut parameters. Hence, a general infrared-safe method is introduced to obtain the giant K factors for all the Fock states relevant for $J/\psi$ and $\chi_{cJ}$ production in section \ref{sec:generalSTOP}. Finally, we draw our conclusions in section \ref{sec:summary}. An instruction on how to use {\sc\small HELAC-Onia}~\cite{Shao:2012iz,Shao:2015vga} to perform the calculations done in this paper is given in appendix \ref{app:helaconia}. The appendix \ref{app:moreplots} contains supplemental figures. \section{Remainders of P-wave counterterms\label{sec:pCTs}} It is well-known that the remaining infrared divergences in the SDC computations for the productions and decays of P-wave Fock states should be cancelled by the P-wave counterterms arising from the renormalization group running of S-wave LDMEs beyond LO in $\alpha_s$, which is analogous to the remaining collinear divergences absorbed by the PDF counterterms in a peturbative QCD calculation. The renormalization of NRQCD operators links the S-wave LDMEs with the P-wave LDMEs as shown in Eq.(150) of Ref.~\cite{Petrelli:1997ge}. Such counterterms, after cancelling infrared divergences with the real and virtual matrix elements, will leave finite remainders proportional to the S-wave SDCs and P-wave LDMEs. The introduction of the P-wave counterterms is crucial especially in the case that the S-wave SDCs are much larger than the P-wave SDCs. In particular, the negative P-wave SDCs for heavy quarkonium hadroproduction at high $P_T$ could be attributed to these negative remainders. We have implemented the following finite remainders of P-wave counterterms: \begin{eqnarray} d\sigma^{\mathcal{C}}({\bigl.^3\hspace{-1mm}P^{[8]}_J})&=&d\hat{\sigma}^{\rm Born}({\bigl.^3\hspace{-1mm}S^{[1]}_1})\times\left(\frac{4}{3}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^3\hspace{-1mm}P^{[8]}_J}) \rangle\nonumber\\ &&+d\hat{\sigma}^{\rm Born}({\bigl.^3\hspace{-1mm}S^{[8]}_1})\times\left(\frac{5}{9}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^3\hspace{-1mm}P^{[8]}_J}) \rangle,\nonumber\\ d\sigma^{\mathcal{C}}({\bigl.^3\hspace{-1mm}P^{[1]}_J})&=&d\hat{\sigma}^{\rm Born}({\bigl.^3\hspace{-1mm}S^{[8]}_1})\times\left(\frac{8}{27}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^3\hspace{-1mm}P^{[1]}_J}) \rangle,\nonumber\\ d\sigma^{\mathcal{C}}({\bigl.^1\hspace{-1mm}P^{[8]}_1})&=&d\hat{\sigma}^{\rm Born}({\bigl.^1\hspace{-1mm}S^{[1]}_0})\times\left(\frac{4}{3}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^1\hspace{-1mm}P^{[8]}_1}) \rangle\nonumber\\ &&+d\hat{\sigma}^{\rm Born}({\bigl.^1\hspace{-1mm}S^{[8]}_0})\times\left(\frac{5}{9}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^1\hspace{-1mm}P^{[8]}_1}),\nonumber\\ d\sigma^{\mathcal{C}}({\bigl.^1\hspace{-1mm}P^{[1]}_1})&=&d\hat{\sigma}^{\rm Born}({\bigl.^1\hspace{-1mm}S^{[8]}_0})\times\left(\frac{8}{27}\frac{\alpha_s}{\pi}\frac{\log{\frac{m_Q^2}{4\mu_{\Lambda}^2}}}{m_Q^2}\right)\times \langle \mathcal{O}({\bigl.^1\hspace{-1mm}P^{[1]}_1}) \rangle, \end{eqnarray} where $m_Q$ is the mass of the heavy quark and $\mu_{\Lambda}$ is the NRQCD scale. In the following, we will set $\mu_{\Lambda}=m_Q$ as usually done in the complete NLO calculations. These remainders have already been implemented in the {\sc\small HELAC-Onia}~\cite{Shao:2012iz,Shao:2015vga}. They are necessary ingredients to reproduce the NLO results, which we will show in the following two sections. \section{A first step towards NLO\label{sec:simpleNLO}} From the discussion in the section~\ref{sec:intro}, it is known that large NLO QCD corrections to the $J/\psi$ production at a high-energy hadron collider are mainly due to the emergence of new $P_T$ power-enhanced fragmentation contributions. Hence, all S- and P-wave Fock states except ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ receive giant K factors from NLO QCD calculations. We first introduce the following basic phase space cuts in order to take into account the hard radiations without using virtual amplitudes. In real part at $\mathcal{O}(\alpha_s^4)$, exact $2$ light-flavoured jets~\footnote{We mean ``light-flavoured jet" here as a cluster of gluon, up, down, strange (anti-)quarks. Similarly, the light-flavoured partons are defined as gluon, up, down, strange (anti-)quarks.} satisfying $P_T(j)>P_{T}^{\rm min}$ and $|y(j)|<y^{\rm max}$ are required, which is denoted as $d\sigma^{\mathcal{R}_0}$. The phase space integrations of Born $d\sigma^{\mathcal{B}}$ ($\mathcal{O}(\alpha_s^3)$) and the remainders of the NRQCD P-wave counterterms $d\sigma^{\mathcal{C}}$ ($\mathcal{O}(\alpha_s^4)$) are infrared safe with $P_T(\rm{onium})$ larger than a given positive value $P_T^{\rm min}(\rm{onium})$. We call the summed results of $d\sigma^{\mathcal{B}}+d\sigma^{\mathcal{R}_0}+d\sigma^{\mathcal{C}}$ as approximated NLO (aNLO). In the following, we take $P_T^{\rm min}(\rm{onium})=5$ GeV, and light-flavoured jets are clustered with anti-$k_T$ algorithm~\cite{Cacciari:2008gp} using radius $R=0.5$ and $|y(j)|<5,P_T(j)>P_{T}^{\rm min}$ by {\sc\small FastJet}~\cite{Cacciari:2011ma}. We will vary $P_{T}^{\rm min}$ from 3 GeV to 6 GeV as a way to estimate the infrared-cut dependence. We have shown the spin-summed double differential distributions for the $c\bar{c}$ Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ in Fig.~\ref{Fig:aNLOvsNLO} with the $13$ TeV proton-proton collisions, while the distributions for the 5 Fock states ${\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[8]}_J},{\bigl.^3\hspace{-1mm}P^{[1]}_0},{\bigl.^3\hspace{-1mm}P^{[1]}_1},{\bigl.^3\hspace{-1mm}P^{[1]}_2}$ are displayed in Fig.~\ref{Fig:aNLOvsNLO2} as our supplemental material. The complete NLO curves (denoting as NLO) from Refs.~\cite{Ma:2010vd,Ma:2010yw} are also shown in order to have a comparison. The red-hatched bands represent the infrared cut variations $P_{T}^{\rm min}\in\ [3, 6]$ GeV, and the grey bands are the uncertainty from the independent variations of renormalization and factorization scales $\mu_R,\mu_F$ around the central value $\mu_0=\sqrt{P_{T}^2({\rm{onium}})+4m_c^2}$ by a factor of 2. It is interesting to notice that the scale uncertainty in general captures the missing virtual and soft/collinear pieces. The agreements between NLO and aNLO are improved as $P_{T}({\rm{onium}})$ increases. A similar behaviour can be observed for the spin-dependent differential cross sections shown in Fig.~\ref{Fig:aNLOvsNLOSpin} for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ and in Figs.~\ref{Fig:aNLOvsNLOSpin2}, \ref{Fig:aNLOvsNLOSpin3} for ${\bigl.^3\hspace{-1mm}P^{[8]}_J},{\bigl.^3\hspace{-1mm}P^{[1]}_1},{\bigl.^3\hspace{-1mm}P^{[1]}_2}$ Fock states, where the NLO curves are from Refs.~\cite{Chao:2012iv,Shao:2014fca}. The spin-density matrix elements of the scalars ${\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[1]}_0}$ are trivial.~\footnote{The spin-density matrix elements $\frac{d\sigma_{J_zJ_z}}{dP_T}$ shown in this paper are defined in the usual helicity frame.} We have utilized CTEQ6M PDF~\cite{Pumplin:2002vw} to be consistent with the NLO results. For the reproducible purpose, the values of LDMEs for the distributions of the Fock states are listed in Table~\ref{tab:LDMEs}. \begin{figure}[ht!] \vspace{-1cm} \centering \includegraphics[width=.99\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_aNLO-crop.pdf} \caption{Comparison of spin-summed differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between our aNLO calculations and the complete NLO calculations.\label{Fig:aNLOvsNLO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma00_aNLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma11_aNLO-crop.pdf} \caption{Comparisons of spin-dependent differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between our aNLO calculations and the complete NLO calculations.\label{Fig:aNLOvsNLOSpin}} \end{figure} \begin{table}[ht!] \begin{center} \begin{tabular}{{|c|}*{5}{c|}}\hline $\langle\mathcal{O}({\bigl.^3\hspace{-1mm}S^{[1]}_1})\rangle$ & $\langle\mathcal{O}({\bigl.^3\hspace{-1mm}S^{[8]}_1})\rangle$ & $\langle\mathcal{O}({\bigl.^1\hspace{-1mm}S^{[8]}_0})\rangle$ & $\langle\mathcal{O}({\bigl.^3\hspace{-1mm}P^{[8]}_J})\rangle/(2J+1)$ & $\langle\mathcal{O}({\bigl.^3\hspace{-1mm}P^{[1]}_J})\rangle/(2J+1)$\\\hline $1.16$ GeV$^3$ & $9.03\cdot 10^{-3}$ GeV$^3$ & $1.46\cdot 10^{-2}$ GeV$^3$ & $3.43\cdot 10^{-2}$ GeV$^5$ & $0.107$ GeV$^5$ \\\hline \end{tabular} \end{center} \caption{The values of LDMEs used in the differential distributions of various Fock states.\label{tab:LDMEs}} \end{table} Because the LP in $P_T$ for ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ already exists at Born $d\sigma^{\mathcal{B}}$ (i.e. $\mathcal{O}(\alpha_s^3)$) from the gluon fragmentation, it is expected that the scale uncertainty at LO would already give a reliable estimate of the missing NLO QCD corrections, which is indeed observed from the left-panel of Fig.~\ref{Fig:aNLO3S18vsNLO}. In such a case, a request of $2$ light-flavoured jets in the computation of $d\sigma^{\mathcal{R}_0}$ is insufficient to obtain an infrared-safe differential cross section. From the right-panel of Fig.~\ref{Fig:aNLO3S18vsNLO}, the aNLO $P_T$ spectra are too hard compared to the complete NLO ones. The reason is because of the large logarithms arising from the very asymmetric dijet system $P_T(j_1)\gg P_T(j_2)$. Such a configuration is suppressed in other Fock states, because the leading fragmentation topologies require at least one light-flavoured parton along with the quarkonium direction at high $P_T$. The weights of the asymmetric dijet events will be enhanced due to the unphysical logarithm $\log{\frac{P_T(j_1)}{P_T(j_2)}}$ in the aNLO calculations of ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$, which should be in principle cancelled by the virtual contributions because of the unitarity. Therefore, one must introduce a more general infrared-safe method to avoid these large logarithms, and at meantime one should maintain the hard radiations from the real contributions. \begin{figure}[hbt!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_LO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_nozasym-crop.pdf} \caption{Comparisons of spin-summed differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ between LO (left), aNLO (right) calculations and the complete NLO calculations.\label{Fig:aNLO3S18vsNLO}} \end{figure} \section{A general infrared-safe method\label{sec:generalSTOP}} \subsection{Infrared-safe cutoffs\label{sec:stopcuts}} \begin{figure}[hbt!] \centering \includegraphics[width=.90\textwidth,draft=false]{./figures/pp2oniumX-crop-compressed.pdf} \caption{Schematic depiction of inclusive quarkonium ${\cal O}_n$ production.\label{Fig:pp2oniumX}} \end{figure} \begin{figure}[hbt!] \centering \includegraphics[width=.90\textwidth,draft=false]{./figures/jets_infrared-crop-compressed.pdf} \caption{Infrared unsafe configurations to be considered in inclusive quarkonium production, where the first 3 subfigures are for the onium jet $j_{{\cal O}_n}$ and the last one is for the light-flavoured jets.\label{Fig:infraredconfigs}} \end{figure} Let us assume a LO fragmentation process for a given Fock state ${\cal O}_n$ is accompanying with $k$ final massless partons: \begin{eqnarray} {\cal F}: p_0 \rightarrow P_{{\cal O}_n}+p_1+\cdots+p_k. \label{eq:fragprocess} \end{eqnarray} For a given observable, one needs to consider ${\cal O}_n$ plus $i$ recoiling partons. For example, in the case of the transverse-momentum distribution for a single quarkonium production (schematically depicted in Fig.~\ref{Fig:pp2oniumX}), the quarkonium at least recoils against one parton at the lowest order (bar the zero transverse momentum bin). The soft- and collinear-safe calculations can be achieved based on pure tree-level matrix elements via the following conditions: \begin{enumerate} \item The number of jets is larger than $i+1$ with the transverse momentum of jet $P_T(j)>P_{T}^{\rm min}$ and the rapidity $|y(j)|<y^{\rm max}$. ${\cal O}_n$ is also taken into account in the jet-clustering procedure. One should make sure that there is exactly one jet containing ${\cal O}_n$ passing the above $P_T$ and rapidity cuts. Such a jet is called an {\it onium-jet} here. \item In the onium-jet, there are at least $k$ light-flavoured partons to fulfill the above fragmentation process. Let us say there are $m$ light-flavoured partons inside the onium-jet with $m\ge k$.~\footnote{At LO one should have $m=k$ since the configuration of $i+k-m<i$ recoiling partons is zero by definition for the given observable when $m>k$.} \item If $m \ge 2$, each pair of parton $1$ and $2$ inside the onium-jet should pass the following soft drop condition~\cite{Larkoski:2014wba} \begin{eqnarray} \frac{{\rm min}(p_{T,1},p_{T,2})}{p_{T,1}+p_{T,2}} > z_{\rm cut}\left(\frac{\Delta R_{12}}{R_0}\right)^\beta~\label{eq:softdropcut} \end{eqnarray} where $p_{T,i}$ is the transverse momentum of parton $i$ and $\Delta R_{12}=\sqrt{\Delta \phi_{12}^2+\Delta y_{12}^2}$. The above cut already excludes the soft singularity as long as $z_{\rm cut}>0$, while the requirement of the collinear safety is guaranteed by choosing $\beta<0$. $R_0$ is the original jet radius, which is an order one number. \end{enumerate} The condition Eq.(\ref{eq:softdropcut}) in item 3 is chosen to kill the infrared unsafe configurations (a) and (b) given in Fig.~\ref{Fig:infraredconfigs}. Either when partons $1$ and $2$ are close to be collinear $\Delta R_{12}\ll R_0$ or if one parton is soft $p_{T,2}\ll p_{T,1}$, Eq.(\ref{eq:softdropcut}) cannot be fulfilled when $z_{\rm cut}>0,\beta<0$. In practice, the absolute value of $\beta$ is at order one and $z_{\rm cut}$ is at the order of $v^2$. If one goes to extra $o$ radiations (i.e. ${\cal O}_n$ plus $i+k+o$ final light-flavoured QCD partons), one should impose the following additional cuts: \begin{enumerate} \item[4.]There are $i+k+o-m$ partons outside the onium-jet. Each parton should form a single jet within $P_T(j)>P_{T}^{\rm min}$ and $|y(j)|<y^{\rm max}$ to avoid the collinear divergences. In order to get rid of large logarithms from infrared cuts arising from the soft large-angle radiations illustrated in the case (d) in Fig.~\ref{Fig:infraredconfigs}, when $i+k+o-m\ge 2$, it is necessary to impose an asymmetric cut on these parton jets \begin{eqnarray} \frac{{\rm min}\left(P_T(j_1),\cdots, P_T(j_{i+k+o-m})\right)}{{\rm max}\left(P_T(j_1),\cdots, P_T(j_{i+k+o-m})\right)}> z_{\rm cut, a}. \end{eqnarray} The value of $z_{\rm cut,a}$ should be a positive number smaller than 1 but not close to 0. It is important to vary its value in order to assess this cut dependence. \item[5.]If $k=0$~\footnote{For example, ${\cal O}_n=Q\bar{Q}({}^3S_1^{[8]})$ has $g\rightarrow Q\bar{Q}({}^3S_1^{[8]})$ fragmentation at LO.} and $m>0$, each parton $1$ in the onium jet $j_{{\cal O}_n}$ should pass the soft cut \begin{eqnarray} z_1> z_{\rm cut,s} \end{eqnarray} where $z_1$ can be the energy fraction $\frac{E_1}{E(j_{{\cal O}_n})}$, the transverse mass fraction $\frac{p_{T,1}}{\sqrt{P_T^2(j_{{\cal O}_n})+M^2(j_{{\cal O}_n})}}$, the transverse momentum fraction $\frac{p_{T,1}}{P_T(j_{{\cal O}_n})}$ or other similar fractions corresponding to $z_1\rightarrow 0$ when the parton $1$ is soft. This condition is needed in order to kill the case (c) in Fig.~\ref{Fig:infraredconfigs}, where all light-flavoured partons in the onium-jet $j_{{\cal O}_n}$ can be soft and the condition Eq.(\ref{eq:softdropcut}) is still satisfied. Similar to the value of $z_{\rm cut}$ in Eq.(\ref{eq:softdropcut}), the proper value of $z_{\rm cut,s}$ should be $\mathcal{O}(v^2)$ as the effect of the soft radiations should be absorbed into the long-distance part of the quarkonium. \end{enumerate} In fact, the combination of items 1-5 introduces a general infrared-safe method for any Fock state production if $k=0$ is assumed at the beginning.~\footnote{When $k=0$, the cut in item 2 will not be applied.} In other words, we do not need to pay a special attention to which kind of fragmentation process $\mathcal{F}$ is allowed for a given Fock state. We call such cuts as STOP cuts, where ``STOP" is an acronym of ``STabilize quarkOnium Production". In the case of the $P_T$ spectrum of a quarkonium $\mathcal{O}_n$ production at a hadron collider, $i$ is equal to $1$ and the LO process is $\mathcal{O}_n$ plus one parton. For a real emission process $\mathcal{O}_n$ plus $o+1$ partons with $o>0$, we should impose the cuts listed in items 1-5 with $k=0$, where the condition in item 2 is fulfilled automatically. Same as the previous section, we will denote the Born contribution at $\mathcal{O}(\alpha_s^3)$ as $d\sigma^{\mathcal{B}}$ and the remainders of the P-wave counterterms at $\mathcal{O}(\alpha_s^4)$ as $d\sigma^{\mathcal{C}}$. $d\sigma^{\mathcal{R}_{\rm STOP}}$ ($d\sigma^{\mathcal{R}^2_{\rm STOP}}$) stands for the contribution from $\mathcal{O}_n$ plus two (three) partons within the STOP cuts. \subsection{Reproducing NLO results\label{sec:NLOStop}} In this section, we will present the results up to NLO QCD corrections (i.e. $\mathcal{O}(\alpha_s^4)$). In order to differentiate our partial NLO calculations with the complete NLO results, we will denote our partial NLO calculations by imposing STOP cuts as ``nLO", i.e. $d\sigma^{\rm nLO}\equiv d\sigma^{\mathcal{B}}+d\sigma^{\mathcal{C}}+d\sigma^{\mathcal{R}_{\rm STOP}}$. In the following, we will illustrate that the complete NLO results can be reproduced with the tree-level generators under the following setup of the STOP cuts: \begin{eqnarray} &&P_T(j)>P_T^{\rm min}, P_T^{\rm min}\in\ [3, 6]~{\rm GeV}, |y(j)|<5.0,\nonumber\\ &&z_{\rm cut}=0.1, \beta=-1, R_0=1.0,\nonumber\\ &&z_{\rm cut,a} \in\ [0.1,0.7], z_{\rm cut,s}=\frac{0.1}{m}, \end{eqnarray} where $m$ is the number of light-flavoured partons inside the onium jet. Jets are reconstructed with the anti-$k_T$ clustering algorithm using {\sc\small FastJet}. Since there is no infrared divergence in the Born after imposing $P_T({\rm onium})>0$ cut, the STOP cuts will not be applied to the Born and Born-like counterterm events. \subsubsection{Reproducing ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$} After imposing the STOP cuts on ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$, we can reproduce the complete NLO curves within the theoretical uncertainties. They are shown in Fig.~\ref{Fig:nLO3S18vsNLO} and Fig.~\ref{Fig:nLO3S18vsNLOSpin} for the spin-summed and spin-dependent differential cross sections respectively. In the left panel of Fig.~\ref{Fig:nLO3S18vsNLO} and the upper panels of Fig.~\ref{Fig:nLO3S18vsNLOSpin}, we estimate infrared cutoff dependence (the red-hatched bands) via the combined variations of $P_T^{\rm min}\in\ [3, 6]$ GeV and $z_{\rm cut,a} \in\ [0.1,0.7]$. The grey-shadowed bands represent the scale uncertainties. Opposed to the aNLO results in the right panel of Fig.~\ref{Fig:aNLO3S18vsNLO}, it indeed shows that the STOP cuts improve the perturbative calculations, and the transverse-momentum dependence in $\frac{d\sigma^{\rm nLO}}{dP_T}$ is the same as the NLO distributions $\frac{d\sigma^{\rm NLO}}{dP_T}$. It demonstrates that the large logarithmic dependence from the simple cuts in section \ref{sec:simpleNLO} disappears after imposing the STOP cuts. The STOP-cut dependence (the red-hatched bands) is not reduced by increasing the $P_T$ of the quarkonium. It is expected since the LP contribution is already present at LO. The variations of the STOP cut variables only alter the fractions of hard radiations in the real matrix elements, which are not logarithmically enhanced. In fact, a careful tuning of STOP cut parameters can reproduce the NLO results at high precision. In the right panel of Fig.~\ref{Fig:nLO3S18vsNLO} and the lower panels of Fig.~\ref{Fig:nLO3S18vsNLOSpin}, we have calculated the ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ differential distributions after using $z_{\rm cut,a}=0.6$ and $z_{\rm cut,s}=\frac{0.2}{m}$. The comparisons to the full NLO calculations imply that the $P_T$ spectra of ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ in different rapidity intervals can be precisely reproduced as long as $P_T({\bigl.^3\hspace{-1mm}S^{[8]}_1})>10$ GeV. \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_nLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_tune_nLO-crop.pdf} \caption{Comparisons of spin-summed differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ between nLO (left), tunned nLO with $z_{\rm cut,a}=0.6, z_{\rm cut,s}=\frac{0.2}{m}$ (right) calculations and the complete NLO calculations.\label{Fig:nLO3S18vsNLO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_dsigma00_nLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_dsigma11_nLO-crop.pdf}\\ \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_tune_dsigma00_nLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_tune_dsigma11_nLO-crop.pdf} \caption{Comparisons of spin-dependent differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ between nLO (up), tunned nLO with $z_{\rm cut,a}=0.6, z_{\rm cut,s}=\frac{0.2}{m}$ (down) calculations and the complete NLO calculations.\label{Fig:nLO3S18vsNLOSpin}} \end{figure} \subsubsection{Reproducing other Fock states} We are now in the position to check the calculations for the other Fock states with the STOP cuts. Like the case of the simple cuts in section~\ref{sec:simpleNLO}, the general infrared-safe STOP cuts can reproduce the complete NLO results within theoretical uncertainties well. These Fock states do not show LP behaviour at LO. The comparisons of nLO calculations to NLO calculations for spin-summed and spin-dependent differential cross sections in the states ${\bigl.^3\hspace{-1mm}S^{[1]}_1},{\bigl.^1\hspace{-1mm}S^{[8]}_0}$ are displayed in Fig.~\ref{Fig:nLOvsNLO} and Fig.~\ref{Fig:nLOvsNLOSpin} respectively, while we put the supplemental plots Figs.~\ref{Fig:nLOvsNLO2},~\ref{Fig:nLOvsNLOSpin2},~\ref{Fig:nLOvsNLOSpin3} for the other Fock states in the appendix \ref{app:moreplots}. With the scale variations shown by the grey bands, nLO results in general are able to successfully reproduce the NLO calculations in both cases. The only exception is the differential cross section of ${\bigl.^1\hspace{-1mm}S^{[8]}_0}$ (see the right panel of Fig.~\ref{Fig:nLOvsNLO}) at very large $P_T$, i.e. $P_T>90$ GeV. Such a discrepancy in ${\bigl.^1\hspace{-1mm}S^{[8]}_0}$ can be better understood from the LO fragmentation function $g\rightarrow {\bigl.^1\hspace{-1mm}S^{[8]}_0}+g$~\cite{Braaten:1996rp}, which has the functional form \begin{eqnarray} D_{g\rightarrow {\bigl.^1\hspace{-1mm}S^{[8]}_0}}(z_{\rm onium})&\propto&3z_{\rm onium}-2z_{\rm onium}^2+2(1-z_{\rm onium})\log{(1-z_{\rm onium})}, \end{eqnarray} where $z_{\rm onium}$ is the momentum fraction of ${\bigl.^1\hspace{-1mm}S^{[8]}_0}$. The function peaks at $z_{\rm onium}=1$. A finite value of $z_{\rm cut,s}$ in the STOP cuts will remove a non-negligible fraction of radiations in the LP contributions. In fact, we have explicitly checked that if we set $z_{\rm cut,s}=\frac{10^{-2}}{m}$ instead of $z_{\rm cut,s}=\frac{0.1}{m}$, the agreement between nLO and NLO results are significantly improved at large $P_T$, which can be found in Fig.~\ref{Fig:nLO1S08vsNLO}. In the spirit of the NRQCD factorization, the soft gluons from the heavy quark pair with the momentum fraction smaller than $v^2$ should be absorbed into the LDMEs as well as their energy evolutions, where $v^2$ is around $0.3$ for the charmonium. Therefore, without taking into account the relativistic corrections, the resolution of NRQCD in describing the heavy quarkonium production should be not better than $v^2$. Hence, it is not straightforward to judge which is a better choice between the two different values $z_{\rm cut,s}=\frac{0.1}{m}$ and $z_{\rm cut,s}=\frac{10^{-2}}{m}$. In fact, we believe $z_{\rm cut,s}=\frac{0.1}{m}$ is a compromising choice in order to avoid spoiling the perturbative convergence in the fixed-order calculations by a large logarithm $\log{z_{\rm cut,s}}$. We have compared the recent CMS measurement~\cite{Sirunyan:2017qdw} to our nLO calculations (with and without STOP cut tuning on ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$) for $\psi(2S)$ production at 13 TeV LHC in Fig.~\ref{Fig:nLOvsCMS}, where the nonperturbative LDMEs are taken from Eqs.(2.17) and (2.18) in Ref.~\cite{Shao:2014yta}. A factor $10^{-1}$ has been multiplied to the nLO results with tuned ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ in order to improve the visibility between the two theoretical bands. Without surprising, the CMS data agree very well with our nLO calculations, because nLO does a similarly good job as NLO. \begin{figure}[ht!] \vspace{-1cm} \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_nLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_1S08_THerror_nLO-crop.pdf} \caption{Comparisons of spin-summed differential cross sections for the Fock states ${\bigl.^3\hspace{-1mm}S^{[1]}_1},{\bigl.^1\hspace{-1mm}S^{[8]}_0}$ between our nLO calculations and the complete NLO calculations.\label{Fig:nLOvsNLO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma00_nLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma11_nLO-crop.pdf} \caption{Comparisons of spin-dependent differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between our nLO calculations and the complete NLO calculations.\label{Fig:nLOvsNLOSpin}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.99\textwidth,draft=false]{./figures/dpt_psi_LHC13_1S08_THerror_zsoft001_nLO-crop.pdf} \caption{Comparisons of the differential cross sections for ${\bigl.^1\hspace{-1mm}S^{[8]}_0}$ between our nLO calculations with $z_{\rm cut,s}=\frac{10^{-2}}{m}$ and the complete NLO calculations.\label{Fig:nLO1S08vsNLO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.99\textwidth,draft=false]{./figures/dpt_psi_LHC13_psi2S_cmsarXiv171011002-crop.pdf} \caption{Comparisons of the nLO $\psi(2S)$ differential cross sections $\frac{d^2\sigma}{dP_Tdy}$ in $|y|<0.6$ with the CMS measurement~\cite{Sirunyan:2017qdw}.\label{Fig:nLOvsCMS}} \end{figure} \subsection{Going beyond NLO} It is usually believed that the color-octet states for $J/\psi$ hadroproduction will not receive giant K factors beyond NLO as the LP topologies in $P_T$ appear at NLO. On the other hand, the color-singlet Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$, which is LO in $v^2$ expansion, contains the LP single-gluon fragmentation contributions starting from NNLO in $\alpha_s$ (i.e. $\mathcal{O}(\alpha_s^5)$). A giant K factor for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ from NLO to NNLO might be possible in $J/\psi$ production, though the NLO calculation shows that the ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ contribution to $J/\psi$ hadroproduction seems to be negligible compared to the color-octet contributions. If it is true, the extractions of color-octet NRQCD LDMEs solely based on NLO calculations will be questionable. This is one of the reasons why the importance of color-octet contributions in $J/\psi$ hadroproduction is still under debate. Although the accomplishment of NNLO calculations for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ is still beyond state-of-the art, it was indeed suggested in Ref.~\cite{Artoisenet:2008fc} that the partial calculation shows a giant K factor $\frac{d\sigma^{\rm NNLO^\star}}{d\sigma^{\rm NLO}}$. Later on, it was pointed out in Ref.~\cite{Ma:2010jj} that the giant K factor observed in Ref.~\cite{Artoisenet:2008fc} is in fact due to the logarithmic enhancement induced by the infrared cutoff. Such a logarithm is expected to be absent in a full NNLO calculation because of the infrared safety. We have the opportunity to clarify the situation with our infrared-safe STOP cut method. With the same setup done in section \ref{sec:NLOStop}, we have performed the calculations for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ plus three light-flavored jets production at $\mathcal{O}(\alpha_s^5)$ at the $13$ TeV. The spin-summed $P_T$ differential distributions are shown in Fig.~\ref{Fig:nnLOvsNLO}, where we have used ``nnLO" and ``nNLO" for the $\mathcal{O}(\alpha_s^n),n\leq 4$ parts being nLO and NLO cross sections respectively. In other words, we have used $d\sigma^{\rm nnLO}\equiv d\sigma^{\rm nLO}+d\sigma^{\mathcal{R}^2_{\rm STOP}}$ and $d\sigma^{\rm nNLO}\equiv d\sigma^{\rm NLO}+d\sigma^{\mathcal{R}^2_{\rm STOP}}$. In the nNLO results, no theoretical uncertainties are taking into account from the NLO piece $d\sigma^{\rm NLO}$. In contrast to the finding made in Ref.~\cite{Artoisenet:2008fc}, we do not observe any giant K factor up to $P_T\simeq 100$ GeV. In fact, the $P_T$ spectra of nnLO and nNLO are not harder than NLO ones. Such an observation can be explained if the coefficient of the LP $P_T$ part arising from the single-gluon fragmentation is much smaller than the coefficient of the NLP $P_T$ part and/or if the average momentum fraction of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ taking from the original gluon is significantly smaller than $1$. The calculation based on the gluon fragmentation function shows a similar behaviour, and the normalization of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ is significantly smaller than the color-octet contributions~\cite{Brambilla:2004wf}. In our calculation, the K factor $\frac{d\sigma^{\rm nnLO}}{d\sigma^{\rm NLO}}$ is ranging from 1 to 3 depending on the infrared cutoff choices. A similar conclusion can be drawn for the spin-dependent differential distributions from Fig.~\ref{Fig:nnLOvsNLOSpin}. We believe a complete NLO calculations of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ plus two jets will help to reduce the remaining large infrared cutoff as well as the renormalization/factorization scale dependence. \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_nnLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_nNLO2-crop.pdf} \caption{Comparisons of spin-summed differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between nnLO (left), nNLO (right) calculations and the complete NLO calculations.\label{Fig:nnLOvsNLO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma00_nnLO-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma11_nnLO-crop.pdf} \caption{Comparisons of spin-dependent differential cross sections for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between nnLO calculations and the complete NLO calculations.\label{Fig:nnLOvsNLOSpin}} \end{figure} \subsection{Reassessing the charm fragmentation\label{sec:charmfrag}} So far, we have only considered the light-flavoured jet(s) accompanying with the quarkonium, which is usually thought to be dominant because the gluons are more often produced than the heavy quarks at high-energy hadron colliders. However, since the LP $P_T$ contribution from the charm quark fragmentation appears at $\mathcal{O}(\alpha_s^4)$, one should not overlook the associated production processes of a quarkonium plus a heavy quark pair. They were first studied in Ref.~\cite{Artoisenet:2007xi} for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$, in Ref.~\cite{Artoisenet:2009zwa} for ${\bigl.^3\hspace{-1mm}S^{[8]}_1},{\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[8]}_J}$ and in Ref.~\cite{Li:2011yc} for ${\bigl.^3\hspace{-1mm}P^{[1]}_0},{\bigl.^3\hspace{-1mm}P^{[1]}_1},{\bigl.^3\hspace{-1mm}P^{[1]}_2}$. To the best of our knowledge, the existing calculations only focus on the spin-summed differential cross sections, while we will also present the spin-dependent results in this section. In fact, one has to examine the relevance of these contributions if large cancellations between various Fock states happen. In Fig.~\ref{Fig:nLOvsccx3S11} and Fig.~\ref{Fig:nLOvsccx3S11Spin}, we compared the tree-level ${\bigl.^3\hspace{-1mm}S^{[1]}_1}+c\bar{c}$ (tagged as ``$c\bar{c}$") production with the nnLO calculations of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ plus light-flavoured partons. The ${\bigl.^3\hspace{-1mm}S^{[1]}_1}+c\bar{c}$ contribution has a harder $P_T$ spectrum than the nnLO contribution. The former one exceeds the latter one above $P_T\simeq 55$ GeV in the spin-summed case, while such a kind of crossover happens earlier for the spin transverse component $\frac{d\sigma_{11}}{dP_T}$ around $P_T\simeq 20$ GeV. On the other hand, the charm quark associated contributions are orders of magnitude smaller than the light-flavoured jet contributions for ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ productions as clearly shown in Fig.~\ref{Fig:nLOvsccx4CO} and Fig.~\ref{Fig:nLOvsccxSpin4CO} for the spin-summed and spin-dependent distributions. The similar conclusion can be drawn for the other Fock states ${\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[8]}_J},{\bigl.^3\hspace{-1mm}P^{[1]}_0},{\bigl.^3\hspace{-1mm}P^{[1]}_1},{\bigl.^3\hspace{-1mm}P^{[1]}_2}$ as shown in Figs.~\ref{Fig:nLOvsccx4CO2},~\ref{Fig:nLOvsccxSpin4CO2},~\ref{Fig:nLOvsccxSpin4CO3} in the appendix \ref{app:moreplots}. \begin{figure}[ht!] \vspace{-1cm} \centering \includegraphics[width=.99\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_nnLOccx-crop.pdf} \caption{Comparisons of spin-summed differential cross section $\frac{d\sigma}{dP_T}$ for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between our nnLO calculations and the LO charmonium plus charm quark pair calculations.\label{Fig:nLOvsccx3S11}} \end{figure} \begin{figure}[ht!] \vspace{-1cm} \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma00_nnLOccx-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S11_THerror_dsigma11_nnLOccx-crop.pdf} \caption{Comparisons of spin-dependent differential cross section $\frac{d\sigma}{dP_T}$ for the Fock state ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ between our nnLO calculations and the LO charmonium plus charm quark pair calculations.\label{Fig:nLOvsccx3S11Spin}} \end{figure} \begin{figure}[ht!] \vspace{-1cm} \centering \includegraphics[width=.99\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_nLOccx-crop.pdf} \caption{Comparisons of spin-summed differential cross sections $\frac{d\sigma}{dP_T}$ for the Fock state ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ between our nLO calculations and the LO charmonium plus charm quark pair calculations.\label{Fig:nLOvsccx4CO}} \end{figure} \begin{figure}[ht!] \centering \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_dsigma00_nLOccx-crop.pdf} \includegraphics[width=.45\textwidth,draft=false]{./figures/dpt_psi_LHC13_3S18_THerror_dsigma11_nLOccx-crop.pdf} \caption{Comparisons of spin-dependent differential cross sections $\frac{d\sigma}{dP_T}$ for the Fock state ${\bigl.^3\hspace{-1mm}S^{[8]}_1}$ between our nLO calculations and the LO charmonium plus charm quark pair calculations.\label{Fig:nLOvsccxSpin4CO}} \end{figure} \section{Summary and outlooks\label{sec:summary}} After implementing the remainders of P-wave counterterms in section~\ref{sec:pCTs}, we have introduced a general infrared-safe method to estimate the giant K factors in quarkonium production in high $P_T$ region. As a proof of concept, we have validated our approach with the existing complete NLO QCD calculations of the Fock states $\bigl.^3\hspace{-1mm}S^{[1,8]}_1,{\bigl.^1\hspace{-1mm}S^{[8]}_0},{\bigl.^3\hspace{-1mm}P^{[1,8]}_J}$ in both spin-summed and spin-dependent cases. They are relevant for $J/\psi$ and $\chi_{cJ}$ hadroproduction up to $\mathcal{O}(v^7)$. Our approach only requires the tree-level amplitudes provided by {\sc\small HELAC-Onia}. To the best of our knowledge, it is the first time to be able to reproduce the complete NLO spin-dependent results with tree-level amplitudes only. These spin-dependent results can be used to predict the polarization observables. We are also firstly able to obtain the spin-summed NLO results for ${\bigl.^3\hspace{-1mm}P^{[1]}_J}$ without performing complete NLO calculations. With our new approach, we have estimated the partial NNLO contributions at $\mathcal{O}(\alpha_s^5)$ for ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ production. It is believed to be at LP in $P_T$ scaling starting at this order, and is the last missing piece for the heavy quarkonium $P_T$ spectrum up to $\mathcal{O}(v^7)$. In contrast to the NNLO$^\star$ calculations based on the simple invariant-mass cuts~\cite{Artoisenet:2008fc}, we do not observe the similar orders of magnitude enhancement compared to the NLO calculations, while an enhancement factor of 1 to 3 is still possible up to $P_T\simeq 100$ GeV depending on the infrared cutoff choices. We believe the complete NLO calculations of ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ plus 2 jets will reduce this uncertainty. Finally, we have also calculated the charmonium plus a charm quark pair production, where the spin-dependent differential cross sections presented here are new. Their contributions to the inclusive $P_T$ distributions of charmonium are only relevant in the ${\bigl.^3\hspace{-1mm}S^{[1]}_1}$ channel. Our approach stabilizes the QCD corrections in the heavy quarkonium production rate calculations at high $P_T$. It is quite appealing not only because it provides a fast way to perform the phenomenology studies of inclusive quarkonium production but also it can be used to improve the predictions in the associated quarkonium production processes. Together with the controlled perturbative SDCs, it is feasible to study various nonperturbative effects in the heavy quarkonium hadroproduction in an acceptable amount of computation time. Last but not least, with a similar method, we believe that we are able to promote the accuracy of both LP and NLP pieces to NLO level simultaneously with the full one-loop calculations. \begin{acknowledgments} I thank Jean-Philippe Lansberg for useful discussions. This work is supported by the ILP Labex (ANR-11-IDEX-0004-02, ANR-10-LABX-63). The computations in this paper were performed with the help of the computing facilities at IPN Orsay. \end{acknowledgments}

77eeec0df5368d33a40759988e811980c8e5db4b

\section{Introduction} Bicategories in the sense of B\'enabou \cite{Ben1967} include 2-categories and are extensively used in the literature. For morphisms between bicategories we can take the general ones defined in \cite{Ben1967} (also called lax functors \cite{Street1972b}), or choose strict, pseudo or oplax functors (in the terminology of \cite{Kelly1974}) depending on whether the composition and identity constraints are equalities, isomorphisms or have their directions reversed. Here, we will also use a more exotic kind of morphism, 2-sided enriched category \cite{Kelly2002} which is a common generalisation of hom-enriched category and lax functor. In Section \ref{sec:2-sided} we review the necessary definitions and set up notation used in the rest of the paper. Hopf monoidal comonads \cite{Chikhladze2010}, or dually Hopf opmonoidal monads \cite{Bruguieres2011}, generalise tensoring with a Hopf monoid. The former are comonads in the 2-category of monoidal categories, monoidal functors and monoidal natural transformations, in the sense of \cite{Street1972a}, with a Hopf condition: the fusion maps are invertible. This all generalises to comonads in Caten, which is the subject of Section \ref{sec:comonads}. Kan extensions and liftings envelope various categorical notions including adjunctions (hence duals) and (weighted) (co)limits. Yet, not much has been said on when these notions are preserved or reflected. In \cite{Weber2016} Weber described conditions under which the forgetful functor from the bicategory of pseudo-algebras for a 2-monad on a bicategory $\mathscr{B}$ reflects left Kan extensions. The main result of our paper, given in Theorem \ref{thm:createKan}, is that a forgetful functor from the bicategory of coalgebras for a Hopf comonad in Caten creates left extensions, generalising the result of \cite{Chikhladze2010} that the forgetful functor from the category of coalgebras for a Hopf monoidal comonad is strong closed. Furthermore we give conditions for the change of base of enrichment to be of the same type. We then apply the theorem to monoidal comonads whose coalgebras are graded, or differential graded, abelian groups \cite{Nikolic2018}. \section{Enrichment on two sides}\label{sec:2-sided} We now quickly review some material from \cite{Kelly2002}. To shorten notation a little, we sometimes denote the hom category $\mathscr{V}(V,V')$ of a bicategory $\mathscr{V}$ by $\mathscr{V}_V^{V'}$. Horizontal composition in $\mathscr{V}$ will be denoted by tensor product $\otimes$. There is a simple kind of tricategory denoted by Caten. Objects of Caten are bicategories for which we use symbols such as $\mathscr{V}$, $\mathscr{W}$, and so on. Arrows $\mathscr{A}:\mathscr{W}\rightarrow \mathscr{V}$ are called 2-sided enriched categories. Such an $\mathscr{A}$ consists of: \begin{itemize} \item a set $\mathrm{Ob}\mathscr{A}$ of objects $A$, $A'$, and so on, together with a span \begin{equation} \begin{tikzpicture}[scale=1, every node/.style={scale=1},baseline=(current bounding box.center)] \node (Pb) at (0,1) {$\mathrm{Ob}\mathscr{A}$}; \node (S) at (-1.5,0) {$\mathrm{Ob}\mathscr{W}$}; \node (P) at (1.5,0) {$\mathrm{Ob}\mathscr{V}$}; \path[->,font=\normalsize,>=angle 90] (Pb) edge node[above left] {$(-)_{-}$} (S); \path[->,font=\normalsize,>=angle 90] (Pb) edge node[above right] {$(-)_{+}$} (P); \end{tikzpicture} \end{equation} assigning to each object $A$ an object $A_-$ in $\mathscr{W}$ and $A_+$ in $\mathscr{V}$ \item homs $\mathscr{A}(A,A')$, also denoted $\mathscr{A}_A^{A'}$, defined to be functors \begin{equation} \mathscr{A}_A^{A'}:\mathscr{W}_{A_-}^{A'_-}\rightarrow \mathscr{V}_{A_+}^{A'_+} \end{equation} \item unit and composition natural transformations \begin{equation} \begin{tikzpicture}[xscale=1.2, yscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,1) {$1$}; \node (A2) at (1,1) {$\mathscr{W}_{A_-}^{A_-}$}; \node (B2) at (1,-1) {$\mathscr{V}_{A_+}^{A_+}$}; \node (mu) at (0.2,0) {$\scriptstyle\eta_{A}\Rightarrow$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$1_{A_-}$} (A2); \path[->,font=\scriptsize,>=angle 90, bend right] (A1) edge node[below left] {$1_{A_+}$} (B2); \path[,->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\mathscr{A}_{A}^{A}$} (B2); \end{tikzpicture} \begin{tikzpicture}[xscale=1.6, yscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,1) {$\mathscr{W}_{A'_-}^{A^{\prime\prime}_-}\times \mathscr{W}_{A_-}^{A'_-}$}; \node (A2) at (1,1) {$\mathscr{W}_{A_-}^{A^{\prime\prime}_-}$}; \node (B1) at (-1,-1) {$\mathscr{V}_{A'_+}^{A^{\prime\prime}_+}\times \mathscr{V}_{A_+}^{A'_+}$}; \node (B2) at (1,-1) {$\mathscr{V}_{A_+}^{A^{\prime\prime}_+}$}; \node (mu) at (0.2,0) {$\scriptstyle\mu_{AA''}^{A'}\Rightarrow$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\otimes$} (A2) (B1) edge node[below] {$\otimes$} (B2); \path[transform canvas={xshift=9mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$\mathscr{A}_{A'}^{A''}\times \mathscr{A}_A^{A'}$} (B1); \path[transform canvas={xshift=-3mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\mathscr{A}_{A}^{A''}$} (B2); \end{tikzpicture} \end{equation} satisfying unit and associativity laws. \end{itemize} Composition of 2-sided enriched categories is given by composition of the spans (pullback), composition of the functors defining homs, and pasting of the unit and multiplication natural transformations. A 2-cell $\mathscr{F}:\mathscr{A}\rightarrow \mathscr{B}$ is an (enriched) functor consisting of \begin{itemize} \item a map of spans $\mathrm{Ob}\mathscr{F}=:F$ \begin{equation} F:\mathrm{Ob}\mathscr{A}\rightarrow \mathrm{Ob}\mathscr{B} \end{equation} which means a function between the object sets such that \begin{equation}\label{eq:obComp} (FA)_-=A_-,\;\text{and}\;(FA)_+=A_+ \end{equation} \item natural transformations \begin{equation} \mathscr{F}_A^{A'}:\mathscr{A}_A^{A'}\Rightarrow \mathscr{B}_{FA}^{FA'} \end{equation} which are compatible with the unit and multiplication of $\mathscr{A}$ and $\mathscr{B}$. \end{itemize} A 3-cell $\psi:\mathscr{F}\rightarrow \mathscr{E}$ is an (enriched) natural transformation with components \begin{equation} \psi_A:1_{A_+}\Rightarrow \mathscr{B}_{FA}^{EA}1_{A_-} \end{equation} satisfying an enriched naturality condition (the omitted coherence 2-cells in $\mathscr{V}$ can be found included in \cite{Kelly2002}) \begin{equation}\label{diag:enrNat} \begin{tikzpicture}[xscale=3, yscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,1) {$\mathscr{A}_A^{A'}(w)$}; \node (A2) at (1,1) {$\mathscr{B}_{FA'}^{EA'}(1_{A'_-})\otimes \mathscr{B}_{FA}^{FA'}(w)$}; \node (B1) at (-1,-1) {$ \mathscr{B}_{EA}^{EA'}(w)\otimes\mathscr{B}_{FA}^{EA}(1_{A_-})$}; \node (B2) at (1,-1) {$\mathscr{B}_{FA}^{EA'}(w)\,.$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\psi_{A'}\otimes (\mathscr{F}_A^{A'})_w$} (A2) (B1) edge node[below] {$\mu$} (B2); \path[->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$ (\mathscr{E}_A^{A'})_w\otimes\psi_A$} (B1); \path[->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\mu$} (B2); \end{tikzpicture} \end{equation} All axioms, compositions, whiskerings, and the fact that Caten is a tricategory are explained in detail in \cite{Kelly2002}. \begin{example} When $\mathscr{W}=1$, $\mathscr{A}$ is precisely a category enriched in the bicategory $\mathscr{V}$. \end{example} \begin{example} When $\mathrm{Ob}\mathscr{A}=\mathrm{Ob}\mathscr{W}$ and $(-)_-=1$, $\mathscr{A}$ is precisely a lax functor from $\mathscr{W}$ to $\mathscr{V}$, and 2-cells are `icons' as so named in \cite{Lack2010a}. \end{example} \subsection{Modules} Instead of (enriched) functors we could have chosen enriched modules $M:\mathscr{A}\xmrightarrow{}\mathscr{B}$ as 2-cells. They consist of \begin{itemize} \item functors \begin{equation} M_B^A:\mathscr{W}_{B_-}^{A_-}\rightarrow \mathscr{V}_{B_+}^{A_+} \end{equation} \item action natural transformations \begin{equation}\label{diag:actions} \begin{aligned} \begin{tikzpicture}[xscale=1.5, yscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,1) {$\mathscr{W}_{A_-}^{A'_-}\times \mathscr{W}_{B_-}^{A_-}$}; \node (A2) at (1,1) {$\mathscr{W}_{B_-}^{A'_-}$}; \node (B1) at (-1,-1) {$\mathscr{V}_{A_+}^{A'_+}\times \mathscr{V}_{B_+}^{A_+}$}; \node (B2) at (1,-1) {$\mathscr{V}_{B_+}^{A'_+}$}; \node (mu) at (0.2,0) {$\scriptstyle\lambda_{BA'}^{A}\Rightarrow$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\otimes$} (A2) (B1) edge node[below] {$\otimes$} (B2); \path[transform canvas={xshift=9mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$\mathscr{A}_{A}^{A'}\times M_B^{A}$} (B1); \path[transform canvas={xshift=-3mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$M_{B}^{A'}$} (B2); \end{tikzpicture} \begin{tikzpicture}[xscale=1.5, yscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,1) {$\mathscr{W}_{B'_-}^{A_-}\times \mathscr{W}_{B_-}^{B'_-}$}; \node (A2) at (1,1) {$\mathscr{W}_{B_-}^{A_-}$}; \node (B1) at (-1,-1) {$\mathscr{V}_{B'_+}^{A_+}\times \mathscr{V}_{B_+}^{B'_+}$}; \node (B2) at (1,-1) {$\mathscr{V}_{B_+}^{A_+}$}; \node (mu) at (0.2,0) {$\scriptstyle\rho_{BA}^{B'}\Rightarrow$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\otimes$} (A2) (B1) edge node[below] {$\otimes$} (B2); \path[transform canvas={xshift=9mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$M_{B'}^{A}\times \mathscr{B}_B^{B'}$} (B1); \path[transform canvas={xshift=-3mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$M_{B}^{A}$} (B2); \end{tikzpicture} \end{aligned} \end{equation} compatible with each other, and with the units and compositions in $\mathscr{A}$ and $\mathscr{B}$. \end{itemize} A module morphism $\sigma:M\Rightarrow N$ consists of natural transformations \begin{equation} \sigma_B^A:M_B^A\Rightarrow N_B^A \end{equation} compatible with the actions (\ref{diag:actions}). Module morphisms compose, and we get a category of modules between $\mathscr{A}$ and $\mathscr{B}$, which we call $\mathrm{Moden}(\mathscr{W},\mathscr{V})(\mathscr{A},\mathscr{B})$. When $\mathscr{V}$ is locally cocomplete, $\mathrm{Moden}(\mathscr{W},\mathscr{V})$ becomes a bicategory equivalent to the bicategory of enriched modules $\mathrm{Conv}(\mathscr{W},\mathscr{V})\text{-}\mathrm{Mod}$, where Conv denotes the internal hom in Caten for the usual product of bicategories \cite{Kelly2002}. Each functor $\mathscr{F}:\mathscr{A}\rightarrow\mathscr{B}$ defines a module $\mathscr{F}_{*}:\mathscr{A}\xmrightarrow{}\mathscr{B}$ by taking \begin{align}\label{eq:indMod} (\mathscr{F}_*)_B^A=&\mathscr{B}_B^{FA} :\mathscr{W}_{B_-}^{A_-}\rightarrow \mathscr{V}_{B_+}^{A_+}\\ \lambda_{BA'}^{A}= &\otimes(\mathscr{A}_{A}^{A'}\times \mathscr{B}_B^{FA}) \xRightarrow{1(\mathscr{F}_A^{A'}\times 1)} \otimes(\mathscr{B}_{FA}^{FA'}\times \mathscr{B}_B^{FA}) \xRightarrow{\mu_{B,FA'}^{FA}} \mathscr{B}_{B}^{FA'}\otimes \label{eq:indModLam}\\ \rho_{BA}^{B'}= &\otimes(\mathscr{B}_{B'}^{FA}\times \mathscr{B}_{B}^{B'}) \xRightarrow{\mu_{B,FA}^{B'}} \mathscr{B}_{B}^{FA}\otimes \,\label{eq:indModRho} \end{align} which is properly typed because of (\ref{eq:obComp}). Compatibility of $\rho$ with $\mu^{(\mathscr{B})}$ and $\eta^{(\mathscr{B})}$ are just the unit and associativity axioms for $\mu^{(\mathscr{B})}$ and $\eta^{(\mathscr{B})}$. Compatibility of $\lambda$ with $\mu^{(\mathscr{A})}$ and $\eta^{(\mathscr{A})}$ follows by applying compatibility of the functor $\mathscr{F}$ with $\mu^{(\mathscr{A})}$ and $\eta^{(\mathscr{A})}$, followed by unit associativity laws for $\mu^{(\mathscr{A})}$ and $\eta^{(\mathscr{A})}$. Similarly, each natural transformation $\psi:\mathscr{F}\rightarrow \mathscr{E}$ has an induced module morphism, \begin{align} \psi_*:\mathscr{F}_*&\rightarrow \mathscr{E}_*\\ (\psi_*)_B^A:=\mathscr{B}_B^{FA}&\Rightarrow \mathscr{B}_B^{EA} \\ ((\psi_*)_B^A)_w=\mathscr{B}_B^{FA}(w)&\xRightarrow{\psi_A\otimes 1} \mathscr{B}_{FA}^{EA}(1_{A_-})\otimes\mathscr{B}_B^{FA}(w)\label{eq:natToMor1} \\ &\xRightarrow{\mu_{B,EA}^{FA} }\mathscr{B}_B^{EA}(w) \nonumber\,. \end{align} To see that $\psi_*$ is compatible with $\lambda$, tensor diagram (\ref{diag:enrNat}) by $\mathscr{B}_B^{FA}(w')$, whisker the resulting square with $\mu^{(\mathscr{B})}$ on the right, and add obvious commutative squares to yield the compatibility requirement. Compatibility with $\rho$ follows from associativity of $\mu^{(\mathscr{B})}$. Also, every module morphism between modules induced by functors gives rise to a natural transformation. Given \begin{equation} \sigma_B^A:\mathscr{B}_B^{FA}\Rightarrow \mathscr{B}_B^{EA} \end{equation} we can form \begin{equation}\label{eq:morToNat} \sigma_A:1_{A_+}\xRightarrow{\eta_{FA}}\mathscr{B}_{FA}^{FA}(1_{A_-})\xRightarrow{(\sigma_{FA}^{A})_{1_{A_-}}} \mathscr{B}_{FA}^{EA}(1_{A_-}) \end{equation} and the natural transformation axiom (\ref{diag:enrNat}) is shown by commutativity of \begin{equation} \begin{tikzpicture}[xscale=2,yscale=1.1, every node/.style={scale=0.8},baseline=(current bounding box.center)] \node (11) at (-0.5,0) {$\mathscr{A}_A^{A'}(w)$}; \node (12) at (0,-1.5) {$\mathscr{A}_A^{A'}(w)\otimes \mathscr{B}_{FA}^{FA}(1_{A_-})$}; \node (13) at (0,-4) {$\mathscr{A}_A^{A'}(w)\otimes \mathscr{B}_{FA}^{EA}(1_{A_-})$}; \node (21) at (1.5,0) {$\mathscr{B}_{FA'}^{FA'}(1_{A'_-})\otimes\mathscr{A}_A^{A'}(w)$}; \node (23) at (2.5,-4) {$\mathscr{B}_{EA}^{EA'}(w)\otimes\mathscr{B}_{FA}^{EA'}(1_{A_-}) $}; \node (31) at (4,0) {$\mathscr{B}_{FA'}^{EA'}(1_{A'_-})\otimes\mathscr{A}_A^{A'}(w)$}; \node (32) at (4,-2.5) {$\mathscr{B}_{FA'}^{EA'}(1_{A'_-})\otimes\mathscr{B}_{FA}^{FA'}(w)$}; \node (33) at (4.5,-4) {$\mathscr{B}_{FA}^{EA'}(w)$}; \node (i11) at (1,-1) {$\mathscr{B}_{FA}^{FA'}(w)$}; \node (i12) at (1.5,-3) {$\mathscr{B}_{FA}^{FA'}(w)\otimes\mathscr{B}_{FA}^{FA}(1_{A_-})$}; \node (i21) at (2.5,-1) {$\mathscr{B}_{FA'}^{FA'}(1_{A'_-})\otimes\mathscr{B}_{FA}^{FA'}(w)$}; \node (i22) at (3,-3) {$\mathscr{B}_{FA}^{FA'}(w)$}; \path[->,font=\scriptsize,>=angle 90] (11) edge node[left] {$1\otimes\eta_{FA}$} (12) (12) edge node[left] {$1\otimes(\sigma_{FA}^A)_{1_{A_-}}$} (13) (13) edge node[below] {$(\mathscr{E}_A^{A'})_w\otimes 1$} (23) (23) edge node[below] {$\mu$} (33) ; \path[->,font=\scriptsize,>=angle 90] (11) edge node[above] {$\eta_{FA'}\otimes 1$} (21) (21) edge node[above] {$(\sigma_{FA'}^{A'})_{1_{A'_-}}\otimes1$} (31) (31) edge node[right] {$1\otimes (\mathscr{F}_A^{A'})_w$} (32) (32) edge node[right] {$\mu$} (33) ; \path[->,font=\scriptsize,>=angle 90] (11) edge node[below] {$(\mathscr{F}_A^{A'})_w$} (i11) (12) edge node[below left] {$(\mathscr{F}_A^{A'})_w\otimes 1$} (i12) (21) edge node[above right] {$1\otimes (\mathscr{F}_A^{A'})_w$} (i21) (i21) edge node[fill=white,rounded corners=1pt,inner sep=0.3pt] {$(\sigma_{FA'}^{A'})_{1_{A'_-}}\otimes 1$} (32) (i22) edge node[above right] {$(\sigma_{FA}^{A'})_{w}$} (33) ; \path[->,font=\scriptsize,>=angle 90] (i11) edge node[above] {$\eta_{FA'}\otimes 1$} (i21) (i21) edge node[left] {$\mu$} (i22) (i11) edge node[left] {$1\otimes \eta_{FA}$} (i12) (i12) edge node[below] {$\mu$} (i22) (i11) edge node[fill=white,rounded corners=1pt,inner sep=0.3pt] {$1$} (i22) ; \end{tikzpicture} \end{equation} where the hexagon and the bottom right square are compatibility conditions between module morphism $\sigma$ and actions (\ref{eq:indModLam}) and (\ref{eq:indModRho}) respectively. \begin{prop}\label{prop:2-sidY} The functor \begin{equation} (-)_*:\mathrm{Caten}(\mathscr{W},\mathscr{V})(\mathscr{A},\mathscr{B}) \rightarrow \mathrm{Moden}(\mathscr{W},\mathscr{V})(\mathscr{A},\mathscr{B}) \end{equation} is full and faithful. \end{prop} \begin{proof} The processes of turning a natural transformation into a module morphism given by (\ref{eq:natToMor1}) and the one turning a module morphism of convergent modules into a natural transformation, given by (\ref{eq:morToNat}), are inverse to each other, as shown by the commuting diagrams (\ref{diag:yComm}). \end{proof} \begin{equation}\label{diag:yComm} \begin{aligned} \begin{tikzpicture}[xscale=4, yscale=1.5, every node/.style={scale=0.8},baseline=(current bounding box.center)] \node (A1) at (0,0) {$1_{A_+}$}; \node (A2) at (1,0) {$\mathscr{B}_{FA}^{FA}(1_{A_-})$}; \node (B1) at (0,-1) {$\mathscr{B}_{FA}^{EA}(1_{A_-})$}; \node (B2) at (1,-1) {$\mathscr{B}_{FA}^{EA}(1_{A_-})\otimes\mathscr{B}_{FA}^{FA}(1_{A_-})$}; \node (C) at (1,-2) {$\mathscr{B}_{FA}^{EA}(1_{A_-})$}; \path[->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\eta_{FA}$} (A2) (B1) edge node[above] {$1\otimes \eta_{FA}$} (B2) (B1) edge node[below] {$1$} (C); \path[->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$\psi_A$} (B1); \path[->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\psi_A\otimes 1$} (B2) (B2) edge node[right] {$\mu$} (C); \end{tikzpicture} \begin{tikzpicture}[xscale=3.2, yscale=1.5, every node/.style={scale=0.8},baseline=(current bounding box.center)] \node (C) at (0,0) {$\mathscr{B}_{B}^{FA}(w)$}; \node (B2) at (0,-1) {$\mathscr{B}_{FA}^{FA}(1_{A_-})\otimes\mathscr{B}_{B}^{FA}(w)$}; \node (B1) at (1,-1) {$\mathscr{B}_{B}^{FA}(w)$}; \node (A1) at (1,-2) {$\mathscr{B}_{B}^{EA}(w)$}; \node (A2) at (0,-2) {$\mathscr{B}_{FA}^{EA}(1_{A_-})\otimes\mathscr{B}_{B}^{FA}(w)$}; \path[<-,font=\scriptsize,>=angle 90] (A1) edge node[below] {$\mu$} (A2) (B1) edge node[below] {$\mu$} (B2) (B1) edge node[above] {$1$} (C); \path[<-,font=\scriptsize,>=angle 90] (A1) edge node[right] {$(\sigma_B^A)_w$} (B1); \path[<-,font=\scriptsize,>=angle 90] (A2) edge node[left] {$(\sigma_{FA}^A)_{1_{A_-}}\otimes 1$} (B2) (B2) edge node[left] {$\eta_{FA}\otimes 1$} (C); \end{tikzpicture} \end{aligned} \end{equation} \section{Comonads in $\text{Caten}$}\label{sec:comonads} Let $\mathscr{G}:\mathscr{V}\rightarrow\mathscr{V}$ be a comonad in $\text{Caten}$, that is, a 2-sided enriched category with enriched functors \begin{align} 1_\mathscr{V}\xleftarrow{\epsilon}\mathscr{G}\xrightarrow{\delta}\mathscr{G}^2 \end{align} satisfying the three comonoid axioms. The existence of the span morphism $\text{Ob}\mathscr{G}\xrightarrow{\text{Ob}\epsilon} \text{Ob}\mathscr{V}$ forces $G_+=G_-=(\text{Ob}\epsilon) (G)=:G_0$, for all $G$. The two counit axioms give $(\text{Ob}\delta)(G)=(G,G)$. With these simplifications, the remaining data for $\mathscr{G}$ is given by endofunctors \begin{align} \mathscr{G}_G^{G'}:\mathscr{V}_{G_0}^{G'_0}\rightarrow \mathscr{V}_{G_0}^{G'_0} \end{align} and natural transformations with components \begin{align} (\mu_{GG''}^{G'})_{v',v}&:\mathscr{G}_{G'}^{G''}(v')\otimes \mathscr{G}_G^{G'}(v)\Rightarrow \mathscr{G}_G^{G''}(v'\otimes v)\\ \eta_G&:1_{G_0}\Rightarrow \mathscr{G}_G^G(1_{G_0})\\ (\delta_G^{G'})_v&:\mathscr{G}_G^{G'}(v)\Rightarrow (\mathscr{G}_G^{G'})^2(v)\\ (\epsilon_G^{G'})_v&:\mathscr{G}_G^{G'}(v)\Rightarrow v \end{align} satisfying enriched functor compatibility axioms, which together with the comonad axioms, correspond exactly to the monoidal comonad axioms dual to the opmonoidal monad axioms appearing in \cite{Bruguieres2011}. \subsection{The bicategory of $\mathscr{G}\text{-}$coalgebras} Each hom $\mathscr{G}_G^{G'}$ becomes a comonad in the usual sense (in $\text{Cat}$). Let $\mathscr{V}^\mathscr{G}$ denote a (soon to become) bicategory with the same objects as $\mathscr{G}$ and with homs the categories of EM-coalgebras \begin{equation} \mathscr{V}^\mathscr{G}(G,G'):=\mathscr{V}(G_0,G'_0)^{\mathscr{G}(G,G')}\,. \end{equation} The identity coalgebra is $(1_{G_0},\eta_G)$ and composition is given on coalgebras by \begin{align} \mathscr{V}^\mathscr{G}(G',G'')\times \mathscr{V}^\mathscr{G}(G,G')&\rightarrow \mathscr{V}^\mathscr{G}(G,G'')\label{eq:tensorCoalg}\\ (v',\gamma_{v'}),(v,\gamma_{v})&\mapsto (v'\otimes v, (\mu_{GG''}^{G'})_{v',v}\circ (\gamma_{v'}\otimes \gamma_v))\,.\nonumber \end{align} The assigned pair does form a coalgebra: compatibility with $\delta$ is witnessed by commutativity of\footnote{When indices are omitted they can be deduced from the context. For example, $\mathscr{G}_G^{G'}(v)$ is the full notation.} \begin{equation} \begin{tikzpicture}[xscale=4,yscale=1.3, every node/.style={scale=1},baseline=(current bounding box.center)] \node (M1) at (0,-1) {$v'\otimes v$}; \node (TM1) at (0,-2) {$\mathscr{G}v' \otimes \mathscr{G}v$}; \node (TE1) at (0,-3) {$\mathscr{G}(v' \otimes v)$}; \node (23) at (1,-3) {$\mathscr{G}(\mathscr{G}v' \otimes \mathscr{G}v)$}; \node (TTE) at (2,-3) {$\mathscr{G}^2(v' \otimes v)$}; \node (TM3) at (2,-1) {$\mathscr{G}(v' \otimes v)$}; \node (N) at (1,-1) {$\mathscr{G}v' \otimes \mathscr{G}v$}; \node (TN) at (1,-2) {$\mathscr{G}^2 v' \otimes \mathscr{G}^2 v$}; \path[->,font=\scriptsize,>=angle 90] (M1) edge node[left] {$\gamma\otimes\gamma$} (TM1) (TM1) edge node[left] {$\mu$} (TE1) (TE1) edge node[below] {$\mathscr{G}(\gamma\otimes \gamma)$} (23) (23) edge node[below] {$\mathscr{G}\mu$} (TTE); \path[->,font=\scriptsize,>=angle 90] (TM3) edge node[right] {$\delta$} (TTE); \path[->,font=\scriptsize,>=angle 90] (N) edge node[above] {$\mu$} (TM3) (M1) edge node[above] {$\gamma\otimes\gamma$} (N) (TM1) edge node[below] {$\mathscr{G}\gamma\otimes\mathscr{G}\gamma$} (TN) (N) edge node[left] {$\delta\otimes \delta$} (TN) (TN) edge node[left] {$\mu$} (23) (TN) edge node[above] {$\mu^{(\mathscr{G}^2)}$} (TTE); \end{tikzpicture} \end{equation} where the upper left square is a componentwise compatibility of local coalgebras $\gamma$ with comultiplication, the bottom left square is naturality of $\mu$, the triangle is the definition of composition for the composite category, and the remaining square is compatibility of the enriched functor $\delta$ with compositions in its source and target, which one can also identify as a typical bimonoid (bialgebra) axiom. Similarly, $\epsilon$ being an enriched functor implies compatibility of (\ref{eq:tensorCoalg}) with $\epsilon$. The assignment extends to coalgebra morphisms, which follows directly from naturality of $\mu$. The unitors and associators are inherited from $\mathscr{V}$, they are coalgebra morphisms, and satisfy the usual monoidale axioms as they do in $\mathscr{V}$. There is an underlying (strict) functor $\mathscr{U}:\mathscr{V}^\mathscr{G} \rightarrow\mathscr{V}$ sending $G$ to the underlying object $G_0$ in $\mathscr{V}$, and disregarding the colagebra structure on homs. By construction, each $\mathscr{U}_V^{V'}$ has a right adjoint $\mathscr{R}_V^{V'}$, and by Theorem 2.7 of \cite{Kelly2002} the right adjoints are part of a 2-sided enriched category $\mathscr{R}:\mathscr{V}\rightarrow \mathscr{V}^\mathscr{G}$ which has the same objects as $\mathscr{G}$, with span legs given by $G_-=G_0$ and $G_+=G$, with unit and multiplication given by \begin{align} (\mu_{GG''}^{(\mathscr{R})G'})_{v',v}:&\mathscr{R}_{G'}^{G''}(v')\otimes \mathscr{R}_G^{G'}(v)\Rightarrow \mathscr{R}_{G}^{G''}(v'\otimes v)\\ =&(\mathscr{G}v'\otimes \mathscr{G}v,\mu_{v',v}\circ (\delta_{v'}\otimes \delta_v))\xRightarrow{\mu_{v',v}} (\mathscr{G}(v'\otimes v),\delta_{v'\otimes v})\\ \eta^{(\mathscr{R})}_G:&1_{G}\Rightarrow \mathscr{R}_G^G(1_{G_0})\\ =&(1_{G_0},\eta_G)\xRightarrow {\eta_G} (\mathscr{G}1_{G_0},\delta_{1_{G_0}})\,. \end{align} Now we have an adjunction in $\mathrm{Caten}$. \begin{equation}\label{diag:catenAdj} \begin{tikzpicture}[scale=1.5, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-1,0) {$\mathscr{V}^\mathscr{G}$}; \node (A2) at (1,0) {$\mathscr{V}$}; \node (ad) at (0,0) {$\bot$}; \path[,->,font=\scriptsize,>=angle 90,bend left] (A1) edge node[above] {$\mathscr{U}$} (A2); \path[<-,font=\scriptsize,>=angle 90,bend right] (A1)edge node[below] {$\mathscr{R}$} (A2); \end{tikzpicture} \end{equation} The counit and the unit of the adjunction are given by the enriched functors \begin{align} \mathscr{U}\circ\mathscr{R}=\mathscr{G}&\xrightarrow{\epsilon} 1_\mathscr{V}\\ 1_{\mathscr{V}^\mathscr{G}}&\xrightarrow{\gamma} \mathscr{R}\circ\mathscr{U}\label{eq:gammaFunct}\\ (\text{ob}\gamma)(G)&=(G,G)\nonumber\\ (v,\gamma_v:v\rightarrow \mathscr{G}_G^{G'}v)&\xRightarrow{\gamma_v} (\mathscr{G}_G^{G'}v,\delta_v)\,.\nonumber \end{align} Now we present a version of Beck's theorem that we are going to use in the rest of the chapter. \begin{prop}\label{prop:comonadic} Any 2-sided enriched category $\mathscr{L}:\mathscr{W}\rightarrow \mathscr{V}$ such that \begin{itemize} \item $\mathscr{L}$ has a right adjoint $\mathscr{R}$ in $\text{Caten}$ \item $\mathscr{L}$ is locally conservative \item $\mathscr{W}$ has, and $\mathscr{L}$ preserves, local $\mathscr{L}$-split equalizers \end{itemize} gives rise to an equivalence to $\mathscr{W}\simeq\mathscr{V}^{\mathscr{G}}$, where $\mathscr{G}$ is the generated comonad $\mathscr{G}=\mathscr{L}\circ \mathscr{R}$. \end{prop} \begin{proof} As was shown in \cite{Kelly2002}, $\mathscr{L}$ has a right adjoint if and only if it is a pseudo-functor and each functor $\mathscr{L}(L,L')$ has a right adjoint, call it $\mathscr{R}(L,L')$. Then, the right adjoint $\mathscr{R}$ has the same objects as $\mathscr{L}$ (and $\mathscr{W}$, since $\mathscr{L}$ is a pseudo-functor), and homs are precisely $\mathscr{R}(L,L')$. From the usual Beck (co-)monadicity theorem it follows that $\mathscr{W}\simeq \mathscr{V}^{\mathscr{G}}$: they have the same objects and equivalent homs. \end{proof} The category $\text{Caten}(\mathscr{X},\mathscr{V})$ has an induced comonad $\text{Caten}(\mathscr{X},\mathscr{G})$ on it. In particular, when $\mathscr{X}=\mathscr{V}^\mathscr{G}$ there is a natural coalgebra structure on $\mathscr{U}$ given by an enriched functor \begin{equation}\label{eq:gammadef} \mathscr{U}\xrightarrow{\mathscr{U}\circ\gamma} \mathscr{G}\circ \mathscr{U} \end{equation} whose components are exactly $\gamma_v:v\Rightarrow \mathscr{G}v$. \begin{lemma}\label{lem:whisk} Let $\mathscr{X}$ be a bicategory. Whiskering with $\mathscr{U}$ \begin{equation} \mathrm{Moden}(\mathscr{X},\mathscr{V}^\mathscr{G})(\mathscr{A},\mathscr{B}) \xrightarrow{\mathscr{U}\circ-} \mathrm{Moden}(\mathscr{X},\mathscr{V})(\mathscr{U}\circ\mathscr{A},\mathscr{U}\circ\mathscr{B}) \end{equation} is conservative and the source has, and $(\mathscr{U}\circ-)$ preserves, $(\mathscr{U}\circ-)\text{-}$split equalizers. \end{lemma} \begin{proof} Let $M,N:\mathscr{A}\xmrightarrow{}\mathscr{B}$ be modules, and $\sigma:M\Rightarrow N$ a module morphism, with components \begin{equation} (\sigma_B^A)_x:M_{B}^{A}(x)\Rightarrow N_{B}^{A}(x) \end{equation} which are 2-cells in $\mathscr{V}^\mathscr{G}$, natural in $x\in \mathscr{X}_{B_-}^{A_-}$. Let $\psi:\mathscr{U}\circ M\Rightarrow \mathscr{U}\circ N$ be an inverse of $\mathscr{U}\circ \sigma$. This precisely means that the component \begin{equation} (\psi_B^A)_x:\mathscr{U}N_{B}^{A}(x)\Rightarrow \mathscr{U}M_{B}^{A}(x) \end{equation} is an inverse of the component 2-cell $(\sigma_B^A)_x$ in $\mathscr{V}$. Since $\mathscr{U}$ is locally conservative, $(\psi_B^A)_x$ is also a coalgebra morphism. Hence, naturality squares for $\psi_B^A$ consist of the same arrows regardless of whether it is seen as a morphism from $\mathscr{U}N_{B}^{A}$ to $\mathscr{U}M_{B}^{A}$, or from $N_{B}^{A}$ to $M_{B}^{A}$. Compatibility of $\psi$ with actions for $M$ and $N$ follows from the same compatibility conditions for $\sigma$ and the fact that they are inverse of each other. Consider a pair $\sigma,\chi:M\Rightarrow N$ with a split equalizer \begin{equation} \begin{tikzpicture}[xscale=0.8, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (3.5,0.7) {$\mathscr{U}\circ N$}; \node (A2) at (0,0.7) {$\mathscr{U}\circ M$}; \node (A3) at (-2.7,0.7) {$E$}; \path[transform canvas={yshift=1.7mm},<-,font=\scriptsize,>=angle 90] (A1) edge [thick] node[transform canvas={yshift=-1.3mm},above] {$\mathscr{U}\circ \sigma$} (A2); \path[transform canvas={yshift=-1.7mm},<-,font=\scriptsize,>=angle 90] (A1)edge [thick] node[transform canvas={yshift=-1.3mm},above] {$\mathscr{U}\circ \chi$} (A2); \path[transform canvas={yshift=0mm},->,font=\scriptsize,>=angle 90] (A1)edge [bend left] node[transform canvas={yshift=-1mm},above] {$\psi$} (A2); \path[transform canvas={yshift=0mm},<-,font=\scriptsize,>=angle 90] (A2) edge [thick] node[transform canvas={yshift=-1mm},above] {$\xi$} (A3); \path[transform canvas={yshift=0mm},->,font=\scriptsize,>=angle 90] (A2)edge [bend left] node[transform canvas={yshift=-1mm},above] {$\phi$} (A3); \end{tikzpicture} \end{equation} meaning that we have the following componentwise formulas: \begin{align} (\phi_B^A)_x \bullet (\xi_B^A)_x &=1_{E_B^A(x)}\\ (\psi_B^A)_x \bullet (\sigma_B^A)_x &= 1_{\mathscr{U}M_B^A(x)}\\ (\psi_B^A)_x \bullet (\chi_B^A)_x &= (\xi_B^A)_x \bullet (\phi_B^A)_x\,. \end{align} This in particular means that the pair $(\sigma_B^A)_x,(\chi_B^A)_x:M_B^A(x)\Rightarrow N_B^A(x)$ has a $\mathscr{U}_{B_+}^{A_+}\text{-}$split equalizer in $\mathscr{V}_{B_{+0}}^{A_{+0}}$. Since $\mathscr{U}_{B_+}^{A_+}$ is comonadic, $(\xi_B^A)_x$ is an equalizer of $(\sigma_B^A)_x$ and $(\chi_B^A)_x$ in $\mathscr{V}_{B_{+}}^{A_{+}}$, with an algebra structure on its source \begin{align} \gamma_{E_B^A(x)}:=E_B^A(x) &\xRightarrow{(\xi_B^A)_x} \mathscr{U}M_B^A(x)\nonumber\\ &\xRightarrow{\gamma_{M_B^A(x)}} \mathscr{G}_{B_+}^{A_+}\mathscr{U}M_B^A(x)\\ &\xRightarrow{\mathscr{G}_{B_+}^{A_+}(\phi_B^A)_x} \mathscr{G}_{B_+}^{A_+} E_B^A(x)\,.\nonumber \end{align} The action components for the module $E$ are coalgebra morphisms, the proof for $\lambda$ (dually for $\rho$) comes from the diagram (\ref{diag:actEcoalgMorph}) (all indices can be deduced from the top left term). \begin{equation}\label{diag:actEcoalgMorph} \begin{tikzpicture}[xscale=1.8,yscale=1.7, every node/.style={scale=1},baseline=(current bounding box.center)] \node (E) at (0,0) {$\mathscr{A}_A^{A'}(x') \otimes E_B^A(x)$}; \node (M1) at (0,-1) {$\mathscr{G}\mathscr{A} \otimes M$}; \node (TM1) at (0,-2) {$\mathscr{G}\mathscr{A} \otimes \mathscr{G}M$}; \node (TE1) at (0,-3) {$\mathscr{G}\mathscr{A} \otimes\mathscr{G}E$}; \node (23) at (2.25,-3) {$G(\mathscr{A}\otimes E)$}; \node (TTE) at (4.5,-3) {$\mathscr{G}E$}; \node (M2) at (4.5,0) {$E$}; \node (TM3) at (4.5,-1) {$M$}; \node (TE2) at (4.5,-2) {$\mathscr{G}M$}; \node (N) at (2.25,-1) {$\mathscr{A}\otimes M$}; \node (TN) at (2.25,-2) {$\mathscr{G}(\mathscr{A}\otimes M)$}; \path[->,font=\scriptsize,>=angle 90] (E) edge node[left] {$\gamma\otimes\xi$} (M1) (M1) edge node[left] {$1\otimes\gamma$} (TM1) (TM1) edge node[left] {$1\otimes\mathscr{G}\phi$} (TE1) (TE1) edge node[below] {$\mu$} (23) (23) edge node[below] {$\mathscr{G}\lambda$} (TTE); \path[->,font=\scriptsize,>=angle 90] (E) edge node[above] {$\lambda$} (M2) (M2) edge node[right] {$\xi$} (TM3) (TM3) edge node[right] {$\gamma$} (TE2) (TE2) edge node[right] {$\mathscr{G}\phi$} (TTE); \path[->,font=\scriptsize,>=angle 90] (E) edge node[above] {$1\otimes \xi$} (N) (N) edge node[above] {$\lambda$} (TM3) (N) edge node[above] {$\gamma\otimes\gamma$} (TM1) (TM1) edge node[below] {$\mu$} (TN) (TN) edge node[above] {$\mathscr{G}\lambda$} (TE2) (N) edge node[left] {$\gamma$} (TN) (TN) edge node[left] {$\mathscr{G}(1\otimes \phi)$} (23); \end{tikzpicture} \end{equation} Diagrams for compatibility of actions of $E$ with units and multiplications in $\mathscr{A}$ and $\mathscr{B}$ are the same as the ones for $\mathscr{U}\circ \mathscr{A}$ and $\mathscr{U}\circ \mathscr{B}$. This proves that $E:\mathscr{A}\xmrightarrow{} \mathscr{B}$ is a module. Components $(\xi_B^A)_x$ are natural in $x$, and compatible with actions of $E$ as coalgebra morphisms because they are natural and compatible as usual arrows. This proves that $\xi$ is a module morphism between $E$ (with coalgebra structure) and $M$. It remains to show that $\xi$ is an equalizer of $\sigma$ and $\chi$, so assume $L\xrightarrow{\omega}M$ is another $\mathscr{V}^\mathscr{G}$-module morphism satisfying $\sigma\bullet \omega=\chi\bullet\omega$. The components of $\phi\bullet\omega$, obtained by composing components of $\phi$ and $\omega$, are coalgebra maps since $\mathscr{U}$ is locally comonadic, and naturality in $x$ and compatibility with actions follows as for $\xi$. \end{proof} \begin{cor}\label{cor:whiskU} Whiskering with $\mathscr{U}$ \begin{equation} \mathrm{Caten}(\mathscr{X},\mathscr{V}^\mathscr{G})(\mathscr{A},\mathscr{B}) \xrightarrow{\mathscr{U}\circ-} \mathrm{Caten}(\mathscr{X},\mathscr{V})(\mathscr{U}\circ\mathscr{A},\mathscr{U}\circ\mathscr{B}) \end{equation} is conservative and the source has, and $(\mathscr{U}\circ-)$ preserves, $(\mathscr{U}\circ-)\text{-}$split equalizers. \end{cor} \begin{proof} This is a direct consequence of Proposition \ref{prop:2-sidY}, Lemma \ref{lem:whisk}, and commutativity of $(-)_*$ with $\mathscr{U}\circ-$. \end{proof} \begin{prop} The bicategory $\mathscr{V}^{\mathscr{G}}$ is an $\mathrm{EM}\text{-}$object for the comonad $\mathscr{G}$ in $\mathrm{Caten}$. \end{prop} \begin{proof} Mapping out of $\mathscr{X}$, \begin{equation} \mathrm{Caten}(\mathscr{X},-):\mathrm{Caten}\rightarrow 2\text{-}\mathrm{CAT} \end{equation} is a pseudo-functor, therefore preserves adjunctions. In particular, applying it to (\ref{diag:catenAdj}) gives \begin{equation} \begin{tikzpicture}[scale=1, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-2,0) {$\mathrm{Caten}(\mathscr{X},\mathscr{V}^\mathscr{G})$}; \node (A2) at (2,0) {$\mathrm{Caten}(\mathscr{X},\mathscr{V})\,.$}; \node (ad) at (0,0) {$\bot$}; \path[,->,font=\scriptsize,>=angle 90,bend left] (A1) edge node[above] {$\mathscr{U}':=\mathrm{Caten}(\mathscr{X},\mathscr{U})$} (A2); \path[<-,font=\scriptsize,>=angle 90,bend right] (A1)edge node[below] {$\mathscr{R}':=\mathrm{Caten}(\mathscr{X},\mathscr{R})$} (A2); \end{tikzpicture} \end{equation} The composite is isomorphic to $\mathrm{Caten}(\mathscr{X},\mathscr{G})$, and what remains to show is that $\mathscr{U}'$ is comonadic in the sense of Proposition \ref{prop:comonadic}. It has a right adjoint $\mathscr{R}'$, and the rest follows from Corollary \ref{cor:whiskU}. \end{proof} \section{Hopf comonads and extension creation}\label{sec:extCreation} \begin{defn} A comonad $\mathscr{G}$ is left Hopf if, for all $G$, $G'$, $G''$, $v\in \mathscr{V}^\mathscr{G}(G,G')$ and $v'\in \mathscr{V}(G'_0,G''_0)$, the fusion map \begin{equation} \mathrm{v}_{v',v}:\mathscr{G}v'\otimes \mathscr{U}v\xRightarrow{1\otimes \gamma_v} \mathscr{G}v'\otimes \mathscr{G}\mathscr{U}v \xRightarrow{(\mu_{G,G''}^{G'})_{v',\mathscr{U}v}} \mathscr{G}(v'\otimes \mathscr{U}v) \end{equation} is invertible. This is equivalent to (left) Hopf maps \begin{equation} \mathrm{h}_{v',v}:\mathscr{R}v'\otimes v \xRightarrow{1\otimes \gamma_v} \mathscr{R} v'\otimes \mathscr{R}\mathscr{U}v \xRightarrow{(\mu_{G,G''}^{G'})_{v',v}} \mathscr{R}(v'\otimes \mathscr{U}v) \end{equation} being invertible. \end{defn} \begin{prop} The inverse fusion maps are $\mathscr{G}$-compatible in the first variable, meaning \begin{equation} \begin{tikzcd} \mathscr{G}(v'\otimes v) \arrow[rightarrow]{dr}{\mathrm{v}_{v',v}^{-1}}& \\ v'\otimes v \arrow[leftarrow]{u}{\epsilon_{v'\otimes v}} & \mathscr{G} v' \otimes v \arrow[rightarrow]{l}{\epsilon_{v'}\otimes 1} \end{tikzcd} \begin{tikzcd} \mathscr{G}(v'\otimes v) \arrow[rightarrow]{r}{\mathrm{v}_{v',v}^{-1}}& \mathscr{G} v' \otimes v \arrow[rightarrow]{r}{\delta_{v'}\otimes 1} & \mathscr{G}^2 v' \otimes v \arrow[leftarrow]{d}{\mathrm{v}_{\mathscr{G}v',v}^{-1}} \\ \mathscr{G}^2(v'\otimes v) \arrow[leftarrow]{u}{\delta_{v'\otimes v}} \arrow[rightarrow]{rr}{\mathscr{G}\mathrm{v}_{v',v}^{-1}} & & \mathscr{G}(\mathscr{G} v' \otimes v ) \end{tikzcd} \end{equation} as well as compatible with any coalgebra structure existing on $v'$, in the sense \begin{equation} \begin{tikzcd} \mathscr{G}(v'\otimes v) \arrow[rightarrow]{dr}{\mathrm{v}_{v',v}^{-1}} & \\ v'\otimes v \arrow[rightarrow]{u}{\gamma_{v'\otimes v}} & \mathscr{G} v' \otimes v \;. \arrow[leftarrow]{l}{\gamma_{v'}\otimes 1} \end{tikzcd} \end{equation} \end{prop} \begin{proof} The statements follow directly from the commuting diagrams \eqref{me1}, \eqref{me2}, \eqref{me3}. \end{proof} \begin{equation}\label{me1} \begin{tikzcd} \mathscr{G}(v'\otimes v) \arrow[leftarrow]{r}{\mu_{v',v}}& \mathscr{G}v'\otimes \mathscr{G}v \arrow[leftarrow]{d}{1\otimes\gamma_v} \arrow[rightarrow]{dl}{\epsilon_{v'}\otimes\epsilon_{v}} \\ v'\otimes v \arrow[leftarrow]{u}{\epsilon_{v'\otimes v}} & \mathscr{G} v' \otimes v \arrow[rightarrow]{l}{\epsilon_{v'}\otimes 1} \end{tikzcd} \end{equation} \begin{equation}\label{me2} \begin{tikzcd} \, &\mathscr{G} v' \otimes v \arrow[rightarrow]{dr}{\delta\otimes 1} \arrow[rightarrow]{d}{1\otimes \gamma} \arrow[rightarrow,swap]{ddr}{\delta\otimes\gamma} & \\ \mathscr{G}(v'\otimes v) \arrow[leftarrow]{r}{\mu}& \mathscr{G} v' \otimes \mathscr{G}v & \mathscr{G}^2 v' \otimes v \arrow[rightarrow]{d}{1\otimes\gamma} \\ & \mathscr{G}^2 v' \otimes \mathscr{G}^2v \arrow[leftarrow]{u}{\delta\otimes\delta} & \mathscr{G}^2 v' \otimes \mathscr{G}v \arrow[rightarrow]{d}{\mu} \arrow[rightarrow]{l}{1\otimes \mathscr{G}\gamma} \\ \mathscr{G}^2(v'\otimes v) \arrow[leftarrow]{uu}{\delta} & \mathscr{G}(\mathscr{G} v' \otimes \mathscr{G}v ) \arrow[rightarrow]{l}{\mathscr{G}\mu} \arrow[leftarrow]{u}{\mu} & \mathscr{G}(\mathscr{G} v' \otimes v ) \arrow[rightarrow]{l}{\mathscr{G}(1\otimes\gamma)} \end{tikzcd} \end{equation} \begin{equation}\label{me3} \begin{tikzcd} \mathscr{G}(v'\otimes v) \arrow[leftarrow]{r}{\mu_{v',v}} & \mathscr{G}v'\otimes \mathscr{G}v \arrow[leftarrow]{d}{1\otimes\gamma_v} \arrow[leftarrow]{dl}{\gamma_{v'}\otimes\gamma_{v}} \\ v'\otimes v \arrow[rightarrow]{u}{\gamma_{v'\otimes v}} & \mathscr{G} v' \otimes v \arrow[leftarrow]{l}{\gamma_{v'}\otimes 1} \end{tikzcd} \end{equation} \begin{thm}\label{thm:createKan} If the comonad $\mathscr{G}$ is left Hopf, then the underlying (pseudo-)functor $\mathscr{U}:\mathscr{V}^\mathscr{G}\rightarrow \mathscr{V}$ creates left Kan extensions. \end{thm} \begin{proof} Consider two coalgebras $(u,\gamma_u)$ and $(v,\gamma_v)$ whose underlying arrows have a left extension $k=\text{lan}_v u$ as shown \begin{equation} \begin{tikzpicture}[xscale=1.5,yscale=1.2, every node/.style={scale=1},baseline=(current bounding box.center)] \node (T0) at (0,1) {$G$}; \node (T1) at (1,1) {$G'$}; \node (T2) at (1,2) {$G''$}; \node (U0) at (0,-1) {$G_0$}; \node (U1) at (1,-1) {$G'_0$}; \node (U2) at (1,-0) {$G''_0$}; \node (dot) at (1.5,-1) {$.$}; \path[->,font=\scriptsize,>=angle 90] (-0.5,0.5) edge node [left] {$\mathscr{U}$} (-0.5,0); \path[->,font=\scriptsize,>=angle 90] (T0) edge node [above left] {$(u,\gamma_u)$} (T2) (T0) edge node [below] {$(v,\gamma_v)$} (T1); \path[dashed,->,font=\scriptsize,>=angle 90] (T1) edge node [right] {$(k,\gamma_k)$} (T2) ; \path[->,font=\scriptsize,>=angle 90] (U0) edge node [above left] {$u$} (U2) (U0) edge node [below] {$v$} (U1); \path[dashed,->,font=\scriptsize,>=angle 90] (U1) edge node [right] {$k$} (U2) ; \node[font=\scriptsize] (kpa) at (0.65,-0.65) {$\kappa\Rightarrow$}; \end{tikzpicture} \end{equation} The universal property of left Kan extensions says there is a bijection \begin{align} \phi:u\Rightarrow l\otimes v \\ \cline{1-1} \bar \phi:k\Rightarrow l \end{align} such that $\phi=(\bar{\phi} \otimes 1)\bullet \kappa$. In particular, there is a map $\gamma_k:k\rightarrow \mathscr{G}k$ corresponding to $u\xRightarrow{\gamma_u}\mathscr{G}u\xRightarrow{\mathscr{G}\kappa}\mathscr{G}(k\otimes v)\xRightarrow{\mathrm{v}_{k,v}^{-1}}\mathscr{G}k\otimes v$ such that the diagram below commutes. \begin{equation}\label{eq:gammaKanprop} \begin{tikzcd} u \arrow[rightarrow]{r}{\gamma_u} &\mathscr{G}u \arrow[rightarrow]{r}{\mathscr{G}\kappa} & \mathscr{G}(k\otimes v) \arrow[rightarrow]{d}{\mathrm{v}_{k,v}^{-1}} \\ k\otimes v \arrow[leftarrow]{u}{\kappa} \arrow[rightarrow]{rr}{\gamma_k\otimes 1} & & \mathscr{G}(k)\otimes v \\ \end{tikzcd} \end{equation} The obtained arrow $\gamma_k$ defines a coalgebra structure on $k$, where the compatibility with $\delta$ and $\epsilon$ follows from \begin{equation} \begin{tikzcd} \, & u \arrow[rightarrow,swap]{ldd}{\kappa} \arrow[rightarrow]{dd}{\gamma} \arrow[rightarrow]{rd}{\gamma} \arrow[rightarrow]{rrrd}{\kappa} &&& \\ & & \mathscr{G} u \arrow[rightarrow]{ddr}{\mathscr{G}\kappa} \arrow[rightarrow,swap]{d}{\mathscr{G}\gamma} && k\otimes v \arrow[rightarrow]{ddd}{\gamma\otimes 1} \\ k\otimes v \arrow[rightarrow,swap]{ddr}{\gamma\otimes 1} & \mathscr{G} u \arrow[rightarrow]{d}{\mathscr{G}\kappa} \arrow[rightarrow]{r}{\delta} & \mathscr{G}^2 u \arrow[rightarrow,swap]{d}{\mathscr{G}^2\kappa} && \\ & \mathscr{G}(k\otimes v) \arrow[rightarrow]{d}{\mathrm{v}^{-1}} \arrow[rightarrow]{r}{\delta} & \mathscr{G}^2(k\otimes v) \arrow[rightarrow,swap]{d}{\mathscr{G}\mathrm{v}^{-1}} & \mathscr{G}(k\otimes v) \arrow[rightarrow]{dr}{\mathrm{v}^{-1}} \arrow[rightarrow]{dl}{\mathscr{G}(\gamma \otimes 1)} & \\ & \mathscr{G}k\otimes v \arrow[rightarrow,swap]{dr}{\delta\otimes 1} & \mathscr{G}(\mathscr{G}k\otimes v) \arrow[rightarrow]{d}{\mathrm{v}^{-1}} & & \mathscr{G}k\otimes v \arrow[rightarrow]{dll}{\mathscr{G}\gamma\otimes 1} \\ & & \mathscr{G}^2 k\otimes v && \end{tikzcd} \end{equation} \begin{equation} \begin{tikzcd} u \arrow[rightarrow]{rrr}{\kappa} \arrow[rightarrow]{dr}{1} \arrow[rightarrow,swap]{rdd}{\gamma} \arrow[rightarrow,swap]{ddd}{\kappa} && & k\otimes v \\ & u \arrow[rightarrow]{urr}{\kappa} & & \\ & \mathscr{G} u \arrow[rightarrow,swap]{u}{\epsilon} \arrow[rightarrow]{r}{\mathscr{G}\kappa} & \mathscr{G}(k\otimes v) \arrow[rightarrow,swap]{uur}{\epsilon} \arrow[rightarrow,swap]{dr}{\mathrm{v}^{-1}} & \\ k\otimes v \arrow[rightarrow,swap]{rrr}{\gamma\otimes 1} & & & \mathscr{G}k\otimes v \arrow[rightarrow,swap]{uuu}{\epsilon\otimes 1} \end{tikzcd} \end{equation} The 2-cell $\kappa$ is a coalgebra morphism, which is obvious after substituting $\mathrm{v}^{-1}$ in (\ref{eq:gammaKanprop}). To see that $\kappa$ exhibits $(k,\gamma_k)$ as a left extension of $(u,\gamma_u)$ through $(v,\gamma_v)$, consider a coalgebra $(l,\gamma_l):G'\rightarrow G''$, and a coalgebra morphism $\phi:u\Rightarrow l\otimes v$. In $\mathscr{V}$, the Kan extension universal property gives $\bar\phi:k\Rightarrow l$. Using the commuting diagram \begin{equation} \begin{tikzcd} \, & u \arrow[rightarrow,swap]{ld}{\kappa} \arrow[rightarrow]{d}{\gamma} \arrow[rightarrow,swap]{rrd}{\phi} \arrow[rightarrow]{r}{\kappa} & k\otimes v \arrow[rightarrow]{rd}{\bar{\phi}\otimes 1} && \\ k\otimes v \arrow[rightarrow,swap]{ddr}{\gamma\otimes 1} & \mathscr{G} u \arrow[rightarrow]{d}{\mathscr{G}\kappa} \arrow[rightarrow]{rd}{\mathscr{G}\phi} & & l\otimes v \arrow[rightarrow]{ld}{\gamma} \arrow[rightarrow]{ldd}{\gamma\otimes 1} & \\ & \mathscr{G}(k\otimes v) \arrow[rightarrow]{d}{\mathrm{v}^{-1}} \arrow[rightarrow,swap]{r}{\mathscr{G}(\bar{\phi}\otimes 1)} & \mathscr{G}(l\otimes v) \arrow[rightarrow,swap]{d}{\mathrm{v}^{-1}} & & \\ & \mathscr{G}k\otimes v \arrow[rightarrow,swap]{r}{\mathscr{G}\bar{\phi}\otimes 1} & \mathscr{G} l\otimes v & & \end{tikzcd} \end{equation} we see that $\bar\phi$ is a coalgebra morphism. \end{proof} Recall that $m$ having an adjoint is equivalent to the existence of a left Kan extension $\text{lan}_m 1_M$ which is respected by $m$; that is, $m\circ\text{lan}_m 1_M=\text{lan}_m m$. \begin{cor} With a Hopf-comonadic $\mathscr{U}:\mathscr{N}\rightarrow\mathscr{M}$, an arrow $n\in\mathscr{N}(N,N')$ has a right adjoint if and only if $\mathscr{U}n$ does. \end{cor} \begin{proof} Being a pseudo functor, $\mathscr{U}$ preserves adjoints. The other way around, assume $Un$ has a right adjoint, that is both $\text{lan}_{Un} 1_{UN}$ and $\text{lan}_{Un} Un$ exist. From the previous theorem, $\text{lan}_{n} 1_N$ exists and $n\circ\text{lan}_{n} 1_N$ is taken to $U(n\circ\text{lan}_{n} 1_N)\cong Un\circ\text{lan}_{Un} 1_{UN}\cong \text{lan}_{Un} Un$ which creates $\text{lan}_{n} n$. \end{proof} \begin{thm}\label{thm:comonadicModU} If $\mathscr{V}$ is locally cocomplete then the induced underlying functor is well-defined \begin{align}\label{eq:modenBeck} \mathscr{N}:= \text{Moden}(\mathscr{X},\mathscr{V}^\mathscr{G}) \xrightarrow{\widetilde{\mathscr{U}}:=\text{Moden}(\mathscr{X},\mathscr{U})} \mathscr{M:=}\text{Moden}(\mathscr{X},\mathscr{V}) \end{align} and it is comonadic in $\text{CATEN}$. Denote its right adjoint by $\widetilde{\mathscr{R}}$. If $\mathscr{R}$ preserves local colimits, and $\mathscr{G}$ is Hopf, then the induced comonad $\widetilde{\mathscr{G}}:=\widetilde{\mathscr{U}}\circ\widetilde{\mathscr{R}}$ is also Hopf. \end{thm} \begin{proof} We will consider the case when $\mathscr{X}$ is the terminal bicategory: then $\mathscr{N}=\mathscr{V}^\mathscr{G}\text{-Mod}$, and $\mathscr{M}=\mathscr{V}\text{-Mod}$. By Proposition 7.5 of \cite{Kelly2002}, $\widetilde{\mathscr{U}}$ is a lax functor. First we show that it has local right adjoints $\widetilde{\mathscr{R}}_\mathscr{A}^\mathscr{B}$ given by \begin{align} \mathscr{M}(\mathscr{U}\circ \mathscr{A},\mathscr{U}\circ \mathscr{B}) &\xrightarrow{\widetilde{\mathscr{R}}_\mathscr{A}^\mathscr{B}} \mathscr{N}(\mathscr{A},\mathscr{B}) \\ (\mathscr{U}\circ \mathscr{A}\xmrightarrow{M}\mathscr{U}\circ \mathscr{B},\alpha) &\mapsto ( \mathscr{A}\xmrightarrow{\widetilde{\mathscr{R}}M} \mathscr{B},\widetilde{\mathscr{R}}\alpha)\\ (\sigma:M\Rightarrow N)&\mapsto (\widetilde{\mathscr{R}}\sigma:\widetilde{\mathscr{R}}M\Rightarrow \widetilde{\mathscr{R}}N) \end{align} where $\alpha$ denotes a 2-sided action (the analogous 1-sided actions are denoted by $\lambda$ and $\rho$) and the assignments are defined by \begin{align} (\widetilde{\mathscr{R}}M)_B^A&:=\mathscr{R}M_B^A\\ (\widetilde{\mathscr{R}}\alpha)_{B,A'}^{B',A}&:=\mathscr{A}_A^{A'}\otimes\mathscr{R}M_{B'}^A\otimes\mathscr{B}_B^{B'}\\ &\xRightarrow{\gamma\otimes 1\otimes \gamma} \mathscr{R}\mathscr{U}\mathscr{A}_A^{A'} \otimes \mathscr{R}M_{B'}^A \otimes \mathscr{R}\mathscr{U}\mathscr{B}_B^{B'}\nonumber\\ &\xRightarrow{\mu^{(\mathscr{R})}} \mathscr{R}(\mathscr{U}\mathscr{A}_A^{A'} \otimes M_{B'}^A \otimes \mathscr{U}\mathscr{B}_B^{B'})\nonumber\\ &\xRightarrow{\mathscr{R}(\alpha)} \mathscr{R}M_B^{A'}\nonumber\nonumber\\ (\widetilde{\mathscr{R}}\sigma)_B^{A}&:=\mathscr{R}\sigma_B^{A}=\sigma_B^A\,. \end{align} Actions $\widetilde{\mathscr{R}}\alpha$, (or separately $\widetilde{\mathscr{R}}\lambda$ and $\widetilde{\mathscr{R}}\rho$) are compatible with unit and composition in $\mathscr{A}$ and $\mathscr{B}$. For example, compatibility of $\rho$ with composition is shown by commutativity of the diagram (\ref{diag:rhoMuComp}). \begin{equation}\label{diag:rhoMuComp} \begin{tikzpicture}[xscale=2,yscale=2, every node/.style={scale=0.8},baseline=(current bounding box.center)] \node (11) at (0,0) {$\mathscr{R}M_{B''}^A\otimes \mathscr{B}_{B'}^{B''}\otimes \mathscr{B}_B^{B'}$}; \node (12) at (0,-1) {$\mathscr{R}M\otimes \mathscr{R}\mathscr{U}\mathscr{B}\otimes \mathscr{B}$}; \node (13) at (0,-2) {$\mathscr{R}(M\otimes \mathscr{U}\mathscr{B})\otimes \mathscr{B}$}; \node (14) at (0,-3) {$\mathscr{R}M\otimes \mathscr{B}$}; \node (24) at (1.5,-3) {$\mathscr{R}M\otimes \mathscr{R}\mathscr{U}\mathscr{B}$}; \node (34) at (3,-3) {$\mathscr{R}(M\otimes \mathscr{U}\mathscr{B})$}; \node (44) at (4.5,-3) {$\mathscr{R}\mathscr{B}$}; \node (41) at (4.5,0) {$\mathscr{R}M\otimes \mathscr{B}$}; \node (42) at (4.5,-1) {$\mathscr{R}M\otimes \mathscr{R}\mathscr{U}\mathscr{B}$}; \node (43) at (4.5,-2) {$\mathscr{R}(M\otimes \mathscr{U}\mathscr{B})$}; \node (i11) at (1.25,-0.5) {$\mathscr{R}M\otimes \mathscr{R}\mathscr{U}\mathscr{B}\otimes \mathscr{R}\mathscr{U}\mathscr{B}$}; \node (i21) at (3.25,-0.5) {$\mathscr{R}M\otimes \mathscr{R}\mathscr{U}(\mathscr{B}\otimes \mathscr{B})$}; \node (i12) at (1.25,-1.5) {$\mathscr{R}(M\otimes \mathscr{U}\mathscr{B})\otimes \mathscr{R}\mathscr{U}\mathscr{B}$}; \node (i22) at (2.5,-2.5) {$\mathscr{R}(M\otimes \mathscr{U}\mathscr{B}\otimes \mathscr{U}\mathscr{B})$}; \node (ii1) at (2.5,-1) {$\mathscr{R}M\otimes \mathscr{R}(\mathscr{U}\mathscr{B}\otimes \mathscr{U}\mathscr{B})$}; \node (ii2) at (3.5,-1.5) {$\mathscr{R}(M\otimes \mathscr{U}(\mathscr{B}\otimes \mathscr{B}))$}; \path[->,font=\scriptsize,>=angle 90] (11) edge node[left] {$1\otimes \gamma\otimes 1$} (12) (12) edge node[left] {$\mu^{(\mathscr{R})}\otimes 1$} (13) (13) edge node[left] {$\mathscr{R}\rho\otimes 1$} (14) (14) edge node[below] {$1\otimes \gamma$} (24) (24) edge node[below] {$\mu^{(\mathscr{R})}$} (34) (34) edge node[below] {$\mathscr{R}\rho$} (44); \path[->,font=\scriptsize,>=angle 90] (11) edge node[above] {$1\otimes\mu^{(\mathscr{B})}$} (41) (41) edge node[right] {$1\otimes\gamma$} (42) (42) edge node[right] {$\mu^{(\mathscr{R})}$} (43) (43) edge node[right] {$\mathscr{R}\rho$} (44); \path[->,font=\scriptsize,>=angle 90] (11) edge node[above right] {$1\otimes \gamma\otimes \gamma$} (i11) (12) edge node[above left] {$1\otimes 1\otimes\gamma$} (i11) (i11) edge node[above] {$1\otimes \mu^{(\mathscr{R}\circ\mathscr{U})}$} (i21) (i21) edge node [fill=white,rounded corners=1pt,inner sep=0.3pt] {$1\otimes\mathscr{R}\mathscr{U}\mu^{(\mathscr{B})}$} (42) (13) edge node[above left] {$1\otimes \gamma$} (i12) (i11) edge node[left] {$\mu^{(\mathscr{R})}\otimes 1$} (i12) (i12) edge node[left] {$\mathscr{R}\rho \otimes 1$} (24); \path[->,font=\scriptsize,>=angle 90] (i12) edge node[below left] {$\mu^{(\mathscr{R})}$}(i22) (i22) edge node[right] {$\mathscr{R}(\rho \otimes 1)$} (34) (i22) edge node[below right] {$\mathscr{R}(1\otimes \mu^{(\mathscr{U}\circ\mathscr{B})})$} (43) ; \path[->,font=\scriptsize,>=angle 90] (i11) edge node[below left] {$1\otimes\mu^{(\mathscr{R})}$}(ii1) (ii1) edge node[left] {$1\otimes \mathscr{R}\mu^{(\mathscr{U})}$} (i21) (ii1) edge node[above left] {$\mu^{(\mathscr{R})}$} (i22) (i21) edge node[right] {$\mu^{(\mathscr{R})}$}(ii2) (i22) edge node[fill=white,rounded corners=1pt,inner sep=0.3pt]{$\mathscr{R}(1\otimes \mu^{(\mathscr{U})})$} (ii2) (ii2) edge node[below left] {$\mathscr{R}(1 \otimes \mathscr{U}\mu^\mathscr{B})$} (43) ; \end{tikzpicture} \end{equation} In this, the non-obvious equalities might be the top pentagon, which is just stating that components of $\mu^{(\mathscr{B})}$ are coalgebra morphisms, naturality of $\mu$ squares, and the bottom right square obtained by applying $\mathscr{R}$ to compatibility of $\rho$ with $\mu^{(\mathscr{U}\circ \mathscr{B})}$. Similarly, components of $\eta^{(\mathscr{B})}$ being coalgebra morphisms leads to compatibility of $\widetilde{\mathscr{R}}\rho$ with $\eta^{(\mathscr{B})}$. Compatibility of $\widetilde{\mathscr{R}}\sigma$ with $\widetilde{\mathscr{R}}\rho$ (and $\widetilde{\mathscr{R}}\lambda$) follows directly from the compatibility of $\sigma$ with $\rho$ (and $\lambda$). The components of the unit and counit of the local adjunctions are given by components of $\gamma$ and $\epsilon$: \begin{align} \tilde{\eta}_\mathscr{A}^\mathscr{B}:1_{\mathscr{N}(\mathscr{A},\mathscr{B})} & \Rightarrow\widetilde{\mathscr{R}}_\mathscr{A}^\mathscr{B}(\mathscr{U}\circ -)\\ ((\tilde{\eta}_\mathscr{A}^\mathscr{B})_N)_B^A=\gamma_{N_B^A}:N_B^A &\xRightarrow{}\mathscr{R}\mathscr{U}N_B^A\\ \tilde{\epsilon}_{\mathscr{A}}^{\mathscr{B}}: \mathscr{U}\circ\widetilde{\mathscr{R}}_\mathscr{A}^\mathscr{B}( -) & \Rightarrow 1_{\mathscr{M}(\mathscr{U}\circ\mathscr{A},\mathscr{U}\circ\mathscr{B})} \\ ((\tilde{\epsilon}_\mathscr{A}^\mathscr{B})_M)_B^A=\epsilon_{M_B^A}: \mathscr{U}\mathscr{R} M_B^A &\xRightarrow{}M_B^A\,. \end{align} They form module morphisms, as proved by diagrams \begin{equation} \begin{tikzpicture}[xscale=6, yscale=1.5, every node/.style={scale=1},baseline=(current bounding box.center)] \node (nb) at (0,0) {$N_B^A\otimes \mathscr{B}_B^{B'}$}; \node (runb) at (1,0) {$\mathscr{R}\mathscr{U}N_B^A\otimes \mathscr{B}_{B'}^B$}; \node (n) at (0,-3) {$N_{B'}^A$}; \node (runrub) at (1,-1) {$\mathscr{R}\mathscr{U}N_B^A\otimes \mathscr{R}\mathscr{U}\mathscr{B}_{B'}^B$}; \node (runub) at (1,-2) {$\mathscr{R}(\mathscr{U}N_B^A\otimes \mathscr{U}\mathscr{B}_{B'}^B)$}; \node (run) at (1,-3) {$\mathscr{R}\mathscr{U}N_{B'}^A$}; \node (irunb) at (0.35,-1.5) {$\mathscr{R}\mathscr{U}(N_B^A\otimes \mathscr{B}_B^{B'})$}; \path[->,font=\scriptsize,>=angle 90] (nb) edge node[left] {$\rho^{(N)}$} (n) (n) edge node[below] {$\gamma$} (run) (nb) edge node[above] {$\gamma\otimes 1$} (runb) (nb) edge node[right] {$\gamma$} (irunb) (runrub) edge node[above left] {$\mu^{(\mathscr{R}\mathscr{U})}$} (irunb) (irunb) edge node[below left] {$\mathscr{R}\mathscr{U}\rho^{(N)}$} (run) (runub) edge node[above right] {$\mathscr{R}\mu^{(\mathscr{U})}$} (irunb) (runb) edge node[right] {$1 \otimes \gamma$} (runrub) (nb) edge node[above right] {$\gamma\otimes \gamma$} (runrub) (runrub) edge node[right] {$\mu^{(\mathscr{R})}$} (runub) (runub) edge node[right] {$\mathscr{R}\rho^{(\mathscr{U}N)}$} (run) ; \end{tikzpicture} \end{equation} \begin{equation} \begin{tikzpicture}[xscale=8, yscale=1.5, every node/.style={scale=1},baseline=(current bounding box.center)] \node (urmub) at (0,0) {$\mathscr{U}\mathscr{R}M_B^A\otimes \mathscr{U}\mathscr{B}_{B'}^B$}; \node (urmb) at (0,-1) {$\mathscr{U}(\mathscr{R}M_B^A\otimes\mathscr{B}_{B'}^B)$}; \node (urmrub) at (0,-2) {$\mathscr{U}(\mathscr{R}M_B^A\otimes \mathscr{R}\mathscr{U}\mathscr{B}_{B'}^B)$}; \node (urmub2) at (0,-3) {$\mathscr{U}\mathscr{R}(M_B^A\otimes \mathscr{U}\mathscr{B}_{B'}^B)$}; \node (urm) at (0,-4) {$\mathscr{U}\mathscr{R}M_{B'}^A$}; \node (mub) at (1,-1) {$M_B^A\otimes\mathscr{U}\mathscr{B}_{B'}^B$}; \node (m) at (1,-2) {$M_{B'}^A$}; \node (urmurub) at (0.5,-1.25) {$\mathscr{U}\mathscr{R}M_B^A\otimes \mathscr{U}\mathscr{R}\mathscr{U}\mathscr{B}_{B'}^B$}; \path[->,font=\scriptsize,>=angle 90] (urmub) edge node [left] {$\mu^{(\mathscr{U})}$} (urmb) (urmb) edge node [left] {$\mathscr{U}(1\otimes \gamma)$} (urmrub) (urmrub) edge node [left] {$\mathscr{U}\mu^{(\mathscr{R})}$} (urmub2) (urmub2) edge node [left] {$\mathscr{U}\mathscr{R}\rho^{(M)}$} (urm) (urm) edge node [below] {$\epsilon$} (m) (urmub) edge node [above] {$\epsilon\otimes 1$} (mub) (mub) edge node [right] {$\rho^{(M)}$} (m) (urmub) edge node [right] {$1\otimes \mathscr{U}\gamma$} (urmurub) (urmurub) edge node [above left] {$\epsilon\otimes\epsilon$} (mub) (urmurub) edge node [above] {$\mu^{(\mathscr{U})}$} (urmrub) (urmurub) edge node [right] {$\mu^{(\mathscr{U}\mathscr{R})}$} (urmub2) ; \path[->,font=\scriptsize,>=angle 90] (urmub2) edge node [below right] {$\epsilon$} (mub) ; \end{tikzpicture} \end{equation} and they satisfy the adjunction axioms because $\gamma$ and $\epsilon$ do. Since $\widetilde{\mathscr{U}}$ has local right adjoints, it preserves local colimits, which, together with pseudofunctoriality of $\mathscr{U}$, gives sufficient conditions for pseudofunctoriality of $\widetilde{\mathscr{U}}$, \begin{equation} \begin{tikzpicture}[scale=1.5, every node/.style={scale=0.9},baseline=(current bounding box.center)] \node (A1) at (-3.5,0.7) { $\sum\mathscr{U}M_{B'}^A\otimes\mathscr{U}\mathscr{B}_B^{B'}\otimes\mathscr{U}N_{C'}^B$}; \node (A2) at (0,0.7) { $\sum\mathscr{U}M_{B}^A\otimes\mathscr{U}N_{C'}^B$}; \node (A3) at (2.7,0.7) { $(\mathscr{U}N\circ_{\mathscr{U}\mathscr{B}}\mathscr{U}M)_{C'}^A$}; \node (B1) at (-3.5,-0.7) { $\sum \mathscr{U}(M_{B'}^A\otimes\mathscr{B}_B^{B'}\otimes N_{C'}^B)$}; \node (B2) at (0,-0.7) { $\sum \mathscr{U}(M_{B}^A\otimes N_{C'}^B)$}; \node (B3) at (2.7,-0.7) { $\mathscr{U}(N\circ_{\mathscr{B}}M)_{C'}^A\,.$}; \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\mathscr{U}\rho\otimes 1$} (A2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (A1)edge node[below] {$1\otimes \mathscr{U}\lambda$} (A2); \path[->,font=\scriptsize,>=angle 90] (A2) edge node[above] {$\text{coeq}$} (A3); \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[above] {$\mathscr{U}(\rho\otimes 1)$} (B2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[below] {$\mathscr{U}(1\otimes\lambda)$} (B2); \path[->,font=\scriptsize,>=angle 90] (B2) edge node[below] {$\mathscr{U}(\text{coeq})$} node[above] {$\text{coeq}$} (B3); \path[transform canvas={xshift=19mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$\sum \mu^{(\mathscr{U})}$} (B1); \path[transform canvas={xshift=-13mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\sum \mu^{(\mathscr{U})}$} (B2); \path[dashed,->,font=\scriptsize,>=angle 90] (A3) edge node[right] {$\mu^{(\widetilde{\mathscr{U}})}$} (B3); \end{tikzpicture} \end{equation} Since $\widetilde{\mathscr{U}}$ is a pseudofunctor and has local right adjoints, by Proposition 2.7 of \cite{Kelly2002}, $\widetilde{\mathscr{R}}$ extends to a 2-sided enriched category which is a right adjoint to $\widetilde{\mathscr{U}}$. $\widetilde{\mathscr{U}}$ also satisfies the other two conditions of Proposition \ref{prop:comonadic} as stated in Lemma \ref{lem:whisk}. This completes the proof that $\widetilde{\mathscr{U}}$ is comonadic. Explicitly, the unit for $\widetilde{\mathscr{R}}$ is a module morphism defined using the enriched functor (\ref{eq:gammaFunct}) \begin{align} \eta_{\mathscr{A}}^{(\widetilde{\mathscr{R}})}:=(\gamma\circ \mathscr{A})_* \end{align} and the multiplication components \begin{equation} \widetilde{\mathscr{R}}(N)\circ_\mathscr{B}\widetilde{\mathscr{R}}(M) \xRightarrow{\mu^{(\widetilde{\mathscr{R}})}} \widetilde{\mathscr{R}}(N\circ_{\mathscr{U}\mathscr{B}}M) \end{equation} are given by the right column of \begin{equation} \hspace{-2cm} \begin{tikzpicture}[scale=1.5, every node/.style={scale=1},baseline=(current bounding box.center)] \node (A1) at (-3.5,0.7) { $\sum\mathscr{R}M_{B'}^A\otimes\mathscr{B}_B^{B'}\otimes\mathscr{R}N_{C'}^B$}; \node (A2) at (0,0.7) { $\sum\mathscr{R}M_{B}^A\otimes\mathscr{R}N_{C'}^B$}; \node (A3) at (2.7,0.7) { $(\mathscr{R}N\circ_{\mathscr{B}}\mathscr{R}M)_{C'}^A$}; \node (B1) at (-3.5,-0.7) { $\mathscr{R}\sum M_{B'}^A\otimes\mathscr{U}\mathscr{B}_B^{B'}\otimes N_{C'}^B$}; \node (B2) at (0,-0.7) { $\mathscr{R}\sum M_{B}^A\otimes N_{C'}^B$}; \node (B3) at (2.7,-0.7) { $\mathscr{R}(N\circ_{\mathscr{U}\mathscr{B}}M)_{C'}^A$};- \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\widetilde{\mathscr{R}}\rho\otimes 1$} (A2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (A1)edge node[below] {$1\otimes \widetilde{\mathscr{R}}\lambda$} (A2); \path[->,font=\scriptsize,>=angle 90] (A2) edge node[above] {$\text{coeq}$} (A3); \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[above] {$\mathscr{R}(\rho\otimes 1)$} (B2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[below] {$\mathscr{R}(1\otimes\lambda)$} (B2); \path[->,font=\scriptsize,>=angle 90] (B2) edge node[below] {$\mathscr{R}(\text{coeq})$} (B3); \path[transform canvas={xshift=19mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$(\mathscr{R}i_{BB'}\bullet\mu^{(\mathscr{R})}\bullet 1\otimes\gamma\otimes 1)_{BB'}$} (B1); \path[transform canvas={xshift=-13mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$(\mathscr{R}i_{B}\bullet \mu^{(\mathscr{R})})_{B}$} (B2); \path[dashed,->,font=\scriptsize,>=angle 90] (A3) edge node[right] {$\mu^{(\widetilde{\mathscr{R}})}$} (B3); \end{tikzpicture} \end{equation} where the top line is defining composition of modules in $\mathscr{N}$, the bottom line is $\mathscr{R}$ applied to the defining composition of modules in $\mathscr{M}$, $i_{B}$ and $i_{BB'}$ are the coproduct inclusions, and $(-)_B$ denotes the induced map for mapping out of a coproduct. For modules $M\in \mathscr{N}(\mathscr{A},\mathscr{B})$ and $N\in \mathscr{M}(\mathscr{U}\circ\mathscr{B},\mathscr{U}\circ\mathscr{C})$, the (left) Hopf map, given by the right column of \begin{equation} \hspace{-0.5cm} \begin{tikzpicture}[scale=1.5, every node/.style={scale=1},baseline=(current bounding box.center)] \node (C1) at (-3.5,2.1) { $\sum M_{B'}^A\otimes\mathscr{B}_B^{B'}\otimes\mathscr{R}N_{C'}^B$}; \node (C2) at (0,2.1) { $\sum M_{B}^A\otimes\mathscr{R}N_{C'}^B$}; \node (C3) at (2.7,2.1) { $(\mathscr{R}N\circ_{\mathscr{B}} M)_{C'}^A$}; \node (A1) at (-3.5,0.7) { $\sum\mathscr{R}\mathscr{U}M_{B'}^A\otimes\mathscr{B}_B^{B'}\otimes\mathscr{R}N_{C'}^B$}; \node (A2) at (0,0.7) { $\sum\mathscr{R}\mathscr{U}M_{B}^A\otimes\mathscr{R}N_{C'}^B$}; \node (A3) at (2.7,0.7) { $(\mathscr{R}N\circ_{\mathscr{B}}\mathscr{R}\mathscr{U}M)_{C'}^A$}; \node (B1) at (-3.5,-0.7) { $\sum \mathscr{R}(\mathscr{U}M_{B'}^A\otimes\mathscr{B}_B^{B'}\otimes N_{C'}^B)$}; \node (B2) at (0,-0.7) { $\sum \mathscr{R}(\mathscr{U}M_{B}^A\otimes N_{C'}^B)$}; \node (B3) at (2.7,-0.7) { $\mathscr{R}(N\circ_{\mathscr{U}\mathscr{B}}\mathscr{U}M)_{C'}^A$};- \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (C1) edge node[above] {$\widetilde{\mathscr{R}}\rho\otimes 1$} (C2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (C1)edge node[below] {$1\otimes \widetilde{\mathscr{R}}\lambda$} (C2); \path[->,font=\scriptsize,>=angle 90] (C2) edge node[above] {$\text{coeq}$} (C3); \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (A1) edge node[above] {$\widetilde{\mathscr{R}}\rho\otimes 1$} (A2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (A1)edge node[below] {$1\otimes \widetilde{\mathscr{R}}\lambda$} (A2); \path[->,font=\scriptsize,>=angle 90] (A2) edge node[above] {$\text{coeq}$} (A3); \path[transform canvas={yshift=1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[above] {$\mathscr{R}(\rho\otimes 1)$} (B2); \path[transform canvas={yshift=-1mm},->,font=\scriptsize,>=angle 90] (B1) edge node[below] {$\mathscr{R}(1\otimes\lambda)$} (B2); \path[->,font=\scriptsize,>=angle 90] (B2) edge node[below] {$\mathscr{R}(\text{coeq})$} node[above] {$\text{coeq}$} (B3); \path[transform canvas={xshift=15mm},->,font=\scriptsize,>=angle 90] (C1) edge node[left] {$\sum \gamma\otimes 1\otimes 1$} (A1); \path[transform canvas={xshift=-9mm},->,font=\scriptsize,>=angle 90] (C2) edge node[right] {$\sum \gamma\otimes 1$} (A2); \path[->,font=\scriptsize,>=angle 90] (C3) edge node[left] {$(1\otimes \gamma)_{C'}^A$} (A3); \path[transform canvas={xshift=15mm},->,font=\scriptsize,>=angle 90] (A1) edge node[left] {$\sum\mu^{(\mathscr{R})}\bullet 1\otimes\gamma\otimes 1$} (B1); \path[transform canvas={xshift=-9mm},->,font=\scriptsize,>=angle 90] (A2) edge node[right] {$\sum \mu^{(\mathscr{R})}$} (B2); \path[->,font=\scriptsize,>=angle 90] (A3) edge node[left] {$\mu^{(\widetilde{\mathscr{R}})}$} (B3); \end{tikzpicture} \end{equation} is invertible because the other two columns are invertible: they are determined by the Hopf maps for $\mathscr{G}$. \end{proof} \begin{example} A review of (differential) graded abelian groups is available in \cite{Nikolic2018}. In the diagram \begin{equation}\label{diag:adjs} \begin{tikzpicture}[scale=2, every node/.style={scale=1},baseline=(current bounding box.center)] \node (d) at (-2,0) {$\mathrm{DGAb}$}; \node (g) at (0,0) {$\mathrm{GAb}$}; \node (a) at (2,0) {$\mathrm{Ab}$}; \node (sym1) at (-1,0.2) {$\bot$}; \node (sym2) at (-1,-0.2) {$\bot$}; \node (sym3) at (1,-0.2) {$\bot$}; \path[,->,font=\scriptsize,>=angle 90,bend right] (g) edge node[above] {$L$} (d); \path[,->,font=\scriptsize,>=angle 90] (d) edge node[below left] {$U$} (g); \path[->,font=\scriptsize,>=angle 90,bend left] (g)edge node[below] {$R$} (d); \path[,->,font=\scriptsize,>=angle 90,bend left] (a) edge node[below] {$C$} (g); \path[,->,font=\scriptsize,>=angle 90] (g) edge node[above] {$\Sigma$} (a); \end{tikzpicture} \end{equation} functors $U$ and $\Sigma$ forget differential and take a sum of all components of the graded abelian group. They are both Hopf-comonadic, and by Theorem \ref{thm:createKan} create duals and cohoms. An abelian group $A$ has a dual if and only if it is finitely generated and projective \cite{Street2007}. As a consequence of $\Sigma$ being Hopf-comonadic, a graded abelian group $A$ has a dual if and only if it has finitely many non-zero components each of which is finitely generated and projective. As a consequence of $U$ being Hopf-comonadic, a chain complex $A$ has a dual if and only if its underlying graded abelian group does. \end{example} \begin{example} Since $U$ is a left adjoint it preserves colimits, so by the Theorem \ref{thm:comonadicModU} the change of base functor $\widetilde{U}$ creates Cauchy modules. \end{example} \bibliographystyle{acm}

7ed2bcf35782a83367e3c6b42f6c79fe41cf24d4

\section{Introduction} In 1914 Kempner \cite{Ke14} introduced a variant of the harmonic series which excluded from its sum all those positive integers that contain the digit $9$ in their base-$10$ expansions. Unlike the familiar harmonic series, Kempner's modified series converges (the limit later shown to be $\approx 22.92$, see \cite{Ba79}). A simple generalisation of Kempner's original argument shows that convergence occurs as long as any non-empty set of digits is excluded, and that this result holds in any base (see \cite{BS08}, for example). Let us introduce some notation to describe these results in general. Fix an integer $b \geqslant 2$ and a subset of integers $S \subseteq [0,b-1]$. Here and throughout the paper, for two integers $x$ and $y$ we use $[x,y]$ to denote the set $\{n\in\mathbb{Z}:x \leqslant n\leqslant y\}$. We then define the \emph{Kempner set} $\mathcal{K}(S,b)$ to be the set of non-negative integers that, when written in base $b$, contain only digits from $S$. Thus $\mathcal{K}([0,8],10)$ denotes the set originally studied by Kempner. We will assume throughout that $0 \in S$, to avoid the ambiguity of leading zeros, and require $S \neq [0,b-1]$, to preclude the trivial set $\mathcal{K}([0,b-1],b)$ (which is nothing more than $\mathbb{Z}_{\geqslant 0}$). These sets $S$ will be referred to as the \emph{permitted} sets $S$, and the related Kempner sets $K(S,b)$ as \emph{proper} Kempner sets. The arithmetic properties of proper Kempner sets have been the object of considerable study in recent years, beginning with the work of Erd\H{o}s, Mauduit, and S\'ark\"ozy, who studied the distribution of residues in $\mathcal{K}(S,b)$ moduli small numbers \cite{EMS98} and proved the existence of integers in $\mathcal{K}(S,b)$ with many small prime factors \cite{EMS99}. Notable recent work includes Maynard's proof \cite{May16} that the sets $\mathcal{K}(S,b)$ contain infinitely many primes whenever $b-\vert S\vert$ is at most $b^{23/80}$, provided $b$ is sufficiently large. In this paper we consider the additive structure of proper Kempner sets. In particular, we consider the following extremal question: \emph{what is the length of the longest arithmetic progression in a proper Kempner set with a fixed given base?} Our methods will be combinatorial, rather than analytic (as in Maynard's work, \cite{May16}). A well known conjecture of Erd\H{o}s-Tur\'{a}n (first given in \cite{ErTu36}) states that any set of positive integers with a divergent harmonic sum contains arithmetic progressions of arbitrary (finite) length. Since proper Kempner sets have convergent harmonic sums, this might suggest that the lengths of arithmetic progressions in a given proper Kempner set are uniformly bounded. This is indeed the case. Let us say that a set $T\subset \mathbb{Z}$ is $k$-free if $T$ contains no arithmetic progression of length $k$. By a simple argument, given in Proposition \ref{prop:ell_finite}, one may show that the proper Kempner set $\mathcal{K}(S,b)$ is $(b^2-b+1)$-free for any $b \geqslant 2$. The main purpose of this article is to understand how close this trivial upper bound is to the truth. \\ In our main theorem, we improve this bound for all $b> 2$, obtaining a tight result that expresses the length of the longest arithmetic progression in $\mathcal{K}(S,b)$ in terms of the prime factorisation of $b$. To state this theorem, we need to introduce some arithmetic functions. If $n$ and $b$ are natural numbers, let $\rho(n)$ denote the square-free radical of $n$ (ie. the product of all distinct primes dividing $n$), and let $\beta(b)$ denote the largest integer less than $b$ such that $\rho(\beta(b))\vert b$. For example, $\beta(10) = 8$, and $\beta(p^k) = p^{k-1}$ for any prime power $p^k$. In other words, $\beta(b)$ is the greatest integer less than $b$ that divides some power of $b$. Finally, let $\ell(b)$ be the length of the longest arithmetic progression contained in some proper Kempner set of base $b$. Our main theorem gives an exact evaluation of $\ell(b)$. \begin{theorem} \label{thm:main theorem} For all $b\geqslant 2$, one has $\ell(b) = (b-1)\beta(b)$. \end{theorem} \noindent For example, $\ell(10)=72$. One particular set that achieves this bound is Kempner's original set, $\mathcal{K}([0,8],10)$, which contains the $72$-term arithmetic progression $\{0,125,250,375,\cdots,8875\}$.\\ The arithmetic functions $\beta(b)$ and $\ell(b)$ are of independent interest, but do not appear to have been considered seriously before.\footnote{The sequence $\beta(b)$ is entry A079277 on the Online Encyclopedia of integer Sequences.} We establish average order results for $\beta(b)$ which show that, for most $b$, the trivial upper bound on $\ell(b)$ from Proposition \ref{prop:ell_finite} is asymptotically correct. \begin{theorem} \label{thm: ell nearly always maximal epsilon free} There is a set of integers $A\subset \mathbb{Z}$ with natural density $1$, i.e. with \[ \lim\limits_{N\rightarrow \infty} \frac{1}{N}\vert A\cap [1,N]\vert = 1,\] such that $\ell(b)\sim b^2$ as $b\rightarrow \infty$ in $A$. \end{theorem} \noindent \textit{Notation:} For $x\in \mathbb{R}$, let $\{x\}$ denote the fractional part of $x$ and let $ \lfloor x \rfloor$ denote the greatest integer that is at most $x$. For a natural number $n$, we let $[n]$ denote the set of integers $\{1,\cdots,n\}$. As mentioned previously, for two integers $x$ and $y$ we use $[x,y]$ to denote the set $\{ n\in \mathbb{Z}: x\leqslant n\leqslant y\}$. We use the notation $\log_q p$ to denote the logarithm of $p$ to base $q$ (as opposed to any iterations of logarithms).\\ \section{Progressions of Maximal Length in Kempner Sets} In this section we give our proof of Theorem \ref{thm:main theorem}, which is an exact evaluation of $\ell(b)$ and the main result of this paper. This will be done in two parts: a constructive lower bound and a proof that this lower bound is sharp. Before that, as promised, we give a simple proof that the function $\ell(b)$ is at least well-defined, i.e. that Kempner sets do not contain arbitrarily long arithmetic progressions. \begin{proposition} \label{prop:ell_finite} For all $b\geqslant 2$, we have $\ell(b) \leqslant (b-1)b$. \end{proposition} \begin{proof} Suppose that $A \subset \mathcal{K}(S,b)$ is a finite arithmetic progression of $\vert A\vert$ terms with common difference $\Delta$. Choose $k \geqslant 0$ such that $b^k \leqslant \Delta < b^{k+1}$. If $I$ denotes the shortest interval of integers containing $A$, then $\vert I \vert = (\vert A \vert-1)\Delta+1$, hence $\vert A\vert = 1+ (\vert I\vert -1)/\Delta$. If $A$ excludes the digit $d$, the upper bound $\Delta < b^{k+1}$ confines $A$ within the interval $[0,db^{k+1} - 1]$ or within an interval of the form \[[b^{k+2} m + (d+1) b^{k+1}, b^{k+2}(m+1)+ db^{k+1}-1],\] for some $m\in \mathbb{Z}_{\geqslant 0}$. Thus $\vert I\vert \leqslant b^{k+2}-b^{k+1}$, which yields \[\vert A\vert \leqslant 1 + \frac{b^{k+2}-b^{k+1}-1}{\Delta} < 1 + \frac{b^{k+2} - b^{k+1}}{b^{k}} \leqslant b^2-b+1,\] hence $\vert A\vert \leqslant b^2-b$ as claimed. \end{proof} The bound in the previous proposition is simple and -- as a consequence -- occasionally weak. In particular, it neglects the potentially compounding effects of digit exclusion at different orders of magnitude, and the arithmetic properties of orbits in the group $\mathbb{Z}/b\mathbb{Z}$. This structure can affect the bounds dramatically, as seen most clearly in the case when the base $b$ is prime. \begin{proposition} \label{prime prop} Let $p$ be prime. Then $\ell(p) \leqslant p-1$. \end{proposition} \begin{proof} Suppose that $\mathcal{K}(S,p)$ contains the progression $A = \{k+ j \Delta : j \in [p]\}$ with $\Delta\neq 0$. By the pigeonhole principle, there exist distinct $i,j\in [p]$ with $k+j \Delta \equiv k+i \Delta \!\!\mod p$ for some $i\neq j$, hence $p \mid \Delta$ (since $p$ is prime). By deleting the rightmost digits of the elements of $A$ we obtain a new progression in $\mathcal{K}(S,p)$ with common difference $\Delta/p$; in particular, the progression $$\left\{\left\lfloor \frac{k}{p} \right\rfloor + j\frac{\Delta}{p}: j\in [p]\right\}.$$ The new common difference is strictly smaller, and we obtain a contradiction by infinite descent. \end{proof} With a little more bookkeeping this proof generalizes to prime powers, and implies that $\ell(p^k) \leqslant p^{k-1}(p^k-1)$. So certainly $\ell(b)$ is not asymptotic to $b^2$ as $b$ ranges over all integers; some restriction in Theorem \ref{thm: ell nearly always maximal epsilon free} is required.\\ We now begin the proof of Theorem \ref{thm:main theorem}. Searching for long progressions in $\mathcal{K}([0,8],10)$, one might happen across the example noted earlier, namely the first $71$ multiples of $125$, which -- together with $0$ -- form an arithmetic progression of length $72$, none of whose members contain the digit 9. This example succeeds due to properties of the prime factorisation of $1000/125$, in relation to the base $10$. These properties generalise, and one may use this to construct long digit-excluding arithmetic progressions in arbitrary bases. \begin{proposition} \label{prop:AP construction} For all $b \geqslant 2$, the Kempner set $\mathcal{K}([0,b-2],b)$ contains an arithmetic progression of length $(b-1)\beta(b)$. Hence $\ell(b) \geqslant (b-1)\beta(b)$. \end{proposition} \begin{proof} Let $K\geqslant 1$ be the smallest natural number such that $\beta(b)\vert b^K$. We claim that all the members of the arithmetic progression $$ A = \frac{b^K}{\beta(b)}[0,(b-1)\beta(b)-1]$$ exclude the digit $b-1$ from their base-$b$ expansions. To see this, let $k$ satisfy $0\leqslant k\leqslant K-1$. Then $\gcd(b^{k+1}\beta(b),b^{K})\geqslant b^{k+1}>b^k\beta(b)$, which implies that $\gcd(b^{k+1},b^K/\beta(b))>b^k$ (by dividing through by $\beta(b)$). In particular, for all integers $x$ and $y$, either \begin{equation} \label{separation condition} \left\vert x \frac{b^K}{\beta(b)} - y b^{k+1}\right\vert>b^k \qquad \text{or} \qquad x \frac{b^K}{\beta(b)} = y b^{k+1}. \end{equation} This observation implies that none of the $K$ rightmost digits of any integer of the form $xb^K/\beta(b)$ can be equal to $b-1$. Indeed, in base $b$, the $b^{k}$ digit of $x b^K/\beta(b)$ is the unique integer $d$ in the range $0\leqslant d\leqslant b-1$ such that \[\left\{\frac{xb^K/\beta(b)}{b^{k+1}} \right\}\in \left[\frac{d}{b},\frac{d+1}{b}\right).\] Yet~\eqref{separation condition} implies that $\{\frac{x b^K}{b^{k+1}\beta(b)}\} \in \{0\} \cup (\frac{1}{b}, \frac{b-1}{b})$ for each $0 \leqslant k \leqslant K-1$. Since this is disjoint from $[\frac{b-1}{b},1)$, we conclude that none of the $K$ rightmost digits of any integer of the form $x b^K/\beta(b)$ can be equal to $b-1$. We now fix $x \in [0,(b-1)\beta(b)-1]$ and consider the leftmost digits of $x b^K/\beta(b)$. Certainly $x b^K/\beta(b) < (b-1)b^K$. From this upper bound we see that the $b^{K}$ digit of $x b^K/\beta(b)$ lies in $[0,b-2]$ and that the digits associated to larger powers of $b$ are all $0$. Combining this with our previous observations, we conclude that $xb^K/\beta(b)$ omits the digit $(b-1)$ for all $x \in [0,(b-1) \beta(b)-1]$, so $A \subset \mathcal{K}([0,b-2],b)$ as claimed. Since $\vert A\vert = (b-1)\beta(b)$, we have $\ell(b) \geqslant (b-1)\beta(b)$. \end{proof} We now proceed with the second half of our evaluation of $\ell(b)$, the verification that this lower bound is exact. This requires a more technical argument. \begin{proposition} \label{prop:upper_bound} For all $b\geqslant 2$, we have $\ell(b) \leqslant (b-1)\beta(b)$. \end{proposition} \begin{proof} Without loss of generality, let $S\subset [0,b-1]$ be any set of $b-1$ digits (containing $0$), and let $A = \{x+ j \Delta : j \in [0,\ell(b)-1]\}$ be an arithmetic progression in $\mathcal{K}(S,b)$ of maximal length, in which $\Delta > 0$ is taken minimally over all arithmetic progressions of length $\ell(b)$. Let $\Delta = d_K b^K + \ldots + d_1 b + d_0$ denote the base $b$ expansion of $\Delta$, where $K$ is chosen such that $d_K \neq 0$. For notational convenience, let $\Delta_k:= d_k b^k + \ldots +d_1b + d_0$ for each $k \geqslant 0$. (Note that $\Delta_k = \Delta$ for $k \geqslant K$.) We may assume without loss of generality that $d_0 \neq 0$, else by removing the rightmost digit from all elements of $A$ one constructs an arithmetic progression contained in $\mathcal{K}(S,b)$ of common difference $\Delta/b$, contradicting our minimality assumption on $\Delta$. (This is the same device as we used in the proof of Proposition \ref{prime prop}). Our proof of Proposition~\ref{prop:upper_bound} rests on the following claim, whose peculiar statement arises naturally from an inductive argument. \begin{claim} Consider the following statements: \begin{itemize} \item[\emph{C}$1$:] $\ell(b)\leqslant (b-1)\beta(b)$; \item[\emph{C}$2$\emph{(}k\emph{)}:] there exist coprime integers $\lambda_k,\mu_k \in [1,b-1]$ satisfying\\ $\lambda_k \Delta_k = \mu_k b^{k+1}$. \end{itemize} Then either \emph{C1} holds or \emph{C2(}k\emph{)} holds for all $k \geqslant 0$. \end{claim} This claim immediately settles the theorem, since the statement C2($k$) cannot possibly hold for all $k\geqslant 0$. Indeed, we have $\lambda_k \Delta_k < b \Delta$, while $\mu_k b^{k+1}$ grows in $k$ without bound. \end{proof} \begin{proof}[Proof of Claim] We prove this claim by induction, showing that for every $k\geqslant 0$, either C1 holds or C2($k^\prime$) holds for all $k^\prime\leqslant k$. For the base case $k=0$, note that $\Delta_k = d_0$. If $(d_0,b)=1$, then $d_0$ generates the additive group $\mathbb{Z}/b \mathbb{Z}$ and the elements $\{x + j \Delta : j \in [0,b-1]\}$ have $b$ distinct units digits. Thus $\ell(b) \leqslant (b-1) \leqslant (b-1)\beta(b)$, so C1 holds. Otherwise, $(d_0,b)>1$, which implies that there exists $\lambda \in [1,b-1]$ for which $\lambda d_0 \equiv 0 \!\! \mod b$. Thus $\lambda d_0 = \mu b$ for some $\mu$, and we may assume that $(\lambda,\mu)=1$ by dividing through by common factors. This concludes the base case. Proceeding to the inductive step, let $k\geqslant 1$ and assume that the inductive hypothesis C2($k^\prime$) holds for all smaller $k^\prime$. In particular, $\Delta_{k-1} = (\mu_{k-1}/\lambda_{k-1}) b^k$ for some coprime integers $\lambda_{k-1},\mu_{k-1}\in [1,b-1]$, and hence $\Delta_k = d_k b^k + (\mu_{k-1}/\lambda_{k-1}) b^k$. Let $\lambda_k$ denote the order of $\Delta_k/b^{k+1}$ in the additive group $\mathbb{R}/\mathbb{Z}$, and let $\mu_k$ denote the integer $\lambda_k (\Delta_k/b^{k+1})$. We see that $(\lambda_k,\mu_k)=1$, as one could divide through by any common factors of $\lambda_k$ and $\mu_k$ to contradict the fact that $\lambda_k$ is the order of $\Delta_k/b^{k+1}$ in $\mathbb{R}/\mathbb{Z}$. Now, if $\lambda_k < b$, then $\mu_k < b$ as well, since $\Delta_k/b_{k+1}<1$ for any $k$. In this case, $\lambda_k$ and $\mu_k$ satisfy the conditions listed in C2($k$). Therefore C2($k^\prime$) holds for all $k^\prime \leqslant k$. It remains to address the case $\lambda_k \geqslant b$. By usual facts about finite subgroups of $\mathbb{R}/\mathbb{Z}$, we note that the orbit of $\Delta_k/b^{k+1}$ in $\mathbb{R}/\mathbb{Z}$ is exactly the set of fractions with denominator dividing $\lambda_k$. In particular, the set of values \[T=\left\{\frac{x}{b^{k+1}} + \frac{\Delta_k j}{b^{k+1}} \!\! \mod 1 : j \in [0,\lambda_k -1]\right\}\] are equally spaced, with gaps of size $1/\lambda_k$. Since $\lambda_k \geqslant b$, for any integer $d \in [0,b-1]$ at least one member of $T$ lies in the half-open interval $[\frac{d}{b},\frac{d+1}{b})$. In other words, at least one member of the progression $x + \Delta_k[0,\lambda_k-1]$ has $b^{k}$ digit equal to $d$. This information immediately implies that $x+\Delta[0,\lambda_k-1]$ is not contained in any proper Kempner set $\mathcal{K}(S,b)$, and hence $\ell(b) \leqslant \lambda_k -1$. However, more can be said with a slight refinement to our analysis. Equal spacing implies that at least $\lfloor \lambda_k/b \rfloor$ members of $T$ lie in the interval $[\frac{d}{b},\frac{d+1}{b})$. We are left with the stronger bound $\ell(b) \leqslant \lambda_k - \lfloor \lambda_k/b\rfloor$. We now establish an upper bound on the function $\lambda_k - \lfloor \lambda_k/b\rfloor$, given the known constraints on $\lambda_k$. For starters, the inductive hypothesis implies that $\lambda_{k-1} \mid \mu_{k-1} b^{k}$, hence $\lambda_{k-1} \mid b^{k}$ (since $\lambda_{k-1}$ and $\mu_{k-1}$ are coprime). Since $\lambda_{k-1} < b$ and $\lambda_{k-1}$ divides a power of $b$, this implies that $\lambda_{k-1} \leqslant \beta(b)$. Secondly, the inductive hypothesis allows us to write \[\frac{\Delta_k}{b^{k+1}} = \frac{d_k \lambda_{k-1} + \mu_{k-1}}{b \lambda_{k-1}},\] which implies that $b\lambda_{k-1} (\Delta_k/b^{k+1}) \equiv 0 \!\! \mod 1$. This implies that $b \lambda_{k-1}$ is a multiple of the order of $(\Delta_k/b^{k+1}) \!\!\mod 1$, ie. $\lambda_k \mid b \lambda_{k-1}$. We conclude that $\lambda_k \leqslant b \lambda_{k-1} \leqslant b\beta(b)$. The function $\lambda \mapsto \lambda - \lfloor \lambda/b \rfloor$ is non-decreasing as $\lambda$ increases over integers, hence \[\ell(b) \leqslant \lambda_k - \left\lfloor \frac{\lambda_k}{b}\right\rfloor \leqslant b \beta(b) - \left\lfloor \frac{b \beta(b)}{b}\right\rfloor =b \beta(b) - \beta(b)= (b-1)\beta(b),\] which implies that C1 holds. This completes the inductive step, and so completes the proof of Theorem \ref{thm:main theorem}. \end{proof} \section{Asymptotic Analysis} In this section we analyse the function $\beta(b)$, with the ultimate goal of proving Theorem \ref{thm: ell nearly always maximal epsilon free}. We begin with the following simple observation. \begin{proposition} \label{prop:betabounds1} We have \[\liminf_{n \to \infty} \frac{\beta(n)}{n} =0 \quad \text{and} \quad \limsup_{n \to \infty} \frac{\beta(n)}{n} =1.\] \end{proposition} \begin{proof} The first claim follows from the observation that $\beta(p) = 1$ for all primes $p$. For the second, we note that $\beta(2^k+2)=2^k$ for all $k>1$. \end{proof} It is clear from this proposition that the behaviour of $\beta(n)$ is erratic as $n$ varies. However, its calculation may be understood as a certain integer programming problem, as illustrated by the following example. \begin{example} \label{example:beta24} In this example, we calculate $\beta(24)$ using techniques from mixed integer programming. We may write $\beta(24)=2^a\cdot 3^b$, with $a,b \in \mathbb{N}$. It follows that $a \log 2 + b \log 3 < \log 24$, and $(a,b)$ may be visualized as a lattice point in the following figure (Figure \ref{fig1}). The equation of the line is $f(x) = \log_3 24-x\log_3 2 $. \begin{figure}[h] \centering \includegraphics[width=0.8\textwidth]{plot1} \caption{Lattice points $(a,b)$ corresponding to $\beta(24)=2^a\cdot3^b$.} \label{fig1} \end{figure} Let us restrict our attention to the set $S$ of lattice points of the form $(a,b)$, in which $b$ is taken maximally for fixed $a$. If $(a,b) \in S$, the vertical distance from $(a,b)$ to the diagonal in Figure \ref{fig1} is then given by $\{\log_3 24 - a\log_3 2\}$. We also note that $a\log 2 + b \log 3$ is maximized among the lattice points below the line when $(a,b) \in S$ and $\{\log_3 24 - a \log_3 2\}$ is minimized (as a function of $a$). In our example, minimization occurs at $(a,b)=(1,2)$, and so we obtain $\beta(24)=2^1 \cdot 3^2=18$. \end{example} The technique of Example \ref{example:beta24} generalizes easily: if $n$ has $k$ prime divisors $p_1,\ldots,p_k$, we may associate to $n$ a set of lattice points in $\mathbb{Z}^k$, namely \[ \{ (a_1,\cdots,a_k) \in \mathbb{Z}_{\geqslant 0}^k : a_1\log p_1 + \cdots + a_k \log p_k < \log n\}.\] The lattice point $(a_1,\ldots, a_k)$ that minimizes distance to the the hyperplane \[x_1 \log p_1 + \ldots x_k \log p_k = \log n\] determines $\beta(n)$ by the formula $\beta(n)= \prod_{i=1}^k p_i^{a_i}$. \\ Combining this idea with well-known equidistribution results gives the following. \begin{lemma} \label{thm:beta_asymp_along_AP} We have $\beta(n) \sim n$ as $n \to \infty$ within $N\mathbb{Z}$ if and only if $N$ is not a prime power. \end{lemma} \begin{proof} If $N=p^k$ is a prime power, then $N\mathbb{Z}$ contains the subsequence $\{p^{kn}\}_{n \geqslant 1}$. Since $\beta(p^{kn})=p^{kn-1}$, we cannot have $\beta(n) \sim n$ within $N\mathbb{Z}$. Otherwise, let $p$ and $q$ be distinct primes dividing $N$, and fix a positive constant $\varepsilon$. As $\log_q p$ is irrational, the sequence $(\{u_n\})_{n=0}^\infty$ given by $u_n := n\cdot\log_q p$ is equidistributed mod $1$ (by the Equidistribution Theorem: see Proposition 21.1 of \cite{IwKo04}, say). In particular, there exists a positive parameter $L_\varepsilon$ such that $l>L_\varepsilon$ implies that the sequence $(\{ u_n\})_{n=0}^l$ contains at least one element in each interval mod $1$ of length $\varepsilon$. Now let $m$ be a natural number and let $l = \lfloor \log_p m \rfloor$. From the above remarks, there exists a positive parameter $M_\varepsilon$ such that, for each $m>M_\varepsilon$, the shifted sequence $(\{\log_q m - u_n \})_{n=0}^l$ contains some element in the interval $(0,\varepsilon)$. In other words there exists $n_0$ at most $l$ (but dependent on $l$) such that \[0 < \{\log_q m - n_0 \cdot \log_q p\}<\varepsilon.\] Also note that $\log_q m - n_0 \cdot \log_q p $ is positive. Now, assume $pq \mid m$ and consider $(a,b):=(n_0,\lfloor \log_q m - n_0 \cdot \log_q p \rfloor )$. We have $\beta(m) \geqslant p^a \cdot q^b$ by construction. So \[\beta(m)\geqslant p^a \cdot q^b = q^{\log_q m - \{\log_q m - n_0 \cdot \log_q p \}} > q^{\log_q m - \varepsilon}=m \cdot q^{-\varepsilon}.\] Thus $q^{-\varepsilon} < \beta(m)/m < 1$, for all $m$ satisfying $m>M_\varepsilon$ and $pq \mid m$. Since $\varepsilon$ was arbitrary, and $q$ fixed, it follows that $\beta(m) \sim m$ within $pq\mathbb{Z}$, and hence within $N\mathbb{Z}$. \end{proof} By considering $N=6$, for example, we obtain a set of density $1/6$ (namely, $6\mathbb{Z}$) on which $\ell(b) \sim b^2$ as $b$ tends to infinity within that set. Any finite union of such sets $N_i\mathbb{Z}$, where $N_i$ has two distinct prime factors $p_i$ and $q_i$, will also have this property, and one may show with relative ease that such a union may be arranged to have natural density arbitrarily close to $1$. However, by quantifying estimates made in the previous lemma, we can do slightly better, and show the existence of a set with the desired property that has density $1$, thereby proving Theorem \ref{thm: ell nearly always maximal epsilon free}. \begin{proof}[Proof of Theorem \ref{thm: ell nearly always maximal epsilon free}] Let $f(N)$ be a function that satisfies $f(N) \to \infty$ as $N \to \infty$ (to be further specified later). For integers $j \geqslant 0$, let $D_j$ denote the set of $n \in (2^{j-1},2^j]$ such that $n$ has at least two distinct prime factors $p,q \leqslant f(2^{j-1})$. Let \[D:= \bigcup_{j\geqslant 0} D_j.\] The set $D$ is our candidate set for use in Theorem \ref{thm: ell nearly always maximal epsilon free}. \begin{lemma} If $f$ grows slowly enough, the set $D$ has natural density $1$. \end{lemma} \begin{proof} We begin by fixing $j \geqslant 0$ and bounding the size of $D_j$ from below. For convenience, we write $N$ for $2^{j-1}$. To produce this lower bound, we find an upper bound for $(N,2N]\setminus D_j$. Indeed, by a standard application of a small sieve (e.g. the Selberg sieve, in particular Theorem 9.3.10 of \cite{Mu08}), one may show that the number of $n \in (N,2N]$ without any prime factor $p $ less than $f(N)$ is \[O\bigg( N \prod_{p < f(N)} \left(1-\frac{1}{p} \right)\bigg),\] provided $f(N)$ grows slowly enough. By Mertens' Third Theorem, this quantity is $O(N/\log f(N))$. By using a union bound and the sieve above, we bound the number of $n \in (N,2N]$ with exactly one prime factor $p < f(N)$ by \[O\Bigg( \sum_{p < f(N)} \frac{N}{p} \prod_{\substack{q<f(N) \\q \neq p}} \left(1-\frac{1}{q}\right)\!\Bigg).\] This quantity is $O(N \log\log f(N)/\log f(N))$ (again by Mertens' theorems), and we conclude by exclusion that \[\lvert D_j \rvert = N\left(1 - O \left(\frac{\log \log f(N)}{\log f(N)}\right)\right).\] This already establishes that $D$ has full upper Banach density. To show that $D$ has \emph{natural} density $1$, we fix $\varepsilon > 0$ and note that, since $f(N) \to \infty$ as $j \to \infty$, there exists $j_0(\varepsilon)$ such that $\lvert D_j \rvert \geqslant 2^{j-1}(1-\varepsilon)$ for all $j \geqslant j_0(\varepsilon)$. In particular, \begin{align*} \sum_{\substack{n \leqslant X \\ n \in D}} 1 &\geqslant \sum_{j_0(\varepsilon) \leqslant j \leqslant \lceil \log_2 X \rceil} \vert D_j \vert - \sum_{X<n \leqslant 2^{\lceil \log_2 X \rceil}} 1. \\ &\geqslant (1-\varepsilon) \left(2^{\lceil \log_2 X \rceil} - 2^{j_0(\varepsilon)-1}\right) + X - 2^{\lceil \log_2 X \rceil}. \end{align*} Simplifying, we see that \[ \liminf_{X \to \infty} \frac{\vert D \cap [1,X]\vert}{X} \geqslant \liminf_{X \to \infty} \frac{X - \varepsilon 2^{\lceil \log_2 X \rceil} - 2^{j_0(\varepsilon)}}{X} \geqslant 1 -2\varepsilon,\] which implies that $D$ has natural density $1$, since $\varepsilon$ was arbitrary. \end{proof} Secondly, we prove that $\beta(n)$ is asymptotically large within $D$. \begin{lemma} \label{second_lemma} If $f$ grows slowly enough, then $\beta(n) \sim n$ as $n \to \infty$ within $D$. \end{lemma} \begin{proof} Our proof presents a more quantitative adaptation of the argument used in Lemma~\ref{thm:beta_asymp_along_AP}. Let $\varepsilon > 0$, and fix $n \in D$. By the definition of $D$, there exist distinct primes $p,q < f(n)$ for which $p,q \mid n$. We will show that, provided $n$ is large enough in terms of $\varepsilon$, there exist non-negative integers $a$ and $b$ for which \[e^{\varepsilon} \geqslant \frac{n}{p^a q^b} > 1.\] Since $p^a q^b \leqslant \beta(n) < n$, and $\varepsilon$ is arbitrary, this will complete the proof. Taking logarithms, it suffices to find non-negative integers $a$ and $b$ for which \[\frac{\varepsilon}{\log q} \geqslant \log_q n - a \log_q p - b > 0.\] Setting $L = \lfloor \log_p n \rfloor$, it will be enough to prove that the sequence of fractional parts $\{ \{a \log_q p\} : a \in [1,L]\}$ contains an element in every interval modulo $1$ of length $\varepsilon/\log q$. Since $p,q \leqslant f(n)$, we reduce our theorem to the following claim: \begin{claim} Let $L'=\lfloor \log n/\log f(n) \rfloor$. Then $S=\{ \{a \log_q p\} : a \in [1,L']\}$ contains an element in every interval modulo $1$ of length $\varepsilon/\log f(n)$, provided $f(n)$ grows slowly enough. \end{claim} The proof of this claim follows from the Erd\H{o}s-Tur\'an inequality (Corollary 1.1 of \cite{Mo94}). Indeed, for any interval $I$ modulo $1$ of length $\varepsilon/\log f(n)$, we have \begin{align} \label{eq:erdos_turan_bound} \left\vert \vert S\cap I\vert - \frac{\varepsilon L'}{\log f(n)} \right\vert \ll \frac{L'}{K+1} + \sum_{k \leqslant K} \frac{1}{k} \bigg\vert \sum_{a=1}^{L'} e^{2\pi i a k\log_q p} \bigg\vert \end{align} for any integer $K \geqslant 1$. It suffices to show that we may choose a $K$ such that the right-hand side in~\eqref{eq:erdos_turan_bound} is $o(L'/\log f(n))$ as $n \to \infty$. Choosing $K = \lfloor \log^2 f(n)\rfloor $ ensures that $L'/(K+1) = o(L'/\log f(n))$. As for the second term in~\eqref{eq:erdos_turan_bound}, bounding the sum over $a$ as a geometric series gives \[\sum_{k \leqslant K} \frac{1}{k} \bigg\vert \sum_{a=1}^{L'} e^{2\pi i a k\log_q p} \bigg\vert \leqslant G(K, p,q)\] for some function $G$ that is independent of $L^\prime$. We may assume without loss of generality that $G$ is increasing in each variable. Then $$G(K,p,q) \ll G( \log^2 f(n), f(n), f(n) ),$$ so it suffices to show that \begin{equation} \label{growth function} G\left(\log^2 f(n), f(n), f(n)\right) = o\left(\frac{L'}{\log f(n)}\right). \end{equation} Recalling the definition of $L'$, this is equivalent to showing \[ G\left(\log^2 f(n), f(n), f(n)\right) \cdot \log^2 f(n)= o\left(\log n \right).\] Yet $G$ is simply some absolute function, so if $f$ grows slowly enough then (\ref{growth function}) will hold. (If one so wished, one could quantify this growth condition using Baker's result \cite{Ba68} on linear forms of logarithms of primes). This proves the claim, and hence the lemma. \end{proof} Combining Lemma~\ref{second_lemma} with Theorem~\ref{thm:main theorem} yields Theorem \ref{thm: ell nearly always maximal epsilon free}. \end{proof} \vspace{5 mm} \bibliographystyle{plain}

8bbd6cebae034e0a72f171a1e39fd3c02eeca894

\section{Introduction} The Ising model, invented almost a century ago, is one of the most studied model of statistical mechanics. A fundamental tool to analyze the Ising model is the random cluster model, or FK model, invented by Fortuin and Kasteleyn around 1969 (see the reference book \cite{Grimmett}). Results on this dependent percolation model can be transferred towards the Ising model via a coupling construction due to Edwards and Sokal \cite{Sokal}. This machinery works extremely well, but so far, it has essentially been employed to study the systems at equilibrium. Yet another rich facet of the Ising model is the dynamics. There exist several microscopic dynamics on spin configurations, which give rise to the so--called stochastic Ising model, and whose equilibrium are described by the Ising Gibbs measure: the Metropolis dynamics, the heat--bath dynamics, the Kawasaki dynamics, to name a few of them. On one hand, these dynamics provide a basic model to study fundamental questions on dynamics, for instance to model the metastability phenomenon \cite{OS}. On the other hand, there is a hope that the understanding of the dynamics will shed light on the equilibrium measure. Percolation models possess also a natural dynamics. For the Bernoulli percolation model, it consists in updating independently the edges, and it leads to beautiful difficult problems \cite{Olle,Steif}. For the FK model, the construction of a dynamics is more subtle and it involves typically non--local computations \cite{SCM}. One naturally wonders whether the stochastic Ising model and dynamical percolation could help to understand each other. Our note is a little investigation into the possibility of building a coupling for the dynamics on Ising and percolation models, which could help to understand the dynamics of the Ising model. We focus here on the case of a Glauber type dynamics, that is a local dynamics which modifies at most one spin at a time. The precise definition of such dynamics, together with several examples, are given in section~\ref{glauber}. We dream of building a coupling dynamics on edge--spins configurations such that: \smallskip \noindent $\bullet$ The marginal on the spins is a Glauber dynamics. \noindent $\bullet$ The marginal on the edges is a simple Markovian evolution. \noindent $\bullet$ The coupling dynamics on edge--spins configurations is reversible with respect to the coupling measure between the Ising and the FK models. \smallskip \noindent Notice that if we drop the first constraint, then we could simply consider a time continuous Markov process on the edge--spins configurations which, after an exponential time of parameter one, jumps on a new independent edge--spins configuration, drawn according to the Ising--FK coupling measure. However the first constraint is essential for us, indeed our hope is to build a dynamical coupling which would help to study a Glauber dynamics of the Ising model. The definition of the Ising--FK coupling measure is recalled in section~\ref{IFK}. The stochastic Ising model is defined in section~\ref{glauber}. The FK dynamics is defined in section~\ref{fkdyna}. The above conditions are precisely stated in section~\ref{goal}. Unfortunately, we did not succeed in building a dynamics which fulfill these three conditions, even if we weaken partially the last one. However, we manage to build couplings which satisfy the third condition. The most basic coupling is a dynamics which changes at most one object at each step, i.e., either one spin or one edge is modified at a time. We prove that, for this type of coupling, the first condition can never be satisfied. We build such a coupling dynamics in section~\ref{one}. Another natural dynamics consists in changing simultaneously the spin at one vertex, together with the edges incident to this vertex. Such a dynamics is presented in section~\ref{two}. The good news is that the marginal of this dynamics on the spins is a Glauber dynamics. However the marginal of this type of dynamics on the edges is not Markovian. In section~\ref{general}, we show that our dream process is not feasible in general. We present a third dynamics in section~\ref{three}, which is non local and fulfills the first and the third conditions. We finally present in section~\ref{four} a localized version of this third dynamics, which can be seen as a contraction of the first dynamics. \section{Ising--FK coupling measure} \label{IFK} In this section, we recall some classical notation and we define the Ising--FK coupling measure~$IP$. We consider a finite graph $(V,E)$, where $V$ is the set of the vertices and $E$ the set of the edges. The edges are unoriented, and $E$ is a subset of the set of pairs of points of $V$. Throughout the paper, we suppose that the set of edges $E$ is not empty. An edge configuration $\eta$ is an element of $\{0,1\}^E$ (where $0$ stands for closed and $1$ for open). A spin configuration $\sigma$ is an element of $\{-1,1\}^V$. An edge $e$ with endvertices $x$ and $y$ is written as $e=\langle x,y\rangle$ or $e=\langle y,x\rangle$. We define $$\forall\sigma\in \{-1,1\}^V \quad\forall e=\langle x,y\rangle\in E\qquad \delta_{\sigma}(e)=\BBone_{\sigma(x)=\sigma(y)} \,.$$ An edge configuration $\eta$ and a spin configuration $\sigma$ are said to be compatible if we have $$\forall e\in E\qquad \eta(e) \leq \delta_{\sigma}(e) \,,$$ i.e., if the endvertices of any open edge in $\eta$ have the same spins in $\sigma$. We denote by $\cC$ the set of the pairs of compatible configurations: $$ \cC=\big\{(\eta,\sigma)\in \{0,1\}^E\times \{-1,1\}^V: \eta(e) \leq \delta_{\sigma}(e)\,\,\,\text{for any}\,\,\, e\in E\big\}. $$ Let $p$ be a fixed number in $[0,1]$. The Ising-Percolation measure $IP$ is the probability measure on the product space $\{0,1\}^E\times \{-1,+1\}^V$ defined as follows. For any $(\eta,\sigma)$ in $\{0,1\}^E\times\{-1,+1\}^V$, we set $$ {IP}(\eta,\sigma)= \frac{1}{Z} \prod_{e\in E}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0} \right), $$ where $Z$ is the partition function, i.e., the normalising constant that makes ${IP}$ a probability measure on $\{0,1\}^E\times\{-1,+1\}^V$, given by $$ Z=\sum_{(\eta,\sigma)\in \{0,1\}^E\times\{-1,+1\}^V}\prod_{e\in E}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0} \right). $$ The Ising-Percolation measure can also be written as \begin{equation}\label{ipsimply} {IP}(\eta,\sigma)= \frac{1}{Z} p^{|\{e\in E:\, {\eta(e)=1}\}|} (1-p)^{|\{e\in E:\, {\eta(e)=0}\}|}\BBone_{(\eta,\sigma)\in \cC}, \end{equation} where the absolute value of a set denotes its cardinality. In fact, the measure $IP$ is the Bernoulli product measure on edge-spin configurations with parameters $(p,1/2)$ conditioned to the set $\cC$ of the compatible configurations. It is well--known that the probability measure ${IP}$ is a coupling of the Ising measure $\mu_{\beta}$ defined on $\{-1,+1\}^V$, together with the random-cluster measure $\phi_{p,2}$ defined on $\{0,1\}^E$, with the relation $p=1-e^{-\beta}$. In fact, the first marginal measure $\phi_{p,2}$ on $\{0,1\}^E$ is given by \begin{equation}\label{tramsf} \forall\,\,\eta\in \{0,1\}^E\qquad \phi_{p,2}(\eta)=\sum_{\sigma \in \{-1,+1\}^V}{IP}(\eta,\sigma)= \frac{1}{Z_{RC}} \left(\prod_{e\in E} p^{\eta(e)}(1-p)^{1-\eta(e)}\right)2^{k(\eta)}. \end{equation} Here $k(\eta)$ denotes the number of connected components (or open clusters) of the graph having for vertices $V$ and for edges the open edges in the configuration $\eta$ (recall that an edge $e$ is said to be open in $\eta$ if $\eta(e)=1$ and closed if $\eta(e)=0$). Naturally, $Z_{RC}$ is the partition function that makes $ \phi_{p,2}$ a probability measure on $\{0,1\}^E$. The second marginal measure is given by \begin{equation}\label{transf} \forall\,\, \sigma\in \{-1,+1\}^V\qquad \sum_{\eta \in \{0,1\}^E}{IP}(\eta,\sigma)= \frac{1}{Z} (1-p)^{|\{\,e\in E:\delta_\sigma(e)=0\,\}|}\,. \end{equation} Notice that $$|\{\,e\in E:\delta_\sigma(e)=0\,\}|\,=\, \sum_{e\in E}\big(1-\delta_{\sigma}(e)\big)\,=\, |E|- \sum_{e\in E}\delta_{\sigma}(e) \,.$$ Let $\beta>0$ be such that $p=1-e^{-\beta}$. Rewriting the formula~\eqref{transf} with these notations, we see that the second marginal is the measure $\mu_{\beta}$ on $\{-1,+1\}^V$ given by \begin{equation}\label{mubeta} \forall\,\, \sigma\in \{-1,+1\}^V\qquad \mu_{\beta}(\sigma) \,=\, \frac{1}{Z_{I}} \exp\left({\beta} \sum_{e\in E}\delta_{\sigma}(e)\right) ,\,\,\, \end{equation} where the partition function $Z_I$ is $$Z_I\,=\,Z\,\exp(\beta|E|) \,.$$ Let us rewrite the Hamiltonian in a more classical way. For an edge $e$ with endpoints $x,y$, we have \begin{equation}\label{reta} \delta_\sigma(e)\,=\,\frac{1}{2}\big(1+\sigma(x)\sigma(y)\big)\,. \end{equation} Summing over $e\in E$, we get $$ \sum_{e\in E}\delta_{\sigma}(e) \,=\, \sum_{\genfrac{}{}{0pt}{1}{\scriptstyle x,y\in V}{\scriptstyle \{x,y\}\in E}} \frac{1}{2}\big(1+\sigma(x)\sigma(y)\big) \,=\, \frac{1}{2}|E|+ \frac{1}{2} \sum_{\genfrac{}{}{0pt}{1}{\scriptstyle x,y\in V}{\scriptstyle \{x,y\}\in E}} \sigma(x)\sigma(y) \,.$$ The standard Hamiltonian of the Ising model is defined as \begin{equation}\label{hcbeta} \forall \sigma\in \{-1,+1\}^V\qquad H(\sigma)\,=\, -\frac{1}{2} \sum_{\genfrac{}{}{0pt}{1}{\scriptstyle x,y\in V}{\scriptstyle \{x,y\}\in E}} \sigma(x)\sigma(y)\,. \end{equation} In the end, we have \begin{equation}\label{mucbeta} \forall\,\, \sigma\in \{-1,+1\}^V\qquad \mu_{\beta}(\sigma) \,=\, \frac{1}{Z_{\beta}} \exp\left(-{\beta} H(\sigma)\right) ,\,\,\, \end{equation} where the partition function $Z_\beta$ is given by $$Z_\beta \,=\, \sum_{\sigma\in \{-1,+1\}^V} \exp\left(-{\beta} H(\sigma)\right) \,=\,Z_I\,\exp\Big(-\frac{1}{2}\beta|E|\Big) \,.$$ We refer to the book of Grimmett \cite{Grimmett} and the references therein for more details about the coupling measure $IP$ between random-cluster and Ising measures, its history and usefulness. We refer to the paper of Schonmann \cite{Sch} for a nice presentation of the Gibbs measure of the Ising model and the associated dynamics. \section{The stochastic Ising model} \label{glauber} In this section, we define Glauber type dynamics for the Ising model. Each of these dynamics defines a Markov process $(\sigma_t)_{t\geq 0}$ on the spin configurations which is called a stochastic Ising model. A Markov process $(\sigma_t)_{t\geq 0}$ on the spin configurations is classically defined through its infinitesimal generator $L$ (see \cite{liggett}). This generator $L$ acts on functions $f$ of the spin configuration and it is of the following form. For any function $f:\{-1,+1\}^V\rightarrow \BBr$, we have $$ \forall\,\, \sigma\in \{-1,+1\}^V\qquad Lf(\sigma)= \sum_{\sigma'\in\{-1,+1\}^V}c(\sigma,\sigma')(f(\sigma')-f(\sigma))\,. $$ The quantity $c(\sigma,\sigma')$ is the rate at which the configuration $\sigma$ is transformed into $\sigma'$ when the system is in the state $\sigma$. We are usually interested in local dynamics which modify at most one spin at a time. We call this type of dynamics Glauber type dynamics. For a configuration $\sigma$ and a vertex $x$, we denote by $\sigma^x$ the configuration obtained from $\sigma$ by flipping the spin at the site $x$. So we require that $c(\sigma,\sigma')$ vanishes when $\sigma$ and $\sigma'$ differ in more than one spin and the quantity $c(\sigma,\sigma^x)$, sometimes denoted by $c(x,\sigma)$, is the rate at which the spin at the site $x$ flips when the system is in the state $\sigma$. The transition rates $c(\cdot,\cdot)$ are defined on $\{-1,1\}^V\times \{-1,1\}^V$ and have thus the following properties: for any $\sigma\in \{-1,+1\}^V$ and $x\in V$, \begin{eqnarray}\label{tauxglauber} && {c}(\sigma,\sigma') = 0\,\,\,{\mbox{if}}\,\,\, |\{x\in V,\,\sigma(x)\neq \sigma'(x)\}|\geq 2, {\nonumber}\\ && {c}(\sigma,\sigma^x)>0\,\,{\mbox{for any}}\,\, x\in V, \\ && c(\sigma,\sigma)=-\sum_{y\in V}c(\sigma,\sigma^y). {\nonumber} \end{eqnarray} We say that the transition rates $c(\cdot,\cdot)$ satisfy the detailed balance condition with respect to the Ising measure $\mu_{\beta}$ if \begin{equation}\label{balanceising} \forall\,\sigma\in\{-1,1\}^V\quad \forall\,x\in V\qquad \mu_{\beta}(\sigma)c(\sigma,\sigma^x)=\mu_{\beta}(\sigma^x)c(\sigma^x,\sigma)\,. \end{equation} When this detailed balance condition holds, the associated dynamics is reversible with respect to the Ising measure $\mu_{\beta}$. Several rates satisfy the conditions~(\ref{tauxglauber}) and the detailed balance condition (\ref{balanceising}). These dynamics and their fundamental properties are presented in the paper of Schonmann \cite{Sch}. In the same paper appears the graphical construction of these dynamics, which provides a natural and intuitive picture of the associated Markov processes. Let us present some classical choices. Recalling the definition of the Ising Hamiltonian~\eqref{hcbeta}, we define further, for $\sigma\in \{-1,+1\}^V$ and $x\in V$, \begin{equation}\label{delta} \Delta_xH(\sigma)\,=\, H(\sigma^x)- H(\sigma)\,. \end{equation} For dynamics which modify at most one spin at a time, we set $$c(x,\sigma)= c(\sigma,\sigma^x)\,.$$ Here are some possible choices for the rates: \noindent {\bf The Metropolis dynamics:} $$c(x,\sigma)\,=\, \exp\Big(-\beta\max\big(\Delta_xH(\sigma),0\big)\Big)\,.$$ \noindent {\bf The Heat bath dynamics:} $$c(x,\sigma)\,=\, \frac{1}{1+ \exp\Big(\beta\max\big(\Delta_xH(\sigma),0)\big)\Big)}\,.$$ \noindent {\bf Unnamed dynamics:} $$c(x,\sigma)\,=\, \exp\Big(-\frac{\beta}{2}\Delta_xH(\sigma)\Big)\,.$$ These three choices satisfy the detailed balance condition~\eqref{balanceising} and correspond to one spin flip dynamics. We call them Glauber type dynamics. Let us try to express the rates with the help of the functions $\delta_\sigma$ on the edges. Using the identity~\eqref{reta}, we have, for any $x\in V$, $$ \Delta_xH(\sigma)\,=\, \sum_{y\in V: \{x,y\}\in E} \sigma(x)\sigma(y) \,=\,\sum_{e\in E_x}\big(2\delta_{\sigma}(e)-1\big) \, $$ where the set $E_x$ is the set of the edges $e\in E$ having $x$ as endvertex, i.e., $$ E_x\,=\,\big\{\,e=\langle x,y\rangle\,\in E:\, y\in V\,\big\}\,. $$ For the unnamed dynamics, we obtain \begin{equation*} c(x,\sigma)\,=\, \exp\left(- \beta \sum_{e\in E_x}\delta_{\sigma}(e) +\frac{\beta}{2}|E_x| \right) \,. \end{equation*} On the lattice $\mathbb{Z}^d$, we have $|E_x|=2d$ for any $x\in\mathbb{Z}^d$. Up to a constant multiplicative factor, the previous rates are equal to \begin{equation}\label{GD} c(x,\sigma)\,=\, \exp\left(- \beta \sum_{e\in E_x}\delta_{\sigma}(e) \right) \,. \end{equation} We will mostly use this choice when trying to build a dynamical coupling between the Ising model and the FK model. \section{FK dynamics} \label{fkdyna} In this section, we define the counterpart of a Glauber type dynamics for the FK model. We are interested in dynamics which modify at most one edge at a time. We build a Markov process on percolation configurations, which is defined through its infinitesimal generator $L$ (see \cite{liggett}). This generator $L$ acts on functions $g$ of the percolation configuration and it is of the following form. For any function $g: \{0,1\}^E \rightarrow \BBr$, we have $$\forall \eta\in \{0,1\}^E \qquad Lg(\eta)\,=\, \sum_{e\in E} c(\eta,\eta^e)(g(\eta^e)-g(\eta))\,,$$ where $\eta^e$ is the configuration obtained from $\eta$ by changing the state of the edge~$e$ and the quantity $c(\eta,\eta^e)$ is the rate at which the edge $e$ changes its state when the system is in the state $\eta$. The transition rates $c(\cdot,\cdot)$ are defined on $\{0,1\}^E \times \{0,1\}^E$ and have the following properties: for any $\eta\in \{0,1\}^E$ and $e\in E$, \begin{eqnarray}\label{tauxpglauber} && {c}(\eta,\eta') = 0\,\,\,{\mbox{if}}\,\,\, |\{e\in E,\,\eta(x)\neq \eta'(e)\}|\geq 2, {\nonumber}\\ && {c}(\eta,\eta^e)>0\,\,{\mbox{for any}}\,\, e\in E, \\ && c(\eta,\eta)=-\sum_{e\in E}c(\eta,\eta^e). {\nonumber} \end{eqnarray} We say that the transition rates $c(\cdot,\cdot)$ satisfy the detailed balance condition with respect to the FK measure $\phi_{p,2}$ if \begin{equation}\label{balancefk} \forall\eta\in \{0,1\}^E \qquad \forall\,e\in E \qquad \phi_{p,2}(\eta) c(\eta,\eta^e)\,=\, \phi_{p,2}(\eta^e) c(\eta^e,\eta)\,. \end{equation} When this detailed balance condition holds, the associated dynamics is reversible with respect to the FK measure $\phi_{p,2}$. Several rates satisfy the conditions~(\ref{tauxpglauber}) and the detailed balance condition (\ref{balancefk}). A natural choice is the following. For an edge $e=\langle x,y\rangle\in E$ and a configuration $\eta\in \{0,1\}^E$, we set $\gamma_{\eta}(e)=1$ if the endpoints $x,y$ of $e$ are connected by a path of open edges in $\eta$ which does not use the edge $e$ itself, and we set $\gamma_{\eta}(e)=0$ otherwise. We define then \begin{equation*} c(e,\eta):=c(\eta,\eta^e) \,=\, \begin{cases} 1-p&\text{ if }\eta(e)=1\,,\cr p&\text{ if }\eta(e)=0\text{ and }\gamma_\eta(e)=1\,,\cr p/2&\text{ if }\eta(e)=0\text{ and }\gamma_\eta(e)=0\,.\cr \end{cases} \end{equation*} One can check that these rates satisfy~\eqref{tauxpglauber} and~ \eqref{balancefk}. \section{Our dream process} \label{goal} The main purpose of this paper is to construct a Markov process on the space $\{0,1\}^E\times \{-1,+1\}^V$ whose marginal on the spins is a Glauber dynamics and which is reversible with respect to the coupling measure ${IP}$. Suppose that $(\eta_t,\sigma_t)_{t\geq 0}$ is such a Markov process and let $q((\eta,\sigma),(\eta',\sigma'))$ be its transition rates. These transition rates satisfy the following usual conditions: \begin{equation}\label{qmatrix1} (\eta,\sigma)\neq (\eta',\sigma')\quad \Longrightarrow \quad q((\eta,\sigma), (\eta',\sigma'))\geq 0, \end{equation} and for any $(\eta,\sigma)\in \{0,1\}^E\times \{-1,+1\}^V$, \begin{equation}\label{qmatrix} \sum_{(\eta',\sigma')\in \{0,1\}^E\times \{-1,+1\}^V}q((\eta,\sigma), (\eta',\sigma'))=0\,. \end{equation} The reversibility property with respect to the coupling measure ${IP}$ is equivalent to the following detailed balance equation: \begin{multline}\label{eqreversible} \forall (\eta,\sigma),(\eta',\sigma')\in\{0,1\}^E\times \{-1,1\}^V\qquad\hfill\cr IP(\eta,\sigma)\,q((\eta,\sigma), (\eta',\sigma'))= IP(\eta',\sigma')\,q((\eta',\sigma'), (\eta,\sigma)). \end{multline} In general, the marginals of a Markov process are not themselves Markov processes. Now, we know from Ball and Yeo \cite{Ball} (see theorem \ref{theoball} of the appendix) that the second marginal $(\sigma_t)_{t\geq 0}$ of the dynamics is a Markov jump process with rates ${c}(\sigma,\sigma')$ if and only if for any $\eta \in \{0,1\}^E$, $\sigma,\sigma' \in \{-1,1\}^V$, the sum $\sum_{\eta'\in \{0,1\}^E} q((\eta,\sigma),(\eta',\sigma'))$ does not depend on $\eta$. We have then \begin{equation}\label{secondmargin} \forall\eta \in \{0,1\}^E\quad\forall\sigma,\sigma' \in \{-1,1\}^V\qquad \sum_{\eta'\in \{0,1\}^E} q((\eta,\sigma),(\eta',\sigma'))= {c}(\sigma,\sigma')\,. \end{equation} Suppose that this is the case. The second marginal $(\sigma_t)_{t\geq 0}$ will then be the Glauber dynamics with the transition rates (\ref{GD}) as soon as the sum in (\ref{secondmargin}) is equal, up to some positive constant, to the transitions rates given in (\ref{GD}). More precisely, we should have, up to some positive multiplicative constant, \begin{equation}\label{marginal} {c}(\sigma,\sigma') = \left\{ \begin{array}{cc} (1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}\,\, & \text{if} \,\, \sigma'=\sigma^x \,,\\ -\sum_{x\in V}(1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}\,\, & \text{if} \,\, \sigma'=\sigma \,,\\ 0 & \text{otherwise,} \end{array} \right. \end{equation} where $E_x$ denotes the set of the edges $e$ in $E$ having $x$ as endvertex. Ideally, we would also wish that the first marginal $(\eta_t)_{t\geq 0}$ is a Markov jump process with rates ${\tilde c}(\eta,\eta')$. This will be the case if and only if, for any $\eta \in \{0,1\}^E$, $\eta' \in \{0,1\}^E$, $\sigma \in \{-1,1\}^V$, the sum $\sum_{\sigma'\in \{-1,1\}^V} q((\eta,\sigma),(\eta',\sigma'))$ does not depend on $\sigma$. We would have then \begin{equation}\label{premieremargin} \forall\eta, \eta' \in \{0,1\}^E\quad \forall\sigma \in \{-1,1\}^V\qquad \sum_{\sigma'\in \{-1,1\}^V} q((\eta,\sigma),(\eta',\sigma'))={\tilde c}(\eta,\eta')\,. \end{equation} So, to sum up, we dream of constructing a Markov process on $\{0,1\}^E\times \{-1,1\}^V$ with transition rates satisfying (\ref{qmatrix1}), (\ref{qmatrix}), (\ref{eqreversible}), (\ref{secondmargin}), (\ref{marginal}) and~(\ref{premieremargin}), that is a Markov jump process $(\eta_t,\sigma_t)_{t\geq 0}$ which is reversible with respect to the $IP$ measure, whose marginal on the spins is a Glauber dynamics, and whose marginal on the edges is Markovian. \section{One change at a time} \label{one} We present here our first try. We build a simple dynamics which updates at most one site or one edge at each time, which is reversible with respect to the coupling measure $IP$ and whose marginal on the spins has the same rates as a Glauber dynamics, although it is not Markovian. Recall that, for $x\in V$, $E_x$ is the set of the edges $e$ in $E$ having $x$ as endvertex, i.e., $$ E_x\,=\,\big\{\,e=\langle x,y\rangle\,\in E:\, y\in V\,\big\}\,. $$ For $\sigma\in \{-1,1\}^V$ and $x\in V$, $\sigma^x$ is the element of $\{-1,1\}^V$ obtained from $\sigma$ by reversing the spin at $x$, i.e., $$ \sigma^x(x)=-\sigma(x)\,,\qquad \forall\, y\in V\setminus \{x\}\quad \sigma^x(y)=\sigma(y)\,. $$ For $\eta\in \{0,1\}^E$ and $e\in E$, we denote by $\eta^e$ the element of $\{0,1\}^E$ obtained from $\eta$ by changing the value of $\eta(e)$, i.e., $$\eta^e(e)=1-\eta(e)\,,\qquad \forall f\in E\setminus\{\,e\,\}\quad \eta'(f)=\eta(f)\,.$$ Before introducing the transition rates of the one change dynamics, we give a condition that the transition rates have to fulfill in order to have the properties announced at the beginning of the section. \begin{lem}\label{lemn} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process, which updates at most one site or one edge at each time and which is reversible with respect to the coupling measure IP. Let us denote by $q((\eta,\sigma), (\eta',\sigma'))$ the transition rates of $(\eta_t,\sigma_t)_{t\geq 0}$. Let $(\eta,\sigma)$ be a pair of compatible configurations. If there exists $x\in V$ such that $q((\eta,\sigma), (\eta,\sigma^x))\neq 0$, then all the edges of $E_x$ are closed in~$\eta$. \end{lem} \begin{proof} The detailed balance equation (\ref{eqreversible}) yields that, for any $\eta\in \{0,1\}$, $\sigma\in \{-1,1\}^V$ and any $x\in V$, $$ IP(\eta,\sigma)q((\eta,\sigma),(\eta,\sigma^x))= IP(\eta,\sigma^x)q((\eta,\sigma^x),(\eta,\sigma)), $$ which gives, thanks to (\ref{ipsimply}), $$ \BBone_{(\eta,\sigma)\in {\cal C}}\,q((\eta,\sigma),(\eta,\sigma^x))= \BBone_{(\eta,\sigma^x)\in {\cal C}}\, q((\eta,\sigma^x),(\eta,\sigma)). $$ Since $(\eta,\sigma)$ are supposed to be compatible, the last equality implies that $$ q((\eta,\sigma),(\eta,\sigma^x))= q((\eta,\sigma^x),(\eta,\sigma))\prod_{e\in E_x}\BBone_{\eta(e)\leq 1-\delta_{\sigma}(e)}\,. $$ Consequently, if $ q((\eta,\sigma),(\eta,\sigma^x))\neq 0$ then necessarily $ \eta(e)\leq 1-\delta_{\sigma}(e)$ for any $e\in E_x$. Since the configurations $\eta$ and $\sigma$ are compatible, necessarily $\eta(e)= 0$ for any $e\in E_x$. \end{proof} \noindent We define next the transition rates of our dynamics. \begin{defi}\label{def0} Let $(\eta,\sigma)$, $(\eta',\sigma')$ be two elements of $\{0,1\}^E\times\{-1,1\}^V$. We consider several cases: \noindent $\bullet$ If $\eta'=\eta$ and there exists $x\in V$ such that $\sigma'=\sigma^x$, $\eta(e)=0$ for any $e\in E_x$, then we define $ c((\eta,\sigma), (\eta',\sigma'))= 1\,. $ \noindent $\bullet$ If $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$, then we set $$c((\eta,\sigma), (\eta',\sigma'))= p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1} \,.$$ \noindent $\bullet$ Otherwise, if $(\eta,\sigma)\neq (\eta',\sigma')$, then we set $c((\eta,\sigma), (\eta',\sigma'))= 0$. \noindent $\bullet$ Finally, if $(\eta,\sigma)= (\eta',\sigma')$, then we set $$ c((\eta,\sigma), (\eta,\sigma))= -\sum_{(\eta',\sigma'),\, (\eta',\sigma')\neq (\eta,\sigma)}c((\eta,\sigma), (\eta',\sigma'))\,. $$ \end{defi} \begin{pro}\label{prorev1} The transition rates introduced in definition \ref{def0} satisfy the detailed balance equation (\ref{eqreversible}). \end{pro} \begin{proof} We check first that, for any $(\eta,\sigma)$ in $\{0,1\}^E\times \{-1,1\}^V$ and $e\in E$, \begin{equation}\label{t2} IP(\eta,\sigma)c((\eta,\sigma), (\eta^e,\sigma))= IP(\eta^e,\sigma)c((\eta^e,\sigma), (\eta,\sigma)) \end{equation} Let $(\eta,\sigma)$ in $\{0,1\}^E\times \{-1,1\}^V$ and $e\in E$ be fixed. We have, since $\eta(e)=1-\eta^e(e)$, \begin{eqnarray}\label{t1} && \left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0} \right)\times\left(p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1} \right){\nonumber}\\ &&= \left(p\BBone_{\eta^e(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta^e(e)=0} \right)\times\left(p\BBone_{\eta^e(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta^e(e)=1} \right). \end{eqnarray} We also have, since $\eta^e=\eta$ on $E\setminus\{e\}$, \begin{eqnarray}\label{nrev} && \prod_{f\in E\setminus\{e\}}\left(p\BBone_{\eta(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta(f)=0} \right){\nonumber}\\ &&= \prod_{f\in E\setminus\{e\}}\left(p\BBone_{\eta^e(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta^e(f)=0} \right). \end{eqnarray} Noting that $$ IP(\eta,\sigma)=\frac{1}{Z}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0} \right)\prod_{f\in E\setminus\{e\}}\left(p\BBone_{\eta(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta(f)=0} \right), $$ we get (\ref{t2}) by multiplying each side of~\eqref{nrev} by~\eqref{t1}. We now prove that, for any $(\eta,\sigma)$ in $\{0,1\}^E\times \{-1,1\}^V$ and any $x\in V$, \begin{equation}\label{t3} IP(\eta,\sigma)c((\eta,\sigma), (\eta,\sigma^x))= IP(\eta,\sigma^x)c((\eta,\sigma^x), (\eta,\sigma)). \end{equation} Noting that $$c((\eta,\sigma), (\eta,\sigma^x))= \prod_{e\in E_x}\BBone_{\eta(e)=0},$$ and that $\delta_{\sigma}(e)=\delta_{\sigma^x}(e)$ for any $e\in E\setminus E_x$, we deduce that \begin{eqnarray*}\label{e1} &&c((\eta,\sigma), (\eta,\sigma^x))IP(\eta,\sigma)= {\nonumber}\\ &&= \frac{1}{Z}\left(\prod_{e\in E_x}\BBone_{\eta(e)=0}\right)\left(\prod_{e\in E}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e) + (1-p)\BBone_{\eta(e)=0}\right)\right){\nonumber}\\ &&=\frac{1}{Z}\left(\prod_{e\in E_x}((1-p)\BBone_{\eta(e)=0})\right)\left(\prod_{e\in E\setminus E_x}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e) + (1-p)\BBone_{\eta(e)=0}\right)\right) {\nonumber}\\ &&= \frac{1}{Z}\left(\prod_{e\in E_x}((1-p)\BBone_{\eta(e)=0})\right)\left(\prod_{e\in E\setminus E_x}\left(p\BBone_{\eta(e)=1}\delta_{\sigma^x}(e) + (1-p)\BBone_{\eta(e)=0}\right)\right) {\nonumber} \\ &&= \frac{1}{Z}\left(\prod_{e\in E_x}\BBone_{\eta(e)=0}\right)\left(\prod_{e\in E}\left(p\BBone_{\eta(e)=1}\delta_{\sigma^x}(e) + (1-p)\BBone_{\eta(e)=0}\right)\right) {\nonumber}\\ &&= c((\eta,\sigma^x), (\eta,\sigma))IP(\eta,\sigma^x). \end{eqnarray*} This finishes the proof of condition~(\ref{t3}). Since the transition rates $c((\eta,\sigma), (\eta',\sigma'))$ given in definition~\ref{def0} allow to perform at most one change at a time, the reversibility condition reduces to the two conditions~(\ref{t2}) and~(\ref{t3}), so the proof of Proposition \ref{prorev1} is complete. \end{proof} \noindent The following proposition shows that, if the initial condition $(\eta_0,\sigma_0)$ is distributed as $IP$, then the infinitesimal behavior of the marginal on the spins of the process $(\eta_t,\sigma_t)_{t\geq 0}$, whose rates are introduced in definition \ref{def0}, is the same as the transition rates given in (\ref{marginal}). \begin{pro}\label{marg} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process with the transition rates introduced in definition \ref{def0}. Suppose that the pair $(\eta_0,\sigma_0)$ is distributed according to $IP$. Then, for any $x\in V$, for any $\sigma\in \{-1,1\}^V$ and any $s\geq 0$, \begin{equation}\label{limite} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma)= (1-p)^{|\{e\in E_x,\,\delta_{\sigma}(e)=1\}|}. \end{equation} \end{pro} \begin{proof} We have, for any $\sigma\in \{-1,1\}^V$, $x\in V$ and $s\geq 0$, \begin{eqnarray}\label{ee1} && \BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) = \sum_{\eta\in \{0,1\}^E}\sum_{\eta'\in \{0,1\}^E}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta',\eta_s=\eta|\sigma_s=\sigma){\nonumber}\\ &&= \sum_{\eta\in \{0,1\}^E}\sum_{\eta'\in \{0,1\}^E}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta'|\sigma_s=\sigma, \eta_s=\eta) \BBp( \eta_s=\eta|\sigma_s=\sigma). \end{eqnarray} Since $(\eta_t,\sigma_t)_{t\geq 0}$ is a Markov jump process with the transition rates $c((\eta,\sigma),(\eta',\sigma'))$ introduced in definition \ref{def0}, then \begin{eqnarray}\label{ee2} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta'|\sigma_s=\sigma, \eta_s=\eta) = c((\eta,\sigma),(\eta',\sigma^x))=\BBone_{\eta'=\eta}\prod_{e\in E_x}\BBone_{\eta(e)=0}. \end{eqnarray} Hence we get, combining (\ref{ee1}) and (\ref{ee2}), for any $s\geq 0$, \begin{eqnarray*} && \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma)=\sum_{\eta\in \{0,1\}^E}\left(\prod_{e\in E_x}\BBone_{\eta(e)=0}\right)\BBp( \eta_s=\eta|\sigma_s=\sigma). \end{eqnarray*} The measure $IP$ is, by proposition \ref{prorev1}, a reversible and a stationary measure for the process $(\eta_t,\sigma_t)_{t\geq 0}$. Hence, if $(\eta_0,\sigma_0)$ is distributed as $IP$, then, for any $s\geq 0$, $(\eta_s,\sigma_s)$ is also distributed as $IP$. Consequently, under the hypothesis of proposition \ref{marg}, we have $$ \BBp( \eta_s=\eta|\sigma_s=\sigma)=\frac{\BBp( \eta_s=\eta,\sigma_s=\sigma)}{\BBp(\sigma_s=\sigma)}=\frac{IP(\eta,\sigma)} {\sum_{\eta'\in \{0,1\}^E}IP(\eta',\sigma)}\,.$$ We conclude that \begin{equation}\label{sumtaux} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) \, = \, \frac{\sum_{\eta\in \{0,1\}^E,\,\, \eta\equiv 0\, on\, E_x}IP(\eta,\sigma)} {\sum_{\eta'\in \{0,1\}^E}IP(\eta',\sigma)} \,. \end{equation} We compute these sums, starting with the very definition of the coupling measure $IP$. We have \begin{equation}\label{sumaux} \sum_{\eta\in \{0,1\}^E,\,\, \eta\equiv 0\, on\, E_x}IP(\eta,\sigma) \,=\, \frac{1}{Z} (1-p)^{|\{\,e\in E:\delta_\sigma(e)=0\,\}|} (1-p)^{|\{\,e\in E_x:\delta_\sigma(e)=1\,\}|} \,. \end{equation} The sum in the denominator has already been computed in equation~\eqref{transf}. Equations~\eqref{transf}, \eqref{sumtaux} and~\eqref{sumaux} together yield the desired result~\eqref{limite}. \end{proof} We consider next the marginal on the edges and we obtain a similar result. \begin{pro}\label{marv} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process with the transition rates introduced in definition \ref{def0}. Suppose that the pair $(\eta_0,\sigma_0)$ is distributed according to $IP$. Then, for any $e\in E$, for any $\eta\in \{0,1\}^E$ and any $s\geq 0$, \begin{equation}\label{gimite} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\eta_{t+s}=\eta^e|\eta_s=\eta)= (1-p)\BBone_{\eta(e)=1}+ p\BBone_{\eta(e)=0,\, \gamma_\eta(e)=1 } + \frac{p}{2}\BBone_{\eta(e)=0,\, \gamma_\eta(e)=0 }\,. \end{equation} \end{pro} \begin{proof} We have, for any $\eta\in \{0,1\}^E$, $e\in E$ and $s\geq 0$, \begin{eqnarray}\label{ff1} && \BBp(\eta_{t+s}=\eta^e|\eta_s=\eta) = \sum_{\sigma\in \{-1,1\}^V}\sum_{\sigma'\in \{-1,1\}^V} \BBp(\sigma_{t+s}=\sigma', \eta_{t+s}=\eta^e, \sigma_s=\sigma | \eta_s=\eta {\nonumber})\\ &&= \sum_{\sigma\in \{-1,1\}^V}\sum_{\sigma'\in \{-1,1\}^V} \BBp(\sigma_{t+s}=\sigma', \eta_{t+s}=\eta^e|\sigma_s=\sigma, \eta_s=\eta) \BBp( \sigma_s=\sigma | \eta_s=\eta). \end{eqnarray} Since $(\eta_t,\sigma_t)_{t\geq 0}$ is a Markov jump process with the transition rates $c((\eta,\sigma),(\eta',\sigma'))$ introduced in definition \ref{def0}, then \begin{multline}\label{ff2} \lim_{t\rightarrow 0}\frac{1}{t} \BBp(\sigma_{t+s}=\sigma', \eta_{t+s}=\eta^e|\sigma_s=\sigma, \eta_s=\eta) = c((\eta,\sigma),(\eta^e,\sigma'))\hfill\cr =\BBone_{\sigma'=\sigma} \big(p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1}\big) \,. \end{multline} Combining (\ref{ff1}) and (\ref{ff2}), we get, for any $s\geq 0$, \begin{multline*} \lim_{t\rightarrow 0}\frac{1}{t} \BBp(\eta_{t+s}=\eta^e|\eta_s=\eta) =\hfill\cr \sum_{\sigma\in \{-1,1\}^V} \big(p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1}\big) \BBp( \sigma_s=\sigma | \eta_s=\eta). \end{multline*} The measure $IP$ is, by proposition \ref{prorev1}, a reversible and a stationary measure for the process $(\eta_t,\sigma_t)_{t\geq 0}$. Hence, if $(\eta_0,\sigma_0)$ is distributed as $IP$, then, for any $s\geq 0$, $(\eta_s,\sigma_s)$ is also distributed as $IP$. Consequently, under the hypothesis of proposition \ref{marv}, we have $$ \BBp( \sigma_s=\sigma | \eta_s=\eta) =\frac{\BBp( \eta_s=\eta,\sigma_s=\sigma)}{\BBp(\eta_s=\eta)}=\frac{IP(\eta,\sigma)} {\sum_{\sigma'\in \{-1,1\}^V}IP(\eta,\sigma')}\,.$$ We conclude that \begin{equation}\label{sumfaux} \lim_{t\rightarrow 0}\frac{1}{t} \BBp(\eta_{t+s}=\eta^e|\eta_s=\eta) \, = \, \frac{\sum_{\sigma\in \{-1,1\}^V} \big(p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1}\big) IP(\eta,\sigma)} {\sum_{\sigma'\in \{-1,1\}^V}IP(\eta,\sigma')} \,. \end{equation} We compute these sums, starting with the very definition of the coupling measure $IP$. We have \begin{multline}\label{tumaux} {\sum_{\sigma\in \{-1,1\}^V} \big(p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1}\big) IP(\eta,\sigma)} \,=\,\hfill\cr \Big( p\BBone_{\eta(e)=0,\, \gamma_\eta(e)=1 } + \frac{p}{2}\BBone_{\eta(e)=0,\, \gamma_\eta(e)=0} + (1-p)\BBone_{\eta(e)=1} \Big) \Phi_{p,2}(\eta) \,. \end{multline} The sum in the denominator has already been computed in equation~\eqref{tramsf}, it is equal to $\Phi_{p,2}(\eta)$. Equations~\eqref{tramsf}, \eqref{sumfaux} and~\eqref{tumaux} together yield the desired result~\eqref{gimite}. \end{proof} \noindent Although the infinitesimal behavior of the marginal on the spins of $(\eta_t,\sigma_t)_{t\geq 0}$ is the same as for the Glauber dynamics with the transition rates given in (\ref{marginal}), the process $(\sigma_t)_{t\geq 0}$ does not evolve according to a Glauber dynamics, because it is not a Markovian process. Similarly, the marginal on the edges is not a Markovian process. We summarize these results in the following theorem. \begin{theo}\label{theo20} We suppose that the set of edges $E$ is not empty. Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process with the transition rates introduced in definition \ref{def0}. Then $(\eta_t,\sigma_t)_{t\geq 0}$ is reversible with respect to the coupling measure IP. However its two marginal processes $(\eta_t)_{t\geq 0}$ and $(\sigma_t)_{t\geq 0}$ are not Markovian jump processes. \end{theo} \begin{proof} The reversibility property has been proved in proposition \ref{prorev1}. Yet the sum $$\sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma^x))$$ depends on $\eta$ for some $\sigma\in \{-1,1\}^V$ and $x\in V$. In fact, we have, for any $\sigma\in \{-1,1\}^V$ and $x\in V$, $$ \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma^x))=c((\eta,\sigma), (\eta,\sigma^x))=\prod_{e\in E_x}\BBone_{\eta(e)=0}. $$ Similarly, the sum $$\sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta^e,\sigma'))$$ depends on $\sigma$ for some $\eta\in \{0,1\}^E$ and $e\in E$. In fact, we have, for any $\eta\in \{0,1\}^E$ and $e\in E$, $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta^e,\sigma'))= c((\eta,\sigma), (\eta^e,\sigma))= p\BBone_{\eta(e)=0}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=1}. $$ The proof of Theorem \ref{theo20} is complete thanks to Theorem 3.1 in Ball and Yeo \cite{Ball} (see theorem \ref{theoball} in the appendix). \end{proof} \noindent Theorem \ref{theo20} shows that the Markov jump process with the transition rates given in definition \ref{def0} does not fulfill our dream, which was to build a Markov jump process on $\{0,1\}^E\times \{-1,1\}^V$, reversible with respect to the $IP$ measure, with a Glauber dynamics as the marginal on the spins and such that the marginal on the edges is Markovian. The following theorem shows that there is no hope to realize this dream with a Markov jump process $(\eta_t,\sigma_t)_{t\geq 0}$ which updates at most one site or one edge at each time. \begin{theo}\label{theoonechange} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process which updates at most one site or one edge at each time and whose marginal on the spins is a Markov jump process with transition rates satisfying (\ref{tauxglauber}). Then $(\eta_t,\sigma_t)_{t\geq 0}$ cannot be reversible with respect to the coupling measure $IP$. \end{theo} \begin{proof} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process satisfying the hypothesis of theorem \ref{theoonechange} and let $q((\eta,\sigma),(\eta',\sigma'))$ be its transition rates. Since this Markov process updates at most one site or one edge at each time, then \begin{equation} \sum_{\eta'\in \{0,1\}^E} q((\eta,\sigma),(\eta',\sigma^x))= q((\eta,\sigma),(\eta,\sigma^x)). \end{equation} Since the marginal on the spins of $(\eta_t,\sigma_t)_{t\geq 0}$ is a Markov process with transition rates $c(\sigma,\sigma')$ satisfying (\ref{tauxglauber}), Theorem 3.1 in Ball and Yeo \cite{Ball} gives that $$\forall \eta\in \{0,1\}^E \qquad \sum_{\eta'\in \{0,1\}^E} q((\eta,\sigma),(\eta',\sigma^x)))=c(\sigma,\sigma^x)\,.$$ Hence, for any $\eta\in \{0,1\}^E$, any $\sigma\in \{-1,1\}^V$ and any $x\in V$ $$ q((\eta,\sigma),(\eta,\sigma^x))= c(\sigma,\sigma^x). $$ Suppose now that the detailed balance equation (\ref{eqreversible}) is satisfied. We deduce then from (\ref{tauxglauber}) and lemma \ref{lemn} that necessarily $\eta(e)=0$ for $e\in E_x$. We conclude that the detailed balance equation (\ref{eqreversible}) is not satisfied for every $\eta\in \{0,1\}^E$ and every $\sigma\in \{-1,1\}^V$. The proof of Theorem \ref{theoonechange} is complete. \end{proof} \noindent Theorem~$3.2$ in \cite{Ball} (see theorem \ref{theoball2} in the appendix) implies that, for a process like the above one, in which at most one marginal process moves at each time step, the marginal processes are Markovian if and only if they are independent. If that were the case, then the equilibrium measure would be a product measure. Therefore it is impossible to realize our dream with a process which changes only one edge or one spin at a time. In the next section, we construct a process which can change simultaneously one spin and its incident edges. \section{One site and the incident edges} \label{two} We build here a dynamics which updates at most one site and its incident edges at each time, which is reversible with respect to the coupling measure $IP$ and whose marginal on the spins is a Glauber dynamics. For $x\in V$, recall that $E_x$ is the set of the edges $e$ in $E$ having $x$ as endvertex, i.e., $$ E_x=\{e\in E,\, e=\langle x,y\rangle,\, y\in V\}, $$ and define $$ \cC_x=\big\{(\eta,\sigma)\in \{0,1\}^E\times \{-1,1\}^V: \eta(e) \leq \delta_{\sigma}(e)\,\,\,\text{for any}\,\,\, e\in E_x\big\}. $$ The following definition gives the transition rates of the dynamics. \begin{defi}\label{def1} Let $(\eta,\sigma)$, $(\eta',\sigma')$ be two elements of $\{0,1\}^E\times\{-1,1\}^V$. We consider several cases: \noindent $\bullet$ If there exists $x\in V$ such that $\sigma'=\sigma^x$ and $\eta'(e)=\eta(e)$ for any $e\in E\setminus E_x$, then we define $$ c((\eta,\sigma), (\eta',\sigma'))= (1-p)^{|\{e\in E_x,\,\, \eta'(e)=0\}|}p^{|\{e\in E_x,\,\,\, \eta'(e)=1\}|}\BBone_{(\eta',\sigma')\in \cC_x}\,. $$ \noindent $\bullet$ Otherwise, if $(\eta,\sigma)\neq (\eta',\sigma')$, then we set $c((\eta,\sigma), (\eta',\sigma'))= 0$. \noindent $\bullet$ Finally, if $(\eta,\sigma)= (\eta',\sigma')$, then we set $$ c((\eta,\sigma), (\eta,\sigma))= -\sum_{(\eta',\sigma'),\, (\eta',\sigma')\neq (\eta,\sigma)}c((\eta,\sigma), (\eta',\sigma'))\,. $$ \end{defi} \begin{pro}\label{proreversible} The transition rates introduced in definition \ref{def1} satisfy the detailed balance equation (\ref{eqreversible}). \end{pro} \begin{proof} Let $(\eta,\sigma)$ and $(\eta',\sigma')$ be fixed in $\{0,1\}^E\times \{-1,1\}^V$. We have to prove (\ref{eqreversible}) only in the case where there exists $x\in V$ for which \begin{equation}\label{vbn} \sigma'=\sigma^x\,,\qquad \forall e\in E\setminus E_x\quad \eta'(e)= \eta(e)\,, \end{equation} because in all the other cases, the detailed balance equation (\ref{eqreversible}) is trivially satisfied. Let us fix $x\in V$. Suppose first that $(\eta,\sigma)$ and $(\eta',\sigma')$ are two elements of $\cC_x$ satisfying~(\ref{vbn}). We get then, for $e\in E_x$, $$\BBone_{\eta(e)=1}\delta_{\sigma}(e)= \BBone_{\eta(e)=1}\,,\qquad \BBone_{\eta'(e)=1}\delta_{\sigma^x}(e)= \BBone_{\eta'(e)=1}\,,$$ therefore \begin{eqnarray*} &&\kern-23pt\left(\prod_{e\in E_x}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0} \right)\right)c((\eta,\sigma), (\eta',\sigma^x))\\ &&= p^{|\{e\in E_x,\,\,\, \eta(e)=1\}|}(1-p)^{|\{e\in E_x,\,\, \eta(e)=0\}|}(1-p)^{|\{e\in E_x,\,\, \eta'(e)=0\}|}p^{|\{e\in E_x,\,\,\, \eta'(e)=1\}|}\\ &&=\left(\prod_{e\in E_x}\left(p\BBone_{\eta'(e)=1}\delta_{\sigma^x}(e)+ (1-p)\BBone_{\eta'(e)=0} \right)\right)c((\eta',\sigma^x), (\eta,\sigma)). \end{eqnarray*} Now we multiply each member of the last equality by $$\prod_{e\in E\setminus E_x}\left(p\BBone_{\eta(e)=1}\delta_{\sigma}(e)+ (1-p)\BBone_{\eta(e)=0}\right) \,=\, \prod_{e\in E\setminus E_x}\left(p\BBone_{\eta'(e)=1}\delta_{\sigma^x}(e)+ (1-p)\BBone_{\eta'(e)=0}\right) \,$$ and we obtain the detailed balance equation (\ref{eqreversible}). Suppose now that $(\eta,\sigma)\in \cC_x$ and $(\eta',\sigma^x)\notin \cC_x$. We have then, by the definition of the transition rates, $$ IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma^x))=0\,. $$ Moreover $IP(\eta',\sigma^x)=0$ because $(\eta',\sigma^x)\notin \cC$, therefore $IP(\eta',\sigma^x)c((\eta',\sigma^x), (\eta,\sigma))=0$. So here again equation (\ref{eqreversible}) is satisfied. Finally suppose that $(\eta,\sigma)\notin \cC_x$. Then $IP(\eta,\sigma)=0$ and so $IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma^x))=0$ which is equal to $IP(\eta',\sigma^x)c((\eta',\sigma^x),(\eta,\sigma))$, since the transition rates $c((\eta',\sigma^x),(\eta,\sigma))$ vanish as soon as $(\eta,\sigma)\notin \cC_x$. We conclude that the detailed balance equation (\ref{eqreversible}) is always satisfied. \end{proof} \noindent We compute next the marginal dynamics on the spins. \begin{pro}\label{pro1} For any $\eta$ in $\{0,1\}^E$ and any $\sigma,\sigma'$ in $\{-1,1\}^V$, the sum $$\sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma'))$$ does not depend on $\eta$ and it is equal to $$ c(\sigma,\sigma') \,=\,\left\{ \begin{array}{cc} (1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}\,\, & if \,\, \sigma'=\sigma^x\,, \\ -\sum_{x\in V}(1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}\,\, & if \,\, \sigma'=\sigma\,, \\ 0 & otherwise\,. \end{array} \right. $$ \end{pro} \begin{proof} Let $\eta\in \{0,1\}^E$ be fixed. If $|\{y\in V, \sigma(y)\neq \sigma'(y)\}|\geq 2$ then, by definition \ref{def1}, \begin{equation}\label{a1} \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma'))=0. \end{equation} Suppose now that there exists $x\in V$ such that $\sigma'=\sigma^x$. Then, by definition \ref{def1}, \begin{multline*} \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma'))\,=\,\\ \sum_{\eta'\in \{0,1\}^E,\,\, \eta'\equiv \eta\, on \, E\setminus E_x}(1-p)^{|\{e\in E_x,\,\, \eta'(e)=0\}|}p^{|\{e\in E_x,\,\,\, \eta'(e)=1\}|}\BBone_{(\eta',\sigma^x)\in \cC_x}.\\ \end{multline*} Our next task is to calculate the last sum. To this end, we use the same proof as for (\ref{transf}) (recall that $IP(\eta,\sigma)$ is also given by (\ref{ipsimply})). We get \begin{equation}\label{a2} \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma^x))= (1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}. \end{equation} Finally, we have to calculate $ \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma))$. We have, from definition \ref{def1}, \begin{align}\label{a3} \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma),& (\eta',\sigma)) = c((\eta,\sigma), (\eta,\sigma)) {\nonumber} \\ &= -\sum_{(\eta',\sigma'),\, (\eta',\sigma')\neq (\eta,\sigma)}c((\eta,\sigma), (\eta',\sigma')) {\nonumber} \\ &= -\sum_{\eta'\in \{0,1\}^E}\sum_{x\in V}c((\eta,\sigma), (\eta',\sigma^x)){\nonumber} \\ &=-\sum_{x\in V}\sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma^x)){\nonumber} \\ &=-\sum_{x\in V}(1-p)^{|\{e\in E_x,\,\, \delta_{\sigma}(e)=1\}|}. \end{align} Equalities (\ref{a1}), (\ref{a2}) and (\ref{a3}) yield the statement of proposition \ref{pro1}. \end{proof} \noindent We compute next the marginal dynamics on the edges. \begin{pro}\label{tro2} Let $\eta$ and $\eta'$ be two different elements of $\{0,1\}^E$. \noindent $\bullet$ If there exists $e=\langle x,y\rangle\in E$ such that $\eta'(e)\neq \eta(e)$ and $\eta'\equiv\eta\,\, {\mbox{on}}\,\, E\setminus \{e\}$, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta',\sigma'))=\sum_{z\in \{x,y\}}(1-p)^{|\{e\in E_z,\,\, \eta'(e)=0\}|}p^{|\{e\in E_z,\,\, \eta'(e)=1\}|}\BBone_{(\eta',\sigma^z)\in \cC_z}\,. $$ \noindent $\bullet$ If there exists $x\in V$ such that $\eta'\equiv \eta$ on $E\setminus E_x$ and the configurations $\eta,\eta'$ differ in more than one edge, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta',\sigma'))=(1-p)^{|\{e\in E_x,\,\, \eta'(e)=0\}|}p^{|\{e\in E_x,\,\, \eta'(e)=1\}|} \BBone_{(\eta',\sigma^x)\in \cC_x}\,. $$ \noindent $\bullet$ If for any $x \in V$, $\eta'\neq \eta\,\, {\mbox{on}}\,\, E\setminus E_x$, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta',\sigma'))=0\,. $$ \end{pro} \begin{proof} Let $\eta$ and $\eta'$ be two fixed different elements of $\{0,1\}^E$. If for any $x\in V$, $\eta$ and $\eta'$ are different on $E\setminus E_x$ then we get, by definition \ref{def1}, $$ \forall\,\, \sigma'\in \{-1,1\}^V\qquad c((\eta,\sigma),(\eta',\sigma'))=0\,. $$ Consequently, \begin{equation}\label{e1tro2} \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma),(\eta',\sigma'))=0. \end{equation} Suppose now that there exists $x\in V$ such that $\eta\equiv \eta'$ on $E\setminus E_x$. Let us define the set $${\cal V}\,=\,\big\{\,y\in V,\,\, \eta\equiv \eta'\,\, on\,\, E\setminus E_y\,\big\}\,.$$ We have \begin{equation}\label{epro4} \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma),(\eta',\sigma'))= \sum_{y\in {\cal V}}c((\eta,\sigma),(\eta',\sigma^y)). \end{equation} Let us set $${\cal E}\,=\,\big\{f\in E,\,\, \eta(f)\neq \eta'(f)\,\big\}\,.$$ By hypothesis, this set is not empty and it is a subset of $E_x$ because $\eta\neq \eta'$ while $\eta\equiv \eta'$ on $E\setminus E_x$. Suppose first that ${\cal E}$ contains only one element, say ${\cal E}=\{e\}$. Necessarily one endvertex of $e$ is $x$. Let $y$ be the other endvertex of $e$, so that $e=\langle x,y\rangle$. We have then ${\cal V}=\{x,y\}$. Indeed, if there exists a vertex $z$ in $V$ such that $$ z\neq x\,\,\,\text{and}\,\, \eta\equiv \eta'\,\,\text{on}\,\, E\setminus E_z\,, $$ then, since $e\in E_x$ and $\eta(e)\neq \eta'(e)$, we have also $e\in E_z$, whence $e=\langle x,z\rangle$ and $z=y$. Equation (\ref{epro4}) can now be rewritten as \begin{equation}\label{e2tro2} \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma),(\eta',\sigma'))= c((\eta,\sigma),(\eta',\sigma^x))+ c((\eta,\sigma),(\eta',\sigma^y))\,. \end{equation} Suppose that ${\cal E}$ contains more than one element. We claim that, in this case, the set ${\cal V}$ is reduced to $\{x\}$. Indeed, let $e_1$ and $e_2$ be two different elements of ${\cal E}$. Let $y\in {\cal V}$. Since $\eta\equiv \eta'\,\,\text{on}\,\, E\setminus E_y$, necessarily both $e_1$ and $e_2$ belong to $E_y$. We know also that both $e_1$ and $e_2$ belong to $E_x$. Consequently $e_1\in E_x\cap E_y$ and $e_2\in E_x\cap E_y$. It follows that $y=x$ since $e_1$ and $e_2$ are different. From (\ref{epro4}), we deduce that, whenever ${\cal E}$ contains more than one element, \begin{equation}\label{e3tro2} \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma),(\eta',\sigma'))= c((\eta,\sigma),(\eta',\sigma^x))\,. \end{equation} Definition \ref{def1} and the three cases of equations~(\ref{e1tro2}), (\ref{e2tro2}) and (\ref{e3tro2}) complete the proof of Proposition \ref{tro2}. \end{proof} \noindent From propositions \ref{proreversible}, \ref{pro1} and \ref{tro2}, we deduce the following theorem. \begin{theo}\label{theo2} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process with the transition rates introduced in definition \ref{def1}. Then $(\eta_t,\sigma_t)_{t\geq 0}$ is reversible with respect to the coupling measure IP. Its second marginal process $(\sigma_t)_{t\geq 0}$ is a Markov process evolving according to a Glauber dynamics while its first marginal $(\eta_t)_{t\geq 0}$ is a non-Markovian jump process. \end{theo} \begin{proof} The reversibility property is deduced from Proposition \ref{proreversible}. Proposition \ref{pro1} shows that the sum $$\sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma'))$$ does not depend on $\eta$ for any $\sigma,\sigma'\in \{-1,1\}^V$, moreover this sum is equal to the transition rates given in (\ref{marginal}). The presence of the indicator function $\BBone_{(\eta',\sigma^x)\in \cC_x}$ in the formula of proposition \ref{tro2} shows that the sum $\sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta',\sigma'))$ depends on $\sigma$ for some $\eta,\eta'\in \{0,1\}^E$. The proof of Theorem \ref{theo2} is complete thanks to Ball and Yeo \cite{Ball}. \end{proof} So this second dynamics does not fulfill our dream. \section{The dream is not feasible}\label{general} A Markov process on $\{0,1\}^E\times \{-1,1\}^V$ is said to be a Markovian coupling if both its marginals are Markov processes. In this section, we will work with the following additional hypothesis on the graph $(V,E)$. \begin{hypo}\label{adhy} We suppose that there exists a vertex ${\overline{x}}\in V$ which belongs to at least four edges. We suppose that there exists a spin configuration $\overline{\sigma}$ such that $\delta_{\overline{\sigma}}(e)=0$ for any $e\in E$. \end{hypo} \noindent For instance, when the graph is bipartite, the spin configuration $\overline{\sigma}$ can be obtained by setting minuses on one part of the graph and pluses on the other part. Theorem \ref{theo1} below shows that, under hypothesis~\ref{adhy}, we cannot build a Markovian coupling which is reversible with respect to the $IP$ measure and whose marginal on the spins is a one spin flip Markov jump process, i.e., with transition rates satisfying (\ref{tauxglauber}). \begin{theo}\label{theo1} Suppose that hypothesis~\ref{adhy} holds. Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markovian jump process with rates $q((\eta,\sigma), (\eta',\sigma'))$ satisfying (\ref{qmatrix1}), (\ref{qmatrix}) and the detailed balance equation (\ref{eqreversible}). If the second marginal $(\sigma_t)_{t\geq 0}$ is a Markov jump process with transition rates satisfying (\ref{tauxglauber}), then the first marginal $(\eta_t)_{t\geq 0}$ is a continuous-time non-Markovian jump process. \end{theo} \begin{proof} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markovian jump process with transition rates $q((\eta,\sigma), (\eta',\sigma'))$ satisfying (\ref{qmatrix1}), (\ref{qmatrix}) and the detailed balance equation (\ref{eqreversible}). Suppose also that the first marginal $(\eta_t)_{t\geq 0}$ is a Markov jump process with transition rates ${\tilde c}(\eta,\eta')$. Suppose also that the second marginal $(\sigma_t)_{t\geq 0}$ is a Markov jump process with transition rates $c(\sigma,\sigma')$ satisfying (\ref{tauxglauber}). From Ball and Yeo \cite{Ball} (see theorem \ref{theoball} of the appendix), we have \begin{equation} \label{mar1} \forall \eta,\eta' \in \{0,1\}^E\qquad \forall \sigma \in \{-1,1\}^V\qquad \sum_{\sigma'\in \{-1,1\}^V} q((\eta,\sigma),(\eta',\sigma'))= {\tilde c}(\eta,\eta')\,, \end{equation} \begin{equation} \label{mar2} \forall \sigma,\sigma' \in \{-1,1\}^V\qquad \forall \eta \in \{0,1\}^E\qquad \sum_{\eta'\in \{0,1\}^E} q((\eta,\sigma),(\eta',\sigma'))= {c}(\sigma,\sigma')\,. \end{equation} Let $\overline{\sigma}$ be a configuration as in hypothesis~\ref{adhy} and let ${\overline{x}}$ be a vertex of $E$ belonging to at least four distinct edges. Let $\overline{\eta}$ be the configuration where all the edges are closed and let $\overline{\eta}^{{\overline{x}}}$ be the configuration where the edges exiting from ${\overline{x}}$ are opened, while all the other edges are closed. From formula~\eqref{mar2} applied to $\overline{\eta},\overline{\sigma}$ and $\overline{\sigma}^{{\overline{x}}}$, we have \begin{equation*} \sum_{\eta'\in \{0,1\}^E} q((\overline{\eta}^{\overline{x}},\overline{\sigma}^{\overline{x}}),(\eta',\overline{\sigma}))= {c}(\overline{\sigma}^{\overline{x}},\overline{\sigma})\,. \end{equation*} The only configuration $\eta'$ compatible with $\overline{\sigma}$ is $\overline{\eta}$, thus the above identity reduces to $$q((\overline{\eta}^{\overline{x}},\overline{\sigma}^{\overline{x}}),(\overline{\eta},\overline{\sigma}))= {c}(\overline{\sigma}^{\overline{x}},\overline{\sigma})\,.$$ The conditions~\eqref{tauxglauber} implies that ${c}(\overline{\sigma}^{\overline{x}},\overline{\sigma})$ is positive, therefore $q((\overline{\eta}^{\overline{x}},\overline{\sigma}^{\overline{x}}),(\overline{\eta},\overline{\sigma}))$ is positive. Since we assumed that the dynamics is reversible, then the rate $q((\overline{\eta},\overline{\sigma}),(\overline{\eta}^{\overline{x}},\overline{\sigma}^{\overline{x}}))$ is also positive. From formula~\eqref{mar1} applied to $\overline{\eta},\overline{\eta}^{\overline{x}}$ and $\overline{\sigma}$, we have \begin{equation*} \sum_{\sigma'\in \{-1,1\}^V} q((\overline{\eta},\overline{\sigma}),(\overline{\eta}^{\overline{x}},\sigma'))= {\tilde c}(\overline{\eta},\overline{\eta}^{\overline{x}})\,, \end{equation*} therefore the rate ${\tilde c}(\overline{\eta},\overline{\eta}^{\overline{x}})$ is positive. Let now $\widehat{\sigma}$ be a spin configuration in which the spins of the vertices connected to ${\overline{x}}$ contain at least two negative spins and at least two positive spins. This is possible because we assumed that ${\overline{x}}$ is connected to at least four distinct vertices. Applying again formula~\eqref{mar1}, this time to $\overline{\eta},\overline{\eta}^{\overline{x}}$ and $\widehat{\sigma}$, we have \begin{equation*} \sum_{\sigma'\in \{-1,1\}^V} q((\overline{\eta},\widehat{\sigma}),(\overline{\eta}^{\overline{x}},\sigma'))= {\tilde c}(\overline{\eta},\overline{\eta}^{\overline{x}})\,. \end{equation*} Since the rate ${\tilde c}(\overline{\eta},\overline{\eta}^{\overline{x}})$ is positive, then there exists a spin configuration $\sigma'$ such that $$q((\overline{\eta},\widehat{\sigma}),(\overline{\eta}^{\overline{x}},\sigma'))\,>\,0\,.$$ In the configuration $\overline{\eta}^{\overline{x}}$, all the edges emanating from ${\overline{x}}$ are opened, hence the neighbours of ${\overline{x}}$ are connected. Since the configuration $\sigma'$ has to be compatible with $\overline{\eta}^{\overline{x}}$, then all the neighbours of ${\overline{x}}$ have the same spin in $\sigma'$, hence the configuration $\sigma'$ has to differ from $\widehat{\sigma}$ in at least two vertices. From the conditions~\eqref{tauxglauber}, we should therefore have that $c(\widehat{\sigma},\sigma')=0$. Yet formula~\eqref{mar2} applied to $\overline{\eta},\overline{\sigma}$ and $\overline{\sigma}^{{\overline{x}}}$ yields \begin{equation*} {c}(\widehat{\sigma},\sigma')\,=\, \sum_{\eta'\in \{0,1\}^E} q((\overline{\eta},\widehat{\sigma}),(\eta',\sigma')) \,\geq\, q((\overline{\eta},\widehat{\sigma}),(\overline{\eta}^{\overline{x}},\sigma'))\,>\,0\,, \end{equation*} which is contradictory. \end{proof} \noindent Notice that hypothesis~\ref{adhy} holds for a cubic box on the $d$--dimensional lattice when $d\geq 2$. Thus our dream process is not realizable in the lattice $\mathbb{Z}^d$ for $d\geq 2$. \section{One edge and one incident cluster} \label{three} We drop here the requirement that the dynamics is local. We build a dynamics which updates at each time at most one edge and the vertices belonging to the open cluster of one of the endpoints of the edge. This dynamics will be reversible with respect to the coupling measure $IP$ and its marginal on the edges is the FK dynamics described in section~\ref{fkdyna}. For $x\in V$, recall that $E_x$ is the set of the edges $e$ in $E$ having $x$ as endvertex, i.e., $$ E_x=\{e\in E,\, e=\langle x,y\rangle,\, y\in V\}\,. $$ The following definition gives the transition rates of the dynamics. \begin{defi}\label{def7} Let $(\eta,\sigma)$, $(\eta',\sigma')$ be two elements of $\{0,1\}^E\times\{-1,1\}^V$. We consider several cases: \noindent $\bullet$ If $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$ and $\gamma_\eta(e)=1$, then we set $$ c((\eta,\sigma), (\eta',\sigma'))= \big( (1-p)\BBone_{\eta(e)=1}+ p\BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \,. $$ \noindent $\bullet$ If $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$, $\gamma_\eta(e)=0$ and $\delta_\sigma(e)=1$, then we set $$ c((\eta,\sigma), (\eta',\sigma'))\,= \, \frac{1}{2} \big( (1-p)\BBone_{\eta(e)=1}+ p \BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \,. $$ \noindent $\bullet$ If there exists $x\in V$ and $e\in E_x$ such that $\eta'=\eta^e$, $\gamma_\eta(e)=0$, $\eta(e)=\delta_\sigma(e)$ and $\sigma'$ is given by $$\forall y\in V\qquad\sigma'(y)\,=\, \begin{cases} \sigma(y)&\text{ if $y$ is not connected to $x$ in $\eta\setminus\{ e\}$}\,,\cr -\sigma(y)&\text{ if $y$ is connected to $x$ in $\eta\setminus\{ e\}$}\,,\cr \end{cases} $$ then we set $$ c((\eta,\sigma), (\eta',\sigma'))\,= \, \frac{1}{4} \big( (1-p)\BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + p \BBone_{\eta(e)=0} \BBone_{(\eta',\sigma')\in \cC} \big) \,. $$ \noindent $\bullet$ Otherwise, if $(\eta,\sigma)\neq (\eta',\sigma')$, then we set $c((\eta,\sigma), (\eta',\sigma'))= 0$. \noindent $\bullet$ Finally, if $(\eta,\sigma)= (\eta',\sigma')$, then we set $$ c((\eta,\sigma), (\eta,\sigma))= -\sum_{(\eta',\sigma'),\, (\eta',\sigma')\neq (\eta,\sigma)}c((\eta,\sigma), (\eta',\sigma'))\,. $$ \end{defi} \noindent We first check that the dynamics associated to these rates is reversible with respect to the measure $IP$. \begin{pro}\label{qroreversible} The transition rates introduced in definition \ref{def7} satisfy the detailed balance equation (\ref{eqreversible}). \end{pro} \begin{proof} Let $(\eta,\sigma)$ and $(\eta',\sigma')$ be fixed in $\{0,1\}^E\times \{-1,1\}^V$. We have to prove (\ref{eqreversible}). We consider several cases, as in definition~\ref{def7}. Suppose first that $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$ and $\gamma_\eta(e)=1$. We have then \begin{multline*} IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma))\,=\,\cr \kern-23pt \frac{1}{Z} \left(\prod_{f\in E}\left(p\BBone_{\eta(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta(f)=0} \right)\right) \big( (1-p)\BBone_{\eta(e)=1}+ p\BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \cr \,=\, \frac{1}{Z} \left(\prod_{f\in E\setminus\{e\}} \left(p\BBone_{\eta(f)=1}+ (1-p)\BBone_{\eta(f)=0} \right)\right) (1-p)p\, \BBone_{(\eta,\sigma)\in \cC}\,. \end{multline*} Since $\eta'=\eta^e$ and $\gamma_\eta(e)=1$, then ${(\eta,\sigma)\in \cC}$ if and only if ${(\eta',\sigma)\in \cC}$. Also, the product in the last formula is the same for $\eta$ and $\eta'$, thus we obtain the detailed balance equation (\ref{eqreversible}). Suppose next that $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$, $\gamma_\eta(e)=0$ and $\delta_\sigma(e)=1$. We have then \begin{multline*} IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma))\,=\,\cr \kern-23pt \frac{1}{Z} \left(\prod_{f\in E}\left(p\BBone_{\eta(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta(f)=0} \right)\right) \frac{1}{2} \big( (1-p)\BBone_{\eta(e)=1}+ p\BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \cr \,=\, \frac{1}{Z} \left(\prod_{f\in E\setminus\{e\}} \left(p\BBone_{\eta(f)=1}+ (1-p)\BBone_{\eta(f)=0} \right)\right) \frac{1}{2} (1-p)p\, \BBone_{(\eta,\sigma)\in \cC}\,. \end{multline*} Since $\eta'=\eta^e$ and $\delta_\sigma(e)=1$, then ${(\eta,\sigma)\in \cC}$ if and only if ${(\eta',\sigma)\in \cC}$. Also, the product in the last formula is the same for $\eta$ and $\eta'$, thus we obtain the detailed balance equation (\ref{eqreversible}). Suppose finally that there exists $x\in V$ and $e\in E_x$ such that $\eta'=\eta^e$, $\gamma_\eta(e)=0$, $\eta(e)=\delta_\sigma(e)$ and $\sigma'$ is obtained from $\sigma$ by reversing all the spins of the sites which are connected to $x$ by an open path in $\eta\setminus\{e\}$. We have then $$\displaylines{ IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma'))\,=\,\hfill\cr \frac{1}{Z} \left(\prod_{f\in E}\left(p\BBone_{\eta(f)=1}\delta_{\sigma}(f)+ (1-p)\BBone_{\eta(f)=0} \right)\right) \frac{1}{4} \big( (1-p)\BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + p \BBone_{\eta(e)=0} \BBone_{(\eta',\sigma')\in \cC} \big) \cr \,=\, \frac{1}{Z} \left(\prod_{f\in E\setminus\{e\}} \left(p\BBone_{\eta(f)=1}+ (1-p)\BBone_{\eta(f)=0} \right)\right) \frac{1}{4} (1-p)p\, \big( \BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + \BBone_{\eta(e)=0} \BBone_{(\eta',\sigma')\in \cC} \big) \,. }$$ To remove the symbol $\delta_\sigma(f)$ in the last line, we have used the fact that, if $\eta(e)=0$ and $(\eta',\sigma')\in \cC$, then $(\eta,\sigma)\in \cC$. Since we have $$ \BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + \BBone_{\eta(e)=0} \BBone_{(\eta',\sigma')\in \cC} \,=\, \BBone_{\eta'(e)=1} \BBone_{(\eta',\sigma')\in \cC} + \BBone_{\eta'(e)=0} \BBone_{(\eta,\sigma)\in \cC}\,, $$ then we can conclude that $$IP(\eta,\sigma)c((\eta,\sigma),(\eta',\sigma'))\,=\, IP(\eta',\sigma')c((\eta',\sigma'),(\eta,\sigma))$$ and the detailed balance equation holds also in this case. We conclude that the detailed balance equation (\ref{eqreversible}) is always satisfied. \end{proof} \noindent We compute next the marginal dynamics on the edges. \begin{pro}\label{qro2} Let $\eta$ be an element of $\{0,1\}^E$ and let $e\in E$. Let also $\sigma$ be an element of $\{-1,1\}^V$ such that $(\eta,\sigma)\in \cC$. \noindent $\bullet$ If $\gamma_\eta(e)=1$, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta^e,\sigma'))\,=\, p\BBone_{\eta(e)=0}+ (1-p)\BBone_{\eta(e)=1}\,. $$ \noindent $\bullet$ If $\gamma_\eta(e)=0$ and $\eta(e)=0$, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta^e,\sigma'))\,=\, \frac{\displaystyle p}{\displaystyle 2}\,. $$ \noindent $\bullet$ If $\gamma_\eta(e)=0$ and $\eta(e)=1$, then $$ \sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta^e,\sigma')) \,=\, 1-p\,. $$ \end{pro} \begin{proof} Let $\eta,e,\sigma$ be as in the statement of the proposition. We use the expression for the rates given in the definition~\ref{def7}. If $\gamma_\eta(e)=1$, then $c((\eta,\sigma), (\eta^e,\sigma'))$ is null unless $\sigma'=\sigma$. Thus the sum reduces to $c((\eta,\sigma), (\eta^e,\sigma))$, which is equal to $p$ if $\eta(e)=0$ and to $1-p$ if $\eta(e)=1$. Suppose next that $\gamma_\eta(e)=0$ and $\eta(e)=0$. We consider two further cases. If $\delta_\sigma(e)=1$, then the sum reduces again to $c((\eta,\sigma), (\eta^e,\sigma))$, which in this case is equal to $p/2$. If $\delta_\sigma(e)=0$, then the sum contains two terms, corresponding to the two spin configurations $\sigma'$ obtained from $\sigma$ by reversing all the spins of the sites which are connected to one extremity of $e$ by an open path in $\eta\setminus\{e\}$. These two rates $c((\eta,\sigma), (\eta^e,\sigma'))$ are both equal to $p/4$, and their sum is again equal to $p/2$. Suppose finally that $\gamma_\eta(e)=0$ and $\eta(e)=1$. The sum contains three terms, one corresponding to $\sigma'=\sigma$ and two corresponding to the two spin configurations $\sigma'$ obtained from $\sigma$ by reversing all the spins of the sites which are connected to one extremity of $e$ by an open path in $\eta\setminus\{e\}$. The rate $c((\eta,\sigma), (\eta^e,\sigma))$ is equal to $(1-p)/2$. The two other rates $c((\eta,\sigma), (\eta^e,\sigma'))$ are equal to $(1-p)/4$. In total, the sum is equal to $1-p$. \end{proof} \noindent The marginal dynamics on the spins is quite complicated, but it is not a Markov process. Let us consider an edge configuration $\eta$ and a spin configuration $\sigma$ such that there exist $x\in V$ and $e\in E_x$ such that $\gamma_\eta(e)=\eta(e)=\delta_\sigma(e)=0$. Let $\sigma'$ be the spin configuration defined by $$\forall y\in V\qquad\sigma'(y)\,=\, \begin{cases} \sigma(y)&\text{ if $y$ is not connected to $x$ in $\eta\setminus\{ e\}$}\,,\cr -\sigma(y)&\text{ if $y$ is connected to $x$ in $\eta\setminus\{ e\}$}\,.\cr \end{cases} $$ We have then $$ \sum_{\eta'\in \{0,1\}^E}c((\eta,\sigma), (\eta',\sigma'))\,=\, c((\eta,\sigma), (\eta^e,\sigma'))\,= \, \frac{p}{4}\,. $$ However the configuration $\sigma'$ depends on $\eta$. Indeed, if we change the status of some edges $f$ in $E_x\setminus \{e\}$, we will change the set of the vertices which are connected to $x$ in $\eta\setminus\{e\}$, and we will obtain an edge configuration $\overline{\eta}$ for which $$\sum_{\eta'\in \{0,1\}^E}c((\overline{\eta},\sigma), (\eta',\sigma'))\,=\, c((\overline{\eta},\sigma), (\overline{\eta}^e,\sigma'))\,= \,0\,.$$ Thus the above sum depends on the configuration $\eta$ and the marginal process on the spin configurations is not a Markov process. \noindent This remark, together with propositions \ref{qroreversible} and~\ref{qro2}, yield the following theorem. \begin{theo}\label{theo3} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process on $\cC$ with the transition rates introduced in definition \ref{def1}. Then $(\eta_t,\sigma_t)_{t\geq 0}$ is reversible with respect to the coupling measure IP. Its first marginal $(\eta_t)_{t\geq 0}$ is a Markov process evolving according to the FK dynamics while its second marginal process $(\sigma_t)_{t\geq 0}$ is a non-Markovian jump process. \end{theo} \begin{proof} The reversibility property is deduced from Proposition \ref{qroreversible}. Proposition \ref{qro2} shows that the sum $\sum_{\sigma'\in \{-1,1\}^V}c((\eta,\sigma), (\eta',\sigma'))$ does not depend on $\sigma$ for any $\eta,\eta'\in \{0,1\}^E$, moreover this sum is equal to the transition rates given in (\ref{marginal}). The proof of Theorem \ref{theo3} is complete thanks to the remark before the theorem and the result of Ball and Yeo \cite{Ball}. \end{proof} So this third dynamics does not fulfill our dream, although it is not even local. \section{One edge and one spin} \label{four} A notable inconvenient of the previous dynamics is that it might reverse simultaneously all the spins associated to an open cluster of the edge configuration. We introduce here a slight modification of this dynamics to ensure that at most one spin is changed at a time. For $x\in V$, recall that $E_x$ is the set of the edges $e$ in $E$ having $x$ as endvertex, i.e., $$ E_x=\{e\in E,\, e=\langle x,y\rangle,\, y\in V\}\,. $$ The following definition gives the transition rates of the dynamics. \begin{defi}\label{def8} Let $(\eta,\sigma)$, $(\eta',\sigma')$ be two elements of $\{0,1\}^E\times\{-1,1\}^V$. We consider several cases: \noindent $\bullet$ If $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$ and $\gamma_\eta(e)=1$, then we set $$ c((\eta,\sigma), (\eta',\sigma'))= \big( (1-p)\BBone_{\eta(e)=1}+ p\BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \,. $$ \noindent $\bullet$ If $\sigma'=\sigma$ and there exists $e\in E$ such that $\eta'=\eta^e$, $\gamma_\eta(e)=0$ and $\delta_\sigma(e)=1$, then we set $$ c((\eta,\sigma), (\eta',\sigma'))\,= \, \frac{1}{2} \big( (1-p)\BBone_{\eta(e)=1}+ p \BBone_{\eta(e)=0}\big) \BBone_{(\eta,\sigma)\in \cC} \,. $$ \noindent $\bullet$ If there exists $x\in V$ and $e\in E_x$ such that $\sigma'=\sigma^x$, $\eta'=\eta^e$, $\gamma_\eta(e)=0$, $\eta(e)=\delta_\sigma(e)$ and $\eta(f)=0$ for $f\in E_x\setminus\{e\}$, then we set $$ c((\eta,\sigma), (\eta',\sigma'))\,= \, \frac{1}{4} \big( (1-p)\BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + p \BBone_{\eta(e)=0} \BBone_{(\eta',\sigma')\in \cC} \big) \,. $$ \noindent $\bullet$ Otherwise, if $(\eta,\sigma)\neq (\eta',\sigma')$, then we set $c((\eta,\sigma), (\eta',\sigma'))= 0$. \noindent $\bullet$ Finally, if $(\eta,\sigma)= (\eta',\sigma')$, then we set $$ c((\eta,\sigma), (\eta,\sigma))= -\sum_{(\eta',\sigma'),\, (\eta',\sigma')\neq (\eta,\sigma)}c((\eta,\sigma), (\eta',\sigma'))\,. $$ \end{defi} \noindent The difference compared to the rates of definition~\ref{def7} is that, in definition~\ref{def8}, we allow to reverse the spins of a cluster only when it is reduced to a single vertex. This dynamics possesses the same properties as the dynamics of the previous section, namely, it is reversible with respect to the coupling measure IP, its first marginal process on the edges is a Markov process while its second marginal process on the spins is a non-Markovian jump process. The marginal dynamics on the edges differs from the FK dynamics in the following way. An edge can be opened between two different clusters only if one the two clusters is reduced to a single vertex. More precisely, the corresponding rates are the following. For $x\in V$ and $e\in E_x$, for any configuration $\eta$, \begin{equation*} c(\eta,\eta^e) \,=\, \begin{cases} 1-p&\text{ if }\eta(e)=1\,,\cr p&\text{ if }\eta(e)=0\text{ and }\gamma_\eta(e)=1\,,\cr p/2&\text{ if }\eta(f)=0\text{ for $f\in E_x$}\,.\cr \end{cases} \end{equation*} Funnily enough, this fourth dynamics can be seen as a contraction of the first dynamics, which could make only one change at a time. \noindent The following proposition shows that, if the initial condition $(\eta_0,\sigma_0)$ is distributed as $IP$, then the infinitesimal behavior of the marginal on the spins of the process $(\eta_t,\sigma_t)_{t\geq 0}$, whose rates are introduced in definition \ref{def8}, is quite similar to the transition rates of the Glauber dynamics as given in (\ref{marginal}). The difference lies in the factor $|E_x|$. On the cubic lattice $\mathbb{Z}^d$, the factor $|E_x|$ is equal to $2d$ and it is independent of $x$. \begin{pro}\label{umarg} Let $(\eta_t,\sigma_t)_{t\geq 0}$ be a Markov jump process with the transition rates introduced in definition \ref{def8}. Suppose that the pair $(\eta_0,\sigma_0)$ is distributed according to $IP$. Then, for any $x\in V$, for any $\sigma\in \{-1,1\}^V$ and any $s\geq 0$, \begin{equation}\label{ulimite} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma)\,=\, \frac{1}{4} |E_x| p(1-p)^{|\{\,e\in E_x:\delta_\sigma(e)=1\,\}|}\,. \end{equation} \end{pro} \begin{proof} We have, for any $\sigma\in \{-1,1\}^V$, $x\in V$ and $s\geq 0$, \begin{eqnarray}\label{uee1} && \BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) = \sum_{\eta\in \{0,1\}^E}\sum_{\eta'\in \{0,1\}^E}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta',\eta_s=\eta|\sigma_s=\sigma){\nonumber}\\ &&= \sum_{\eta\in \{0,1\}^E}\sum_{\eta'\in \{0,1\}^E}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta'|\sigma_s=\sigma, \eta_s=\eta) \BBp( \eta_s=\eta|\sigma_s=\sigma). \end{eqnarray} Recall that $(\eta_t,\sigma_t)_{t\geq 0}$ is a Markov jump process with transition rates $c((\eta,\sigma),(\eta',\sigma'))$, thus \begin{eqnarray}\label{uee2} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x, \eta_{t+s}=\eta'|\sigma_s=\sigma, \eta_s=\eta) \,=\, c((\eta,\sigma),(\eta',\sigma^x))\,. \end{eqnarray} Using the expression of the rates $c((\eta,\sigma),(\eta',\sigma'))$ given in definition \ref{def8}, together with formulas~(\ref{uee1}) and (\ref{uee2}), we obtain that, for any $s\geq 0$, \begin{multline}\label{try} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) \,=\, \sum_{\eta\in \{0,1\}^E} \sum_{e\in E_x}\left(\prod_{f\in E_x\setminus\{e\}}\BBone_{\eta(f)=0}\right) \cr \times \frac{1}{4} \big( (1-p)\BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + p \BBone_{\eta(e)=0} \BBone_{(\eta^e,\sigma^x)\in \cC} \big) \BBp( \eta_s=\eta|\sigma_s=\sigma) \,. \end{multline} The measure $IP$ is, by proposition \ref{prorev1}, a reversible and a stationary measure for the process $(\eta_t,\sigma_t)_{t\geq 0}$. Hence, if $(\eta_0,\sigma_0)$ is distributed as $IP$, then, for any $s\geq 0$, $(\eta_s,\sigma_s)$ is also distributed as $IP$. Consequently, under the hypothesis of proposition \ref{umarg}, we have \begin{equation}\label{fracte} \BBp( \eta_s=\eta|\sigma_s=\sigma)=\frac{\BBp( \eta_s=\eta,\sigma_s=\sigma)}{\BBp(\sigma_s=\sigma)}=\frac{IP(\eta,\sigma)} {\sum_{\eta'\in \{0,1\}^E}IP(\eta',\sigma)}\,. \end{equation} Equation~\eqref{transf} yields that \begin{equation}\label{fructe} {\sum_{\eta'\in \{0,1\}^E}IP(\eta',\sigma)} \,=\, \frac{1}{Z} (1-p)^{|\{\,e\in E:\delta_\sigma(e)=0\,\}|}\,. \end{equation} Reporting~\eqref{fracte} and~\eqref{fructe} in \eqref{try}, we conclude that \begin{multline} \label{atx} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) \, = \, {Z} (1-p)^{-|\{\,e\in E:\delta_\sigma(e)=0\,\}|}\times \cr \sum_{\eta\in \{0,1\}^E} \sum_{e\in E_x}\left(\prod_{f\in E_x\setminus\{e\}}\BBone_{\eta(f)=0}\right) \frac{1}{4} \big( (1-p)\BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} + p \BBone_{\eta(e)=0} \BBone_{(\eta^e,\sigma^x)\in \cC} \big) IP(\eta,\sigma) \,. \end{multline} We compute these sums as follows. We have, for $e\in E_x$, $$\displaylines{ \sum_{\eta\in \{0,1\}^E} \left(\prod_{f\in E_x\setminus\{e\}}\BBone_{\eta(f)=0}\right) \BBone_{\eta(e)=1} \BBone_{(\eta,\sigma)\in \cC} IP(\eta,\sigma) \,=\, \hfill\cr \hfill \frac{1}{Z} (1-p)^{|E_x|-1 +|\{\,e\in E\setminus E_x:\delta_\sigma(e)=0\,\}|} p\,\delta_\sigma(e)\,,\cr \sum_{\eta\in \{0,1\}^E} \left(\prod_{f\in E_x\setminus\{e\}}\BBone_{\eta(f)=0}\right) \BBone_{\eta(e)=0} \BBone_{(\eta^e,\sigma^x)\in \cC} IP(\eta,\sigma) \,=\,\hfill\cr \hfill\frac{1}{Z} (1-p)^{|E_x| +|\{\,e\in E\setminus E_x:\delta_\sigma(e)=0\,\}|} \big(1-\delta_\sigma(e)\big)\,. }$$ Reporting the values of these sums in formula~\eqref{atx}, we obtain \begin{multline} \label{ayx} \lim_{t\rightarrow 0}\frac{1}{t}\BBp(\sigma_{t+s}=\sigma^x|\sigma_s=\sigma) \, = \, \frac{1}{4} (1-p)^{-|\{\,e\in E:\delta_\sigma(e)=0\,\}|}\times \cr \sum_{e\in E_x} \Big( (1-p)^{|E_x| +|\{\,e\in E\setminus E_x:\delta_\sigma(e)=0\,\}|} p\,\delta_\sigma(e) + p(1-p)^{|E_x| +|\{\,e\in E\setminus E_x:\delta_\sigma(e)=0\,\}|} \big(1-\delta_\sigma(e)\big)\Big)\cr \,=\, \frac{1}{4} (1-p)^{-|\{\,e\in E:\delta_\sigma(e)=0\,\}|} \sum_{e\in E_x} p(1-p)^{|E_x| +|\{\,e\in E\setminus E_x:\delta_\sigma(e)=0\,\}|} \cr \,=\, \frac{1}{4} \sum_{e\in E_x} p(1-p)^{|\{\,e\in E_x:\delta_\sigma(e)=1\,\}|} \,=\, \frac{1}{4} |E_x| p(1-p)^{|\{\,e\in E_x:\delta_\sigma(e)=1\,\}|} \,. \end{multline} This yields the desired result~\eqref{ulimite}. \end{proof} \section*{Appendix} \noindent Let $(V,E)$ be a finite graph and let $(X_t)_{t\geq 0}=((\eta_t,\sigma_t))_{t\geq 0}$ be a time--homogeneous continuous--time Markov process with state space $\Omega=\{0,1\}^E\times \{-1,1\}^V$ and with infinitesimal generator ${{Q}}$ defined as follows. For any function $f$ defined on $\Omega$ with values in $\mathbb R$, we have $$ {{Q}}f(\eta,\sigma)=\sum_{(\eta',\sigma')\in \Omega} {{q}}((\eta,\sigma),(\eta',\sigma')) \left(f(\eta',\sigma')-f(\eta,\sigma)\right), $$ where ${{q}}((\eta,\sigma),(\eta',\sigma'))$ are the transition rates defined by, for $(\eta,\sigma)\neq (\eta',\sigma')$, $$ {{q}}((\eta,\sigma),(\eta',\sigma'))= \lim_{t\rightarrow 0^+}\frac{1}{t} \BBp\left(X_t=(\eta',\sigma') | X_0=(\eta,\sigma) \right))\,, $$ and for $(\eta,\sigma)=(\eta',\sigma')$, \begin{equation}\label{E} {{q}}((\eta,\sigma),(\eta,\sigma))= - \sum_{(\eta',\sigma'):\, (\eta,\sigma)\neq (\eta',\sigma') }{{q}}((\eta,\sigma),(\eta',\sigma')). \end{equation} Our goal is to present some conditions on the transition rates ${{q}}((\eta,\sigma),(\eta',\sigma'))$ under which $(\sigma_t)_{t\geq 0}$ or $(\eta_t)_{t\geq 0}$ are Markov processes. This is a lumpability problem discussed in Ball and Yao (1993). We first recall the following definition of lumpability. \begin{defi} Let $\Omega$ be a countable set, $\{\Omega_1,\cdots,\Omega_r\}$ be a partition of $\Omega$ and $h$ the function from $\Omega$ to $\{1,\cdots,r\}$ defined by $h(x) = j$ if $x \in \Omega_j$. An homogeneous Markov chain $(X_t)_{t\geq 0}$ with state space $\Omega$ is lumpable with respect to the partition ${\Omega_1,\cdots,\Omega_r}$ if $(h(X_t))_{t\geq 0}$ is an homogeneous Markov chain for every initial distribution $y_0$ on $\{1,\cdots,r\}$ and its transition rates do not depend on the choice of the initial distribution $y_0$. \end{defi} \noindent Let us take the following partition of $\Omega$: $$ \Omega\,= \bigcup_{\sigma\in \{-1,1\}^V} \Omega_{\sigma} \,,\qquad \Omega_{\sigma}\,=\,\big\{\,(\eta,\sigma)\in \Omega: \eta\in \{0,1\}^E\,\big\}\,. $$ In this case, the function $h$ is the projection from $\Omega$ to $\{-1,1\}^V$ defined by $h((\eta,\sigma))=\sigma$. Our purpose is to obtain necessary and sufficient conditions under which the Markov process $((\eta_t,\sigma_t))_{t\geq 0}$ is lumpable with respect to the partition $(\Omega_{\sigma})_{\sigma \in \{-1,1\}^V}$. \\ \\ It is well known (see for instance Theorem 3.1 in Ball and Yao (1993)\footnote{Ball and Yao's result is stated under their condition 2.2 which is satisfied in the actual context. }) that a Markov process is lumpable with respect to the partition $(\Omega_j)_{1\leq j\leq r}$ if and only if there exist positive real numbers $\lambda_{i,j}$, $1\leq i,j\leq r$, such that $$ \forall i\neq j\,,\quad \forall \, x\in \Omega_i\,,\qquad \sum_{y\in \Omega_j}q_{x,y}= \lambda_{i,j}\,. $$ In order to simplify the reading of this paper, we adapt this criterion of lumpability to the case of the partition $(\Omega_{\sigma})_{\sigma \in \{-1,1\}^V}$ and the Markov process $((\eta_t,\sigma_t))_{t\geq 0}$. We summarize it in the following theorem. \begin{theo}\label{theoball} Let $((\eta_t,\sigma_t))_{t\geq 0}$ be a continuous-time Markov process with finite state space $\{0,1\}^E\times \{-1,1\}^V$ and transition rates ${{q}}((\eta,\sigma),(\eta',\sigma'))$. Then the following statements are equivalent: \begin{enumerate} \item[(i)] For any $\sigma\neq \sigma'$, there exist positive numbers $c(\sigma,\sigma')$ such that $$ \forall\, \eta\in \{0,1\}^E\,,\qquad \sum_{\eta'\in \{0,1\}^E}q((\eta,\sigma), (\eta',\sigma'))= c(\sigma,\sigma')\,. $$ \item[(ii)] The spin marginal $(\sigma_t)_{t\geq 0}$ of the Markov process $((\eta_t,\sigma_t))_{t\geq 0}$ is a Markov process for every initial distribution $\sigma_0$ on $\{-1,1\}^V$ and its transition rates $c(\sigma,\sigma')$ do not depend on the choice of the initial distribution of $\sigma_0$. \end{enumerate} \end{theo} \noindent The analog of Theorem \ref{theoball} holds for the edge marginal process of $((\eta_t,\sigma_t))_{t\geq 0}$. For a proof, we refer to Theorem 3.1 in Ball and Yeo \cite{Ball}. We also recall Theorem 3.2 of Ball and Yeo \cite{Ball}, that we adapt to our context. \begin{theo}\label{theoball2} Suppose that the transition rates $q((\eta,\sigma), (\eta',\sigma'))$ of the Markov process $(\eta_t,\sigma_t)_{t\geq 0}$ satisfy $$ q((\eta,\sigma), (\eta',\sigma'))\neq 0 \,\,\,\Longrightarrow\,\,\, (\eta,\sigma)=(\eta',\sigma')\,\,{\mbox{or}}\,\, (\eta=\eta'\,\,{\mbox{and}}\,\,\sigma\neq \sigma') \,\,{\mbox{or}}\,\, (\eta\neq \eta'\,{\mbox{and}}\,\,\sigma= \sigma'). $$ Then $(\eta_t)_{t\geq 0}$ and $(\sigma_t)_{t\geq 0}$ are both Markov processes if and only if they are mutually independent. \end{theo} \bibliographystyle{plain}

9f0fea7dc0b3219810b60cf1b59b3df95f5503e8

\section{Introduction} Agile methods are increasingly used in industry as they are established to support projects in their success \citep{serrador2015does}. \citet{cockburn2001agilepeoplefactor} argue that individual competency is an important success factor in agile projects. In agile methods ``the emphasis on people and their talent, skill, and knowledge becomes evident.'' Even on team-level, they argue that the emphasis is again ``on competency rather than process.'' Literature suggests that we progress through two major stages during the development of a cognitive skill, a declarative knowledge stage and a procedural knowledge stage \citep{anderson1982acquisition}. While the former can be acquired by reading text books (e.g.\ learn how to lead a team), the latter, the procedural knowledge, can only be acquired in process (e.g.\ by actually leading a team and learning from mistakes). Hence, it seems that success in agile projects depends on individual skills, that are developed in individuals over time. Many studies in software engineering have focused on explaining the individual skills (see e.g.\ \citet{turley1994identifying}), which implies that the individual non-technical skills are believed to predict team-level performance in relation to collaborative aspects. However, no studies have looked at explicit connections between agile practices and individual non-technical skills. As we assume technical skills to always be a precondition for a successful software development endeavour, non-technical skills seem to be important in such endeavours as well \citep{Bender2014}. Such non-technical individual skills have been stated as especially important in agile practices, since they focus more on individuals than processes \citep{cockburn2001agilepeoplefactor}. Therefore, this paper investigates the connections between (i.e.\ the predictive power of) thirteen self-assessed non-technical individual skills and the intended mature use of a set of eight common agile practices. In this paper, we define \emph{practice maturity} by assuming it to be possible to describe by an evolutionist model comprising of a progressive and directional set of changes. As the practices mature, they increase in perfection or complexity over time \citep{king1984evolution}. We assume the end goal of the development of the agile practices to be to what degree a practice is implemented in its intended way, as measured by the degree of agreement to items in the \emph{perceptive agile measurement} \citep{so}. We did not investigate the use of agile practices by having subjects tick the practices they use from a list, but instead used the perceptive agile measurement \citep{so} to assess the degree of agile behavior prescribed by an agile practice. We assume different degrees of such agile behavior to be equivalent to different levels of agile practice maturity. The perceptive agile measurement was created to assess the social-psychological effect of agile practices, which provides a higher resolution of the actual behaviour shown in these agile teams. The sample consisted of agile team members from seven organizations in Brazil and The Netherlands. We set up eight hypotheses regarding the associations between all individual skills and each agile practice. In order to assess the hypotheses, we conducted a diversity of analyses: First, we checked our survey data for normality. Second, we ran eight independent ANOVAs. Third, we built three new regression models for the significant models and analyzed their effect sizes. The results show non-significant or negligible effect sizes for all our analyses. We therefore conclude that looking at non-technical individual skills is not the optimal level of analysis when wanting to understand the use of agile practices, and argue for using the ``agile team'' as the level of analysis instead. To clarify the outline of this paper, the next section (Section~\ref{sec:related_work}) presents related work with regards to individual and team level skills in agile software development. Section~\ref{constructs} gives an overview of the measurements and constructs (i.e., non-technical individual skills and agile practice) used in this paper. Section~\ref{sec:methodology} depicts the method used to measure agility and non-technical individual skills and correlate these measurements in a set of multiple linear regression models, Section~\ref{sec:results} presents the results, which are discussed in Section~\ref{sec:discussion}, followed by conclusions and future work in Section~\ref{sec:conclusions}. \section{Related Work}\label{sec:related_work} In this section, we provide a definition of what skills are, provide an overview of research on individual skills in the software engineering context, and present the few studies looking at a higher level of analysis in relation to skills in the software engineering domain. \subsection{What Is a Skill?}\label{skillsdef} We define a skill as ``an ability to do something well; expertise'' \citep{dictionary} in this study in its broader sense. However, in order to understand what this means, we need to look at related research in what characterizes a skill and how they are acquired. As mentioned in the introduction, the acquisition of a cognitive skill can be described by two major stages: a declarative knowledge stage and a procedural knowledge stage \citep{anderson1982acquisition}. These stages are intimately connected to what \citet{argyris2000double} calls single- and double-loop learning. We can learn to repeat new information without integrating it on a deeper level. With such shallow learning, we can not apply our acquired knowledge in new situations and we fail to translate the new knowledge to similar cases \citep{argyris2000double}. Therefore, we asked the participants of this study about their own satisfaction in relation to specific skills, which then are more in relation to the procedural knowledge (double-loop) rather than shallower declarative (single-loop) knowledge in their work context. In this study, we also define a team-level skill as the capability of the entire team in relation to non-technical skills, i.e.\ even if all members can not plan well for themselves, the team as an entity might have good planning skills anyways as a product of collaboration and reciprocal help. Another scientific field that has gone through similar phases in research is the medical field, and \citet{fletcher2002role} shows how surgical teams have gone from only focusing on technical skills to also realizing the importance of the non-technical skills needed for successful treatment of patients. \citet{fletcher2002role} divides the non-technical skills into two categories; (1) cognitive and mental skills (e.g.\ decision-making, planning, etc.) and (2) social or interpersonal skills (e.g.\ team-working, communication, leadership, etc.). This research highlights the importance of individual non-technical skills, which has also been common in the software engineering research and is presented next. \subsection{Research on Individual Skills in Agile Software Development}\label{skills} Agile methods are described as having a strong focus on individuals and their skills \citep{cockburn2001agilepeoplefactor}. \citet{cockburn2001agilepeoplefactor} conclude that individual skills seem more important than team characteristics -- even on the team level they argue that the emphasis is on individual competency. Strengths as well as weaknesses of individual skills need to be taken into consideration as both can have an influence on the success of an organization. \citet{conboy2011peopleoverprocess} report an increased exposure of capabilities and reliance on social skills. They report that exposing weaknesses of team members can often be counter-productive and even highly respected and performing team members can be bullied, challenging existing organizations. Furthermore, an understanding of the skills necessary within a team can help team members in their development. In agile software development, the individual skills of software engineers need to be considered, according to many studies. For example, \citet{turley1994identifying} identified 38 essential competencies of software engineers. Among the top ten are: (1) Team Oriented, (2) Seeks Help, (3) Helps Others, (4) Use of Prototypes, (5) Writes\slash Automates Tests with Code, (6) Knowledge, (7) Obtains Necessary Training\slash Learning, (8) Leverages\slash Reuses Code, (9) Communication/Uses Structured Techniques for Communication, and (10) Methodical Problem Solving. In addition to the general skills that a software engineer should possess, skills of requirements engineers play an important role in agile software development. The focus in agile software development is on the customer who decides on what is of value and who is supposed to be on site to clarify requirements \citep{Fricker2010}. Hence, contrary to traditional RE, in agile development any member from the development team can directly interact with the customer and collect requirements \citep{Sillitti2005}. Furthermore, RE represents an area in which it is especially important to consider social\slash non-technical skills and theories \citep{Malley2004}. Hence, non-technical skills of requirements engineers are important and required for each individual team member in agile software development. Negotiation is an inevitable element in RE \citep{Nuseibeh2000, Grunbacher2005}. For example in requirements elicitation and analysis it plays a major role as it supports handling stakeholder conflicts concerning requirements \citep{Ahmad2011}. Negotiation promotes a shared vision, shared knowledge, and cooperation among stakeholders \citep{Ahmad2011}. Furthermore, communication is an inevitable skill in RE \citep{Coughlan2003}. Team members need to communicate with each other for many reasons, i.e. bugs discussion, code issues, code reviews, code refactoring, code synchronization, coordination, management, support issues, sprint planning, quality, user story clarification and user story negotiation \citep{Inayat2015a}. In agile requirements engineering most activities depend on communication between different parties, their input and judgment. Hence, it is highly dependent on the skills of the team members. Further skills for phased as well as agile RE are: 1) Dividing bigger tasks into small ones, 2) giving up control -- as code is developed and changes by different people, 3) writing meaningful tests, 4) conversation, 5) object-oriented design, 6) fast cycle times \citep{Kovitz2003}. \subsection{Research on Team-Level Skills in Agile Software Development}\label{org} Most existing research focuses on individual skills, as presented in Section~\ref{skills}. A few studies also report about the importance of non-technical skills on a team level (e.g., with regards to team work and setting up teams). \citet{Lalsing2012} identified the underlying people factors for a team to be effective in agile software development, and several agile projects were in the study. Results showed that for projects exceeding the project budget, issues were e.g.\ related to team communication and collaboration (i.e.\ trust and interaction). They concluded that it is crucial to select the ``right people for the right team,'' and not only the ``right people.'' \citet{Tanner2014} study success and failure of agile projects in waterfall environments. From studying two cases of agile projects and literature on this topic, they concluded that the following factors have an influence on the projects success or failure of agile projects: 1) Culture, 2) Customer Involvement and Mandate, 3) Stakeholder Involvement and Buy-In, 4) Team Structure and Team Logistics, 5) Project Type and Project Planning, and 6) Skill Level and Attitude of Team Members. \citet{Crowder2015} identified communication, coordination, trust and team orientation as the most important team factors for distributed agile teams, based on survey data. From a systematic literature review, they identified team orientation, shared leadership, mutual performance monitoring, backup behavior, feedback, team autonomy, team learning, coordination, communication, trust, collective culture, ease of use of technology, and team familiarity as teamwork factors. \citet{grenjss2} also found that team maturity (from a group dynamics perspective) is a key factor in the success of building agile teams in large organizations. A challenge when trying to understand productivity in general is to get data from the right level of analysis \citep{hackman2003learning}. Software engineering research on human factors can learn a great deal from the journey social psychology has done from the individual to the group as a level of analysis (cf.\ \citet{hogg2000we}), a journey that has largely been repeated by organizational science \citep{klein2000micro}. If we want to understand agile practices, the individual skills level might be too much on a micro level based on research in other fields, and we could study the agile teams and how software developers and other team-members behave collectively instead, i.e.\ explained variance in data might come from the meso level of the team as an entity. Recent research in social psychology \citep{woolley2010evidence,engel2014reading} has shown that, in general, the performance of teams on a diversity of tasks is set on group-level independent of the intelligence of the individuals. The intelligence of groups have instead been shown to be more dependent of social sensitivity (i.e.\ a person's ability to read emotions in facial expression), and conversational turn-taking (i.e.\ groups were less collectively intelligent if a few individuals dominated the conversations). Such findings contradict the conclusions drawn in a lot of software engineering research on the importance of non-technical individual skills (see for example \citet{turley1994identifying}). To clarify such contradictions, we conduct a study in the software engineering domain aiming to shed light on this contradiction of individual vs.\ group level. As software engineering mainly looks at individual level, this is the level we chose to investigate in this current study. According to \citet{hackman2003learning}, one should preferably look at the macro (i.e.\ organizational) level in addition to the micro and meso levels. While we recognize that variance possibly could be explained on the organizational level, we focus on discussing the micro and meso levels in this study. \section{Measurements, Constructs, and Research Hypotheses}\label{constructs} In this section we describe the measurements used in the current study and their operationalization based on previous studies. We first present theory on common agile practices and how they can be measured, and then we suggest a measurement of self-assessed non-technical individual skills based on such studies in the agile context. \subsection{Common Agile Practices} Agility as a concept can be difficult to delineate \citep{laanti2013definitions}. Agility emerged in practice and has been discussed across different scientific domains such as manufacturing and logistics \citep{boothharmer1994agile}, business management \citep{vanOosterhout2010business}, information systems literature \citep{Conboy2009}, and sports science \citep{sheppard2006agility}, which makes it difficult to define. Fuelled by the difficulties in definition, an analysis by \citet{laanti2013definitions} suggests to look at agility as a set of concrete practices to understand agility. Other studies have followed that track and investigated the usage and perceptions of practices perceived as agile within software development teams \citep{williams,so}. In line with \citet{salvato2009capabilities} we believe that such concrete, routinized activities have a huge impact on the effectiveness and sustainability of project management processes -- and, as such, that there must be some kind of behaviour that can be considered ``more'' agile than other behaviour connected to more traditional project management groups. Early Scrum literature generally describes to the following practices \citep{schwaberBeedle2003agile}: (1) Collocated scrum teams, (2) Daily scrum stand-up meetings (3) Iteration planning in sprint planning meetings, (4) Iterative development in sprints, and (5) Sprint reviews. Based on previous research and perceptions of practitioners, \citet{so} constructed an instrument for quantitative analyses of social-psychological effects of the above mentioned agile practices, however categorized a bit differently than those Scrum practices above. They created scales for the eight of the core agile practices, namely (1) Iteration planning, (2) Iterative development, (3) Continuous integration and testing, (4) Stand-up meetings, (5) Customer access, (6) Customer acceptance tests, (7) Retrospectives, and (8) Collocation. The framework of \citet{so} is particularly useful as they provide for the first time a scientifically validated psychometric instrument covering these eight core agile practices, i.e.\ their tool tries to capture agile behaviour in connection to the practices and therefore claims to measure the actual practices. The measured practices Customer access, and Customer acceptance tests, also extend their survey to include aspects of agile requirements engineering, which adds a focus on the entire development chain from planning to delivering in the agile context. We use the questionnaire suggested by \citet{so} to understand the connection of agile practices and individual skills with the reason that we need social-psychological measurements in our study. In the following, we will elaborate on the eight practices included in the instrument by \citet{so}. Furthermore, in the text boxes below each of the depicted practice we reproduce the exact items (i.e., the agile practice and sub-questions used to cover this agile practice) used in our questionnaire. \paragraph{Iteration Planning} In a Collaborative planning workshop the deliverables and scope of an iteration is defined with all team members being present. It is executed at the beginning of each iteration, sometimes also referred to as \emph{Planning Game}. The practice generally consists of two stages: In the first stage, requirements are gathered in the form of user stories to serve as a medium for discussions between the customer and the developers. In the second stage, the stories are revised, estimated and prioritized into an iteration backlog \citep{liu2005environment}. The active participation of technical team members in definition as well as estimation of user stories is considered as an indication for a mature application of the practice \citep{wang2012assimilation,sharp2004ethnographic}. \begin{framed} \scriptsize \textbf{Iteration Planning:} (1) All members of the technical team actively participated during iteration planning meetings. (2) All technical team members took part in defining the effort estimates for requirements of the current iteration. (3) When effort estimates differed, the technical team members discussed their underlying assumption. (4) All concerns from team members about reaching the iteration goals were considered. (5) The effort estimates for the iteration scope items were modified only by the technical team members. (6) Each developer signed up for tasks on a completely voluntary basis. (7) The customer picked the priority of the requirements in the iteration plan. \end{framed} \paragraph{Iterative Development} Routinized delivery of sub-results (working software) in short and iterations of fixed length~\citep{petersen2}. Although the practice has been popularized with the dawn of agile methods, the application of iterative software development dates as far back as the mid-1950s \citep{larman2003iterative}. Short iterations of 30 days or less together with continuous integration, have been found as the two practices considered most essential for a team to be considered agile \citep{williams}. Iterative development is a shared practice in agile methods as well as in user-centered design~\citep{chamberlain2006towards}. \begin{framed} \scriptsize \textbf{Iterative Development:} (1) We implemented our code in short iterations. (2) The team rather reduced the scope than delayed the deadline. (3) When the scope could not be implemented due to constraints, the team held active discussions on re-prioritization with the customer on what to finish within the iteration. (4) We kept the iteration deadlines. (5) At the end of an iteration, we delivered a potentially shippable product. (6) The software delivered at iteration end always met quality requirements of production code. (7) Working software was the primary measure for project progress. \end{framed} \paragraph{Continuous Integration and Testing} \citet{holck2003continuous} define continuous integration as follows: (1) access of development team members to add contributions to the development version at any time, and (2) obligation of team members to integrate their own contributions properly. In order to enable such a continuous integration of ongoing development into a software system, the practices are ofter linked to (automated) testing methods to enable a timely verification of the system \citep{hellmann2012agile}. \begin{framed} \scriptsize \textbf{Continuous Integration and Testing:} (1) The team integrated continuously. (2) Developers had the most recent version of code available. (3) Code was checked in quickly to avoid code synchronization/integration hassles... (4) The implemented code was written to pass the test case. (5) New code was written with unit tests covering its main functionality. (6) Automated unit tests sufficiently covered all critical parts of the production code. (7) For detecting bugs, test reports from automated unit tests were systematically used to capture the bugs. (8) All unit tests were run and passed when a task was finished and before checking in and integrating. (9) There were enough unit tests and automated system tests to allow developers to safely change any code. \end{framed} \paragraph{Stand-up meetings} Frequent team coordination meetings in which team members provide a status update to their colleagues. The meetings are generally hold standing up and are time boxed to 5-15 minutes to frame its short and focused nature. Each coaching session starts with a team stand-up where each group was asked the three common questions: ``What have you done since the last meeting?'' ``What are you planning on doing until the next meeting?'' and ``What issues and impediments are you facing that prevent you from accomplishing these things?'' An ethnographic account of the practice is provided by \citet{sharp2004ethnographic}. \citet{stray2012investigating} investigated the application of daily stand-up team coordination meetings. They found that only 24\% of each of the meetings they studied focused on coordination. Rather, 35\% of the meeting time was spent on content-discussions elaborating problem issues and discussing possible solutions. \begin{framed} \scriptsize \textbf{Stand-Up Meetings:} (1) Stand up meetings were extremely short (max. 15 minutes). (2) Stand up meetings were to the point, focusing only on what had been done and needed to be done on that day. (3) All relevant technical issues or organizational impediments came up in the stand up meetings. (4) Stand up meetings provided the quickest way to notify other team members about problems. (5) When people reported problems in the stand up meetings, team members offered to help instantly. \end{framed} \paragraph{Customer Access} Availability of customers for product feedback and clarification of requirements is integral to effective agile teams and has been found as one of the critical success factors when implementing agile methods \citep{misra2009identifying,chow2008survey}. Especially when moving away from traditional software development customer access can be a challenge as customers might not be used to close interaction \citep{misra2009identifying}. \begin{framed} \scriptsize \textbf{Customer Access:} (1) The customer was reachable. (2) The developers could contact the customer directly or through a customer contact person without any bureaucratic hurdles. (3) The developers had responses from the customer in a timely manner. (4) The feedback from the customer was clear and clarified his requirements or open issues to the developers. \end{framed} \paragraph{Customer Acceptance Tests} Acceptance tests defined by the customer present a means to the developers to determine which iteration goals have been achieved at the end of each iteration \citep{so2010making}. \begin{framed} \scriptsize \textbf{Customer Acceptance Tests:} (1) How often did you apply customer acceptance tests? (2) A requirement was not regarded as finished until its acceptance tests (with the customer) had passed. (3) Customer acceptance tests were used as the ultimate way to verify system functionality and customer requirements. (4) The customer provided a comprehensive set of test criteria for customer acceptance. (5) The customer focused primarily on customer acceptance tests to determine what had been accomplished at the end of an iteration. \end{framed} \paragraph{Retrospectives} Workshop at the end of each iteration to improve the process and incorporate successful practices for the next iteration. In order to enable continuous improvement of the practices applied each team member lists ``What went well'' and ``What could be improved'' \citep{maham2008planning}. The impact, however, dependent largely on their implementation \citep{mchugh2012agile}. For example, retrospectives should take place at the end of each iteration and systematically assign all improvement points to responsible individuals \citep{so2010making}. \begin{framed} \scriptsize \textbf{Retrospectives:} (1) How often did you apply retrospectives? (2) All team members actively participated in gathering lessons learned in the retrospectives. (3) The retrospectives helped us become aware of what we did well in the past iteration(s). (4) The retrospectives helped us become aware of what we should improve in the upcoming iteration(s). (5) In the retrospectives (or shortly afterward), we systematically assigned all important points for improvement to responsible individuals. (6) Our team followed up intensively on the progress of each improvement point elaborated in a retrospective. \end{framed} \paragraph{Collocation} Close proximity of development team members is reported as one of the critical success factors when implementing agile methods \citep{lindvall2002empirical}. It is recognized as one of the important vehicles for successful communication and knowledge creation \citep{misra2009identifying}. \begin{framed} \scriptsize \textbf{Collocation:} (1) Developers were located majorly in... (2) All members of the technical team (including QA engineers, db admins) were located in... (3) Requirements engineers were located with developers in... (4) The project/release manager worked with the developers in... (5) The customer was located with the developers in... \end{framed} \subsection{Common Non-Technical Skills} We use common denominators on skills derived from previous studies as presented in Section \ref{skills}. In that analysis, we found 13 non-technical individual skills that have been identified in at least two papers as important for software developers to be successful. Table \ref{skillstable} presents the non-technical skills used in this study and supporting literature depicting the importance of this skill. \begin{table \renewcommand{\arraystretch}{1.5} \caption{Non-technical individual skills used in the present study} \label{skillstable} \centering \begin{tabular}{ll} \hline \bfseries Non-technical skills & \bfseries Studies including skill \\ \hline (1) Communication Skills & \citep{lee2006managers, turley1994identifying, Coughlan2003, Lalsing2012, Kelle2015, napier2009projectskills, Crowder2015} \\ \hline (2) Teamwork Skills & \citep{lee2006managers, turley1994identifying, cockburn2001agilepeoplefactor, Tanner2014, napier2009projectskills, Crowder2015} \\ \hline (3) Collaboration Skills & \citep{turley1994identifying, Lalsing2012, Crowder2015} \\ \hline (4) Ability to Meet Project Goals & \citep{turley1994identifying, chen2001validation, Crowder2015} \\ \hline (5) Customer Orientation & \citep{turley1994identifying, Fricker2010, Kelle2015, napier2009projectskills, Tanner2014} \\ \hline (6) Requirements Management Skills & \citep{Sillitti2005, Ahmad2011} \\ \hline (7) Planning Skills & \citep{lee2006managers, turley1994identifying, napier2009projectskills, Tanner2014} \\ \hline (8) Leadership Skills & \citep{lee2006managers, Kelle2015, napier2009projectskills, Crowder2015, Cohn2004} \\ \hline (9) Decision-Making Skills & \citep{napier2009projectskills, janis1977dma} \\ \hline (10) Business-Minded Skills & \citep{turley1994identifying, Crowder2015} \\ \hline (11) Problem-Solving Skills & \citep{turley1994identifying, napier2009projectskills} \\ \hline (12) Organizing Skills & \citep{lee2006managers, turley1994identifying} \\ \hline (13) Negotiation Skills & \citep{turley1994identifying, Grunbacher2005, Ahmad2011, Shell2001, Crowder2015} \\ \hline \end{tabular} \end{table} \subsection{Research Hypotheses} Based on the literature, in this current paper, we investigate the assumption that the use of agile practices is positively connected to non-technical individual skills. All the items were self-assessed by team members in agile software development projects. The hypotheses were therefore the following: \begin{itemize} \item \textbf{H$_1$:} The mature use of the agile practice Iteration Planning is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_2$:} The mature use of the agile practice Iterative Development is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_3$:} The mature use of the agile practice Continuous Integration and Testing is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_4$:} The mature use of the agile practice Stand-Up Meetings is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_5$:} The mature use of the agile practice Customer Access is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_6$:} The mature use of the agile practice Customer Acceptance Tests is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_7$:} The mature use of the agile practice Retrospectives is positively associated with agile team members' self-assessed non-technical skills. \item \textbf{H$_8$:} The mature use of the agile practice Collocation is positively associated with agile team members' self-assessed non-technical skills. \end{itemize} \section{Method}\label{sec:methodology} In order to investigate the connections between the individual skills and agile practices, we created a survey using the agile practices suggested by \citet{so} and our aggregation of non-technical individual skills (both derived in the previous section). \subsection{Participants} All the participants were members of agile teams in the participating companies (as stated by our company contacts). We explicitly asked all team members to answer the survey, but to skip questions they were not able to assess. As a consequence, most respondents were intimately connected to development code in one way or the other. The reason why we did not asked specifically for only software developers, were that some employees that conduct software development might have another title, like for example ``system engineer.'' This study was carried out at six organizations in total, three companies in Brazil and three in The Netherlands (two companies and one public sector IT department). These companies were selected because the first and third authors had direct or indirect research connections to people within the organizations. We wanted the participating organizations to be as diverse as possible in order to being able to generalize our study to the broader population of agile team members in the software development context, i.e.\ we intended to survey agile team members from different continents, different sizes of organization, as well as the public and private sectors (see Table \ref{participants}). \begin{landscape} \begin{table} \renewcommand{\arraystretch}{1.5} \caption{The Participating Organizations} \label{participants} \centering \begin{tabular}{cccccccccccc} \hline \bfseries Organization & \bfseries Country &\bfseries \# of employees &\bfseries \# of responses \\ \hline Company 1 & Brazil & ~100 & 22 \\ \hline Company 2 & Brazil & ~5,000 & 57 \\ \hline Company 3 & Brazil & ~35 & 11 \\ \hline Company 4 & The Netherlands & ~3,500 & 4 \\ \hline Company 5 & The Netherlands & ~50,000 & 7 \\ \hline Public Sector Organization & The Netherlands & ~5,000 & 12 \\ \hline & & & Total: 113* \\ \end{tabular} \end{table} *After list-wise deletion due to missing values, we had at least 99 valid responses to use in our regression analyses. \end{landscape} The Brazilian sub-sample contained data points from IT departments at a large on-line media and social networking enterprise with around 5,000 employees, a smaller software consultancy company with around 35 employees, and a company that provides programming courses to individuals and companies, with around 100 employees. All the agile team members received the surveys via their managers but their replies were anonymous. The response rate was 92\% for the Brazilian sub-sample due to the fact that they were collected on-site in paper form. The Dutch sub-sample consisted of four groups across three organizations: an IT service provider (around 3,500 employees), banking and financial services (around 50,000 employees), and an IT service department in the public sector (around 5,000 employees). The response rate for the Dutch sub-sample was 81\%. All the participating companies said they use an agile development approach in the software development conducted, but with teams of different maturity levels in that process. The total number of respondents in the first measurement was 158 agile team members. \subsection{Survey items} In the survey these skills were separately put as questions in the following form: ``How satisfied are you with your [skill]?'' The reason why we used personal satisfaction of a certain non-technical skill was that, in agile teams, such skills are perceived as utterly important. Therefore, a satisfaction of a skill should be more related to a perceived peer-evaluation by a team-member than if asked to only rate their own individual skills. In addition, as mentioned previously, we aimed at investigating procedural rather than declarative skills, and in order for a participant to rate their individual skill high in context, we believe such rating will be more in relation to their declarative knowledge. We used the survey suggested by \citet{so} to measure the mature use of the eight agile practices included in their study, i.e.\ higher scores on their survey imply higher maturity of a practice since the questions are in relation to the intended, and therefore mature, use of a practice. The entire construct used is presented in Section~\ref{constructs}. Due to all the different definitions and ambiguity of ``agility'' \citep{laanti2013definitions}, we chose this survey since it captures the social-psychological behavior in connection to what the different practices try to achieve. It is also the only tool we have found that is validated through a factor analysis \citep{fabrigar} and a reliability analysis (using the Cronbach's $\alpha$ \citep{cronbach}) with a sample of $N=227$ \citep{so}. The agile items in the questionnaire were assessed on a 7-point Likert scale (1 = never and 7 = always), with one exception being the Collocation items that were rated from 1 = the same room to 5 = different timezones. These scales were used for the simple reason that these measurements were developed and validated using those exact scales. The skills were assessed from 1 = completely dissatisfied and 9 = completely satisfied. In order to validate our results further, we also built a regression model using perceived code quality as a response variable. The perceived code quality aspect was measured using the single question: ``How would you rate the code quality in your product(s)?'' rated from 1 (very poor) to 4 (excellent). The reason for this validation was to investigate if non-technical individual skills were connected to a completely different aspect of the software developed (i.e.\ other than the agile practices). If that were the case, we would have support for the usefulness of measuring non-technical individual skills outside the scope of agile practices. \subsection{Data collection and analysis} The questionnaires were distributed in paper form by one of the managers and collected on site by the first and third authors (depending on the country). To evaluate if the data was normally distributed, we plotted frequency histograms for all the multiple linear regression models (for one example, see Figure~\ref{d}). Figure~\ref{d} shows that the residuals are randomly scattered around the regression line. However, we saw some issues with the dependent variable Retrospectives, which turned out to have a set of outliers. When these outliers were removed and the data looked normally distributed, the ANOVA for that category was still not significant, meaning that the outliers did not affect the result. We also checked the variance inflation factor (VIF) for each regression model and all values were below 3, which is acceptable for these kinds of analysis since a common rule of thumb is below 10 \citep{marquaridt1970generalized}. \begin{figure} \centerline{\includegraphics[scale=0.5]{ipfreq.pdf}} \caption{Frequency histogram with Iteration Planning as dependent variable and all the individual skills as factors.} \label{d} \end{figure} In order to investigate the connections between the two concepts we first ran eight ANOVAs with all the skills as factors and the agile practices as response variables one by one. The purpose was to investigate the predictive power of knowing the agile team members perceived individual skills on all the agile practices separately (plus the quality question in its own analysis as a validation). This means that we investigated how much of the variance in a measured agile practice that was explained by the non-technical individual skills together, and therefore, opted to use linear regression analysis with all the skills as factors. It is important to note the differences between predictive and causal models and in this study we only claim the former. However, since we have the theoretical assumption that skills predict agile practices maturity and not the other way around, we have the skills as independent variables and the agile practices measurements as dependent variables. If the ANOVA was significant at an alpha level of 5\%, we proceeded and built a multiple linear regression model to see which skills (factors) were significant. As a measurement of effect size we used $\eta^2$ (often called $R^2$ in regression analysis) for each omnibus test (i.e.\ ANOVA) \citep{coheneffect}. \section{Results}\label{sec:results} To assess the predictive power of non-technical individual skills on the agile practices, we ran eight independent ANOVA omnibus tests of which only the three response variables ``Iteration Planning'' ($F=2.166, p=0.017, N=105$), ``Customer Access'' ($F=2.415, p=0.008, N=102$), and ``Customer Acceptance Tests'' ($F=2.940, p=0.001, N=99$) were significant. All the other ANOVAs using ``Iterative Development'' ($F=1.307, p=0.224, N=103$), ``Continuous Integration and Testing'' ($F=0.664, p=0.792, N=100$), ``Stand-Up Meetings'' ($F=0.862, p=0.595, N=99$), ``Retrospectives'' ($F=0.946, p=0.510, N=101$), and ``Collocation'' ($F=0.968, p=0.488, N=103$) as dependent variables were not significant at an alpha level of 5\%. We therefore failed to reject the null-hypotheses in favour of H$_2$, H$_3$, H$_4$, H$_7$, H$_8$ using the agile practices as response variables, i.e.\ they were not significant at an alpha level of 5\%. Therefore, we also conclude that we have weak support, but still reject the null hypotheses in favour of H$_1$,H$_5$, and H$_6$. Next, we explain why the support was weak. The agile practices Iteration Planning, Customer Access, and Customer Acceptance Tests had significant ANOVAs and we therefore ran further analyses using the significant factors (i.e.\ non-technical individual skills) in order to evaluate the size of the effects found. For these three significant omnibus tests, we built new models based on the significant factors and calculated effect sizes, which were found to be low or very low. The results were: Iteration Planning (adjusted $R^2=11.7\%$), using planning and teamwork skills as factors (see Table~\ref{planning}), Customer Access (adjusted $R^2=6.0\%$), using business-minded skills and organizing skills as factors (see Table~\ref{cust}), and finally Customer Acceptance Tests (adjusted $R^2=4.6\%$), using organizing skills as a factor (see Table~\ref{acc}). The only model with explained variance over ten percent was predicting ``Iteration Planning'' by using planning and teamwork skills skills. The regression models built using Customer Access and Customer Acceptance Tests had effect sizes under 10\%, showing low predictive power of non-technical individual skills, even though we rejected the null hypotheses. Higher teamwork and planning skills were connected to better iteration planning, which is the only result that makes sense. Actually, organizing skills were negatively correlated to Customer Access and Customer Acceptance Tests, which questions the relevance of the results in general. This would then mean that good organizing skills would be bad for mature Customer Access and Customer Acceptance Tests practice, which seems odd. However, we believe these results also point out that individual satisfaction of e.g.\ organizing skills is a poor predictor of agile maturity of a practice. Additionally, these effect sizes should be seen as irrelevant according to \citet{coheneffect}, while above ten percent is considered only a small effect. Therefore, the significant connections between teamwork and planning skills and iteration planning are the only ones relevant to analyze further. The variance in the measurement of Iteration Planning, could be explained by 11.7\%, which is considered a small effect in these types of studies \citep{coheneffect}. As mentioned before, it makes sense that the individual skills of teamwork and planning would be connected to how well a team plans for an iteration, however, it makes equally much sense that individual communication, collaboration, decision-making, problem-solving, organizing, and negotiation skills would be connected to e.g.\ the team ability to develop iteratively. In addition, the effect was barely over ten per cent, which we still consider much lower that would be the case if all these agile team practices depended on individual non-technical skills. \begin{table} \renewcommand{\arraystretch}{1.5} \caption{Linear Regression Coefficients (Dependent Variable: Iteration Planning with 112 valid cases). Adjusted $R^2=11.7\%$} \label{planning} \centering \begin{tabular}{cccccccccccc} \hline \bfseries Model & \bfseries Unstandardized B &\bfseries Std. Error &\bfseries Standardized B &\bfseries t &\bfseries $p$-value \\ \hline (Constant) & 3.373 & 0.528 & & 6.389 & 0.000* \\ \hline Teamwork Skills & 0.175 & 0.074 & 0.778 & 2.354 & 0.020* \\ \hline Planning Skills & 0.133 & 0.062 & 0.208 & 2.148 & 0.034* \\ \hline {*p$<$.05} \\ \end{tabular} \end{table} \begin{table} \renewcommand{\arraystretch}{1.5} \caption{Linear Regression Coefficients (Dependent Variable: Customer Access with 108 valid cases). Adjusted $R^2=6.0\%$} \label{cust} \centering \begin{tabular}{cccccccccccc} \hline \bfseries Model & \bfseries Unstandardized B &\bfseries Std. Error &\bfseries Standardized B &\bfseries t &\bfseries $p$-value \\ \hline (Constant) & 5.251 & 0.692 & & 7.590 & 0.000* \\ \hline Business-Minded Skills & 0.173 & 0.080 & 0.204 & 2.150 & 0.034* \\ \hline Organizing Skills & -0.181 & 0.077 & -0.222 & -2.342 & 0.021* \\ \hline {*p$<$.05} \\ \end{tabular} \end{table} \begin{table} \renewcommand{\arraystretch}{1.5} \caption{Linear Regression Coefficients (Dependent Variable: Customer Acceptance Tests with 104 valid cases). Adjusted $R^2=4.6\%$} \label{acc} \centering \begin{tabular}{cccccccccccc} \hline \bfseries Model & \bfseries Unstandardized B &\bfseries Std. Error &\bfseries Standardized B &\bfseries t &\bfseries $p$-value \\ \hline (Constant) & 5.811 & 0.757 & & 7.679 & 0.000* \\ \hline Organizing Skills & -0.269 & 0.111 & -0.234 & -2.431 & 0.017* \\ \hline {*p$<$.05} \\ \end{tabular} \end{table} \paragraph{Validating individual skills against perceived code quality} As a validation of the non-technical individual skills measurements we also built a model using the perceived code quality in products as a response variable. The non-technical individual skills showed no connection to the agile team members' perceived code quality ($F=1.172, p=0.314, N=99$), which means that non-technical individual skills also failed to predict the agile team members' perceived quality of the code. \paragraph{Summary of statistical results} To summarize the results above, we first analyzed our survey data for normality and plotted frequency histograms for multiple regression models using non-technical individual skills as factors and the agile practices (one-by-one), and perceived code quality, as response variables. To assess the predictive power of non-technical individual skills on the agile practice and perceived quality, we ran nine independent ANOVA omnibus tests of which only the three response variables ``Iteration Planning,'' ``Customer Access,'' and ``Customer Acceptance Tests'' were significant. For these three significant omnibus tests, we build new models based on the significant factors and calculate effect sizes, which were found to be low or very low. The only model with explained variance over ten percent was predicting ``Iteration Planning'' by using teamwork and planning skills. This result makes sense, however, so would many other predictions which were not significant, and in addition, higher organizing skills were connected to lower Customer Access and Customer Acceptance Tests, which questions the relevance of such a measurement. The data analyses therefore have shown that looking at non-technical individual skills is not the optimal level of analysis when wanting to predict agile maturity. What this means and the implications of the low predictive power of non-technical individual skills in connection agile practices will be discussed next. It is important to, again, highlight the differences between correlation\slash predictive and causal models and in this study we only claim the former. \section{Discussion}\label{sec:discussion} The results show that there is very little predictive power when using self-assessed non-technical individual skills to understand the perceived maturity of the agile practices. Based on previous work in software engineering we would \emph{not} expect such a result. This means that we can not look at self-assessed non-technical individual skills when trying to predict the intended and mature use of agile practices. If there is little value in looking at individual non-technical skills when understanding or improving agile practices, what is then the option? As mentioned in the introduction, \citet{turley1994identifying} identified 38 essential competencies of software engineers on different abstraction levels, which we interpret as an indication of the issues shown in our present study with using non-technical individual skills in order to increase the use of agile practices in software development teams, i.e.\ they are simply too many. In accordance with \citet{Tanner2014} and \citet{Crowder2015} who investigated agile project success in relation to team orientation, shared leadership, backup behavior, feedback, team autonomy, team learning, coordination, communication, trust, collective culture, team familiarity, customer involvement and mandate, stakeholder involvement and buy-in, and team structure and team logistics, we have also found empirical support for looking at other levels of analysis than the individual, when wanting to optimize the benefits from an agile approach. Since project success is connected to agility \citep{serrador2015does} and \citet{so} suggest a measurement for the intended use of the agile practices, we assume that higher scores on the agile practices measurement do imply a higher probability of project success. \citet{Lalsing2012} also state that it is of utter importance to find the ``right people for the right team'' and not only the ``right people,'' which seems to be a key when building teams that can leverage agile practices in the way that they are intended. \citet{hackman2003learning} underlines the importance of crossing levels in organizational research and in this present study, we have shown that the individual level does not explain much variance and we instead need to investigate the team as the level of analysis. The cross-section between what is team and organization could be hard to define and e.g. organizational routines could be defined as both \citep{pentland2005organizational}. However, the distinctiveness between teams in organizations are often possible to find, and we argue the team-level needs to be in focus instead of the individual level, but preferably also in relation to the organizational level. As a comparison, we looked at how \citet{edum2000developing} present individual skills needed in the construction industries. They suggest skills not far from what is suggested in software development, which supports our claim that these non-technical individual skills are too general to be useful in predictions of the dynamics of the specific organizational case. \citet{edum2000developing} divide the needed abilities into primary and secondary knowledge and skill elements for developing project management competencies. Among the primary knowledge and skill elements, they report: (1) Planning and scheduling, (2) Construction management activities, (3) Basic technical knowledge in own field, (4) Productivity and cost control, (5) Leadership, (6) Delegation, (7) Negotiation, (8) Decision making, (9) Motivation and promotion, (10) Team working, (11) Time management, (12) Top management relations, (13) Establishing budgets, (14) Reporting systems, (15) Drafting contracts, (16) Communication skills Presentation, (17) General and business correspondence, (18) Report writing, (19) Chairing meetings, and (20) Understanding of organization. We believe our findings imply that the teams need all the abilities that the non-technical individuals skills try to capture, but should be seen, and investigated, as a team capacity instead. This means that there is a difference between an individual having team working skills and the skills the team as a whole possesses. Our present study supports the findings presented in Section~\ref{org}, that team skills are key to implementing and using agile practices. The collective intelligence is a property of the team itself \citep{woolley2010evidence} and, therefore, also the agile practices, i.e.\ just like the collective intelligence is unrelated to individual intelligence \citep{woolley2010evidence}, individual non-technical skills seem to be unrelated to non-technical team skills. In addition, personalities can be consciously changed over time \citep{hudson2015volitional} and depend on our group membership \citep{reynolds2001role}. From the software development context, \citet{grenjss2} present an interesting quote in their study saying that the interviewees were surprised by how vocal some, previously very quiet, programmers get on some of their agile teams. Such a finding has extensive empirical support from social psychology and the studies of how context and its social expectations and interactions form reality \citep{snyder1984belief}. \subsection{Threats to Validity}\label{sec:validity} We reflect on the threats to validity using internal, external, construct, and conclusion validity following the guidelines by \citet{Wohlin2000Experimentation}. \paragraph{Internal Validity} We only used the self-assessed (i.e.\ perceived) non-technical individual skills. It is difficult to determine the correlation of these agile team members' self-assessed skills to their actual skills, or peer-assessed skills. However, it has been proven in psychology research that people overestimate their skills systematically (see e.g.\ \citet{alicke1995personal}). Hence, we assume our self-assessment to be a valid measurement when building associative models with relative associations between variables. A potential other threat is the operationalization of ``skills'' into asking about the agile team members' satisfaction of their own skill. As mentioned in the method, the reason for using the personal satisfaction of a certain non-technical skills was the fact that, in agile teams, such skills are perceived as utterly important and a satisfaction of a skill should then be more related to a peer-evaluation then if asked to only rate their individual skills. Yet, we recognize that the reported individual skills could differ from ``real'' or the actual skills perceived by peers. We did not include participants from companies not following agile practices for a comparison between other types of software development work practices. The rational behind this was our focus on companies using agile practices and the fact that we prioritized having a high number of participants than comparing the results to participants from organizations not following agile practices. We would also have had to specify and measure other work practices, which then became out of scope for this study. However, we would not be surprised if non-technical individual skills would turn out to be weakly connected to other types of high performance work practices, both in other software development methods, but also in other fields. \paragraph{External validity} The sampling in this study represents a convenience sampling procedure, as the authors of this paper had direct or indirect research connections to people within the organizations that chose participants to the study. To mitigate this threat, we tried to diversify the sample of participants to include as a representative sampling of agile teams as possible. We involved participants from seven organizations from different continents, different sizes of organizations, as well as the public and private sectors. In addition, in order to correlate skills to levels of agile maturity, the organizations, and the participating agile teams, were all on different agile maturity levels, both within and between organizations. Four organizations were based in Brazil and three organizations were from Europe, The Netherlands, including two companies and one public sector IT department. We therefore believe that our sample is representative for agile team members since we sampled from IT organizations of different types and sizes. Again, we aimed at looking for correlations, but allowing the skills measurements to co-vary, which makes such a sample appropriate. We also opted not to specify that we only wanted software developers to answer the survey, since we knew that many team members conducting software development work, might have other titles. We instead explicitly asked all members to respond, but told them to skip questions they could not assess. Therefore, our sample reflects all the agile team members that were involved in software development. \paragraph{Conclusion validity} Due to our large sample size of 113 survey responses, we had a decent sample for building our linear regression models. We also made sure all the assumptions were fulfilled when building such statistical models. \paragraph{Construct validity} To ensure that we do not just base our measurements on our assumptions, we conducted a literature review on individual skills in software development and agile projects, as well as on agile practices. From this review, we presented the most related work we found and step-by-step derived representative non-technical individual skills as well as agile practices. We only included a measurement if we found evidence for it in at least two publications. In addition, and more importantly, we failed to reject most of our hypothesis, which was not our initial intention, of course, but our data turned out to provide us with different results than we expected, i.e.\ we reported our results and discussed it without trying to fish for significant $p$ values in data. A common construct validity threat is hypothesis guessing. Since the participants filled out their surveys on paper and totally anonymous (i.e.\ not even stating their role or gender), we believe they were given a chance to answer honestly. We also did not inform participants about any of our hypotheses, but instead introduced our research topic of finding drivers behind agile practices in general over a large sample from more than one country. \section{Conclusions and Future Work}\label{sec:conclusions} This paper set out to investigate the assumption that non-technical individual skills are positively connected to the mature use of agile practices. Through building a set of multiple linear regression models using a total of 113 survey responses, we analyzed the predictive power in measuring individual skills in relation to agile practices. We found that there is very low power in using non-technical individual skills to predict the maturity of agile practices in software development teams. We, therefore, conclude that looking at non-technical individual skills is not the optimal level of analysis when trying to understand and predict the use of agile practices in the software development context. Future studies should focus more on the team-level when understanding the use of agile practices and build upon such theories when understanding the dynamics of agile teams in addition to trying to validated the result of the present study using external, or peer-assessed, measurements of skills and agile practices. In future studies, instead of asking questions about the non-technical individual skills, we would suggest items regarding non-technical team-level skills. For example, given that members of the team have enough talent and experience for the kind of work that is conducted, items suggested by \citet{wageman2005team} might be useful, like for example, whether everyone in the team has the special skills that are needed for team work. Future studies should also include the macro level of analysis in order to investigate if aspect of team agility could be explained on the organizational level, which there are indications of \citep{roth}. \section*{Acknowledgements} We would like to thank all the participating companies as well as colleagues helping out in different phases of this research. \bibliographystyle{model5-names}

b4b7b1fa9cdcb61f8b0909f0b79f0b794e221b4a

\section{Introduction} In the standard model of cosmology \cite{wie}, \cite{peeb}, the dark energy component which due to observations constitutes about $72\%$ of the universe is represented by the cosmological constant $\Lambda$. This component of the universe constituents is responsible for the current speeding up expansion of of the universe. The other leading constituent in this model is the cold dark and baryonic matter, so is known as the $\Lambda$$CDM$ model, and although it agrees very well with observations, this model faces some debatable issues. For instance, the huge gab between the observed value of the vacuum energy density and that expected from the quantum field theory, in fact $120$ orders of magnitude. Also the appeared coincidence of the order of magnitude between the measured vacuum energy density for the present time and the matter energy density although the former is presumed to be constant in time while the latter decreases with it. This leads cosmologists to visualize other scenarios. One of these scenarios is the dark fluid models. Dynamical dark energy models are promising. Representative of these models there are the Chaplygin gas and generalized Chaplygin gas \cite{baf}-\cite{gor}, K-essence \cite{arm}, the tachyonic field \cite{gib}-\cite{sen}, and the condensate cosmology \cite{bas}. In these models both dark energy and dark matter are represented by one single fluid, dark fluid, consequently solving the coincidence problem. This dark fluid has the dynamical property to behave as dark matter in the early time and as dark energy in the late time. As real fluids naturally show dissipative phenomena, dissipative cosmological fluid is one line of research. As early as 1967, Zel'dovich \cite{zel} on calculating the universe's entropy shew that the present specific entropy of the universe can be calculated by considering the action of the dissipative processes in the early universe. In 1987, the most simple model of the viscous universe was proposed by Padmanabhan and Chitre \cite{padm}. They considered a universe model dominated by dust with constant viscosity coefficient. They came to the conclusion that viscosity can be neglected at early times, while at late times it causes the universe to enter a late inflationary era with exponential accelerated expansion. A lot of works was then done considering viscous cosmology. Fabris et. al. \cite{fabris} studied the possibility that the present accelerated expansion of the universe is driven by a viscous fluid. Their fluid was controlled by Eckart's formalism for bulk viscosity. They shew that although their model leads to the same results of generalized Chaplygin gas (GCG) model for some choices of the parameters, their model shows absence of instabilities in the power spectrum for any choice of the parameter $\nu$ of the viscosity coefficient, $\left( \zeta(\rho)=\zeta_0 \rho^\nu \right)$, a problem characterizes the GCG model. They also shew that their model has a more normal situation since the viscosity grows with density. There are many other authors, e.g., Avelino and Nucamendi \cite{art}, and Li and Barrow \cite{lib}, considered the possibility that the present acceleration of the universe is driven by bulk viscous pressure. The problem of instabilities of the power spectrum of the GCG model was also studied by Sandvik et. al. \cite{sandv}. They shew that on the perturbation level the matter power spectrum data strongly constrain the parameters of the CGC model leaving the narrow room of allowed such models indistinguishable from $\Lambda$CDM model. However, it has been shown by Reis et. al. \cite{reisetal} that the ad hoc inclusion of entropy perturbations to the CGC model enlarges the parameter space, so that for a wide range of the parameter space the instabilities and oscillations disappear and results are compatible with Large Scale Structure and CMB observations. many authors, e.g., Hip$\grave{o}$lito-Ricaldi et.al. \cite{hopo}, \cite{hopoetal}, Fabris et. al. \cite{fabrisetal}, and Borges et. al. \cite{borges}, then considered viscosity as a natural candidate for intrinsic entropy perturbation. The authors shew that the perturbation dynamics for their models avoid short scale oscillations or instabilities. They also shew how the viscous dark fluid models well competitive with $\Lambda$CDM model. The effect of viscosity on the evolution of different cosmological parameters like scale factor, the Hubble parameter, and the statefinder parameters is studied by Mostafapoor and Gr{\o}n \cite{mosp}. They shew that the viscosity of the cosmic fluid causes the energy density of the universe to converge to a finite value and stay constant for large time in contrast to the case of perfect fluid where the universe ends up as an empty space. A study on the effect of different types of viscosity on the decay of anisotropy of the universe is made by Brevik and Gr{\o}n \cite{brev}. The authors came to the conclusion that the existence of viscosity affects the early and late time evolution of the universe importantly and tends to smooth out anisotropies in the universe. A general viscous isotropic flat Friedman universe was studied by Norman and Brevik \cite{norman}. An important result of their work is the calculation of the present day viscosity coefficient of the cosmological fluid constrained by Hubble parameter observations to a non-zero value. Many authors, e.g., Velten and Schwartz \cite{velt}, Wang and Meng \cite{wang}, Brevik \cite{brvi}, and Sasidharan and Mathew \cite{sasi} also came to the same result. Many authors also considered dissipative phenomena and its important roles for many events during the evolution of the universe. Bamba et. al. \cite{bamb} considered the bulk viscous fluid description for inflationary universe. They concluded that representing inflation through fluid instead of a scalar field can equally explain the observational results acquired by Plank Satellite. Buoninfante and Lambiase \cite{buon} considering the problem of the late abundance of gravitinos shew that dissipative effects allow to avoid the late over production of gravitinos. In this work, we extend our previous work on the unified model we began in \cite{esra}. In that work we considered a viable model for the universe in which the two dark sectors are considered as one single fluid. That fluid was a perfect one with an equation of state (EoS) in the form of a correction to the vacuum EoS by a one asymptotes between two power laws, so that it has the advantage of interpolation between the two equations of state of the DE and DM. In this work we extend our previous work by adding dissipative effects in the form of bulk viscosity. Shear viscosity is considered negligible as observations proved that the universe is isotropic. The paper is organized as follows. Section \ref{dynamics} introduces the basic dynamics of the bulk viscous cosmology and applying it to the unified model studied in this work. In section \ref{phase} we analyze the model using the phase space methodology and study the evolution of the our cosmological model. Section \ref{cosmpar} is to constrain the parameters of the model using its asymptotic behavior and many cosmological observations. In section \ref{test} we study what the model expects for the universe evolution which confronts the model with many basic physical and observational tests. In section \ref{conc} we present our final conclusion. \section{Dynamics of the Model} \label{dynamics} A flat homogeneous and isotropic universe is described by the metric \begin{equation} ds^2=dt^2-a^2(t) \delta_{ij} dx^i dx^j \label{metric} \end{equation} where we consider units with $c=1$. The energy momentum tensor for the fluid is \begin{equation} T_{\mu\nu}=\rho U_\mu U_\nu+\left(p-\theta \zeta \right)h_{\mu\nu} \label{tensor} \end{equation} where $\rho$ is the energy density of the cosmic fluid as a whole, $\theta=3 H$ is the volume expansion rate of the fluid with $H$ is the Hubble parameter, $\zeta=\zeta(\rho)$ is the bulk viscous coefficient that arises in the fluid which is restricted to be positive, and $h_{\mu\nu}=U_\mu U_\nu-g_{\mu\nu}$ is the projection tensor to the $3-$space orthogonal to the fluid element, where in comoving coordinates the four-velocity $U_\mu={\delta^0}_\mu$. With the metric (\ref{metric}) Einstein's equation will lead to the Friedman equations \begin{align} \frac{\dot{a}^2}{a^2}=\frac{8 \pi G}{3}\rho \label{dot}\\ \frac{\ddot{a}}{a}=-\frac{4 \pi G}{3}(\rho+3P) \label{ddot} \end{align} Considering units with $8 \pi G=1$, the above eqns reduce to \begin{align} H^2=\frac{1}{3} \rho \label{hsq} \\ H^2+\dot{H}=-\frac{1}{6}(\rho+3P) \label{hddot} \end{align} where $H=\frac{\dot{a}}{a}$ is the Hubble parameter. The effective pressure $P$ consists of an adiabatic part $p$ plus a viscous term \begin{equation} P=p-\theta \zeta(\rho) \label{pres} \end{equation} The adiabatic part $p$ is considered to be a barotropic one that can asymptote between dust and DE in the early and late times \cite{esra} \begin{equation} p=-\rho + \frac{\gamma \rho^n}{1+\delta \rho^m} \label{van1} \end{equation} where $\gamma$, $\delta$, $n$, and $m$ are free parameters. As mentioned in \cite{esra}, this form of the adiabatic pressure has the advantage that it enables interpolation between different powers for the density, which allows for smooth phase transitions during the universe evolution. It also has the advantage that it has a general EoS for DE that enables the cosmological constant as a special case. The conservation equation ${{T_\nu}^\mu};\mu=0$ gives \begin{equation} \dot{\rho}+3 \frac{\dot{a}}{a} (\rho + P)=0 \label{cons} \end{equation} This can be written as \begin{equation} \dot{\rho}+3 \frac{\dot{a}}{a} (\rho + p)=\theta^2 \zeta(\rho) \label{cons1} \end{equation} which shows that the viscosity term may play the role of a coupling term between matter and dark energy. This relation can be written in terms of Hubble parameter as \begin{equation} \dot{H}+\frac{1}{2} (\rho + p) = \frac{1}{2} \theta \zeta(\rho) \label{rodot} \end{equation} The bulk viscous coefficient is considered to take the form \begin{equation} \zeta(\rho) = \zeta_0 \rho^\nu \label{vis} \end{equation} where $\zeta_0$ and $\nu$ are constants. In this work we focus on the simple ansatz of $\nu=0$ so that we have a constant bulk viscus coefficient $\zeta(\rho) = \zeta_0$. \section{Cosmological Model Evolution in View of the Theory of Dynamical System} \label{phase} The theory of dynamical systems, see for ex. \cite{brin}, \cite{lyn}, is originated to study the long term behavior of evolving systems. In this theory, the phase space is a multidimensional space in which each dimension represents one degree of freedom of the dynamical system. For example, a single particle moving in one dimension is represented by a two dimensional phase space, or a phase plane, since it has two degrees of freedom, the position $x(t)$ and the momentum $p(t)$. A dynamical system can in general be described by an autonomous, does not explicitly depending on time, system of differential equations $\bf{\dot{x}}=\bf{f}\left(\bf{x}\right)$, where in the n-dimensional system the vectors $\textbf{x}$ and $\textbf{f}\left(\textbf{x}\right)$ are given by $\textbf{x}=\left(x_1, x_2, .. x_n\right)$ and $\textbf{f}\left(\textbf{x}\right)=\left(f_1\left(\textbf{x}\right), f_2\left(\textbf{x}\right), .. f_n\left(\textbf{x}\right)\right)$. In the phase space, each point represents one possible state of the system. Accordingly, the phase space of some dynamical system is a representation of all possible states of the system. The analysis of the phase space of the dynamical system is some sort of an abstract view to the system. The motion of the system tracing a curve in its phase space, with time as an implicit parameter. This curve represents a solution characterized by a set of initial conditions for the dynamical system. Such curves are called trajectories or orbits. The development of the solution with time is indicated by arrows. The representation of these trajectories in the phase space forming the phase portrait of the dynamical system. The phase portrait is thus a graphical tool visualizing how the solution of differential equations of the dynamical system would behave on the long run. Many important features of the motion can be guessed from the phase portrait of the dynamical system without solving the equations of motion in detail. One of the most salient features of the phase portrait is the existence of the fixed points (known also as equilibrium points, critical points, or stationary points). At these points the solution is a stationary solution for the system where we have $\bf{f}\left(\bf{x}\right)=\bf{0}$. It is important to test the stability of the fixed points under a small perturbation about $\bf{f}\left(\bf{x}\right)$. If the trajectories tend to move away from the fixed point in all directions under a small perturbation, the fixed point is considered to be a source or repeller and is unstable. If, on the other hand, the trajectories tend to move towards the fixed point from all directions under a small perturbation, the fixed point is considered to be a sink or attractor and is stable. If they move towards the point from one direction and away from the other it is called a saddle point and is half or semi-stable. The application of the theory of dynamical systems to cosmological scenarios is very powerful, see for ex. \cite{col}-\cite{aawad1}. In most of the cosmological models, the cosmological equations, in spite of the difficulty of their analytical solution, may have many solution branches due to different initial conditions. The application of the phase space method allows the extraction of essential information about different solutions of the system, and about the evolution of the cosmological model and its asymptotic dynamics. Accordingly, expectations about the origin and fate of the Universe can be extracted. In addition, it allows the discussion of the stability of the model solution (a comprehensive discussion about the application of the theory of dynamical systems to the FLRW cosmology and consequences of the existence of fixed points on finite time singularities of different types is found in \cite{aawad}). Let's now apply the theory to our model. We begin by combining eq.(\ref{van1}) to eq.(\ref{rodot}) to get \begin{equation} \dot{H}=-\frac{1}{2}\left(\frac{\alpha H^r}{1+\beta H^s}- 3 \zeta_0 H\right) \label{dhdt} \end{equation} where $\alpha$, $\beta$, $r$, and $s$ are constants related to those of (\ref{van1}) through the relations \begin{equation} r=2 n \; ; \; \; \; s=2 m \; ; \; \; \; \alpha=3^n \gamma \; ; \; \; \; and \; \; \; \beta=3^m \delta \label{const} \end{equation} As mentioned in \cite{rich}, one of the unwritten rules that one has to follow when choosing the variables of the phase space is that they should be dimensionless. Accordingly, we consider the dimensionless parameters $h=H/H_0$ and $\tau=H_0 t$, whence relation (\ref{dhdt}) reduces to \begin{equation} \frac{d h}{d \tau}=-\frac{1}{2} \alpha \left[\frac{h^{s+2}}{H_0^{-s}+\beta h^s} - \left(\frac{\zeta_0}{\beta H_0}\right) h\right] \label{htau} \end{equation} Now, as the effective equation of state parameter is calculated using the relation $P = \omega_{eff} \; \rho$, then the use of (\ref{pres}) and (\ref{van1}) together with (\ref{hsq}) and (\ref{const}) gives \begin{equation} \omega_{eff}=-1 + \frac{\alpha}{3 h^2} \left[ \frac{h^{s+2}}{H_0^{-s}+\beta h^s}- \left(\frac{\zeta_0}{\beta H_0}\right) h \right] \label{effwh} \end{equation} Accordingly, relation (\ref{htau}) can also be written as \begin{equation} \frac{d h}{d \tau}=\frac{-3}{2} \left( 1 + \omega_{eff}\left(\tau\right) \right) h^2 \label{phas} \end{equation} Relations (\ref{htau}) and (\ref{phas}) trace the trajectory of the system in its phase space. Fig.\ref{fig:phase} shows the phase portrait for our viscus fluid model where $\zeta(\rho)=\zeta_0$ against the corresponding perfect fluid model of \cite{esra} where $\zeta(\rho)=0$. The red boundary represents the zero acceleration ansatz. The shaded area represents deceleration phase of the evolution where $q \succ 0$. Inside this region, expansion takes place if $H \succ 0$, while contraction takes place if $H \prec 0$. Outside this region we find the acceleration phase, where $q \prec 0$, and where expansion accelerated if $H \succ 0$ while contraction accelerated if $H \prec 0$. \begin{figure} [h] \vspace{-1cm} \includegraphics[width=12.5cm, height=10cm] {PhaseDiagram.eps} \vspace{-1.1cm} \caption{\scriptsize{Phase diagram of our model, solid blue curve, for $\zeta(\rho)=\zeta_0$, compared to previous work with $\zeta(\rho)=0$ \cite{esra}, dash-dot green curve. The zero acceleration ansatz, dash red curve, separates the deceleration region, shaded area, from acceleration regions. The two purple diamonds show the transition between deceleration and accelerated expansion regions}} \vspace{-.2cm} \label {fig:phase} \end{figure} We can see from the Fig. that the bulk viscous pressure trajectory characterized by two fixed points. One of these points is the null solution where $(h,f(h))=(0,0)$. This is an unstable fixed point, (although it is semi-stable point with respect to the perfect fluid model). It represents an unstable transient Minkowskinian empty space since the matter density drops to zero asymptotically, $h \rightarrow 0$. To reach this point, the trajectory exhibits positive slope, which, due to (\ref{phas}), means that $\omega_{eff} \prec -1$. The universe then crosses the phantom line in this region. The other fixed point, $(h,0)$, is a stable future attractor. Referring to relation (\ref{phas}), since $h \neq 0$ at this point then $\frac{d h}{d \tau}=0$ when $\omega_{eff}=-1$ . Accordingly, this point represents a stable de Sitter universe dominated by dark energy. It is the cosmological constant asymptote of the model. Now, fixed points are stationary, so, if the dynamical system starts at a fixed point, it'll remain there forever, These are de Sitter Cosmologies. It is thus clear that we have a multi-branch solution, so that different initial conditions lead to different solutions. One of these solutions exists in the negative-$H$ patch. However, this does not match our universe. We know from observations that $H$ is positive, we live in an expanding universe. In the other solution the system evolves between the two fixed points. If the system starts at a point between the two fixed points it'll evolve smoothly without any singularities. However, although this solution exists in the positive $H$ branch, such universe is again not ours. It has properties which contradict observations, such as the positivity of $\dot H$ and the nonexistence of the transition phases. A third solution evolves in the positive $H$ patch till it reaches the fixed de Sitter point. The physics of the evolution of this universe matches ours. It evolves from a Big Bang singularity to a de Sitter space. It evolves from decelerated to accelerated expansion. It is practically evolving to a de Sitter space free from future singularities of types $I, II$ or $III$ since $f(h)$ is continuous and differentiable \cite{aawad}. Accordingly, this third solution is what we'll consider. \section{Cosmological Parameters of the Model} \label{cosmpar} We are about to constrain the model parameters. Our model is a unified dark fluid model which can describe the evolution of the universe and is able to interpolate smoothly between dust at early time and dark energy at late time. The barotropic part of the model is some sort of a correction to the vacuum EoS by a one which asymptotes between two power laws that are able to describe the two phases of dust and DE. Specifically, at the late time we have a more general EoS for DE which enables cosmological constant as a special case. In this section we are going to constrain the model parameters by first studying the asymptotic behavior of the model and applying the conditions that satisfying the two phases of the cosmic fluid at early and late times. The remaining parameters are then constrained by making use of different cosmological observations. \subsection{Asymptotic Behavior} \label{adiab} Let's now study the asymptotic behavior of the EoS and constrain the parameters in order to realize the two asymptotes of the cosmic fluid, e.g., dust at early times and DE at late times. The parameter $s$ is considered as positive as this is the successful case of study in \cite{esra}. Now, at early times the adiabatic part of (\ref{dhdt}) possesses the asymptotic form \begin{equation} \dot{H}=-\frac{\alpha}{2 \beta} H^{r-s} \label{asymp} \end{equation} This constrains the parameters to satisfy the dust equation of state at that time, such that \begin{equation} r-s=2 \; ; \; \; \; and \; \; \; \; \frac{\alpha}{\beta}=3 \label{par} \end{equation} On the other hand, the late time asymptotic form is \begin{equation} \dot{H}=-\frac{-\alpha}{2} H^{r} \label{asymp2} \end{equation} This gives no new information. Eq (\ref{par}) lets two out of four parameters for this adiabatic part as free, and, of course, we still have the parameter of the viscosity coefficient $\zeta_0$. \subsection{Constraining the Bulk Viscosity Coefficient with $q_0$} \label{decepar} The present day value of deceleration parameter (DP), $q_0$, is one of the most important cosmological parameters. In his (1970) paper, Alan Sandage \cite{sand} defined the observational cosmology as the search for two parameters, the Hubble parameter $H_0$ and the deceleration parameter $q_0$. DP is defined as a dimensionless dynamical parameter given by \begin{equation} q(z) = -\frac{\ddot{a}}{a} \frac{1}{H^2} \label{dec1} \end{equation} Now using \begin{equation} \frac{\ddot{a}}{a}=H^2+\dot{H} \label{dec2} \end{equation} relation (\ref{dec1}) will take the form \begin{equation} q(z) = -1 - \frac{\dot{H(z)}}{H^2(z)} \label{dec3} \end{equation} Using (\ref{dhdt}), we have for our model \begin{equation} q(z)=-1+\frac{1}{2}\left(\frac{\alpha H^s}{1+\frac{\alpha}{3} H^s}- 3 \frac{\zeta_0}{H}\right) \label{qofz} \end{equation} where we used eqn(\ref{par}) for the parameters $r$ and $\beta$. Solving for $\zeta_0$ at $z=0$ we get \begin{equation} \zeta_0=\frac{2 H_0}{3}\left[\left(-1-q_0\right)+\frac{1}{2}\left(\frac{\alpha {H_0}^s}{1+\frac{\alpha}{3} {H_0}^s}\right)\right] \label{zeta} \end{equation} This equation constrains the bulk viscosity coefficient $\zeta_0$. If we are able to constrain the two parameters $\alpha$ and $s$, we can calculate $\zeta_0$ through this relation using the two observational values of $q_0$ and $H_0$. \subsection{Constrainig the Remaining two parameters} \label{othpar} The remaining two parameters $\alpha$ and $s$ can be also constrained based on other cosmological observations like cosmic deceleration-acceleration transition redshift and the age of the universe. \subsubsection{Cosmic Deceleration-Acceleration Transition} \label{dectran} Observations from type Ia Supernovae (SNe Ia) and cosmic microwave background (CMB) support the scenario of the current speeding up expansion of the universe due to the domination of dark energy budget, and a slowing up expansion of earlier times due to the domination of cold dark and baryonic matter. This means that the universe underwent a dynamical phase transition from deceleration to acceleration at some transition redshift $z_{tr}$. At this value of the redshift, deceleration parameter, $q(z_{tr})$, is zero. Using the value of $z_{tr}$ due to observations in eq(\ref{qofz}), we can constrain one of the remaining two parameters, $\alpha$ or $s$. \subsubsection{Age of the Universe} \label{age} The age of the universe is a powerful tool for examining cosmological models and also for adjusting their parameters. The lower limit to the age of the universe is obtained by dating the oldest stellar populations. Of special interest in this regard are globular clusters, the oldest objects in our galaxy. Each cluster has a chemically homogeneous populations of stars all born nearly simultaneously. There are three ways to reliably infer the age of the oldest stars in the galaxy\cite{kra}: radioactive dating, white dwarf cooling and the main sequence turnoff time scaling. A summary for the universe age estimates due to different models and measurements is given by Spergel et. al. $(2003)$\cite{sprg}. The range $[11-16]$ Gyr is estimated due to globular clusters age. Radioactive dating estimates resulted in the range $[9.5-20]$ Gyr, while white dwarfs put a lower limit of $12.5\pm0.7$ Gyr. Krauss and Chaboyer $(2003)$\cite{kra} estimated the age of globular clusters using Monte Carlo simulation. They estimated a range of $[11-16]$ Gyr to the age of the universe. Kristiansen and Elgaroy\cite{kri} used a combination of cosmic microwave background (CMB), large scale structure (LSS), and SNe Ia data to get a lower limit of $12.58\pm0.26$ Gyr to the expansion age of the universe. In this context, we'll use the recent data for the age of the universe to constrain the last parameter in our model. The age of the universe can be calculated through the relation \begin{equation} {\int_0}^{t_0} dt = - {\int_{H_0}}^\infty \frac{dH}{\dot{H}} \label{t0} \end{equation} Using (\ref{dhdt}) with (\ref{par}) we get for our model \begin{equation} t_0= {\int_{H_0}}^\infty \frac{2\left(1+\frac{\alpha}{3} H^s\right)}{\alpha H^{s+2}- 3 \zeta_0 H \left(1+\frac{\alpha}{3} H^s\right)} \label{aget0} \end{equation} Using $t_0=13.8$ Gyr from Plank 2015 results \cite{plank} the last parameter of the model can be constrained. Proceeding this way, all of our parameters are now constrained. In our calculations the Hubble constant, $H_0$, is taken as $70$ km/s/Mpc \cite{coll}, deceleration parameter as $q_0 = -0.57$ \cite{varg}, and deceleration acceleration transition redshift as $z_{tr} = 0.76$ \cite{far}. The parameters of the model are constrained due to these observations to the values $\alpha = 9.6\times 10^{-8}$, $s=3.92$, and $\zeta_0=4.68$. And to attain a correct asymptotic cosmological behavior for the model we must, due to (\ref{par}), have $\beta=\alpha/3$ and $r=s+2$. \section{Expectations for the Universe Evolution} \label{test} In the following we study what the model expects for the universe evolution. First we apply the diagnostic $Om(z)$ test to define our model category. We then confront our model with many physical and cosmological observational tests. \subsection{The $Om(z)$ Diagnostic} \label{digtest} The behavior of the cosmological models is well defined through the behavior of their cosmological parameters such as the Hubble parameter, the deceleration parameter, and the EoS parameter. However, all acceptable DE model have a positive Hubble parameter $H(z)$, and a deceleration parameter $q(z)$ switches sign during evolution from $+ve$ to $-ve$ indicating a phase transition from deceleration to acceleration at a given value of $z$ known from observations. Accordingly, these two parameters can not differentiate quintessence-like from phantom-like DE models. Even the effective EoS parameter for dynamical dark fluid models may not be enough to effectively differentiate models. As $H$ is a function of $\dot a$ and $q$ is a function of $\ddot a$, one way to differentiate models is to use higher time derivatives for the scale parameter. There are two parameters that are functions of $\dddot{a}$ which are called statefinder parameters \cite{alam}, \cite{sah}, usually denoted as $\left\{r, s\right\}$. These two parameters do, in fact, differentiate models. However, these are functions of the $\dddot{a}$, so they need somewhat heavy calculations. A more simpler, while also effective, way to do the job is to apply the $Om(z)$ diagnostic test. It relies only on the first order derivative and so demands less effort. The $Om(z)$ diagnostic test was introduced by Sahni et. al. $(2008)$ \cite{sah} to distinguish the behavior of DE in dynamical models away from the EoS. His relation stems from the redshift dependence of the function $H^2$ and is given by \begin{equation} Om(z)=\frac{\left(H/{H_0}\right)^2-1}{\left(1 + z\right)^3 -1} \label{omofz} \end{equation} In unified models, the two dark sectors are treated as one entity, so that the density of cosmic fluid is given by \begin{eqnarray} \rho_{tot}(z)= \rho_{m}(z) + \rho_{de}(z) \label{rhotot} \\ = \rho_{{m}_0} \left(1 + z\right)^3 + \rho_{de}(z) \label{rhoto0} \end{eqnarray} Using (\ref{hsq}) and divide by the critical density, $\rho_c(z)$, we get \begin{equation} H^2={H_0}^2\left[\Omega_{{m}_0}\left(1 + z\right)^3 + \Omega_{{de}_0} g(z)\right] \label{gofz} \end{equation} where $g(z)$ is defined through the relation $\rho_{de}(z)=\rho_{{de}_0} g(z)$. It follows that \begin{equation} Om(z)=\frac{1}{\left(1 + z\right)^3 -1}\left[\Omega_{{m}_0}\left(1 + z\right)^3 + \Omega_{{de}_0} g(z) -1\right] \label{omofz1} \end{equation} For a spatially flat universe we then have \begin{equation} Om(z)=\Omega_{{m}_0} + \frac{\left(1-\Omega_{{m}_0}\right)\left(g(z)-1\right)}{\left(1 + z\right)^3 -1} \label{omofz2} \end{equation} The $\Lambda CDM$ model considers a cosmological constant $\Lambda$ for DE and hence has $g(z)=1$, so that it has $Om(z)=\Omega_{{m}_0}$, i.e., it is just a null test of cosmological constant. For any other dynamical model, $Om(z)\succ\Omega_{{m}_0}$ where it represents quintessence, or $Om(z)\prec\Omega_{{m}_0}$ where it represents phantom. \begin{figure} [h] \hspace{2cm} \includegraphics[width=7cm, height=7cm] {Diagnostic.eps} \vspace{-.6cm} \caption{\scriptsize{Evolution of the $Om(z)$ parameter}} \vspace{-.2cm} \label {fig:omofz} \end{figure} Fig. \ref{fig:omofz} shows the evolution of the $Om(z)$ for our model together with the $\Lambda CDM$ ansatz. The Fig. reflects the fact that our model originally situated in the quintessence region but slides through the phantom barrier before it continues as quintessence in the near present. \subsection{The Hubble Parameter} \label{hubtest} Hubble parameter, $H(z)$, is the parameter that measures the cosmological expansion rate. In our model, we can solve for a relation between the Hubble parameter and redshift. Relation (\ref{cons}) together with (\ref{hsq}) result in the relation \begin{equation} \frac{dH}{da}=-\frac{1}{2a}\left(\frac{\alpha H^{s+1}}{1+\frac{\alpha}{3} H^s}- 3 \zeta_0\right) \label{dhda} \end{equation} This has the solution \begin{equation} z= -1 + exp\left[-2 {\int_H}^{H_0} \left(\frac{1+ \frac{\alpha}{3} H^s}{\alpha H^{s+1}- \zeta_0 \alpha H^s- 3 \zeta_0}\right)dH\right] \label{hd} \end{equation} The integration in the above equation can be solved numerically to get a relation between $H$ and $z$. Recently, Farooq et al. \cite{far} compiled updated list of $38$ measurements of $H(z)$ for $0.07 \leq z \leq 2.36$. To compare the general behavior of our results with observations, we used his list against our results from relation (\ref{hd}). Fig. \ref{fig:hubble} shows a plot for $H(z)$ from our model compared to observations from Farooq et al. \cite{far}. Results due to $\Lambda CDM$ model are also shown. The Fig. shows that the model can represent the data well and in fact it coincides with results from $\Lambda CDM$ model especially at larg $z$. \begin{figure} [h] \hspace{2cm} \includegraphics[width=7cm, height=7cm] {HubbleParameter.eps} \vspace{-.6cm} \caption{\scriptsize{General behavior of Hubble Parameter due to our model compared to observations from Farooq et al. \cite{far}. Shown also results due to $\Lambda CDM$ model.}} \vspace{-.2cm} \label {fig:hubble} \end{figure} \subsection{Evolution of the Deceleration Parameter} \label{dtest} Deceleration parameter (DP) is the parameter which was supposed to be a measure of the deceleration of the expansion of the universe due to gravity. However, cosmological observations of the high redshift supernovae in (1998) presented convincing evidence of the fact that the expansion of universe is instead accelerated \cite{ries}, \cite{perl}. Since our universe is accelerated if we have have $\frac{\ddot{a}}{a} \succ 0$, accordingly, due to the definition of DP in (\ref{dec1}), if $q \prec 0$, the expansion is accelerating, while if $q \succ 0$ it is decelerating. The redshift evolution of the deceleration parameter can be visualized using relation (\ref{ddot}) for standard cosmology in eq. (\ref{dec1}). This gives \begin{equation} q(z)=\frac{1}{H^2} \frac{4 \pi G}{3} \left(\rho+3 \omega \rho\right) \label{qstan} \end{equation} Applying (\ref{hsq}) and considering units with $8 \pi G = 1$ one gets \begin{equation} q(z)=\frac{1}{2} \left(1 + 3 \omega \right) \label{qst} \end{equation} We can see from the above relation that $q(z)$ changes from $+\frac{1}{2}$ for matter dominated era to $-1$ for dark energy dominated era. For our model, deceleration parameter can be calculated directly from relation (\ref{qofz}) using the numerical integration of relation (\ref{hd}). Fig. \ref{fig:dec} shows a plot for $q(z)$ for our model compared to $\Lambda CDM$ model. \begin{figure} \hspace{2cm} \includegraphics[width=7cm, height=7cm] {DecelParameter.eps} \vspace{-.6cm} \caption{\scriptsize{Deceleration Parameter from our model compared to that of $\Lambda CDM$ model.}} \vspace{-.2cm} \label {fig:dec} \end{figure} We can see from the Fig. that the evolution of deceleration parameter due to our model has the behavior that is well known in the literature. We can also see the values of $q_0=-0.57$ and $z_{tr}=0.761$. \subsection{Density of the Universe} \label{dentes} Referring to relations (\ref{rhotot})-(\ref{gofz}), the density parameters of the fluid components are given by \begin{align} \Omega_{m}(z)= \frac{{H_0}^2}{H^2} \Omega_{{m}_0}\left(1 + z\right)^3 \label{Omegam} \\ \Omega_{de}(z)= \frac{{H_0}^2}{H^2} \Omega_{{de}_0} g(z) \label{Omegade} \end{align} using the numerical integration of relation (\ref{hd}), the matter and dark energy density parameters can be graphically represented. The parameter $\Omega_{{de}_0}$ is taken as $0.718$ \cite{wmap}. Fig.\ref{fig:densty} shows $\Omega_i$ for each component due to our model compared to the $\Lambda CDM$ model. \begin{figure} \hspace{2cm} \includegraphics[width=7cm, height=7cm] {Omega-i.eps} \vspace{-.6cm} \caption{\scriptsize{Matter and dark energy density parameters for our model compared to $\Lambda CDM$ model.}} \vspace{-.2cm} \label {fig:densty} \end{figure} The Fig. again represents the behavior that is well known in the literature. It is also clear that while our model has the same trend as the $\Lambda CDM$ model at low redshift, they behave differently at high redshift. Specifically, the dark energy density of our model decays more slowly than that of the $\Lambda CDM$ model which reflects the smooth transition of our model. We can also see that the model expects a redshift of $z=0.35$ for both dark sectors to share half the energy density of the universe. \subsection{Effective Equation of State Parameter} \label{effeq} While the cosmological constant $\Lambda$ represents DE candidate for the $\Lambda CDM$ model, dynamical models have effective time evolving DE candidate with a non-trivial parametrized equation of state. Consider the conservation equation (\ref{cons}) for dark energy \begin{equation} \dot{\rho_{de}}+3 \frac{\dot{a}}{a} (\rho_{de} + P_{de})=0 \label{decon} \end{equation} This can be written in the form \begin{equation} \dot{\rho_{de}}+3 \frac{\dot{a}}{a} (1 + \omega_{de}(a))\rho_{de}=0 \label{decon1} \end{equation} where $\omega_{de}$ is the parameter of the underlying dark energy EoS, $P_{de}=\omega_{de} \; \rho_{de}$. Now using \begin{equation} \frac{da}{a}=-\frac{dz}{1+z} \label{dadz} \end{equation} We can integrate to get the evolving dark energy density. This will be given by \begin{equation} \rho_{de}(z)= \rho_{{de}_0} exp\left[3 {\int_0}^z \frac{1 + \omega_{de}(z)}{1+z} dz\right] \label{rhode} \end{equation} Dividing by $\rho_c$, this can be substituted in (\ref{gofz}) to give \begin{equation} H^2={H_0}^2 \left\{\Omega_{{m}_0}\left(1 + z\right)^3 + \Omega_{{de}_0} exp\left[3 {\int_0}^z \frac{1 + \omega_{de}(z)}{1+z} dz\right]\right\} \label{hsq1} \end{equation} where the comparison with eq(\ref{gofz}) gives \begin{equation} g(z)= exp\left[3 {\int_0}^z \frac{1 + \omega_{de}(z)}{1+z} dz\right] \label{gofz1} \end{equation} To get $\omega_{de}(z)$ we make use of the pressure of the total fluid of the universe, $P$, which can be written as \begin{equation} P= P_{m}+P_{de}=P_{de} \label{pre1} \end{equation} Accordingly \begin{equation} {\omega}_{eff} \; \rho_{tot} = \omega_{de} \; \rho_{de} \label{omgs} \end{equation} which on dividing by $\rho_c(z)$ leads to \begin{equation} \omega_{de}(z)= \frac{{\omega }_{eff}(z)}{\Omega_{de}(z)} \label{omgs1} \end{equation} The effective EoS parameter for the universe fluid, ${\omega }_{eff}(z)$, is calculated by considering relations(\ref{pres}) and (\ref{van1}) where the pressure of the universe fluid is given by \begin{equation} P=-\rho + \frac{\gamma \rho^n}{1+\delta \rho^m}- 3 \zeta_0 H = \omega_{eff} \; \rho \label{effp} \end{equation} Using (\ref{hsq}) and (\ref{const}) one finally gets \begin{equation} \omega_{eff}=-1 + \frac{\alpha H^s}{3 + \alpha H^s} - \frac{\zeta_0}{H} \label{wmod} \end{equation} Fig. \ref{fig:weff} shows a plot of the model effective EoS parameter together with the parameter of the underlying DE EoS. \begin{figure} [h] \hspace{2cm} \includegraphics[width=7cm, height=7cm] {w-eff.eps} \vspace{-.6cm} \caption{\scriptsize{Effective EoS parameters for the model and DE.}} \vspace{-.2cm} \label {fig:weff} \end{figure} As it was expected by the $Om(z)$ diagnostic test, the Fig. shows that the DE slides through the phantom barrier where it reaches a minimum parameter value of $\omega_{de}=-1.167$ then bounces up to quintessence again in the near present. It shows a quintessence behavior in the present time where it comes to the current value of $\omega_{de}(z=0)=-0.99$. We can also see that $\omega_{de}\rightarrow -1$ in the future where $z\rightarrow -1$ in consistence with the expectations that the universe will be fully dominated with DE. It is also clear from the Fig. that the effective EoS of the universe fluid follows the dust EoS at large $z$ while drops to $\omega_{eff}(z=0)=-0.713$ at the present time while tends to the value of $-1$ as $z \rightarrow -1$ so that it represents the DE EoS. The universe effectively crosses the value of $\omega_{eff}=-1/3$ at the transition redshift $z=0.76$ where $q=0$ in agreement with the expectations at this transition time. \subsection{Estimating the Viscosity of the Cosmic Fluid} Due to Velten and Schwartz \cite{velt}, viscus dark matter allowed to have a bulk viscosity $\preceq 10^7$ Pa.s. On the other hand, Wang and Meng \cite{wang} considered a model with time dependent bulk viscosity coefficient, $\zeta=\zeta(t)$. Their analysis resulted in a current bulk viscosity given in the average by $\zeta_0 \approx 10^5$ Pa.s, in agreement with Velten and Schwartz. Another work is done with Sasidharan and Mathew \cite{sasi} in which they perform a phase space analysis of a universe dominated with bulk viscous matter. They considered a bulk viscosity coefficient which is a function of Hubble parameter and its first derivative. Their $\chi^2$ analysis to Supernovae data for the constant bulk viscosity coefficient model resulted in a current bulk viscosity of $\zeta_0 \approx 7.68 \times 10^7$ Pa.s. In our work we estimated the value of the parameter $\zeta_0$ as $4.68$. We are using astronomical units, so that our $\zeta_0$ has the units of Hubble parameter. Accordingly in $SI$ units, we estimate a value of $8\times10^6$ Pa.s as the value of the viscosity of the cosmic fluid. We can see that this result agrees with the work of the previous authors. \section{Conclusions} \label{conc} A dissipative unified dark fluid model is considered which has an adiabatic EoS part that can asymptotes between two power laws, so that it has the ability of smooth transition between dust and DE equations of state. The dissipative part is described by bulk viscosity with constant coefficient where the shear viscosity is excluded due to the isotropy of the universe. The model is analyzed using the theory of dynamical system where the phase space method can give expectations about the evolution of the model and the asymptotic behavior of its different classes of solution without exactly solving the equations. The phase portrait of the model showed that it has three solution classes. Two of these solutions contradict observational constraints of our universe. The third solution which has $+ve \; H$ and $-ve \; \dot H$ matches the dynamics of our universe, and, due to \cite{aawad}, it is free from future finite time singularities of types $I$, $II$, or $III$. The parameters of the model are constrained through its asymptotic behavior as a first step. This shows that two out of the five parameters are not independent, we chose them to be $\beta$ and $r$. Accordingly, only three parameters will remain free, $\alpha$, $s$, and $\zeta_0$. We constrain these parameters using many observational constraints like today's value of deceleration parameter, the redshift value of deceleration-acceleration transition, and the age of the universe. The $Om(z)$ diagnostic test is used to identify the model category, whether it is quintessence-like or phantom-like model. It shows that it is basically quintessence although it slides through the phantom barrier before it continues as a quintessence in the region near present. Expectations for the universe evolution due to the model are studied which confront the model with many physical and cosmological observational tests. We first solve the Hubble parameter equation which could be solved numerically and plotted against the new complied data of Farook et. al. \cite{far}. The results show that the model is able to represent the trend of the observational data well and its results coincide with those from $\Lambda CDM$ model especially at large $z$. The evolution of the deceleration parameter is also studied and its graphical representation reflects the behavior that is well known in the literature. It shows also the values of the present day deceleration parameter as $q_0=-0.57$ and the value of the transition red-shift as $z_{tr}=0.761$. The density parameters for the fluid components are calculated and plotted as functions of redshift. Again the graphical representation reflects the well known behavior in the literature. Besides, due to the ability of our model to asymptote smoothly between the two extreme ends, pure matter and pure DE, the DE density of our model decays more slowly than that of the $\Lambda CDM$ model at large $z$. We also got the matter-DE equality to occurs at the redshift value of $z=0.35$. We also studied the evolution of the effective EoS of the universe fluid as a whole, $\omega_{eff}$, and also the DE underlying EoS, $\omega_{de}$. The graphical representation of these parameters shows many important features. It manifests the fact that the effective EoS of the universe fluid follows the dust EoS at large $z$ and the DE EoS to the future. It also shows that the current value of the effective EoS parameter is $\omega_{eff}(z=0)=-0.713$. Another important note is that due to our results, the universe effectively crosses the value of $\omega_{eff}=-1/3$ at the transition redshift $z=0.76$ where $q=0$ in agreement with the expectations at that transition time. On the other hand, the representation of the DE EoS shows that, in agreement with the $Om(z)$ test, the DE slides through the phantom barrier where it reaches a minimum value of $\omega_{de}=-1.167$ then bounces up to quintessence again just before the present time. It shows a quintessence behavior in the present time with the current value of $\omega_{de}(z=0)=-0.99$. It also tends to the the value of $\omega_{de}\rightarrow -1$ to the future in agreement with the expectations that the universe will be fully dominated with DE. The value of the viscosity coefficient of the cosmic fluid is also estimated. We get the value of $8\times10^6$ Pa.s, in agreement with the predictions of many authors that the cosmic fluid has a viscosity $\preceq 10^7$ Pa.s. Finally, We can see that when our model parameters are well constrained through observations, the model could pass very important cosmological and observational tests. Graphical representations of different physical quantities such as $Om(z)$, Fig. \ref{fig:omofz}, and the effective EoS parameter, Fig. \ref{fig:weff}, show how our model is different form the $\Lambda CDM$ model, but it has the ability to fit the observations in different known regions, see Figs. \ref{fig:hubble}, \ref{fig:dec}, and \ref{fig:densty}. We conclude that our model is able to describe the behavior of the universe evolution. \section{Aknowledgment} \label{akn} The author would like to thank Dr. Adel Awad and Dr. Waleed El Hanafy for valuable discussions and guidance.

1376c4033bc723f18f488954cdd415f7471e45fb

\section{Introduction} Iwahori-Hecke algebras are deformations of the group algebras of finite Coxeter groups $W$ with nonzero parameters. They play an important role in the study of representations of finite groups of Lie type. In \cite{gr}, Geck and Rouquier gave a basis of the center of Iwahori-Hecke algebras. The basis is closely related to minimal length elements in the conjugacy classes of $W$. The 0-Hecke algebra was used by Carter and Lusztig in \cite{cl} in the study of $p$-modular representations of finite groups of Lie type. $0$-Hecke algebras are deformations of the group algebras of finite Coxeter groups with zero parameter. In \cite{xh}, He gave a basis of the center of $0$-Hecke algebras associated to finte Coxeter groups. The basis is closely related to maximal length elements in the conjugacy classes of $W$. Affine Hecke algebras are deformations of the group algebras of affine Weyl groups $W^{\text{aff}}$. They appear naturally in the representation theory of reductive $p$-adic groups. In \cite{gl89}, Lusztig gave a basis of the center of affine Hecke algebras. In \cite{xh}, He also mentioned a similar proof could be applied to give a basis of the center of affine $0$-Hecke algebras. The basis is closely related to finite conjugacy classes in $W^{\text{aff}}$. Let $\mathbf{G}$ be a connected reductive group over a $p$-adic field $F$. The study of mod-$p$ representations of $\mathbf{G}(F)$ naturally involves the pro-$p$-Iwahori Hecke algebra of $\mathbf{G}(F)$. Let $R$ be a commutative ring. In \cite{vm1}, Vigneras discussed the $R$-algebra $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ which generalizes the pro-$p$-Iwahori Hecke algebra of $\mathbf{G}(F)$. In \cite{vm2}, Vigneras gave a basis of the center of $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ by using the Bernstein relation and alcove walks (the definition of alcove walk can be found in \cite{ug}). The basis of center is closely related to finite conjugacy classes in $W(1)$. In general, the expression of the center in \cite{vm2} is complicated if we want to write it out explicitly by Iwahori-Matsumoto presentation. But for $R$-algebras $\mathcal{H}_R(0,c_{\tilde{s}})$, we can give an explicit description of the center by Iwahori-Matsumoto presentation. This is the main result of this paper. In Section $2$, we review the definition of $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ and some properties of the group $W(1)$. In Section $3$, we define a new operator $r_{v,w}$. In Section $4$, we show maximal length terms of a central element in $\mathcal{H}_R(0,c_{\tilde{s}})$ comes from finite conjugacy classes in $W(1)$. In Section $5$, we prove some technical results regarding $r_{v,w}$, where $w$ is in some finite conjugacy class. In Section $6$, we give a basis of the center of $\mathcal{H}_R(0,c_{\tilde{s}})$. In Section $7$, we give some examples to show how the main result works. \section{Preliminary} The symbols $\mathbb{N},\mathbb{Z},\mathbb{R}$ refers to the natural numbers, the integers and the real numbers. Let $\Sigma$ be a reduced and irreducible root system with simple system $\Delta$. Let $W_0$ be the finite Weyl group of $\Sigma$, and $S_0$ be the set of simple reflections corresponding to $\Delta$. Then $S_0$ is a generating set of $W_0$. Let $\mathcal{V}=\mathbb{Z}\Sigma^{\vee}\otimes_{\mathbb{Z}}\mathbb{R}$ be the $\mathbb{R}$-vector space spanned by the dual root system $\Sigma^{\vee}$. Let $\Sigma^{\text{aff}}$ be the affine root system associated to $\Sigma$, i.e. the set $\Sigma+\mathbb{Z}$ of affine functionals on $\mathcal{V}$. The term hyperplane always means the null-set of an element of $\Sigma^{\text{aff}}$. Choose a special vertex $\mathfrak{v}_0\in\mathcal{V}$ such that $\mathfrak{v}_0$ is stabilized by the action of $W_0$. Let $\mathfrak{C}_0$ be the Weyl chamber at $\mathfrak{v}_0$ corresponding to $S_0$ and let $\mathfrak{A}_0\in\mathfrak{C}_0$ be the alcove for which $\mathfrak{v}_0\in\bar{\mathfrak{A}}_0$ where $\bar{\mathfrak{A}}_0$ is the closure of $\mathfrak{A}_0$. Let $W^{\text{aff}}$ be the affine Weyl group of $\Sigma^{\text{aff}}$ and $S^{\text{aff}}$ be the set of affine reflections corresponding to walls of $\mathfrak{A}_0$. Then $S^{\text{aff}}$ is a generating set of $W^{\text{aff}}$ extended from $S_0$. Denote by $\ell: W^{\text{aff}}\rightarrow\mathbb{N}$ the length function relative to the generating set $S^{\text{aff}}$. The group $W^{\text{aff}}$ is also equipped with the Bruhat order $\leq$. Let $F$ be a non-archimedean local field and let $\mathbf{G}$ be a connected reductive $F$-group. Let $\mathbf{T}\subseteq \mathbf{G}$ be a maximal $F$-split torus and set $\mathbf{Z}$ and $\mathbf{N}$ be $\mathbf{G}$-centralizer and $\mathbf{G}$-normalizer of $\mathbf{T}$ respectively. Let $\mathbf{G}(F),\mathbf{T}(F),\mathbf{Z}(F),\mathbf{N}(F)$ be the groups of $F$-points of $\mathbf{G},\mathbf{T},\mathbf{Z},\mathbf{N}$. Then the group $\mathbf{Z}(F)$ admits a unique parahoric subgroup $\mathbf{Z}(F)_0$. We may define the Iwahori-Weyl group of $(\mathbf{G},\mathbf{T})$ to be the quotient $W:=\mathbf{N}(F)/\mathbf{Z}(F)_0$. There are two ways to express the Iwahori-Weyl group as a semidirect product. By the work of Bruhat and Tits, it is known that there exists a reduced root system $\Sigma$ such that the corresponding affine Weyl group is a subgroup of $W$. Denoting by $W_0$ the finite Weyl group of $\Sigma$, it can be shown that $W=\Lambda\rtimes W_0$ and that $W=W^{\text{aff}}\rtimes\Omega$. The action of $W^{\text{aff}}$ on $\mathcal{V}$ extends to an action of $W$. The subgroup $\Lambda$ acts on $\mathcal{V}$ by translations and the subgroup $\Omega$ acts on $\mathcal{V}$ by invertible affine transformations that stabilize the base alcove $\mathfrak{A}_0$ in $\mathcal{V}$. The group $\Omega$ stabilizes $S^{\text{aff}}$. We can extend the length function $\ell: W^{\text{aff}}\rightarrow\mathbb{N}$ to $W$ by inflation along the projection $W^{\text{aff}}\rtimes\Omega\rightarrow W^{\text{aff}}$. Then the subgroup of length $0$ elements in $W$ is $\Omega$. The Bruhat order on $W$ can also be defined. Let $v=v'\tau,w=w'\tau'$ be two elements in $W$ where $v',w'\in W^{\text{aff}}$ and $\tau,\tau'\in\Omega$, then $v\leq w$ if and only if $v\leq w$ and $\tau=\tau'$. The group $\Lambda$ is finitely generated and abelian and the action of $\Lambda$ on $\mathcal{V}$ is given by the homomorphism $$\nu: \Lambda\rightarrow\mathcal{V}$$ such that $\lambda\in\Lambda$ acts as translation by $\nu(\lambda)$ in $\mathcal{V}$. The group $\Lambda$ is normalized by $x\in W_0$: $x\lambda x^{-1}$ acts as translation by $x(\nu(\lambda))$. The length $\ell$ is constant on each $W_0$-conjugacy class in $\Lambda$. By Lemma $2.1$ in \cite{vm2}, a conjugacy class of $W$ is finite if and only if it is contained in $\Lambda$, and infinite if and only if it is disjoint from $\Lambda$. We'll later use the following geometric characterization of length (see Lemma $5.1.1$ in \cite{sr}): \begin{lem}\label{13} Let $w\in W$ and $s\in S^{\text{aff}}$. If $H_{s}$ is the hyperplane stabilized by $s$, then \begin{itemize} \item $\ell(sw)>\ell(w)$ if and only if $\mathfrak{A}_0$ and $w(\mathfrak{A}_0)$ are on the same side of $H_s$, \item $\ell(ws)>\ell(w)$ if and only if $\mathfrak{A}_0$ and $w(\mathfrak{A}_0)$ are on the same side of $w(H_s)$. \end{itemize} \end{lem} We will also use the following result on Bruhat order. \begin{lem}\label{14} Let $x,y\in W$ with $x\leq y$. Let $s\in S^{\text{aff}}$. Then \begin{itemize} \item $\min\{x,sx\}\leq\min\{y,sy\}$ and $\max\{x,sx\}\leq\max\{y,sy\}$. \item $\min\{x,xs\}\leq\min\{y,ys\}$ and $\max\{x,xs\}\leq\max\{y,ys\}$. \end{itemize} \end{lem} \begin{proof} When $\Omega$ is trivial, this is all well-known: see Corollary $2.5$ in \cite{gl03}. The more general statement is immediate by definition of the Bruhat order on $W$ because $W=W^{aff}\rtimes\Omega$. \end{proof} Let $Z$ be an arbitrary abelian group, and let $W(1)$ be an extension of $W$ by $Z$ given by the short exact sequence $$1\rightarrow Z\rightarrow W(1)\stackrel{\pi}\rightarrow W\rightarrow 1.$$ We denote by $X(1)$ the inverse image in $W(1)$ of a subset $X\subseteq W$. Then $W^{\text{aff}}(1),\Lambda(1)$ are normal in $W(1)$ and $W(1)=W^{\text{aff}}(1)\Omega(1)=\Lambda(1)W_0(1),Z=W^{\text{aff}}(1)\cap\Omega(1)=\Lambda(1)\cap W_0(1)$. The length function on $W$ inflates to a length function on $W(1)$, still denoted by $\ell$, such that $\ell(\tilde{w})=\ell(\pi(\tilde{w}))$ for $\tilde{w}\in W(1)$. Let $R$ be a commutative ring. For $\tilde{w}\in W(1)$ and $t\in Z$, $\tilde{w}(t)=\tilde{w}t\tilde{w}^{-1}$ depends only on the image of $\tilde{w}$ in $W$ because $Z$ is commutative. By linearity the conjugation defines an action $$(\tilde{w},c)\mapsto \tilde{w}(c): W(1)\times R[Z]\rightarrow R[Z]$$ of $W(1)$ on $R[Z]$ factoring through the map $\pi: W(1)\rightarrow W$. We recall the definition of the generic algebra $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ introduced in \cite{vm1}. \begin{thm} Let $(q_{\tilde{s}},c_{\tilde{s}})\in R\times R[Z]$ for all $\tilde{s}\in S^{\text{aff}}(1)$. Suppose \begin{itemize} \item $q_{\tilde{s}}=q_{\tilde{s}t}=q_{\tilde{s'}}$, \item $c_{\tilde{s}t}=c_{\tilde{s}}t$ and $\tilde{w}(c_{\tilde{s}})=c_{\tilde{w}\tilde{s}\tilde{w}^{-1}}=c_{\tilde{s'}},$ \end{itemize} where $t\in Z,\tilde{w}\in W(1),\tilde{s},\tilde{s'}\in S^{\text{aff}}(1)$ and $\tilde{s'}=\tilde{w}\tilde{s}\tilde{w}^{-1}$. Then the free $R$-module $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ of basis $(T_{\tilde{w}})_{\tilde{w}\in W(1)}$ admits a unique $R$-algebra structure satisfying \begin{itemize} \item the braid relations: $T_{\tilde{w}}T_{\tilde{w'}}=T_{\tilde{w}\tilde{w'}}$ for $\tilde{w},\tilde{w'}\in W(1), \ell(\tilde{w})+\ell(\tilde{w'})=\ell(\tilde{w}\tilde{w'})$, \item the quadratic relations: $T_{\tilde{s}}^2=q_{\tilde{s}}T_{\tilde{s}^2}+c_{\tilde{s}}T_{\tilde{s}}$ for $\tilde{s}\in S^{\text{aff}}(1)$, \end{itemize} where $c_{\tilde{s}}=\sum_{t\in Z}c_{\tilde{s}}(t)t\in R[Z]$ is identified with $\sum_{t\in Z}c_{\tilde{s}}(t)T_t$. \end{thm} The algebra $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ is called the $R$-algebra of $W(1)$ with parameters $(q_{\tilde{s}},c_{\tilde{s}})$. The basis of the center of $\mathcal{H}_R(q_{\tilde{s}},c_{\tilde{s}})$ given in \cite{vm2} can be very complicated when written explicitly by Iwahori-Matsumoto presentation. But when $q_{\tilde{s}}=0$, we can write out the basis explicitly. In this paper, all our discussions are under the condition of $q_{\tilde{s}}=0,\forall \tilde{s}\in S^{\text{aff}}(1)$, that is, the algebra $\mathcal{H}_R(0,c_{\tilde{s}})$. In this case, the quadratic relations become $T_{\tilde{s}}^2=c_{\tilde{s}}T_{\tilde{s}}$. For convenience, we define a $W(1)$-action on $\mathcal{H}_R(0,c_{\tilde{s}})$ given by $\tilde{w}\bullet T_{\tilde{w'}}=T_{\tilde{w} \tilde{w'} \tilde{w}^{-1}}$ for any $\tilde{w}, \tilde{w'}\in W(1)$, extended linearly to all elements in $\mathcal{H}_R(0,c_{\tilde{s}})$. The following lemma is useful in later discussion: \begin{lem}\label{1} Let $\tilde{w}_1,\tilde{w}_2,\tilde{v}_1,\tilde{v}_2\in W(1), \tilde{s}_1,\tilde{s}_2\in S^{\text{aff}}(1)$, and suppose $\tilde{w}_1 \tilde{s}_1 \tilde{v}_1=\tilde{w}_2 \tilde{s}_2 \tilde{v}_2$ and $\pi(\tilde{w}_1 \tilde{v}_1)=\pi(\tilde{w}_2 \tilde{v}_2)$. Then $\tilde{w}_1 c_{\tilde{s}_1} \tilde{v}_1=\tilde{w}_2 c_{\tilde{s}_2}\tilde{v}_2$. \end{lem} \begin{proof} Since $\pi(\tilde{w}_1 \tilde{v}_1)=\pi(\tilde{w}_2 \tilde{v}_2)$, we have $\tilde{w}_2 \tilde{v}_2=\tilde{w}_1 t \tilde{v}_1$ for some $t\in Z$, hence $\tilde{w}_1^{-1} \tilde{w}_2=t \tilde{v}_1 \tilde{v}_2^{-1}$. Then $\tilde{s}_1=\tilde{w}_1^{-1}\tilde{w}_2 \tilde{s}_2 \tilde{v}_2 \tilde{v}_1^{-1}=t(\tilde{v}_1 \tilde{v}_2^{-1})\tilde{s}_2(\tilde{v}_1 \tilde{v}_2^{-1})^{-1}$, so $c_{\tilde{s}_1}=t(\tilde{v}_1 \tilde{v}_2^{-1})c_{\tilde{s}_2}(\tilde{v}_1 \tilde{v}_2^{-1})^{-1}=\tilde{w}_1^{-1} \tilde{w}_2 c_{\tilde{s}_2} \tilde{v}_2 \tilde{v}_1^{-1}$, i.e., $\tilde{w}_1 c_{\tilde{s}_1} \tilde{v}_1=\tilde{w}_2 c_{\tilde{s}_2} \tilde{v}_2$. \end{proof} \section{A New Operator} In this section, we will define an operator $r_{v,w}$ for any pair $(v,w)\in W\times W$ with $v\leq w$. This operator is the main ingredient of this paper. For every $s\in S^{\text{aff}}$, pick a lifing $\tilde{s}$ in $S^{\text{aff}}(1)$, and for every $\tau\in\Omega$, pick a lifting $\tilde{\tau}$ in $\Omega(1)$. Let $w\in W$ with $\ell(w)=n$ and $\underline{w}=s_{i_1}s_{i_2}\dotsm s_{i_n}\tau$ be a reduced expression of $w$. A subexpression of $\underline{w}$ is a word $s_{i_1}^{e_{i_1}}s_{i_2}^{e_{i_2}}\dotsm s_{i_n}^{e_{i_n}}\tau$ with $(e_{i_1},e_{i_2},\dotsm,e_{i_n})\in\{0,1\}^n$. A subexpression is called non-decreasing if $\ell(s_{i_1}^{e_{i_1}}s_{i_2}^{e_{i_2}}\dotsm s_{i_n}^{e_{i_n}}\tau)=\sum_{k=1}^n e_{i_k}$. Let $v\leq w$, then there exists $(e_{i_1},e_{i_2},\dotsm,e_{i_n})\in\{0,1\}^n$ such that $\underline{v}_{\underline{w}}=s_{i_1}^{e_1}s_{i_2}^{e_2}\dotsm s_{i_n}^{e_n}\tau$ equals $v$ and is also a non-decreasing subexpression of $\underline{w}$. Let $\tilde{w}\in W(1)$ be a lifting of $w$, then $\tilde{w}$ has an expression $\underline{\tilde{w}}=t\tilde{s}_{i_1}\tilde{s}_{i_2}\dotsm\tilde{s}_{i_n}\tilde{\tau}$ for some $t\in Z$. Then the operator, $$r_{\underline{v}_{\underline{w}}}: \bigoplus_{\tilde{w}\in W(1),\pi(\tilde{w})=w} RT_{\tilde{w}}\longrightarrow\bigoplus_{\tilde{v}\in W(1),\pi(\tilde{v})=v} RT_{\tilde{v}}$$ is defined term by term and extended linearly, where $$r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=T_t T_{\tilde{s}_{i_1}}^{e_1}(-c_{\tilde{s}_{i_1}})^{1-e_1}T_{\tilde{s}_{i_2}}^{e_2}(-c_{\tilde{s}_{i_2}})^{1-e_2}\dotsm T_{\tilde{s}_{i_n}}^{e_n}(-c_{\tilde{s}_{i_n}})^{1-e_n}T_{\tilde{\tau}}.$$ In other words, we fix $T_{\tilde{s}_{i_k}}$'s for $e_k=1$, and replace all the other $T_{\tilde{s}_{i_k}}$'s with $-c_{\tilde{s}_{i_k}}$'s. It is easy to see that $r_{\underline{v}_{\underline{w}}}$ is independent of choice of liftings. \begin{example} In $SL_3$ case, $W$ is generated by three elements $s_0,s_1,s_2$ with relations $s_i^2=1$ for all $i$ and $s_is_js_i=s_js_is_j$ if $i\neq j$. Let $\tilde{s}_0,\tilde{s}_1,\tilde{s}_2$ be liftings of $s_0,s_1,s_2$ respectively. Let $\underline{w}=s_0s_1s_2s_0s_1s_2$, $\underline{\tilde{w}}=t\tilde{s}_0\tilde{s}_1\tilde{s}_2\tilde{s}_0\tilde{s}_1\tilde{s}_2$ for some $t\in Z$. Let $(e_1,e_2,e_3,e_4,e_5,e_6)=(1,1,1,0,1,0)$ so that $\underline{v}_{\underline{w}}=s_0s_1s_21s_11$. Then $$r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=T_tT_{\tilde{s}_0}T_{\tilde{s}_1}T_{\tilde{s}_2}(-c_{\tilde{s}_0})T_{\tilde{s}_1}(-c_{\tilde{s}_2})=T_tT_{\tilde{s}_0}T_{\tilde{s}_1}T_{\tilde{s}_2}c_{\tilde{s}_0}T_{\tilde{s}_1}c_{\tilde{s}_2}.$$ \end{example} A priori, $r_{\underline{v}_{\underline{w}}}$ depends not only on the choice of reduced expression $\underline{w}$ but also on the choice of non-decreasing subexpression $\underline{v}_{\underline{w}}$. In the following part, we will show that, in fact, $r_{\underline{v}_{\underline{w}}}$ is independent of these choices, so the notation $r_{v,w}$ makes sense. \begin{lem} Let $w\in W$ with $\ell(w)=n$ and let $\underline{w}=s_{i_1}s_{i_2}\dotsm s_{i_n}\tau$ be a reduced expression of $w$. Let $\tilde{w}\in W(1)$ be a lifting of $w$ with $\underline{\tilde{w}}=t\tilde{s}_{i_1}\tilde{s}_{i_2}\dotsm\tilde{s}_{i_n}\tilde{\tau}$ for some $t\in Z$. Let $v\leq w$, and let $\underline{v}_{\underline{w}}=s_{i_1}^{e_1}s_{i_2}^{e_2}\dotsm s_{i_n}^{e_n}\tau$ and $\underline{v}'_{\underline{w}}=s_{i_1}^{f_1},s_{i_2}^{f_2}\dotsm s_{i_n}^{f_n}\tau$ be two non-decreasing subexpressions of $\underline{w}$ which both equal $v$. Then $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})$. \end{lem} \begin{proof} We show this by induction on $l=\ell(w)+\ell(v)$. If $l=0$, then $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})=T_{t\tilde{\tau}}$. If $l=1$, then $\ell(w)=1$ and $\ell(v)=0$, so $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})=T_t(-c_{\tilde{s}_{i_1}})T_{\tilde{\tau}}$. Now suppose that the statement is correct for $l<k$, and we consider the case when $l=k$. \begin{itemize} \item If $e_1=f_1$, then by induction, the statement is correct. \item If $e_1\neq f_1$, then without loss of generality, we may assume that $e_1=1,f_1=0$, then \begin{align}\nonumber r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}}) &=T_tT_{\tilde{s}_{i_1}}T_{\tilde{s}_{i_2}}^{e_2}(-c_{\tilde{s}_{i_2}})^{1-e_2}\dotsm T_{\tilde{s}_{i_n}}^{e_n}(-c_{\tilde{s}_{i_n}})^{1-e_n}T_{\tilde{\tau}},\\ r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}}) &=T_t(-c_{\tilde{s}_{i_1}})T_{\tilde{s}_{i_2}}^{f_2}(-c_{\tilde{s}_{i_2}})^{1-f_2}\dotsm T_{\tilde{s}_{i_n}}^{f_n}(-c_{\tilde{s}_{i_n}})^{1-f_n}T_{\tilde{\tau}}.\nonumber \end{align} Let $\ell(v)=m$, then we may assume that $f_{i_j}=1$ for $j\in\{j_1,\dotsm,j_m\}\subseteq\{2,\dotsm,n\}$ and $f_{i_j}=0$ otherwise in subexpression $\underline{v}'_{\underline{w}}$. But $s_{i_1}v<v$, so by strong exchange condition, $s_{i_1}v=s_{i_{j_1}}\dotsm\widehat{s_{i_{j_d}}}\dotsm s_{i_{j_m}}$ for some $j_d$. Then by induction, $$r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})=T_tT_{\tilde{s}_{i_1}}T_{\tilde{s}_{i_2}}^{e'_2}(-c_{\tilde{s}_{i_2}})^{1-e'_2}\dotsm T_{\tilde{s}_{i_n}}^{e'_n}(-c_{\tilde{s}_{i_n}})^{1-e'_n}T_{\tilde{\tau}},$$ where $e'_{i_j}=1$ for $j\in\{j_1,\dotsm,\widehat{j_d},\dotsm,j_m\}$ and $e'_{i_j}=0$ otherwise. Now the only difference between $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})$ and $r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})$ is that $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})$ has $T_{\tilde{s}_{i_1}}$ and $-c_{\tilde{s}_{i_{j_d}}}$ as factors in the first and $j_d$th position respectively, while $r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})$ has $-c_{\tilde{s}_{i_1}}$ and $T_{\tilde{s}_{i_{j_d}}}$ as factors in the first and $j_d$th position respectively. Factors in all other positions are the same for $r_{\underline{v}_{\underline{w}}}(T_{\tilde{w}})$ and $r_{\underline{v}'_{\underline{w}}}(T_{\tilde{w}})$. Since $c_{\tilde{s}}$ is just a $R$-linear combination of elements in $Z$, it suffices to show that $$\tilde{s}_{i_1}t_1\tilde{s}_{i_{j_1}}t_2\tilde{s}_{i_{j_2}}\dotsm t_{j_d}c_{\tilde{s}_{i_{j_d}}}\dotsm t_m\tilde{s}_{i_{j_m}}=c_{\tilde{s}_{i_1}}t_1\tilde{s}_{i_{j_1}}t_2\tilde{s}_{i_{j_2}}\dotsm t_{j_d}\tilde{s}_{i_{j_d}}\dotsm t_m\tilde{s}_{i_{j_m}}$$ for any $m$-tuple $(t_1,\dotsm,t_m)\in Z^m$, which holds by Lemma \ref{1}. \end{itemize} This finishes the proof. \end{proof} This lemma tells us that $r_{\underline{v}_{\underline{w}}}$ is independent of the choice of non-decreasing subexpression $\underline{v}_{\underline{w}}$. So we can rewrite the operator as $r_{v,\underline{w}}$. \begin{thm} Let $w\in W$ with $\ell(w)=n$ and let $\underline{w}_1=s_{11}s_{12}\dotsm s_{1n}\tau$ and $\underline{w}_2=s_{21}s_{22}\dotsm s_{2n}\tau$ be two reduced expressions of $w$. Let $\tilde{w}\in W(1)$ be a lifting of $w$ with two corresponding expressions $\underline{\tilde{w}}_1=t_1\tilde{s}_{11}\tilde{s}_{12}\dotsm \tilde{s}_{1n}\tilde{\tau}$ and $\underline{\tilde{w}}_2=t_2\tilde{s}_{21}\tilde{s}_{22}\dotsm\tilde{s}_{2n}\tilde{\tau}$ for some $t_1,t_2\in Z$ respectively. Let $v\leq w$ with $\ell(v)=m$, then $r_{v,\underline{w}_1}(T_{\tilde{w}})=r_{v,\underline{w}_2}(T_{\tilde{w}})$. \end{thm} \begin{proof} Since $\underline{w}_1$ and $\underline{w}_2$ are two reduced expressions of $w$, then by Theorem $1.9$ in \cite{gl03} there exists a sequence $$\underline{w}_1=(\underline{w})_1,(\underline{w})_2,...,(\underline{w})_d=\underline{w}_2$$ of reduced expressions of $w$ such that $(\underline{w})_i$ and $(\underline{w})_{i+1}$ differ only by a braid relation. So without loss of generality, we may assume that $\underline{w}_1$ and $\underline{w}_2$ differ only by a braid relation, and even more we may assume $n,m$ are both even and other cases for $n,m$ follow by similar proofs. Then \begin{align}\nonumber \underline{\tilde{w}}_1 &=\underbrace{\tilde{s}_{\alpha}\tilde{s}_{\beta}\dotsm\tilde{s}_{\alpha}\tilde{s}_{\beta}}_{n},\\\nonumber \underline{\tilde{w}}_2 &=t\underbrace{\tilde{s}_{\beta}\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}\tilde{s}_{\alpha}}_{n},\\\nonumber v&=\underbrace{s_{\alpha}s_{\beta}\dotsm s_{\alpha}s_{\beta}}_m.\nonumber \end{align} for some $t\in Z$. Therefore, \begin{align}\nonumber r_{v,\underline{w}_1}(T_{\tilde{w}}) &=\underbrace{T_{\tilde{s}_{\alpha}}\dotsm T_{\tilde{s}_{\beta}}}_{m}\underbrace{(-c_{\tilde{s}_{\alpha}})\dotsm(-c_{\tilde{s}_{\beta}})}_{n-m}\\\nonumber &=\underbrace{T_{\tilde{s}_{\alpha}}\dotsm T_{\tilde{s}_{\beta}}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\beta}}}_{n-m},\\\nonumber r_{v,\underline{w}_2}(T_{\tilde{w}}) &=T_t(-c_{\tilde{s}_{\beta}})\underbrace{T_{\tilde{s}_{\alpha}}\dotsm T_{\tilde{s}_{\beta}}}_{m}\underbrace{(-c_{\tilde{s}_{\alpha}})\dotsm(-c_{\tilde{s}_{\alpha}})}_{n-m-1}\\\nonumber &=T_tc_{\tilde{s}_{\beta}}\underbrace{T_{\tilde{s}_{\alpha}}\dotsm T_{\tilde{s}_{\beta}}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-1}.\nonumber \end{align} It is enough to show that $\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\beta}}}_{n-m}=tc_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-1}$. But $t\tilde{s}_{\beta}=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-1}\tilde{s}_{\beta}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-1}$, so $tc_{\tilde{s}_{\beta}}=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-1}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-1}$. Therefore \begin{align}\nonumber tc_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-1} &=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-1}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-1}\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-1}\\\nonumber &=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-m-1}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-m-1}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-1}\\\nonumber &=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}c_{\tilde{s}_{\alpha}}\underbrace{\tilde{s}_{\beta}\dotsm\tilde{s}_{\alpha}}_{n-m-2}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\beta}^{-1}}_{n-m-2}\underbrace{c_{\tilde{s}_{\beta}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-2}\\\nonumber &=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}c_{\tilde{s}_{\alpha}}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-m-3}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-m-3}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\alpha}}}_{n-m-3}\\\nonumber &\dotsm\\\nonumber &=\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\beta}}_{m}\underbrace{c_{\tilde{s}_{\alpha}}\dotsm c_{\tilde{s}_{\beta}}}_{n-m}. \end{align} The third equality holds since $$\underbrace{\tilde{s}_{\alpha}\dotsm\tilde{s}_{\alpha}}_{n-m-1}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\alpha}^{-1}}_{n-m-1}c_{\tilde{s}_{\alpha}}=c_{\tilde{s}_{\alpha}}\underbrace{\tilde{s}_{\beta}\dotsm\tilde{s}_{\alpha}}_{n-m-2}c_{\tilde{s}_{\beta}}\\\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\beta}^{-1}}_{n-m-2}$$ which is true becasue $\underbrace{\tilde{s}_{\beta}\dotsm\tilde{s}_{\alpha}}_{n-m-2}c_{\tilde{s}_{\beta}}\underbrace{\tilde{s}_{\alpha}^{-1}\dotsm\tilde{s}_{\beta}^{-1}}_{n-m-2}\in R[Z]$ and $\tilde{s}_{\alpha}t'\tilde{s}_{\alpha}^{-1}c_{\tilde{s}_{\alpha}}=c_{\tilde{s}_{\alpha}t'\tilde{s}_{\alpha}^{-1}\tilde{s}_{\alpha}}=c_{\tilde{s}_{\alpha}t'}=c_{\tilde{s}_{\alpha}}t'$ for any $t'\in Z$. And all subsequent equalities hold for a similar reason. \end{proof} As the main result of this section, this theorem guarantees that $r_{v,\underline{w}}$ is independent of the choice of reduced expression of $w$. So we can rewrite the operator as $r_{v,w}$, which is what we need and will be used later. By definition of the operator, we can easily get the following propositions. \begin{prop} Let $u,v,w\in W$ and suppose $u\leq v\leq w$, then $$r_{u,v}r_{v,w}=r_{u,w}.$$ \end{prop} \begin{prop}\label{2} Let $u,v,w\in W$ and $\tilde{u},\tilde{w}\in W(1)$ be liftings of $u,w$ respectively. \begin{itemize} \item[(1)] If $v\leq w$ and $\ell(uv)=\ell(u)+\ell(v), \ell(uw)=\ell(u)+\ell(w)$, then $$T_{\tilde{u}}r_{v,w}(T_{\tilde{w}})=r_{uv,uw}(T_{\tilde{u}\tilde{w}}).$$ \item[(2)] If $v\leq w$ and $\ell(vu)=\ell(v)+\ell(u), \ell(wu)=\ell(w)+\ell(u)$, then $$r_{v,w}(T_{\tilde{w}})T_{\tilde{u}}=r_{vu,wu}(T_{\tilde{w}\tilde{u}}).$$ \end{itemize} \end{prop} \section{Maximal Length Elements} Let $\mathcal{Z}_R(0,c_{\tilde{s}})$ be the center of $\mathcal{H}_R(0,c_{\tilde{s}})$ and $h\in \mathcal{Z}_R(0,c_{\tilde{s}})$. Then $$h=\sum_{\tilde{w}\in W(1)} a_{\tilde{w}}T_{\tilde{w}},\quad\text{for some}\ \ a_{\tilde{w}}\in R.$$ Set $\text{supp}(h)=\{\tilde{w}\in W(1)| a_{\tilde{w}}\neq 0\}$. Let $\text{supp}(h)_{\max}$ be the set of maximal length elements in $\text{supp}(h)$. The following theorem tells what $\text{supp}(h)_{\max}$ is comprised of. \begin{thm}\label{11} Suppose $h\in\mathcal{Z}_R(0,c_{\tilde{s}})$, then $\text{supp}(h)_{\max}$ is a finite union of finite conjugacy classes in $W(1)$. \end{thm} This theorem comes from the following results. \begin{lem}\label{3} Let $\tilde{s}\in S^{\text{aff}}(1)$, $h\in \mathcal{Z}_{R}(0,c_{\tilde{s}})$ and $\tilde{w}\in \text{supp}(h)_{\max}$. If $\ell(\tilde{s}\tilde{w})>\ell(\tilde{w})$ or $\ell(\tilde{w}\tilde{s})>\ell(\tilde{w})$, then $\tilde{s}\tilde{w}\tilde{s}^{-1}\in\text{supp}(h)_{\max}$ and $a_{\tilde{s}\tilde{w}\tilde{s}^{-1}}=a_{\tilde{w}}$. \end{lem} \begin{proof} Without loss of generality, we may assume that $\ell(\tilde{s}\tilde{w})>\ell(\tilde{w})$. Then $\tilde{s}\tilde{w}\in \text{supp}(T_{\tilde{s}} h)=\text{supp}(h T_{\tilde{s}})$, and $$\text{supp}(T_{\tilde{s}} h)_{\max}=\{\tilde{s}\tilde{x}|\tilde{x}\in\text{supp}(h)_{\max},\ell(\tilde{s}\tilde{x})>\ell(\tilde{x})\},$$ $$\text{supp}(h T_{\tilde{s}})_{\max}=\{\tilde{y}\tilde{s}|\tilde{y}\in\text{supp}(h)_{\max},\ell(\tilde{y}\tilde{s})>\ell(\tilde{y})\}.$$ Both sets are nonempty because $\tilde{s}\tilde{w}\in \text{supp}(T_{\tilde{s}}h)_{\max}$. Therefore, $\tilde{s}\tilde{w}\tilde{s}^{-1}\in\text{supp}(h)_{\max}$ and $\ell(\tilde{s}\tilde{w}\tilde{s}^{-1})=\ell(\tilde{w})$. The $R$-coefficient of $T_{\tilde{s}\tilde{w}}$ in $T_{\tilde{s}} h$ is $a_{\tilde{w}}$ and the $R$-coefficient of $T_{\tilde{s}\tilde{w}}$ in $h T_{\tilde{s}}$ is $a_{\tilde{s}\tilde{w}\tilde{s}^{-1}}$. Thus $a_{\tilde{s}\tilde{w}\tilde{s}^{-1}}=a_{\tilde{w}}$. \end{proof} We recall Main Theorem in \cite{sr}: \begin{thm}\label{12} Fix $w\in W$. If $w\notin\Lambda$ then there exists $s\in S^{\text{aff}}$ and $s_1,\dotsm,s_n\in S^{\text{aff}}$ such that, setting $w'\stackrel{\text{def}}=s_n\dotsm s_1 w s_1\dotsm s_n$, \begin{itemize} \item $\ell(s_i\dotsm s_1 w s_1\dotsm s_i)=\ell(w)$ for all $i$, \item $\ell(sw's)>\ell(w')$. \end{itemize} \end{thm} \begin{lem}\label{4} Suppose $h\in \mathcal{Z}_{R}(0,c_{\tilde{s}})$ and $\tilde{w}\in\text{supp}(h)_{\max}$, then $\tilde{w}\in \Lambda(1)$. \end{lem} \begin{proof} We prove by contradiction. Assume $\tilde{w}\in\text{supp}(h)_{\max}$ but $\tilde{w}\notin \Lambda(1)$. By Theorem \ref{12}, $\exists \tilde{s}\in S^{\text{aff}}(1)$ and $\tilde{s}_1,\tilde{s}_2,\dotsm,\tilde{s}_n\in S^{\text{aff}}(1)$ such that \begin{itemize} \item $\ell(\tilde{s}_i\dotsm\tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\dotsm\tilde{s}_i^{-1})=\ell(\tilde{w})$ for all $i$, \item $\pi(\tilde{s}_i\tilde{s}_{i-1}\dotsm \tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\dotsm \tilde{s}_{i-1}^{-1}\tilde{s}_i^{-1})\neq \pi(\tilde{s}_{i-1}\dotsm \tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\dotsm \tilde{s}_{i-1}^{-1})$ for all $i$, \item $\ell(\tilde{s}\tilde{w'}\tilde{s}^{-1})>\ell(\tilde{w'})$, where $\tilde{w'}=\tilde{s}_n\dotsm \tilde{s}_2\tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\tilde{s}_2^{-1}\dotsm \tilde{s}_n^{-1}$. \end{itemize} By Lemma \ref{3}, $\tilde{s}_i\dotsm \tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\dotsm\tilde{s}_i^{-1}\in \text{supp}(h)_{\max}$ for all $i$, in particular, $\tilde{w'}=\tilde{s}_n\dotsm\tilde{s}_2\tilde{s}_1\tilde{w}\tilde{s}_1^{-1}\tilde{s}_2^{-1}\dotsm\tilde{s}_n^{-1}\in \text{supp}(h)_{\max}$. By Lemma \ref{3} again, $\tilde{s}\tilde{w'}\tilde{s}^{-1}\in \text{supp}(h)_{\max}$. But $\ell(\tilde{s}\tilde{w'}\tilde{s}^{-1})>\ell(\tilde{w'})$, which is a contradiction. \end{proof} \begin{proof}[Proof of Theorem \ref{11}] It suffices to show that if $h\in \mathcal{Z}_{R}(0,c_{\tilde{s}})$, $\tilde{w}\in\text{supp}(h)_{\max}$ and $Cl(\tilde{w})$ is the $W(1)$-conjugacy class of $\tilde{w}$ in $W(1)$, then $Cl(\tilde{w})\subseteq\text{supp}(h)_{\max}$ and $a_{\tilde{w'}}=a_{\tilde{w}}$ for any $\tilde{w'}\in Cl(\tilde{w})$. By Lemma \ref{3} and Lemma \ref{4}, $\tilde{x}\tilde{w}\tilde{x}^{-1}\in\text{supp}(h)_{\max}$ and $a_{\tilde{x}\tilde{w}\tilde{x}^{-1}}=a_{\tilde{w}}$ for any $\tilde{x}\in W^{\text{aff}}(1)$. It remains to show that $\tilde{\tau} \tilde{w}\tilde{\tau}^{-1}\in\text{supp}(h)_{\max}$ and $a_{\tilde{\tau} \tilde{w}\tilde{\tau}^{-1}}=a_{\tilde{w}}$ for any $\tilde{\tau}\in\Omega(1)$. But $\tilde{\tau}\tilde{w}\in \text{supp}(T_{\tilde{\tau}} h)=\text{supp}(h T_{\tilde{\tau}})$, and $$\text{supp}(T_{\tilde{\tau}} h)_{\max}=\{\tilde{\tau}\tilde{x}|\tilde{x}\in\text{supp}(h)_{\max}\},$$ $$\text{supp}(h T_{\tilde{\tau}})_{\max}=\{\tilde{y}\tilde{\tau}|\tilde{y}\in\text{supp}(h)_{\max}\}.$$ Both sets are nonempty because $\tilde{\tau}\tilde{w}\in \text{supp}(T_{\tilde{\tau}}h)_{\max}$. Therefore, $\tilde{\tau}\tilde{w}\tilde{\tau}^{-1}\in\text{supp}(h)_{\max}$. The $R$-coefficient of $T_{\tilde{\tau}\tilde{w}}$ in $T_{\tilde{\tau}} h$ is $a_{\tilde{w}}$ and the $R$-coefficient of $T_{\tilde{\tau}\tilde{w}}$ in $h T_{\tilde{\tau}}$ is $a_{\tilde{\tau}\tilde{w}\tilde{\tau}^{-1}}$. Thus $a_{\tilde{\tau}\tilde{w}\tilde{\tau}^{-1}}=a_{\tilde{w}}$. \end{proof} It is well-known that a conjugacy class $C$ of $W$ is finite if and only if $C$ is contained in $\Lambda$. The same is true for $W(1)$ and $\Lambda(1)$. So $\text{supp}(h)$ is a finite union of some conjugacy classes in $\Lambda(1)$. In fact, Theorem $4.1$ is valid even without the restriction on $q_{\tilde{s}}$, because the proof does not use the fact that $q_{\tilde{s}}=0$. \section{Some Technical Results} Let $C$ be a finite conjugacy class in $W(1)$. Set $$h_{\lambda,C}=\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}T_{\tilde{\lambda}},$$ for every $\lambda\in\pi(C)$. We may just write $h_{\lambda}$ for $h_{\lambda,C}$ when there is no ambiguity. Now we prove some properties of $r_{x,\lambda}(h_{\lambda})$. \begin{lem}\label{6} Let $C$ be a finite conjugacy class in $W(1)$, $\lambda\in\pi(C)$ and $s\in S^{\text{aff}}$. Let $x\in W$ with $x<sx$ or $x<xs$. Suppose that $x\leq \lambda$ and $x\leq s\lambda s$, then $$r_{x,\lambda}(h_{\lambda})=r_{x,s\lambda s}(h_{s\lambda s}).$$ \end{lem} \begin{proof} Without loss of generality, we may assume $x<sx$. If $s\lambda s=\lambda$, then it is clearly true. If $s\lambda s\neq\lambda$, then without loss of generality, we may assume $s\lambda<\lambda$. In this case, $x\leq s\lambda$ by Lemma \ref{14}. Thus $$r_{x,\lambda}(h_{\lambda})=r_{x,s\lambda}(r_{s\lambda,\lambda}(h_{\lambda})),\quad r_{x,s\lambda s}(h_{s\lambda s})=r_{x,s\lambda}(r_{s\lambda,s\lambda s}(h_{s\lambda s})).$$ It suffices to show that $r_{s\lambda,\lambda}(h_{\lambda})=r_{s\lambda,s\lambda s}(h_{s\lambda s})$. Since $c_{\tilde{s}^{-1}}\in R[Z]$, we may assume that $$c_{\tilde{s}^{-1}}=\sum_{t\in Z}b_t t,\quad\text{for some}\ \ b_t\in R.$$ Then \begin{align}\nonumber &r_{s\lambda,\lambda}(h_{\lambda})=r_{s\lambda,s\lambda s}(h_{s\lambda s})\\\nonumber \Longleftrightarrow&\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-c_{\tilde{s}^{-1}}T_{\tilde{s}\tilde{\lambda}}=\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-T_{\tilde{s}\tilde{\lambda}}c_{\tilde{s}^{-1}}\\\nonumber \Longleftrightarrow&\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-c_{\tilde{s}^{-1}}T_{\tilde{s}\tilde{\lambda}}=\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-T_{\tilde{s}\tilde{\lambda}}(\tilde{s}c_{\tilde{s}^{-1}}\tilde{s}^{-1})\\\nonumber \Longleftrightarrow&\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-(\sum_{t\in Z}b_t t)T_{\tilde{s}\tilde{\lambda}}=\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-T_{\tilde{s}\tilde{\lambda}}(\tilde{s}(\sum_{t\in Z}b_t t)\tilde{s}^{-1})\\\nonumber \Longleftrightarrow&\sum_{t\in Z}b_t\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-T_{t\tilde{s}\tilde{\lambda}}=\sum_{t\in Z}b_t\sum_{\tilde{\lambda}\in\pi^{-1}(\lambda)\cap C}-T_{\tilde{s}\tilde{\lambda}(\tilde{s}t\tilde{s}^{-1})}.\nonumber \end{align} We want to show the last equation. It suffices to show that \begin{align}\nonumber &\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}t\tilde{s}\tilde{\lambda}=\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{s}\tilde{\lambda}(\tilde{s}t\tilde{s}^{-1})\\\nonumber \Longleftrightarrow&\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{s}^{-1}t\tilde{s}\tilde{\lambda}=\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{\lambda}(\tilde{s}t\tilde{s}^{-1})\\\nonumber \Longleftrightarrow&(\tilde{s}^{-1}t\tilde{s})(\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{\lambda})(\tilde{s}^{-1}t\tilde{s})^{-1}=\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{\lambda}\nonumber \end{align} for any $t\in Z$. The last equation holds because $\sum_{\tilde{\lambda}\in \pi^{-1}(\lambda)\cap C}\tilde{\lambda}$ is fixed by $Z$. \end{proof} For $w,w'\in W$, we write $w\stackrel{s}\rightarrow w'$ if $w'=sws$ and $\ell(w')=\ell(w)-2$. For $w,w'\in W$, we write $w\stackrel{s}\sim w'$ if $w'=sws$, $\ell(w')=\ell(w)$, and $sw>w$ or $ws>w$. We write $w\sim w'$ if $\exists$ a sequence $$w=w_0,w_1,...,w_n=w'$$ such that $w_{i-1}\stackrel{s_i}{\sim} w_i$ for every $i$ and some $s_i\in S^{\text{aff}}$. If $\lambda,\lambda'$ are in the same finite conjugacy class in $W$, then $\lambda'=w\lambda w^{-1}$ for some $w\in W$. Since $W=\Lambda\rtimes W_0$, we can write $w=w_0\lambda''$ for some $w_0\in W_0$ and $\lambda''\in\Lambda$. Thus by commutativity of $\Lambda$, $\lambda'=(w_0\lambda'')\lambda(w_0\lambda'')^{-1}=w_0\lambda w_0^{-1}$. Therefore, $\lambda\sim\lambda'.$ \begin{lem}\label{7} Let $C$ be a finite conjugacy class in $W(1)$ and $\lambda\in\pi(C)$. Let $x,x'\in W$ and $x\leq\lambda$. Suppose $x\sim x'$ is given by $$x=x_0\stackrel{s_1}\sim x_1\stackrel{s_2}\sim\dotsm\stackrel{s_n}\sim x_n=x',$$ for some $s_i\in S^{\text{aff}}$. Let $w=s_n\dotsm s_1$. Then $\exists \lambda'\sim\lambda, \tilde{w}\in W(1)$ a lifting of $w$, such that $x'\leq\lambda'$ and $$\tilde{w}\bullet(r_{x,\lambda}(h_{\lambda}))=r_{x',\lambda'}(h_{\lambda'}).$$ \end{lem} \begin{proof} It suffices to prove the case where $x\stackrel{s} \sim x'$ for some $s\in S^{\text{aff}}$, i.e. $x'=sxs$. Without loss of generality, we may assume that $sx>x$. \begin{itemize} \item If $s\lambda>\lambda$, then by Lemma \ref{14} $sxs\leq s\lambda s$. It is enough to show that $T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})=r_{sxs,s\lambda s}(h_{s\lambda s})T_{\tilde{s}}$ for any $\tilde{s}\in S^{aff}(1)$ with $\pi(\tilde{s})=s$. But \begin{align}\nonumber T_{\tilde{s}}r_{x,\lambda}(h_{\lambda}) &=r_{sx,s\lambda}(T_{\tilde{s}}h_{\lambda})\\\nonumber &=r_{sx,s\lambda}(h_{s\lambda s}T_{\tilde{s}})\\\nonumber &=r_{sxs,s\lambda s}(h_{s\lambda s})T_{\tilde{s}}.\nonumber \end{align} The second the equality holds because $\tilde{s}\bullet h_{\lambda}=h_{s\lambda s}$, and the other equalities hold by Proposition \ref{2}. Therefore, $$\tilde{s}\bullet(r_{x,\lambda}(h_{\lambda}))=r_{x',s\lambda s}(h_{s\lambda s}).$$ \item If $s\lambda<\lambda$, then by Lemma \ref{14} $sxs<sx\leq \lambda$ and $x\leq s\lambda< s\lambda s$. Therefore, for any $\tilde{s}\in S^{aff}(1)$ with $\pi(\tilde{s})=s$, we have \begin{align}\nonumber T_{\tilde{s}}(r_{x,\lambda}(h_{\lambda})) &=T_{\tilde{s}}(r_{x,s\lambda s}(h_{s\lambda s}))\\\nonumber &=(r_{sxs,\lambda}(h_{\lambda})).\nonumber \end{align} The first equality holds by Lemma \ref{6} and the second equality holds by a similar proof as the last step. Thus $$\tilde{s}\bullet(r_{x,\lambda}(h_{\lambda}))=r_{x',\lambda}(h_{\lambda}).$$ \end{itemize} This finishes the proof. \end{proof} Recall that $\nu$ is the homomorphism which defines the action of $\Lambda$. Set $\Lambda^+=\{\lambda\in\Lambda| \beta(\nu(\lambda))\geq 0, \forall \beta\in\Sigma^+\}$ where $\Sigma^+$ is the set of positive roots in $\Sigma$. A element in $\Lambda$ is called dominant if it is contained in $\Lambda^+$. Let $\mu_0\in\Lambda^+$ and $\lambda\in\Lambda$. Let $\lambda_0$ be the unique dominant element in $\{\lambda'\in\Lambda| \lambda'\sim\lambda\}$. Suppose $\mu_0\leq\lambda$, then by Theorem $1.3$ in \cite{hh} and Proposition $6.1$ in \cite{th}, $\mu_0\leq\lambda_0$. We have the following result. \begin{lem}\label{8} Let $\mu_0\in\Lambda^+ and \lambda\in\Lambda$. Let $\lambda_0$ be the unique dominant element in $\{\lambda'\in\Lambda| \lambda'\sim\lambda\}$. Suppose $\mu_0\leq\lambda$, then $\exists$ a sequence $$\lambda_0,\lambda_1,\dotsm,\lambda_n=\lambda$$ such that $\lambda_{i-1}\stackrel{s_i}\sim\lambda_i$ for every $i$ and some $s_i\in S_0$, and $\mu_0\leq\lambda_i$ for all $i$. \end{lem} \begin{proof} Since $\lambda\sim\lambda_0$, there exists $w\in S_0$ such that $\lambda=w\lambda_0 w^{-1}$. We prove the statement by induction on $l=\ell(w)$. If $l=0,1$, then it is obvious. Now suppose that the statement is correct for $l<k$, and we consider the case when $l=k$. Let $w=s_{i_k}\dotsm s_{i_1}$ and it suffices to show that $\mu_0\leq s_{i_k}\lambda s_{i_k}$. If $s_{i_k}\lambda s_{i_k}=\lambda$, then it is obvious. If $s_{i_k}\lambda s_{i_k}\neq \lambda$, then by $s_{i_k}w<w$ and Lemma \ref{13}, $w(\mathfrak{A}_0)$ and $\mathfrak{A}_0$ are on different sides of $H_{s_{i_k}}$. On the other hand, $s_{i_k}\lambda s_{i_k}\neq \lambda$, then $\nu(\lambda)=\nu(w\lambda_0 w^{-1})=w(\nu(\lambda_0))\in w(\bar{\mathfrak{C}}_0)\backslash H_{s_{i_k}}$. Thus $\lambda(\mathfrak{A}_0)=\mathfrak{A}_0+\nu(\lambda)$ and $\mathfrak{A}_0$ are on different sides of $H_{s_{i_k}}$, i.e. $s_{i_k}\lambda<\lambda$ by Lemma \ref{13}. We also have $s_{i_k}\mu_0>\mu_0$, thus by Lemma \ref{14} $\mu_0\leq s_{i_k}\lambda<s_{i_k}\lambda s_{i_k}$, which finishes the proof. \end{proof} \begin{thm}\label{9} Let $C$ be a finite conjugacy class in $W(1)$, $\lambda_1,\lambda_2\in \pi(C)$ and $x\in W$. Suppose $x\leq\lambda_1,\lambda_2$, then $$r_{x,\lambda_1}(h_{\lambda_1})=r_{x,\lambda_2}(h_{\lambda_2}).$$ \end{thm} \begin{proof} We prove it by induction on $d=\ell(\lambda_1)-\ell(x)=\ell(\lambda_2)-\ell(x)$. If $d=0$, then it is obvious since $x=\lambda_1=\lambda_2.$ Now suppose $d>0$. \begin{itemize} \item If $x\notin \Lambda$, then by Theorem \ref{12} there exist $s_1,s_2,\dotsm,s_n,s'\in S^{\text{aff}}$ such that $s_is_{i-1}\dotsm s_1xs_1\dotsm s_{i-1}s_i\stackrel{s_{i+1}}\sim s_{i+1}s_is_{i-1}\dotsm s_1xs_1\dotsm s_{i-1}s_is_{i+1}$ for all $i$ and $s's_ns_{n-1}\dotsm s_1xs_1\dotsm s_{n-1}s_ns'\stackrel{s'}\rightarrow s_ns_{n-1}\dotsm s_1xs_1\dotsm s_{n-1}s_n$. Let $\tilde{w}\in W^{\text{aff}}(1)$ be a lifting of $s_ns_{n-1}\dotsm s_1$ and $x'=s_ns_{n-1}\dotsm s_1xs_1\dotsm s_{n-1}s_n$. Then by Lemma \ref{7}, $$\tilde{w}\bullet(r_{x,\lambda_1}(h_{\lambda_1}))=r_{x',\lambda_1'}(h_{\lambda_1'}),\quad \tilde{w}\bullet(r_{x,\lambda_2}(h_{\lambda_2}))=r_{x',\lambda_2'}(h_{\lambda_2'}),$$ for some $\lambda_1'\sim\lambda_1, \lambda_2'\sim\lambda_2$. We have $\lambda_1'\sim\lambda_2'$ because $\lambda_1\sim\lambda_2$. It suffices to show that $r_{x',\lambda_1'}(h_{\lambda_1'})=r_{x',\lambda_2'}(h_{\lambda_2'}).$ It can be easily checked by Lemma \ref{14} that $s'x'\leq \lambda_j'$ or $s'\lambda_j's'$ for $j=1,2$. Without loss of generality, we may assume that $s'x'\leq \lambda_1',\lambda_2'$, then $$r_{x',\lambda_1'}(h_{\lambda_1'})=r_{x',s'x'}(r_{s'x',\lambda_1'}(h_{\lambda_1'}))=r_{x',s'x'}(r_{s'x',\lambda_2'}(h_{\lambda_2'}))=r_{x',\lambda_2'}(h_{\lambda_2'}),$$ where the second equality holds by induction. If $s'x'\leq s'\lambda_j's'$, then $x'<s'\lambda_j's'$. By Lemma \ref{6}, $$r_{x',\lambda_j'}(h_{\lambda_j'})=r_{x',s'\lambda_j's'}(h_{s'\lambda_j's'})=r_{x',s'x'}(r_{s'x',s'\lambda_j's'}(h_{s'\lambda_j's'})),$$ and we can apply a similar proof as above. \item If $x\in \Lambda$, then there exist $w\in S_0$ such that $x_0=wxw^{-1}\in\Lambda^+$. Let $\tilde{w}\in W(1)$ be a lifting of $w$, then by Lemma \ref{7}, $$\tilde{w}\bullet(r_{x,\lambda_1}(h_{\lambda_1}))=r_{x_0,\lambda_1'}(h_{\lambda_1'}),\quad \tilde{w}\bullet(r_{x,\lambda_2}(h_{\lambda_2}))=r_{x_0,\lambda_2'}(h_{\lambda_2'}),$$ for some $\lambda_1'\sim\lambda_1,\lambda_2'\sim\lambda_2$. We have $\lambda_1'\sim\lambda_2'$ because $\lambda_1\sim\lambda_2$. It suffices to show that $r_{x_0,\lambda_1'}(h_{\lambda_1'})=r_{x_0,\lambda_2'}(h_{\lambda_2'})$. By Lemma \ref{7} and \ref{8}, $r_{x_0,\lambda_1'}(h_{\lambda_1'})=r_{x_0,\lambda_0}(h_{\lambda_0})=r_{x_0,\lambda_2'}(h_{\lambda_2'})$ where $\lambda_0\in\Lambda^+$ and $\lambda_0\sim\lambda_1',\lambda_0\sim\lambda_2'$. \end{itemize} This finishes the proof. \end{proof} \section{Center of $\mathcal{H}_R(0,c_{\tilde{s}})$} Let $C$ be a finite conjugacy class in $W(1)$. Then $C\subset\Lambda(1), \pi(C)\subset\Lambda$ and there is a unique element $\lambda_0\in\pi(C)\cap\Lambda^+$. Set $$\text{Adm}(C)=\text{Adm}(\lambda_0)=\{w\in W| w\leq\lambda\ \ \text{for some}\ \ \lambda\in\pi(C)\}.$$ We define $$h_C=\sum_{w\in \text{Adm}(C)} h_w,$$ where $h_w=h_{w,C}$ if $w\in\pi(C)$ and otherwise $h_{w}=r_{w,\lambda}(h_{\lambda})$ for any $\lambda\in \pi(C)$ with $\lambda>w$. By Theorem \ref{9}, $h_C$ is well defined. \begin{lem} Suppose $C$ be a finite conjugacy class in $W(1)$. Then $h_C\in\mathcal{Z}_R(0,c_{\tilde{s}})$. \end{lem} \begin{proof} For any $\tilde{\tau}\in\Omega(1)$ with $\pi(\tilde{\tau})=\tau$, \begin{align}\nonumber T_{\tilde{\tau}}h_C &=\sum_{w\in\text{Adm}(C)}T_{\tilde{\tau}}h_w\\\nonumber &=\sum_{w\in\text{Adm}(C)}h_{\tau w\tau^{-1}}T_{\tilde{\tau}}\\\nonumber &=(\tilde{\tau}\bullet(\sum_{w\in\text{Adm}(C)}h_w))T_{\tilde{\tau}}\\\nonumber &=h_C T_{\tilde{\tau}}.\nonumber \end{align} The second equality holds by definition of $h_C$ and Proposition \ref{2}, and the third equality holds because $h_C$ is stable under the action of $W(1)$. It remains to show that for any $\tilde{s}\in S^{\text{aff}}(1)$ with $\pi(\tilde{s})=s$, $T_{\tilde{s}}h_C=h_CT_{\tilde{s}}$. The left hand side $$T_{\tilde{s}}h_C=\sum_{w\in\text{Adm}(C)}T_{\tilde{s}}h_w=\sum_{x,sx\in\text{Adm}(C)}T_{\tilde{s}}h_x+\sum_{y\in\text{Adm}(C),sy\notin\text{Adm}(C)}T_{\tilde{s}}h_y.$$ If $x,sx\in\text{Adm}(C)$, then without loss of generality, we may assume $x<sx\leq \lambda\in\pi(C)$. In this case, \begin{align}\nonumber T_{\tilde{s}}h_x+T_{\tilde{s}}h_{sx} &=T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})+T_{\tilde{s}}r_{sx,\lambda}(h_{\lambda})\\\nonumber &=T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})+c_{\tilde{s}}r_{sx,\lambda}(h_{\lambda})\\\nonumber &=T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})+T_{\tilde{s}}(-r_{x,sx}(r_{sx,\lambda}(h_{\lambda})))\\\nonumber &=T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})+T_{\tilde{s}}(-r_{x,\lambda}(h_{\lambda}))\\\nonumber &=0.\nonumber \end{align} The second equality holds because $T_{\tilde{s}}T_{\tilde{s}\tilde{x}}=c_{\tilde{s}}T_{\tilde{s}\tilde{x}}$ for any $\tilde{x}\in W(1)$ with $\pi(\tilde{x})=x$. The third equality holds because $c_{\tilde{s}}T_{\tilde{s}}=T_{\tilde{s}}c_{\tilde{s}}$ and $c_{\tilde{s}}T_{\tilde{s}\tilde{x}}=T_{\tilde{s}}(c_{\tilde{s}}T_{\tilde{x}})=T_{\tilde{s}}(-r_{x,sx}(T_{\tilde{s}\tilde{x}}))$ for any $\tilde{x}\in W(1)$ with $\pi(\tilde{x})=x$. The fourth equality holds by Proposition \ref{2}. Therefore, $$T_{\tilde{s}}h_C=\sum_{x\in\text{Adm}(C),sx\notin\text{Adm}(C)}T_{\tilde{s}}h_x.$$ Similarly, $$h_CT_{\tilde{s}}=\sum_{x\in\text{Adm}(C),xs\notin\text{Adm}(C)}h_xT_{\tilde{s}}.$$ But it is easy to check by Lemma \ref{14} that there is a one-to-one correspondence between the two sets $\{x\in\text{Adm}(C)|sx\notin\text{Adm}(C)\}$ and $\{x\in\text{Adm}(C)|xs\notin\text{Adm}(C)\}$, i.e., $y\in\{x\in\text{Adm}(C)|sx\notin\text{Adm}(C)\}$ if and only if $sys\in\{x\in\text{Adm}(C)|xs\notin\text{Adm}(C)\}$. Therefore, it is enough to show that if $x\in\text{Adm}(C)$ and $sx\notin\text{Adm}(C)$, then $$T_{\tilde{s}}h_x=h_{sxs}T_{\tilde{s}}.$$ Now $x<sx$, and we suppose $x\leq\lambda\in\pi(C)$. If $s\lambda>\lambda$, then by Lemma \ref{14} $sxs\leq s\lambda s$, thus \begin{align}\nonumber T_{\tilde{s}}h_x &=T_{\tilde{s}}r_{x,\lambda}(h_{\lambda})\\\nonumber &=r_{sx,s\lambda}(T_{\tilde{s}}h_{\lambda})\\\nonumber &=r_{sx,s\lambda}(h_{sxs}T_{\tilde{s}})\\\nonumber &=r_{sxs,s\lambda s}(h_{s\lambda s})T_{\tilde{s}}\\\nonumber &=h_{sxs}T_{\tilde{s}}.\nonumber \end{align} The second and fourth equalities hold by Proposition \ref{2}. The third equality holds because $\tilde{s}\bullet h_{\lambda}=h_{s\lambda s}$. If $s\lambda<\lambda$, then by Lemma \ref{14} $sx\leq\lambda$, but $\lambda<\lambda s$ so by Lemma \ref{14} again $sxs\leq\lambda$ and $sx\leq\lambda s$, therefore $x\leq s\lambda s$. Now let $y=sxs$, then $y\leq\lambda$ and $sys\leq s\lambda s$, therefore applying a similar proof as above, we have $h_yT_{\tilde{s}}=T_{\tilde{s}}h_{sys}$, i.e., $T_{\tilde{s}}h_x=h_{sxs}T_{\tilde{s}}$. This finishes the proof. \end{proof} \begin{thm} The center $\mathcal{Z}_R(0,c_{\tilde{s}})$ of $\mathcal{H}_R(0,c_{\tilde{s}})$ has a basis $\{h_C\}_{C\in\mathcal{F}(W(1))}$, where $\mathcal{F}(W(1))$ is the family of finite conjugacy classes in $W(1)$. \end{thm} \begin{proof} First, $\{h_C\}_{C\in\mathcal{F}(W(1))}$ is linearly independent since $\text{supp}(h_C)_{\max}$ differs. Next, we show that $\bigoplus_{C\in\mathcal{F}(W(1))}Rh_C = \mathcal{Z}_R(0,c_{\tilde{s}})$. We prove this by contradiction, so assume that $\bigoplus_{C\in\mathcal{F}(W(1))}Rh_C\subsetneqq \mathcal{Z}_R(0,c_{\tilde{s}})$. Let $h$ be an element in $\mathcal{Z}_R(0,c_{\tilde{s}})-\bigoplus_{C\in\mathcal{F}(W(1))}Rh_C$ and $\max_{w\in\text{supp}(h)}\ell(w)\leq \max_{w\in \text{supp}(h')}\ell(w)$ for any $h'\in \mathcal{Z}_R(0,c_{\tilde{s}})-\bigoplus_{C\in\mathcal{F}(W(1))}Rh_C$. Let $h=\sum_{\tilde{w}\in W(1)}a_{\tilde{w}}T_{\tilde{w}}$ and $M=\{\tilde{w}\in W(1)| \tilde{w} \ \ \text{maximal length with}\ \ a_{\tilde{w}}\neq 0\}$. By Theorem \ref{11}, $M$ is a union of some finite conjugacy class $C_i$'s. If $\tilde{w},\tilde{w}'\in C_i$ for some $i$, then $a_{\tilde{w}}= a_{\tilde{w}'}$, so we set $a_{C_i}=a_{\tilde{w}}$ for any $\tilde{w}\in C_i$. Let $h'=h-\sum_i a_{C_i} h_{C_i}$, then $h'\in \mathcal{Z}_R(0,c_{\tilde{s}})-\bigoplus_{C\in\mathcal{F}(W(1))}Rh_C$. But $\max_{w\in\text{supp}(h')}\ell(w)<\max_{w\in\text{supp}(h)}\ell(w)$. That is a contradiction. \end{proof} \section{Examples} Given a finite conjugacy class $C$ in $W(1)$, we can write out the corresponding central element $h_C$ as follow. First we know $\pi(C)$, so we can write out $h_{\lambda,C}$ for each $\lambda\in\pi(C)$. For other $x\in\text{Adm}(C)$, it is easy to find a $\lambda\in\pi(C)$ such that $x<\lambda$. Then we can apply the operator $r_{x,\lambda}$ on $h_{\lambda,C}$ by changing some factors $T_{\tilde{s}}$ to $-c_{\tilde{s}}$. Adding up all these terms, we get $h_C$. In this section, we give two examples to show how the above process works. \begin{example} In $GL_2$ case, the Iwahori Weyl group $W=W^{\text{aff}}\rtimes\Omega$. The affine Weyl group $W^{\text{aff}}$ is generated by $S^{\text{aff}}=\{s_0,s_1\}$. The group $\Omega$ is generated by $\tau$ and $\tau s_0=s_1\tau,\tau s_1=s_0\tau$. Suppose $C_1$ is a finite conjugacy class in $W(1)$ with $$\pi(C_1)=\{s_0s_1s_0s_1,s_1s_0s_1s_0\}.$$ Then $\text{Adm}(C_1)=\{s_0s_1s_0s_1,s_1s_0s_1s_0,s_0s_1s_0,s_1s_0s_1,s_0s_1,s_1s_0,s_0,s_1,1\}$. Suppose $$h_{s_0s_1s_0s_1,C_1}=\sum_{t\in Z_1}T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0\tilde{s}_1t},$$ for some subset $Z_1\subseteq Z$. Then $$h_{s_1s_0s_1s_0,C_1}=\sum_{t\in Z_1}T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1t\tilde{s}_0},$$ where $\tilde{s}_1\tilde{s}_0\tilde{s}_1t\tilde{s}_0$ is indeed a lifting of $s_1s_0s_1s_0$. Since $s_0s_1s_0,s_1s_0s_1<s_0s_1s_0s_1$, we have $$h_{s_0s_1s_0,C_1}=r_{s_0s_1s_0,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}-T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0}c_{\tilde{s}_1t},$$ $$h_{s_1s_0s_1,C_1}=r_{s_1s_0s_1,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}-c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1t}.$$ Since $s_0s_1,s_1s_0<s_0s_1s_0s_1$, we have $$h_{s_0s_1,C_1}=r_{s_0s_1,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}c_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_0\tilde{s}_1t},$$ $$h_{s_1s_0,C_1}=r_{s_1s_0,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_0}c_{\tilde{s}_1t}.$$ Since $s_0,s_1<s_0s_1s_0s_1$, we have $$h_{s_0,C_1}=r_{s_0,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}-T_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}c_{\tilde{s}_1t},$$ $$h_{s_1,C_1}=r_{s_1,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}-c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}T_{\tilde{s}_1t}.$$ Since $1<s_0s_1s_0s_1$, we have $$h_{1,C_1}=r_{1,s_0s_1s_0s_1}(h_{s_0s_1s_0s_1,C_1})=\sum_{t\in Z_1}c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}c_{\tilde{s}_1t}.$$ We can easily tell that the parity of of sign is determined by length difference. Therefore the corresponding central element is \begin{align}\nonumber h_{C_1}&=\sum_{t\in Z_1}T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0\tilde{s}_1t}+T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1t\tilde{s}_0}-T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0}c_{\tilde{s}_1t}\\\nonumber &-c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1t} +c_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_0\tilde{s}_1t}+c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_0}c_{\tilde{s}_1t}\\\nonumber &-T_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}c_{\tilde{s}_1t}-c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}T_{\tilde{s}_1t}+c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}c_{\tilde{s}_1t}. \end{align} Suppose $C_2$ is another finite conjugacy class in $W(1)$ with $$\pi(C_2)=\{s_0s_1s_0\tau,s_1s_0s_1\tau\}.$$ Then $\text{Adm}(C_2)=\{s_0s_1s_0\tau,s_1s_0s_1\tau,s_0s_1\tau,s_1s_0\tau,s_0\tau,s_1\tau,\tau\}$. Suppose $$h_{s_0s_1s_0\tau,C_2}=\sum_{t\in Z_2}T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0\tilde{\tau}t},$$ for some subset $Z_2\subseteq Z$. Then $$h_{s_1s_0s_1\tau,C_2}=\sum_{t\in Z_2}T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1\tilde{s}_0\tilde{\tau}t\tilde{s}_1^{-1}}=\sum_{t\in Z_2}T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1\tilde{\tau}(\tilde{\tau}^{-1}\tilde{s}_0\tilde{\tau}t\tilde{s}_1^{-1})},$$ where $(\tilde{\tau}^{-1}\tilde{s}_0\tilde{\tau}t\tilde{s}_1^{-1})$ is an element in $Z$. So $\tilde{s}_1\tilde{s}_0\tilde{s}_1\tilde{\tau}(\tilde{\tau}^{-1}\tilde{s}_0\tilde{\tau}t\tilde{s}_1^{-1})$ is indeed a lifting of $s_1s_0s_1\tau$. Since $s_0s_1\tau,s_1s_0\tau<s_0s_1s_0\tau$, we have $$h_{s_0s_1\tau,C_2}=r_{s_0s_1\tau,s_0s_1s_0\tau}(h_{s_0s_1s_0\tau,C_2}),\quad h_{s_1s_0\tau,C_2}=r_{s_1s_0\tau,s_0s_1s_0\tau}(h_{s_0s_1s_0\tau,C_2}).$$ Since $s_0\tau,s_1\tau<s_0s_1s_0\tau$, we have $$h_{s_0\tau,C_2}=r_{s_0\tau,s_0s_1s_0\tau}(h_{s_0s_1s_0\tau,C_2}),\quad h_{s_1\tau,C_2}=r_{s_1\tau,s_0s_1s_0\tau}(h_{s_0s_1s_0\tau,C_2}).$$ Since $\tau<s_0s_1s_0\tau$, we have $$h_{\tau,C_2}=r_{\tau,s_0s_1s_0\tau}(h_{s_0s_1s_0\tau,C_2}).$$ Therefore the corresponding central element is \begin{align}\nonumber h_{C_2}&=\sum_{t\in Z_2}T_{\tilde{s}_0\tilde{s}_1\tilde{s}_0\tilde{\tau}t}+T_{\tilde{s}_1\tilde{s}_0\tilde{s}_1\tilde{\tau}(\tilde{\tau}^{-1}\tilde{s}_0\tilde{\tau}t\tilde{s}_1^{-1})}-T_{\tilde{s}_0\tilde{s}_1}c_{\tilde{s}_0}T_{\tilde{\tau}t}-c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_0\tilde{\tau}t}\\\nonumber &+c_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_0\tilde{\tau}t}+c_{\tilde{s}_0}T_{\tilde{s}_1}c_{\tilde{s}_0}T_{\tilde{\tau}t}-c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_0}T_{\tilde{\tau}t}. \end{align} \end{example} \begin{example} In $SL_3$ case, the Iwahori Weyl group $W=W^{\text{aff}}$. The affine Weyl group $W^{\text{aff}}$ is generated by $S^{\text{aff}}=\{s_0,s_1,s_2\}$ with braid relations $s_is_js_i=s_js_is_j$ for $i\neq j$. Suppose $C$ is a finite conjugacy class in $W(1)$ with $$\pi(C)=\{s_0s_1s_2s_1,s_1s_0s_1s_2,s_2s_0s_2s_1,s_1s_2s_1s_0,s_2s_1s_0s_1,s_1s_2s_0s_2\}.$$ Then $$\text{Adm}(C)=\{s_0s_1s_2s_1,s_1s_0s_1s_2,s_2s_0s_2s_1,s_1s_2s_1s_0,s_2s_1s_0s_1,s_1s_2s_0s_2,$$$$s_1s_2s_1,s_1s_0s_1,s_2s_0s_2,s_0s_1s_2,s_0s_2s_1,s_1s_0s_2,s_1s_2s_0,s_2s_1s_0,s_2s_0s_1,$$$$s_0s_1,s_0s_2,s_1s_2,s_2s_1,s_1s_0,s_2s_0,s_0,s_1,s_2,1\}.$$ Suppose $$h_{s_0s_1s_2s_1,C}=\sum_{t\in Z'}T_{\tilde{s}_0\tilde{s_1}\tilde{s}_2\tilde{s}_1t},$$ for some subset $Z'\subseteq Z$. Then \begin{align}\nonumber h_{s_1s_0s_1s_2,C}&=\sum_{t\in Z'}T_{\tilde{s}_1t\tilde{s}_0\tilde{s_1}\tilde{s}_2},\\\nonumber h_{s_2s_0s_2s_1,C}&=\sum_{t\in Z'}T_{\tilde{s}_2\tilde{s}_0\tilde{s_1}\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1}}=\sum_{t\in Z'}T_{\tilde{s}_2\tilde{s}_0\tilde{s}_2\tilde{s}_1(\tilde{s}_1^{-1}\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1})},\\\nonumber h_{s_1s_2s_1s_0,C}&=\sum_{t\in Z'}T_{\tilde{s_1}\tilde{s}_2\tilde{s}_1t\tilde{s}_0},\\\nonumber h_{s_2s_1s_0s_1,C}&=\sum_{t\in Z'}T_{\tilde{s}_2\tilde{s}_1t\tilde{s}_0\tilde{s_1}},\\\nonumber h_{s_1s_2s_0s_2,C}&=\sum_{t\in Z'}T_{\tilde{s}_2^{-1}\tilde{s_1}\tilde{s}_2\tilde{s}_1t\tilde{s}_0\tilde{s}_2}=\sum_{t\in Z'}T_{(\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1}\tilde{s}_1^{-1})\tilde{s}_1\tilde{s}_2\tilde{s}_0\tilde{s}_2},\nonumber \end{align} where $\tilde{s}_1^{-1}\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1},\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1}\tilde{s}_1^{-1}$ are elements in $Z$. So the elements $\tilde{s}_2\tilde{s}_0\tilde{s}_2\tilde{s}_1(\tilde{s}_1^{-1}\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1})$ and $(\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1}\tilde{s}_1^{-1})\tilde{s}_1\tilde{s}_2\tilde{s}_0\tilde{s}_2$ are indeed liftings of $s_2s_0s_2s_1$ and $s_1s_2s_0s_2$ respectively. Since $s_1s_2s_1,s_0s_1s_2,s_0s_2s_1<s_0s_1s_2s_1;s_1s_0s_1,s_1s_0s_2<s_1s_0s_1s_2;s_2s_0s_2<s_2s_0s_2s_1;s_1s_2s_0,s_2s_1s_0<s_1s_2s_1s_0;s_2s_0s_1<s_2s_1s_0s_1$, we have $$h_{s_1s_2s_1,C}=r_{s_1s_2s_1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),h_{s_1s_0s_1,C}=r_{s_1s_0s_1,s_1s_0s_1s_2}(h_{s_1s_0s_1s_2,C}),$$ $$h_{s_2s_0s_2,C}=r_{s_2s_0s_2,s_2s_0s_2s_1}(h_{s_2s_0s_2s_1,C}),h_{s_0s_1s_2,C}=r_{s_0s_1s_2,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),$$ $$h_{s_0s_2s_1,C}=r_{s_0s_2s_1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),h_{s_1s_0s_2,C}=r_{s_1s_0s_2,s_1s_0s_1s_2}(h_{s_1s_0s_1s_2,C}),$$ $$h_{s_1s_2s_0,C}=r_{s_1s_2s_0,s_1s_2s_1s_0}(h_{s_1s_2s_1s_0,C}),h_{s_2s_1s_0,C}=r_{s_2s_1s_0,s_1s_2s_1s_0}(h_{s_1s_2s_1s_0,C}),$$ $$h_{s_2s_0s_1,C}=r_{s_2s_0s_1,s_2s_1s_0s_1}(h_{s_2s_1s_0s_1,C}).$$ Since $s_0s_1,s_0s_2,s_1s_2,s_2s_1<s_0s_1s_2s_1;s_1s_0,s_2s_0<s_1s_2s_1s_0$, we have $$h_{s_0s_1,C}=r_{s_0s_1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),h_{s_0s_2,C}=r_{s_0s_2,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),$$ $$h_{s_1s_2,C}=r_{s_1s_2,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),h_{s_2s_1,C}=r_{s_2s_1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),$$ $$h_{s_1s_0,C}=r_{s_1s_0,s_1s_2s_1s_0}(h_{s_1s_2s_1s_0,C}),h_{s_2s_0,C}=r_{s_2s_0,s_1s_2s_1s_0}(h_{s_1s_2s_1s_0,C}).$$ Since $s_0,s_1,s_2<s_0s_1s_2s_1$, we have $$h_{s_0,C}=r_{s_0,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),h_{s_1,C}=r_{s_1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}),$$ $$h_{s_2,C}=r_{s_2,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}).$$ Since $1<s_0s_1s_2s_1$, we have $$h_{1,C}=r_{1,s_0s_1s_2s_1}(h_{s_0s_1s_2s_1,C}).$$ Therefore the corresponding central element is \begin{align}\nonumber h_C&=\sum_{t\in Z'}T_{\tilde{s}_0\tilde{s}_1\tilde{s}_2\tilde{s}_1t}+T_{\tilde{s}_1t\tilde{s}_0\tilde{s_1}\tilde{s}_2}+T_{\tilde{s}_2\tilde{s}_0\tilde{s}_2\tilde{s}_1(\tilde{s}_1^{-1}\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1})}\\\nonumber &+T_{\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_0}+T_{\tilde{s}_2\tilde{s}_1t\tilde{s}_0\tilde{s}_1}+T_{(\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1}\tilde{s}_1^{-1})\tilde{s}_1\tilde{s}_2\tilde{s}_0\tilde{s}_2}\\\nonumber &-c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_2\tilde{s}_1t}-T_{\tilde{s}_1t\tilde{s}_0\tilde{s_1}}c_{\tilde{s}_2}-T_{\tilde{s}_2\tilde{s}_0\tilde{s}_2}c_{\tilde{s}_1(\tilde{s}_1^{-1}\tilde{s}_2^{-1}\tilde{s}_1\tilde{s}_2\tilde{s}_1t\tilde{s}_2^{-1})}\\\nonumber &-T_{\tilde{s}_0\tilde{s}_1\tilde{s}_2}c_{\tilde{s}_1t}-T_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_2\tilde{s}_1t}-T_{\tilde{s}_1t\tilde{s}_0}c_{\tilde{s_1}}T_{\tilde{s}_2}\\\nonumber &-T_{\tilde{s}_1\tilde{s}_2}c_{\tilde{s}_1t}T_{\tilde{s}_0}-c_{\tilde{s}_1}T_{\tilde{s}_2\tilde{s}_1t\tilde{s}_0}-T_{\tilde{s}_2}c_{\tilde{s}_1t}T_{\tilde{s}_0\tilde{s}_1}\\\nonumber &+T_{\tilde{s}_0\tilde{s}_1}c_{\tilde{s}_2}c_{\tilde{s}_1t}+T_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_2}c_{\tilde{s}_1t}+c_{\tilde{s}_0}T_{\tilde{s}_1\tilde{s}_2}c_{\tilde{s}_1t}+c_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_2\tilde{s}_1t}\\\nonumber &+T_{\tilde{s}_1}c_{\tilde{s}_2}c_{\tilde{s}_1t}T_{\tilde{s}_0}+c_{\tilde{s}_1}T_{\tilde{s}_2}c_{\tilde{s}_1t}T_{\tilde{s}_0}-T_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_2}c_{\tilde{s}_1t}\\\nonumber &-c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_2}T_{\tilde{s}_1t}-c_{\tilde{s}_0}c_{\tilde{s}_1}T_{\tilde{s}_2}c_{\tilde{s}_1t}+c_{\tilde{s}_0}c_{\tilde{s}_1}c_{\tilde{s}_2}c_{\tilde{s}_1t}. \end{align} \end{example}

9b636cb87668da206e80b93531ccf95edb9d4a64

\section{Introduction} Recent advances of modern technology makes redundant data collection available for \textit{time-to-event} information, which facilitates observing and tracking the event of interests. However, due to different reasons, many events would lose tracking during observation period, which makes the data \textit{censored}. We only know that the true time to the occurrence of the event is larger or smaller than, or within the observation time, which have been defined as survivorship bias categorized into \textit{right-censored}, \textit{left-censored} and \textit{internal-censored} respectively \cite{lee2003statistical}. Survival analysis, a.k.a. time-to-event analysis \cite{lee2018deephit}, is a typical statistical methodology for modeling time-to-event data while handling censorship, which is a traditional research problem and has been studied over decades. The goal of survival analysis is to estimate the time until occurrence of the particular event of interest, which can be regarded as a regression problem \cite{lee2003statistical,wu2015predicting}. It can also be viewed as to predict the probability of the event occurring over the whole timeline \cite{wang2016functional,lee2018deephit}. Specifically, given the information of the observing object, survival analysis would predict the probability of the event occurrence at each time point. Nowadays, survival analysis has been widely used in real-world applications, such as clinical analysis in medicine research \cite{zhu2017wsisa,luck2017deep,katzman2018deepsurv} taking diseases as events and predicting survival time of patients; customer lifetime estimation in information systems \cite{jing2017neural,grob2018recurrent} which estimates the time until the next visit of users; market modeling in game theory fields \cite{wu2015predicting,wang2016functional} that predicts the event (i.e., winning) probability over the whole referral space. Because of the essential applications in the real world, the researchers in both academic and industrial fields have devoted great efforts to studying survival analysis in recent decades. Many works of survival analysis are from the view of traditional statistic methodology. Among them, Kaplan-Meier estimator \cite{kaplan1958nonparametric} bases on non-parametric counting statistics and forecasts the survival rate at coarse-grained level where different observing objects may share the same forecasting result, which is not suitable in recent personalized applications. Cox proportional hazard method \cite{cox1992regression} and its variants such as Lasso-Cox \cite{tibshirani1997lasso} assume specific stochastic process or base distribution with semi-parametric scaling coefficients for fine-tuning the final survival rate prediction. Other parametric methods either make specific distributional assumptions, such as Exponential distribution \cite{lee2003statistical} and Weibull distribution \cite{ranganath2016deep}. These methods pre-assume distributional forms for the survival rate function, which may not generalize very well in real-world situations. Recently, deep learning, i.e., deep neural network, has been paid huge attention and introduced to survival analysis in many tasks \cite{ranganath2016deep,grob2018recurrent,lee2018deephit}. However, in fact, many deep learning models for survival analysis \cite{katzman2018deepsurv,ranganath2016deep} actually utilize deep neural network as the enhanced feature extraction method \cite{lao2017deep,grob2018recurrent} and, worse still, rely on some assumptions of the base distributions for the survival rate prediction, which also suffers from the generalization problem. Lately, \citet{lee2018deephit} proposed a deep learning method for modeling the event probability without assumptions of the probability distribution. Nevertheless, they regard the event probability estimation as a pointwise prediction problem, and ignores the sequential patterns within neighboring time slices. Moreover, the gradient signal is too sparse and has little effect on most of the prediction outputs of this model, which is not effective enough for modeling time-to-event data. With the consideration of all the drawbacks within the existing literatures, in this paper we propose our Deep Recurrent Survival Analysis (DRSA) model for predicting the survival rate over time at fine-grained level, i.e., for each individual sample. To the best of our knowledge, this is the first work utilizing auto-regressive model for capturing the sequential patterns of the feature over time in survival analysis. Our model proposes a novel modeling view for time-to-event data, which aims at flexibly modeling the survival probability function rather than making any assumptions for the distribution form. Specifically, DRSA creatively predicts the \textit{conditional} probability of the event at each time given that the event \textit{non}-occurred before, and combines them through probability chain rule for estimating both the probability density function and the cumulative distribution function of the event over time, eventually forecasts the survival rate at each time, which is more reasonable and mathematically efficient for survival analysis. We train DRSA model by end-to-end optimization through maximum likelihood estimation, not only on the observed event among uncensored data, but also on the censored samples to reduce the survivorship bias. Through these modeling methods, our DRSA model can capture the sequential patterns embedded in the feature space along the time, and output more effective distributions for each individual sample at fine-grained level. The comprehensive experiments over three large-scale real-world datasets demonstrate that our model achieves significant improvements against state-of-the-art models under various metrics. \section{Related Works}\label{sec:related-works} \subsection{Learning over Censored Data} The event occurrence information of some samples may be lost, due to some limitation of the observation period or losing tracks during the study procedure \cite{wang2017machine}, which is called data \textit{censorship}. When dealing with time-to-event information, a more complex learning problem is to estimate the probability of the event occurrence at each time, especially for those samples without tracking logs after (or before) the observation time which is defined as \textit{right-censored} (or \textit{left-censored}) \cite{wang2017machine}. Survival analysis is a typical statistical methodology for modeling time-to-event data while handling censorship. There are two main streams of survival analysis. The first view is based on traditional statistics scattering in three categories. (i) Non-parametric methods including Kaplan-Meier estimator \cite{kaplan1958nonparametric} and Nelson-Aalen estimator \cite{andersen2012statistical} are solely based on counting statistics, which is too coarse-grained to perform personalized modeling. (ii) Semi-parametric methods such as Cox proportional hazard model \cite{cox1992regression} and its variants Lasso-Cox \cite{tibshirani1997lasso} assumes some base distribution functions with the scaling coefficients for fine-tuning the final survival rate prediction. (iii) Parametric models assume that the survival time or its logarithm result follows a particular theoretical distribution such as Exponential distribution \cite{lee2003statistical} and Weibull distribution \cite{ranganath2016deep}. These methods either base on statistical counting information or pre-assume distributional forms for the survival rate function, which generalizes not very well in real-world situations. The second school of survival analysis takes from machine learning perspective. Survival random forest which was first proposed in \cite{gordon1985tree} derives from standard decision tree by modeling the censored data \cite{wang2016functional} while its idea is mainly based on counting-based statistics. Other machine learning methodologies include Bayesian models \cite{ranganath2015survival,ranganath2016deep}, support vector machine \cite{khan2008support} and multi-task learning solutions \cite{Li2016mtlsa,alaa2017deep}. Note that, deep learning models have emerged in recent years. \citet{faraggi1995neural} first embedded neural network into Cox model to improve covariate relationship modeling. From that, many works applied deep neural networks into well-studied statistical models to improve feature extraction and survival analysis through end-to-end learning, such as \cite{ranganath2016deep,luck2017deep,lao2017deep,katzman2018deepsurv,grob2018recurrent}. Almost all the above models assume particular distribution forms which also suffers from the generalization problem in practice. \citet{biganzoli1998feed} utilized neural network directly to predict the survival rate for each sample and \citet{lisboa2003bayesian} extended it to a Bayesian network method. In \cite{lee2018deephit} the authors proposed a feed forward deep model to directly predict the probability density values at each time point and sum them for estimating the survival rate. However, in that paper, the gradient signal is quite sparse for the prediction outputs from the neural network. Moreover, to our best knowledge, none of the related literatures considers the sequential patterns within the feature space over time. We propose a recurrent neural network model predicting the \textit{conditional} probability of event at each time and estimate the survival rate through the probability chain rule, which captures the sequential dependency patterns between neighboring time slices and back-propagate the gradient more efficiently. \subsection{Deep Learning and Recurrent Model} Due to its adequate model capability and the support of big data, deep learning, a.k.a. deep neural network, has drawn great attention ranging from computer vision \cite{krizhevsky2012imagenet} and speech recognition \cite{graves2013speech} to natural language processing \cite{bahdanau2014neural} during the recent decades. Among them, recurrent neural network (RNN) whose idea firstly emerged two decades ago and its variants like long short-term memory (LSTM) \cite{hochreiter1997long} employ memory structures to model the conditional probability which captures dynamic sequential patterns. In this paper we borrow the idea of RNN and well design the modeling methodology for survival function regression. \section{Deep Recurrent Survival Analysis}\label{sec:methodology} In this section, we formulate the survival analysis problem and discuss the details of our proposed model. We take the view of \textit{right-censorship} which is the most common scenario in survival analysis \cite{kaplan1958nonparametric,cox1992regression,wang2017machine,lee2018deephit}. \subsection{Problem Definition}\label{sec:prob-def} We define $z$ as the variable of the true occurrence time for the event of interest if it has been tracked. We just simplify the occurrence of the event of interest as \textit{event}, and define the probabilistic density function (P.D.F.) of the true event time as $p(z)$, which means the probability that the event truely occurs at time $z$. Now that we have the P.D.F. of the event time, we can derive the \textit{survival rate} at each time $t$ as the C.D.F. as \begin{equation}\small S(t) = \text{Pr}(z > t) = \int_t^{\infty} p(z) dz ~, \label{eq:survival-definition} \end{equation} which is the probability of the observing object surviving, i.e., event not occurring, until the observed time $t$. Then the straightforward definition of the \textit{event rate}, i.e., the probability of event occurring before the observing time $t$, as that \begin{equation}\small W(t) = \text{Pr}(z \leq t) = 1 - S(t) = \int_0^t p(z) dz ~. \label{eq:win-prob-definition} \end{equation} The data of the survival analysis logs are represented as a set of $N$ triples $\{ \bs{x}, z, t \}_{i=1}^N$, where $t > 0$ is the observation time for the given sample. Here $z$ is left unknown (and we marked $z$ as null) for the censored samples without the observation of the true event time. $\bs{x}$ is the feature vector of the observation which encodes different information under various scenarios. Our goal is to model the distribution of the true event time $p(z)$ over all the historical time-to-event logs with handling the censored data of which the true event time is unknown. So the main problem of survival analysis is to estimate the probability distribution $p(z | \bs{x})$ of the event time with regard to the sample feature $\bs{x}$, for each sample. Formally speaking, the derived model is a ``mapping'' function $T$ which learns the patterns within the data and predicts the event time distribution over the time space as \begin{equation}\label{eq:task-function} p( z | \bs{x}) = T(\bs{x}) ~. \vspace{-5pt} \end{equation} \subsection{Discrete Time Model}\label{sec:disc-time} First of all, we present the definition of the \textit{conditional hazard rate} over continuous time as \begin{equation}\label{eq:hazard} h(t) = \lim_{\triangle t \rightarrow 0} \frac{\text{Pr}(t < z \leq t + \triangle t ~|~ z > t)}{\triangle t} ~, \end{equation} which models the instant occurrence probability of the event at time $t$ given that the event has not occurred before. Note that the concept of \textit{hazard rate} has been commonly utilized in many survival analysis literatures \cite{cox1992regression,faraggi1995neural,luck2017deep}. In the discrete context, a set of $L$ time slices $0 < t_1 < t_2 < \ldots < t_L$ is obtained which arises from the finite precision of time determinations. Analogously we may also consider the grouping of continuous time as $l = 1, 2, \ldots, L$ and uniformly divide disjoint intervals $V_l = (t_{l-1}, t_{l}]$ where $t_0 = 0$ and $t_l$ is the last observation interval boundary for the given sample, i.e., the tracked observation time in the logs. $V_L$ is the last time interval in the whole data space. This setting is appropriately suited in our task and has been widely used in clinical research \cite{Li2016mtlsa,lee2018deephit}, information systems \cite{jing2017neural,grob2018recurrent} and other related fields \cite{wu2015predicting,wang2016functional}. As such, our event rate function and survival rate function over discrete time space is \begin{equation}\label{eq:win-lose-def} \begin{aligned} W(t_l) & = \text{Pr}(z \leq t_l) = \sum_{j \leq l} \text{Pr}(z \in V_j) ~, \\ S(t_l) & = \text{Pr}(z > t_l) = \sum_{j > l} \text{Pr}(z \in V_j) ~, \end{aligned} \vspace{-5pt} \end{equation} where the input to the two functions is the observed time $t_l$ from the log. And the discrete event time probability function at the $l$-th time interval is \begin{equation}\label{eq:discrete-pdf-def}\small p_l = \text{Pr}(z \in V_l) = W(t_{l}) - W(t_{l-1}) = S(t_{l-1}) - S(t_{l}) ~. \end{equation} The discrete conditional hazard rate $h_l$, defined as the conditional probability as \begin{equation}\label{eq:discrete-instant-win-func}\small h_l = \text{Pr}(z \in V_l | z > t_{l-1}) = \frac{\text{Pr}(z \in V_l)}{\text{Pr}(z > t_{l-1})} = \frac{p_l}{S(t_{l-1})} ~, \end{equation} which approximates the continuous conditional hazard rate function $h(t_l)$ in Eq.~(\ref{eq:hazard}) as the intervals $V_l$ become infinitesimal. \subsection{Deep Recurrent Model}\label{sec:drsa} Till now, we have presented the discrete time model and discuss the death (i.e., event) and survival probability over the discrete time space. We here propose our DRSA model based on recurrent neural network with the parameter $\bs{\theta}$, which captures the sequential patterns for conditional probability $h^i_l$ at every time interval $V_l$ for the $i^{\text{th}}$ sample. The detailed structure of DRSA network is illustrated in Figure~\ref{fig:drsa}. At each time interval $V_l$, the $l$-th RNN cell predicts the instant hazard rate $h^i_l$ given the sample feature $\bs{x}^i$ and the current time $t_l$ conditioned upon the previous events as \begin{equation}\label{eq:discrete-hazard-function} \begin{aligned} h^i_l &= \text{Pr}(z \in V_l ~|~ z > t_{l-1}, ~ \bs{x}^i; \bs{\theta}) \\ &= f_{\bs{\theta}}(\bs{x}^i, t_l ~|~ \bs{r}_{l-1}) ~, \end{aligned} \end{equation} where $f_{\bs{\theta}}$ is the RNN function taking $(\bs{x}^i, t_l)$ as input and $h^i_l$ as output. $\bs{r}_{l-1}$ is the hidden vector calculated from the RNN cell at the last time step which contains the information about the conditional. It is quite natural for using the recurrent cell to model the conditional probability over time \cite{bahdanau2014neural}. In our paper we implement the RNN function as a standard LSTM unit \cite{hochreiter1997long}, which has been widely used in sequence data modeling. The details of the implementation of our RNN architecture can be referred in our supplemental materials and our reproductive code published in the experiment part. \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth]{figures/drsa.pdf} \caption{Detailed illustration of Deep Recurrent Survival Analysis (DRSA) model. Note that only the uncensored logs have the true event time and can calculate $p_z$ for the loss $L_z$. The calculation of $p_z$ and $S(t)$ have been derived in Eqs.~(\ref{eq:pdf-calc}) and (\ref{eq:discrete-survival-def}) respectively.}\label{fig:drsa} \vspace{-10pt} \end{figure} From Eqs.~(\ref{eq:win-lose-def}), (\ref{eq:discrete-instant-win-func}) and (\ref{eq:discrete-hazard-function}), we can easily derive the survival rate function $S(t)$ through \textit{probability chain rule}, and the corresponding event rate function $W(t)$ at the time $t$ for the $i^{\text{th}}$ individual sample as \begin{equation}\small \begin{aligned} S(t | \bs{x}^i; \bs{\theta}) &= \text{Pr}(t < z | \bs{x}^i; \bs{\theta}) \\ &= \text{Pr}(z \not\in V_1, z \not\in V_2, \dots, z \not\in V_{l^i} | \bs{x}^i; \bs{\theta}) \\ &= \text{Pr}(z \not\in V_1 |\bs{x}^i; \bs{\theta}) \cdot \text{Pr}(z \not\in V_2 | z \not\in V_1, \bs{x}^i; \bs{\theta}) \cdots \\ & ~~~~~~~~~~~~~\cdot \text{Pr}(z \not\in V_{l^i} | z \not\in V_1, \ldots, z \not\in V_{l^i-1}, \bs{x}^i; \bs{\theta}) \\ & = \prod_{l: l \leq l^i} \left[ 1- \text{Pr}(z \in V_l ~|~ z > t_{l-1} ,~ \bs{x}^i; \bs{\theta}) \right] \\ &= \prod_{l: l \leq l^i} (1 - h^i_l) ~, \\ \label{eq:discrete-survival-def} \end{aligned} \end{equation} \begin{equation}\small \begin{aligned} W(t | \bs{x}^i; \bs{\theta}) &= \text{Pr}(t \geq z | \bs{x}^i; \bs{\theta})= 1- \prod_{l: l \leq l^i} (1 - h^i_l)~, \label{eq:discrete-win-def} \end{aligned} \end{equation} where $l^i$ is the time interval index for the $i^{\text{th}}$ sample at $t^i$. Moreover, taking Eqs.~(\ref{eq:discrete-pdf-def}) and (\ref{eq:discrete-instant-win-func}) into consideration, the probability of time $z$ lying in the interval of $V_{l^i}$ for the $i^\text{th}$ sample is \begin{equation}\label{eq:pdf-calc} p^i_l = \text{Pr}(z \in V_{l^i} | \bs{x}^i; \bs{\theta}) = h^i_{l^i} \prod_{l: l < l^i} (1 - h^i_l) ~. \end{equation} By means of probability chain rule, it connects all the outputs of the conditional hazard rate $h$ at each individual time to the final prediction, i.e., the probability $p(z)$ of the true event time $z$ and the survival rate $S(t)$ at each time $t$. This feed-forward calculation guarantees that the gradient signal from the loss function can be transmitted through back-propagation more effectively comparing with \cite{lee2018deephit}, which will be discussed below. \subsection{Loss Functions} Since there is no ground truth of either the event time distribution or survival rate, here we maximize the log-likelihood over the empirical data distribution to learn our deep model. Specifically, we take three objectives as our losses. The first loss is to minimize the negative log-likelihood of the true event time $z = z^i$ over the \textit{uncensored} logs as \begin{equation}\label{eq:market-price-objective} \begin{aligned} L_z &= - \log \prod_{(\bs{x}^i, z^i) \in \mathbb{D}_{\text{uncensored}}} \text{Pr}(z \in V_{l^i} | \bs{x}^i; \bs{\theta}) \\ &= - \log \prod_{(\bs{x}^i, z^i) \in \mathbb{D}_{\text{uncensored}}} h^i_{l^i} \prod_{l: l < l^i} (1 - h^i_l) \\ &= - \sum_{(\bs{x}^i, z^i) \in \mathbb{D}_{\text{uncensored}}} \left[ \log h^i_{l^i} + \sum_{l: l < l^i} \log(1 - h^i_l) \right] ~, \end{aligned} \end{equation} where $l^i$ is the index of the interval of the true event time $z^i \in V_{l^i}$. The second loss is to minimize the negative partial log-likelihood of the event rate over the \textit{uncensored} logs as \begin{equation}\label{eq:winning-objective} \begin{aligned} L_{\text{uncensored}} &= - \log \prod_{(\bs{x}^i, t^i) \in \mathbf{D}_{\text{uncensored}}} \text{Pr}(t^i \geq z | \bs{x}^i; \bs{\theta}) \\ &= - \log \prod_{(\bs{x}^i, t^i) \in \mathbb{D}_{\text{uncensored}}} W(t^i | \bs{x}^i; \bs{\theta}) \\ &= - \sum_{(\bs{x}^i, t^i) \in \mathbb{D}_{\text{uncensored}}} \log \Big[ 1- \prod_{l: l \leq l^i} (1 - h^i_l) \Big] ~. \end{aligned} \end{equation} This loss adds more supervisions onto the predictions over the time range $(z^i,t^i)$ for the uncensored data than those \cite{katzman2018deepsurv,lee2018deephit,tibshirani1997lasso} merely supervise on the true event time $z^i$. Though the censored logs do not contain any information about the true event time, we would only know that the true event time $z$ is greater than our logged observing time $t^i$ then. Here we incorporate the partial log-likelihood embedded in the \textit{censored} logs as the third loss to correct the learning bias of our model as \begin{equation}\label{eq:losing-objective} \begin{aligned} L_{\text{censored}} &= - \log \prod_{(\bs{x}^i, t^i) \in \mathbb{D_{\text{censored}}}} \text{Pr}(z > t^i | \bs{x}^i; \bs{\theta}) \\ &= - \log \prod_{(\bs{x}^i, t^i) \in \mathbb{D_{\text{censored}}}} S(t^i | \bs{x}^i; \bs{\theta}) \\ &= - \sum_{(\bs{x}^i, t^i) \in \mathbb{D_{\text{censored}}}} \sum_{l: l \leq l^i} \log (1 - h^i_l) ~. \end{aligned} \end{equation} \subsection{Model Realization}\label{sec:model-realization} In this section, we unscramble some intrinsic properties of our deep survival model. First of all, we analyze the model effectiveness of DRSA. In \cite{lee2018deephit}, the proposed deep model directly predicts the event probability $p(t)$ and combines to estimate the survival rate as $S(t) = \sum_{t'\leq t}p(t')$ while ignoring the sequential patterns. As a result, the gradient signal would only have effect on the prediction at time $t$ individually. On the contrary, from Figure~\ref{fig:drsa}, we can see that our DRSA model is obviously more effective since the supervision would be directly back-propagated through the chain rule calculation to all the units with strict mathematical derivation, which guarantees to transmit the gradient more efficiently and effectively. We also explicitly model the sequential patterns by conditional hazard rate prediction and we will illustrate the advantage of that in the experiments. Then we take the view of censorship prediction of our methodology. As is known that there is a censoring status as an indicator of survival at the given time, for each sample as \begin{equation}\label{eq:lose-indicator} \small c^i(t^i)= \begin{cases} 0, & \text{if}~~~~~~~~~~ t^i \geq z^i ~, \cr 1, & \text{otherwise}~~ t^i < z^i ~. \end{cases} \end{equation} In the tracking logs, each sample $(\bs{x}^i, z^i, t^i)$ is uncensored where $c^i = 0$. While for the censored logs losing tracking at the observation time, the true event time $z^i$ is unknown but the tracker only has the idea that $z^i > t^i$, thus $c^i = 1$. Moreover, for the uncensored data, it is natural to ``push down'' the probability of survival $S(t^i)$. And for the censored data, it needs to ``pull up'' $S(t^i)$ since we ``observe event not occurred'' at time $t^i$. However, using only $L_z$ to supervise the prediction of $p(z^i)$ at time $z^i$ in Eq.~(\ref{eq:market-price-objective}) is insufficient. So that we incorporate the two partial likelihood losses $L_{\text{uncensored}}$ and $L_{\text{censored}}$ in Eqs.~(\ref{eq:winning-objective}) and (\ref{eq:losing-objective}). Therefore, taking Eqs.~(\ref{eq:winning-objective}) and (\ref{eq:losing-objective}) altogether, we may find that the combination of $L_{\text{uncensored}}$ and $L_{\text{censored}}$ describes the classification of survival status at time $t^i$ of each sample as \begin{align} \tiny & L_{\text{c}} = L_{\text{uncensored}} + L_{\text{censored}} \nonumber \\ &= - \log \prod_{(\bs{x}^i, t^i) \in \mathbb{D}_{\text{full}}} \left[ S(t^i | \bs{x}^i; \bs{\theta}) \right]^{c^i} \cdot \left[ 1 - S(t^i | \bs{x}^i; \bs{\theta}) \right]^{1-c^i} \label{eq:classification-loss} \\ &= - \sum_{(\bs{x}^i, t^i) \in \mathbb{D}_{\text{full}}} \left\{ c^i \cdot \log S(t^i | \bs{x}^i; \bs{\theta}) \right. \nonumber \\ &~~~~~~~~~~~~~~~~~~~~~~~ + \left. (1-c^i) \log \left[ 1 - S(t^i | \bs{x}^i; \bs{\theta}) \right] \right\} ~, \nonumber \end{align} which is the cross entropy loss for predicting the survival status at time $t^i$ given $\bs{x}^i$ over all the data $\mathbb{D}_{\text{full}} = \mathbb{D}_{\text{uncensored}} \bigcup \mathbb{D}_{\text{censored}}$. Combining all the objective functions and our goal is to minimize the negative log-likelihood over all the data samples including both uncensored and censored data as \begin{equation}\label{eq:total-loss} \arg \min_{\bs{\theta}}~~ \alpha L_z + (1-\alpha) L_{\text{c}} ~, \end{equation} where $\bs{\theta}$ is the model parameter in Eq.~(\ref{eq:discrete-hazard-function}) and the hyperparameter $\alpha$ controls the loss value balance between them. Specifically, $\alpha$ controls the magnitudes of the two losses at the same level to stabilize the model training. We also analyze the model efficiency in the supplemental material of this paper and the time complexity of model inference is the same as the traditional RNN model which has proven practical efficiency in the industrial applications \cite{zhang2014sequential}. \section{Experiments}\label{sec:exps} We evaluate our model with strong baselines in three real-world tasks. Moreover, we have published the implementation code for reproductive experiments\footnote{Reproductive code link: https://github.com/rk2900/drsa.}. \subsection{Tasks and Datasets} We evaluate all the compared models in three real-world tasks. We also published the processed full datasets\footnote{We have put sampled data in the published code. The three processed full datasets link: https://goo.gl/nUFND4.}. \begin{description} \item[CLINIC] is a dataset for tracking the patient clinic status \cite{knaus1995support}. Here the goal of survival analysis is to estimate the time till the event (death), and predict the probability of the event with waning effects of baseline physiologic variables over time. \item[MUSIC] is a user lifetime analysis dataset \cite{jing2017neural} that contains roughly 1,000 users with entire listening history from 2004 to 2009 on last.fm, a famous online music service. Here the event is the user visit to the music service and the goal is to predict the time elapsed from the last visit of one user to her next visit. \item[BIDDING] is a real-time bidding dataset in the computational advertising field \cite{ren2018bid,wang2016functional}. In this scenario, the time is correspondent to the bid price of the bidder and the event is just winning of the auction. The feature contains the auction request information. Many researchers \cite{wu2015predicting,wang2016functional} utilized survival analysis for unbiased winning probability estimation of a single auction while handling the losing (censored) logs without knowing the true winning price. \end{description} The statistics of the three datasets are provided in Table~\ref{tab:stat}. We split the CLINIC and MUSIC datasets to training and test sets with ratio of 4:1 and 6:1, respectively. For feature engineering, all the datasets have been one-hot encoded for both categorical and numerical features. The original BIDDING data have already been feature engineered and processed as training and test datasets. Note that, the true time of the event of all the testing data have been preserved for the performance evaluation. In these datasets, since all the time is integer value, we bucketize the discrete time interval as interval size $s_{\text{intv}}=1$ and the maximal time interval number $L$ is equal to the largest integer time in each dataset. The discussion about various interval sizes has been included in the supplemental materials. \begin{table}[t] \centering \caption{The statistics of the datasets. \#: number; AET: averaged event time. There are 9 independent subsets in BIDDING dataset so that we provide the overall statistics in this table and present the details in supplemental materials.}\label{tab:stat} \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|r|r|r|r|r|r|r|} \hline \multirow{2}*{Dataset} & \multirow{2}{*}{Total~\#~~} & Censored~ & {Censored} & AET~~~ & AET~~~~~~ & AET~~~~ & Feature \\ & & Data \# ~~ & Rate~~~ & ($\mathbb{D}_{\text{full}}$)~~ & ($\mathbb{D}_{\text{uncensored}}$) & ($\mathbb{D}_{\text{censored}}$)~ & \# ~~~ \\ \hline CLINIC & 6,036 & 797 & 0.1320 & 9.1141 & 5.3319 & 33.9762 & 14 \\ MUSIC & 3,296,328 & 1,157,572 & 0.3511 & 122.1709 & 105.2404 & 153.4522 & 6 \\ BIDDING & 19,495,974 & 14,848,243 & 0.7616 & 82.0744 & 25.0484 & 99.9244 & 12 \\ \hline \end{tabular} } \end{table} \subsection{Evaluation Metrics} The first evaluation metric is the \textit{time-dependent concordance index} (\textbf{C-index}), which is the most common evaluation used in survival analysis \cite{Li2016mtlsa,luck2017deep,lee2018deephit} and reflects a measure of how well a model predicts the ordering of sample event times. That is, given the observing time $t$, two samples $d^1 = (\bs{x}^1, z^1)$ with large event time $z^1 > t$ and $d^2 = (\bs{x}^2, z^2)$ with small event time $z^2 \leq t$ should be ordered as $d^1 \prec d^2$ where $d^1$ is placed before $d^2$. This evaluation is quite similar to the area under ROC curve (AUC) metric in the binary classification tasks \cite{wang2017machine}. From the ranking view of event probability estimation at time $t$, C-index assesses the ordering performance among all the uncensored and censored pairs at $t$ among the test data. We also use \textit{average negative log probability} (\textbf{ANLP}) to evaluate the regression performance among different forecasting models. ANLP is to assess the likelihood of the co-occurrence of the test sample with the corresponding true event time, which is correspondent to the event time likelihood loss $L_z$ in Eq.~(\ref{eq:market-price-objective}). Here we compute ANLP as \begin{equation} \bar{P} = - \frac{1}{|\mathbb{D}_{\text{test}}|} \sum_{(\bs{x}^i, z^i) \in \mathbb{D}_{\text{test}}} \log p(z^i|\bs{x}^i) ~, \end{equation} where $p(z|\bs{x})$ is the learned time-to-event probability function of each model. Finally, we conduct the significance test to verify the statistical significance of the performance improvement of our model w.r.t. the baseline models. Specifically, we deploy a MannWhitney U test \cite{mason2002areas} under C-index metric, and a t-test \cite{bhattacharya2002median} under ANLP metric. \subsection{Compared Settings} We compare our model with two traditional statistic methods and five machine learning methods including two deep learning models. \begin{itemize}[leftmargin=2mm] \item \textbf{KM} is Kaplan-Meier estimator \cite{kaplan1958nonparametric} which is a statistic-based non-parametric method counting on the event probability at each time over the whole set of samples. \item \textbf{Lasso-Cox} is a semi-parametric method \cite{zhang2007adaptive} based on Cox proportional hazard model \cite{cox1992regression} with $l1$-regularization. \item \textbf{Gamma} is a parametric gamma distribution-based regression model \cite{zhu2017gamma}. The event time of each sample is modeled by a unique gamma distribution with respect to its features. \item \textbf{STM} is a survival random tree \cite{wang2016functional} model which splits the data into small segments using between-node heterogeneity and utilizes Kaplan-Meier estimator to derive the survival analysis for each segment. \item \textbf{MTLSA} is the recently proposed multi-task learning with survival analysis model \cite{Li2016mtlsa}. It transforms the original survival analysis problem into a series of binary classification subproblems, and uses a multi-task learning method to model the event probability at different time. \item \textbf{DeepSurv} is a Cox proportional hazard model with deep neural network \cite{katzman2018deepsurv} for feature extraction upon the sample covariates. \item \textbf{DeepHit} is a deep neural network model \cite{lee2018deephit} which predicts the probability $p(z)$ of event over the whole time space with the input $\bs{x}$. This method achieved state-of-the-art performance in survival analysis problem. \item \textbf{DRSA} is our proposed model which has been described above. The implementation details can be referred to supplemental materials and the published code. \end{itemize} \subsection{Results and Analysis} We present the evaluation results according to the category which the compared models belong to. KM and Lasso-Cox model are statistic-based methods, while Gamma, STM and MTLSA are machine learning based models. The rest models are deep neural network models with end-to-end learning paradigm. \begin{table}[t] \centering \caption{Performance comparison on C-index (the \textit{higher}, the better) and ANLP (the \textit{lower}, the better). (* indicates p-value $<10^{-6}$ in significance test)}\label{tab:results} \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|c|c|c||c|c|c|} \hline \multirow{2}{*}{\textbf{Models}} & \multicolumn{3}{c||}{\textbf{C-index}} & \multicolumn{3}{c|}{\textbf{ANLP}}\cr & CLINIC & MUSIC & BIDDING & CLINIC & MUSIC & BIDDING \cr \hline KM & 0.710 & 0.877 & 0.700 & 9.012 & 7.270 & 15.366 \cr Lasso-Cox & 0.752 & 0.868 & 0.834 & 5.307 & 28.983 & 38.620 \cr \hline\hline Gamma & 0.515 & 0.772 & 0.703 & 4.610 & 6.326 & 6.310 \cr STM & 0.520 & 0.875 & 0.807 & 3.780 & 5.707 & 5.148 \cr MTLSA & 0.643 & 0.509 & 0.513 & 17.759 & 25.121 & 9.668 \cr \hline\hline DeepSurv & 0.753 & 0.862 & 0.840 & 5.345 & 29.002 & 39.096 \cr DeepHit & 0.733 & 0.878 & 0.858 & 5.027 & 5.523 & 5.544 \cr DRSA & \bf 0.774\textsuperscript{*} & \bf 0.892\textsuperscript{*} & \bf 0.911\textsuperscript{*} & \bf 3.337\textsuperscript{*} & \bf 5.132\textsuperscript{*} & \bf 4.774\textsuperscript{*} \cr \hline \end{tabular} } \end{table} \subsubsection{Estimation for Event Rate over Time} The left part of Table~\ref{tab:results} has illustrated the performance of the event rate estimation, i.e., C-index metric. From the table, we may observe the following facts. (i) Deep learning models illustrated relatively better C-index performance, which may be caused by the higher model capacity for feature extraction. (ii) Not only within the deep learning models, but also over all the compared methods, our DRSA model achieved the best C-index scores with significant improvements on all the datasets, which proves the effectiveness of our model. (iii) The traditional statistical models, i.e., KM and Lasso-Cox provided stable performance over the three datasets. (iv) The models with pre-assumptions about the event time distribution, i.e., Gamma and DeepSurv, did not perform well because the strong assumptions lack generalization in real-world tasks. (v) Within the deep models, no assumption about the latent distribution of time-to-event data makes DeepHit and DRSA flexibly model the data and perform better. \subsubsection{Event Time Prediction} ANLP is a metric to measure the regression performance on the true event time prediction, i.e., the forecasting of the likelihood of the true event time. From the right part of Table~\ref{tab:results}, it also shows the similar findings to the C-index results discussed above, e.g., our DRSA model has the best performance among all the methods. Moreover, STM segmented the data well so it achieved relatively better performance than other normal machine learning methods. With effective sequential pattern mining over time, our DRSA model performed relatively better than other deep models. Note that DeepHit directly predicts the probability of time-to-event while our modeling method is based on hazard rate prediction and optimize through probability chain rule. The results reflect the advantage of the sequential pattern mining with the novel modeling perspective of our model. \begin{figure}[h] \centering \includegraphics[width=0.8\columnwidth]{figures/learning_curve_of_new_dataset.pdf} \caption{Learning curves. Here ``epoch'' means one iteration over the whole training data and $\alpha=0.25$ in Eq.~(\ref{eq:total-loss}). Learning curves over BIDDING dataset can be referred in our supplemental material.}\label{fig:learn-curve} \end{figure} \subsubsection{Model Convergence} To illustrate the model training and convergence of DRSA model, we plot the learning curves and the C-index results on CLINIC and MUSIC datasets in Figure~\ref{fig:learn-curve}. Recall that our model optimizes over two loss functions, i.e., the ANLP loss $L_z$ and the cross entropy loss $L_{\text{c}}$. From the figure, we may find that DRSA converges quickly and the values of both loss function drop to stable convergence at about the first complete iteration over the whole training dataset. Moreover, the two losses are alternatively optimizing and facilitate each other during the training, which proves the learning stability of our model. \subsubsection{Model Forecasting Visualization} \begin{figure}[t] \centering \includegraphics[width=1.0\columnwidth]{figures/survival_curve_24.pdf} \caption{A comprehensive visualization of survival rate $S(t|\bs{x}^i)$ estimation and event time probability $p(z|\bs{x}^i)$ prediction over different models. The vertical dotted line is the true event time $z^i$ of this sample.}\label{fig:visualization} \end{figure} Figure~\ref{fig:visualization} illustrates the estimated survival rate curve $S(t|\bs{x}^i)$ over time and the forecasted event time probability $p(z|\bs{x}^i)$ for an arbitrarily selected test sample $(\bs{x}^i, z^i, t^i)$. Note that the KM model makes the same prediction for all the samples in the dataset, which is not personalized well. Our DRSA model accurately placed the highest probability on the true event time $z^i$, which explains the result of ANLP metric where DRSA achieved the best ANLP scores. Since DeepHit directly predicted the probability of the event time $p(z^i)$ without any considerations of the previous conditional. And it has no supervision onto the predictions in the time range $(z^i, t^i)$ which makes the gradient signal too sparse only onto the true event time $z^i$. As a result, DeepHit did not place the probability well over the whole time space. \subsubsection{Ablation Study on the Losses} \begin{table}[t] \centering \caption{Ablation study on the losses.}\label{tab:ab-results} \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|c|c|c||c|c|c|} \hline \multirow{2}{*}{\textbf{Models}} & \multicolumn{3}{c||}{\textbf{C-index}} & \multicolumn{3}{c|}{\textbf{ANLP}}\cr & CLINIC & MUSIC & BIDDING & CLINIC & MUSIC & BIDDING \cr \hline DeepHit & 0.733 & 0.878 & 0.858 & 5.027 & 5.523 & 5.544 \cr DRSA$_{\text{unc}}$ & 0.765 & 0.881 & 0.823 & 3.441 & 5.412 & 12.255 \cr DRSA$_{\text{cen}}$ & 0.760 & 0.882 & 0.900 & \textbf{3.136} & 5.459 & 4.990 \cr DRSA & \bf 0.774 & \bf 0.892 & \bf 0.911 & 3.337 & \bf 5.132 & \bf 4.774 \cr \hline \end{tabular} } \end{table} In this ablation study, we compare the model performance on the three losses. DRSA$_{\text{unc}}$ optimizes under $(L_z + L_{\text{uncensored}})$ over only the uncensored data, and DRSA$_{\text{cen}}$ optimizes under $(L_z + L_{\text{censored}})$ without the loss $L_{\text{uncensored}}$. Note that our full model DRSA optimizes under all the three losses $(L_z + L_{\text{uncensored}} + L_{\text{censored}})$ as stated in Eq.~(\ref{eq:total-loss}). From Table~\ref{tab:ab-results}, we may find that both two partial likelihood losses $L_{\text{uncensored}}$ and $L_{\text{censored}}$ contribute to the final prediction. Moreover, our DRSA over all the three losses achieved the best performance, which reflects the effectiveness of our classification loss $L_c = L_{\text{uncensored}} + L_{\text{censored}}$ as that in Eq.~(\ref{eq:classification-loss}), which optimizes the C-index metric directly. \section{Conclusion}\label{sec:conclusion} In this paper, we comprehensively surveyed the survival analysis works from the modeling view and discussed the pros and cons of them. To make flexibly modeling over time, we proposed a deep recurrent neural network with novel modeling view for conditional hazard rate prediction. And probability chain rule connects the predicted hazard rate at each time, for the event time probability forecasting and survival rate estimation. The experiments on three large-scale datasets in three real-world tasks from different fields illustrated the significant advantages of our model against the strong baselines including state-of-the-art model. For the future work, it is natural to apply our model for competing risks prediction \cite{alaa2017deep,lee2018deephit} with shared feature embedding at the base architecture and multi-task learning for loss function. \section{Acknowledgments} The corresponding authors Weinan Zhang and Yong Yu thank the support of National Natural Science Foundation of China (61632017, 61702327, 61772333), Shanghai Sailing Program (17YF1428200).

d3ddabb94c582ff8354afbe40ae2c03e79bbb2cf

\section{Radar Interference Statistics} \label{sec:analysis} As highlighted in Sec.~\ref{sec:sysModel}, a statistical description of the aggregate interference is required to properly set the radar detection threshold. Considering omni-directional antennas, classical stochastic geometry results characterise $\mathcal I$ as L\`{e}vy distributed for $\alpha=4$, while no closed-form solution is known for other path-loss exponents. In order to derive compact and insightful expressions for a broader range of parameters, we follow a different approach, tuning $\Theta$ based on the power statistics of the strongest interferer only. For the sake of tractability, we furthermore initially focus on an ideal cone antenna pattern, assuming the gain to be $\mathcal G_m$ over a beamwidth $\varphi$ and zero elsewhere,\footnote{For any beamwidth, the gain is set as $\mathcal G_m = 4\pi/\varphi^2$. The impact of more realistic antenna patterns as well as the tightness of the strongest interferer approximation will be discussed in detail in Sec.~\ref{sec:results}.} to obtain: \begin{prop} Let \ensuremath{{\mathcal{I}_s}}\ be the power of the strongest interferer at the typical node over a detection slot. Then, under the cone antenna pattern assumption, we have \begin{equation} \ensuremath{F_{\mathcal{I}_s}}(i) := \ensuremath{\mathbb P}\{\ensuremath{{\mathcal{I}_s}} \leq i \} = \exp\left( -\frac{\lambda \,\delta \,\varphi^2\, \Omega \,\omega^{2/\alpha}}{4\pi} \cdot i^{-2/\alpha}\right) \label{eq:cdfInt} \end{equation} \label{prop:cdf} \end{prop} where $\omega := P_t \,\mathcal G_m^2 \ell$, $\Omega=1$ for the no-fading case and $\Omega = \Gamma(1+2/\alpha)$ in the presence of Rayleigh fading, with $\Gamma(x)= \int_{0}^{\infty}x^{t-1} e^{-x}\,dt$. \\ \begin{proof} For the considered model, a transmitter active over the observed slot interferes at the typical receiver only if the alignment of their two randomly oriented antenna patterns overlaps. By simple geometrical arguments, this event has probability $\varphi^2/(4\pi^2)$. By virtue of this further thinning, the aggregate interference is driven by the homogeneous PPP $\Phi_{\mathcal I}'$ of intensity $\delta \lambda \varphi^2/(4 \pi^2)$. Following a reasoning similar to \cite{Haenggi14_TIT}, let us introduce a new PPP \mbox{$\Xi := \{ \xi_i = (\,\left\Vert \mathbf x_i \right\Vert^{-\alpha} \zeta_{\mathbf x_i} )^{-1}, \mathbf x_i \in \Phi_{\mathcal I}' \}$} over $\mathbb R^+$, whose elements are ordered s.t. $\xi_i < \xi_j$, $\forall \,i<j$. From the mapping theorem, $\Xi$ is a non-homogeneous process of intensity measure $\Lambda([0,r]) = \delta \lambda \varphi^2 r^{2/\alpha}\, \Omega /(4 \pi)$, \mbox{$\forall \, r >0$}, where \mbox{$\Omega := \mathbb E[\zeta_{\mathbf x}^{2/\alpha}]$} evaluates to $1$ with no-fading (i.e. $\zeta_{\mathbf x} \equiv 1$) and to \mbox{$\int_{0}^{\infty} t^{2\alpha} e^{-t} dt = \Gamma(1+2/\alpha)$} with Rayleigh fading (i.e. $\zeta_{\mathbf x}\sim \exp(1)$). Let us indicate as $\mathbf x_{s} \in \Phi_{\mathcal I}'$ the coordinates of the strongest interferer at the typical receiver over the observed slot. Recalling the definition of $\omega$, we get \mbox{$\ensuremath{F_{\mathcal{I}_s}}(i) = \ensuremath{\mathbb P}\{ (\left \Vert \mathbf x_s \right\Vert^{-\alpha} \zeta_{\mathbf x_s} )^{-1} \geq \omega / i \}$}. Noting that $(\left \Vert \mathbf x_s \right\Vert^{-\alpha} \zeta_{\mathbf x_s} )^{-1} = \xi_1$, $\ensuremath{F_{\mathcal{I}_s}}(i)$ can be computed as the probability of not having any point of the process $\Xi$ in the interval $[0,\omega/i)$, i.e. $\ensuremath{F_{\mathcal{I}_s}}(i) = \exp(-\Lambda([0,\omega/i)))$, giving \eqref{eq:cdfInt}. \end{proof} Approximating the aggregate interference experienced by a radar receiver with its strongest component, Prop.~\ref{prop:cdf} allows us to derive a compact formulation of the detection threshold needed to achieve a target false alarm probability. Namely, solving \mbox{$\ensuremath{\mathsf{P}_{\mathsf{fa}}} = 1 - \ensuremath{F_{\mathcal{I}_s}}(\Theta)^{M-1}$} we get \begin{equation} \Theta = \omega \, \left( \frac{-\Omega \,(1-\delta) \,\lambda \,\varphi^2}{4\pi \ln(1-\ensuremath{\mathsf{P}_{\mathsf{fa}}})}\right)^{\alpha/2}. \label{eq:threshold} \end{equation} The simple closed-form expression in \eqref{eq:threshold} provides guidance on how to tune the radar receiver for any path-loss exponent \eq{\alpha>2} as well as for both the Rayleigh-fading and the no-fading case, pinpointing the impact of all relevant system parameters. Remarkably, the presence of fading on the interfering signals is embedded through the sole scaling factor $\Omega$. \section{Impact of Noise on Detection} \label{app:noise} In order to gauge the effect of noise on detection performance, let us focus for simplicity on the no-fading case, and assume that a radar receiver is subject to a mixture of interference and random noise. Following a typical approach \cite{Nartasilpa16}, the latter can be modelled at the symbol level as a complex normal random variable with zero mean and variance $P_n = k_B T B F$, where $k_B$ is the Boltzmann constant, $T$ is the receiver operating temperature, $F$ is the receiver noise figure and $B$ is the bandwidth. In this case, detection is affected over a slot of interest by a non target-related power \mbox{$\mathcal Z := W+\mathcal I$}, where $W$ is an exponential r.v. of parameter $1/P_n$, i.e. \mbox{$W\sim \rm{exp}(1/P_n)$}, and $\mathcal I$ is the interference level captured by the stochastic geometry model presented in Sec.~\ref{sec:analysis}. Observing that $W$ and $\mathcal I$ are statistically independent, and relying on the strongest interferer approximation for $\mathcal I$ (see Theorem~1), the cumulative distribution function of $\mathcal Z$ can be derived by means of convolution operations as: \begin{align} F_{\mathcal Z}(z) &:= \mathbb P\{ W + \mathcal I_s < z \} \nonumber\\ &= \frac{1}{P_n} \! \int_{0}^{z} \exp\left( -\frac{\lambda \delta \varphi^2 \omega^{\frac{2}{\alpha}}}{4\pi} (z - w)^{-\frac{2}{\alpha}} - \frac{w}{P_n} \right) dw\,. \label{eq:jointCDF} \end{align} The detection threshold $\Theta$ needed to guarantee a desired false alarm rate \ensuremath{\mathsf{P}_{\mathsf{fa}}}\ can then be derived numerically from \eqref{eq:jointCDF}, solving $1 - F_{\mathcal Z}(\Theta)^{M-1} = \ensuremath{\mathsf{P}_{\mathsf{fa}}}$. In turn, the maximum detectable range may be obtained by setting $\mathcal S = \Theta$ and solving the equation with respect to the target distance, leading to \begin{equation} d_{\mathsf m} = \left( \frac{\kappa \sigma P_t \mathcal G^2 \ell}{4\pi \Theta} \right)^{\frac{1}{2\alpha}} . \label{eq:dmNoiseAndInt} \end{equation} The closed-form expression in \eqref{eq:dmNoiseAndInt} clarifies that, in the presence of noise, transmission power $P_t$, operating frequency $f$ and bandwidth $B$ (embedded in $\Theta$) start to play a role in determining system performance. \begin{figure} \centering \includegraphics[width=.5\columnwidth]{dMaxVsLambda_noFading_randNoise.pdf} \caption{\ensuremath{d_{\mathsf{m}}}\ vs. radar density, no-fading case ($\alpha=2$, $f=60 \rm{GHz}$), with antenna beamwidth $\varphi=\pi/6$. The solid line shows results obtained assuming $\mathcal I \gg W$ and disregarding noise (Fig. $2$, p. $3$). Instead, the dashed blue line indicates the detectable range assuming a random noise component. The dash-dotted line shows performance with random noise, and in the absence of interference.} \vspace{-3mm} \label{fig:dmNoiseAndInt} \end{figure} Fig.~\ref{fig:dmNoiseAndInt} shows the trend of $d_{\mathsf m}$ against the network density $\lambda$, considering an antenna beamwidth $\varphi = \pi/6$, and setting, in the no-fading case under study ($\alpha=2$, $f = 60 \,\rm{GHz}$), $B = 125 \,\rm{MHz}$, $P_t = 20 \,\rm{dBm}$, $T=290 \,\rm{K}$, $F=10$. For convenience, the solid line replicates results obtained assuming an interference-limited system (i.e. disregarding noise) and already discussed in Fig.~\ref{fig:dmVsLambda}, whereas the dashed line is obtained assuming random noise as discussed above ($W\sim \exp(1/P_n)$). As expected, when the device density increases, the effect of noise becomes negligible, and the two curves coincide. Conversely, for very low densities -- which are likely to not be of practical interest for most applications -- the detection behaviour becomes noise-limited. In this case, $d_{\mathsf m}$ saturates to a value that can easily be computed analytically. Namely, in the absence of interference, the detection threshold can be determined by setting \mbox{$\ensuremath{\mathsf{P}_{\mathsf{fa}}} = 1 - \mathbb P\{ W < \Theta \}^{M-1}$} to obtain \begin{align} \Theta = -P_n \ln \left( 1 - (1-\ensuremath{\mathsf{P}_{\mathsf{fa}}})^{\frac{\delta}{1-\delta}} \right) \end{align} and, solving $\mathcal S = \Theta$ with respect to the target distance: \begin{align} d_{\mathsf m,\text{no-int}} = \left( -\frac{P_t \mathcal G^2 \kappa \sigma \ell}{4 \pi P_n \ln \left( 1 - (1-\ensuremath{\mathsf{P}_{\mathsf{fa}}})^{\frac{\delta}{1-\delta}} \right)} \right)^{\frac{1}{2\alpha}} \end{align} which is shown by the dash-dotted line in Fig.~\ref{fig:dmNoiseAndInt}. The results in Fig.~\ref{fig:dmNoiseAndInt} confirm the interference-limited nature of the networks under study for reasonable network densities, and underpin the ability of the presented approach to predict the fundamental trends and the effect of interference on radar detection performance. \section{Introduction} \label{sec:intro} Compact low-cost radar devices are set to become pervasive for providing short-range environmental awareness in emerging applications such as automotive~\cite{Kandeepan17_ITS}, enhanced localization~\cite{Guidi16_TMC}, or radio resource optimization~\cite{Simic16_RadCom}. In future heterogeneous networks, low-power radars are also envisioned to share spectrum -- and possibly be co-located -- with communication devices~\cite{Guidi16_TMC, Heath18_TVT}, in e.g. the $60$~GHz unlicensed band. These scenarios give rise to coexistence of a multitude of radar devices, randomly oriented over a large area, sharing a frequency band in an uncoordinated fashion. For such radar networks, it is paramount to understand the effect of mutual interference on the achievable detection and false alarm rates. Mutual radar interference has been thoroughly studied in simple two-node topologies~\cite{Brooker07_TEC}. Recent research has started to address more general configurations. The first results were reported in~\cite{Jondral13_EURASIP}, focusing on OFDM radars. The achievable detection probability for different network densities was investigated, relying on a Gaussian approximation of the aggregate interference. A step forward was taken in~\cite{Kandeepan17_ITS}, with the introduction of a stochastic geometry framework to study the performance of a linear automotive radar network. The authors considered an SIR-based detection model, and derived results under different statistical distributions for the radar positions, only for the no-fading case, with a path-loss exponent of $2$, and effectively omni-directional antennas. Random access for the radars was also proposed to mitigate interference. However, the fundamental performance limits in terms of detection range for a radar immersed in a planar field of interferers are not yet fully characterised. In this work, we propose a simple analytical approach to study a network of independent pulsed-radar devices. Exploiting a strongest-interferer approximation, we derive compact closed-form expressions for the radar detection performance, both for the no-fading and Rayleigh-fading cases and for any path-loss exponent. Our model clearly and comprehensively captures the effect of key system parameters. We study the tradeoff between a desired detection/false-alarm performance and the radar network density and antenna directivity, yielding practical insight into system design rules for tuning the radar network. \section{Radar Performance Evaluation} \label{sec:results} The derived interference statistics allow us not only to properly tune the threshold $\Theta$, but also to evaluate the impact of radar-to-radar interference on the detection performance. To better stress the key tradeoffs for the system under study, we focus on two use-cases of practical interest. Firstly, we consider a setting where links are line-of-sight ($\alpha\simeq 2$) and not affected by fading, representative of mm-wave propagation. We then complement our discussion delving into a scenario characterised by larger path-loss exponents and Rayleigh fading for both radar echo and interference, drawing conclusions that are applicable to traditional cellular and WiFi frequency bands. A radar cross section $\sigma=10 \,\rm{m}^2$ and a processing gain $\kappa=10$ are assumed for the target, while devices transmit with power $P_t=10\,\rm{dBm}$ and pulse repetition frequency $\delta=10^{-2}$. Unless otherwise specified, $\ensuremath{\mathsf{P}_{\mathsf{fa}}}=0.1$. \begin{figure} \centering \includegraphics[width=.5\columnwidth]{pdVsDistance_noFading.pdf} \caption{\ensuremath{\mathsf{P}_{\mathsf{d}}}\ vs. distance in the no-fading (\mbox{$\alpha=2$}, \mbox{$f=60 \,\rm{GHz}$} -- Sec.~\ref{sec:noFad}) and Rayleigh-fading case ($\alpha=3$, $\alpha=4$, $f=2.4 \,\rm{GHz}$ --Sec.~\ref{sec:fading}). Lines report analytical results (strongest interferer approximation), while markers simulation outcomes considering the aggregate interference. In all setups, $\varphi = \pi/6$, $\lambda=10^{-4}$.} \label{fig:pdVsDistance} \vspace{-1em} \end{figure} \vspace{-3mm} \subsection{No-Fading Case}\label{sec:noFad} In the absence of fading, the presented framework offers a closed-form expression of the target detection probability. Relying on the strongest interferer approximation, we readily get $\ensuremath{\mathsf{P}_{\mathsf{d}}} = 1 - \ensuremath{F_{\mathcal{I}_s}}(\Theta - S)$, where $\ensuremath{F_{\mathcal{I}_s}}(i)$ and $\Theta$ follow from \eqref{eq:cdfInt} and \eqref{eq:threshold} setting $\Omega=1$. The trends obtained for $\alpha = 2$,\footnote{Strictly speaking, analytical results are reported for $\alpha\rightarrow 2$.} and $f=60\,\rm{GHz}$ are presented by the solid line in Fig.~\ref{fig:pdVsDistance}, assuming a beamwidth $\varphi=\pi/6$ and a density $\lambda=10^{-4}$ [radar/m$^2$]. Given the deterministic nature of the incoming echo power, a target is detected with probability $1$ as long as it is close enough to satisfy the condition $\mathcal S > \Theta$. Conversely, when the sole target reflection is not sufficient to exceed the detection threshold, the plot highlights a sharp drop of \ensuremath{\mathsf{P}_{\mathsf{d}}}. This behaviour stems from the strong attenuation undergone by the radar echo, which follows a path-loss power law of exponent $2\alpha$. For $\mathcal S\rightarrow 0$, a detection occurs with probability \eq{\ensuremath{\mathbb P}\{\mathcal I>\Theta\}=1-(1-\ensuremath{\mathsf{P}_{\mathsf{fa}}})^{\delta/(1-\delta) } }, which is the asymptotic value the solid curve in Fig.~\ref{fig:pdVsDistance} converges to. In order to verify the accuracy of the strongest interferer approximation, dedicated simulations were performed. Specifically, multiple PPP instances were generated, so as to extract the statistics of the \emph{aggregate} interference at the typical node and thus compute the detection threshold matching the desired false alarm rate. The probability of target detection was then extracted accounting once more for the whole interference affecting the receiver. The outcome of the simulation study is shown by circle markers in Fig.~\ref{fig:pdVsDistance}. The results show a very tight match with the analytical prediction, confirming that the behaviour of the system is indeed driven by the disruptive effect of having an interferer close to the detecting radar, and further supporting the proposed framework as a simple tool to accurately tune the detection threshold and predict the achievable radar performance. Along this line of reasoning, the trend exhibited by \ensuremath{\mathsf{P}_{\mathsf{d}}}\ pinpoints the existence of a critical distance \ensuremath{d_{\mathsf{m}}}, separating accurate and missed detection. Setting \eq{\mathcal S =\Theta}, we readily get \begin{equation} \ensuremath{d_{\mathsf{m}}} = \left( \frac{\kappa \sigma}{4\pi}\right)^{\frac{1}{2\alpha}} \left(\frac{-4\pi \ln(1-\ensuremath{\mathsf{P}_{\mathsf{fa}}})}{(1-\delta)\lambda \,\varphi^2} \right)^{\frac{1}{4}}. \label{eq:dm} \end{equation} The expression offers interesting insights. Firstly, \eqref{eq:dm} clarifies that the maximum detectable range does not depend on the operating frequency and radar transmission power.\footnote{For very low $\lambda$, i.e. when interference no longer plays a role, detection performance eventually becomes limited by noise power, as discussed in App.~\ref{app:noise}. We also assume $\delta$ to be picked so that the detection range is limited by incoming power rather than by the unambiguous range.} Secondly, the limited impact on \ensuremath{d_{\mathsf{m}}}\ of the pulse repetition frequency $\delta$ emerges. This insight is non-trivial, as it settles a critical tradeoff. Lower values of $\delta$ potentially favour correct detection of farther targets, both due to the longer echo waiting time and to the weaker interference expected over a single reception slot -- driven by a thinned PPP of intensity $\delta\lambda [\varphi^2/(4\pi^2)]$. Conversely, reducing the pulse repetition frequency calls for a higher $\Theta$ (i.e. a stronger echo for detection) to grant a desired \ensuremath{\mathsf{P}_{\mathsf{fa}}}, since the greater number of slots spent listening favour the occurrence of a false alarm event. The two effects eventually balance each other out, contributing to \ensuremath{d_{\mathsf{m}}}\ through a scaling factor \eq{(1-\delta)^{-1/4}}, which is negligible for practical values of $\delta$. Finally, the result emphasises the minor role of the path-loss exponent $\alpha$ in determining \ensuremath{d_{\mathsf{m}}}, unless large radar cross sections come into play. \begin{figure} \centering \includegraphics[width=.5\columnwidth]{dMaxVsLambda_noFading.pdf} \caption{\ensuremath{d_{\mathsf{m}}}\ vs. radar density, no-fading case (\mbox{$\alpha=2$}, \mbox{$f=60 \,\rm{GHz}$}). For the cone antenna model, lines show analytical results, while circle markers their verification via simulations. Square markers show simulation results for a realistic antenna patterns, $\varphi=\pi/6$.} \label{fig:dmVsLambda} \vspace{-1em} \end{figure} The maximum detectable distance computed via \eqref{eq:dm} for different beamwidths is shown by solid and dashed lines against $\lambda$ in Fig.~\ref{fig:dmVsLambda}. The accuracy of the analytical approach is again confirmed for all the considered setups by the circle markers, which show the results of system simulations accounting for aggregate interference. The plot clearly highlights the paramount role played by the density of radars sharing the same channel, stressing that for large values of $\lambda$ the detection performance deteriorates sharply. The strong impact of antenna directionality is also apparent. With omni-directional nodes, interference generated by close-by transmitters is disruptive, and targets farther than a few meters can be detected reliably only in very sparse networks. Conversely, narrower beams beneficially filter out undesired signals, enabling detection for \ensuremath{d_{\mathsf{m}}}\ and $\lambda$ of practical interest already for $\varphi=\pi/6$. Two further points follow from the observed trends. Firstly, the considered idealised cone antenna model neglects potentially harmful energy transmitted and received from secondary lobes. To gauge the influence of this, we performed dedicated simulations employing realistic antenna patterns. Specifically, we used the \textsc{Matlab} \emph{Phased Array System Toolbox} to generate a pattern with half-power beamwidth of approximately $25^{\circ}$ and main sidelobes with $10\, \rm{dB}$ attenuation, corresponding to a $4\times 4$ uniform quadratic array of isotropic elements with half-wavelength spacing. The achievable \ensuremath{d_{\mathsf{m}}}\ is presented in Fig.~\ref{fig:dmVsLambda} with square markers, showing an extremely close match to the analytical results. \begin{figure} \centering \includegraphics[width=.5\columnwidth]{dMaxVsBW_noFading.pdf} \caption{\ensuremath{d_{\mathsf{m}}}\ vs. antenna beamwidth, no-fading case (\mbox{$\alpha=2$}, \mbox{$f=60 \,\rm{GHz}$}, \eq{\lambda=10^{-4}}). A cone antenna model is assumed. Lines show analytical results, while markers simulation results.} \label{fig:dmVsBeamwidth} \vspace{-1em} \end{figure} Secondly, the need to tune the antenna beamwidth to achieve a desired \ensuremath{d_{\mathsf{m}}}\ emerges from Fig.~\ref{fig:dmVsLambda} as a critical design choice. This aspect is explored in Fig.~\ref{fig:dmVsBeamwidth}, which depicts the maximum detectable target distance against $\varphi$ for a reference density $\lambda=10^{-4}$ [radar/m$^2$]. Notably, for narrow beamwidths, ranges in the order of $100\,\rm{m}$ are achievable even when immersed in a field of interferers. Finally, the plot also examines the impact of the targeted false alarm probability, reporting \ensuremath{d_{\mathsf{m}}}\ for three distinct values of \ensuremath{\mathsf{P}_{\mathsf{fa}}}. The study confirms that more stringent requirements in terms of erroneous detection come at the expense of a severe reduction in the detectable range, stressing even more the need for highly directional radar operations. \vspace{-3mm} \subsection{Rayleigh-Fading Case} \label{sec:fading} We extend our study to a setup in which signals undergo multiple reflections caused by the environment. As a worst case of the multipath-induced random fluctuations in the received power, we consider a higher path-loss exponent and Rayleigh fading on the interfering signals and the target echo.\footnote{We note that realistic multipath-induced signal fluctuations would result in performance bounded by the no-fading and Rayleigh-fading cases.} In this case, the detection probability can be computed conditioning on the interference level, i.e. \eq{\ensuremath{\mathsf{P}_{\mathsf{d}}}= \mathbb E_{\mathcal I}[\ensuremath{\mathbb P}\{\zeta \geq (\Theta -\mathcal I) \,4\pi d^{2\alpha} /(\omega \kappa \sigma) \,|\, \mathcal I\}]}, with \eq{\zeta\sim\exp(1)}. Under the strongest interferer approximation and for a cone antenna pattern we thus have \begin{equation} \ensuremath{\mathsf{P}_{\mathsf{d}}} = 1 - \ensuremath{F_{\mathcal{I}_s}}(\Theta) + \int_{0}^{\Theta} \!\!e^{\mathlarger{-\frac{(\Theta-i) 4\pi d^{2\alpha}}{\omega \kappa \sigma}}} \, \ensuremath{f_{\mathcal{I}_s}}(i) \,di \label{eq:pdFading} \end{equation} where $\ensuremath{f_{\mathcal{I}_s}}(i) := d\ensuremath{F_{\mathcal{I}_s}}/di$ follows from \eqref{eq:cdfInt}, whereas the term $1-\ensuremath{F_{\mathcal{I}_s}}(\Theta)$ accounts for the event of having an interference level above the threshold, which marks a target as present regardless of the actual echo power $\mathcal S$. The detection rate can then be computed by means of numerical integration for any configuration of the system parameters. Moreover, inspection of \eqref{eq:pdFading} reveals that the analytical expression derived for \ensuremath{\mathsf{P}_{\mathsf{d}}}\ is independent of the transmission power and carrier frequency also in the presence of fading. Namely, $\Theta$ and $\mathcal I_s$ are proportional to $\omega$, which embeds the effect of $P_t$ and $f$ and cancels out in the integrand function (\emph{c.f.} \eqref{eq:threshold}). The achievable detection probability for different target distances is depicted by dashed lines in Fig.~\ref{fig:pdVsDistance} for $\varphi = \pi/6$ and path-loss exponents $\alpha=3$ and $4$. Analytical trends are verified also in this case by means of simulations accounting for the level of aggregate interference both in the threshold tuning and the performance evaluation, the results of which are shown by square and triangle markers. As expected, the step transition between accurate and poor detection discussed for the no-fading case is replaced by a smooth degradation, caused by the statistical fluctuations of the echo power. On the other hand, Fig.~\ref{fig:pdVsDistance} confirms the small impact of varying the path-loss exponent when aiming for reliable detection also in the presence of random signal variations. Indeed, for sufficiently high values of \ensuremath{\mathsf{P}_{\mathsf{d}}}\ (e.g. $\ensuremath{\mathsf{P}_{\mathsf{d}}}\geq0.9$), the beneficially lower level of interference brought by harsher propagation conditions is counterbalanced by the weaker incoming echo. In turn, the latter effect prevails for larger target distances due to the path-loss power law of exponent $2\alpha$ for the useful reflection, leading to the sharper performance decay observed for $\alpha=4$. \begin{figure} \centering \includegraphics[width=.5\columnwidth]{PdVsPfa_Fading.pdf} \caption{Receiver Operating Characteristic (ROC) diagram, Rayleigh-fading case. Markers differentiate radar densities, while solid and dashed lines target distances. In all setups, $\varphi=\pi/6$, $\alpha=3$, \mbox{$f=2.4 \,\rm{GHz}$}.} \label{fig:pdVsPfa_fading} \vspace{-1em} \end{figure} Further insights on the system behaviour are offered by the ROC diagram in Fig.~\ref{fig:pdVsPfa_fading}, showing \ensuremath{\mathsf{P}_{\mathsf{d}}}\ vs. \ensuremath{\mathsf{P}_{\mathsf{fa}}}\ for different network densities. In sparse setups, e.g. $\lambda=10^{-5}$ [radar/m$^2$], a slight degradation of the detection performance can be effectively traded off for a lower \ensuremath{\mathsf{P}_{\mathsf{fa}}}, offering a useful design choice. By contrast, for larger $\lambda$, the plot emphasises that no reliable target identification is possible for the considered distances. This outcome highlights once more the strong role played by the transmitter density, strongly suggesting the need to design medium access strategies tailored for radar networks. \section{System Model and Preliminaries} \label{sec:sysModel} We consider a population of pulsed-radar devices, modelling their locations as a homogeneous Poisson point process (PPP) $\Phi = \{ \mathbf{x}_i\}$ of intensity $\lambda$ over $\mathbb R^2$. Time is divided in slots of equal duration, and devices continuously operate with a common pulse repetition frequency $\delta = 1/M$ [slot$^{-1}$]. Namely, each device follows cycles of duration $M$, transmitting a pulse over one slot, and then waiting for a target echo over the subsequent $M-1$ slots.\footnote{For a bandwidth $B$, a slot duration $1/B$ can be assumed; $B$ influences the radar resolution, investigating which is out of scope of this work.} While slot-level synchronisation eases system modelling, the uncoordinated nature of the radar population is captured by allowing random offsets among the operating cycles of different nodes. Accordingly, each PPP element is assigned an independent mark \mbox{$m_i \sim \mathcal U \{0,M-1\}$}, and transmits its pulses at slots $m_i + kM$, $k\in \mathbb N$. Within this setup, we are interested in characterising the impact of radar-to-radar interference on the detection performance. We focus without loss of generality on the \emph{typical} node located at the origin of the plane, and consider a power-based detection rule. Specifically, a target is declared present if the aggregate incoming power over any of the $M-1$ slots spent listening for the echo exceeds a threshold $\Theta$. All radars transmit with power $P_t$, and have a planar antenna pattern with maximum gain $\mathcal G_m$, where the boresight direction of each device is modelled as an independent and uniform r.v. in $[0,2\pi)$. Signal propagation undergoes path-loss with exponent $\alpha > 2$, and we investigate both the no-fading and Rayleigh-fading cases. For a target of radar cross section $\sigma$ in the boresight direction of the typical node, the incoming reflected power $\mathcal S$ follows from the well-known range equation as \begin{equation} \mathcal S = \frac{P_t \, \mathcal G_m^2 \, \kappa \,\sigma \,\ell}{4\pi} \cdot \zeta \,d^{-2\alpha} \,. \label{eq:echoPower} \end{equation} Here, $\ell := [\,c/(4\pi f)\,]^2$, $c$ is the speed of light, $f$ is the carrier frequency, $\kappa$ is a signal processing gain, $d$ is the target distance, and $\zeta \sim \exp(1)$ in the case of Rayleigh fading or is set to $1$ otherwise. By the properties of thinning for PPP, the aggregate interference $\mathcal I$ the typical radar experiences over each of the $M-1$ detection slots is i.i.d., taking the form \begin{equation} \mathcal I = \sum\nolimits_{\mathbf x \in \Phi_{\mathcal I}} P_t \,\mathcal G_t \mathcal G_r \, \ell \, \zeta_{\mathbf x} \, \Vert \mathbf x \Vert^{-\alpha}. \label{eq:aggIntDef} \end{equation} In \eqref{eq:aggIntDef}, $\Phi_{\mathcal I}$ is a thinned version of $\Phi$ with intensity \eq{\lambda' = \delta \lambda} capturing the active interferers over the observed slot. In turn, $\mathcal G_t$ and $\mathcal G_r$ are the transmit and receive antenna gains for the link between the radar transmitter located at $\mathbf x$ and the typical receiver (thus depending on the random boresight orientation of the radars). Finally, for the same link, $\zeta_{\mathbf x}$ is once more either an exponential r.v. of unit mean in the case of Rayleigh fading, or is set to $1$ when no fading is considered. For the introduced system model, a target is in general correctly detected with detection probability \ensuremath{\mathsf{P}_{\mathsf{d}}}, expressed as $\ensuremath{\mathbb P}\left\{ \mc{S} + \mc{I} + W \geq \Theta \right\}$, where $W$ is a random variable accounting for noise. Given the interference-limited nature of the networks under consideration, discussed in more depth in App.~\ref{app:noise}, we will however assume in the remainder $\mathcal I \gg W$, and disregard the noise component. Following this approach, a \emph{false alarm} is triggered when the power over at least one of the observed slots exceeds the detection threshold in the absence of the desired echo, i.e. with probability \mbox{$\ensuremath{\mathsf{P}_{\mathsf{fa}}} = 1 - F_{\mathcal I}(\Theta)^{M-1}$}, where $F_{\mathcal I}(i) = \ensuremath{\mathbb P}\left\{ \mc{I} \leq i \right\}$ is the cumulative distribution function of $\mathcal I$. Following common practice in radar design, we will set $\Theta$ so as to achieve a tolerable false alarm rate \ensuremath{\mathsf{P}_{\mathsf{fa}}}, and evaluate the performance of the system in terms of the corresponding detection probability.

83893c1ea8d141ec3046b5f2b470101c37f24372

\section{Introduction} In an article in the journal Science in 2013, Xie et al. (\cite{Xie}) stated ``the restorative function of sleep may be a consequence of the enhanced removal of potentially neurotoxic waste products that accumulate in the awake central nervous system'' indicating the value of sleep in changing the parameters of the brain's functioning. We can choose to consider the brain as a server in a queueing system which decreases its service rate over time but recovers after it has a rest (vacation). Vacation queueing systems have been studied by many authors with different models (\cite{Dos}, \cite{TianZhang}, \cite{GuoHas}, \cite{IsiO}, \cite{ZhHou}). Working vacations, introduced by Servi and Finn(2002)\cite{SerFin}, refer to a time period, during which the service slows but does not stop. Servers would gradually get exhausted during continuous work, but the service rate could increase after a vacation of the server. We include two types of systems in this paper. The first kind of system is the regular $M/M/1$ system, in which the server works without vacations, and the service rate is a constant with a relatively low value \cite{GroHar}. The other kind of system also has exponential interarrival and service times. However, the service rate changes after each state transition. When the service rate decreases to a certain value, the server stops working and has a vacation, after which the service rate would return to the highest level. To compare the performances of different queueing systems, two commonly used measures are the expected waiting time of a customer, $E(W)$), and the expected number of customers $E(L)$ in the system. These are related via Little's formula \cite{GroHar}. In this paper, we only use $E(L)$ to measure performance of different queueing systems. Values of $E(L)$ are obtained using matrix-analytic methods (\cite{LatG}, \cite{HeQ}). We show that the system with vacations performs better than the regular $M/M/1$ system under certain conditions. \section{Quasi Birth and Death Processes with 4 Phases} In this section, we compare a queueing system with working vacations with a regular $M/M/1$ system having a constant service rate. Consider the decrease of service rate over time as working vacations (\cite{SerFin}), during which the server works with lower efficiency. The number of customers in the system and states of the server form a continuous time Markov process $\{(X(t),Y(t)),t\ge 0\}$, where $X$ is the level variable (number of customers) and $Y$ is the phase variable (server efficiency with low values indicating a high efficiency). Each state $(n,i)$ with $X(t)=n>0$ and $Y(t)=i<4$ moves to $(n-1,i+1)$ with rate $\mu_i$, or to $(n+1,i+1)$ with rate $\lambda$. For all $n\in \mathbb{N}$, state $(n,4)$ will always go to $(n+2,1)$ with rate $\lambda/2$. State $(0,i)$ will always go to $(1,i+1)$ with rate $\lambda$, $i=1,2,3$. Set $\mu_1=\mu$, $\mu_2=a\mu$ and $\mu_3=b\mu$ ($0<b<a<1$). The process can be shown by the following network. The system takes a working vacations when $Y=2$ or $3$, having a regular vacation with interval $\sim Exp(\lambda/2)$ when $Y=4$. \[ \xymatrix{ (0,1)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3) {\mu}(1,1)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3) {\mu}(2,1)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){\mu}(3,1) & \dots \\ (0,2)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){a\mu}(1,2)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){a\mu}(2,2)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){a\mu}(3,2) & \dots \\ (0,3)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){b\mu}(1,3)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){b\mu}(2,3)\ar[rd]^(0.3)\lambda & \ar[ld]_(0.3){b\mu}(3,3) & \dots \\ (0,4) \ar"1,3"|{\frac{\lambda}{2}} & (1,4) \ar"1,4"|{\frac{\lambda}{2}} & (2,4)\ar"1,5"|{\frac{\lambda}{2}} & (3,4) & \dots } \] The motivation for the model is that each arrival or service completion takes time and reduces the server's efficiency so Y (phase) increases at each step ($i=1,2,3$). For $i=4$, the server is exhausted and even though there may be customers to be served, the server takes a vacation long enough for 2 more customers to arrive, and then begins service with renewed vigor and first level of efficiency. Setting up the model in this way keeps the transition between states exponential at all times. The interarrival rate for customers is $\lambda$ so the expected time between until the next customer is $1/\lambda$. The expected time for two customers to arrive is $2/\lambda$ so we take the arrival rate to be $\lambda/2$ to move from state $(n,4)$ to state $(n+2,1)$ (vacation time). Another approach could have used the sum of two exponentials (each with rate $\lambda$) but we can keep our model simpler by using rate $\lambda/2$ to have 2 customers arrive. The two approaches are not identical, though the mean times are the same, but we keep our state space more tractable using our approach. \begin{Theorem} The system is stable if $\lambda< \dfrac{\frac{\mu}{\lambda+\mu}+\frac{a\mu}{\lambda+a\mu}+\frac{b\mu}{\lambda+b\mu}+0\cdot\frac{2}{\lambda}} {\frac{1}{\lambda+\mu}+\frac{1}{\lambda+a\mu}+\frac{1}{\lambda+b\mu}+\frac{2}{\lambda}}$ \end{Theorem} \begin{proof} It is sufficient to consider the situation when the level is large as that determines the stability condition. For states with phase variable $Y=i$ ($i=1,2,3,4$), and level $X$ large, let $v_i$ be the state transition rate, and let $w_i$ be proportion of sojourn time in those states. \begin{equation} w_i=\frac{\frac{1}{v_i}}{\sum_{i=1}^4\frac{1}{v_i}} \end{equation} where $v_1=\lambda+\mu$, $v_2=\lambda+a\mu$, $v_3=\lambda+b\mu$, $v_4=\lambda/2$. The average service rate of the system (for large level $X$) should be calculated as a weighted average. \begin{align} \bar \mu &= \sum_{i=1}^4w_i\mu_i \notag\\ &= \frac{\frac{\mu}{\lambda+\mu}+\frac{a\mu}{\lambda+a\mu}+\frac{b\mu}{\lambda+b\mu}+0\cdot\frac{2}{\lambda}} {\frac{1}{\lambda+\mu}+\frac{1}{\lambda+a\mu}+\frac{1}{\lambda+b\mu}+\frac{2}{\lambda}} \end{align} The system is stable if $\lambda<\bar{\mu}$. The result follows. \end{proof} We note in the previous proof that $\bar{\mu}$ is a function of $\lambda$. To emphasize this, we define \begin{equation*} g(\lambda) \stackrel{\bigtriangleup}{=} \frac{\frac{\mu}{\lambda+\mu}+\frac{a\mu}{\lambda+a\mu}+\frac{b\mu}{\lambda+b\mu}+0\cdot\frac{2}{\lambda}} {\frac{1}{\lambda+\mu}+\frac{1}{\lambda+a\mu}+\frac{1}{\lambda+b\mu}+\frac{2}{\lambda}}. \end{equation*} Unfortunately, for our 4 phase model, it turns out that regardless of $\lambda$, $\mu$, $a$, $b$, the expected number of customers will be shorter under a regular $M/M/1$ model with service rate $b\mu$ than under our model that allows for a vacation, at the cost of two customers arriving. We prove this as follows. \begin{Theorem} For the 4 phase model which is stable (i.e. $g(\lambda)>\lambda$), $g(\lambda)$ is always smaller than $b\mu$. \end{Theorem} \begin{proof} First note that in our 4 phase model, states (0,1), (0,2) and (1,1) are not recurrent. Further, the average service rate that appears for large level $X$ is an upper bound on the rate for small levels (like 1). So we will work with the service rate for large levels. Now \begin{equation*} g(\lambda)-b\mu=-\mu\cdot\frac{\lambda^3(4b-a-1)+2\lambda^2\mu(ab+b-a+2b^2)+3\lambda\mu^2b^2(a+1)+2ab^2\mu^3}{3\lambda\mu^2(a+b+ab)+4\lambda^2\mu(a+b+1)+5\lambda^3+2ab\mu^3} \end{equation*} The denominator is always positive so we define \begin{equation*} f(\lambda) \stackrel{\bigtriangleup}{=}\lambda^3(4b-a-1)+2\lambda^2\mu(ab+b-a+2b^2)+3\lambda\mu^2b^2(a+1)+2ab^2\mu^3, \end{equation*} Note that $g(\lambda)-b\mu<0 \Leftrightarrow f(\lambda)>0$.\\ We will view the situation graphically by considering $f(\lambda)$ which is usually a cubic in $\lambda$. \\ Case 1: $4b-a-1=0$. Then $f(\lambda)$ becomes a quadratic. Also $a=4b-1$. The coefficient of $\lambda^2$ in $f(\lambda)$ is $2\mu(ab+b-a+2b^2)=2\mu(a(b+1)+b+2b^2)$, which is $>0$, os the quadratic is convex. The two real roots of $f(\lambda)=0$ are $-b\mu$ and $-\frac{b\mu(4b-1)}{6b^2-4b+1}$, which are both negative. So the value of $f(\lambda)$ is positive for value of $\lambda$ which is greater than the largest root so $f(\lambda)>0$ for $\lambda>0$ , as desired. \\ Case 2: When $4b-a-1 < 0$, $f(\lambda)$ is a cubic with a negative coefficient for the $\lambda^3$ term. Let $A=\sqrt{9a^2b^2-4a^2b-14ab^2+4a^2-4ab+9b^2}$. The 3 roots of $f(\lambda)=0$ are $-b\mu$,$-\frac{\mu(3ab-2a+3b+A)}{2(4b-a-1)}$ and $-\frac{\mu(3ab-2a+3b-A)}{2(4b-a-1)}$. Two of the three roots of $f(\lambda)$ are negative with the largest root $-\frac{\mu(3ab-2a+3b+A)}{2(4b-a-1)}$. So the cubic $f(\lambda)$ will be positive between the second largest root and the largest root, after which it becomes negative. But for $\lambda$ greater than the largest root, we have $g(\lambda)>\lambda$ so we are outside the stable region of the system. So our result is still true.\\ Case 3: When $4b-a-1 > 0$, $f(\lambda)$ is a cubic with a positive coefficient for the $\lambda^3$ term. Again, we get 3 roots of $f(\lambda)=0$. The largest of the three roots is $-\frac{\mu(3ab-2a+3b-A)}{2(4b-a-1)}$. However, the largest root would be a negative number under the following analysis. \begin{align} 4b-a-1>0 &\Rightarrow b<a<4b-1 \notag \\ &\Rightarrow b<4b-1 \notag \\ &\Rightarrow b\in(\frac{1}{3},1) \notag \\ 0<a<1 &\Rightarrow \frac{a}{a+1}\in(0,\frac{1}{2}) \notag \\ &\Rightarrow \frac{2a}{3a+3}\in(0,\frac{1}{3}) \notag \\ &\Rightarrow b>\frac{a}{a+1} \notag \\ &\Rightarrow 3ab-2a+3b>0 \notag \end{align} Thus, there would be \begin{align} &3ab-2a+3b-\sqrt{9a^2b^2-4a^2b-14ab^2+4a^2-4ab+9b^2}<0 \notag \\ &\Leftrightarrow (3ab-2a+3b)^2<9a^2b^2-4a^2b-14ab^2+4a^2-4ab+9b^2 \notag \\ &\Leftrightarrow ab(4b-a-1)<0 \notag \end{align} Since $f(\lambda)$ is a cubic with a positive coefficient for $\lambda^3$, then $f(\lambda)$ must be positive for all $\lambda$ larger than the largest root of $f(\lambda)=0$ so $f(\lambda)>0$ for all $\lambda>0$. The result follows. \end{proof} Hence, when the service rate of a regular $M/M/1$ system equals the lowest service rate in the 4 phase system, the 4 phase system will always have a lower overall average service rate than the $M/M/1$ system. This means that the M/M/1 system will have a lower expected number of customers than the 4 phase system and there is no advantage in using the 4 phase system. As a result, we move to consider a 5 phase system. \section{Quasi-Birth-and-Death Process with 5 States of Service} Add one more phase standing for $c\mu$ ($0<c<b<a<1$) as service rate to the former system, and change the constant service rate in $M/M/1$ to $c\mu$. The proportion of time when the server stops working in the new system would decrease. With fixed $a$, $b$, $c$ and $\mu$, there would be a range of $\lambda$ such that the average service rate in the new system is higher than that in $M/M/1$, and the expected number of customers would be reduced when servers take some time to rest. The statement could be proved more succinctly by numerical methods rather than analytical ones. \subsection{Matrix-Analytic Methods for Calculating the Expected Number of Customers} For the 5-phase system, let the states be\\ $(0,1),(0,2),(0,3),(0,4),(0,5), (1,1),(1,2),(1,3),(1,4),(1,5), (2,1),(2,2),\dots$. The Q-matrix (infinitesimal matrix) of the system with 5 states of service is \[ Q1= \begin{pmatrix} A_{00} & A_{01} & A_{02} & & & & \\ A_{10} & A_{11} & A_{01} & A_{02} & & & \\ & A_{10} & A_{11} & A_{01} & A_{02} & & \\ & & A_{10} & A_{11} & A_{01} & A_{02} & \\ & & & \ddots & \ddots & \ddots & \ddots \end{pmatrix} \] where \begin{align*} A_{00} =\begin{pmatrix} -\lambda & & & & \\ & -\lambda & & & \\ & & -\lambda & & \\ & & & -\lambda & \\ & & & & -\lambda/2 \end{pmatrix}_{5\times 5} , A_{01} =\begin{pmatrix} 0& \lambda & & & \\ & 0& \lambda & & \\ & & 0 & \lambda & \\ & & & 0 & \lambda\\ & & & &0 \end{pmatrix}_{5\times 5} \\ A_{02} =\begin{pmatrix} 0 & 0 & \dots & 0 \\ \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \dots & 0 \\ \lambda/2 & 0 &\dots & 0 \end{pmatrix}_{5\times 5} , A_{10} =\begin{pmatrix} 0 & \mu & & & \\ & 0 & a\mu & & \\ & & 0 & b\mu & \\ & & & 0 & c\mu \\ & & & &0 \end{pmatrix}_{5\times 5} \\ A_{11}=\begin{pmatrix} -(\lambda+\mu) & & & & \\ & -(\lambda+a\mu) & & & \\ & & -(\lambda+b\mu) & & \\ & & & -(\lambda+c\mu) & \\ & & & & -\lambda/2 \end{pmatrix}_{5\times 5} \end{align*} Since states $(0,1)$, $(0,2)$ and $(1,1)$ are not positive recurrent, we delete the corresponding rows and columns from Q1. Let \[ A_0= \begin{pmatrix} A_{02} & 0 \\A_{01} & A_{02} \end{pmatrix}, A_1= \begin{pmatrix} A_{11} & A_{01} \\A_{10} & A_{11} \end{pmatrix}, A_2= \begin{pmatrix} 0 & A_{10} \\ 0 & 0 \end{pmatrix}\] After that, the Q matrix could be written as \begin{equation} Q= \begin{pmatrix} B_{11} & B_{12} & & & & \\ B_{21} & A_1 & A_0 & & & \\ & A_2 & A_1 & A_0 & & \\ & & A_2 & A_1 & A_0 & \\ & & &\ddots&\ddots&\ddots \end{pmatrix} \label{Qmatrix} \end{equation} where \[ B_{11}=\begin{pmatrix} -\lambda &0 &0 &0 &0 &-\lambda &0\\ 0 &-\lambda &0 &0 &0 &0 &-\lambda \\ 0 &0 & -\lambda/2 &0 &0 &0 &0 \\ a\mu &0 &0 &-(\lambda+a\mu) &0 &0 &0 \\ 0 &b\mu &0 &0 &-(\lambda+b\mu) &0 &0 \\ 0 &0 &c\mu &0 &0 &-(\lambda+c\mu) &0 \\ 0 &0 &0 &0 &0 &0 &-\lambda/2 \\ \end{pmatrix}_{7 \times 7} \]\[ B_{12}=\begin{pmatrix} 0 & & & & & & & &\\ 0 &0 & & & & & & &\\ \lambda/2 &0 &0 & & & & & &\\ 0&0 &\lambda &0 & & & &\\ 0&0 &0 &\lambda &0 & & &\\ 0&0 &0 &0 &\lambda &0 & &\\ 0&0 &0 &0 &0 &-\lambda/2 &0 & \end{pmatrix}_{7\times 10} ,\quad B_{21}=\begin{pmatrix} & &\mu &0 &0 &0 \\ & & &a\mu &0 &0 \\ & & & &b\mu & 0\\ & & & & &c\mu \\ & & & & & \\ & & & & & \end{pmatrix}_{10 \times 7} \] Note that $Q$ has the form of a quasi birth and death process while $Q1$ did not. \\ Let $\vec{\pi}_0 = (\pi_{(0,3)}, \pi_{(0,4)}, \pi_{(0,5)}, \pi_{(1,2)}, \pi_{(1,3)}, \pi_{(1,4)}, \pi_{(1,5)})$. For $j\ge1$, let $\vec{\pi}_j = (\pi_{(j+1,1)}, \dots, \pi_{(j+1,5)}, \pi_{(j+2,1)}, \dots, \pi_{(j+2,5)})$. Let $\vec{\pi}=(\vec{\pi}_0, \vec{\pi}_1,\dots)$. From $\vec{\pi} Q = \vec{0}$, we have: \begin{align} \vec{\pi}_0B_{11} +\vec{\pi_1}B_{21} = 0 \label{pi1} \\ \vec{\pi}_0B_{12} +\vec{\pi}_1(A_1+RA_2) = 0 \label{pi2} \end{align} Also, \begin{align} \vec{\pi}_j = \vec{\pi}_1R^{j-1},\quad \forall j \ge 1\label{piR} \\ R^2A_2 + RA_1 + A_0 = 0 \end{align} where the $R$ matrix ($10 \times 10$) can be found using iteration. \begin{align} R(0) &= [0], \notag \\ R(n+1) &= -\sum\limits_{k=0,k\ne 1}^{\infty}R^k(n)A_kA_1^{-1}, n \ge 0 \\ &= -(A_0A_1^{-1}+R^2(n)A_2A_1^{-1}). \notag \end{align} Let $\vec{e}$ be a column vector if 1's of various lengths, as appropriate. Using the expression in equation (\ref{piR}), $\vec{\pi} \vec{e} =1$ implies \begin{equation}\vec{\pi}_0 e + \vec{\pi}_1 (I-R)^{-1}e = 1. \label{pi3}\end{equation} Using (\ref{pi1}), (\ref{pi2}) and (\ref{pi3}), $\vec{\pi}_0$ and $\vec{\pi}_1$ can be obtained. From these, limiting probabilities for all states are obtained using (\ref{piR}). Next \begin{align} E(L)&=\sum\limits_{j=1}^\infty \vec{\pi}_1 R^{j-1}(j\vec{e}+(1,1,1,1,1,2,2,2,2,2)^T) \notag \\ &=\vec{\pi}_1(\sum\limits_{j=1}^\infty jR^{j-1}\vec{e} + \sum\limits_{j=1}^\infty R^{j-1} (1,1,1,1,1,2,2,2,2,2)^T) \\ &=\vec{\pi}_1((I-R)^{-2}\vec{e}+(I-R)^{-1}(1,1,1,1,1,2,2,2,2,2)^T)\notag \end{align} \subsection{Numerical Example of Comparing Two Systems} Set $a = 0.99$, $b = 0.98$ and $c = 0.1$. Then the expected number in the two systems (5 phase system vs M/M/1 with lowest service rate of the 5 phase system) in terms of $\lambda$ and $\mu$ is shown in figure \ref{3D}. \begin{figure}[htb] \centering \includegraphics[scale=0.5]{3D_new} \caption{Expected Numbers of Customers Varying with $\lambda$ \& $\mu$} \label{3D} \end{figure} The expected numbers of the 5 phse system are plotted in red, and those of the $M/M/1$ system are plotted in blue. We see Figure \ref{2D} that the new system is better than the regular one only when the load $\frac{\lambda}{\mu}$ is within a certain range $(k_1, k_2)$, where $k_2 = c$. The value of $\lambda/\mu$ such that two systems have the same $E(L)$ is $k_1$. \begin{figure}[htb] \centering \includegraphics[scale=0.6]{rho.eps} \caption{Expected Numbers of Customers Varying with $\rho=\lambda/\mu$} \label{2D} \end{figure} The value of $k_1$ could be estimated using MATLAB (see Appendix A). For $a=0.99$, $b=0.98$ and $c=0.1$, $k_1$ is calculated to be around 0.02358. Thus, with $\lambda/\mu \in (0.02358,0.10000)$, the 5 phase system performs better than the regular $M/M/1$ system. \section{Conclusion} Through comparisons on mean service rates and expected numbers of customers, we are able to state that, with two kinds of working vacations and one phase for regular rest, a 4 phase queueing system can never outperform the regular $M/M/1$ system with the minimal service rate. However, after we add another phase for the working vacation, it is possible for the queueing system to outperform the regular $M/M/1$ system, but only when the ratio of $\lambda$ and $\mu$ is within a certain range. The boundary of that range depends on the service rate decrease during working vacations. Basically, there is evidence that sleep is a valuable tool in allowing the brain to recuperate to its normal functioning. In a better model of the brain's recovery system, there would be a larger number of phases and the service rate would be large initially and drop off close to zero in the final phase. Our limited 5 phase model indicates that there is a real possibility for improved functioning with a good sleep cycle. The exact parameters of such a cycle would need to be estimated by a large data set, but the analysis here suggests that such a data collection is a valuable resource. \section{Acknowledgment} We acknowledge funding and support from MITACS Global Internship program, University of Windsor, Central South University, CSC Scholarship. \quad

369859db03be889bd43e35cdbf19bfca3aace7d6

\section{Introduction} A broad spectrum is required when generating ultrashort laser pulses for exploring very fast molecular processes. Flamentation in gas \cite{chin2005,berge2007,cou2007} or in hollow-core fibers \cite{wadsworth_supercontinuum_2002, zheltikov_let_2006} is particularly utilized for spectral broadening to generate few cycle white light laser pulses with high time resolution. Such short pulses were for example applied to obtain electronic wave packets \cite{li_coherent_2015}. Tailoring of laser pulses is another current achievement with high potential since by using longer shaped pulses, it was already demonstrated that excitation pathways can be selected so that ionization \cite{wollenhaupt2002interferences, lindinger2004isotope}, dissociation \cite{assion1998control, kling_control_2006}, or fluorescence \cite{roth2009quantum} can be controlled. The optimal pulse shapes were often found in a closed loop scheme. Recently, pulse shaping was for the first time used with white light laser pulses from filaments in gases to optimize the ionization ratio of small silver clusters \cite{hagemann_2015} and for the energy deposition in rare gases \cite{gat2018}. Laser filamentation in gases is a fascinating process involving dispersion, multiple nonlinear and spatial effects, as well as ionization dynamics. For the necessary cancellation of self-focusing and plasma defocusing high peak intensities are required, either by using high powered lasers or extremely short pulses. For many applications a white light supercontinuum is generated by using sapphire plates \cite{imran2010efficient, jarman2005supercontinuum}. In contrast, we utilize an intense Ti:Sa laser with a two-stage flamentation in air to create the broad spectrum \cite{matt2018,stein1999}. This approach has the advantage of providing comparatively high pulse energies. Ionization processes depend on the maximal pulse intensity, the wavelength and the ionization potential. The Keldysh parameter \cite{Keldysh1965} is relevant to decide whether ionization follows mainly multi-photon ionization (MPI) or tunnel-ionization (TI). Theories were developed (ADK \cite{ammosov_tunnel_1987}, PPT \cite{perelomov_ionization_1966}) in order to describe the ionization rates \cite{talebpour1999semi}. They enabled the simulation of filamentation and laser pulse propagation and provided insight into the dynamics resulting from spatial and temporal effects. Spectral blue shifting \cite{le_blanc_spectral_1993}, temporal pulse breaking \cite{bree2010plasma} and filamentation over long distances \cite{durand_kilometer_2013} have been observed. Laser filamentation is highly nonlinear and sensitively relies on the incoming pulse shape. Recent investigations of the influence of linear chirp and of the temporal and spatial focusing enabled the shift and control of the spatial onset of the filament \cite{kasparian_white-light_2003} and the pulse shaping of the filaments \cite{heck2006,acker2006,zah2014}. In this paper we want to study the influence of the shape of ultrabroadband pulses on the resulting spectra after filamentation in argon. Our unique laser system with pulse shaping unit allows us to arbitrarily change the phase and amplitude of octave-spanning white light laser pulses \cite{Hagemann2013}. The spectra at the filament output will provide insight into the interplay of the occuring effects and provide guidance to get various spectral forms by tuning the initial pulse parameters. This contribution is partially received from the doctoral thesis of A. Patas \cite{Patas}. \section{Experimental Setup} \begin{figure}[!] \centering \includegraphics[width=0.5\textwidth]{figure_1.pdf} \caption{TG-FROG traces of chirped and unchirped laser pulses before the argon chamber. a) Negatively chirped pulse with $-20\,fs^2$. b) Transform-limited pulse. c) Laser pulse which is positively chirped with $20\,fs^2$. Higher wavelengths are underrespresented due to the calibration of the TG-FROG.} \label{TG-Frog} \end{figure} The laser system used in this experiment is capable of producing few cycle, white light laser pulses and was extensively described earlier \cite{Hagemann2013}. It consists of a Ti:Sa oscillator (Femtosource Compact, Spectra-Physics) followed by a multipass amplifier (Odin C, Quantronix) resulting in pulses at a rate of $1\,kHz$ with a central wavelength of $800\,nm$ and a spectral width of about $100\,nm$. The laser beam diameter amounts to 8$\pm0.3$ $mm$. The laser power is reduced using an aperture to $490\,mW$ which has shown to give the optimal spectral broadening after a first $f=4\,m$ filamentation stage in air. The spectrally broadenend and chirped pulses are then compressed by a pair of broadband chirped mirrors. Then a second $(f=3\,m)$ filamentation is used to further broaden the laser spectrum to the octave-spanning result. The pulses are again compressed with another pair of chirped mirrors before entering the pulse shaping setup. The folded 4f setup uses cylindrical focusing mirrors instead of lenses to avoid chromatic abberation. We use a pulse shaper (SLM640, CRI) with two liquid crystal arrays of 640 pixels each. These arrays have their optical axes oriented at plus and minus 45 degrees with respect to the laser polarization. This allows for phase and amplitude modulation of the complete white light spectrum. Extensive tests and calibration were required to ensure a correct phase and amplitude response at each wavelength. A wire-grid polarizer is used after the pulse shaping setup to allow for amplitude modulation if required. We were able to verify the performance of the setup with the help of a TG-FROG \cite{Hagemann2013}. Using the pulse shaper to compensate for the remaining pulse-chirp we repeatedly measured TG-FROG lengths of $5-7\,fs$ (FWHM) with pulse energies up to $50\,\mu J$ just before the measurement chamber. Since the TG-FROG is very sensitive to the peak-intensity, we were not able to measure pulses with more complex shapes. As a demonstration of the capability of the setup Fig. \ref{TG-Frog} shows positive and negative linearly chirped pulses as well as the short pulse as a reference. The corresponding laser spectrum is shown in Fig. \ref{TLvsRef} as a dashed red line. After a couple of beam steering mirrors the beam is focused by a $250\,mm$ off axis focusing mirror within a gas chamber which is filled with 9 bar of argon (for details see \cite{matt2018}). The spectrum was measured after the laser beam passed the exit window of the chamber and hit a Teflon beam block with the spectrometer (USB2000 UV/VIS, Ocean Optics) pointing at the beam block. This method enables to detect the entire spectrum. Teflon is used because it reduces the amount of interference from multiple reflecting surfaces. \begin{figure}[ht] \centering \includegraphics[width=0.8\textwidth]{figure_2.pdf} \caption{Laser reference spectrum used for the experiments (dashed red line). The spectrum was recorded directly at the entrance of the chamber filled with 9 bar of argon. The strongly modified spectrum for the same transform-limited laser pulse after transmission through the argon chamber is also shown (solid black line).} \label{TLvsRef} \end{figure} The chirp induced by the gas as well as the entrance windows had to be accounted for each day, since small variations measurably changed the obtained pulse length. Therefore, an electret microphone was placed close to the filament position where it recorded the acoustic shock wave released by the pulse. This value is known to be correlated to the free carrier density produced by the laser pulse \cite{yu2003}. With the help of the PRISM \cite{Wu2011} algorithm the phase was optimized for the maximum acoustic shock before each experiment. Simulations show that this method is indeed a very suitable approach for finding the correct dispersion compensating offset phase at the onset of a filament. \section{Results} The dispersion compensating offset phase found by the PRISM algorithm was used to create precompensated pulses with the pulse shaper. These pulses are close to transform-limited at the onset of the filament. Fig. \ref{TLvsRef} (black line) shows the spectrum obtained after the chamber when using this offset phase. This phase compensates the dispersion of the gas and the optical elements until the position of the filament, \begin{figure}[t] \centering \includegraphics[width=0.8\textwidth]{figure_3.pdf} \caption{Contour plot of the spectra behind the chamber as vertical lines depending on the linear pulse chirp. A broadening of the spectra is obtained close to zero chirp and modulations occur in the low wavelength range for positive and negative chirps likewise.} \label{lscan} \end{figure} so that the spatial focus coincides with the temporal focussing of the laser pulse. Thereby the maximal peak intensity and ionization rate is reached in the beginning of the filament, resulting in the largest microphone signal. Dispersion phase compensation done at 3, 7 and 9 bar, with or without extra window in front of the chamber allowed us to find the exact dispersion matching one bar argon or a single chamber-window. The obtained values are $610.5\,fs^2/m$ for GVD and $326.8 \, fs^3/m$ for GDD. The obvious pulse parameters to examine are the linear and quadratic chirp. Therefore spectra were taken while tuning the linear and quadratic chirp ($b_3$) from $-200\,fs^2$ to $200\,fs^2$ (Fig. \ref{lscan}) and $-400\,fs^3$ to $400\,fs^3$ (Fig. \ref{qscan}), respectively. The center wavelength for the chirp expansion was chosen to be $761.22\,nm$ which coincided with the center of the input spectrum. It is known from former studies \cite{Sprangle2002,berti2015nonlinear} that only a combination of self-phase modulation, self-steepening and plasma effects leads to the measured spectra. The plasma acts as a defocussing element on the beam and thereby counteracts the self-focussing by the Kerr-effect. This was shown to result in repetitive focussing and defocussing \cite{mlejnek1998dynamic, talebpour1999re}. \begin{figure}[b] \centering \includegraphics[width=0.8\textwidth]{figure_4.pdf} \caption{Spectra modified by nonlinear and plasma effects in the chamber. The x-axis denotes the quadratic chirp of the pulses at the focal position. Strong spectral bands are observed which differ for positive and negative chirps. The asymmetry between positive and negative chirps is visible, resulting from the asymmetry of the pulse shapes.} \label{qscan} \end{figure} The linear chirp scan shows modulations which shift to higher energies for smaller values in absolute chirp. Close to the transform-limit nonlinear effects lead to the strongest spectral broadening. The asymmetry in the broadening for positive and negative linear chirps can be explained by a red-shifted rising front and a blue-shifted trail for positive chirps which leads to increased broadening in a nonlinear medium and vice versa less broadening for negative chirps. It should be mentioned that, moreover, a small blue shift due to the onset of free electron plasma generation occurs for all pulse shapes \cite{wood91}. This shift depends on the laser pulse induced ionization rate for generating free electrons. \begin{figure}[t] \centering \includegraphics[width=0.8\textwidth]{figure_5.pdf} \caption{Experimental results for wavelength-shifted third order phase functions with a scaling factor of $b_3 = 2\cdot10^2 fs^3$. The x-axis denotes the center wavelength of the chirp expansion. With a decreasing antisymmetry wavelength a spectral band shifts accordingly to lower wavelengths.} \label{l0shift} \end{figure} \begin{figure}[ht] \centering \includegraphics[width=0.6\textwidth]{figure_6.pdf} \caption{Recorded spectra for selected third order phase values and center wavelengths of expansion (see Fig. \ref{l0shift}). a) Spectra for pulses of $100\,fs^3$ third order phase factor and $\lambda_c$ at $800\,nm$ and $1000\,nm$, respectively. b) Corresponding spectra for laser pulses of $200\, fs^3$ third order phase factor. In these cases, narrower spectral features and shifts to longer wavelengths are observed.} \label{l0shiftcut} \end{figure} As a next step to understand the influence of pulse modulation we investigated the spectral dependence on the quadratic chirp (see Fig. \ref{qscan}). We observe two dominant asymmetric spectral bands which shift towards higher frequencies for lower absolute chirp values. The differences in the spectra for positively and negatively chirped pulses originate from the temporal pulse asymmetry. For laser pulses with $b_3>0\,fs^3$ the pulse has a slow leading and a steep trailing flank while for $b_3<0\,fs^3$ the argon atoms experience a steep rise in the electric field strength followed by a long decrease. The slope determines the wavelength shift by self-phase modulation which is visible in Fig. \ref{qscan}. For positively quadratically chirped pulses, one observes a narrow band shifted far towards higher frequencies (due to the steep trailing flank) and a broad spectral band shifted to lower frequencies (resulting from the slow leading flank). These effects decrease for larger chirps because it is primarily dependent on the peak intensity. Besides the shifted band we find a broad dip in the spectrum for positive quadratic chirps which broadens for pulses with lower quadratic chirp and higher peak intensities. This feature can be used to optimize intensity ratios between spectral regions of filament spectra when those are experimentally required. In order to see whether these features could be controlled in an easy parametric way, we employed a parametrization consisting of a quadratic chirp expanded around a selectable frequency $\omega_c$. \begin{equation} \varphi(\omega) = \frac{b_3}{6} (\omega-\omega_c)^3 \end{equation} This is an antisymmetric phase function which was shown to optimize second and third order processes around the wavelength of antisymmetry \cite{pastirk_selective_2003}. The parametrization was chosen because the dominating $\chi^{(3)}$ processes in white light generation are expected to be controllable in a similar way. The obtained spectral modifications in Fig. \ref{l0shift} can be understood by taking a closer look at the turning points of the third order phase function. At these points of antisymmetry ($\omega_c$) the first derivative of the phase (the instantaneous frequency) $\partial\varphi(t)/\partial t$ becomes 0. This implies that, at the peak intensity, frequency-components around $\omega_c$ are dominant. Hence, the spectral maxima in the $\lambda_c$ shift scans can be seen as beeing partially moved in correspondance to these antisymmetry points. The given explanation is supported by similar slopes of the center wavelength and the wavelengths of the maxima. It can further be observed that a shift of $\lambda_c$ to lower wavelengths results in a shift of the main peak to lower wavelengths. By comparing Fig. \ref{l0shiftcut} a) and b) it is obvious that larger quadratic chirps lead to sharper spectral features which can be explained by a spectrally narrower central zero phase region. These features can also be regarded as a demonstration of the feasible control of the nonlinear filament properties by multiphoton intrapulse interference effects \cite{pastirk_selective_2003} utilizing third order phase functions. A smaller shift of the spectral peak compared to the central wavelength is noticed for pulses with $b_3 = 200fs^3$ (Fig. \ref{l0shiftcut} b)) compared to the lower third order phase factor of $100 fs^3$ in Fig. \ref{l0shiftcut}. This observation of a minor maxima shift compared to the central wavelength is due to more dominating third order phases corresponding with reduced peak intensity. In Fig. \ref{l0shiftcut} b) it is e.g. visible that for $\lambda_c = 1000\,nm$ the received less intense pulse generates only minor spectral modulations by self-phase modulation or ionization. As a useful method for controlled spectral modification one can choose the desired spectral position by adjusting the center wavelength and one can select the spectral width by modifying the prefactor. This fast parametric spectral modulation can favorably be utilized for further applications. \begingroup \squeezetable \begin{table}[b] \caption{Results of the optimizations for the given spectral ranges. A large difference is obtained between both solutions.} \vspace{0.4cm} \begin{tabular}{c | c | c | c | c | c} $(600-650)\,nm$ & $(650-700)\,nm$ & $\lambda_c$ & $b_2$ & $b_3$ & $b_4$\\ \hline min & max & $631.9\,nm$ & $69.6\,fs^2$ & $1152\,fs^3$ & $4455\,fs^4$\\ max & min & $733.0\,nm$ & $-2.1\,fs^2$ & $386\,fs^3$ & $1422\,fs^4$\\ \end{tabular} \end{table} \endgroup After understanding the basic effects of parametric pulse shaping on the generated filament spectra we employed an evolutionary algorithm \cite{Patas} to optimize a selected ratio of two spectral bands. The utilized parameter set consisted of a chirp expansion up to the third order around a center wavelength $\lambda_c$. Table I shows the results of two optimizations either maximizing or minimizing the ratio of the integrated spectra between $600\,nm-650\,nm$ and $650\,nm-700\,nm$ and Fig. \ref{specAreas} displays the corresponding spectra. If the spectral band of higher wavelengths is maximized the optimization proceeds for 27 generations of the evolutionary algorithm before converging to an intensity ratio of approximately 11.7. The most favorable optimization for the inverted fitness function yielded an intensity ratio of 4.1 after 33 generations, which corresponds to an overall modulation of the ratio by a factor of about 48. Each generation has 30 individuals including one survivor from the previous generation. This result reveals how evolutionary algorithms can be used in connection with parametric pulse shaping to find optimized pulse shapes for filamentation for various experimental applications. \begin{figure}[t] \centering \includegraphics[width=0.8\textwidth]{figure_7.pdf} \caption{Optimizations of the ratios of spectral areas, where the area between $(600-650)\,nm$ was maximized and between $(650-700)\,nm$ was minimized (dashed red curve), and in contrast where $(600-650)\,nm$ was minimized and $(650-700)\, nm$ was maximized (solid black curve).} \label{specAreas} \end{figure} The inset of Fig. \ref{specAreas} presents the optimized phase functions (without the offset phase used for compensation) which lead to the displayed spectra. The spectral features can be explained with the wavelengths of the inflection points which can be regarded as local antisymmetry points. The solid phase function for maximizing $650 - 700\,nm$ shows two local antisymmetry points whereas the dashed phase function shows one inflection point in between which corresponds to spectral maxima, respectively, as can be observed in Fig. \ref{specAreas}. Hence, the parametric optimizations utilize the spectral positions of local antisymmetry points to specifically modify the spectra. This proves that higher order terms are relevant for precise spectral control. We want to conclude that filaments can be well modified in a controlled way. Pulse shapes have an immense impact on the spectrum obtained from a filament and should be examined more closely in this regard in future. Laser pulse shaping with filamentation could become a versatile tool when looking for custom spectral shapes for further experiments. \section*{Acknowledgements} The authors thank Prof. Dr. L. W\"oste for his encouragement. A. L. acknowledges the Klaus Tschira Foundation (KTS) for financial support (project 00.314.2017). M.M. acknowledges funding from MHV fellowship grant number: PMPDP2-145444 and NCCR MUST Womens Postdoc Award. S. H. acknowledges cofunding under FP7-Marie Skłodowska-Curie – NCCR MUST program (200021-117810). The work was supported by the ERC advanced grant "Filatmo" and the SNF NCCR MUST grant.

0a1b8a6a164481c00fdd1dc4ed7cf7cd56a56bfa

\section{Motivations} The Standard Model has proven to be a solid framework to describe elementary particles, although in recent times more and more evidences demand an extension of the current theory. Together with a lack of direct detection of new particles at colliders, the current situation requires an increase in the precision of measurements and theoretical calculations to investigate indirect signatures of beyond the Standard Model physics. The Higgs boson represents a good candidate to pursue this research project: its properties and couplings are still under investigation, and on general grounds we expect a change in the $ggH$ coupling due to new physics at $\mathcal{O}(1\,\textup{TeV})$ to be around $5\%$. This sets the goal for theoretical uncertainties to be below $\mathcal{O}(1\%)$. Considering gluon fusion, the main channel of production of Higgs bosons at the LHC, the theoretical uncertainty (see \cite{Anastasiou:2016cez,Patrignani:2016xqp,Mistlberger:2018etf}) at present originates from $\mathcal{O}(2\%)$ residual scale uncertainty in pure QCD contributions, $\mathcal{O}(1\%)$ uncertainty caused by unknown mass effects of $b$ and $c$ quarks in higher order QCD corrections, and $\mathcal{O}(1\%)$ uncertainty in QCD-EW contributions. The case of QCD-EW corrections is peculiar. Indeed, although the LO ($\alpha_{~}^2 \alpha_S$) contribution is know for arbitrary values of $m_H$ and $m_{W,Z}$ \cite{Aglietti:2004nj,Aglietti:2006yd}, QCD corrections ($\alpha_{~}^2 \alpha_S^2$) to this kind of diagrams are known only using different approximations: either in the limit $m_H \ll m_{W,Z}$ \cite{Anastasiou:2008tj} or using factorization approaches \cite{Actis:2008ug,Actis:2008ts}. Corrections to the fermionic channel has been computed in \cite{Degrassi:2004mx}. To decrease the corresponding theoretical uncertainty it is therefore necessary to evaluate the NLO contributions to mixed QCD-EW corrections to Higgs production in gluon fusion for physical values of $m_H$ and $m_{W,Z}$ and full QCD structure \cite{Bonetti:2016brm,Bonetti:2017ovy,Bonetti:2018ukf}. \section{Amplitude} To compute the NLO cross-section we start considering QCD virtual corrections to the LO QCD-EW contributions to $gg \to H$, as shown in Fig.~\ref{QCD-EW}. \begin{figure \centering \subfloat[][LO.] {\includegraphics[width=0.47\textwidth]{Imm/Gluon_fusion_EW_LO_BW}} \qquad \subfloat[][Virtual NLO.] {\includegraphics[width=0.47\textwidth]{Imm/Gluon_fusion_EW_NLO_BW}} \caption{Sketch of Feynman diagrams for mixed QCD-EW contributions to $gg \to H$.} \label{QCD-EW} \end{figure} In variance with the pure QCD case, where the main contribution is given by the top quark coupling to the Higgs boson, in this class of QCD-EW corrections the Higgs boson always couples to EW vector bosons, which subsequently couple to the QCD part through a fermionic loop. The first effect of this modified coupling is that the top quark contribution is suppressed with respect to the contribution of light quarks \cite{Degrassi:2004mx}, since the enhancing factor given by the Yukawa coupling is not present. We can therefore consider light quarks only (the first two generations for the diagrams containing $W$ bosons, all quarks but the top in case of $Z$ bosons), and work in the approximation of massless quarks. Moreover, since $W$ and $Z$ bosons never appear together in the same diagram, each single diagram depends on two parameters only: the square of the mass of the EW vector boson $m_V^2$ and the square of the mass of the Higgs boson $s=m_H^2=(p_1+p_2)^2$. There are 47 Feynman diagrams both for $W$ and $Z$ bosons coupling to $H$. The amplitude consists of a single form factor \begin{equation} \mathcal{M}^{c_1 c_2}_{\lambda_1 \lambda_2} = \delta^{c_1 c_2} \epsilon_{\lambda_1}(\mathbf{p}_1) \cdot \epsilon_{\lambda_2}(\mathbf{p}_2) \mathcal{F}(s,m_W,m_Z). \end{equation} In particular \begin{equation} \label{eqa} \mathcal{F}(s,m_W,m_Z) = -\mathrm{i}\frac{\alpha_{~}^2 \alpha_S(\mu) v}{64 \mathrm{\pi} \sin^4\theta_W} \sum_{V=W,Z} C_V A(m_V^2/s,\mu^2/s), \end{equation} where \begin{equation} C_W = 4, \qquad\qquad C_Z = \frac{2}{\cos^4\theta_W}\left( \frac{5}{4}-\frac{7}{3}\sin^2\theta_W+\frac{22}{9}\sin^4\theta_W \right). \end{equation} The quantity $A$ in Eq.~(\ref{eqa}) can be naturally expanded in powers of $\alpha_S$. This gives \begin{equation} A(m^2/s,\mu^2/s) = A_\textup{2L}(m^2/s) + \frac{\alpha_S(\mu)}{2\mathrm{\pi}} A_\textup{3L}(m^2/s,\mu^2/s) + \mathcal{O}\left( \alpha_S^2 \right). \end{equation} The three-loop contribution $A_\textup{3L}(m^2/s,\mu^2/s)$ can be written as a linear combination of 95 three-loop scalar \emph{master integrals} $\mathcal{I}(s,m,\epsilon)$, where $\epsilon=(4-D)/2$ is the dimensional regularization parameter. \section{Differential equations and uniformly transcendental functions} It is useful to change variables from $(s,m^2)$ to $(s,y)$, where $y$ is defined as \begin{equation} y = \frac{\sqrt{1-4m^2/s}-1}{\sqrt{1-4m^2/s}+1}. \end{equation} With these new variables, the $s$-dependence of the master integrals can be determined simply by dimensional analysis of the integrals. This gives \begin{equation} \mathcal{I}_n(s,y,\epsilon) = (-s-\mathrm{i}0)^{a_n-3\epsilon}\mathcal{J}_n(y,\epsilon). \end{equation} We determine the functions $\mathcal{J}_n(y,\epsilon)$ using differential equations. We derive them by differentiating the master integrals with respect to $y$ and we include in cascade DEs for subtopologies appearing in the r.h.s. until we obtain a closed system of first-order differential equations. We then proceed to solve this system of equations. As shown in Ref.~\cite{Henn:2013pwa,Argeri:2014qva}, in many cases of interest it is possible to express the master integrals using so-called \emph{uniformly transcendental functions}, i.e. functions that admit series expansion in $\epsilon$ with coefficient of the $\epsilon^n$ term having weight $n$. A function $F_n(y)$ has \emph{weight} $n$ if it can be written as $n$ nested integrations over $\mathrm{d}\log R_n(\xi)$, where $R_n(\xi)$ is a rational function in $\xi$ \begin{equation} F_n(y) = \int_0^y \dots \int_0^{\xi_{n-1}} \,\mathrm{d}\log R_n(\xi_n) \dots \mathrm{d}\log R_1(\xi_1) \qquad \Rightarrow \qquad w(F_n) = n. \end{equation} This definition of weight admits an unambiguous extension to constants: weight $n$ functions evaluated in rational points give weight $n$ constants. Moreover, a function which is a product of two functions with weights $n_1$ and $n_2$ has weight $n_1+n_2$. Uniformly transcendental functions satisfy a characteristic Cauchy problem \cite{Remiddi:1999ew,Henn:2013pwa,Lee:2014ioa}. Their differential equations can be written in the so-called \emph{canonical fuchsian} form \begin{equation} \frac{\mathrm{d}}{\mathrm{d}y}\mathbf{F}(y,\epsilon) = \epsilon \sum_{a=1}^A B_a \frac{\mathrm{d}\log R_a(y)}{\mathrm{d}y}\mathbf{F}(y,\epsilon), \end{equation} where the $\epsilon$ dependence is completely factorized, $B_a$ are matrices which elements are rational numbers, and kinematic variables are present only in simple poles coming from the $\mathrm{d}\log R(y)/\mathrm{d}y$ structures, where $R(y)$ is a rational function of $y$. Furthermore, it is much easier to fix the integration constants for uniformly transcendental functions: by comparing the solution of the differential equations with a boundary value at a rational point $y_0$ it is possible to express the integration constants in terms of simple rational combinations of a small set of constant with weight given by the order in the $\epsilon$ expansion. In other words, the boundary value $L(y,\epsilon)$ in $y \to y_0$ is a uniformly transcendental sum of constants. In order to exploit these useful properties of uniformly transcendental functions, we proceed to tune our system of differential equations and boundary conditions in order to obtain a canonical system of equations and uniformly transcendental boundary functions. In our case we explicitly cast the system of differential equations into a canonical fuchsian form, while we investigate directly the boundary conditions only for the simplest subtopologies, common to all the master integrals. This procedure has proven to be sufficient in our case to find uniformly transcendental expressions for all our master integrals. Starting from the Cauchy problem for uniformly transcendental functions, its solution can be written in terms of a Dyson series in $\epsilon$ thanks to the canonical structure of the DEs \cite{Remiddi:1999ew,Henn:2013pwa,Argeri:2014qva,Lee:2014ioa} \begin{multline} \mathbf{F}(y,\epsilon) = \mathcal{P}_y\mathrm{e}^{\epsilon \int A(\xi)\,\mathrm{d}\xi}\mathbf{F}_0(\epsilon) = F_0^{(0)} + \left[ \int_y A(\xi_1) \mathbf{F}_0^{(0)}\,\mathrm{d}\xi_1 + \mathbf{F}_0^{(1)} \right] + \\ + \left[ \int_y A(\xi_1) \int_{\xi_1} A(\xi_2) \mathbf{F}_0^{(0)}\,\mathrm{d}\xi_2\mathrm{d}\xi_1 + \int_y A(\xi_1) \mathbf{F}_0^{(1)}\,\mathrm{d}\xi_1 + \mathbf{F}_0^{(2)} \right] + \mathcal{O}\left( \epsilon^3 \right) \end{multline} where $A(\xi) = \sum_{a=1}^A B_a \mathrm{d}\log R_a(\xi)/\mathrm{d}\xi$. Nested integrations are naturally expressed in terms of \emph{Goncharov Polylogarithms} \cite{Goncharov} \begin{equation} G(a_n,a_{n-1},\dots,a_1;y) := \int_0^y \frac{1}{\xi-a_n} G(a_{n-1},\dots,a_1;\xi) \,\mathrm{d}\xi \end{equation} with $G(z) := 1$ and $G(0_n,\dots,0_1;y) := \log^n y / n!$. The integration constants $F_0^{(n)}$ are, order by order in $\epsilon$, simple rational linear combinations of fixed weight constants. The constants that appear in our calculation are enlisted in Table~\ref{table1}. \begin{table \centering \begin{tabular}{ccc} \toprule Weight &\multicolumn{2}{c}{Values} \\ \midrule 0 &\multicolumn{2}{c}{$1$} \\ 1 &\multicolumn{2}{c}{$\emptyset$} \\ 2 &\multicolumn{2}{c}{$\mathrm{\pi}^2$} \\ 3 &\multicolumn{2}{c}{$\zeta(3)$} \\ 4 &\multicolumn{2}{c}{$\mathrm{\pi}^4$} \\ 5 &$\mathrm{\pi}^2 \zeta(3)$ &$\zeta(5)$ \\ 6 &$\mathrm{\pi}^6$ &$\zeta^2(3)$ \\ \bottomrule \end{tabular} \caption{Values appearing in the constant terms at each weight.} \label{table1} \end{table} Given their small number, it is possible to fit them using numerical evaluation with high number of digits at the boundary and to compare the result to the expected boundary value $\mathbf{L}(y,\epsilon)$ \begin{equation} \lim_{y \to 1} \left[ \mathbf{F}(y,\epsilon)-\mathbf{L}(y,\epsilon) \right]. \end{equation} \section{Calculation of the virtual NLO QCD-EW contributions} The system of DEs for the 95 MIs of $A_\textup{3L}(m^2/s,\mu^2/s)$ in canonical fuchsian form reads \begin{multline} \mathrm{d}\mathbf{F}(y,\epsilon) = \epsilon \left[ B_+ \, \mathrm{d}\log(1-y) + B_r \, \mathrm{d}\log(y^2-y+1) + B_- \, \mathrm{d}\log(y+1) + B_0 \, \mathrm{d}\log y \right] \mathbf{F}(y,\epsilon). \end{multline} The $\mathrm{d}\log$ structures appearing in the the system (the same appearing in the differential equations for $A_\textup{2L}(m^2/s)$) can be related to the cuts on the diagrams: $(1-y)$ corresponds to the cut of all massless lines, $(y^2-y+1)$ to the cut of just one massive line, $(y+1)$ to the cut of two massive lines, and $y$ to the residue at $s \to \infty$. To fix the integration constants we choose to compare the results from the differential equations with an independent evaluation at the point $y \to 1$. This particular point allows for a fast numerical evaluation of GPLs at high precision, as well as for a clear physical interpretation of the result, given by the fact that $y \to 1$ corresponds to the limit $m^2 \gg s$. We can therefore compute the boundary values of the integrals for the matching by performing a \emph{large-mass expansion} of our functions \cite{Smirnov:2002pj}. We compute the large-mass expansion for the master integrals in the following way. MIs depend on two different scales: the external momenta $p_{1,2}$ and the internal mass $m$. The large-mass expansion corresponds to the \emph{mathematical limit} $p_1 \sim p_2 \sim \sqrt{s} \ll m$, where configurations of loop momenta $\{k\}_I$ that can produce a non-vanishing contribution correspond to internal momenta either scaling as $\sqrt{s}$ or $m$. These contributions must also satisfy a ``large-momentum conservation law'', which states that large momentum cannot be created, destroyed or provided by external legs. This forces lines carrying large momentum to be internal ones and to be arranged in a closed circuit. After Taylor expanding the integrand in all small parameters up to the required order, we sum over all non-vanishing integrals to obtain the large-mass expansion. Notice that this procedure allows for a diagrammatic approach, giving us a collection of tadpoles, massless bubbles and massless triangles, all available in the literature. As an example, consider the integral \begin{equation} \imineq{Imm/MI7}{12}. \end{equation} Its large-mass expansion consists of two terms (thick lines indicate large momentum flow) \begin{align} \imineq{Imm/MI7_case1}{10} \quad\to\quad& \imineq{Imm/BMI7}{10} \quad\times\quad \imineq{Imm/T2MI7}{10}, \\ \imineq{Imm/MI7_case2}{10} \quad\to\quad& \imineq{Imm/T3MI7}{10} \quad+\quad s\frac{2(1+\epsilon)}{2-\epsilon}\,\imineq{Imm/T3MI7_2}{10} \quad+\quad \mathcal{O}\left(\frac{(-s)^2}{(M^2)^{4}}\right). \end{align} All the MIs have been numerically checked against \texttt{SecDec} \cite{Borowka:2015mxa} and \texttt{pySecDec} \cite{Borowka:2017idc} for multiple values of $y$. Agreement was found in all cases. \section{NLO cross-section} As expected, virtual NLO contribution shows both UV and IR divergencies. The UV divergencies are fixed by renormalization of $\alpha_S$ (since we are considering QCD corrections to LO amplitude). The IR divergent part has to cancel against real corrections. Its form is given by Catani's formula \cite{Catani:1998bh} \begin{gather} A_\textup{3L} = \mathbf{I}_g^{(1)} A_\textup{2L} + A_\textup{3L}^\textup{fin}, \\ \mathbf{I}^{(1)}=\left(\frac{-s-i0}{\mu^2}\right)^{-\epsilon} \frac{\mathrm{e}^{\epsilon\gamma_E}}{\Gamma(1-\epsilon)} \left[-\frac{C_A}{\epsilon^2}-\frac{\beta_0}{\epsilon}\right]. \end{gather} where $A_\textup{2L}$ up to $\mathcal{O}\left(\epsilon^2\right)$ was computed in \cite{Bonetti:2016brm}. Taking $\sqrt{s}=\mu=m_H=125.09\,\textup{GeV}$, $m_W=80.385\,\textup{GeV}$, $m_Z=91.1876\,\textup{GeV}$, $N_C=3$ and $N_f=5$ we obtain \begin{equation} \label{eq1} \begin{array}{lcll} A_\text{LO}(m_Z^2/m_H^2,1) &= &-6.880846 &- \mathrm{i}\, 0.5784119\,, \\ A_\text{LO}(m_W^2/m_H^2,1) &= &-10.71693 &- \mathrm{i}\, 2.302953\,, \\ A^\text{fin}_\text{NLO}(m_Z^2/m_H^2,1) &= &-2.975801 &- \mathrm{i}\, 41.19509\,, \\ A^\text{fin}_\text{NLO}(m_W^2/m_H^2,1) &= &-11.31557 &- \mathrm{i}\, 54.02989\,. \end{array} \end{equation} It is possible to see that the difference between the imaginary parts of the amplitude is much bigger than the difference in the real parts. This behavior can be understood since imaginary parts of the amplitude are related to the possibility of producing on-shell intermediate particles. At the level of single diagrams, crossing both $s=0$ and $s=m_V^2$ generates imaginary parts ($s=0$ for on-shell massless fermions and $s=m_V^2$ for one on-shell massive vector boson). At the level of the amplitude (summing all diagrams) the contributions at $s=0$ for the 2-loop case vanish, since the only possible cut generates processes of the form $gg \to q \bar{q} \quad|\quad q \bar{q} \to H$, and the Higgs boson cannot couple to massless fermions, even at the loop level. This is not the case at three loops, since other cuts can give a non vanishing contribution crossing $s=0$, as depicted in Fig.~\ref{cuts}. \begin{figure \centering \subfloat[][Cut for $s=0$ at LO.] {\includegraphics[width=0.40\textwidth]{Imm/Imag3_2_BW}} \qquad \subfloat[][Non-zero contribution for $s > 0$ at NLO.] {\includegraphics[width=0.40\textwidth]{Imm/Imag4_2_BW}} \caption{Contributions from diagrams like the one on the left vanish in the sum at LO. Diagrams on the right provide a non-zero contribution in the sum at NLO.} \label{cuts} \end{figure} To obtain a value for the NLO QCD-EW contributions to the physical cross-section real corrections are required. A first result can be obtained considering that for Higgs gluon fusion in the pure QCD case it has been observed that the gluonic PDFs suppress real corrections with a highly energetic extra gluon. Assuming this property to be valid also for QCD corrections to QCD-EW processes, we can employ a soft limit for the extra gluon, which allows us to factorize the real emission from the partonic squared amplitude \begin{equation} \lim_{E_4 \to 0} \left| A_\textup{NLO}^\textup{real} \right|^2 = \frac{\alpha_S^{~}}{4\mathrm{\pi}} N_C \frac{2 p_1 \cdot p_2}{p_1 \cdot p_4 p_2 \cdot p_4} \left| A_\textup{LO}^{~} \right|^2 + \mathcal{O}\left( p_4^{-1} \right), \end{equation} where $p_4$ and $E_4$ are the 4-momentum and the energy, respectively, of the extra gluon. At the level of the hadronic cross-section we have \cite{Catani:2001ic,deFlorian:2012za,Bonvini:2013jha} \begin{equation} \label{cs} \sigma = \int_0^1 \int_0^1 f(x_1,\mu) f(x_2,\mu) \sigma_\textup{LO} z G(z,\mu,\alpha_S^{~}) \,\mathrm{d} x_2 \mathrm{d} x_1, \end{equation} where $z:=m_H^2/(S_\textup{h} x_1 x_2)$ is the energy of the core process $gg \to H$, and \begin{gather} G = \delta(1-z) + \frac{\alpha_S}{2\mathrm{\pi}}\left [8 N_C \left( \mathcal{D}_1 + \frac{\mathcal{D}_0}{2} \log \frac{m_H^2}{\mu^2} \right) + \left ( \frac{2 \pi^2}{3} N_C + \frac{\sigma_\textup{NLO}^\textup{fin}}{\sigma_\textup{LO}} \right ) \delta(1-z) \right ], \\ \mathcal{D}_0 := \left[ \frac{\log(1-z)}{1-z} \right]_+, \\ \mathcal{D}_1 := \left[\frac{\log(1-z)}{1-z}\right]_+ + (2-3z + 2 z^2) \frac{\log[ (1-z)/\sqrt{z}]}{1-z} - \frac{\log(1-z)}{1-z}, \end{gather} with $\sigma_\textup{NLO}^\textup{fin}$ being the cross-section contribution coming from the NLO finite remainder from Eq.~(\ref{eq1}). The numerical evaluation of the hadronic cross-section in Eq.~(\ref{cs}) using NNPDF30 for PDFs and the running of $\alpha_S$ gives \begin{equation} \begin{array}{lclcccrl} \sigma_\textup{LO}^\textup{QCD} = 20.6\,\textup{pb}, &~~~~~~~ &\sigma_\textup{LO}^\textup{QCD-EW} = 21.7\,\textup{pb} &~~~ &\Rightarrow &~~~ &+ 5.3 \% &\textup{~at~LO}\\ \sigma_\textup{NLO}^\textup{QCD} = 32.7\,\textup{pb}, &~~~~~~~ &\sigma_\textup{NLO}^\textup{QCD-EW} = 34.4\,\textup{pb} &~~~ &\Rightarrow &~~~ &+5.2 \% &\textup{~at~NLO} \end{array} \end{equation} showing that the enhancement given by QCD corrections is similar between pure QCD and QCD-EW $gg \to H$. \section{Conclusions} We evaluated the NLO mixed QCD-EW corrections to $gg \to H$, employing the soft-gluon limit for real emissions. This provides a modification of the cross-section from LO to NLO of $+5.2\%$, in line with the corresponding enhancement in the pure QCD case. The error previously associated with the QCD-EW contributions, coming from the discrepancy in the results obtained in complete factorization and in EFT for $m_{W,Z} \to +\infty$, has been removed by the present calculation. The next necessary step towards a full result is given by the exact evaluation of the real corrections, featuring $gg \to gH$, $q\bar{q} \to gH$, $qg \to qH$ and $\bar{q}g \to \bar{q}H$ contributions. This is an interesting and challenging task both for physics and mathematics. \bibliographystyle{BiBTeX/BIBLIOSTYLE.bst}

7ed9e1f1c17e8130d9e5db7023517e1e5b2f5026

\section{CIPT calculation of R\MakeLowercase{n} I and O\MakeLowercase{g} I} \label{sec:CIPT} \begin{table*} [t!] \begin{center} \caption{CIPT calculations of excitation spectrum, ionisation potential and electron affinity for Rn I and Og I. Experimental results for Rn I are included for comparison. Here $E_E$ and $E_T$ are experimental and theoretical CIPT excitation energies respectively with $\Delta = E_E - E_T$. We also present the calculated Land\'{e} $g$-factors and the energy difference between the experimental and theoretical excitation energies. \label{tab:RnOgSpectrum}} \begin{tabular}{l@{\hspace{0.5cm}}cc@{\hspace{0.5cm}}r@{\hspace{0.5cm}}r@{\hspace{0.5cm}}r@{\hspace{0.5cm}}r@{\hspace{0.5cm}}|l@{\hspace{0.5cm}}cc@{\hspace{0.5cm}}r@{\hspace{0.5cm}}r@{\hspace{0.5cm}}r} \toprule \toprule \multicolumn{7}{c}{Rn I} & \multicolumn{5}{c}{Og I} \\ & State & $J$ & \multicolumn{1}{c}{\parbox{1cm}{$E_E$\cite{NIST_ASD} \\ (cm$^{-1}$)}} & \multicolumn{1}{c}{\parbox{1cm}{$E_T$ \\ (cm$^{-1}$)}} & \multicolumn{1}{c}{$g_T$} & \multicolumn{1}{c}{\parbox{1cm}{$\Delta$ \\ (cm$^{-1}$)}} & & State & $J$ & \multicolumn{1}{c}{\parbox{1cm}{$E_T$ \\ (cm$^{-1}$)}} & \multicolumn{1}{c}{$g_T$} & \parbox{1.5cm}{Ref. \cite{Indelicato2007}\\ (cm$^{-1}$)} \\ \hline \\ $6s^2 6p^6$ & $^1$S & 0 & 0 & 0 & 0 & & $7s^2 7p^6$ & $^1$S & 0 & 0 & 0 & 0 \\ $6s^2 6p^5 7s$ & $^3$P$^{\rm_o}$ & 2 & 54 620 & 55 323 & 1.50 & -703 & $7s^2 7p^5 8s$ & $^3$P$^{\rm_o}$ & 2 & 33 884 & 1.50 & 34 682 \\ $6s^2 6p^5 7s$ & $^1$P$^{\rm_o}$ & 1 & 55 989 & 56 607 & 1.18 & -618 & $7s^2 7p^5 8s$ & $^1$P$^{\rm_o}$ & 1 & 36 689 & 1.17 & 38 150 \\ $6s^2 6p^5 7p$ & $^3$S & 1 & 66 245 & 67 171 & 1.76 & -926 & $7s^2 7p^5 8p$ & $^3$P & 1 & 49 186 & 1.60 & \\ $6s^2 6p^5 7p$ & $^3$D & 2 & 66 708 & 67 658 & 1.13 & -950 & $7s^2 7p^5 8p$ & $^3$D & 2 & 49 451 & 1.15 \\ $6s^2 6p^5 6d$ & $^1$S$^{\rm_o}$ & 0 & 67 906 & 69 145 & 0 & -1 239 & $7s^2 7p^5 8p$& $^3$D & 3 & 53 777 & 1.33 & \\ $6s^2 6p^5 7p$ & $^3$D & 3 & 68 039 & 68 891 & 1.33 & -852 & $7s^2 7p^5 8p$ & $^3$P & 1 & 53 881 & 1.24 \\ $6s^2 6p^5 7p$ & $^1$P & 1 & 68 332 & 69 313 & 1.09 & -981 & $7s^2 7p^5 7d$ & $^1$S$^{\rm_o}$ & 0 & 54 155 & 0 & 53 556 \\ $6s^2 6p^5 7p$ & $^3$P & 2 & 68 790 & 69 749 & 1.37 & -959 & $7s^2 7p^5 8p$ & $^3$P & 2 & 54 446 & 1.35 & \\ $6s^2 6p^5 6d$ & $^3$P$^{\rm_o}$ & 1 & 68 891 & 70 002 & 1.36 & -1 111 & $7s^2 7p^5 7d$ & $^1$S$^{\rm_o}$ & 1 & 54 725 & 1.33 & 54 927 \\ $6s^2 6p^5 7p$ & $^1$S & 0 & 69 744 & 70 800 & 0 & -1 056 & $7s^2 7p^5 7d$ & $^3$F$^{\rm_o}$ & 4 & 54 938 & 1.25 & 48 474 \\ $6s^2 6p^5 6d$ & $^3$F$^{\rm_o}$ & 4 & 69 798 & 70 742 & 1.25 & -944 & $7s^2 7p^5 7d$ & $^3$D$^{\rm_o}$ & 2 & 55 416 & 1.30 & 49 039 \\ $6s^2 6p^5 6d$ & $^3$D$^{\rm_o}$ & 2 & 70 223 & 71 188 & 1.32 & -965 & $7s^2 7p^5 7d$ & $^3$F$^{\rm_o}$ & 3 & 55 622 & 1.06 & 49 603 \\ $6s^2 6p^5 6d$ & $^3$F$^{\rm_o}$ & 3 & 70 440 & 71 334 & 1.06 & -894 & $7s^2 7p^5 8p$ & $^1$S & 0 & 55 729 & 0 \\ & & & & & & & $7s^2 7p^5 7d$ & $^1$D$^{\rm_o}$ & 2 & 56 317 & 0.98 & 50 410 \\ & & & & & & & $7s^2 7p^5 7d$ & $^5$F$^{\rm_o}$ & 3 & 56 343 & 1.25 & 50 168 \\ & & & & & & & $7s^2 7p^5 7d$ & $^1$P$^{\rm_o}$ & 1 & 57 855 & 0.84 & 58 072 \\ \multicolumn{12}{c}{Ionisation potentials} \\ $6s^2 6p^5$ & $^2$P$^{\rm_o}$ & $3/2$ & 86 693 & 87 721 & 1.33 & -1 028 & $7s^2 7p^5$ & $^2$P$^{\rm_o}$ & 3/2 & 71 508 & 1.33 & 71 320\cite{Jerabek2018} \\ \multicolumn{12}{c}{Electron Affinity} \\ $6s^2 6p^6 7s$ & $^2$S & 1/2 & & 1 868 & 2.00 & & $7s^2 7p^6 8s$ & $^2$S & 1/2 & -773$^{a}$ & 2.00 & -516 \cite{Goidenko2003} \\ \bottomrule \bottomrule \end{tabular} \end{center} \begin{flushleft} $^a$ Negative value indicates the state is bound. \end{flushleft} \end{table*} To calculate the spectra of oganesson we use a combination of the configuration interaction and perturbation theory (CIPT), introduced in ref. \cite{DBHF2017}. This technique has been used to calculate the spectra in open $d$-shell and open $f$-shell atoms with a large number of valence electrons where other many-body methods are unfeasible\cite{DBHF2017, LDFDb2018, Dzuba2018}. Calculations for W I, Ta I and Yb I are in good agreement with experiment. In this section we will give a brief overview of the CIPT method for Rn and Og. For an in depth discussion of the CIPT method refer to refs. \cite{DBHF2017}. \\ We generate the set of complete orthogonal single-electron states for both Rn I and Og I by using the $V^{N-1}$ approximation \cite{Kelly1964, Dzuba2005} (where $N$ is the total number of electrons). The Hartree-Fock (HF) calculations for atomic core are done for the open-shell configurations $6s^2 6p^5$ and $7s^2 7p^5$ for the Rn I and Og I respectively. The single-electron basis sets are calculated in the field of the frozen core using a B-spline technique with 40 B-spline states of order 9 in a box with radius 40 $a_B$ (where $a_B$ is the Bohr radius) with partial waves up to $l_{\text{max}} = 4$ included \cite{Johnson1988}. \\ The many-electron wavefunctions $ |i \rangle = \Phi_i(r_1,\dots,r_{N_e})$ are formed through single and double excitations from low-lying reference configurations. The many-electron wavefunctions are ordered by energy and divided into two sets. The first set represents a small number of low energy states which contribute greatly to the total CI valence wavefunction ($i < N_{\text{eff}}$, where $N_{\text{eff}}$ is the number of included low energy states) and the remaining wavefunctions represent a large number of high energy terms which are small corrections to the valence wavefunction ($N_{\rm eff} < i \le N_{\text{total}}$). The valence wavefunction can be written as \begin{align} | \Psi \rangle &= \sum_{i=1}^{N_{\text{eff}}} c_{i}|i\rangle + \sum_{i = N_{\text{eff}} + 1}^{N_{\text{total}}} c_{i}|i\rangle . \end{align} The off-diagonal matrix elements between the higher order states are neglected ($\langle i | H^{\text{CI}} | j \rangle = 0 $ for $N_{\rm eff} < i,j \le N_{\text{total}}$) which greatly decreases the computation time for a small sacrifice in accuracy. \footnote{It immediately follows from the perturbation theory that contributions of CI matrix elements between high states to low state energy are suppressed by a second power of large energy denominators while the contribution of matrix elements between high and low states are only suppressed by the first power in the denominator.} The matrix elements between high energy and low energy states are included pertubatively by modifying the low energy matrix elements, \begin{equation} \langle i|H^{\rm CI}|j\rangle \rightarrow \langle i|H^{\rm CI}|j\rangle + \sum_k \frac{\langle i|H^{\rm CI}|k\rangle\langle k|H^{\rm CI}|j\rangle}{E - E_k}, \end{equation} where $i,j \le N_{\rm eff}$, $N_{\rm eff} < k \le N_{\rm total}$, $E_k = \langle k|H^{\rm CI}|k\rangle$, and $E$ is the energy of the state of interest. This results in a modified CI matrix and the energies are found through solving the standard eigenvalue problem, \begin{align} \label{eq:CI_diag} (H^{\rm CI} - EI)X=0, \end{align} where $I$ is unit matrix, the vector $X = \{c_1, \dots, c_{N_{\rm eff}}\}$. The CI equations (\ref{eq:CI_diag}) are iterated in the CIPT method. For a detailed discussion of the CIPT precedure see Refs.~\cite{DBHF2017, LDFDb2018}. \\ We included both Breit interaction\cite{Breit1929, Mann1971, DF2016} and QED radiative corrections in our calculation of the Og spectra. The Breit interaction $V_B$ accounts for the magnetic interaction between two electrons and retardation. The QED corrections $V_R$ accounts for the Ueling potential and electric and magnetic formfactors\cite{FG2005}. For the calculation of the even parity states of Og the low energy reference states in the effective matrix were $7s^27p^6$ and $7s^27p^58p$ while for the odd states $7s^2 7p^5 8s$ and $7s^2 7p^5 7d$. For the calculation of the ionisation potential and electron affinity we remove or add one electron from the states in the effective matrices respectively.\\ Each level is presented with an $LS$ notation. These are selected by comparing calculated $g$-factors to the non-relativistic expression, \begin{align} g_{NR} = 1 + \dfrac{J(J + 1) - L(L+1) + S(S+1)}{2J(J+1)}. \end{align} and using the $L$ and $S$ values as fitting parameters. We stress that the presented $LS$ notations are approximations as the states of Og are highly relativistic and strongly mixed. \\ In Table~\ref{tab:RnOgSpectrum} we present the results of our CIPT calculations for Rn I and Og I. We compare the Rn I CIPT calculations to the experimental results. The lack of experimental $g$-factors for Rn I make it difficult to confirm the correct identification of the states and therefore we must rely solely on the order of the energy levels. We find that there is good agreement between the experimental and theoretical states with an agreement with $\Delta \approx -900$ cm$^{-1}$ with the largest discrepancy $\Delta \approx -1239 $~cm$^{-1}$. We expect a similar accuracy for our Og I calculations (also presented in Table~\ref{tab:RnOgSpectrum}). \\ Comparing the spectrum of Rn to Og we see that despite the similar electronic structure (with differing principal quantum numbers) there are significant differences. The Og spectrum is much more dense than Rn with the first excitation lying more than 20~000~cm$^{-1}$ below the equivalent excitation in Rn. This results in an odd parity state which lies in the optical region. This makes the state a good candidate for initial experimental measurement. In the final column of Table~\ref{tab:RnOgSpectrum} we present the states calculated in ref.~\cite{Indelicato2007}. This work also did not present $g$-factors which made comparing states uncertain, therefore we compared them by ordering energies. For 4 of the states there was good agreement with our results lying within $1000$ cm$^{-1}$ however for the other states there was a large discrepancy of $>4000$ cm$^{-1}$. \\ Our calculated value of the ionisation potential of Og in Table~\ref{tab:RnOgSpectrum} is in excellent agreement with the value calculated in Ref.~\cite{Jerabek2018} ($E_{IP}=$71~320~cm$^{-1}$) where a CCSD(T) method was used. It has been shown that Og has a positive electron affinity which is an anomaly in the group of noble gases \cite{EliavOg1996, Goidenko2003, Eliav2015}. This is another consequence of the stabilized $8s$ orbital due to the large relativistic effects. Our calculation presented in Table \ref{tab:RnOgSpectrum} confirms this with an electron affinity of 773 cm$^{-1}$ (0.095 eV) which is in good agreement with the coupled cluster value presented in \cite{Goidenko2003}. For comparison we also present the negative ion calculation for Rn I which is known to be unstable. All other negative ionic states of Og were found to be unstable. \section{Electric dipole transitions of O\MakeLowercase{g} I} \label{sec:E1} While Og follows the expected trend for elements in noble group where each consecutive element has both a smaller IP and first excitation energy. However Og has some properties which can be considered exotic even amongst the Group 18 elements. According to the calculated spectrum in Table~\ref{tab:RnOgSpectrum} it is the only noble element which has an allowed optical electric dipole (E1) transition ($\omega < $40~000 cm$^{-1}$) from the ground state, unlike Rn where the first odd state lies at 57~334 cm$^{-1}$. \\ The E1 transition amplitudes, $A_{\text{E1}}$, between states which satisfy the conditions of opposite parity and $\Delta J \leq 1$ are calculated using the many-electron wavefunctions created in the CIPT method and the self-consistent random-phase approximation which includes polarization of the atomic electron core by an external electromagnetic field. The details of the method are presented in Ref. \cite{Dzuba2018}. \\ The E1 transition rate is calculated using (in atomic units), \begin{align} \label{eq:Transitionrate} T_{E1} = \dfrac{4}{3}\left(\alpha \omega\right)^3\dfrac{ A_{\text{E1}}^2}{2J + 1} \end{align} where $J$ is the angular momentum of the upper state, $\alpha$ is the fine structure constant and $\omega$ is the frequency of the transition in atomic units. The transition amplitudes and transition rates for the allowed E1 transitions in Og are presented in Table~\ref{tab:E1_transitions}. In ref.~\cite{Indelicato2007} the major E1 transition rates were also calculated with a MCDF approach, these are included for comparison Table~\ref{tab:E1_transitions}. \\ \linebreak We calculated the rates of the $(n+1)s \rightarrow np$ transitions in lighter neutral noble elements Kr and Xe and compared them to experimental values, these are presented in Table \ref{tab:E1_comp}. The experimental uncertainties are approximately $2\%$ for Xe I transitions\cite{Xe_BSD} and $~10$-$25\%$ for Kr I transitions\cite{Kr_BSD}. Comparing our calculated values to the experimental values in Table \ref{tab:E1_comp} we see the accuracy for these transitions is from 0.6\% to 17.7\%. We used the experimental energies to calculate the transitions rates of Kr I and Xe I using (\ref{eq:Transitionrate}) and since the uncertainty in the experimental energies are negligible the uncertainty in our calculations compared to experimental results in Table \ref{tab:E1_comp} is equivalent to the uncertainty in the square of the calculated transition amplitude $A_{E1}^2$. For our calculation of the Og I transition rates we needed to take into account the non-negligible uncertainty in the energies of our CIPT calculations. Therefore assuming an accuracy of 18\% for $A_{E1}^2$ and an uncertainty of 3\% in the CIPT energy ($\left|\Delta\right| \approx 1000$~cm$^{-1}$) we expect a transition rate accuracy of 20\% for the $8s \rightarrow 7p$ optical transition ($\omega = 36~689$~cm$^{-1}$) of Og I in Table \ref{tab:E1_transitions}. \\ \begin{table}[h] \center \caption{Comparison of E1 transition rates between experimental and CIPT values for Kr I and Xe I. Here $A_{E1}$ is the transition amplitude in atomic units and $T_{E1}$ is the transition rate. \label{tab:E1_comp}} \begin{tabular}{c@{\hspace{0.5cm}}c@{\hspace{1cm}}c@{\hspace{0.5cm}}c@{\hspace{0.5cm}}c} \toprule \toprule State & $E_{\text{Exp}}$ & $A_{\text{E1}}$ & $T_{\text{E1, CIPT}}$ & $T_{\text{E1,Exp}}$ \\ & (cm$^{-1}$) & (a.u.) & ($\times 10^6$ s$^{-1}$) & ($\times 10^6$ s$^{-1}$) \\ \hline \multicolumn{5}{c}{Kr I} \\ $^1$P$_1^{\rm_o}$ & 80 916 & 0.94 & 314 & 312\cite{Kr_BSD} \\ $^3$P$_1^{\rm_o}$ & 85 846 & 0.87 & 320 & 316\cite{Kr_BSD} \\ \multicolumn{5}{c}{Xe I} \\ $^1$P$_1^{\rm_o}$ & 68 045 & 1.18 & 295 & 273\cite{Xe_BSD} \\ $^3$P$_1^{\rm_o}$ & 77 185 & 0.98 & 298 & 253\cite{Xe_BSD} \\ \bottomrule \bottomrule \end{tabular} \end{table} \begin{table}[h] \center \caption{Electric dipole transition amplitudes of Og I from the ground state $^1$S$_0$ to the excited states of odd parity and angular momenta $J=1$. Here $A_{E1}$ is the transition amplitude in atomic units and $T_{E1}$ is the transition rate. We include results of MCDF calculations from ref. \cite{Indelicato2007} for comparision. There is significant disagreement for the third transition however there is another transition in \cite{Indelicato2007} which has a rate ($986 \times 10^{6}$ s$^{-1}$) close to our calculated value. So, the disagreement may be the result of a misprint in \cite{Indelicato2007}.\label{tab:E1_transitions}} \begin{tabular}{c@{\hspace{0.5cm}}c@{\hspace{1cm}}c@{\hspace{0.5cm}}c@{\hspace{0.5cm}}c} \toprule \toprule State & $E_{\text{CIPT}}$ & $A_{\text{E1}}$ & $T_{\text{E1, CIPT}}$ & $T_{\text{E1,MCDF}}$ \cite{Indelicato2007} \\ & (cm$^{-1}$) & (a.u.) & ($\times 10^6$ s$^{-1}$) & ($\times 10^6$ s$^{-1}$) \\ \hline $^1$P$_1^{\rm_o}$ & 36 689 & 2.09 & 145 & 204\\ $^1$S$_1^{\rm_o}$ & 54 725 & 0.727 & 58.4 & 55.3 \\ $^1$P$_1^{\rm_o}$ & 57 855 & -2.67 & 936 & 9.9, 986* \\ \bottomrule \bottomrule \end{tabular} \end{table} Only the first transition in Table \ref{tab:E1_transitions} lies in the optical region and therefore it has the highest likelihood of being measured first. The large rate of the transition $^1$S$_0$~$\rightarrow$~$^1$P$_1^{\rm_o}$ is also promising for experimental measurement. \section{Electron density of O\MakeLowercase{g} } \label{sec:Relativistic} \begin{figure} \includegraphics[scale=0.58]{Ogplot.eps} \caption{Radial electron density, $4\pi\rho(r)r^2$ plot for Og I in both relativistic and non-relativistic approximations. The solid blue line and the dashed red line are non-relativistic and relativistic approximations respectively. The principle quantum peaks have been labeled for the non-relativic plot.\label{Og_plot}} \end{figure} \begin{figure} \includegraphics[scale=0.55]{Ogplot_zoom.eps} \caption{Lower right section of Figure~\ref{Og_plot}.\label{Og_plot_zoom}} \end{figure} It has been shown in Ref. \cite{Jerabek2018} using fermion localization that the electron density of Og is smoother than other group 18 analogues which have distinct atomic shells . The cause of this is the large relativistic effects in SHE which effectively smear out the shells into a smoother electron density (the same was shown for the nucleon density). The relativistic effects can also be seen by looking at the radial electron densities with relativistic and non-relativistic approximations. The Hartree-Fock radial electron density for Og is plotted on a logarithmic scale in Figure~\ref{Og_plot} in both the relativistic and non-relativistic approximations. There are a total of 7 peaks in the radial densities corresponding to the principle quantum numbers $n$ where lower shells have distinct peaks in both the relativistic and non-relativistic approximations. As expected, in the relativistic approximation the inner shells ($n=1,2,3$) shift closer to the nucleus however higher shells are relatively unaffected ($n \geq 4$). This results in a similar density profile for the electrons a large distance away from the nucleus. \\ In Figure~\ref{Og_plot_zoom} we plot the tail of the density function in Figure~\ref{Og_plot}. Here we see that, while spread out, the principle shell peaks still exist in the non-relativistic approximation. However in the relativistic approximation the density has been smoothed out to such a degree that there are no discernible peaks. This supports the results in ref.~\cite{Jerabek2018} where they calculated the electron shell structure of Og I and found that it disappears for external shells due to the high relativistic effects. This can be explained as the large spin-orbit splitting doubles the number of sub-shells which overlap making the overall distribution smooth.\\ \section{Conclusion} In this work we calculated the spectrum and E1 transitions for Og I including the ionisation potential. We demonstrated the accuracy of the calculations by comparing similar calculations of Rn I to experimental data and expect an uncertainty of no more than $|\Delta| \approx 1000$~cm$^{-1}$. We found the spectrum of Og I is dense compared to other elements in group 18 with significantly lower ionisation potential and excited states which follows the periodic trend. This compact spectrum introduces an allowed optical E1 transition which does not exist in other group 18 elements which presents a possibility for future experimental measurements. Our work also supports recent findings\cite{Jerabek2018} which suggest the electron shell structure of Og I is less prominent than lighter elements due to large relativistic effects which results in the outer electron density to becoming smooth. \\ This work was funded in part by the Australian Research Council. \bibliographystyle{apsrev4-1}

3e7d008fad27a5d88d79ac49b21f2c0085f40ef5

\section{Introduction} {Nonlinear wave processes are observed in a variety of engineering and physics applications such as acoustics~\citep{Hamilton_NLA_1998, Lighthill_JSV_1978}, combustion noise~\citep{Bonciolini_PRE_2017, DouasbinEtAl_JCP_2018}, jet noise~\citep{Baars_JETNOISE_2014, Von_PRE_2012, Baars_ExpFluids_2013}, thermoacoustics~\citep{ScaloLH_JFM_2015, GuptaLS_JFM_2017}, surface waves~\citep{AblowitzClarkson1991}, and plasma-physics~\citep{Gurbatov2012}, requiring nonlinear evolution equations to describe the dynamics of perturbations}. In the case of high amplitude planar acoustic wave propagation, two main nonlinear effects are present: acoustic streaming~\citep{Lighthill_JSV_1978, Gedeon_1997_Cryocoolers} and wave steepening~\citep{Hamilton_NLA_1998, Naugolnykh1998}. Acoustic streaming is an Eulerian mean flow and is attributed to the kinematic nonlinearities~\cite{Lighthill_JSV_1978}. Convective derivatives of velocity in momentum conservation equation cause wave induced Reynolds stresses~\citep{ScaloLH_JFM_2015} which have non-zero mean values in time. In one dimension, longitudinal stresses are generated which cause steady mass flow due to wave propagation. On the other hand, wave steepening occurs due to local gradients in the wave speed associated with thermodynamic nonlinearities~\citep{Hamilton_NLA_1998}. Wave steepening entails generation of smaller length scales via a nonlinear energy cascade, {which can be exemplified by developing the product of two truncated Fourier series,} \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{TGV_fig.pdf} \put(-510,200){$(a)$} \put(-240,200){$(b)$} \put(-510,110){$(c)$} \put(-375,110){$(d)$} \put(-240,110){$(e)$} \put(-120,110){$(f)$} \caption{ $Q$-criterion iso-surfaces colored with the local velocity magnitude obtained from a direct numerical simulation of a Taylor-Green vortex in a triply-periodic domain $\left[-\pi, \pi\right]^3$~\cite{Chapelier_jcp_2017}, exhibiting breakdown into hydrodynamic turbulence $(a)$, velocity perturbation field in a high amplitude nonlinear traveling acoustic wave (TW) $(b)$, evolution of normalized spatial average of $u^2$ $(c),(e)$, and velocity spectra $|\widehat{u}_k|^2$ $(d), (f)$ at times $t_0, t_1, t_2, t_3$. The spectral broadening occurs due to the nonlinear terms in the governing equations generating smaller length scales resulting in energy cascade from larger to smaller length scales. } \label{fig: SpectraTurbulenceNLA} \end{figure*} \begin{eqnarray} \left(\sum^{n}_{k=-n}a_k e^{2\pi ikx}\right)\left(\sum^{m}_{l=-m}b_le^{2\pi ilx}\right) = \nonumber\\ ={\sum_k a_k b_{-k}} +{\underset{k+l\neq 0}{\sum_k \sum_l} a_k b_l e^{2\pi i(k+l)x}}. \label{eq: IntroTrig} \end{eqnarray} The left hand side of the Eq.~\eqref{eq: IntroTrig} represents a generic quadratic nonlinear term appearing in a governing equation. Continued nonlinear evolution results in further generation of smaller length scales, as depicted by the second term on the right hand side of Eq.~\eqref{eq: IntroTrig}, ultimately leading to spectral broadening. In the case of nonlinear acoustic waves, the shock thickness is the smallest length scale present in the flow, governed by the viscous dissipation. {The latter causes saturation of the spectral broadening process, hence establishing an energy flow (primarily) directed from large scales to small scales. Identical spectral energy dynamics are observed in classic hydrodynamic turbulence~\cite{tennekes1972first}, where nonlinear processes such as vortex stretching and tilting (only existing in three-dimensions) cause spontaneous generation of progressively smaller vortical structures (i.e. eddies), until velocity gradients become sufficiently large for viscous dissipation to become relevant (see Figure~\ref{fig: SpectraTurbulenceNLA}).} {In previous numerical investigations~\citep{GuptaLS_JFM_2017} -- inspired by the experimental setups in \cite{BiwaEtAl_JASA_2014, YazakiIMT_PhysRevLett_1998} -- the present authors have demonstrated the existence of an equilibrium spectral energy cascade in quasi-planar weak shock waves sustained by thermoacoustic instabilities in a resonator. The latter inject energy only at scales comparable to the resonator length (large scales); harmonic generation then takes place, leading to spectral broadening and progressive generation of smaller scales until viscous losses, occurring at the shock-thickness scale, dominate the energy cascade. Building upon the findings in \citep{GuptaLS_JFM_2017}, in the present work, we mathematically formalize the dynamics of nonlinear acoustic spectral energy cascade in a more canonical setup neglecting thermoacoustic energy sources and focusing on purely planar waves. Due to the absence of physical sources of energy, the energy of nonlinear acoustic waves (if correctly defined) decays monotonically in time due to viscous dissipation at small length scales, analogous to freely decaying hydrodynamic turbulence (see Figure~\ref{fig: SpectraTurbulenceNLA})}. We study the spatio-temporal and spectro-temporal evolution of such finite amplitude planar nonlinear acoustic waves in three canonical configurations in particular: traveling waves (TW), standing waves (SW), and randomly initialized Acoustic Wave Turbulence (AWT). {In spite of it's theoretical nature, planar nonlinear acoustic theory is still commonly used in practical investigations such as sonic boom propagation (TW)~\cite{Crow_JFM_1969}, rotating detonation engines (SW)~\cite{Schwinn_CandF_2018}, combustion chamber noise (AWT)~\cite{Culik_AGARD_2006}, and thermoacoustics ~\cite{BiwaEtAl_JASA_2014, YazakiIMT_PhysRevLett_1998}.} We study these configurations for pressure amplitudes and viscosities spanning three orders of magnitudes. Utilizing the second order nonlinear acoustics approximation, we derive analytical expressions for the spectral energy, energy transfer function, and dissipation. Analogous to the study of small scale generation in hydrodynamic turbulence, well quantified by the K41 theory~\citep{kolmogorov1941a, kolmogorov1941b, kolmogorov1941c}, we also define the relevant length scales associated to fully developed nonlinear acoustic waves elucidating the scaling features of the energy spectra. To this end, we perform the direct numerical simulations (DNS) resolving all the relevant length scales~\cite{pope2000turbulent} of nonlinear acoustic wave propagation. Usually, problems in nondispersive nonlinear wave propagation are studied utilizing the model Burgers equation \citep{Whitham2011, Gurbatov_1981_JETP, Naugolnykh1998, Gurbatov2012}. The spectral energy and decay dynamics of one dimensional Burgers turbulence have been studied extensively by Kida~\citep{Kida_1979_JFM}, Gurbatov et al.~\citep{Gurbatov_JFM_1997, Gurbatov1991}, Woyczynski~\citep{Woyczynski_Gottingen_2006}, Fournier and Frisch~\cite{Fournier1983Burgers}, and Burgers~\cite{Burgers1974nonlinear}. However, the equations of second order nonlinear acoustics can be reduced to Burgers equation only assuming planar TW, thus limiting its applicability. Generalized problems involving an ensemble of acoustic waves of different amplitudes, such as AWT, have also been subjects of detailed analysis~\citep{Kida_1979_JFM, Newell_2011_ARFM}. Such studies primarily involve formulation of the kinetic equations of complex amplitudes of weakly nonlinear harmonic waves~\cite{Zakharov2012}. Utilizing the kinetic equations wave interaction potentials are defined in the context of wave-wave interactions. However, such analysis are restricted to complex harmonic representation of waves in space and time and hence, fail to elucidate the inter-scale energy transfer dynamics due to general nonlinear wave interactions.{\color{black}{ In this work, we utilize the continuum gas dynamics governing equations to elucidate the spectral energy cascade and decay dynamics of nonlinear acoustics. The nonlinear equations governing high amplitude acoustics yield novel analytical expressions for spectral energy, spectral energy flux, and spectral dissipation rate valid for planar nonlinear acoustic waves with general phasing}}. The dissipation causes power law decay of energy in time due to gradual increase of the dissipative length scale. Such decay dynamics occur due to separation of energy containing and diffusive length scales and resemble those of decaying homogeneous isotropic turbulence (HIT) \citep{batchelor1948decay, batchelor1949nature, hinze1975turbulence, Ishihara_ARFM_2009}. We present a framework for studying nonlinear acoustic wave propagation phenomenon in one dimension utilizing the second order nonlinear acoustics equations and DNS of compressible 1D Navier-Stokes (resolving all length scales). We derive the former utilizing the entropy scaling considerations for weak shocks as discussed in Section~\ref{sec: EntropyScaling}. In Section~\ref{sec: EnergyCorollary}, we derive a novel perturbation energy corollary for nonlinear acoustic perturbations utilizing the second order governing equations yielding a new perturbation energy function. Its spatial average defines the Lyapunov function of the system and decays monotonically in the presence of dissipation and absence of energy sources which is confirmed through the DNS data shown in Section~\ref{sec: numerics} along with brief explaination of numerical technique utilized. In Section~\ref{sec: ScalesAcousticCascade}, we derive the spectral energy conservation equation thus identifying the spectral energy flux and spectral dissipation utilizing the energy corollary. Furthermore, we discuss the evolution of the primary length scales involved in the spectral energy cascade and decay. Finally, in Section~\ref{sec: SpectralScaling}, we show the scaling of spectral energy, spectral energy flux, and spectral dissipation. Throughout, the theoretical results are supported utilizing the DNS of three specific cases of acoustic waves namely, single harmonic traveling wave (TW), single harmonic standing wave (SW), and random broadband noise (AWT). \section{Governing equations and scaling analysis} \label{sec: EntropyScaling} {In this section, we derive the governing equations for nonlinear acoustics truncated up to second order (in the acoustic perturbation variables) for a single-component ideal gas. We begin with fully compressible one-dimensional Navier-Stokes equations for continuum gas dynamics and analysis of entropy scaling with pressure jumps in weak shocks formed due to the steepening of nonlinear acoustic waves~(Section~\ref{sec: 1DNS}). We then briefly discuss the variable decomposition and non-dimensionalization in Section~\ref{sec: 2ndOrderEntropy}, followed by the derivation of second order governing equations for nonlinear acoustics in Section~\ref{sec: 2ndOrderNLA}.} \subsection{Fully compressible 1D Navier-Stokes and entropy scaling in weak shocks} \label{sec: 1DNS} \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{Figure1.pdf} \put(-250,200){$(a)$} \put(-250,97){$(b)$} \caption{Weak shock wave structure $(a)$ pressure $p^*$ and $(b)$ entropy $s^*$ propagating with a speed $a^*_s>a^*_0$ obtained from DNS (see Section~\ref{sec: numerics}). $\Delta s^*/R^*$ and $s^*_{\mathrm{max}}/R^*$ are the entropy jump and maximum entropy respectively. With increasing viscosity, the peak in entropy remains constant. The DNS data has been obtained for base state viscosity values given in Table~\ref{tab: test_cases}.} \label{fig: EntropyJumps} \end{figure} One dimensional governing equations of continuum gas dynamics (compressible Navier-Stokes) for an ideal gas are given by, \begin{align} &\frac{\partial \rho^*}{\partial t^*} + \frac{\partial (\rho^* u^*) }{\partial x^*} = 0,\label{eq: NS_eqns1} \\ &\frac{\partial}{\partial t^*}\left(\rho^* u^*\right) + \frac{\partial}{\partial x^*}\left(\rho^* u^{*2}\right) = -\frac{\partial p^*}{\partial x^*} \nonumber \\ &+ \frac{\partial }{\partial x}\left(\left(\frac{4}{3}\mu^* + \mu^*_B\right)\frac{\partial u^*}{\partial x^*}\right),\label{eq: NS_eqns2} \\ &\rho^* T^*\left(\frac{\partial s^*}{\partial t^*} + u^*\frac{\partial s^*}{\partial x^*}\right) = \frac{\partial }{\partial x^*}\left( \frac{\mu^*C^*_p}{Pr} \frac{\partial T^*}{\partial x^*}\right) + \nonumber \\ &\left(\frac{4}{3}\mu^*+\mu^*_B\right)\left(\frac{\partial u^*}{\partial x^*}\right)^2, \label{eq: NS_eqns3} \end{align} which are closed by the ideal gas equation of state, \begin{align} p^* = \rho^* R^* T^*,\label{eq: IdealGas} \end{align} where $p^*, u^*, \rho^*, T^*, s^*$ respectively denote total pressure, velocity, density, temperature, and entropy of the fluid, $ x^*$ and $t^*$ denote space and time, and $\mu^*$ denotes dynamic viscosity. In this work, we perform DNS of Eqs.~\eqref{eq: NS_eqns1}-\eqref{eq: NS_eqns3} to resolve all the length scales of planar nonlinear acoustic waves. For our simulations (see Section~\ref{sec: numerics}), {\color{black}{we choose the gas specific constants for air at standard temperature and pressure (STP), \begin{equation} R^* = 287.105~\frac{\mathrm{m}^2}{\mathrm{s}^2\cdot K},\quad\quad\mu^*_B=0,\quad\quad Pr=0.72. \end{equation} \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{EntropyJumps.pdf} \caption{{\color{black}{Entropy jump $\Delta s^*=s^*_2-s^*_1$ and maximum entropy generated $s^*_{\mathrm{max}}$ versus pressure jump $\Delta p^*$ across a planar shock wave. In the labeled region ($\Delta p^*/\gamma p^*_1 < 1$, referred as `weak shocks' hereafter), the entropy jump $\Delta s^*$ scales as $\mathcal{O}\left(\Delta p^{*3}\right)$, whereas the maximum entropy generated $s^*_{\mathrm{max}}$ scales as $\mathcal{O}\left(\Delta p^{*2}\right)$, approximately. Markers denote DNS data (see Section~\ref{sec: numerics}), (\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,circle,fill=black!20!black](){};}},\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,regular polygon, regular polygon sides=4,fill=black!20!black](){};}}) $\mu^*=7.5\times10^{-3}~$kg$\cdot$m$^{-1}\cdot$s$^{-1}$; (\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,circle,fill=black!20!gray](){};}},\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,regular polygon, regular polygon sides=4,fill=black!20!gray](){};}}), $\mu^*=7.5\times10^{-4}~$kg$\cdot$m$^{-1}\cdot$s$^{-1}$; (\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,circle,fill=none](){};}},\protect\raisebox{0.5pt}{\tikz{\node[draw,scale=0.4,regular polygon, regular polygon sides=4,fill=none](){};}}), $\mu^*=7.5\times10^{-5}~$kg$\cdot$m$^{-1}\cdot$s$^{-1}$ for varying values of $\Delta p^*$. Solid lines correspond to Eqs.~\eqref{eq: EntropyMach} and~\eqref{eq: EntropyMax}.}}} \label{fig: EntropyScaling} \end{figure} Planar nonlinear acoustic waves steepen and form weak shocks. \revision{For weak shocks, the smallest length scale (shock-thickness) is also significantly larger than the molecular length scales. Hence, in this work, we neglect the molecular vibrational effects in the single component ideal gas ($\mu^*_B=0$) , typically modeled via bulk viscosity effects~\cite{Cramer_PoF_2012}}. Across a freely propagating planar weak shock (Fig.\ref{fig: EntropyJumps}), the entropy jump ($\Delta s^* = s^*_2 - s^*_1$) is given by the classical gas-dynamic relation~\cite{LiepmannRoshko}, \begin{align} \frac{\Delta s^*}{R^*} &= \frac{1}{\gamma - 1}\ln\left(1 + \frac{2\gamma}{\gamma + 1}\left(M^2 -1 \right)\right) \nonumber \\ &- \frac{\gamma}{\gamma - 1}\ln\left(\frac{\gamma + 1}{\gamma - 1 + 2/M^2}\right), \label{eq: EntropyMach} \end{align} where $M$ is the Mach number, given by, \begin{equation} \frac{\Delta p^*}{\gamma p^*_1} = \frac{p^*_2 - p^*_1}{\gamma p^*_1} = \frac{2}{\gamma + 1}\left(M^2-1\right), \label{eq: PressureMach} \end{equation} and $\Delta p^* = p^*_2 - p^*_1$ is the pressure jump with $p^*_1$ and $p^*_2$ being the pre-shock and post-shock pressures, respectively.}} Near the inflection point of the fluid velocity profile, the entropy reaches a local maximum $(s^* = s^*_\mathrm{max})$. According to Morduchow and Libby~\cite{Morris_JASc_1949}, maximum entropy $s^*_{\mathrm{max}}$ assuming $\mu^*_B = 0$ and $Pr = 3/4$, can be obtained as, \begin{equation} \frac{s^*_{\mathrm{max}}}{R^*} = \frac{1}{\gamma - 1}\ln\left(1 + \frac{\gamma - 1}{2}M^2\left(1-\xi\right)\xi^{\frac{\gamma - 1}{2}}\right), \label{eq: EntropyMax} \end{equation} where, \begin{equation} \xi = \frac{\gamma - 1}{\gamma + 1} + \frac{2}{\gamma + 1}\frac{1}{M^2}. \end{equation} For weak shock waves, ($\Delta p^*/\gamma p^*_1 <1$), entropy jump $\Delta s^*$ and maximum entropy $s^*_{\mathrm{max}}$ scale with pressure jumps as (cf. Fig.~\ref{fig: EntropyScaling}), \begin{equation} {\Delta s^*} = \mathcal{O}\left(\Delta p^{*3}\right), \quad s^*_{\mathrm{max}} = \mathcal{O}\left({\Delta p}^{*2}\right), \end{equation} independent of $\mu^*$ (cf. Eqs.~\eqref{eq: EntropyMach} and~\eqref{eq: EntropyMax}). The overall entropy jump $\Delta s^*$ is due to irreversible thermoviscous losses occurring within the shocks. However, the overshoot in entropy ($s^*_{\mathrm{max}}>\Delta s^*$) is due to both reversible and irreversible processes, and is not in violation of the second law of thermodynamics~\cite{Morris_JASc_1949}. Moreover, in the range of pressure jumps considered in the DNS (see Section~\ref{sec: numerics}), the maximum Mach number of the shock is around $M\approx 1.4$, which is well within the limits of validity of the continuum approach~\cite{Bird_MolecularGasDynamicsBook_1994}. Hence, it is physically justified to draw conclusions regarding the smallest length scales through the governing equations based on continuum approach and assuming thermodynamic equilibrium. \subsection{Perturbation variables and non-dimensionalization} \label{sec: 2ndOrderEntropy} {In this section, we utilize the previous consideration on the second order scaling of the maximum entropy $s^*_{\mathrm{max}}$ inside a weak shock wave to derive second order nonlinear acoustics equations. To this end, we decompose the variables in base state and perturbation fields and derive equations containing only linear and quadratic terms in perturbation fields. Denoting the base state with the superscript $( )_0$ and the perturbation fields with the superscript $({ })'$, we obtain, } \begin{subequations} \begin{align} & \rho^* = \rho^{*}_0 + {\rho^*}',\quad p^* = p^*_0 + {p^*}', \\ & u^* = {u^*}', \quad s^* = {s^*}', \quad T^* = T^*_0 + {T^*}', \label{eq: PertDecom} \end{align} \end{subequations} where no mean flow $u^*_0 = 0$ is considered and $s^*_0$ is arbitrarily set to zero. We neglect the fluctuations in the dynamic viscosity as well, i.e., \begin{equation} \mu^* = \mu^*_0. \end{equation} {\color{black}{While in classic gas dynamics, pre-shock values are used to normalize fluctuations or jumps across the shock (e.g. see Eq.~\eqref{eq: PressureMach}), hereafter we choose base state values to non-dimensionalize the nonlinear acoustics equations, \begin{subequations} \begin{align} & \rho = \frac{{\rho^*}}{\rho^*_0} = 1 + \rho ' ,~~p = \frac{p^*}{\gamma p^*_0} = \frac{1}{\gamma} + p', \\ & u = \frac{u^*}{a^*_0} = u', ~~s = \frac{s^*}{R^*} = s',~~ T = \frac{T^*}{T^*_0} = 1 + T',\\ &x = \frac{x^*}{L^*}, ~~ t = \frac{a^*_0 t^*}{L^*}. \end{align} \label{eq: per_norm} \end{subequations} where $L^*$ is the length of the one-dimensional periodic domain. \revision{As also typically done in classical studies of homogeneous isotropic turbulence~\cite{Ishihara_ARFM_2009,Batchelor1953theory,pope2000turbulent,monin1971statistical,Hosokawa_PRE_2008, Burattini_PRE_2006}, periodic boundary conditions represent a common (yet not ideal) way to approximate infinite domains; as such, a spurious interaction between the flow physics that one wishes to isolate and the periodic box size may occur. For the TW and SW test cases analyzed herein, $L^*$ corresponds to the initial (and hence largest) reference length scale of the acoustic perturbation; in the AWT case, the value of $L^*$ should be chosen as much larger than the integral length scale $\ell$ or Taylor microscale $\lambda$ (see Section~\ref{sec: ScalesAcousticCascade}), which truly define the state of turbulence.}} {Due to thermodynamic nonlinearities, wave propagation velocity increases across a high-amplitude compression front, resulting in wave-steepening ~\citep{Whitham2011} and hence generation of small length scales associated with increasing temperature and velocity gradients responsible for thermoviscous dissipation. Increase in thermoviscous dissipation results in positive entropy perturbations peaking within the shock structure. For pressure jumps $\Delta p^*/\gamma p^*_1<1$, the maximum entropy scales approximately as $\mathcal{O}\left(\Delta p^{*2}\right)$ (cf. Fig.~\ref{fig: EntropyScaling}). Moreover, as we discuss in the later section (see Section~\ref{sec: EnergyCorollary}), the second order nonlinear acoustic equations impose a strict limit of $|p'|<1/\gamma$ ($\simeq 0.714$ for $\gamma = 0.72$) for base state normalized (Eq.~\eqref{eq: per_norm}) (not pre-shock state normalized (Eq.~\eqref{eq: PressureMach})) perturbations. Hence, in our simulations~(see Section~\ref{sec: numerics}), we consider a suitable range of $10^{-3}<p'<10^{-1}$, which satisfies the aforementioned constraints. Thus, the second order scaling of entropy holds in our simulations. Below, we utilize this entropy scaling to derive the correct second order nonlinear acoustics equations governing the spatio-temporal evolution of dimensionless perturbation variables $p'$ and $u'$, as defined in Eq.~\eqref{eq: per_norm}.} \subsection{Second order nonlinear acoustics equations} \label{sec: 2ndOrderNLA} For a thermally perfect gas, the differential in dimensionless density $\rho$ can be related to differentials in dimensionless pressure $p$ and dimensionless entropy $s$ as, \begin{align} d\rho &= \left(\frac{\partial \rho}{\partial p}\right)_{s} dp + \left(\frac{\partial \rho}{\partial s}\right)_{p} ds,\nonumber\\ & = \frac{\rho}{\gamma p}dp - \frac{\rho(\gamma - 1)}{\gamma}ds. \label{eq: ConstitutiveEquation} \end{align} Nondimensionalizing the continuity Eq.~\eqref{eq: NS_eqns1} and substituting Eq.~\eqref{eq: ConstitutiveEquation}, we obtain, \begin{align} &\frac{\partial \rho}{\partial t} + \frac{\partial \rho u}{\partial x} = 0,\nonumber\\ &\frac{\partial p}{\partial t} + u\frac{\partial p}{\partial x} + \gamma p\frac{\partial u}{\partial x} = (\gamma -1)p\left(\frac{\partial s}{\partial t} + u\frac{\partial s}{\partial x}\right). \label{eq: PressureStep1} \end{align} Substituting the dimensionless forms of Eqs.~\eqref{eq: NS_eqns3} and~\eqref{eq: IdealGas} and utilizing the decomposition given in Eqs.~\eqref{eq: per_norm}, we obtain the following truncated equation for pressure perturbation $p'$, \begin{align} \frac{\partial p'}{\partial t} + \frac{\partial u'}{\partial x} + \gamma p' \frac{\partial u'}{\partial x} &+ u'\frac{\partial p'}{\partial x} = \nu_0\left(\frac{\gamma - 1}{Pr}\right)\frac{\partial^2 p'}{\partial x^2} \nonumber \\ & + \mathcal{O}\left(p's', s'^2, p'^3, \left(\frac{\partial u'}{\partial x}\right)^2\right).\label{eq: pressure} \end{align} Similarly, the truncated equation for velocity perturbation $u'$ is obtained as, \begin{align} \frac{\partial u'}{\partial t} + \frac{\partial p'}{\partial x} + \frac{\partial}{\partial x}\left(\frac{u'^2}{2} - \frac{p'^2}{2}\right) &= \frac{4}{3}\nu_0\frac{\partial^2 u'}{\partial x^2} \nonumber \\ &+\mathcal{O}\left(\rho'^2p',\rho'^3p' \right). \label{eq: velocity} \end{align} In Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity}, $\nu_0$ is the dimensionless kinematic viscosity given by, \begin{equation} \nu_0 = \frac{\mu^*_0}{\rho^*_0 a^*_0L^*}, \label{eq: DimensionlessVisc} \end{equation} and quantifies viscous dissipation of waves relative to propagation. Equations~\eqref{eq: pressure} and~\eqref{eq: velocity} constitute the nonlinear acoustics equations truncated up to second order, governing spatio-temporal evolution of finite amplitude acoustic perturbations $p'$ and $u'$. The entropy scaling ($s^*_{\mathrm{max}}=\mathcal{O}\left(\Delta p^{*2}\right)$) discussed previously results in the dissipation term on the right hand side of Eq.~\eqref{eq: pressure}. Left hand side of Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} contains terms denoting linear and nonlinear isentropic acoustic wave propagation. {Detailed derivation of Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} is given in Appendix~\ref{sec: appedixA}, where we also show that the nonlinear terms on the left hand side (LHS) of Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} are independent of the thermal equation of state. The functional form of the second order perturbation energy norm ($E^{(2)}$, Eq \eqref{eq: Energy_norm}) -- being exclusively dictated by such terms (see Section \ref{sec: EnergyCorollary}) -- is independent of the thermal equation of state of the gas. The results shown in this work focus on ideal-gas simulations merely for the sake of simplicity, with no loss of generality pertaining to inviscid nonlinear (up to second order) spectral energy transfer dynamics.} We note that, Eq.~\eqref{eq: pressure} consists of the velocity derivative term ($\gamma p'{\partial u'}/{\partial x}$), and is different from those obtained by Naugol'nykh and Rybak~\cite{Naugol1975spectrum}, which in dimensionless form read, \begin{align} &\frac{\partial p'}{\partial t} - (\gamma -1)p'\frac{\partial p'}{\partial t} + \frac{\partial u'}{\partial x} + p'\frac{\partial u'}{\partial x} = 0,\\ &\frac{\partial u'}{\partial t} + \frac{\partial p'}{\partial x} + \frac{\partial }{\partial x}\left(\frac{u^2}{2} - \frac{p^2}{2}\right) = 0. \end{align} We adopt Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} throughout the study since they represent the truncated governing equations exactly. Unlike Naugol'nykh and Rybak~\cite{Naugol1975spectrum}, we do not approximate the density $\rho$ using the Taylor series and only use the total differential form given in Eq.~\eqref{eq: ConstitutiveEquation}. Additionally, we note that Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} can be combined into the Westervelt's equation~\cite{Hamilton_NLA_1998} only if the Lagrangian defined as, \begin{equation} \mathcal{L} = \frac{u' ^2}{2} - \frac{p' ^2}{2}, \end{equation} is zero, which holds only for linear pure traveling waves. The derivation of the Burgers equation in nonlinear acoustics follows from the Westervelt's equation~\cite{Hamilton_NLA_1998}. Hence, it is inadequate in modeling general nonlinear acoustics phenomena involving mixed phasing of nonlinear waves which occurs in the Standing Wave (SW) and Acoustic Wave Turbulence (AWT) cases analysed here in.\section{Second order perturbation energy} \label{sec: EnergyCorollary} In this section we derive a new perturbation energy function for nonlinear acoustic waves utilizing Eqs.~\eqref{eq: pressure} and \eqref{eq: velocity}. To this end, we derive the perturbation energy conservation relation (energy corollary) for high amplitude acoustic perturbations. We show that the spatial average of the perturbation energy function satisfies the definition of the Lyapunov function for high amplitude acoustic perturbations and evolves monotonically in time (cf.~Fig.~\ref{fig: EnergyTime}). Utilizing the energy corollary, we derive spectral energy transport relations in further sections. Multiplying Eqs.~\eqref{eq: pressure} and \eqref{eq: velocity} with $p'$ and $u'$ respectively and adding, we obtain, \begin{align} &\frac{\partial }{\partial t}\left(\frac{p' ^2}{2} + \frac{u' ^2}{2}\right) + \frac{\partial }{\partial x}\left(u'p' + \frac{u'^3}{3}\right) + \gamma p' ^2\frac{\partial u'}{\partial x} = \nonumber \\ &\nu_0\left(\frac{\gamma - 1}{Pr}\right)p'\frac{\partial^2 p'}{\partial x^2} + \frac{4}{3}\nu_0 u'\frac{\partial^2 u'}{\partial x^2}. \quad\quad \label{eq: EnergyCorollary_Open} \end{align} Spatial averaging of Eq.~\eqref{eq: EnergyCorollary_Open} over a periodic domain $[0, L]$ yields, \begin{align} \frac{d\left\langle E^{(1)}\right\rangle}{dt} = -\left\langle \gamma p'^2\frac{\partial u'}{\partial x}\right\rangle &- \nu_0\left(\frac{\gamma - 1}{Pr}\right)\left\langle\left(\frac{\partial p'}{\partial x}\right)^2\right\rangle\nonumber\\ &-\frac{4}{3}\nu_0\left\langle\left(\frac{\partial u'}{\partial x}\right)^2\right\rangle, \label{eq: LinearAcousticEnergy} \end{align} where $\left\langle . \right\rangle$ is the spatial averaging operator, \begin{equation} \left\langle . \right\rangle = \frac{1}{L}\int^L_0\left(.\right)dx, \end{equation} and \begin{equation} E^{(1)} = \frac{u'^2}{2} + \frac{p'^2}{2}, \end{equation} is the first order isentropic acoustic energy. Equation~\eqref{eq: LinearAcousticEnergy} suggests that, in a lossless medium ($\nu_0\rightarrow 0$), $\left\langle E^{(1)}\right\rangle$ would exhibit spurious non-monotonic behavior in time due to the first term on right hand side. Such non-monotonic behaviour is confirmed by the DNS results shown in Fig.~\ref{fig: EnergyTime}. Consequently, the linear acoustic energy norm $E^{(1)}$ does not quantify the perturbation energy correctly for high amplitude perturbations since the spatial average $\left\langle E^{(1)}\right\rangle$ supports spurious growth and decay in the absence of physical sources of energy. \begin{figure}[!b] \includegraphics[width=\linewidth]{EnergyFunction.pdf} \caption{Comparison of perturbation energy function for nonlinear acoustic waves $E^{(2)}$ with the linear acoustic energy $E^{(1)}$ in the case of $p'=u'$ (assumed for illustrative purpose). The correction $f(p')$ is independent of $u'$.} \label{fig: EnergyComp} \end{figure} The corrected perturbation energy function can be obtained upon recursively evaluating the velocity derivative term ($\gamma p'^2\partial u'/\partial x$ in Eq.~\eqref{eq: EnergyCorollary_Open}) utilizing Eq.~\eqref{eq: pressure} as, {\color{black}{ \begin{align} \gamma p'^2\frac{\partial u'}{\partial x} &= -\frac{\partial }{\partial t}\left(\frac{\gamma p'^3}{3}\right) - \frac{\partial}{\partial x}\left(\frac{\gamma u' p'^3 }{3}\right) -\nonumber \\ &\gamma\left(\gamma - \frac{1}{3}\right)p'^3\frac{\partial u'}{\partial x} - \frac{\nu_0(\gamma - 1)}{Pr}\gamma p'^3\frac{\partial p'}{\partial x}. \label{eq: Rec1} \end{align} Furthermore, the third term in above Eq.~\eqref{eq: Rec1} on the right can be further evaluated as, \begin{align} &\gamma\left(\gamma - \frac{1}{3}\right)p'^3\frac{\partial u'}{\partial x} = -\frac{\partial }{\partial t}\left(\frac{\gamma}{4}\left(\gamma - \frac{1}{3}\right)p'^4\right) \nonumber \\ &- \frac{\partial}{\partial x}\left(\frac{\gamma}{4}\left(\gamma - \frac{1}{3}\right)u'p'^4\right) - \gamma\left(\gamma - \frac{1}{3}\right)\left(\gamma - \frac{1}{4}\right)p'^4\frac{\partial u'}{\partial x}\nonumber \\ &-\frac{\nu_0(\gamma -1 )}{Pr}\gamma\left(\gamma - \frac{1}{3}\right)p'^3\frac{\partial^2 p'}{\partial x^2}, \label{eq: Rec2} \end{align} and so on}}. Continued substitution according to Eqs.~\eqref{eq: Rec1} and~\eqref{eq: Rec2} yields the closure of the system and the following energy corollary, \begin{equation} \frac{\partial E^{(2)}}{\partial t} + \frac{\partial I}{\partial x}= \nu_0\left(\frac{\gamma - 1}{Pr}\right) h(p')\frac{\partial^2 p'}{\partial x^2} + \frac{4}{3}\nu_0 u'\frac{\partial^2 u'}{\partial x^2} , \label{eq: energy_cons} \end{equation} where, \begin{equation} I(p',u') = p' u' + \frac{u'^3}{3} + u' f(p'), \label{eq: Intensity} \end{equation} is the intensity (energy flux) of the field, $h(p')$ is given by, \begin{equation} h(p') = p' + \frac{\partial f(p')}{\partial p'} = \frac{\partial E^{(2)}}{\partial p'}. \end{equation} and $E^{(2)}$ is given by, \begin{equation} E^{(2)}(p',u') = \frac{u' ^2}{2} + \frac{p' ^2}{2} + f(p') = E^{(1)} + f(p'), \label{eq: Energy_norm} \end{equation} and defines the second order perturbation energy for high amplitude acoustic perturbations. The energy corollary Eq.~\eqref{eq: energy_cons} is mathematically exact for the governing Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity}. \begin{figure*}[!t] \centering \includegraphics[width=\textwidth]{EnergyLyapunov.pdf} \put(-510,220){$(a)$} \put(-330,220){$(b)$} \put(-160,220){$(c)$} \caption{Spatial profile of finite amplitude waves (top) for TW($a$), SW($b$), and AWT($c$). Evolution of the average perturbation energy (< $E^{(2)}$ > (--); < $E^{(1)}$ > ($--$)) evaluated from the DNS data (bottom) scaled by the initial value against scaled time $t/\tau$ (cf. Eq.~\eqref{eq: SteepeningTime}) for increasing values of perturbation amplitude $A_{\mathrm{rms}}$ defined in Eq.~\eqref{eq: PcDef} at $\nu_0 = $ 1.836$\times$10$^{-7}$ (see Table~\ref{tab: test_cases}). The curves are shifted vertically by 0.25 for illustrative purpose only. With increasing perturbation amplitude $A_{\mathrm{rms}}$, the variation of linear acoustic energy norm < $E^{(1)}$ > becomes increasingly non-monotonic. The vertical dashed line (bottom) highlights the end of approximately inviscid spectral energy cascade regime. In this regime, the energy is primarily redistributed in the spectral space due to the nonlinear propagation ($\epsilon\simeq 0$).} \label{fig: EnergyTime} \end{figure*} The correction term $f(p')$ in $E^{(2)}$ appears due the thermodynamic nonlinearities and can be derived in the closed form as, \begin{equation} f(p') = \sum^{\infty}_{n=2}T_n = \sum^{\infty}_{n=2}(-1)^{n+1}\frac{\gamma p'^{n+1}}{n+1}\prod^{n}_{i=3}\left(\gamma - \frac{1}{i}\right), \label{eq: EnergyCorrection_app} \end{equation} {where $T_2$ and $T_3$ can be identified in Eqs.~\eqref{eq: Rec1} and \eqref{eq: Rec2}, respectively.} Isolating the $n^{th}$ term of the above infinite series as, \begin{equation} T_n = (-1)^{n+1}\frac{\gamma p'^{n+1}}{n+1}\underbrace{\left(\gamma - \frac{1}{3}\right)\left(\gamma - \frac{1}{4}\right)\cdots\left(\gamma - \frac{1}{n}\right)}_{n-2~\mathrm{terms}}. \label{eq: nth-term} \end{equation} Multiplied fractions in the Eq.~\eqref{eq: nth-term} above yield the $n^{th}$ term as, \begin{equation} T_n = -\frac{2\gamma}{\left(\gamma-1\right)\left(2\gamma - 1\right)}\left(\gamma p'\right)^{n+1}\binom{1/\gamma}{n+1}. \end{equation} Finally, the energy correction $f(p')$ can be recast as, \begin{align} f(p') = \sum^{\infty}_{n=2}T_n = &-\frac{2\gamma}{\left(\gamma-1\right)\left(2\gamma - 1\right)}\Big(\left(1+\gamma p'\right)^{1/\gamma} - 1 - \nonumber \\ & p'+ \frac{\left(\gamma - 1\right)p'^2}{2}\Big). \label{eq: EnergyCorrection} \end{align} The correction function $f(p')$ defined in Eq.~\eqref{eq: EnergyCorrection} accounts for second order isentropic nonlinearities and is not a function of entropy perturbation. Hence, $E^{(2)}$ accounts for the effect of high amplitude perturbations on perturbation energy isentropically. We note that this separates $E^{(2)}$ fundamentally from generalized linear perturbation energy norms, such as the ones derived by Chu~\cite{Chu1965EnergyNorm} for small amplitude non-isentropic perturbations, and by Meyers~\cite{myers1986exact} for acoustic wave propagation in a steady flow. {Moreover, as discussed in the previous section (and shown in Appendix~\ref{sec: appedixA}), since the isentropic nonlinearities on the LHS of Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity} are independent of the thermal equation of state, the functional form of $E^{(2)}$ and $I$ are also independent of the equation of state. However, the dissipation term on the right hand side of the energy corollary Eq.~\eqref{eq: energy_cons} may change with the thermal equation of state.} The energy correction $f(p')$ {is infinite order in pressure perturbation $p'$} and converges only for perturbation magnitude $|p'|<1/\gamma$ thus naturally yielding the strict limit of validity of second order acoustic equations in modelling wave propagation and wave steepening. Figure~\ref{fig: EnergyComp} shows the newly derived second order perturbation energy $E^{(2)}$ compared against the isentropic acoustic energy $E^{(1)}$. Both $E^{(2)}$ and $E^{(1)}$ are non-negative in the range $|p'|<1/\gamma$ ($p'=u'$ is assumed for illustrative purpose). Furthermore, $E^{(2)}$ is asymmetric in nature, with larger energy in dilatations compared to compressions of same magnitude, as shown in Fig.~\ref{fig: EnergyComp}. Such asymmetry signifies that the medium (\revision{compressible ideal gas} in the present study) relaxes towards the base state faster for finite dilatations compared to compressions. For compact supported or spatially periodic perturbations, the energy conservation Eq.~\eqref{eq: energy_cons} shows that the spatially averaged energy $\left\langle E^{(2)}\right\rangle$ decays monotonically in time (in the absence of energy sources) accounting for the nonlinear interactions i.e., \begin{align} \dot{\mathscr{V}}=\frac{d \left\langle E^{(2)}\right\rangle}{dt} &= -\nu_0\left(\frac{\gamma - 1}{{Pr}}\right)\left\langle \frac{\partial^2 E^{(2)}}{\partial p'^2}\left(\frac{\partial p'}{\partial x}\right)^2\right\rangle \nonumber \\ &- \frac{4}{3}\nu_0\left\langle\left(\frac{\partial u'}{\partial x}\right)^2\right\rangle \nonumber \\ &= \left\langle \mathcal{D}\right\rangle= -\epsilon\leq 0, \label{eq: LyapunovFunction} \end{align} where $\mathcal{D}$ is the perturbation energy dissipation and $\epsilon$ is the negative of its spatial average. The spatial average $\left\langle E^{(2)}\right\rangle$ is non-negative ($E^{(2)}\geq 0$), and Eq.~\eqref{eq: LyapunovFunction} and Fig.~\ref{fig: EnergyTime} confirm that $\left\langle E^{(2)}\right\rangle$ evolves monotonically in time in the absence of physical energy sources. Hence, the spatial average of the perturbation energy function $\left\langle E^{(2)}\right\rangle$ defines the Lyapunov function $\mathscr{V}$ of the nonlinear acoustic system governed by the set of second order governing Eqs.~\eqref{eq: pressure} and \eqref{eq: velocity} exactly. The spatial average $\left\langle E^{(2)}\right\rangle$ should be used for studying the stability of nonlinear acoustic systems~\cite{George2012, Strogatz2018NonlinearDynamics}, which, however, falls beyond the scope of this work. Wave-front steepening entails cascade of perturbation energy into higher wavenumbers thus broadening the energy spectrum. Fully broadened spectrum of acoustic perturbations exhibits energy at very small length scales which causes high thermoviscous energy dissipation. We analyse the separation of length scales and energy decay caused by nonlinear wave steepening and thermoviscous energy dissipation in the following sections. To this end, we utilize the direct numerical integration of Navier-Stokes Eqs.~\eqref{eq: NS_eqns1}-\eqref{eq: IdealGas} resolving all the length scales (DNS) and the exact energy corollary~Eq.~\eqref{eq: energy_cons} for second order truncated Eqs.~\eqref{eq: pressure} and~\eqref{eq: velocity}. \section{High Fidelity Simulations with Adaptive Mesh Refinement} \label{sec: numerics} We perform shock-resolved numerical simulations of 1D Navier-Stokes (DNS) Eqs.~\eqref{eq: NS_eqns1}-\eqref{eq: IdealGas} with Adaptive Mesh Refinement (AMR). We use the perturbation energy $E^{(2)}$ defined in Eq.~\eqref{eq: Energy_norm} to define the characteristic dimensionless perturbation amplitude $A_{\mathrm{rms}}$ as, \begin{equation} A_{\mathrm{rms}} = \sqrt{\left\langle E^{(2)}\right\rangle}, \label{eq: PcDef} \end{equation} which is varied in the range $10^{-3}-10^{-1}$. The dimensionless kinematic viscosity at base state $\nu_0$ is also varied from 1.836$\times$10$^{-5}$ to 1.836$\times$10$^{-7}$. The base state conditions in the numerical simulations correspond to STP, i.e. $p^*_0 = 101325~\mathrm{Pa}$ and $T^*_0 = 300~\mathrm{K}$. \begin{table}[!t] \def1.2{2} \begin{ruledtabular} \begin{tabular}{c c c c} $\nu_0$ & 1.836$\times$10$^{-5}$ & 1.836$\times$10$^{-6}$ & 1.836$\times$10$^{-7}$\\ \hline $A_{\mathrm{rms},0}$ & $10^{-3}$ & $10^{-2}$ & $10^{-1}$ \\ \hline \hline $u^*_{\mathrm{rms}}~\left(\mathrm{m}/\mathrm{s}\right)$ & 0.347 & 3.472& 34.725\\ \hline $p^*_{\mathrm{rms}}~\left(\mathrm{kPa}\right)$ & 0.142 & 1.419 & 14.185\\ \end{tabular} \end{ruledtabular} \caption{Simulation parameter space for TW, SW, and AWT cases listing base state dimensionless viscosity $\nu_0$ (cf. Eq.~\eqref{eq: DimensionlessVisc}), initial characteristic perturbation amplitude $A_{\mathrm{rms},0}$ (cf. Eq.~\eqref{eq: PcDef}), and dimensional characteristic perturbation in velocity $u^*_{\mathrm{rms}}$ and pressure $p^*_{\mathrm{rms}}$ fields (Eq.~\eqref{eq: AcousticReynolds}).} \label{tab: test_cases} \end{table} The goal of spanning $A_{\mathrm{rms}}$ and $\nu_0$ over three orders of magnitude is to achieve widest possible range of energy cascade rate and dissipation within computationally feasible times. Equation~\eqref{eq: PcDef} yields the definitions of the perturbation Reynolds number~$\mathrm{Re}_L$ , characteristic perturbation velocity field~$u^*_{\mathrm{rms}}$, and pressure field $p^*_{\mathrm{rms}}$ as, \begin{equation} \mathrm{Re}_L = \frac{A_{\mathrm{rms}}a^*_0 L^*}{\nu_0^*},~\quad u^*_{\mathrm{rms}} = a^*_0A_{\mathrm{rms}},~\quad p^*_{\mathrm{rms}} = \rho^*_0{a_0^*}^2A_{\mathrm{rms}}, \label{eq: AcousticReynolds} \end{equation} where $\mathrm{Re}_L$ denotes ratio of diffusive to wave steepening time scale over the length $L$. {In the simulations, we keep $\mathrm{Re}_L\gg 1$, which corresponds to very fast wave steepening rates compared to diffusion.} In further sections (see Section~\ref{sec: ScalesAcousticCascade}), we define the wave turbulence Reynolds number $\mathrm{Re}_{\ell}$ based on the integral length scale $\ell$. Below, we briefly discuss the numerical scheme utilized for shock-resolved simulations and outline the initialization of the three configurations (TW, SW, and AWT) for numerical simulations. \subsection{Numerical approach} \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{AMR_tree.pdf} \caption{Illustration of the binary tree implementation of Adaptive Mesh Refinement (AMR) technique (top left). The mesh is refined based on the resolution error in pressure field in each cell acting as a node of a binary tree. The pressure field shown (middle) corresponds to the randomly initialized AWT case with $A_{\mathrm{rms}}=10^{-1}$, $\nu_0 = 1.836\times$10$^{-6}$ (Table~\ref{tab: test_cases}) at $t/\tau = 0.04$. The inset shows the resolved shock wave with $(+)$ denoting the cell interfaces. The mesh refinement levels (bottom) show the depth $d$ of the binary tree.} \label{fig: AMR} \end{figure} We integrate the fully compressible 1D Navier-Stokes Eqs.~\eqref{eq: NS_eqns1}-\eqref{eq: NS_eqns3} in time utilizing the staggered spectral difference (SD) spatial discretization approach~\cite{kopriva1996conservative}. In the SD approach, the domain is discretized into cells. Within each cell, the orthogonal polynomial reconstruction of variables allows numerical differentiation with spectral accuracy. We refer the reader to the work by Kopriva and Kolias~\cite{kopriva1996conservative} for further details. To accurately resolve spectral energy dynamics at all length scales, i.e. for resolved weak shock waves, we combine the SD approach with the adaptive mesh refinement (AMR) approach as first introduced by Mavriplis~\cite{mavriplis1994adaptive} for spectral methods. The SD-AMR approach eliminates the computational need of very fine grid everywhere for resolving the propagating shock waves. To this end, we expand the values of a generic variable $\phi$ local to the cell in the Legendre polynomial space as, \begin{equation} \phi = \sum^{N}_{i=1}\hat{\phi}_i \psi_i(x) \end{equation} where $\psi_i(x)$ is the Legendre polynomial of $(i-1)^{\mathrm{th}}$ degree. The polynomial coefficients $\hat{\phi}_i$ are utilized for estimating the local resolution error $\varepsilon$ defined as~\cite{mavriplis1994adaptive}, \begin{equation} \varepsilon = \left(\frac{2 \hat{\phi}^2_N}{2N + 1} + \int^{\infty}_{N+1}\frac{2 f_{\varepsilon}^2(n)}{2n + 1} dn\right)^{1/2}, \quad f_{\varepsilon}(n) = ce^{-\sigma n}, \end{equation} where $f_{\varepsilon}$ is the exponential fit through the coefficients of the last four modes in the Legendre polynomial space. As the estimated resolution error $\varepsilon$ exceeds a pre-defined tolerance, the cell divides into two subcells, which are connected utilizing a binary tree (shown in Fig.~\ref{fig: AMR}). The subcells merge together if the resolution error decreases below a pre-defined limit. \subsection{Initial conditions} \begin{table}[!b] \setlength{\tabcolsep}{0.95em} \def1.2{1.2} \begin{ruledtabular} \begin{tabular}{c c c c} {$~~$} & TW & SW & {AWT}\\ \hline $k_0$ & 1 & 1 & 1 \\ $k_E$ & 1 & 1 & 100 \\ $b_0(k)$ & 0 & 0 & $e^{-(|k|-k_E)^2}$ \\ {$\widehat{E}_k$} & {${A}_{\mathrm{rms}}^2\delta(k_0)$} & {${A}_{\mathrm{rms}}^2\delta(k_0)$} & ${A}_{\mathrm{rms}}^2$\\ \end{tabular}\quad \end{ruledtabular} \caption{Initial spectral compositions for traveling wave (TW), standing wave (SW), and acoustic wave turbulence (AWT). $\delta\left(\cdot\right)$ is the Dirac delta function.} \label{tab: InitialSpectra} \end{table} We utilize the Riemann invariants for compressible Euler equations to initialize the propagating traveling and standing wave cases in the numerical simulations. The Riemann invariants in terms of perturbation variables assuming nonlinear isentropic changes are given by, \begin{align} R_{-} = \frac{2}{\gamma -1 }\left(\left(1 + {\rho '}\right)^{\frac{\gamma - 1}{2}} - 1\right) - u',\label{eq: InvariantL} \\ R_{+} = \frac{2}{\gamma -1 }\left(\left(1 + {\rho '}\right)^{\frac{\gamma - 1}{2}} - 1\right) + u', \label{eq: InvariantR} \end{align} where $R_{-}$ and $R_{+}$ are the left and right propagating invariants, respectively, and $u'$ and ${\rho '}$ are normalized velocity and density perturbations, as defined in Eq.~\eqref{eq: per_norm}. Initial conditions for TW and SW cases correspond to $R_{-} = 0$ and $R_{-} = R_{+}$ respectively. To initialize the broadband noise case, we first choose $p'$ and $u'$ pseudo-randomly from a uniform distribution for the whole set of discretization points in $x$. Low-pass filtering of $p'$ and $u'$ yields, \begin{align} \widetilde{\widehat{p}_k}(t&=0) = \widehat{p}_kb_0(k),\quad \widetilde{\widehat{u}_k}(t=0) = \widehat{u}_kb_0(k)\nonumber\\ b_0(k) &= \begin{cases} 1 & k_0\leq|k|\leq k_E \\ e^{-(|k|-k_E)^2} & |k|> k_E \end{cases}. \label{eq: FilterPU1} \end{align} where ${\widehat{p}_k}$ and ${\widehat{p}_k}$ are the Fourier coefficients of pseudo-random fields $p'$ and $u'$, respectively. $\widetilde{\widehat{p}_k}$ and $\widetilde{\widehat{u}_k}$ are the low-pass filtered coefficients. Inverse Fourier transform of Eq.~\eqref{eq: FilterPU1} yields smooth initial conditions with the initial spectral energy ${\widehat{E}_k}$, as defined in Section~\ref{sec: ScalesAcousticCascade} (cf. Eq.~\eqref{eq: spectral_energy}). For TW and SW, only single harmonic ($k=1$ in the current work) contains all of the initial energy. However, for AWT, $\widehat{E}_k$ is governed by the correlation function of velocity and pressure fields. In Table~\ref{tab: InitialSpectra}, we summarize the initial spectral energy for all three cases based on Eq.~\eqref{eq: FilterPU1}.\section{Scales of acoustic energy cascade and dissipation} \label{sec: ScalesAcousticCascade} In this section, we derive the analytical expressions of spectral energy, energy cascade flux, and spectral energy dissipation utilizing the exact energy corollary Eq.~\eqref{eq: energy_cons} (see Section~\ref{sec: EnergyCorollary}). We then identify the integral length scale $\ell$, the Taylor microscale $\lambda$, and the Kolmogorov length scale $\eta$ for TW, SW, and AWT cases in a periodic domain utilizing the DNS data (see Fig.~\ref{fig: LengthScalesSummary} and Table~\ref{tab: LengthScalesTable}). Temporal evolution laws of these length scales yield energy decay laws, which are used for dimensionless spectral scaling relations (see Section~\ref{sec: SpectralScaling}). \subsection{Spectral energy flux and dissipation rate for periodic perturbations} The exact perturbation energy conservation equation is given by (cf. Eq.~\eqref{eq: energy_cons}), \begin{equation} \frac{\partial E^{(2)}}{\partial t} + \frac{\partial I}{\partial x} = \mathcal{D} \end{equation} Integrating over the periodic domain, the above energy corollary can be converted into the following statement of conservation of perturbation energy in the spectral space, \begin{equation} \frac{d}{d t} \sum_{|k'|\leq k} \widehat{E}_{k'} + \widehat{\Pi}_k = \sum_{|k'|\leq k}\widehat{\mathcal{D}}_{k'}, \label{eq: spectral_conservation} \end{equation} \begin{table}[!t] \setlength{\tabcolsep}{0.2em} \def1.2{1.2} \begin{ruledtabular} \begin{tabular}{c c c c} \multirow{2}{*}{Length scale} & Integral & Taylor & Kolmogorov\\ & length scale & Microscale & length scale\\ &$\ell$ & $\lambda$ & $\eta$\\ \hline \\ {Definition} & $\sqrt{\frac{\sum_k {\widehat{E}_k}/{k^2}}{\sum_k \widehat{E}_k}}$ & $\sqrt{\frac{2\delta\left\langle E^{(2)}\right\rangle }{\epsilon}} $ & $\frac{\delta}{\sqrt{\left\langle E^{(2)}\right\rangle}}$ \\ Characteristic & \multirow{2}{*}{$(k_0,k_E)$} & \multirow{2}{*}{$(k_E, k_{\delta})$} & \multirow{2}{*}{$(k_\delta, \infty)$} \\ spectral range & & & \\ \end{tabular}\quad \end{ruledtabular} \caption{ Summary of the three length scales $\ell$, $\lambda$, and $\eta$, respective definitions, and the range of spectrum characterized by them. The integral length scale characterizes the energy containing range $(k_0, k_E)$. The Taylor microscale is the characteristic of the energy transfer and dissipation range $(k_E, k_{\delta})$. The Kolmogorov length scale corresponds to the highest wavenumber generated as a result of nonlinear acoustic energy cascade.} \label{tab: LengthScalesTable} \end{table} \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{SpectraLengthScaleSummary.pdf} \put(-245,410){$(a)$} \put(-245,295){$(b)$} \put(-245,195){$(c)$} \put(-245,110){$(d)$} \caption{ {Schematic illustrating the global picture of various length scales associated with spectral energy cascade in nonlinear acoustics in both spatial $(a)$ and spectral $(b)-(d)$ space. $(a)$ shows the perturbation velocity $u'$ ($--$) and pressure $p'$ (--) fields in AWT obtained from the DNS data for $\nu_0 = $1.836$\times$10$^{-7}$ and $A_{\mathrm{rms}}=10^{-1}$. $(b)$ shows the corresponding spectral energy $\widehat{E}_k$ in log-log space. The spectral flux $\widehat{\Pi}_k$ and dissipation $\widehat{\mathcal{D}}_k$ are shown in $(c)$ and $(d)$ respectively. The integral length scale $\ell$ corresponds to the characteristic distance between the shock waves traveling in the same direction. The Kolmogorov length scale $\eta$ corresponds to the shock wave thickness. The Taylor microscale $\lambda$ is the diffusive length scale and satisfies $\ell\gg\lambda\gg\eta$. $L$ corresponds to the length of the domain.}} \label{fig: LengthScalesSummary} \end{figure} \begin{figure*}[!t] \centering \includegraphics[width=\textwidth]{EnergyFluxALL.pdf} \put(-510,210){$(a)$} \put(-330,210){$(b)$} \put(-160,210){$(c)$} \caption{Spectro-temporal evolution of $\widehat{E}_k$ (top) and spectral flux $\widehat{\Pi}_k$ (bottom) for a TW ($a$), SW ($b$) and AWT ($c$). Spectral flux $\widehat{\Pi}_k$ for a traveling wave simply increases towards high wavenumbers. For a standing wave, $\widehat{\Pi}_k$ oscillates at low wavenumbers cyclically due to collisions of oppositely traveling shock waves while high wavenumber behaviour resembles that of a traveling shock. For AWT, the spectral broadening occurs for $k>k_E$ {with small fluctuations in time for $k<k_E$}.} \label{fig: EnergyFLux} \end{figure*} where the first term corresponds to the temporal rate of change of cumulative spectral energy density, \begin{equation} \frac{d \widehat{E}_k}{dt} \approx \frac{d}{dt}\left(\frac{|\widehat{u}_k|^2}{2} + \frac{|\widehat{p}_k|^2}{2}\right) + \Re\left(\widehat{p}_{-k}\frac{d\widehat{g}_k}{dt}\right), \label{eq: SpectralEnergyRate} \end{equation} and $\widehat{g}$ is the Fourier transform of $g(p')$ given by, \begin{equation} g(p') = \frac{\gamma}{\gamma-1}\left(\left(1+\gamma p'\right)^{1/\gamma} - 1 -p' \right). \end{equation} The spectral energy $\widehat{E}_k$ is given by, \begin{equation} \widehat{E}_k = \frac{|\widehat{u}_k|^2}{2} + \frac{|\widehat{p}_k|^2}{2} + \Re \left(\widehat{p}_{-k}\left(\widehat{\frac{f(p')}{p'}}\right)_k\right). \label{eq: spectral_energy} \end{equation} It is noteworthy that the correction in spectral energy does not follow directly from the nonlinear correction function $f(p')$ derived in the physical space. In Eq.~\eqref{eq: SpectralEnergyRate}, we have made the following approximation, \begin{equation} \frac{d}{dt}\left( \Re \left(\widehat{p}_{-k}\left(\widehat{\frac{f(p')}{p'}}\right)_k\right)\right) \approx \Re\left(\widehat{p}_{-k}\frac{d\widehat{g}_k}{dt}\right) \end{equation} Second term $\widehat{\Pi}_k$ in Eq.~\eqref{eq: spectral_conservation} is the flux of spectral energy density from wavenumbers $|k'|\leq k$ to $|k'| > k$ and is given by, \begin{align} \widehat{\Pi}_k &= \sum_{|k'|\leq k}\Re\Big(\widehat{p}_{-k'}\left(\frac{\partial (\widehat{u' g})}{\partial x}\right)_{k'} + \widehat{p}_{-k'}\left(\widehat{u'\frac{\partial p'}{\partial x}}\right)_{k'} +\nonumber \\ &\frac{1}{2}\widehat{u}_{-k'}\reallywidehat{\frac{\partial}{\partial x}\left(u' ^2 - p' ^2\right)_{k'}}\Big). \label{eq: SpectralFlux} \end{align} Finally, the spectral dissipation $\widehat{\mathcal{D}}_k$ is given by, {\color{black}{ \begin{align} \widehat{\mathcal{D}}_k = \nu_0\frac{\gamma - 1}{Pr}&\Re\left(\widehat{p}_{-k}\left(\reallywidehat{\left(1 + \frac{\partial g}{\partial p'}\right)\left(\frac{\partial^2 p'}{\partial x^2}\right)}\right)_{k}\right) \nonumber \\ &- {\frac{16\pi^2}{3}\nu_0} k^2|\widehat{u}_k|^2. \label{eq: SpectralDissipation} \end{align} }} Detailed derivation of Eqs.~\eqref{eq: spectral_conservation}-\eqref{eq: SpectralDissipation} is given in appendix~\ref{sec: appedixB}. Figure~\ref{fig: LengthScalesSummary} summarizes the typical shape of the spectral energy $\widehat{E}_k$, spectral energy flux $\widehat{\Pi}_k$, and the spectral dissipation $\widehat{\mathcal{D}}_k$ along with the relative positions of the three relevant length scales, the integral length scale, $\ell$, the Taylor microscale, $\lambda$, and the Kolomogorov length scale $\eta$ in the spectral space. The spectro-temporal evolution of any configuration of nonlinear acoustic waves can be quantified utilizing these length scales and the respective evolution in time which is discussed in detail in the subsections below. Table~\ref{tab: LengthScalesTable} summarizes these length scales and the characteristic spectral range. {In further sections, we discuss all the spectral quantities as functions of absolute value of wavenumbers and drop the $|.|$ notation for convenience.} The spectral energy flux $\widehat{\Pi}_k$, defined in Eq.~\eqref{eq: SpectralFlux}, is in terms of interactions of the Fourier coefficients of the pressure $\widehat{p}_k$ and velocity $\widehat{u}_k$ perturbations. For compact support or periodic perturbations, $\widehat{\Pi}_k$ approaches zero in the limit of very large wavenumbers $k\rightarrow \infty$, \begin{equation} \lim_{k\rightarrow\infty} \widehat{\Pi}_{k} = \left\langle \frac{\partial I}{\partial x}\right\rangle = 0. \end{equation} The last two terms in Eq.~\eqref{eq: SpectralFlux} result in $\widehat{\Pi}_{k}\rightarrow 0$ for large $k$ for general acoustic phasing. Hence, they are most relevant in SW and AWT cases. In a pure traveling wave (TW), $u' = p'$ at first order due to which the last two terms in Eq.~\eqref{eq: SpectralFlux} become negligible. Furthermore, the sequence of $\widehat{\Pi}_k$ also converges monotonically i.e., \begin{equation} \lim_{k\rightarrow\infty}\left(\widehat{\Pi}_{k-1} - \widehat{\Pi}_{k}\right) \rightarrow 0^{+}, \label{eq: MonotonicConv} \end{equation} as shown in Figs.~\ref{fig: LengthScalesSummary}$(c)$ and~\ref{fig: EnergyFLux}. The flattening of the spectral energy flux $\widehat{\Pi}_k$ (Eq.~\eqref{eq: MonotonicConv}) begins at a specific wavenumber $k_\delta$ associated to the {Kolmogorov length scale}, $\eta$, as shown in Fig.~\ref{fig: LengthScalesSummary}. The spectral energy $\widehat{E}_k$ deviates off the $k^{-2}$ decay near the wavenumber $k_\delta$. Figure~\ref{fig: EnergyFLux} shows the spectro-temporal evolution of the spectral energy $\widehat{E}_k$ and the flux $\widehat{\Pi}_k$ for TW, SW, and AWT prior to formation of shock waves. For TW, $\widehat{\Pi}_k$ increases in time due to spectral broadening. In SW, $\widehat{\Pi}_k$, while increasing, also oscillates at low wavenumbers due to the periodic collisions of oppositely propagating shocks. A combination of these processes takes place in a randomly initialized smooth finite amplitude perturbation, which at later times develops into AWT. At later times, nonlinear waves in all three configurations fully develop in to shock waves. Up to the shock formation, the spectral dynamics of all configurations simply involve increase of the spectral flux $\widehat{\Pi}_k$. The dimensionless shock formation time $\tau$ can be estimated as, \begin{equation} \tau = \frac{2}{(\gamma - 1)A_{\mathrm{rms},0}}. \label{eq: SteepeningTime} \end{equation} Upon shock formation, the dynamic evolution of TW and SW remains phenomenologically identical. The isolated shocks propagate and the total perturbation energy of the system decays due to thermoviscous dissipation localized around the shock wave. {\color{black}{However, for AWT, along with collisions of oppositely propagating shocks, those propagating in the same direction coalesce due to differential propagating speeds. As we discuss below, this modifies the energy decay and spectral energy dynamics in AWT significantly compared to TW and SW.}} In the sub-sections below, we elucidate the energy dynamics before and after shock formation for TW, SW, and AWT. To this end, we define and discuss the relevant length scales as mentioned above, namely: the Taylor microscale $\lambda$, the integral length scale $\ell$, and the Kolmogorov length scale $\eta$. Particular focus is given to the AWT case due to modified dynamics caused by shock coalescence. \begin{figure*}[!t] \centering \includegraphics[width=\textwidth]{EnergyDecayALL.pdf} \put(-510,290){$(a)$} \put(-330,290){$(b)$} \put(-160,290){$(c)$} \caption{Temporal evolution of scaled total energy <$E^{(2)}$> ${A^{-2}_{\mathrm{rms},0}}$ (--) (top), dissipation rate $\epsilon {A^{-3}_{\mathrm{rms},0}}$ ($--$) (mid) and normalized Taylor microscale $\lambda/\sqrt{\delta \tau}$ (bottom) for TW ($a$), SW ($b$) and AWT ($c$) against the scaled time $t/\tau$ for varying perturbation Reynolds number $\mathrm{Re}_L$. The time $t_0$ signifies fully broadened spectrum of the perturbation field.} \label{fig: EnergyDecayAll} \end{figure*} \subsection{Taylor microscale} In hydrodynamic turbulence, the Taylor microscale $\lambda$ separates the inviscid length scales from the viscous length scales~\cite{tennekes1972first,pope2000turbulent}. Due to the spectral energy cascade in planar nonlinear acoustics, we note that the spectral energy varies as $\widehat{E}_k\sim k^{-2}$ due to the formation of shocks and the spectral dissipation due to thermoviscous diffusion varies as $\widehat{\mathcal{D}}_k\sim k^2 \widehat{E}_k$. Consequently, the dissipation acts over most of the length scales with $k>k_E$ (Fig.~\ref{fig: LengthScalesSummary}$(d)$), unlike hydrodynamic turbulence where the viscous dissipation dominates only the smaller length scales~\cite{tennekes1972first, pope2000turbulent}. As shown in Fig.~\ref{fig: LengthScalesSummary}$(c)$, length scales in the range $(k_E, k_{\delta})$ exhibit both dissipation $\widehat{\mathcal{D}}_k$ and energy transfer $\widehat{\Pi}_k$. For $k>k_{\delta}$, $\widehat{\Pi}_k$ begins to converge monotonically to 0 and the interval $(k_{\delta}, 1/\eta)$ primarily exhibits dissipation $\widehat{\mathcal{D}}_k$ only. The Taylor microscale $\lambda$ quantifies the length scale associated to the whole dissipation range. Utilizing the definition of the total perturbation energy $\left\langle E^{(2)} \right\rangle$ and the dissipation rate $\epsilon$ (cf. Eq.~\eqref{eq: LyapunovFunction}), the microscale $\lambda$ can be defined as, \begin{equation} \lambda(t) = \sqrt{\frac{2\delta\left\langle E^{(2)}\right\rangle }{\epsilon}}, \label{eq: Microscale} \end{equation} where $\delta$ is the thermoviscous diffusivity, given by, {\color{black}{ \begin{equation} \delta = \nu_0\left(\frac{4}{3} + \frac{\gamma - 1}{\mathrm{Pr}}\right). \label{eq: EffectiveDiff} \end{equation} }} Equation~\eqref{eq: Microscale} indicates that the Taylor microscale can be identified as the geometrical centroid of full energy spectrum, i.e. \begin{equation} \lambda \sim \sqrt{\frac{\sum_k \widehat{E}_k}{\sum_k k^2 \widehat{E}_k}}. \label{eq: lengthscales_centroids} \end{equation} As the smaller length scales (higher harmonics) are generated, the dissipation rate $\epsilon$ tends to increase reaching a maximum in time. The increase of dissipation rate $\epsilon$ implies decrease of the length scale $\lambda$ in time. Minima of $\lambda$ indicates the fully-broadened spectrum of energy limited by the thermoviscous diffusivity at very large scales. Further spatio-temporal evolution of the system is dominated by dissipation thus indicating the purely diffusive nature of the Taylor microscale, i.e., \begin{equation} \lambda\rightarrow C\sqrt{\delta t}. \label{eq: TaylorMicroscaleAssy} \end{equation} The temporal evolution of $\lambda$ is qualitatively similar for TW, SW, and AWT, the constant $C$ in Eq.~\eqref{eq: TaylorMicroscaleAssy} differs for TW and SW compared with AWT due to the different spatial structure of perturbations. The time $t_0$ at which $\lambda$ reaches minimum signifies {fully developed nonlinear acoustic waves}. In case of AWT, it signifies fully developed acoustic wave turbulence. Figure~\ref{fig: EnergyDecayAll} shows the decay of scaled total perturbation energy $\left\langle E^{(2)}\right\rangle A_{\mathrm{rms},0}^{-2}$ and total dissipation rate $\epsilon A_{\mathrm{rms},0}^{-3}$ for the TW, SW, and AWT. We note that the total energy decays as a power law $t^{-2}$ for both TW and SW, whereas, for AWT, the initial decay law is $t^{-2/3}$. Asymptotic evolution (at large $t$) of the Taylor microscale follows from the decay laws as $\lambda = \sqrt{\delta t}$ and $\lambda = \sqrt{3\delta t}$ respectively. Since energy decay law of a single harmonic traveling and standing waves is rather trivial, we focus primarily on the AWT case for further discussion. \subsection{Integral length scale} \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{IntegralLengthScale.pdf} \put(-250,250){$(a)$} \put(-250,125){$(b)$} \caption{Evolution of the integral length scale $\ell$ $(a)$ and the Reynolds number $\mathrm{Re}_{\ell}$ ($b$) defined in Eqs.~\eqref{eq: IntegralLengthScale} and~\eqref{eq: IntegralReynoldsNumber}, respectively, for all the cases of AWT considered. For small thermoviscous diffusivity, $\ell$ increases approximately as $t^{1/3}$ before saturating to the dimensionless domain length $L=1$ and $\mathrm{Re}_{\ell}$ remains approximately constant.} \label{fig: IntegralLengthScaleEvolution} \end{figure} We identify the integral length scale $\ell$ as the characteristic length scale of the energy containing scales. In general, random {smooth} broadband noise (AWT) develops into an ensemble of shocks, propagating left and right in a one-dimensional system. For an ensemble of shock waves distributed spatially along a line, $\ell$ corresponds to the characteristic distance between consecutive shock waves traveling in the same direction, as shown schematically in Fig.~\ref{fig: LengthScalesSummary}. Formally, we define $\ell$ as, \begin{equation} \ell = \sqrt{\frac{\sum_k \frac{\widehat{E}_k}{k^2}}{\sum_k \widehat{E}_k}}, \label{eq: IntegralLengthScale} \end{equation} which is identical to the integral length scale defined in Burgers turbulence~\cite{Gurbatov_JFM_1997}. {\color{black}{Definition in Eq.~\eqref{eq: IntegralLengthScale} yields the centroid wavenumber of the initial energy spectrum (unlike Taylor microscale, which corresponds to the full energy spectrum) and hence is characteristic of the large length scales of fully developed AWT.}} To elucidate the evolution of the total perturbation energy $\left\langle E^{(2)}\right\rangle$ utilizing the integral length scale, we assume the following model spectral energy density $\widehat{E}_k$, \begin{equation} \widehat{E}_k = \begin{cases} C_1 k^n & k_0\leq k \leq k_E \\ C_2 k^{-2} & k_{\delta}> k > k_E \end{cases}, \label{eq: SpectralEnergyAssump} \end{equation} where $k^n$ corresponds to the shape of initialized energy spectrum in the range $(k_0,k_E)$ (Fig.~\ref{fig: LengthScalesSummary}). In this work, we only utilize the white noise initialized AWT cases which correspond to $n=0$ (see Table~\ref{tab: InitialSpectra}). Moreover, \begin{equation} C_2 = C_1 k_E^{n+2}. \end{equation} The wave numbers $k_E$ and $k_\delta$ vary in time due to decaying energy. By definition, the mean of perturbations is zero. Hence, the smallest wavenumber containing energy $k_0$ (cf. Fig.~\ref{fig: LengthScalesSummary}) is the reciprocal of the domain length $L$, i.e., \begin{equation} k_0 = 1/L. \end{equation} We note that the above model spectral energy $\widehat{E}_k$ holds for two primary reasons. Firstly, the energy cascade results in the $k^{-2}$ decay of the spectral energy $\widehat{E}_k$ due to formation of shock waves~\cite{Gurbatov_JFM_1997}. In the limit of vanishing viscosity $\delta \rightarrow 0 $, such decay extends up to $k\rightarrow \infty$ in which case the developed shock waves render the system $C_0$ discontinuous. Secondly, the shape of the spectral energy $\widehat{E}_k$ for $k\rightarrow k_0$ corresponds to $k^{n}$, which is also the shape of initial energy spectral at time $t=0$. Such argument corresponds to the concept of \emph{permanence of large eddies} in hydrodynamic turbulence~\cite{Batchelor1953theory}, which in spectral space can be written as, \begin{equation} \widehat{E}_k(t) \approx \widehat{E}_k(t=0),~~\mathrm{as}~~k\rightarrow k_0. \label{eq: PLE} \end{equation} Gurbatov \emph{et al.}~\cite{Gurbatov_JFM_1997} utilized a similar argument in the context of Burgers turbulence. Combining the Eqs.~\eqref{eq: IntegralLengthScale}-\eqref{eq: PLE}, the integral length scale $\ell$ is given by, \begin{align} \ell &\approx \begin{cases} \sqrt{\frac{n+1}{n-1}\left(\frac{k^{n-2}_E + k^{n-3}_Ek_0 + \cdots k^{n-2}_0}{k^{n}_E + k^{n-1}_Ek_0 + \cdots k^{n}_0}\right)} & n\neq 1 \\ \sqrt{\frac{2\ln(k_E/k_0)}{k^2_E - k^2_0}} & n = 1. \end{cases}, \label{eq: IntegralLengthN} \end{align} where we have used the simplifying approximation of $k_\delta \gg k_E$. We note that the Eq.~\eqref{eq: IntegralLengthN} indicates the dependence of $\ell$ and consequently the energy decay law on $n$. In the present work, we perform numerical simulations for an uncorrelated white noise (filtered) which corresponds to $n=0$, and \begin{equation} \ell \approx \frac{1}{\sqrt{k_0 k_E}}. \label{eq: IntegralApprox} \end{equation} As a result of \emph{permanence of large eddies}, the decay of energy in the initial regime of AWT is associated only to the decreasing $k_E$ or increasing integral length scale $\ell$. Integrating Eq.~\eqref{eq: SpectralEnergyAssump} in the spectral space and differentiating in time yields (for $n=0$ in the current simulations), \begin{align} \frac{d \left\langle E^{(2)} \right\rangle}{dt} &= C_1\left(2\frac{dk_E}{dt}\left(1 - \frac{k_E}{k_\delta}\right) + \left(\frac{k_E}{k_{\delta}}\right)^2\frac{dk_\delta}{dt}\right) \\ &\approx -\frac{2C_1}{k_0 \ell^3}\frac{d\ell}{dt}. \label{eq: EdotLdot} \end{align} Above relation shows that derivation of the energy decay power law amounts to finding the kinetic equations of the integral length scale $\ell$ and the limiting wavenumber $k_\delta$. For TW and SW, $\ell$ remains constant by definition. Consequently, the energy decay rate only depends on decrease of wavenumber $k_\delta$ and the coefficient $C_2$ due to the thermoviscous diffusion (cf. Eq.~\eqref{eq: dEdt1} ). However, for an ensemble of shock waves in AWT, $\ell$ increases monotonically in time, as shown in Fig.~\ref{fig: IntegralLengthScaleEvolution}$(a)$ due to the coalescence of shock waves propagating in the same direction. At large times, the domain consists of only two shock waves propagating in opposite directions. In the context of Burgers turbulence in an infinite one-dimensional domain, Burgers~\cite{Burgers1974nonlinear} and Kida~\citep{Kida_1979_JFM} have derived the appropriate asymptotic evolution laws for the integral length scale $\ell$ based on the dimensional arguments. However, in the present work, the finiteness of the domain renders the asymptotic analysis infeasible. Our numerical results indicate that $\ell\sim t^{1/3}$ ($k_E\sim t^{-2/3}$) for randomly distributed shock waves at various $\mathrm{Re}_L$ values considered, as shown in Fig.~\ref{fig: IntegralLengthScaleEvolution}$(a)$. Equations~\eqref{eq: EdotLdot} and~\eqref{eq: IntegralApprox} show that such scaling is consistent with the observed energy decay law $\left\langle E^{(2)} \right\rangle \sim t^{-2/3}$ thus validating the result in Eq.~\eqref{eq: EdotLdot}. It is noteworthy that decay $k_E\sim t^{-2/3}$ is a result analogous to the one discussed in Burgers turbulence~\cite{Burgers1974nonlinear,Kida_1979_JFM,Gurbatov_JFM_1997} considered in an infinite one-dimensional domain. {\color{black}{Due to infinitely long domain, the average distance between the shocks approaches $1/k_E$ (not $\ell$) simply due to larger number of shocks in the domain separated by the distance $1/k_E$ since $k_E$ corresponds to the largest wavenumber carrying initial energy, thus implying that mean distance between the shocks increases as $t^{2/3}$ as noted by Burgers~\cite{Burgers1974nonlinear}.}} Based on the integral length scale, the Reynolds number $\mathrm{Re}_{\ell}$ can be defined as, \begin{equation} \mathrm{Re}_{\ell} = \mathrm{Re}_{L}\ell, \label{eq: IntegralReynoldsNumber} \end{equation} which captures the ratio of the diffusive time scale to the wave turbulence time. Upon formation of shock waves, the perturbation energy decays due to coalescence. Shock waves coalesce locally thus increasing the characteristic separation between the shock waves thus causing $\ell$ to increase. In this regime, the Reynolds number $\mathrm{Re}_{\ell}$ remains constant (Fig.~\ref{fig: IntegralLengthScaleEvolution}$(b)$) which denotes that the ratio of shock coalescence time scale $(\ell L^*)/(a^*_0A_{\mathrm{rms}})$ and the diffusive time scale $(\ell L^*)^2/\nu^*_0$ remains constant. As the wave turbulence decays further, $\ell \rightarrow L$ with continued decay of energy. Consequently, $\mathrm{Re}_{\ell}$ also begins to decay. \subsection{Kolmogorov length scale} For spectral energy $\widehat{E}_k\sim k^{-2}$ over the intermediate range of wavenumbers, $k\in (k_E, k_\delta)$ (cf. Fig.~\ref{fig: LengthScalesSummary}), the Taylor microscale can be estimated as, \begin{equation} \lambda\sim \frac{1}{\sqrt{k_E k_\delta}}, \label{eq: TaylorMicroscaleApp} \end{equation} utilizing the Eq.~\eqref{eq: lengthscales_centroids}. Equation~\eqref{eq: TaylorMicroscaleApp} shows that $\lambda$, despite being a dissipative scale, is not the smallest scale generated due to the energy cascade. Analogous to the hydrodynamic turbulence, we define the Kolmogorov length scale $\eta$~\cite{tennekes1972first} as the smallest length scale generated as a result of the acoustic energy cascade. The length scale $\eta$ can be approximated by the balance of nonlinear steepening and energy dissipation, i.e., \begin{equation} \frac{{A}^2_{\mathrm{rms}}}{\eta} \sim \delta\frac{A_{\mathrm{rms}}}{\eta^2}, \quad \eta \sim \frac{\delta}{{A}_{\mathrm{rms}}}, \label{eq: KolomogorovScale} \end{equation} where $A_{\mathrm{rms}}$ is defined in Eq.~\eqref{eq: PcDef}. Figure~\ref{fig: LengthScalesSummary} illustrates the integral length scale $\ell$ and the Kolmogorov length scale $\eta$ in a typical AWT field. Visual inspection indicates $\ell \gg \eta$ which is as expected. We note that $\eta$ and $1/k_{\delta}$ evolve in time similarly, differing only by a constant value. For AWT, this is immediately realizable since, Eq.~\eqref{eq: TaylorMicroscaleApp} shows that $k_\delta \sim t^{-1/3}$ and Eq.~\eqref{eq: KolomogorovScale} shows that $\eta \sim t^{1/3}$ which implies $k_\delta \eta$ remains constant when the energy decays. \begin{figure} \centering \includegraphics[width=\linewidth]{TravelingWaveSpectra.pdf} \put(-240,370){$(a)$} \put(-240,250){$(b)$} \put(-240,120){$(c)$} \caption{{\color{black}{Fully developed spectra of compensated energy $(a)$, spectral energy flux $(b)$, and cumulative dissipation $(c)$ for TW at time instant $t_0\approx0.03$. Harmonics with wavenumbers such that $k\eta < 1$ contain all the energy. The spectral energy flux vanishes at $k\eta \approx 1$ thus indicating numerical resolution of all the energy containing harmonics. The marked regime $0.1<k\eta<1$ signifies the dissipation range. The constant $C\approx 0.075$. (--) $A_{\mathrm{rms},0}=10^{-1}$; ($--$) $A_{\mathrm{rms},0}=10^{-2}$; ($\cdots$) $A_{\mathrm{rms},0}=10^{-3}$}}} \label{fig: SpectralFlux_Traveling} \end{figure} For TW and SW, the spectral energy given by Eq.~\eqref{eq: SpectralEnergyAssump} corresponds to the degenerate case of $k_0 = k_E = 1$. For such a form of spectral energy, the energy evolution (cf. Eq.~\eqref{eq: EdotLdot}) changes to, \begin{equation} \frac{d\left\langle E^{(2)}\right\rangle}{dt} = \frac{1}{k_0}\frac{dC_2}{dt}\left(1 - \frac{k_0}{k_\delta}\right) + \frac{C_2}{k^2_\delta}\frac{dk_\delta}{dt}. \label{eq: dEdt1} \end{equation} As shown in Fig.~\ref{fig: EnergyDecayAll}, the Taylor microscale $\lambda \rightarrow \sqrt{\delta t}$. Consequently, for $k_E = k_0$ constant, Eq.~\eqref{eq: TaylorMicroscaleApp} shows that $k_\delta \sim t^{-1}$. Equation~\eqref{eq: dEdt1} shows that the decay of perturbation energy is due to decay in $C_2$ and $k_{\delta}$. Our numerical results (cf.~Fig.~\ref{fig: EnergyDecayAll}) show that for TW and SW, $\left\langle E^{(2)} \right\rangle \sim t^{-2}$ which suggests that $C_2\sim t^{-2}$ for $k_\delta\gg 1$ from Eq.~\eqref{eq: dEdt1}. Hence, the compensated energy spectrum $k^2\widehat{E}_k\sim t^{-2}$ for both TW and SW indicating that dissipation $\widehat{\mathcal{D}}_k$ remains active over all the length scales $k>k_0$ while the energy decays. Equation~\eqref{eq: KolomogorovScale} shows that the Reynolds number based on the Kolmogorov length scale or the shock thickness $Re_{\eta} = \eta \mathrm{Re}_L$ remains constant in time, {\color{black}{ \begin{equation} Re_{\eta} = \frac{\rho^*_0 a^*_0 L^* \eta A_{\mathrm{rms}}}{\mu^*_0} = \frac{4}{3} + \frac{\gamma - 1}{Pr}. \end{equation} }} Above relation shows that $Re_{\eta} = \mathcal{O}\left(1\right)$ indicating that $\eta$ is the length scale at which diffusion dominates the nonlinear wave steepening.\section{Scaling of spectral quantities} \label{sec: SpectralScaling} In this section, we discuss the variation and scaling of the energy $\widehat{E}_k$, the spectral energy flux $\widehat{\Pi}_k$, and the cumulative dissipation $\sum_{k'<k} \widehat{\mathcal{D}}_{k'}$ for high amplitude TW, SW, and AWT cases utilizing the length scale analysis presented in the previous sections. We show that the spectral energy $\widehat{E}_k$ and the cumulative dissipation $\sum_{k'<k} \widehat{\mathcal{D}}_{k'}$ for all the cases can be collapsed on to a common structure versus the reduced wavenumber $k\eta$ however, the flux $\widehat{\Pi}_k$ lacks such a universality. As discussed in previous section (cf. Eq.~\eqref{eq: dEdt1}), the decay of total energy $\left\langle E^{(2)}\right\rangle$ and dissipation rate $\epsilon$ for TW is given by, \begin{equation} \left\langle E^{(2)}\right\rangle \sim t^{-2},~\mathrm{and}~\epsilon\sim t^{-3}, \label{eq: EnergyTimeTraveling} \end{equation} which are well known results for the Burgers equation as well~\citep{Bec_PhyRep_2007}. \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{StandingWaveSpectra.pdf} \put(-240,370){$(a)$} \put(-240,250){$(b)$} \put(-240,120){$(c)$} \caption{Fully developed spectra of compensated energy $(a)$, spectral energy flux $(b)$, and cumulative dissipation $(c)$ for SW averaged over one time cycle after $t_0\approx0.04$. Harmonics with wavenumbers $k\eta < 1$ contain all the energy. The spectral energy flux vanishes at $k\eta \approx 1$ thus indicating numerical resolution of all the energy containing harmonics. The marked regime $0.1<k\eta<1$ signifies the dissipation range. The constant $C\approx 0.075$. (--) $A_{\mathrm{rms},0}=10^{-1}$; ($--$) $A_{\mathrm{rms},0}=10^{-2}$; ($\cdots$) $A_{\mathrm{rms},0}=10^{-3}$} \label{fig: SpectralFlux_Standing} \end{figure} While the results in Eq.~\eqref{eq: EnergyTimeTraveling} are well known, we note that such power law decay results in a universally constant structure of shock waves in the spectral space, as shown in Figs.~\ref{fig: SpectralFlux_Traveling} and~\ref{fig: SpectralFlux_Standing}. Utilizing the estimate of Kolmogorov length scale $\eta$ given in Eq.~\eqref{eq: KolomogorovScale}, the energy dissipation rate $\epsilon$ and the Kolmogorov length scale $\eta$ can be related as, \begin{equation} \epsilon \sim \frac{{A_{\mathrm{rms}}^3}}{\ell},~\mathrm{and}~ \eta \sim \frac{\delta}{\left(\epsilon \ell\right)^{1/3}}. \end{equation} {\color{black}{Hence, the energy spectrum $\widehat{E}_k$ can be written in the following collapsed form (Fig.~\ref{fig: SpectralFlux_Traveling}$a$). \begin{equation} \widehat{E}_kk^{2}\epsilon^{-2/3}\ell^{1/3} \sim C F(k\eta). \label{eq: SpectralForm} \end{equation} In Eq.~\eqref{eq: SpectralForm}, the integral length scale $\ell$ is used for making the left hand expression dimensionless}}. For TW and SW, the integral length scale $\ell$ remains constant by definition ($\ell = L$). Hence, $C$ in Eq.~\eqref{eq: SpectralForm} is constant and can be attributed to the Kolmogorov's universal equilibrium theory for hydrodynamic turbulence. $F(.)$ is a function which decays as the reduced wavenumber $k\eta$ increases to 1. From the numerical simulations for cases listed in Table~\ref{tab: test_cases} we obtain, \begin{equation} C \approx 0.075. \label{eq: ConstantValue} \end{equation} Scaling of $\widehat{\Pi}_k$ with the energy dissipation rate $\epsilon$ shows the relative magnitude of spectral energy flux compared to the energy dissipation.{\color{black}{ For increasing Reynolds numbers $\mathrm{Re}_L$, we note that $\widehat{\Pi}_k/\epsilon$ increases but still remains less than 1 in the energy transfer and dissipation range, as shown in Fig.~\ref{fig: SpectralFlux_Traveling}$(b)$. This highlights the primary difference between energy spectra of nonlinear acoustic waves and hydrodynamic turbulence, in which, the energy transfer range does not exhibit viscous dissipation~\cite{pope2000turbulent}. However, in nonlinear acoustics, the dissipation occurs over all the smaller length scales which do not contain energy initially (Fig.~\ref{fig: LengthScalesSummary}$(d)$). Moreover, for $k\eta\approx 0.1$, the flux $\widehat{\Pi}_k$ rapidly approaches to zero. In the regime $k\eta>0.1$, scaled cumulative dissipation $\sum_{k'<k} \widehat{\mathcal{D}}_{k'}/\epsilon\rightarrow 1$ as $k\eta\rightarrow 1$.}} Such functional forms of spectral energy, spectral energy flux, and cumulative dissipation can also be realized for the SW case. At later times, the nonlinear evolution results in two opposite traveling shock waves which collide with each other twice in one time period. Such collisions cause instantaneous peaks in the dissipation rate $\epsilon$ and corresponding oscillations in the Taylor microscale $\lambda$, as shown in Fig.~\ref{fig: EnergyDecayAll}. However, the total energy $\left\langle E^{(2)} \right\rangle$ decays monotonically by definition. In the spectral space, such collisions generate periodic oscillations in the spectral energy flux $\widehat{\Pi}_k$, as shown in Fig.~\ref{fig: EnergyFLux}. Averaging over one such time cycle yields the energy spectra forms similar to that for TW, as shown in Fig.~\ref{fig: SpectralFlux_Standing}. Such cycle averaging is allowed since the total energy $\left\langle E^{(2)}\right\rangle$ and the dissipation rate $\epsilon$ decay such that averaged behavior is identical to the one of traveling waves. Furthermore, the value of the constant $C$ is identical for SW.{\color{black}{ We further note that for the case with lowest Reynolds number $\mathrm{Re}_L$ ($\nu_0 = 1.836\times10^{-5}$ and $A_{0,\mathrm{rms}} = 10^{-3}$), the spectra exhibit energy for $k\eta > 1$ (Fig.~\ref{fig: SpectralFlux_Standing}$(a)$ since the Eq.~\eqref{eq: KolomogorovScale} underpredicts $\eta$. This suggests that the nonlinear spectral energy transfer is small compared to the spectral dissipation, as shown by Fig.~\ref{fig: SpectralFlux_Standing}$(b)$.}} As discussed in previous sections, the decay phenomenology of AWT is different from that of TW and SW. Typical acoustic field $u'(x,t), p'(x,t)$ for a randomly initialized perturbation at a time after shock formation is shown in Fig.~\ref{fig: LengthScalesSummary}$(a)$. The velocity field corresponds to randomly positioned shocks connected with almost straight slant lines (expansion waves) and the pressure field with identical distribution of shocks but connected with horizontal lines. Shocks traveling in the same direction collide inelastically and coalesce, while those traveling in opposite directions pass through. As discussed previously, the integral length scale $\ell$ defines the average distance between the adjacent shock traveling in the same direction. Due to gradual coalescence of the shocks, $\ell$ increases in time. Moreover, as $t\rightarrow \infty$, it is obvious that two opposite traveling shocks remain in the domain and $\ell \rightarrow L$. We note that such behaviour is similar to the Burgers turbulence~\citep{Kida_1979_JFM}. Figure~\ref{fig: SpectralEnergyTurbulence} shows the fully developed compensated spectra at scaled dimensionless time $t/\tau = t_0\approx 6\times10^{-4}$. For AWT, the compensated energy spectrum $\widehat{E}_kk^{2}\epsilon^{-2/3}\ell^{1/3}$ defined in Eq.~\eqref{eq: SpectralForm} does not remain constant in the energy transfer range of wavenumbers due to decay laws of energy and dissipation derived in the previous section.{\color{black}{ Moreover, for lowest $\mathrm{Re}_L$ case, the spectra exhibit energy for $k\eta>1$ (Fig.~\ref{fig: SpectralEnergyTurbulence}$(a)$) due to underprediction of $\eta$ obtained via balancing of nonlinear wave propagation and thermoviscous dissipation effects. The dissipation acts at large length scales also in the lowest $\mathrm{Re}_L$ case. Consequently, the spectral energy flux $\widehat{\Pi}_k$ is very small compared to dissipation $\epsilon$ and the length scale $\eta$ is primarily governed by diffusion only.}} \begin{figure} \centering \includegraphics[width=\linewidth]{SpectralEnergyTurbulence.pdf} \put(-240,370){$(a)$} \put(-240,250){$(b)$} \put(-240,120){$(c)$} \caption{Fully developed spectra of compensated energy, $(a)$, spectral energy flux $(b)$, and cumulative dissipation $(c)$ against scaled wavenumber $k\eta$ for the randomly initialized broadband noise (AWT) cases with $A_{\mathrm{rms},0}$ and $\nu_0$ listed in Table~\ref{tab: test_cases} at dimensionless time $t/\tau = \tau_0\approx 6\times10^{-4}$. The marked regime $0.1<k\eta<1$ signifies the dissipation range. (--) $A_{\mathrm{rms},0}=10^{-1}$; ($--$) $A_{\mathrm{rms},0}=10^{-2}$; ($\cdots$) $A_{\mathrm{rms},0}=10^{-3}$} \label{fig: SpectralEnergyTurbulence} \end{figure}\section{Concluding remarks} We have studied the spectral energy transport and decay of finite amplitude planar nonlinear acoustic perturbations governed by fully compressible 1D Navier-Stokes equations through shock-resolved direct numerical simulations (DNS) focusing on propagating single harmonic traveling wave (TW), standing wave (SW), and randomly initialized Acoustic Wave Turbulence (AWT). The maximum entropy perturbations scale as $p'^2$ for normalized pressure perturbation $p' \sim \mathcal{O}\left(10^{-3} - 10^{-1}\right)$. Consequently, the second order nonlinear acoustic equations are adequate to derive physical conclusions on spectral energy transfer in the system. Utilizing the second order equations, we derived the analytical expression for corrected energy corollary for finite amplitude acoustic perturbations yielding infinite order correction term in the perturbation energy density. We have shown that the spatial average of the corrected perturbation energy density can be classified as a Lyapunov function for the second order nonlinear acoustic system with strictly monotonic behaviour in time. Utilizing the corrected energy corollary, we derived the expressions for spectral energy, spectral energy flux, and spectral dissipation, analogous to the spectral energy equation studied in hydrodynamic turbulence. Utilizing the spectral expressions, we performed theoretical study of three possible length scales characterizing a general nonlinear acoustic system, namely, the integral length scale $\ell$, the Taylor microscale $\lambda$, and the Kolmogorov length scale $\eta$. In traveling waves (TW) and standing waves (SW), $\ell$ remains constant in the decaying regime. Spatial average of perturbation energy decays as $\left\langle E^{(2)}\right\rangle\sim t^{-2}$ and dissipation rate as $\epsilon \sim t^{-3}$ in time. The Kolmogorov scale increases linearly in time ($\eta\sim t$) in the decaying regime. Moreover, the spectral energy for both traveling and standing waves assumes the self-similar form: $\widehat{E}_k k^2 \epsilon^{-2/3}\ell^{1/3}\sim 0.075 f(k\eta)$. In acoustic wave turbulence (AWT), due to gradual increase of the integral length scale $\ell$ caused by the shock coalescence, the approximate decay laws are $\left\langle E^{(2)}\right\rangle \sim t^{-2/3}$ and $\epsilon \sim t^{-5/3}$, similar to the Burgers turbulence~\citep{Burgers1974nonlinear}. While, various cases for AWT qualitatively collapse with the scaling $\widehat{E}_k k^2 \epsilon^{-2/3}\ell^{1/3}$, quantitative scaling can only be obtained utilizing a statistically stationary ensemble of shock waves combined with random forcing, which falls beyond the current scope. \section{Acknowledgement} We acknowledge the financial support received from the NSF/DOE under Grant No. DE-SC0018156 and Lynn Fellowship at Purdue University. Computations have been run on the high-performance computing resources provide by the Rosen Center for Advanced Computing (RCAC) at Purdue University.

e090e17d2fb46cd144846741bbf81bba44a401c6

\section{INTRODUCTION} Kernel PCA (KPCA) is a non-linear version of PCA proposed in [7]. A major limitation of KPCA is that the eigen vectors of the covariance matrix in the kernel space are linear combinations of all the training data points, which becomes cumbersome for storage as well as querying a new test point. Obtaining sparse coefficients in KPCA is of paramount importance for real world applications. This problem was first addressed in [8], where sparse kernel feature analysis method was proposed to obtain sparse coefficients, by imposing a $l_1$ penalty on the coefficients. Tipping in [9] proposed a Probabilistic PCA([10]) based method for SKPCA. However, such a method is data dependent as it does not ensure sparse solutions for all types of data. Achlioptas \textit{et al.} in [11] propose a sparse unbiased and low variance estimator for the eigenvectors using concentration inequalities. Recently, Wang \textit{et al}. in [3] formulated SKPCA as a regression problem and solved it using Alternating Direction Method of Multipliers (ADMM). Outlier detection is a one class classification problem with possible multiple sub classes within the outlier class. The trained model for this task should be able to distinguish the learned inlier class from an outlier or novel class. One class SVM based approach in [5] and SVDD approach in [6] have previously been used for outlier detection. Another approach for outlier detection has been presented in [2], in which KPCA has been utilized for modeling the training set. The main contributions of this paper are two fold. Firstly, we extend the work done in [1] to the case of SKPCA. We apply their technique in the kernel space to obtain a generalized version of the problem in Theorem 4 of [1] and also present a method to solve it. To the best of our knowledge, this method for SKPCA has not been employed before. To validate our method, we consider the problem of outlier detection for which KPCA has been used before in [2]. By performing extensive experimentation on 5 real world datasets, we show that it performs nearly as well as KPCA (even better than KPCA for 3 datasets) and better than the recently proposed sparse KPCA approach in [3] as well as naive thresholding (simply picking the largest magnitude coefficients). It must be emphasized here that we are not claiming that our SKPCA method performs better than normal KPCA, instead we are merely claiming that it works almost as well as normal KPCA even under high sparsity. Secondly, we provide a novel probabilistic proof to justify the existence of sparse solutions to the KPCA problem using the RBF kernel. To the best of our knowledge, this is the first attempt at mathematically justifying the validity of sparsifying the KPCA problem. The paper is organized as follows - Section \Rmnum{2} gives a brief overview of the outlier detection algorithm in [2] using KPCA. Section \Rmnum{3} describes our method of Sparse KPCA for outlier detection. Section \Rmnum{4} presents our theoretical justification of sparse KPCA for the RBF kernel. Section \Rmnum{5} discusses and shows the results of our experiments. Section \Rmnum{6} concludes the paper. \section{KPCA for Outlier Detection} \begin{comment} {Kernel PCA is an extension of PCA to a higher dimensional feature space} $\phi(x)$ : $R^{d}$ $\rightarrow$ $R^{M}$, where $d \langle M$. The Covariance matrix in the feature space is defined by $C = \frac{1}{N} \sum_{i=1}^{N} \phi(x_i)\phi(x_i)^{T}$, assuming the data is centered in kernel space. Let $d_k$ and $v_k$ be the $k^{th}$ eigen value and eigen vector of $C$ respectively, then $C v_k = d_k v_k$ where $k=1,2,3.....N$. Using the previous formulation of Covariance matrix we get $C v_k = \frac{1}{N} \sum_{i=1}^{N} \phi(x_i)(\langle\phi(x_i), v_k>) = d_k v_k$. It can be shown as in [7] that $v_k = \sum_{i=1}^{N} \alpha_{k,i} \phi(x_i)$, where $\alpha_{k,i}$ is the $i^{th}$ component of the $k^{th}$ eigen vector of the Gram Matrix ($K$) such that $K_{ij} = \langle\phi(x_i),\phi(x_j)>$. Thus, the eigen vectors in Kernel PCA are a linear combination of all the training data points and hence using a sparse representation of $\alpha_k$ would significantly reduce the necessary data storage and can also act as a regularizer. \end{comment} Hoffmann in [2] presented an outlier detection approach using KPCA. In [2], spherical potential of a point $z$ (mapped to $\phi(z)$ in feature space) is defined with respect to a data distribution as $p_s(z) = ||\phi(z) - \phi_0||^{2}$, where $\phi_0$ is the mean of the data in feature space. The reconstruction error is defined as $p(\widetilde{\phi}(z)) = \langle\widetilde{\phi}(z),\widetilde{\phi}(z) \rangle - \langle W\widetilde{\phi}(z),W\widetilde{\phi}(z) \rangle$, where $\widetilde{\phi}(z) = \phi(z) - \phi_0$ and $W$ contains the top $q$ eigen vectors of the data in kernel space, arranged along the rows. We seek to get a sparse representation of $W$ using our SKPCA approach. Specifically, if the reconstruction error of a point is less than a certain threshold, then it is classified as an inlier otherwise as an outlier. When we use an RBF kernel, inliers lie inside a sphere whereas outliers lie outside it. \setcounter{equation}{0} \section{Sparse KPCA For Outlier Detection} In this section, we explain our method of obtaining sparse coefficients for the PCs in kernel space and its application in the outlier detection problem. \subsection{Sparse KPCA (SKPCA) Algorithm} We have $n$ data points $\{x_{1},x_{2},\ldots,x_{n}\}$ which get mapped to $\{{\phi}(x_{1}),{\phi}(x_{2}),\ldots,{\phi}(x_{n})\}$ in the kernel space. We assume that the data in the kernel space is centered. Similar to [1], we first propose the following regression problem to obtain the coefficients of the first PC (${\alpha_1}$): \begin{equation} \begin{aligned} & \underset{\alpha,\beta}{\text{min }} \mathrm \sum_{k=1}^n (\|{\phi}(x_{k}) - v_{1}\langle v_{2},{\phi}(x_{k}) \rangle\|^2 +{\lambda}{\beta_{1,k}}^2) \end{aligned} \end{equation} \vspace{-0.4cm} \begin{equation*} \begin{aligned} \vspace{-0.4cm} \text{where } v_{1} = \sum_{k=1}^n {\alpha_{1,k}}{\phi}(x_{k}) \text{ , } v_{2} = \sum_{k=1}^n {\beta_{1,k}}{\phi}(x_{k}) \end{aligned} \end{equation*} \begin{equation*} \begin{aligned} \vspace{-0.55cm} \text{such that } \langle v_{1},v_{1} \rangle = 1 \implies\ {\alpha_1}^{T}K{\alpha_1} = 1 \end{aligned} \end{equation*} \begin{comment} \begin{multline*} \text{where } v_{1} = \sum_{k=1}^n {\alpha_{k}}{\phi}(x_{k}) \text{ , } v_{2} = \sum_{k=1}^n {\beta_{k}}{\phi}(x_{k}) \\ \text{ and } v_{1}^{T}v_{1} = 1 \end{multline*} \end{comment} \begin{comment} \begin{multline*} v_{1}^{T}v_{1} = 1 \implies\ \sum_{i=1}^n\sum_{j=1}^n {\alpha_{i}}{\alpha_{j}}K(x_{i},x_{j}) = 1 \\ \implies\ {\alpha}^{T}K{\alpha} = 1 \end{multline*} \end{comment} Here, $K$ is the Gram matrix of the centered data~\cite{c7}. Also, $\alpha_1$ and $\beta_1$ are $n$ x $1$ vectors with $\alpha_{1,i}$ and $\beta_{1,i}$ as their $i^{th}$ elements respectively. So, we must minimize equation (1) (denote it by $J$) subject to ${\alpha_1}^{T}K{\alpha_1} = 1$. Substituting the values of $v_{1}$ and $v_{2}$ in $J$, we get: \begin{equation} \begin{aligned} & J = \ \sum_{k=1}^n \|{\phi}(x_{k}) - (\sum_{i=1}^n {\alpha_{1,i}}{\phi}(x_{i})) {\overline{\beta_{1,k}}}\|^2 +{\lambda}\sum_{k=1}^n {\beta_{1,k}}^2 \end{aligned} \end{equation} \vspace{-0.2in} \begin{equation*} \begin{aligned} \text{where } {\overline{\beta_{1,k}}} = \sum_{i=1}^n {\beta_{1,i}}K(x_{k},x_{i}) = \text{$k^{th}$ element of } K{\beta_1} \end{aligned} \end{equation*} \vspace{-0.2cm} Rewriting (2) in matrix form and using ${\alpha_1}^{T}K{\alpha_1} = 1$, $K^{T}=K$, i.e. the Gram matrix is symmetric, we get: \begin{equation} \begin{aligned} & J = tr(K) - 2{\alpha_1}^{T}K^2{\beta_1} + {\beta_1}^{T}K^2{\beta_1} + \lambda{\beta_1}^{T}{\beta_1} \end{aligned} \end{equation} Observe that this is the same as the PCA problem in [1] with the data matrix $X$ replace by the Gram matrix $K$. Even though the optimization problem is exactly the same, the constraint is different, namely ${\alpha_1}^{T}K{\alpha_1} = 1$ instead of ${\alpha_1}^{T}{\alpha_1} = 1$. This is quite intuitive. Partially differentiating with respect to ${\beta}$ and setting it to $0$ : \vspace{-0.03in} \begin{equation} \begin{aligned} \frac{\partial J}{\partial \beta_1} = 0 \implies\ \widehat{\beta_1} = (K^2+\lambda I)^{-1} K^2{\alpha_1} \end{aligned} \end{equation} Substituting $\widehat{\beta_1}$ from (4) in $J$, we get: \vspace{-0.02in} \begin{equation} \begin{aligned} J = tr(K) - {\alpha_1}^{T}K^2(K^2+\lambda I)^{-1}K^2{\alpha_1}\\ \end{aligned} \end{equation} Now, minimizing $J$ with respect to $\alpha_1$ subject to $\alpha_{1}^{T}K\alpha_1 = 1$ is equivalent to the following problem: \begin{equation} \begin{aligned} \widehat{\alpha_1} = \underset{\alpha_1}{\operatorname{argmax}}{\hspace{0.1in}} {\alpha_1}^{T}K^2(K^2+\lambda I)^{-1}K^2{\alpha_1} \end{aligned} \end{equation} \begin{comment} \begin{equation*} \text{ subject to }\alpha_{1}^{T}K\alpha_1 = 1 \end{equation*} \end{comment} Let $H$ denote the Lagrangian cost function obtained by imposing the equality constraint and $\rho$ be the Lagrange multiplier. \begin{equation} \begin{aligned} H = {\alpha_1}^{T}K^{2}(K^{2}+{\lambda}I)^{-1}K^{2}{\alpha_1}-\rho({\alpha_1}^{T}K{\alpha_1}-1) \end{aligned} \end{equation} \vspace{-0.3cm} \begin{equation} \begin{aligned} \frac{\partial H}{\partial \alpha_1} = 0 \implies\ K(K^{2}+{\lambda}I)^{-1}K^{2}{\alpha_1}=\rho {\alpha_1} \end{aligned} \end{equation} Now, let $K=EDE^{T}$ be the eigen decomposition of $K$. \begin{comment} \begin{multline*} \implies\ K(K^{2}+{\lambda}I)^{-1}K^{2}{\alpha_1} \\ = ED(D^2+\lambda I)^{-1}D^{2}E^{T} = E\frac{D^3}{(D^2+\lambda I)}E^{T} \end{multline*} \end{comment} \begin{equation} \implies\ K(K^{2}+{\lambda}I)^{-1}K^{2} = E\frac{D^3}{(D^2+\lambda I)}E^{T} \end{equation} For the sparse PCA problem in [1], we had $D^4$ in the numerator above and not $D^3$. However, $\frac{D^3}{(D^2+\lambda I)}$ is also monotonically increasing and so the value of the objective function will be maximum when $\widehat{\alpha_1}$ is proportional to the eigen vector corresponding to the largest eigen value (as was the case for the sparse PCA problem in [1]), i.e. $\widehat{\alpha_1} = \pm (\frac{E_1}{\sqrt{D_1}})$ where $D_1$ is the maximum eigen value and $E_1$ is the corresponding eigen vector and $ \widehat{\beta_1} = (K^{2}+\lambda I)^{-1}K^2 \widehat{\alpha_1}=(\frac{D_1^2}{D_1^2+ \lambda})(\frac{E_1}{\sqrt{D_1}})$. This was only for the first PC. Now, we will deal with the case of the first $m$ PCs and introduce the Lasso term as done in [1] so as to obtain sparse solutions. For this, we must minimize the following objective function with respect to $\alpha_m$ and $\beta_m$ which are $n$ x $m$ matrices: \begin{multline} \begin{aligned} J = tr(K) - 2tr({\alpha_m}^{T}K^2{\beta_m}) + tr({\beta_m}^{T}(K^2+\\ \lambda I){\beta_m}) +\sum_{j=1}^m \lambda_j |\beta_{m}^{j}|_{1} \text{ subject to } \alpha_m^{T}K\alpha_m = I_{m} \end{aligned} \end{multline} \begin{comment} \begin{multline*} \begin{aligned} J = \sum_{k=1}^n \|{\phi}(x_{k}) - (\Phi\alpha_{m})\langle (\Phi\beta_{m}),{\phi}(x_{k}) \rangle\|^2 +{\lambda}\|{\beta_{m}}\|^2\\ +\sum_{j=1}^m \lambda_j |\beta_{m}^{j}|_{1} \text{ subject to } \alpha_m^{T}K\alpha_m = I_{m} \text{, } \\ \Phi = [\phi(x_{1}),...,\phi(x_{n})] \end{aligned} \end{multline*} \end{comment} Here, $\beta_{m}^{j}$ denotes the $j^{th}$ column in $\beta_m$ and $I_{m}$ denotes the $m$ x $m$ identity matrix. \begin{comment} Note that this is a convex problem with respect to $\alpha_m$ and $\beta_m$ individually, but not jointly convex with respect to both of them. Thus, we shall follow an alternating optimization scheme, i.e. keeping $\alpha_m$ fixed, find the optimal $\beta_m$, then keeping $\beta_m$ fixed, find the optimal $\alpha_m$ and repeat this process until convergence. \end{comment} Note that this is a convex problem with respect to $\alpha_m$ and $\beta_m$ individually and so we shall follow an alternating optimization scheme, i.e. keeping $\alpha_m$ fixed, find the optimal $\beta_m$, then keeping $\beta_m$ fixed, find the optimal $\alpha_m$ and repeat this process until convergence. The solution for the first step of this iterative approach for sparse KPCA, i.e. keeping $\alpha_m$ fixed and finding the optimal $\beta_m$ is the same as that for sparse PCA in [1]. Specifically, $\beta_{m}^{j}$ is obtained by solving the following objective function: \begin{equation} \begin{aligned} \beta_{m}^{j}=\text{} & \underset{\beta^{*}}{\text{argmin}} & \mathrm {\beta}^{*T}(K^2+\lambda I){\beta}^{*}- 2{\alpha_{m}^{j}}^{T}K^2{\beta}^{*}+\lambda_j|\beta^{*}|_{1} \end{aligned} \end{equation} Note that $\alpha_{m}^{j}$ here refers to the $j^{th}$ column of $\alpha_{m}$ which was obtained in the previous iteration. This is the familiar naive elastic net problem [4] in ${\beta^{*}}$, which can be solved using the LARS-EN algorithm [4]. The solution for the second step of this iterative approach for sparse KPCA, i.e. keeping $\beta_{m}$ fixed and finding the optimal $\alpha_{m}$ is different as compared to that for sparse PCA in [1]. Specifically, we must solve the following problem: \begin{equation} \begin{aligned} \alpha_{m} = \text{} & \underset{\alpha^{*}}{\text{argmax}} & \mathrm tr({\alpha^{*}}^{T}K^2{\beta_m}) \text{ subject to } {\alpha^{*}}^{T}K{\alpha^{*}} = I_{m} \end{aligned} \end{equation} Note that in sparse PCA, the objective function for finding the optimal $\alpha_m$ was the same but the constraint was different, namely $\alpha_m^{T}\alpha_m = I_{m}$ . We present a corollary to Theorem 4 of [1] which enables us to solve the aforementioned problem in (12). We first restate Theorem 4 of [1] (renamed as \textit{Theorem 1} here) for ready reference of the reader followed by the corollary (\textit{Corollary 1}). \begin{theorem} Consider the following problem where $\alpha$, $\beta$ are $n$ x $m$ matrices ($n$ $>$ $m$): \begin{equation*} \begin{aligned} \widehat{\alpha} = \text{ } & \underset{\alpha}{\text{argmax}} & \mathrm tr({\alpha}^{T}{\beta}) \text{ subject to } \alpha^{T}\alpha = I_{m} \end{aligned} \end{equation*} The solution to this problem is given by $\widehat{\alpha}=U{V}^{T} \text{ where the SVD of } \beta = U{\Sigma}{V}^{T}$. \end{theorem} \begin{corollary} Consider the following problem where $\alpha$, $\beta$ are $n$ x $m$ matrices ($n$ $>$ $m$) and $Q$ is a $n$ x $n$ symmetric positive definite matrix: \begin{equation*} \begin{aligned} \widehat{\alpha} = \text{ } & \underset{\alpha}{\text{argmax}} & \mathrm tr({\alpha}^{T}{\beta}) \text{ subject to } \alpha^{T}Q\alpha = I_{m} \end{aligned} \end{equation*} The solution to this problem is given by $\widehat{\alpha} = U\Sigma^{-1/2}U^{*}{V^{*}}^{T}$ where the SVD of $\Sigma^{-1/2}U^{T}{\beta} = U^{*}{\Sigma^{*}}{V^{*}}^{T}$ and the SVD of $Q = U\Sigma U^{T}$. \end{corollary} Note that \textit{Corollary 1} can be interpreted as an extension of \textit{Theorem 1} to a generalized inner product space (defined by the matrix $Q$). We now present a proof for \textit{Corollary 1}. Let the SVD decomposition of $Q$ be $U\Sigma U^{T}$. Thus, $\alpha^{T}Q\alpha = (\alpha^{T}U\Sigma^{1/2})(\Sigma^{1/2}U^{T}\alpha)$, where $\Sigma^{1/2}$ is a diagonal matrix with diagonal entries equal to square roots of the corresponding entries of $\Sigma$. Now, denote $\alpha^{*} = (\Sigma^{1/2}U^{T}\alpha)$ and from the constraint, we must have ${\alpha^{*}}^{T}{\alpha^{*}} = I_{m}$. Also, $\alpha = (U\Sigma^{-1/2}{\alpha^{*}})$ and so ${\alpha}^{T}{\beta} = {{\alpha}^{*}}^{T}(\Sigma^{-1/2}U^{T}{\beta}) = {{\alpha}^{*}}^{T}{\beta^{*}}$, where ${\beta^{*}} = \Sigma^{-1/2}U^{T}{\beta}$. Thus, we have reduced the generalized problem to the following equivalent problem: \begin{equation*} \begin{aligned} \widehat{{\alpha^{*}}} = \text{ } & \underset{\alpha^{*}}{\text{argmax}} & \mathrm tr({{\alpha^{*}}}^{T}{{\beta^{*}}}) \text{ subject to } {{\alpha^{*}}}^{T}{\alpha^{*}} = I_{m} \end{aligned} \end{equation*} This is exactly the problem whose solution has been given in Theorem 4 of [1] (\textit{Theorem 1} in this paper). We directly use that result to get: \begin{equation*} \begin{aligned} \widehat{{\alpha^{*}}} = U^{*}{V^{*}}^{T} \text{, SVD of } {{\beta^{*}}} = U^{*}{\Sigma^{*}}{V^{*}}^{T} \end{aligned} \text{, } \widehat{\alpha} = U\Sigma^{-1/2}{\widehat{\alpha^{*}}} \end{equation*} For our algorithm, we put $\alpha=\alpha_m$, $\beta=K^2\beta_m$ and $Q=K$ in \textit{Corollary 1}. This completes the solution for the second stage of the iterative algorithm.\\ \textbf{SKPCA ALGORITHM: \\ 1. Initialize $\alpha_m$ with the $m$ eigen vectors of $K$ corresponding to the $m$ largest eigen values. \\2. First for fixed $\alpha_m$, solve the elastic net problem in (11) for $\beta_{m}$. \\3. Then with the $\beta_m$ obtained in the previous step, update $\alpha_m$ as ${\alpha_m} = U\Sigma^{-1/2}U^{*}{V^{*}}^{T}$ where $\Sigma^{-1/2}U^{T}K^2{\beta_m} = U^{*}{\Sigma^{*}}{V^{*}}^{T}$ and $K = U\Sigma U^{T}$. \\4. Repeat steps 2-3 till convergence. \subsection{Outlier Detection Algorithm using Sparse KPCA} We employ the same strategy presented in [2] for distinguishing outliers from inliers as explained earlier, except that here we use the $\beta_m$ obtained from the algorithm above (i.e. the sparse solutions) to construct the $W$ matrix in the reconstruction error term. \section{Theoretical Justification of Sparse KPCA With RBF Kernel} Assume that we have a set $S_{m}$ of $m$ data points, $\{x_{1},x_{2},\ldots,x_{m}\}$ drawn in i.i.d. fashion from a continuous probability distribution $P$. Consider the subset $S_{n} = \{x_{1},x_{2},\ldots,x_{n}\}$ of $n < m$ points. The elements of $S_{n}$ obey a particular condition, the details of which shall be presented later. We shall prove that the eigenvectors of the covariance matrix (henceforth referred to as $C_{m}$) in the kernel space constructed out of all points in $S_{m}$ can be expressed as a linear combination of just $\{\phi(x_{1}),\phi(x_{2}),\ldots,\phi(x_{n})\}$ with high probability for sufficiently large $m$ and $n$. \begin{comment} Let the Gram matrix for the points in $S_{m}$ and $S_{n}$ be $K_{m}$ and $K_{n}$ respectively. Also, let $v$ be an eigenvector of $K_{n}$ with eigenvalue $\lambda$. Thus, we have $K_{n}v = \lambda v$. We want to see whether $K_{m}\begin{bmatrix} v \\ \vec{0} \end{bmatrix} \approx \lambda \begin{bmatrix} v \\ \vec{0} \end{bmatrix}$ to investigate the existence of sparse approximate eigenvectors of $K_{m}$. We can partition $K_{m}$ as follows:\[ K_{m}= \left[ \begin{array}{c|c} K_{n} & K_{m-n,n} \\ \hline K_{m-n,n}^{T} & K_{m-n} \end{array} \right] \] In the above partition, $K_{m-n}$ is the Gram matrix of the remaining $m-n$ points. Observe that the first $n$ entries of $K_{m}\begin{bmatrix} v \\ \vec{0} \end{bmatrix} = K_{n}v = \lambda v$. Let us denote the $i^{th}$ entry of the vector $K_{m}\begin{bmatrix} v \\ \vec{0} \end{bmatrix}$ by $v_{i}'$. Our claim is that $\{v_{n+1}',v_{n+2}',\ldots,v_{m}'\}$ are individually close to $0$ with high probability for sufficiently large $m$ and $n$. We shall prove this claim next. Let $y = \begin{bmatrix} v_{1} \\ v_{2} \\ \vdots \\ v_{n} \end{bmatrix}$. Then, the $v_{i}'$ ($n+1 \leq i \leq m$) is given by: \begin{equation} v_{i}' = \sum_{j=1}^n {u_{j}k(x_{i},x_{j})} \text{, $k(x,y)$ being the kernel function.} \end{equation} For our case, we choose the RBF kernel and thus $k(x,y) = exp(-\frac{\|x-y\|^2}{2\sigma ^2})$. We want the summation in equation (13) to be close to 0 for all $i > n$. \end{comment} Once again, we assume that the data in the kernel space is centered. An eigenvector, $v$ of $C_{m}$ can be expressed as $v = \sum_{i=1}^m {\alpha_{i}}{\phi}(x_{i})$. Consider a random point $x'$ sampled from $P$. The projection of $x'$ on $v$ in the kernel space is given as $\langle \phi(x'),v \rangle = \sum_{i=1}^m {\alpha_{i}}k(x_{i},x')$, where $k(x,y)$ is the chosen kernel function. For our case, we choose the RBF kernel and thus $k(x,y) = \exp(-\frac{\|x-y\|^2}{2\sigma ^2})$. Without loss of generality, assume $\sigma = 1/\sqrt{2}$ for simplicity. Thus, we have $\langle \phi(x),v \rangle = \sum_{i=1}^n {\alpha_{i}}\exp(-\|x_{i} - x'\|^2) + \sum_{i=n+1}^m {\alpha_{i}}\exp(-\|x_{i} - x'\|^2)$. For $i>n$, define $f(i) = \underset{1 \leq j \leq n}\argmin{ \|x_{i} - x_{j}\|}$. We shall show that $\exp(-\|x_{i} - x'\|^2) \approx \exp(-\|x_{f(i)} - x'\|^2)$ for $i>n$ holds with high probability. For $x'$, define $g(x') = \underset{1 \leq j \leq n}\argmin{ \|x_{j} - x'\|}$. Then, for $i>n$ : $\exp(-\|x_{i} - x'\|^2) = \exp(-\|(x_{i} - x_{f(i)}) +(x_{f(i)} - x') \|^2) = \exp(-\|x_{f(i)} - x'\|^2)\times \exp(-\|x_{i} - x_{f(i)}\|^2) \times \exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x') \rangle)$. The third term in the above factorization can further be written as: $\exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x') \rangle) = \exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')} + x_{g(x')} - x') \rangle) = \exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')}\rangle) \times \exp(-2\langle (x_{i} - x_{f(i)}),(x_{g(x')} - x')\rangle)$. Finally, we get : $\exp(-\|x_{i} - x'\|^2) = \exp(-\|x_{f(i)} - x'\|^2)\times \exp(-\|x_{i} - x_{f(i)}\|^2) \times \exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')}\rangle) \times \exp(-2\langle (x_{i} - x_{f(i)}),(x_{g(x')} - x')\rangle)$. \begin{comment} We will now show that the second, third and fourth terms in the above factorization are individually close to 1, due to which their product is also close to 1 and as a result $\exp(-\|x_{i} - x'\|^2) \approx \exp(-\|x_{f(i)} - x'\|^2)$ with high probability. First we shall show that the second term is close to 1 with high probability. Let $\Pr(\|u - v\| \geq d) = p(d)$ where $u$ \& $v$ are sampled in i.i.d. fashion from $P$. Thus, $\Pr(\|x_{i} - x_{j}\| \leq d \text{ for some } j, 1 \leq j \leq n, i > n) = 1- (p(d))^n$ which intuitively tends to $1$ for large $n$. Thus, $\|x_{i} - x_{f(i)}\| \leq d$ holds with high probability. For $d = 0.1\sigma = 0.1/\sqrt{2}$, $\exp(-\|x_{i} - x_{f(i)}\|^2) \geq 0.995$. We make 2 assumptions here to show this formally. The first one is $\sigma^{2} = \mathrm{E} [\|u - v\|^{2}]$ ($u$ \& $v$ as defined before \& $\sigma = 1/\sqrt{2})$ as it works well in practice. The other one is $\|u - v\|^{2} \leq C\sigma^{2}$ $a.s.$ with $C>1$ and this too usually happens in practice. Now we can directly use Hoeffding's one-sided inequality to get $p(0.1/\sqrt{2}) \geq 1 - \exp(-1.9602/C^{2})$ and so $p(0.1/\sqrt{2})^{n}$ tends to 0 for large enough $n$ which is a function of $C$. Hence, the second term is close to 1 with high probability. \end{comment} We will now show that the second, third and fourth terms in the above factorization are individually close to 1, due to which their product is also close to 1 and as a result $\exp(-\|x_{i} - x'\|^2) \approx \exp(-\|x_{f(i)} - x'\|^2)$ with high probability. First we shall show that the second term is close to 1 with high probability. Let $\Pr(\|u - v\| \geq d) = p(d)$ where $u$ \& $v$ are sampled in i.i.d. fashion from the distribution $P$. Here, we assume that $p(d) < 1$ for $d>0$. From elementary probability, $\Pr(\|x_{i} - x_{j}\| \leq d \text{ for some } j, 1 \leq j \leq n, i > n) = 1- (p(d))^n$ which tends to $1$ for large enough $n$. Thus, $\|x_{i} - x_{f(i)}\| \leq d$ holds with high probability, for sufficiently large $n$. For $d = 0.1\sigma = 0.1/\sqrt{2}$, $\exp(-\|x_{i} - x_{f(i)}\|^2) \geq 0.995$. Hence, the second term is close to 1 with high probability. We shall deal with the fourth term before we analyze the third term. By the Cauchy-Schwarz inequality, $|\langle (x_{i} - x_{f(i)}),(x_{g(x')} - x')\rangle| \leq \|x_{i} - x_{f(i)}\|.\|x_{g(x')} - x'\|$. We showed in the analysis of the second term that $\|x_{i} - x_{f(i)}\| \leq d$ with high probability. We employ the same analysis for $x'$ instead of $x_{i}$ (and $x_{g(x')}$ instead of $x_{f(i)}$) to conclude that $\|x_{g(x')} - x'\| \leq d$ with high probability. Thus, with $d = 0.1/\sqrt{2}$, $0.99 \leq \exp(-2\langle (x_{i} - x_{f(i)}),(x_{g(x')} - x')\rangle) \leq 1.01$. Thus, the fourth term is also close to $1$ with high probability. Finally, we come to the analysis of the third term. Here, $|\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')})\rangle| \leq \|x_{i} - x_{f(i)}\|.\|x_{f(i)} - x_{g(x')}\| \leq d\|x_{f(i)} - x_{g(x')}\|$ from the analysis of the second term. Let $S$ denote a subset of any $n$ distinct points out of the $m$ points in $S_{m}$. There are ${m}\choose{n}$ such subsets. Define $d_{max}^S = \underset{i \in S,j \in S}{\text{max }} \mathrm \|x_{i}-x_{j}\|$. With this definition, $|\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')})\rangle| \leq d.d_{max}^{S_{n}}$. Let $\Pr(d_{max}^S \geq \tilde{d} \text{ for a particular } S) = \tilde{p}(\tilde{d}) = 1-(1-p(\tilde{d}))^{{n}\choose{2}}$. Here again, we have $\tilde{p}(\tilde{d}) < 1$ for $\tilde{d} > 0$ as $p(\tilde{d}) < 1$ for $\tilde{d} > 0$. Thus, $\Pr(d_{max}^S \leq \tilde{d} \text{ for some } S \text{ out of all possible choices}) = 1- (\tilde{p}(\tilde{d}))^{{m}\choose{n}}$ which also tends to $1$ for $m >> n$. Without loss of generality, assume that $S_{n}$ satisfies the condition $d_{max}^{S_{n}} \leq \tilde{d}$. This is the exact condition on the elements of $S_{n}$ which was mentioned in the beginning of this section. Taking $\tilde{d} = 0.3/\sqrt{2}$, $0.97 \leq \exp(-2\langle (x_{i} - x_{f(i)}),(x_{f(i)} - x_{g(x')})\rangle) \leq 1.03$. Thus, even the third term is also close to $1$ with high probability. Thus, $\exp(-\|x_{i} - x'\|^2) \approx \exp(-\|x_{f(i)} - x'\|^2)$ holds with high probability. So, we have $\langle \phi(x),v \rangle \approx \sum_{i=1}^n {\alpha_{i}}\exp(-\|x_{i} - x'\|^2) + \sum_{i=n+1}^m {\alpha_{i}}\exp(-\|x_{f(i)} - x'\|^2)$ with high probability. This can be rewritten as $\langle \phi(x),v \rangle \approx \sum_{i=1}^n {\beta_{i}}\exp(-\|x_{i} - x'\|^2)$, where $\beta_{i} = (\alpha_{i}+\sum_{j>n \mid f(j)=i} {\alpha_{j}}$). Therefore, $v = \sum_{i=1}^m {\alpha_{i}}{\phi}(x_{i}) \approx \sum_{i=1}^n {\beta_{i}}{\phi}(x_{i})$ with high probability, where $\beta_{i}$'s are as defined before. This finishes our proof. Note that one can design a naive algorithm based on the above proof, but its combinatorial nature would render it highly inefficient. We conjecture that our algorithm (or any other sparse KPCA algorithm) does this subset selection task efficiently and perhaps approximately by solving a convex problem instead. \section{Experiments} We have compared the performance of our SKPCA algorithm with standard KPCA, naive thresholding of the $m$ largest magnitude components of the KPCA coefficients (with $m$ = average number of non-zero components per PC obtained through SKPCA) as well the algorithm given in [3]. We have compared our method with naive thresholding because it has been noted in [1] that naive thresholding works almost as well as their sparse PCA method. We have tested these methods on 5 real world datasets - MNIST (Fig. 1), Fashion MNIST[12] (Fig. 2), Satimage2[13] (Fig. 3), ETH-80 dataset (Fig. 4) and Internet Advertisements dataset[14] (Fig. 5). Our experiments reveal that our SKPCA algorithm provides high sparsity without compromising on accuracy. The accuracy using our SKPCA algorithm is nearly the same as KPCA, which is not possible with naive thresholding at low sparsity and the sparsity obtained is in general much more than that obtained from the algorithm in [3], for the same accuracy. Also, our algorithm has just 1 sparsity controlling parameter which is the L1-ratio = $\lambda_j/\lambda$ $\forall j$, whereas the algorithm in [3] has 5 parameters - $\rho,\lambda,\lambda_{1,k},\epsilon^{abs},\epsilon^{rel}$. The metric used to measure accuracy in our experiments is F1-score (standard for outlier detection tasks). We obtained 3 plots in our simulations. The first one is a box plot to show the variability of F1-score over different training and test data subsets (we performed the same experiment 10 times on randomly chosen training and test data subsets out of the entire dataset). However, we did not show the variability of the algorithm in [3] for reasons explained later. The second one is a F1-score vs. sparsity (measured as the percentage of non-zero coefficients in the kernel space PCs) curve for all the aforementioned methods. The F1-score vs. sparsity curves were constructed by trying out the various methods on a range of sparsity controlling parameters for a randomly chosen subset of training and test data. The KPCA horizontal line in the F1-score vs. sparsity curves is provided just for reference and it has all non-zero coefficients. The third one contains ROC curves for all the aforementioned methods. The number of PCs used for MNIST, Fashion MNIST, ETH-80, Satimage2 and Internet Ads were 15, 15, 15, 7 and 24 respectively. We took $\sigma^{2}$ in the RBF kernel to be the average of $\|x_i-x_j\|^2$ for all pairs $i,j$ such that $i \neq j$. \justify \textbf{MNIST:} Similar to the experiment performed in [2], we chose 0 as the inlier class and all digits from 1-9 as the outlier class. We used a training set of 3000 inliers and a test set consisting of 3000 inliers and 3000 outliers. The variability plot is for L1-ratio=0.7. The mean and standard deviation of sparsity for the 10 trials were 3.35\% and 0.13\% respectively. The ROC curve of our algorithm is also for L1-ratio=0.7, while the ROC curve of the algorithm in [3] is for $\rho = 0.02, \lambda=0.001,\lambda_{1,k}=0.01,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$. The area under ROC curve (AUROC) values for our SKPCA algorithm, standard KPCA, the algorithm in [3] and naive thresholding were 0.974, 0.986, 0.962 and 0.957 respectively. \begin{comment} It can be clearly seen that the variance of the accuracy is lower for SKPCA as compared to naive thresholding, which is also mentioned in [1] for sparse PCA. The sparsity of the components for all these trials is also pretty much the same. In the accuracy vs. sparsity graph, the accuracy of our SKPCA method is the closest to ordinary KPCA as compared to the other methods for most of the sparsity axis, except for very high sparsity (i.e. towards the extreme left of the X-axis denoting the average percentage of non-zero components). \end{comment} \begin{comment} \begin{figure*}% \centering \subfigure[Accuracy Variability plot for MNIST]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{MNIST0_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for MNIST]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{MNIST0_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for MNIST]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{MNIST0_acc_300_g2.png}}% \caption{Plots for MNIST} \end{figure*} \begin{figure*}% \centering \subfigure[Accuracy Variability plot for Fashion MNIST]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{FashionMNIST_7_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Fashion MNIST]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{FashionMNIST7_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Fashion MNIST]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{FashionMNIST_acc_300_g2.png}}% \caption{Plots for Fashion MNIST} \end{figure*} \begin{figure*}% \centering \subfigure[Accuracy Variability plot for Satimage2]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{Satimage2_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Satimage2]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{Satimage2_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Satimage2]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{Satimage_acc_300_g2.png}}% \caption{Plots for Satimage2} \end{figure*} \begin{figure*}% \centering \subfigure[Accuracy Variability plot for Fruits ETH-80]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{Fruits_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Fruits ETH-80]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{Fruits_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Fruits ETH-80]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{Fruits_acc_300_g2.png}}% \caption{Plots for ETH-80} \end{figure*} \end{comment} \justify \textbf{Fashion MNIST:} This is a recent dataset which is very similar to MNIST but much more challenging. Here, we took item 0 (t-shirts and some other tops) as the inlier class and item 1 (pants) as the outlier class. Training and test dataset size was the same as in MNIST. The variability plot is for L1-ratio=0.3. The mean and standard deviation of sparsity for the 10 trials were 8.43\% and 0.24\% respectively. The ROC curve of our algorithm is also for L1-ratio=0.3, while the ROC curve of the algorithm in [3] is for $\rho = 0.05, \lambda=0.002,\lambda_{1,k}=0.006,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$. The AUROC values for our SKPCA algorithm, standard KPCA, the algorithm in [3] and naive thresholding were 0.919, 0.898, 0.872 and 0.877 respectively. \begin{comment} For this dataset though, the variance of SKPCA is a bit high, but still less than that of Naive Thresholding. The sparsity of the components for all these trials is nearly the same, even for this dataset. In the accuracy vs. sparsity graph, the accuracy of our SKPCA method is the closest to ordinary KPCA compared to the other methods over the entire sparsity range in consideration. \end{comment} \justify \textbf{Satimage2:} The data was already presented as inliers and outliers. Training set size was 400 inliers and test set consisted of 500 inliers and 71 outliers (only that many were available). The variability plot is for L1-ratio=0.6. The mean and standard deviation of sparsity for the 10 trials were 5.55\% and 0.33\% respectively. The ROC curve of our algorithm is also for L1-ratio=0.6, while the ROC curve of the algorithm in [3] is for $\rho = 0.07, \lambda=0.003,\lambda_{1,k}=0.016,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$. The AUROC values for our SKPCA algorithm, standard KPCA, the algorithm in [3] and naive thresholding were 0.963, 0.958, 0.942 and 0.935 respectively. \begin{comment} The variance of the accuracy is lower for SKPCA as compared to naive thresholding, like the previous datasets. The sparsity of the components for all these trials is pretty much the same. In the accuracy vs. sparsity graph, the accuracy of our SKPCA method is the best, better than even ordinary KPCA over almost the entire sparsity range in consideration (especially for high sparsity), and later converges to the same accuracy as KPCA. \end{comment} \justify \textbf{ETH-80:} We used the apples in this data set as inliers and tomatoes as outliers (they have nearly the same color and shape making it more challenging). Training set size was 300 inliers and test set consisted of 111 inliers and 111 outliers. The variability plot and is for L1-ratio=0.3. The mean and standard deviation of sparsity for the 10 trials were 11.07\% and 0.70\% respectively. The ROC curve of our algorithm is also for L1-ratio=0.3, while the ROC curve of the algorithm in [3] is for $\rho = 0.1, \lambda=0.006,\lambda_{1,k}=0.012,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$. The AUROC values for our SKPCA algorithm, standard KPCA, the algorithm in [3] and naive thresholding were 0.859, 0.853, 0.841 and 0.836 respectively. \justify \textbf{Internet Ads:} The data was already presented as inliers and outliers. Training set size was 600 inliers and test set consisted of 380 inliers and 380 outliers. The variability plot and is for L1-ratio=0.4. The mean and standard deviation of sparsity for the 10 trials were 2.60\% and 0.29\% respectively. The ROC curve of our algorithm is also for L1-ratio=0.4, while the ROC curve of the algorithm in [3] is for $\rho = 0.008, \lambda=0.009,\lambda_{1,k}=0.019,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$. The AUROC values for our SKPCA algorithm, standard KPCA, the algorithm in [3] and naive thresholding were 0.783, 0.785, 0.744 and 0.739 respectively.\\ \begin{comment} The variance of the accuracy is lower for SKPCA as compared to naive thresholding, just as in the previous datasets. The sparsity of the components for all these trials is nearly the same, though the sparsity is lesser in this case due to a lower L1-ratio. Even for this dataset, in the accuracy vs. sparsity graph, the accuracy of our SKPCA method is the best, better than even ordinary KPCA over the entire sparsity range in consideration. \end{comment} \begin{comment} \begin{figure}% \centering \subfigure[Accuracy Variability plot for MNIST]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{MNIST0_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for MNIST]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{MNIST0_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for MNIST]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{MNIST0_acc_300_g2.png}}% \caption{Plots for MNIST} \end{figure} \begin{figure}% \centering \subfigure[Accuracy Variability plot for Fashion MNIST]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{FashionMNIST_7_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Fashion MNIST]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{FashionMNIST7_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Fashion MNIST]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{FashionMNIST_acc_300_g2.png}}% \caption{Plots for Fashion MNIST} \end{figure} \begin{figure}% \centering \subfigure[Accuracy Variability plot for Satimage2]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{Satimage2_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Satimage2]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{Satimage2_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Satimage2]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{Satimage_acc_300_g2.png}}% \caption{Plots for Satimage2} \end{figure} \begin{figure}% \centering \subfigure[Accuracy Variability plot for Fruits ETH-80]{% \label{fig:first}% \includegraphics[width=3cm, height=3cm]{Fruits_var_300.png}}% \qquad \subfigure[Sparsity Variability plot for Fruits ETH-80]{% \label{fig:second}% \includegraphics[width=3cm, height=3cm]{Fruits_comps_300.png}}% \qquad \subfigure[Accuracy vs. Sparsity plot for Fruits ETH-80]{% \label{fig:third}% \includegraphics[width=3cm, height=3cm]{Fruits_acc_300_g2.png}}% \caption{Plots for ETH-80} \end{figure} \end{comment} \begin{comment} \begin{figure} \centering \subfloat[Accuracy Variability plot for MNIST]{ \includegraphics[width=30mm, height=30mm]{MNIST0_var_300.png} } \subfloat[Sparsity Variability plot for MNIST]{ \includegraphics[width=30mm, height=30mm]{MNIST0_comps_300.png} } \subfloat[Accuracy vs. Sparsity plot for MNIST]{ \includegraphics[width=30mm, height=30mm]{MNIST0_acc_300_g2.png}} \caption{Plots for MNIST} \end{figure} \begin{figure}[h] \centering \subfloat[Accuracy Variability plot for Fashion MNIST]{ \includegraphics[width=30mm, height=30mm]{FashionMNIST_7_var_300.png} } \subfloat[Sparsity Variability plot for Fashion MNIST]{ \includegraphics[width=30mm, height=30mm]{FashionMNIST7_comps_300.png} } \subfloat[Accuracy vs. Sparsity plot for Fashion MNIST]{ \includegraphics[width=30mm, height=30mm]{FashionMNIST_acc_300_g2.png}} \caption{Plots for Fashion MNIST} \end{figure} \end{comment} \begin{figure}[!h] \centering \subfloat[F1-Score Variability plot for MNIST]{ \includegraphics[width=36mm, height=39mm]{MNIST0_var.png} } \hspace{0.3 in} \subfloat[F1-Score vs. Sparsity plot for MNIST]{ \includegraphics[width=36mm, height=39mm]{MNIST_F1.png}} \hspace{0.2 in} \subfloat[ROC curves for MNIST]{ \includegraphics[width=36mm, height=39mm]{MNIST_ROC.png}} \caption{Plots for MNIST} \end{figure} \begin{figure}[!h] \centering \subfloat[F1 score Variability plot for Fashion MNIST]{ \includegraphics[width=36mm, height=38mm]{Fashion_MNIST_var_300.png} } \hspace{0.3 in} \subfloat[F1 score vs. Sparsity plot for Fashion MNIST]{ \includegraphics[width=36mm, height=38mm]{fashion_mnist_f1.png}} \hspace{0.2 in} \subfloat[ROC curves for Fashion MNIST]{ \includegraphics[width=36mm, height=39mm]{F_MNIST_ROC.png}} \caption{Plots for Fashion MNIST} \end{figure} \begin{figure}[!h] \centering \subfloat[F1-Score Variability plot for Satimage2]{ \includegraphics[width=36mm, height=38mm]{Satimage2_var.png} } \hspace{0.3 in} \subfloat[F1-Score vs. Sparsity plot for Satimage2]{ \includegraphics[width=36mm, height=38mm]{Satimage2_F1.png}} \hspace{0.2 in} \subfloat[ROC curves for Satimage2]{ \includegraphics[width=36mm, height=39mm]{Satimage2_ROC.png}} \caption{Plots for Satimage2} \end{figure} \begin{figure}[!h] \centering \subfloat[F1 score Variability plot for ETH-80]{ \includegraphics[width=36mm, height=38mm]{ETH_80_var_300.png} } \hspace{0.3 in} \subfloat[F1 score vs. Sparsity plot for ETH-80]{ \includegraphics[width=36mm, height=38mm]{eth.png}} \hspace{0.2 in} \subfloat[ROC curves for ETH-80]{ \includegraphics[width=36mm, height=39mm]{ETH_ROC.png}} \caption{Plots for ETH-80} \end{figure} \begin{figure}[!h] \centering \subfloat[F1-Score Variability plot for Internet Ads]{ \includegraphics[width=36mm, height=38mm]{Internet_Ads_var.png} } \hspace{0.3 in} \subfloat[F1-Score vs. Sparsity plot for Internet Ads]{ \includegraphics[width=36mm, height=38mm]{Internet_Ads_F1.png} } \hspace{0.2 in} \subfloat[ROC curves for Internet Ads]{ \includegraphics[width=36mm, height=39mm]{Internet_Ads_ROC2.png}} \caption{Plots for Internet Ads} \end{figure} It can be seen from the variability plots that the variance of F1-score is lower for SKPCA as compared to naive thresholding for all 5 datasets, which was also mentioned in [1] for sparse PCA. The F1-score vs. sparsity plots show that the F1-score of our SKPCA method is the closest to ordinary KPCA for all 5 datasets (and even better than it for 3 datasets), in comparison to the other 2 methods, over the entire sparsity range in consideration. Also, the AUROC value of our method is more than that of the other 2 methods for all 5 datasets and even more than that of ordinary KPCA for 3 datasets-Fashion MNIST, Satimage2 and ETH-80. We have not added the variability of the method in [3] because we found out in our experiments that the parameters involved in their algorithm are very sensitive to the chosen data subset (resulting in large differences in F1-score and sparsity for different data subsets). This is not the case with our algorithm, i.e. our algorithm does not require much tuning of parameters over randomly chosen data subsets (for the same problem) as compared to [3], for near optimal performance. Table \Rmnum{1} lists the F1-score and sparsity (i.e. the \% of non-zero coefficients per PC) obtained using our SKPCA algorithm and the method in [3] for the MNIST case over 10 randomly chosen training and test data subsets with fixed parameters for both algorithms - L1-ratio=0.7 for our algorithm and $\rho = 0.02, \lambda=0.001,\lambda_{1,k}=0.01,\epsilon^{abs}=0.01,\epsilon^{rel}=0.0001$ for the algorithm in [3]. \begin{comment} \begin{tabu} to 0.4\textwidth { | c | c | c | c | c | } \hline Trial No. & SKPCA Acc. & SKPCA of Non-zero Comps. & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline item 11 & item 12 & item 13 & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline item 11 & item 12 & item 13 & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline item 11 & item 12 & item 13 & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline item 21 & item 22 & item 23 & item 12 & item 13 \\ \hline \end{tabu} \end{comment} \begin{comment} \begin{table}[htb] \caption{hello2 testing} \label{1234} \resizebox{\columnwidth}{!}{% \begin{tabular}{c*{5}{>{$}c\langle{$}}} & \text{Trial No.} & \text{Accuracy} & \text{MLDFSP} & \text{PNDND} & \text{LDFSSVM} \\ Aa & 97.8\pm 2.3 & 97.4\pm 3.0 & 97.4\pm 3.0 & 97.0\pm 3.6 & 99.1\pm 1.8 \\ Al & 69.8\pm 9.5 & 93.6\pm 3.6 & 86.9\pm 8.4 & 91.5\pm 4.5 & 94.9\pm 5.4 \\ Av & 97.0\pm 2.9 & 99.2\pm 1.8 & 99.6\pm 1.4 & 98.3\pm 2.9 & 99.6\pm 1.4 \\ Aw & 97.8\pm 3.1 & 95.3\pm 3.1 & 96.2\pm 2.5 & 92.4\pm 3.3 & 97.5\pm 3.0 \\ Ay & 77.5\pm 7.2 & 90.2\pm 4.1 & 90.2\pm 5.7 & 88.1\pm 4.8 & 91.6\pm 4.9 \\ avg & 88.0 & 95.1 & 94.1 & 93.5 & 96.5 \end{tabular} } \end{table} \end{comment} \begin{table}[!h] \centering \includegraphics[width=0.9\linewidth]{Table_new_cropped.pdf} \caption{F1-score \& sparsity of our SKPCA algorithm and the algorithm in [3] with fixed parameters for MNIST.} \label{tab:ExcelTable} \end{table} \begin{comment} Figure Labels: Use 8 point Times New Roman for Figure labels. Use words rather than symbols or abbreviations when writing Figure axis labels to avoid confusing the reader. As an example, write the quantity ÒMagnetizationÓ, or ÒMagnetization, MÓ, not just ÒMÓ. If including units in the label, present them within parentheses. Do not label axes only with units. In the example, write ÒMagnetization (A/m)Ó or ÒMagnetization {A[m(1)]}Ó, not just ÒA/mÓ. Do not label axes with a ratio of quantities and units. For example, write ÒTemperature (K)Ó, not ÒTemperature/K.Ó \end{comment} \section{CONCLUSION} In this paper, we presented a novel algorithm for sparse KPCA which outperforms the method in [3] and is comparable to KPCA in terms of accuracy, while providing high sparsity. We also mathematically showed the validity of sparsifying KPCA with the RBF kernel, which is the first attempt in this direction, to the best of our knowledge. We showed its successful application for outlier detection on 5 real world data sets. The next step would be to explore the performance of this sparse KPCA algorithm on other applications where KPCA is employed. \addtolength{\textheight}{-12cm}

57ffc2e008a57d16c7f28440619fad7119ee8480

\section{Introduction} \sloppy The problem of the restoring of (pseudo)Riemannian manifold on prescribed curvature operator is the actual direction in the research of curvature operators. The Riemannian locally homogeneous spaces with the prescribed values of the spectrum of Ricci operator have been identified by O.\,Kowalski and S.\,Nikcevic in~\cite{KlepikovaSV:Koval-Ni1996}. The problem of~the existence of~locally homogeneous Lorentzian manifold and prescribed Ricci operator was investigated by G.\,Calvaruso and O.\,Kowalski~in~\cite{KlepikovaSV:CalKow2009}. There are also some papers about this problem in nonhomogeneous case (see~\cite{KlepikovaSV:Calvaruso2008,KlepikovaSV:Kowalski1993}). Similar results were obtained by D.N.~Oskorbin, E.D.~Rodionov, O.P. Khromova for~the~one-dimensional curvature operator and the sectional curvature operator in~the~case of three-dimensional Lie groups with left-invariant Riemannian metrics~\cite{KlepikovaSV:OR2013,KlepikovaSV:GO2013}. The main purpose of this paper is to consider the problem of the precribed sectional curvature operator $\mathcal{K}$ on the three-dimensional Lorentzian locally homogeneous manifolds. Unlike the case of the Riemannian metric, there always exist an orthonormal basis, in~which the matrix of the curvature operator is diagonal, in the case of Lorentzian metric different cases can occur known as \textit{Segre types} (see.~\cite{KlepikovaSV:Segre_types_2000}). Namely, the following cases can~occur: \begin{enumerate} \item Segre type $\{111\}$: the operator $\mathcal{K}$ has three real eigenvalues (possibly coincident), each associated to a one-dimensional eigenspace. \item Segre type $\{1z\overline{z}\}$: the operator $\mathcal{K}$ has one real and two complex the conjugate eigenvalues. \item Segre type $\{21\}$: the operator $\mathcal{K}$ has two real eigenvalues (possibly coincident), the first of which has algebraic multiplicity 2, each associated to a one-dimensional eigenspace. \item Segre type $\{3\}$: the operator $\mathcal{K}$ has one real eigenvalue of algebraic multiplicity 3, associated one-dimensional eigenspace. \end{enumerate} \section{Three-dimensional homogeneous Loretzian manifolds} Let $(M,g)$ be three-dimensional homogeneous manifold, with the Lorenzian metric $g$ of~signature $\left(+,+,-\right)$. We denote by $\nabla$ its Levi-Civita connection and by $R$ its curvature tensor, which defined by following $$R\left(X,Y\right)Z=\left[\nabla_Y,\nabla_X\right]Z+\nabla_{\left[X,Y\right]}Z.$$ The Lorenzian metric $g$ induces a scalar product $\langle\cdot,\cdot\rangle$ in the bundle $\Lambda^2 M$ by the rule $$ \left\langle X_1 \wedge X_2, Y_1 \wedge Y_2 \right\rangle = \mathrm{det}\left(\left\langle X_i,Y_i\right\rangle\right). $$ The curvature tensor $R$ at any point can be considered as an operator $\mathcal{K} \colon \Lambda^2 M \to \Lambda^2 M$, called the \textit{sectional curvature operator} and defined by the equation \begin{equation*}\label{eq:def_sec_curv} \left\langle X \wedge Y,\mathcal{K}\left(Z \wedge T\right)\right\rangle = R\left(X, Y, Z, T\right). \end{equation*} The studying of curvature operators on three-dimensional locally homogeneous Lorentzian spaces is based on the following fact, which was proved by G.\,Calvaruso in~\cite{KlepikovaSV:Calvaruso2007}. \begin{Theorem} Let $(M,g)$ be a three-dimensional connected, simply connected, complete locally homogeneous Lorentzian manifold. Then, either $(M,g)$ is locally symmetric, or it is~locally isometric to a three\nobreakdash-dimensional Lie group equipped with a left-invariant Lorentzian metric. \end{Theorem} Further classification results for three-dimensional Lorentzian Lie groups was obtained~in~\cite{KlepikovaSV:RSCh2006}. \begin{Theorem} Let $G$ be a three-dimensional Lie group with left-invariant Lorentzian metric. Then \begin{itemize} \item If $G$ is unimodular, then there exists a pseudo-orthonormal frame field $\left\{e_1, e_2, e_3\right\}$, such that the metric Lie algebra of $G$ is one of the following: \begin{enumerate} \item \begin{equation*} \mathcal{A}_1: \begin{split} [e_1,e_2] &= \lambda_3e_3, \\ [e_1,e_3] &= -\lambda_2e_2, \\ [e_2,e_3] &= \lambda_1e_1, \end{split} \end{equation*} with $e_1$ timelike; \item \begin{equation*} \mathcal{A}_2: \begin{split} [e_1,e_2] &= \left(1-\lambda_2\right)e_3-e_2, \\ [e_1,e_3] &= e_3-\left(1+\lambda_2\right)e_2, \\ [e_2,e_3] &= \lambda_1e_1, \end{split} \end{equation*} with $e_3$ timelike; \item \begin{equation*} \mathcal{A}_3: \begin{split} [e_1,e_2] &= e_1-\lambda e_3, \\ [e_1,e_3] &= -\lambda e_2-e_1, \\ [e_2,e_3] &= \lambda_1e_1+e_2+e_3, \end{split} \end{equation*} with $e_3$ timelike; \item \begin{equation*} \mathcal{A}_4: \begin{split} [e_1,e_2] &= \lambda_3e_2, \\ [e_1,e_3] &= -\beta e_1-\alpha e_2, \\ [e_2,e_3] &= -\alpha e_1+\beta e_2, \end{split} \end{equation*} with $e_1$ timelike and $\beta \ne 0$. \end{enumerate} \item If $G$ is non-unimodular, then there exists a pseudo-orthonormal frame field $\left\{e_1, e_2, e_3\right\}$, such that the metric Lie algebra of $G$ is one of~the~following: \begin{enumerate} \item \begin{equation*} \mathcal{A}: \begin{split} [e_1,e_2] &= 0, \\ [e_1,e_3] &= \lambda \sin\varphi\,e_1-\mu\cos\varphi\,e_2, \\ [e_2,e_3] &= \lambda \cos\varphi\,e_1+\mu\sin\varphi\,e_2, \end{split} \end{equation*} with $e_3$ timelike and $\sin\varphi\ne 0$, $\lambda+\mu \ne 0$, $\lambda \geqslant 0$, $\mu \geqslant 0$; \item \begin{equation*} \mathcal{B}: \begin{split} [e_1,e_2] &= 0, \\ [e_1,e_3] &= te_1-se_2, \\ [e_2,e_3] &= pe_1+qe_2, \end{split} \end{equation*} with $\left\langle e_2,e_2\right\rangle=-\left\langle e_1,e_3\right\rangle=1$ and otherwise zero, and $q \ne t$; \item \begin{equation*} \mathcal{C}_1: \begin{split} [e_1,e_2] &= 0, \\ [e_1,e_3] &= se_1+pe_2, \\ [e_2,e_3] &= pe_1+qe_2, \end{split} \end{equation*} with $e_2$ timelike and $q\ne s$; \item \begin{equation*} \mathcal{C}_1: \begin{split} [e_1,e_2] &= 0, \\ [e_1,e_3] &= qe_1-re_2, \\ [e_2,e_3] &= pe_1+qe_2, \end{split} \end{equation*} with $e_2$ timelike and $q\ne 0$, $p+r \ne 0$. \end{enumerate} \end{itemize} \end{Theorem} \begin{Remark} There are exactly six nonisomorphic three-dimensional unimodular Lie algebras and the corresponding types of three-dimensional unimodular Lie groups (see~\cite{KlepikovaSV:Milnor1976}). All of them are listed in the Table~\ref{tab_PastukhovaSV:Chibrikova2006} together with conditions on structure constants for~which the Lie algebra has this type. If there is a ``$-$'' in the Table~\ref{tab_PastukhovaSV:Chibrikova2006} at the intersection of the row, corresponding to the Lie algebra, and the column, corresponding to the type, then it means that this type of the basis is impossible for given Lie algebra. For~the~case of Lie algebra $\mathcal{A}_1$ we give only the signs of the triple $\left(\lambda_1,\lambda_2,\lambda_3\right)$ up to reorder and sign change. \end{Remark} \begin{table*}[ht!] \caption{Three-dimensional unimodular Lie algebras} \label{tab_PastukhovaSV:Chibrikova2006} \begin{center} \small \begin{tabular}[c]{|c|c|c|c|c|} \hline \multirow{2}*{Lie algebra} & \multicolumn{4}{c|}{Restrictions on the structure constants} \\ \cline{2-5} & $\mathcal{A}_1$ & $\mathcal{A}_2$ & $\mathcal{A}_3$ & $\mathcal{A}_4$ \\ \hline $su(2)$ & $\left(+,+,+\right)$ & $-$ & $-$ & $-$ \\ \hline $sl(2,\mathbb{R})$ & $\left(+,+,-\right)$ & $\lambda_1 \ne 0$, $\lambda_2 \ne 0$ & $\lambda \ne 0$ & $\lambda_3 \ne 0$ \\ \hline $e(2)$ & $\left(+,+,0\right)$ & $-$ & $-$ & $-$ \\ \hline $e(1,1)$ & $\left(+,-,0\right)$ & \parbox[c]{80pt}{\centering $\lambda_1 = 0$, $\lambda_2 \ne 0$ \\ or \\ $\lambda_1 \ne 0$, $\lambda_2 = 0$} & $\lambda = 0$ & $\lambda_3 = 0$ \\ \hline $h$ & $\left(+,0,0\right)$ & $\lambda_1 = 0$, $\lambda_2 = 0$ & $-$ & $-$ \\ \hline $\mathbb{R}^{3}$ & $\left(0,0,0\right)$ & $-$ & $-$ & $-$ \\ \hline \end{tabular} \end{center} \end{table*} \begin{Remark} We note that similar bases was also constructed by G. Calvaruso, L.A.~Cordero and P.E. Parker in~\cite{KlepikovaSV:Calvaruso2007,KlepikovaSV:CP1997}. \end{Remark} The following classification result for the case of three-dimensional Lorentzian locally symmetric space was proved in~\cite{KlepikovaSV:Calvaruso2007}. \begin{Theorem} \label{th:loc_sym} A connected, simply connected three-dimensional Lorentzian locally symmetric space $(M,g)$ is locally isometric to \begin{enumerate} \item a Lorentzian space form $\mathbb{R}^3_1$, $\mathbb{S}^3_1$ or $\mathbb{H}^3_1$ (with zero, positive and negative sectional curvature respectively), or \item a direct product $\mathbb{R}\times\mathbb{S}^2_1$, $\mathbb{R}\times\mathbb{H}^2_1$, ${\mathbb{S}^2\times\mathbb{R}_1}$ or $\mathbb{H}^2\times\mathbb{R}_1$, or \item a space with a Lorentzian metric $g$, which admitted a local coordinate system $(u_1,u_2,u_3)$ such, that the metric tensor has the following form $$ g= \begin {pmatrix} 0 & 0 & 1 \\ 0 & \varepsilon & 0 \\ 1 & 0 & f(u_2,u_3) \end {pmatrix},$$ where $\varepsilon=\pm 1$, $f(u_2,u_3)=u_2^2\alpha+u_2\beta(u_3)+\xi(u_3)$, $\alpha\in\mathbb{R}$, $\beta$, $\xi$ are arbitrary smooth functions. \end{enumerate} \end{Theorem} \section{Three-dimensional Lorentzian Lie groups} Further, by ``a three-dimensional Lorenzian Lie group $\left(G,\mathfrak{g}\right)$'' we shall mean a~three\nobreakdash-di\-men\-sional Lie group $G$, which equipped with a left-invariant Lorentziam metric~$g$ and having metric Lie algebra $\mathfrak{g}$. Now we can prove the following \begin{Theorem} \label{th:A2} A three-dimensional unimodular Lorentzian Lie group $\left(G,\mathcal{A}_2\right)$ with the~sectional curvature operator $\mathcal{K}$ exist if and only if \begin{enumerate} \item $\mathcal{K}$ has the Segre type $\{111\}$ with the eigenvalues $ k_1=- k_2=- k_3 \geqslant 0$ up to renumeration, or \item $\mathcal{K}$ has the Segre type $\{12\}$ with the eigenvalues \begin{enumerate} \item $ k_1 = k_2 =0$, or \item $ k_2 <0$. \end{enumerate} \end{enumerate} \end{Theorem} \begin{proof} In this case the matrix of the sectional curvature operator $\mathcal{K}$ has the following form \begin{equation*} \mathcal{K}=\begin {pmatrix} \frac34\lambda_1^2-\lambda_1\lambda_2 & 0 & 0 \\ 0 & 2\lambda_2-\lambda_1-\frac14\lambda_1^2 & 2\lambda_2-\lambda_1 \\ 0 & -2\lambda_2+\lambda_1 & -2\lambda_2+\lambda_1-\frac14\lambda_1^2 \end {pmatrix}. \end{equation*} If $\lambda_1=2\lambda_2$, then the matrix of the sectional curvature operator has the~diagonal form with the eigenvalues $k_1=-k_2=-k_3\geqslant0$. Else, the matrix of the sectional curvature operator $\mathcal{K}$ has the following Jordan form: $$\mathcal{K}=\begin {pmatrix} -\lambda_1\lambda_2+\frac34\lambda_1^2 & 0 & 0 \\ 0 & -\frac14\lambda_1^2 & 1 \\ 0 & 0 & -\frac14\lambda_1^2 \end {pmatrix},$$ and the eigenvalues are equal to \begin{gather*} k_1= -\lambda_1\lambda_2+\frac34\lambda_1^2, \quad k_2= -\frac14\lambda_1^2 \leqslant 0. \end{gather*} If $ k_2=0$, then $\lambda_1=0$ and all of the eigenvalues are equal to zero. Suppose that $ k_2<0$. Then, this follows that $\lambda_1 = \pm2\sqrt{- k_2}$. Expressing~$\lambda_2$, we find $$ \lambda_2 = \mp \frac{ k_1+3 k_2}{2\sqrt{- k_2}}. $$ \end{proof} The remaining cases are concerned in a similar way. \section{Three-dimensional Lorentzian locally symmetric spaces} The Theorem~\ref{th:loc_sym} allows us to divide the problem of studying the curvature operators on~three-dimensional locally symmetric Lorentzian manifolds by~three subtasks. At~the~same time, it is obvious that the sectional curvature operator $\mathcal{K}$ is diagonalizable for~Lorentzian manifolds of constant sectional survature $\mathbb{R}^3_1$, $\mathbb{S}^3_1$ and $\mathbb{H}^3_1$ (i.e. $\mathcal{K}$ has the Segre type~$\{111\}$) and $\mathcal{K}$ has three equal eigenvalues (zero, positive or negative respectively). In the case of direct products (case 2 of the Theorem~\ref{th:loc_sym}) the sectional curvature operator $\mathcal{K}$ has the Segre type~$\{111\}$ with two zero and third non-zero eigenvalues. Therefore, only the case 3 of Theorem~\ref{th:loc_sym} is of interest, in which the~metric tensor has the following form in local coordinate system $$ g= \begin {pmatrix} 0 & 0 & 1 \\ 0 & \varepsilon & 0 \\ 1 & 0 & f(u_2,u_3) \end {pmatrix},$$ where $\varepsilon=\pm 1$, $f(u_2,u_3)=u_2^2\alpha+u_2\beta(u_3)+\xi(u_3)$, $\alpha\in\mathbb{R}$, $\beta$, $\xi$ are arbitrary smooth functions. Calculating the matrix of the sectional curvature operator $\mathcal{K}$, we have \begin{equation*} \mathcal{K}= \begin {pmatrix} 0 & 0 & \frac{\alpha}{\varepsilon} \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end {pmatrix}. \end{equation*} Therefore, either the sectional curvature operator $\mathcal{K}$ has the~Segre type $\{111\}$ with the~eigenvalues $k_1=k_2=k_3=0$ for $\alpha=0$, or $\mathcal{K}$ has the Segre type~$\{12\}$ with the~eigenvalues $k_1=k_2=0$ for $\alpha\ne0$. Hence, the following theorem holds. \begin{Theorem} \label{th:K_loc_sym} A connected, simply connected three-dimensional Lorentzian locally symmetric space with the sectional curvature operator $\mathcal{K}$ exist if and only if \begin{enumerate} \item $\mathcal{K}$ has the Segre type $\{111\}$ with the equal eigenvalues, or \item $\mathcal{K}$ has the Segre type $\{111\}$ with the two zero eigenvalues and one nonzero, or \item $\mathcal{K}$ has the Segre type $\{12\}$ with the zero eigenvalues. \end{enumerate} \end{Theorem} \section{Sectional curvature operator of locally homogeneous Lorentzian 3-manifolds} In this section, using the results of the previous sections, we determine under which conditions the different Segre types occur for the sectional curvature operator of a~three\nobreakdash-dimensional locally homogeneous Lorentzian manifold. Next theorems follows from the results on the cases of metric Lie groups and of the locally symmetric spaces. \begin{Theorem} \label{th:loc_hom_nd} A connected, simply connected three-dimensional Lorentzian locally homogeneous manifold $(M, g)$ with non-diagonalizable sectional curvature operator $\mathcal{K}$ exist if and only if $\mathcal{K}$ satisfies one of the following conditions: \begin{enumerate} \item $\mathcal{K}$ has the Segre type $\{12\}$ an \begin{enumerate} \item eigenvalues equal to zero, or \item $k_2<0$; \end{enumerate} \item $\mathcal{K}$ has the Segre type $\{3\}$ with negative eigenvalue; \item $\mathcal{K}$ has the Segre type $\{1z\overline{z}\}$ and \begin{enumerate} \item complex eigenvalues have negative real part, or \item $0 \leqslant \frac{k_2+ k_3}{2} < -k_1$. \end{enumerate} \end{enumerate} \end{Theorem} \begin{Theorem} A connected, simply connected three-dimensional Lorentzian locally homogeneous manifold $(M, g)$ with sectional curvature operator $\mathcal{K}$ with Segre type~$\{111\}$ exist if and only if eigenvalues $k_1$, $k_2$, $k_3$ satisfy one (or more than one) of the following conditions: \begin{enumerate} \item all eigenvalues are equal to each other; \item two eigenvalues are equal to zero and third is nonzero; \item exactly two of $k_1 + k_2$, $k_1 + k_3$ and $k_2 + k_3$ are equal to zero; \item $\left(k_1 + k_2\right)\left(k_1 + k_3\right)\left(k_2 + k_3\right)<0$; \item up to renumeration $$k_2k_3 \leqslant k_1^2<\left(\frac{k_2+k_3}{2}\right)^2\; \text{ and } \;k_1<\frac{k_2+k_3}{2};$$ \item up to renumeration $$k_2<0,\; k_3<0,\; \left|k_1\right|\leqslant\sqrt{k_2k_3};$$ \item up to renumeration $$k_1 < -\left|\frac{k_2+k_3}{2}\right|.$$ \end{enumerate} \end{Theorem}

204bf57431fb44ea8d1c6f789aa00bf49c18c45f

\section{Introduction} Verification and validation (V\&V) of models and simulations of astrophysical phenomena present challenges because the problem of studying these phenomena is largely one of indirectly observing multi-scale, multi-physics events. Other aspects of astrophysics also contribute challenges. The enormous length scales of astrophysical objects and vast distances to most astrophysical events preclude ready experimental access, limiting the availability of validation data. As with a great many applications, models suffer from epistemic uncertainty in the underlying basic physics (e.g.\ turbulence, fluid instabilities, and nuclear reaction rates), which is difficult to control and assess in simulations incorporating multiple interacting physical processes. The large range of length and time scales in many astrophysical problems frequently necessitates capturing sub-grid-scale physics within simulations, relevant examples being thermonuclear flames and turbulent combustion. The requirement of the development of sub-grid-scale models for these physical processes obviously introduces an additional level of complexity to V\&V. Finally, the magnitude of the requisite computations for astrophysical events means that even with sub-grid-scale models, simulations may only capture the bulk effect of the underlying physics and some properties such as detailed compositions must be obtained by post-processing the simulation results with augmenting, higher-fidelity routines. Even with these issues, V\&V are vital parts of computational astrophysics as with any research domain. We present two studies aimed at validating components of Flash, a freely available, parallel, adaptive mesh simulation code used for modeling astrophysical phenomena and other applications. We first present a study of validating the hydrodynamics routines in Flash with experiments designed to replicate the high energy density environments of astrophysics and probe the underlying physics. The investigation formally addresses the issues of concern in validating hydrodynamics and serves as a well-controlled case study. The second study we present addresses physics that is difficult to include in whole-star simulations, due to limits in computing power, but that can be incorporated with approximate models and also calculated by post-processing simulation results. The problem is thermonuclear combustion and describing the overall reactions while including minimal nuclear species, and this work addresses the issue of comparing prohibitively expensive detailed models and simpler models that allow three-dimensional simulations. As we will describe below, the challenges to astrophysical validation made parts of our study incomplete. The effort, however, was rewarding and very much worth the investment. Verification tests quantified the accuracy of code modules for problems with an analytic or accepted result, and the regular application of these tests serves for regression testing as the code is developed. Validation tests, though incomplete, demonstrated reasonable agreement between experiment and simulation for the case of the hydrodynamics study. Comparison between models of increasing sophistication allowed us to quantify the trade-off between fidelity of the method and expense. These studies all led to a deeper understanding of the underlying physics, and while we cannot say the modules and code were completely ``validated," the process greatly increased our confidence in the results. \section{Approach to Verification and Validation} \label{sec:vvapproach} Our methods for V\&V largely follow accepted practices from the fluid dynamics community \citep[see Ch.\ 26 by Roache in this volume]{1998aiaa,roache1998verification,roache98,oberkampf2010verification}. We adopt the following definitions (based on definitions from the American Institute of Aeronautics and Astronautics \citep{1998aiaa}). \begin{quotation} \begin{trivlist} \item Model: A representation of a physical system or process intended to enhance our ability to understand, predict, or control its behavior. \item Simulation: The exercise or use of a model. (That is, a model is used in a simulation). \item Verification: The process of determining that a model implementation accurately represents the developer's conceptual description of the model and the solution of the model. \item Validation: The process of determining the degree to which a model is an accurate representation of the real world from the perspective of the intended uses of the model. \item Uncertainty: A potential deficiency in any phase or activity of the modeling process that is due either to a lack of knowledge (epistemic uncertainty or incertitude) or due to variability or inherent randomness (aleatory uncertainty). \item Error: A recognizable deficiency in any phase or activity of modeling that is not due to lack of knowledge. \item Prediction: Use of a model to foretell the state of a physical system under conditions for which the model has not been validated. \item Calibration: The process of adjusting numerical or physical modeling parameters in the computational model for the purpose of improving agreement with experimental data. \end{trivlist} \end{quotation} Our definition of uncertainty differs from the original definition of the AIAA in that we expand the definition of uncertainty to also include aleatory uncertainty~\citep[see][and references therein]{uqproceedings2018,hoffmanetall2018}. Another perspective comes from \citet{roache98}, who offers a concise, albeit informal, summary of V\&V terminology: \begin{quotation} First and foremost, we must repeat the essential distinction between Code Verification and Validation. Following \citet{boehm81} and \citet{blottner90}, we adopt the succinct description of ``Verification" as ``solving the equations right", and ``Validation" as ``solving the right equations". The code author defines precisely what partial differential equations are being solved, and convincingly demonstrates that they are solved correctly, i.e. usually with some order of accuracy, and always consistently, so that as some measure of discretization (e.g.\ the mesh increments) $\Delta \rightarrow 0$, the code produces a solution to the continuum equations; this is Verification. Whether or not those equations and that solution bear any relation to a physical problem of interest to the code user is the subject of Validation. \end{quotation} Roache also notes that a ``code" cannot be validated, but only a calculation or range of calculations can be validated. Roache also makes a distinction between verifying a code and verifying a calculation, noting that ``use of a verified code is not enough." We also adhere to this explication of V\&V terminology and note that following Roache, validation can be described as probing the range of validity of a code or model~\citep{calder.fryxell.ea:on}. Our approach to verification consists of testing simulation results against analytic or benchmarked solutions and quantifying the error. The comparisons typically consist of simulations performed at increasing spatial and/or temporal resolutions to confirm convergence of the simulation to the correct answer. Details of these tests have appeared in the literature, and many of the tests are incorporated into automated regression testing of Flash \citep{calder.fryxell.ea:on, weirsetal2005a, weirsetal2005b, dwarkadasetal2005, hearnetal2007, dubeyetal2009, dubeyetal2015}. We validate by performing similar tests against data from experiments designed to replicate astrophysical environments. We take a hierarchical approach to validation, beginning by isolating the basic underlying physics and testing how well simulations capture it. We then devise tests of aggregate problems that capture the expected behavior of the astrophysical events. In the case of sub-grid models or post-processed results, we simulate simple problems with these models and compare against either actual validation data or direct numerical simulations. As with verification, we perform convergence tests, though as we describe below the process of demonstrating convergence is difficult for some fluid dynamics problems. Another aspect of our testing concerns quantifying error on the adaptive simulation mesh (described below). Our approach is to test solutions on the finest simulation mesh against data or a solution, but the methodology for quantitatively comparing the solution at the different resolutions of an adaptive mesh is incomplete \citep{li2010,crash,shuetal2011,liwood2017}. We typically check for consistency between simulations on an adaptive mesh and simulations of the same problem on a fully-refined mesh while quantifying the accuracy of the solution on the fully-refined mesh \citep{calder.fryxell.ea:on}. Also, in addition to problems characterizing solutions on an adaptive mesh, just simulating fluids at the extreme Reynolds numbers of astrophysics on adaptive meshes presents challenges~\citep{kritsuknormanpadoan2006,mitran2009}. We describe the difficulties of simulating extreme Reynolds number flow in the discussion of our hydrodynamics method below. We close discussion of our approach to V\&V with a general note on the role of validation in astrophysics. Because of the literally astronomical distances to astrophysical events and extreme conditions involved, experimentally accessing astrophysical phenomena or even just replicating the environments of astrophysics is difficult. Thus one cannot readily perform validation experiments, which typically leads to an incomplete process of validation. Simulations of astrophysical events are therefore generally in the realm of prediction, that is, foretelling the state of a physical system under conditions for which the model has not been validated. Despite this, the process of V\&V in astrophysics serves to build confidence in these predictions even if one cannot conclude that simulations or codes are ``validated." \section{Simulation Instruments} Our principal simulation instrument is the Flash code, which we use for simulating astrophysical events. Fundamentally, Flash simulates problems of fluid dynamics and consists of solvers for hydrodynamics and the additional physics of astrophysical events (described below). With Flash, we construct the numerical implementation of our conceptual model of the astrophysical event, and the act of simulating is the exercise of the model. We note that the exercise of a model is far more than just solving a set of differential equations. Multi-physics applications like astrophysics combine multiple solvers, each of which may rely on possibly uncontrolled assumptions~\cite[See][for a thorough discussion]{winsberg2010science}. For this reason, we take the hierarchical approach to validation of modules in Flash mentioned above. Our second instrument is a nucleosynthetic post-processing toolkit used in tandem with Flash. In the case of supernovae, comparison to observations requires the calculation of light curves (the intensity of light from the object as a function of time) and spectra. However, the yield of a particular element, titanium for example, may be critical for accurate spectra, but mostly unimportant to the energy release. Many elements fall into this category, so that the computation of the explosion is much less expensive when split into two stages. The energy release and explosion is computed with a small number of species in Flash, and is followed by post-processing to obtain all important species. The post-processing tools we present below apply state-of-the-art nuclear reaction networks to Lagrangian thermodynamic histories sampled from the Flash simulation. The resulting abundances are used to calculate light curves and spectra~\citep[e.g.][]{milesetal2016}. \subsection{The Flash Code} The simulation instrument we use for modeling astrophysics events is the Flash code, developed at the University of Chicago \citep{Fryxetal00, calder.curtis.ea:high-performance, dubeyetal2009,dubeyetal2013,dubeyetal2014}. Flash is a parallel, adaptive-mesh, hydrodynamics plus additional physics code originally designed for the compressible fluid flows associated with astrophysics. Flash incorporates multiple hydrodynamics methods \citep{Fryxetal00,LEE2009952,LEE2013269,Lee2017230,leeetall2017} coupled with modules for the requisite additional physics of the applications. In particular, Flash has undergone considerable development for high energy density physics applications \citep{Tzeferacosetal2015}. The hydrodynamics modules solve the Euler equations of compressible hydrodynamics, shown here with gravitational sources as would apply to a self-gravitating problem such as a star. \begin{displaymath} \frac{\partial \rho}{\partial t} + {\bf \nabla} \cdot (\rho {\bf v}) = 0 \end{displaymath} \begin{displaymath} \frac{\partial \rho {\bf v}}{\partial t} + {\bf \nabla} \cdot \left(\rho {\bf v}{\bf v}\right) + {\bf \nabla}P = \rho {\bf g} \end{displaymath} \begin{displaymath} \frac{\partial \rho E}{\partial t} + {\bf \nabla} \cdot \left(\rho E + P\right) {\bf v} \: = \: \rho {\bf v} \cdot {\bf g} + S \; . \end{displaymath} Here $\rho$ is the mass density, ${\bf v}$ is the velocity, $P$ is the pressure, $E$ is the internal energy of the gas, ${\bf g}$ is the gravitational acceleration, and $S$ represents any additional source. The system is closed by an equation of state of the form \begin{displaymath} P = P\left(\rho,E\right) \end{displaymath} and Flash offers choices for particular applications. Flash calculates the acceleration due to gravity from the gravitational potential, \begin{displaymath} {\bf g} = - {\bf \nabla} \Phi \; . \end{displaymath} which is calculated by solving the Poisson equation for self-gravity \begin{displaymath} \nabla^2 \Phi\left({\bf r}\right) = 4\pi G \rho\left({\bf r}\right) \; . \end{displaymath} Here $\phi$ is the gravitational potential and $G$ is Newton's gravitational constant. Flash also tracks different material species by advecting mass scalars, \begin{displaymath} \frac{\partial X\rho}{\partial t} + {\bf \nabla} \cdot (X \rho {\bf v}) = 0 \; , \end{displaymath} where $X$ is the mass fraction of a given species~\citep{Fryxetal00}. Our first validation example addressed the Flash hydrodynamics module (without gravity) for the case of experiments involving fluid instabilities thought to occur during one class of stellar explosions known as a core collapse supernova~\citep{fryxellarnettmueller1991}. The particular hydrodynamic module in Flash used for this study is based on the PROMETHEUS code \citep{FryxMuelArne89} and evolves the Euler equations in one, two, or three dimensions using a modified version of the Piecewise-Parabolic Method (PPM) \citep{Colella1984The-Piecewise-P}. The implementation allows use of general equations of state as is required for simulating stellar material \citep{colellaglaz85}, but this capability was not used in the validation example. PPM is a higher-order version of the method developed by Godunov~\citep{godunov59,godunov62}, a finite-volume conservation scheme that solves the Riemann problem at the interfaces of the control volumes to compute fluxes into each volume. The conserved fluid quantities are treated as cell averages that are updated by the fluxes at the interfaces. This treatment has the effect of introducing explicit non-linearity into the difference equations and permits the calculation of sharp shock fronts and contact discontinuities without introducing significant non-physical oscillations into the flow. In addition, PPM utilizes a dissipative shock capturing scheme to further stabilize shocks and contact discontinuities, and is thus not directly solving the Euler equations \citep{majda84,winsberg2010science}. The adaptive mesh of Flash is block structured and is supported primarily through the Paramesh Library~\citep{paramesh1,paramesh2}, though it is under the process of migrating to the AMReX library~\citep{amrex}. The view of AMReX from other units in the Flash code will remain similar to that of Paramesh, and in the near future the two packages will be available as alternative implementations of the Grid unit. Later, the support for Paramesh may be dropped if it becomes too inefficient on newer platforms. \subsection{The Post-processing Toolkit} The nucleosynthetic post-processing toolkit uses the recorded Lagrangian history of fluid elements to compute the yield of nuclides (elements and their isotopes) produced in a stellar explosion \citep{travaglioetal2004,townetal15}. The Lagrangian thermodynamic history is determined by integrating the position of a conceptual microscopic fluid element by \begin{displaymath} {\bf r}(t) = {\bf r}_0 + \int_0^t {\bf v}({\bf r}, t')\,dt'\ , \end{displaymath} where ${\bf r}_0$ is the initial position and ${\bf v}({\bf r},t)$ is the velocity field as computed by the hydrodynamic simulation. This conceptual fluid element is often called a particle because it moves as a fluid-embedded particle would. From the resulting ${\bf r}(t)$, it is possible to also record the thermodynamic state, namely $T(t) = T({\bf r}(t))$ and $\rho(t)=\rho({\bf r}(t))$, the temperature and density, respectively. Such recorded histories are often called tracks or trajectories because they represent how the fluid element evolves in location and thermodynamic state space as a function of time. Nucleosynthetic post-processing is performed in order to obtain the composition of material after it is processed by combustion and ejected. Composition is parameterized by abundances of various species quantified as the fraction of a unit of mass that is in the form of a particular species. For example, the fraction, by mass, that is in the form of $^{12}$C, may be written $X_{^{12}\rm C}$, and must be between 0 and 1. The abundances are found in post-processing by integrating \begin{displaymath} \label{eq:abund_integral} X_i(t) = X_{i,0} + \int_0^t \dot X_i( \rho(t'), T(t') )\, dt'\ , \end{displaymath} where $\dot X_i(\rho,T)$ are determined by the density and temperature-dependent reaction rates for processes which involve species $i$. Any given specie is typically involved in multiple reactions, forming a network that is used to evaluate each rate. The end of the necessary integrations is typically well defined. As the star expands $T$ and $\rho$ fall until most reactions will become very slow compared to the time being simulated, effectively freezing out. Consideration of further evolution, typically radioactive decay, may be necessary depending on the usage of the resulting abundances. These integrations are typically performed for a large number of tracks which sample the ejected material by a suitably distributed choice of their initial positions ${\bf r}_0$. \subsection{Simulating Reactive Flow} With both Flash and the post-processing toolkit, the goal of simulations is to capture the evolution of stellar material during the course of an astrophysical event. Because stars are essentially self-gravitating gas, the interiors of stars are well described by the equations of fluid flow. During an astrophysical event, thermonuclear reactions drive the evolution by changing the composition and by releasing energy, which changes thermodynamic conditions like pressure and density. This combustion typically occurs in a relatively small region of space, e.g.\ a thin flame, that may be difficult to resolve in simulations of the event. The two validation examples we present address the two principal parts- fluid flow, including shocks and fluid instabilities, and the evolution of the composition. For fluid dynamics problems, there are two fundamental classes of simulation distinguished by whether or not the scales of the numerical grid can resolve viscous effects~\citep[][and references therein]{calder.fryxell.ea:on,winsberg2010science}. Simulations that can resolve viscous effects are said to be ``Direct Numerical Simulations,'' while those that cannot and rely on a (possibly uncontrolled) sub-grid-scale model for viscous effects are referred to as ``Large Eddy Simulations.'' An eddy is a fluid current whose flow direction differs from that of the general flow, and the motion of the fluid is the net result of the movements of the eddies that compose it~\citep{eddy_britanica}. Large eddy simulations do not resolve either the explicit viscosity of the fluid or the contribution to the viscosity from eddies on unresolved scales~\citep [][and references therein]{Fureby1996,ZHIYIN201511}. The issue of convergence of a solution for fluid flows is not as simple as it might seem. The enormous size of objects means that astrophysical regimes typically have Reynolds numbers far in excess of the Reynolds numbers of terrestrial flows, which are themselves higher than can be readily captured in hydrodynamics simulations. Even when run on contemporary supercomputers, simulations cannot capture the possibly $\ge 10^{8}$ Reynolds numbers of astrophysical flows making direct numerical simulations impossible. Thus simulations of astrophysical events are large eddy simulations that can either rely on sub-grid-scale models for turbulent flow or just allow the intrinsic numerical diffusion of the hydrodynamics method to set the limiting Reynolds number. This latter case, know as Implicit Large Eddy Simulation (ILES), is frequently applied and is the approach taken in the studies presented here. In ILES, changing the resolution changes the effective viscosity and hence the Reynolds number, which changes the problem itself and leads to the question of convergence of results with resolution. Considerable study has gone into determining the validity of this approach \citep{margolinrider2002,grinstein2007implicit, margolinshashkov2008, margolin2014}. As our results show, large eddy simulations may not demonstrate convergence of a solution with resolution. \section{Validation Examples} As of this writing, Flash has had 20 years of development by generations of scientists. Much of this effort has been subjected to rigorous V\&V \citep{calder.fryxell.ea:on, cise_val_2004, weirsetal2005a, weirsetal2005b, dwarkadasetal2005, hearnetal2007, dubeyetal2009, dubeyetal2015,townetal15}. In this contribution, we present two examples of validating the Flash code and post-processing toolkit for astrophysical applications. The first example is from early work comparing simulations to laboratory experiments addressing fluid instabilities in high energy density environments similar to the interiors of stars. The second example is ongoing work on computing reaction products in three-dimensional simulations of type Ia supernovae. This study includes comparison between methods for use in the simulations of the events and for calculating detailed abundances from the density and temperature histories of Lagrangian tracers. While this contribution describes two examples of the V\&V efforts for the Flash code, we note that V\&V efforts continue as the capabilities and applications of Flash evolve. A recent survey of software engineering practice in scientific computing includes Flash as a case study and offers an independent perspective on the development of Flash \citep{Storer:2017}. \subsection{Overview of Flash Problems} The Flash code was originally designed to investigate astrophysical thermonuclear flashes, explosive events powered by thermonuclear fusion. These events all involve a close binary star system with matter being transferred (accreted) onto a compact star (either a white dwarf or a neutron star) from a companion~\citep{rosneretal2000}. The three flash problems originally addressed by Flash were type I x-ray bursts~\citep{zingaleetal2001}, classical novae~\citep{alexakisetal2004}, and type Ia supernovae~\citep{plewacalderlamb2004,townsleyetal2007}. X-ray bursts occur when a thermonuclear runaway occurs in a thin $\sim10-100$ m layer of H- or He-rich fuel accreted onto the surface of a neutron star. The radius of the underlying neutron star my be inferred from observations and thereby allow constraints on the properties of dense matter. Classical novae occur when a thermonuclear $\sim 10^4$ m thick layer of H-rich material similarly explodes. In this case, material from the explosion is unbound and these events are thought to synthesize some intermediate-mass elements found in the galaxy. Type Ia supernovae are thought to occur when a pair of white dwarf star merge and/or when a white dwarf accretes enough mass to ignite fusion in the core. In this case, enough energy is added to overcome the gravitational binding and the star is completely disrupted, producing a bright explosion that may be used as an indicator for cosmological distances. (See references in above works for literature on each astrophysical topic, and \citealt{calderetal2013} for an overview of ongoing investigation of Type Ia Supernovae.) As mentioned above, these problems involve reactive flow, and in all cases there is a vast difference between the length scale of the combustion front and the astrophysical object. Hence the need for sub-grid-scale models. Fluid instabilities that may influence the burning rate are also of particular importance~\citep{calderetal2007,zhangetal2007,townetal15}. Accordingly, the validation examples we present address problems of combustion and fluid instabilities. \subsection{Shocks and Fluid Instabilities} The high energy density environments of intense lasers interacting with matter are similar to the interiors of stars, and experiments offer opportunities for a quantitative comparison between data and simulation not possible with observations of astrophysical phenomena. The validation study we present was performed by a collaboration between Flash developers and experimentalists working at the Omega laser at the University of Rochester~\citep{soures96,boehly95, bradley98}. The experiment chosen for the study involved a shock propagating through a multi-layer target with layers of decreasing density and was designed to produce hydrodynamic instabilities thought to arise during an astrophysical event known as a core collapse supernova explosion~\citep{arnett89,fryxellarnettmueller1991}. While this type of supernova is not a thermonuclear flash problem, much of the constituent physics is the same, allowing this experiment to serve for validation. The decreasing-density configuration is subject to the Richtmyer-Meshkov instability that occurs when a shock propagates though a material interface with decreasing density~\citep{richtmyer60,meshkov69}. The configuration is also subject to the Rayleigh-Taylor instability~\citep{taylor+50,chandra+81}, which develops after the passage of the shock and subsequently dominates instability growth. Interest in the problem of fluid instabilities during the process of a core collapse supernova followed from the early observation of radioactive elements from deep in the core of the star in SN 1987A \citep{mulleretal1989}. Stars with a mass of greater than 8-10 times that of the Sun end their lives in a spectacular explosion known as a core collapse supernova. These events are among the most powerful explosions in the cosmos, releasing energy of order $10^{53}$ erg at a rate of $10^{45-46}$ watts, and are important for galactic chemical evolution because they produce and disseminate heavy elements. Core collapses supernovae also signal the birth of neutron stars and black holes, which are the basic building blocks of other astrophysical systems such as pulsars and x-ray binaries. During their lifetimes, stars are powered by the thermonuclear fusion of elements beginning with hydrogen fusing into helium. In a massive star, fusion continues all the way to iron-group elements. A core collapse supernova occurs when the inert iron core can no longer support the weight of the material above it and the core collapses, which releases gravitational binding energy that is in part converted to the energy of an expanding shock that ejects the outer layers of the star. Just prior to the explosion, the interior of the star has an onion-like structure, with iron-group elements in the core, then layers of silicon, magnesium, neon, oxygen, carbon, helium, and finally the outermost layer may be hydrogen. When the supernova explosion occurs, the shock passes through these layers of decreasing density. The early observation of a core element suggests some sort mixing must have occurred during the explosion, and, accordingly, motivated investigation into the effects of fluid instabilities. The laboratory experiment was designed to probe this scenario. The experimental configuration consists of a strong shock driven through a target with three layers of decreasing density. The interface between the first two layers is perturbed while the second interface is flat. An initially planar shock created by the deposition of energy from the laser is perturbed as it crosses the first interface, which excites a Richtmyer-Meshkov instability. As the perturbed shock propagates through the second interface, the perturbation is imprinted on the interface. The material begins to flow, leading to the growth of Rayleigh-Taylor instabilities. The three layers of the target are in a cylindrical shock tube composed of Be, with the density decreasing in the direction of shock propagation. The materials were Cu, polyimide plastic, and carbonized resorcinol formaldehyde (CRF) foam, with thicknesses of 85, 150, and 1500$\ \mathrm{\mu m}$ and densities 8.93, 1.41, and 0.1$\ \mathrm{g\ cm^{-3} }$, respectively. The shock tube delays the lateral decompression of the target, keeping the shock planar. The surface of the Cu layer was machined with a sinusoidal ripple of wavelength $200\ \mathrm{\mu m}$ and amplitude $15\ \mathrm{\mu m}$ to perturb the shock as it passes this interface. \begin{figure}[h] \includegraphics[angle=0, width=\columnwidth]{f1.eps} \caption{ Results of the three-layer target experiment. Shown are side-on X-ray radiographs at 39.9 ns (left) and 66.0 ns (right). The long, dark ``fingers'' are spikes of expanding Cu, and the horizontal band of opaque material to the right of the spikes of Cu is the brominated plastic tracer showing the imprinted instability growth at the plastic-foam interface. From \citet{calder.fryxell.ea:on} \copyright\ AAS. Reproduced with permission. } \label{fig:shock_exp} \end{figure} The experiment was driven by 10 beams of the laser with the target configured so that the laser beams impinge a thin section of CH ablator to prevent direct illumination and pre-heating of the target. The experimental diagnostics were X-ray radiographs taken at different times during a ``shot." The Be shock tube, polymide plastic, and CRF foam are transparent to X-rays, while the Cu layer is opaque to X-rays. Embedded within the polyimide layer was a tracer strip of brominated CH that is also opaque to X-rays. This tracer layer provided the diagnostic for polymide-foam interface, allowing visualization of the shock-imprinted structure. Figure \ref{fig:shock_exp} shows X-ray radiographs of the experiment at two times, one relatively early at 39.9 ns (left) and one relatively late at 66.0 ns (right). These images were from two different shots. The long, dark ``fingers" are spikes of expanding Cu, and the vertical band of opaque material to the right of the spikes of Cu is the brominated plastic tracer, showing the imprinted instability growth at the plastic-foam interface. The radiographs illustrate the configuration at early and late times in the evolution of the shocked target. The outer regions of the Cu and brominated strip show the effects of the shock tube, but the central part is largely immune to these effects. Making a quantitative comparison between the simulations and the experiments and determining the uncertainty in the study required close collaboration between experimentalists and theorists. This is an important point worth stressing. Without the contribution of both to interpreting and quantifying the experiments and simulations, there would have been little chance for a meaningful quantitative comparison. The data from the experiments are the radiographs, and finding a meaningful measurement for comparison to the simulation results required understanding the accuracy of the diagnostics and sources of uncertainty in the experiment. The metric for comparison between simulation and experiment was chosen as the length of the copper spikes, which are fairly obviously seen in the radiograph, but which required a deep understanding of the experiments to quantify. The paragraphs below summarize the sources of error and uncertainty in the experiments and the reader is referred to the original paper for complete details~\citep{calder.fryxell.ea:on}. A cautionary note concerning these details is warranted, however. The intervening years between these experiments and this writing have seen enormous progress in diagnosing high-energy-density experiments and the experiments described here are not the current state of the art~\citep{gamboa2012,stoecletal2012,gamboa2014a}. \begin{figure}[t] \includegraphics[angle=0, width=\columnwidth]{f2.eps} \caption{ Schematic of the three-layer target simulation initial conditions. Shown are the locations of the three materials, Cu, CH, and C, the shock, and the details of the sinusoidal perturbation of the Cu-CH interface. The schematic is not to scale. From \citet{calder.fryxell.ea:on} \copyright\ AAS. Reproduced with permission. } \label{fig:shock_init} \end{figure} The lengths of the Cu spikes in the experimental radiographs were determined by three methods. The first was a straightforward visual inspection of the images using a spatial reference grid located just below the images of Figure \ref{fig:shock_exp}. The second used a contour routine to better quantify the uncertainty in the location of the edges of the spikes. The third method was consistent with the analysis of the simulations. A section in the center of the images was vertically averaged to produce a single spatial lineout of optical depth through the region occupied by the Cu and CH. The same 5\% and 90\% threshold values were used to quantitatively determine the extent of the Cu spikes. Taking the average of all three methods, values of 330 $\pm$ $25\ \mathrm{\mu m}$ and 554 $\pm$ $25\ \mathrm{\mu m}$ are obtained at 39.9 and 66.0 ns, respectively. Sources contributing to uncertainty in these experimental measurements include the spatial resolution of the diagnostic, the photon statistics of the image, target alignment and parallax, and the specific contrast level chosen to define the length of the Cu spikes. These considerations allowed calculation of the experimental error bars presented in the figure (described below) that compares the experimental results to the simulation results. In addition to the spatial uncertainty, there were also several sources of uncertainty in the temporal accuracy. These arise from target-to-target dimensional variations, shot-to-shot drive intensity variations, and the intrinsic timing accuracy of the diagnostics. The experimental uncertainty in the timing is, however, relatively small, and is approximately indicated by the width of the symbols used in the comparison figure (below). The Flash simulations were two-dimensional with a three-layer arrangement of Cu, polyimide CH, and C having the same densities as those of the experimental target to model the experiment. The initial conditions for the Flash simulations represent the configuration 2.1 ns after the laser shot. At this point, the laser has deposited its energy and the shock is approaching the Cu-CH interface and the evolution is purely hydrodynamic. The initial conditions (thermodynamic profiles) for the Flash simulation were obtained from simulations of the laser-material interaction performed with a one-dimensional radiation hydrodynamics code~\citep{larsen94} that was able to describe the process of energy deposition occurring in the initial 2.1 ns. These one-dimensional profiles were mapped onto the two-dimensional grid with a sinusoidal perturbation added to the Cu-CH interface. Figure \ref{fig:shock_init} illustrates the initial configuration of the Flash simulations. For convenience, the simulations used periodic boundary conditions on the transverse boundaries and zero-gradient outflow boundary conditions on the boundaries in the direction of the shock propagation. The materials were treated as gamma-law gases, with $\gamma =$ 2.0, 2.0, and 1.3 for the Cu, CH, C, respectively. These values for gamma were chosen to give similar shock speeds to the shock speeds observed in the experiments. From these initial conditions, the simulations were evolved to approximately 66 ns. \begin{figure}[h] \includegraphics[angle=0, width=\columnwidth]{f3.eps} \caption{ Simulated radiographs from the 6 levels of refinement (effective resolution of $512 \times 256$) simulation of the three-layer target experiment. The simulated radiographs were created from the fluid abundances at times corresponding approximately to those of the images from the experiment, 39.9 ns (left) and 66.0 ns (right). Shown are the parts of the simulation domain that match the regions in the experimental results. From \citet{calder.fryxell.ea:on} \copyright\ AAS. Reproduced with permission. } \label{fig:shock_sim} \end{figure} Figure \ref{fig:shock_sim} shows simulated radiographs from a simulation at an intermediate resolution, allowing visual comparison to the experimental results. The figure presents simulated radiographs at approximately the two times corresponding to the images from the experiment, 39.9 ns (left panel) and 66.0 ns (right panel). The simulation in Figure \ref{fig:shock_sim} had 6 levels of mesh refinement corresponding to an effective resolution of $1024 \times 512$ grid zones. The simulated radiographs were created from the abundances of the three materials assigning an artificial opacity to each abundance and applying the opacity to an artificial ``beam." In addition, the abundances were de-resolved to match the resolution of the pixels in the experimental images and random Poisson-distributed `noise' was added to the intensity. An obvious qualitative difference between the simulated and experimental radiographs is readily observed in the curvature of the experimental instabilities that is not present in the simulations instabilities. The use of periodic boundary conditions in the transverse directions in the simulation was not consistent with the boundary conditions of the experiment, which was performed with the three materials of the target inside a cylindrical Be shock tube. The experiment results show the influence of the shock tube walls as a curving or pinching of the outer Cu spikes, while the simulations did not consider these boundary effects. Comparison of the simulated radiographs to the radiographs from the experiment show that the simulations captured the bulk behavior of the materials, particularly the growth of Cu spikes and the development of C bubbles. We can conclude from this comparison that the simulations resemble the experimental results. Assessment of the comparison as``good" or ``bad" is difficult, however, with only a visual comparison, especially one that indicates a difference due to a boundary condition effect. What is needed is a quantitative comparison, and for that we apply the same techniques as we apply to verification, a convergence study to show the simulations converge with resolution and a quantitative comparison to the experimental results. To test convergence of the solutions, a suite of simulations was performed at increasing resolution. The effective resolutions of the simulations were 128 $\times$ 64, 256 $\times$ 128, 512 $\times$ 256, 1024 $\times$ 512, 2048 $\times$ 1024, and 4096 $\times$ 2048, corresponding to 4, 5, 6, 7, 8, and 9 levels of adaptive mesh refinement. As noted above, the lengths of the Cu spikes were chosen as the metric for quantitative comparison to the experiments. Flash solves an advection equation for each abundance, which allowed tracking the flow of each material with time. The spike lengths in the simulations were measured by averaging the CH abundance in the $y$-direction across the simulation domain then smoothing the resulting one-dimensional array slightly to minimize differences that would occur owing to very small scale structure. The length of the Cu spikes was then determined by the average distance spanned by minimum locations of average abundances 0.05 and 0.9. The results were reasonably robust to the amount of smoothing and threshold values. The results of testing the convergence of the Cu spike length measurements are shown in Figure \ref{fig:shock_conv}, which depicts percent differences from the highest resolution simulation, 9 levels of adaptive mesh refinement, as functions of time. The trend is that the difference decreases with increasing mesh resolution, with the 7 and 8 level of adaptive mesh refinement simulations always demonstrating agreement to within five percent. The trend of decreasing difference with increasing mesh resolution demonstrates a convergence of the flow, but it is subject to caveats. We note that the trend does not describe the behavior at all points in time (that is, the percent difference curves sometimes cross each other), and this average measurement is an integral property of the flow and in no way quantifies the differences in small scale structure observed in the abundances. In particular, we note that the difference curve for the simulation with 8 levels of adaptive mesh refinement crosses the curves of both the 7 and 6 level simulations, suggesting that higher-resolution simulations may deviate further from these results and produce degraded agreement with the experiment. This result is in keeping with the above-mentioned concerns with ILES. \begin{figure}[h] \includegraphics[angle=0, width=\columnwidth]{f4.eps} \caption{ Percent difference of the Cu spike lengths from those of the highest resolution (9 levels of adaptive mesh refinement) simulation vs.\ time. The percent differences are from the lower resolution simulations of 4, 5, 6, 7, and 8 levels of adaptive mesh refinement, with the corresponding effective resolutions in the legend. We note that the convergence is not perfect. The curve from the 8 levels of refinement simulation crosses those of the 6 and 7 levels of refinement simulations, indicating a higher percent difference. Adapted from \citet{calder.fryxell.ea:on}. } \label{fig:shock_conv} \end{figure} Figure \ref{fig:shock_results} shows the Cu spike length vs.\ time for 4 simulations at increasing resolution. Also shown are the above-mentioned experimental results. The experimental error bars correspond to $\pm 25\ \mathrm{\mu m}$, the spatial error of the experiment. The width of the symbols marking the experimental results indicates approximately the timing error. The figure shows that the simulations quantitatively agree with the experimental results at the early and late times to within the experimental uncertainty. \begin{figure}[h] \includegraphics[angle=0, width=\columnwidth]{f5.eps} \caption{Results from a validation test consisting of a laser-driven shock propagating through a multi-layer target. The lengths of the Cu spikes is plotted vs.\ time from 4 simulations at 6, 7, 8, and 9 levels of adaptive mesh refinement in a convergence study. The effective resolutions are given in the legend. Also shown are the experimental results at two times with spatial error bars of ($\pm 25 \mathrm{\mu m}$). The timing error is about the width of the diamonds marking the experimental result. The differences between the simulations at different resolutions is less than the uncertainty of the experimental results. Adapted from \citet{calder.fryxell.ea:on}. } \label{fig:shock_results} \end{figure} As noted above, this study has previously appeared in the literature. Complete details of the validation study may be found in \citet{calder.fryxell.ea:on,calder2005,calderetal2006} and additional details of the experiments may be found in \citet{calder.pre,robeyetal01}. \subsection{Computation of Reaction Products in Large Eddy Simulations of Supernovae} When a laboratory experiment is available, the distinction between verification and validation is fairly clear, as discussed earlier. However, when creating predictive simulations of astrophysical processes that cannot be reproduced directly in the laboratory, even using appropriate scaling laws, the distinction can become less clear because the task becomes one of confirmation of simulation results without laboratory results. In many situations, notably in stellar combustion, it is possible to have a model that is demonstrably more physically valid but is too expensive or constrained to be used for the desired predictive simulations. Simpler models must be applied to simulate observed phenomena, hence the need for comparison of different methods. Nuclear reaction networks and multi-dimensional simulations present a good example of this confluence of verification and validation. In astrophysical detonations it is possible to compute the steady-state structure of the propagating reaction front with a large reaction network with hundreds of species and thousands of reactions using error-controlled numerical methods \citep[e.g.][]{sharpe1999,mooretownsleybildsten2013} Consider the following question: How many species are necessary to accurately predict the characteristics of the flow such as peak temperature and reaction front width? This is not a verification question. We can use verification techniques to demonstrate that the equations governing the time integration of the reactions are being solved to a desired accuracy. Such a test, however, does not demonstrate whether or not a particular selection of species is sufficient for the stated purpose. So we proceed to compare our model with say three or a dozen ``effective" reactions or species to another model which we believe to be more physically valid because it has more complete reaction physics. This situation is neither verification that our model is being solved correctly (that is already done) nor is it validation against a specific physical experiment. It is, however, validation under the definition introduced in section \ref{sec:vvapproach} above, in that it addresses whether the model is physically correct. Some terminology refers to this as confirmation of one model with a physically more valid model. Since the label depends finely on definitions of terminology, it is useful in discussion to term this type of comparison as something that combines elements of verification and validation (see Ch.\ 41 by Beisbart in this volume). It is a model-to-model comparison, as verification often is, but addresses the physical applicability of the model, as validation does. If integration of thousands of reactions were the only issue, this validation of simplified models might not be worthwhile; instead one would just use the better model directly. There are areas of prediction, however, where direct use of the better model can be infeasible. In explosive astrophysical combustion (which powers type Ia supernova explosions), it is typically desirable to predict the overall products and the speeds at which they are ejected. Unfortunately, a simulation that can predict that information must include the entire star, which may be around $10^9$~cm in size. The reaction front through which the combustion takes place is one cm or less in thickness \citep{townetal15}. Also, the propagation of this front through the star will generally occur in a way that obeys no particularly symmetry, making it necessary to simulate this combustion and ejection of material in three dimensions. The necessity of simulating the whole star in three dimensions presents several challenges from the standpoint of V\&V. First, since the combustion phenomena occur far below the best possible grid scale ($\sim 10^5$~cm), the typical method of verification by convergence study is not valid. Claiming convergence for a numerical solution of differential equations presupposes that the relevant gradients are numerically resolved and become better resolved at higher resolution. This is the very meaning of resolution. However, in the full-scale astrophysical case, an example of the above-mentioned large eddy simulation situation, the composition gradients representing the physical reaction front (the length scale over which the fuel is consumed and converted to products) are never actually resolved. Secondly, while error-controlled methods for ODE integration are well-understood, similar automated control of accuracy is not available in current widely used methods for solution of PDEs, such as in hydrodynamics. Because this control is not built into the method, performing predictive simulations involves a constant process of verification to ensure that solutions obtained do not depend on resolution. That process can be both expensive and time-consuming. Thirdly, it may be computationally infeasible to include hundreds of species and thousands of reactions in the full-scale hydrodynamic simulation, thus even if we were able to verify the methods for reactive hydrodynamics, we would need to use a model for the reactions that we know to have specific deficiencies and would therefore need some form of validation against more physically complete models. Finally, as discussed earlier, because some physical processes such as fluid dissipation due to viscosity is left implicit, a higher-resolution simulation may not only be more numerically accurate but also more physically valid. As a result of these issues, verification and validation of the simulation of a stellar explosion can be mixed in a way that is not always cleanly separable. Here we will present a discussion of ongoing efforts at verification and validation of methods for computing the products of thermonuclear supernova explosions. The full-star simulations use a simplified model of the reactions for computational efficiency, and are necessarily under-resolved. The overall goal is to compare the results from this computational model to computational models of much higher physical and numerical fidelity. In the case of combustion, those are computations with large, complete nuclear reaction networks computed using resolved, error-controlled numerical techniques. The limitation is that the latter methods can only be used under certain flow conditions, specifically, a steady state. We therefore proceed by treating the methods used in the full-star simulation as the model to be validated by comparison to more physical calculations. This is similar to verification by comparison to a benchmark, except that the two models are known to be different by construction. Table \ref{tab:sim_cap} shows a matrix comparing the capabilities of compressible hydrodynamics simulations in various dimensions as well as the fully resolved method, which can only be used in one dimension and for reaction fronts propagating in a steady state through a uniform medium. As shown, a resolved calculation with the full network at all densities relevant to the supernova can only be performed with the steady-state method. However, this method cannot be used to treat transients (e.g. ignition or non-spatially uniform density) or general geometries including the full star. Of the hydrodynamical methods in various spatial dimensions, represented in the other three columns of the table, only one-dimensional calculations can use a full reaction network effectively and resolve the reaction front, though not at all densities. The possible importance of transient effects necessitates a multi-step strategy utilizing cross-comparisons of calculations of reaction front structure among several different methods. For example, we can verify one-dimensional dynamical calculations at uniform densities using comparison to steady-state calculations, and then use one-dimensional calculations with non-uniform density to characterize transient effects. Even for a transient, it is informative to compare to steady-state solutions in order to provide physical insight to the importance of non-uniformities in density. \begin{table} \begin{center} \begin{tabular}{|l|c|c|c|c|} \hline Capability & 3-d & 2-d & 1-d & 1-d steady\\ \hline \hline full reaction network & $\times$ & $\times$ & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark \\ \hline resolved at low density & $\times$ & $\times$ & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark \\ \hline resolved at high density & $\times$ & $\times$ & $\times$ & \cellcolor{green!25} \checkmark \\ \hline transients (dynamical) & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark & $\times$ \\\hline general geometries & \cellcolor{green!25} \checkmark & $\times$ & $\times$ & $\times$ \\\hline full star & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark & \cellcolor{green!25} \checkmark & $\times$ \\\hline \end{tabular} \caption{Capabilities of simulations in various dimensions and assumptions. Comparison of results among simulations is performed in order to validate that full star three-dimensional simulations reproduce the results of more physically valid one- dimensional calculations of steady state properties of detonations.\label{tab:sim_cap}} \end{center} \end{table} Figure \ref{fig:det_comp} shows an example of a comparison of the compositional structure of a propagating detonation reaction front computed with the one-dimensional dynamical method and the one-dimensional steady-state method. The hydrodynamical simulation (dashed lines) was performed at a physical resolution of $10^5$~cm, which corresponds to a hydrodynamical time step of about $10^{-4}$~s. The fuel here is mostly $^{12}$C and $^{16}$O, which is reacted to eventually become $^{56}$Ni. The consumption of $^{12}$C is not shown, but is even faster than that of $^{16}$O. The structure for a detonation propagating in steady state (solid lines) is computed with an error-controlled method using adaptive time stepping and an error tolerance of order $10^{-6}$, and is therefore suitably resolved by construction. The abundance histories from the hydrodynamical model shown here are the result of using a simplified reaction model in the hydrodynamics and then post-processing the resulting density and temperature histories of fluid elements with a larger reaction network \citep{travaglioetal2004,townetal15}. The goal of this comparison is to validate that away from the unresolved portion of the reaction front (timescales $\gtrsim 10^{-3}$~s), the composition history is accurately predicted by the under-resolved calculation with the simplified burning model. This comparison shows that the results are in good agreement for steady-state, planar detonations. For an example of a comparison for non-planar (curved) detonations see \cite{mooretownsleybildsten2013}. \begin{figure}[h] \includegraphics[width=\columnwidth]{f6.eps} \caption{ Comparison of planar steady-state detonation structure simulated hydrodynamically at $10^5$ cm resolution using post-processing of Lagrangian tracers (dashed) with the steady-state structure computed directly using error-controlled integration (solid). Abundances here are given as mass fractions. Similar to comparisons made in \citet{townetal15}. The oxygen consumption structure will remain unresolvable even with more than an order} of magnitude higher resolution in the hydrodynamic simulation. \label{fig:det_comp} \end{figure} The validation of methods for computing astrophysical combustion in large eddy simulations is ongoing. The various possible calculations represented in Table \ref{tab:sim_cap} must be compared for geometries and conditions for which there is overlap in capability. This process also entails ongoing improvement of both the simplified reaction model utilized in the large eddy simulations \citep{townsleyetal2009,willcoxetal2016} as well as improving techniques for computing the final yields \citep{townetal15}. \section{Discussion} The simulational results for the hydrodynamics validation example fell within the temporal and spatial error bars of the experimental results thus showing quantitative agreement between simulation and experiment for the metric of the lengths of the copper spikes. This agreement demonstrates that the hydrodynamics module in Flash captured the bulk properties of the flow and observable morphology. and builds confidence in simulations of astrophysical phenomena. We cannot, however, declare the code ``validated" for a host of reasons: \begin{itemize} \item The experimental configuration produced essentially a two-dimensional result, hence our modeling it with two-dimensional simulations. The experiment was three-dimensional, so correctly describing the fluid instabilities, particularly the amount of small-scale structure in the flow may require three-dimensional simulations. \item The models were incomplete. The three materials were modeled as ideal gasses, a questionable assumption. Also, for convenience, the simulations neglected the presence of the shock tube surrounding the target and assumed periodic boundary conditions. Thus the simulations did not include effects due to the shock tube. \item The experimental diagnostics, radiographs, are really shadows that cannot adequately capture small-scale structure. Even if three-dimensional simulations that better described the fluid instabilities had been performed, comparison to the experimental results is limited by the experimental diagnostics. \item The observed degraded agreement between simulations at the highest resolutions indicates the results are not converged. We attribute this result to the fact that the Euler equations allow a changing numerical viscosity with resolution, which changes the Reynolds number and thus the nature of any turbulence. Additional commentary on this issue may be found in \citet{calder.fryxell.ea:on}. \end{itemize} Even with limitations, the demonstrated ability of the simulations to capture the expected bulk properties of the flow builds confidence in the results of astrophysical simulations, allowing us to conclude that the shocks and fluid instabilities study was a success. The principal differences observed between the results from simulations and the experimental results were in the amount of small scale structure observed in the flow, with the amount of small-scale structure in the simulations increasing with resolution. This behavior is expected because the effective Reynolds number increases with resolution as described above, and we believe this effect is the source of the observed imperfect convergence. Because the experimental data are radiographs and cannot capture the actual amount of small-scale structure in the flow, the correct amount of small-scale structure remains undetermined and even if the convergence of the simulations had been perfect, we could not conclude the solution converged to the correct result. In addition to increasing confidence in the results, the hydrodynamics validation study was well worth the investment because of the lessons learned in comparing the experimental and simulational results. The collaborative process of determining the metric for comparison and extracting the results from the experimental and simulational data resulted in a better understanding of the issues, which also increases confidence in the astrophysical results. The experimentalists also benefited from the process of validation because the process of comparison suggested metrics for future comparisons, provided useful diagnostics, and supplied a virtual model that aided in the design of future experiments. A point worth stressing again in conclusion is the importance of close collaboration between the experimentalists and theorists needed to make a meaningful quantitative comparison. Raw experimental data such as a radiograph alone does not allow for a quantitative comparison to simulational results. Finally, we note that the success of this collaboration seeded interest in high energy density physics among the developers of Flash, which subsequently resulted in an extended course of collaborative research into high energy density physics (see \citet{Tzeferacosetal2015} and references therein). The product of reactive hydrodynamics study gave a look at the process of comparing models of differing fidelity to ensure that macroscopic (three-dimensional) simulations capture the physics of thermonuclear reactions while also allowing the calculation of detailed abundances. Our approach is to test simplified models against higher-fidelity models for a given physical process, here thermonuclear combustion. Simplified models then facilitate three-dimensional simulations that would be intractable otherwise. The results of these studies are also applicable to the problem of determining detailed abundances from the density and temperature histories of Lagrangian tracers. We illustrated this process with a comparison between results from post-processed tracers from a hydrodynamics simulation and a detailed calculation of steady-state burning structure. This study confirmed that our simulations capture the essence of the reactions in whole-star models, and thereby increased confidence in our predictions of the astrophysical events. \section{Conclusions} The cases we present here are but one part of the continuing effort at verifying and validating Flash and associated infrastructure (e.g.\ the post-processing method presented here). The first study of validating the hydrodynamics was performed early in the development of Flash. Though very informative, it could have been continued further with additional quantification of the effect of missing physics as a good next step. Also, further modifications to the code would allow it to capture high energy density phenomena better. Such activities, however, were not critical to the astrophysical problems. Still, the case was very informative and served to increase confidence in the results. The second case, the computation of reaction products in large eddy simulations of supernovae, is very much a work in progress and represents our contemporary effort. Our conclusion from both of these studies is that like any discipline in computational science, V\&V are an essential part of the process of modeling astrophysical phenomena. V\&V in astrophysics can be particularly challenging due to the inaccessibility of the physical conditions attained and limited ancillary measurements available for distant events. As shown here by these examples, however, positive steps that build confidence in models can be taken based on comparisons using related laboratory experiments and more complete physical models where available. \begin{acknowledgement} This contribution was supported in part by the Department of Energy through grant DE-FG02-87ER40317 and the research described here was supported in part by earlier grant B341495 to the Center for Astrophysical Flashes at the University of Chicago. The software used in this work was in part developed by the DOE-supported ASC/Alliances Center for Astrophysical Thermonuclear Flashes at the University of Chicago. This research has made use of NASA's Astrophysics Data System. The authors thank Eric Winsberg and Bruce Fryxell for helpful discussions about this manuscript. \end{acknowledgement} \newcommand{Astrophysical Journal}{Astrophysical Journal} \newcommand{Monthly Notices of the Royal Astronomical Society}{Monthly Notices of the Royal Astronomical Society} \newcommand{Astronomy and Astrophysics}{Astronomy and Astrophysics}

c19def9c26b09428b6c085cb78ab83cfa271b680

\section{Introduction} The ocean is a critical component of the global carbon cycle, presently holding 60 times more carbon than the pre-industrial atmosphere \citep{ciais2013}. Changes in ocean carbon storage also affect atmospheric carbon dioxide (\ce{CO2}), thereby impacting the climate system. From a dynamical perspective, ocean carbon uptake is intricately linked to physical circulations \citep{graven2012}, and recent research emphasizes the role of large-scale fluid advective processes in transporting carbon across the base of the mixed layer \citep{levy2013}. However, few studies have explored the role of small-scale turbulent circulations in ocean carbon uptake, resulting in continued uncertainty regarding parameterizations of air-sea \ce{CO2} fluxes in Earth system models (ESMs). The present study examines the ocean carbon cycle at small scales using large eddy simulations (LES) to model reactive carbonate species evolving in the presence of realistic mixed-layer turbulence. Prior research on reacting flows (e.g., \citep{hamlington2011}) has shown that the strongest interactions between reactions and turbulence occur when chemical and mixing timescales are within an order of magnitude, as is the case for carbonate chemistry in the oceanic mixed layer. In the ocean surface boundary layer, cooling-driven surface convection, wind-driven shear turbulence, and wave-driven Langmuir turbulence \citep{langmuir1938} occur at time scales of roughly \SIrange{1}{100}{\minute}, while wave periods and breaking occur at roughly \SIrange{1}{10}{\second}. Once \ce{CO2} is transferred across the air-sea interface, it reacts with seawater to produce bicarbonate (\ce{HCO3^-}) and carbonate (\ce{CO3^2-}) in a series of reactions whose rate-limiting steps have time scales of roughly \SI{1}{\minute} \citep{zeebe2001}. As a result, the timescales of small-scale ocean turbulence and carbonate chemistry can be of the same order, leading to strong coupling between flow physics and reactions. Larger-scale turbulent processes, such as mixing by meso- and submeso-scale eddies, have much longer timescales and are not likely to interact strongly with carbonate chemistry reactions, although they may link in a similar way to the biological carbon cycle, which also has long timescales. Although there are linkages between inorganic and biological carbon cycles that can cause slow variations in chemical composition, chemical reactions themselves remain fast and are coupled most strongly to correspondingly fast turbulent mixing. Despite the timescale matching between carbonate chemistry and turbulence, however, the present study is the first to simultaneously solve chemical and fluid flow equations in a coupled fashion. Previous studies have shown that Langmuir turbulence produces spatial heterogeneity (or ``patchiness'') at the ocean surface through the aggregation of buoyant tracers such as debris, plankton, nutrients, or oil within the convergence zones of its counter rotating cells \citep{langmuir1938,stommel1949,smayda1970,powell1975,barstow1982,thorpe2000,thorpe2009,lewis2005,qiao2009,smith2016,suzuki2016}. Due to enhanced mixing, Langmuir turbulence also increases the vertical extent over which tracers are distributed \citep{woodcock1950,woodcock1993,shoener1970,johnson1977,ledbetter1979,buranathanitt1982,pinelalloul1995,solow1995,gallager1996,yool1998,bees1998,mcwilliams2000}. The effects of surface waves on sea-surface chemistry have also been examined \citep{sutcliffe1963,williams1967,parsons1973,eisenreich1978,dierssen2009}, largely focusing on the increase in aggregation of nutrients due to windrows and the precipitation of organic particles due to bubble injection from surface wave breaking. In each of these cases, however, reactions were assumed to be sufficiently fast or slow in comparison to the dominant mixing process and, as a result, significant reaction-flow couplings were neglected. Although some studies have examined time-dependent carbonate chemistry within the ocean, primarily focusing on enhancement of air-sea fluxes and uptake of carbon by individual phytoplankton cells, these studies have largely assumed that the flow is laminar, quiescent, or well-mixed by small-scale turbulence \citep{quinn1971,johnson1982,williams1983,jahne1995,wolf-gladrow1999,schulz2006,zeebe2007,schulz2009,guo2011,uchikawa2012}. Many studies have also examined the enhancement of \ce{CO2} exchange rates across the air-sea interface as a function of wind-driven turbulence, wave breaking, and bubble injection; however, no studies have specifically included a time-dependent \ce{CO2} hydration mechanism \citep{bolin1960,hoover1969,pankow1982,goldman1982,asher1986,woolf1993,jahne1995,farmer1995,asher1998,boutin1999,jacobs2002,thorpe2003,kuss2004}. The primary goals of this paper are to examine the effects of Langmuir turbulence on carbonate chemistry by simultaneously considering the time-dependent nature of both processes and to exemplify a coupled chemistry-physics modeling system capable of carrying out this examination. Specifically, this study seeks to determine how Langmuir turbulence affects the amount of dissolved inorganic carbon (DIC) in the oceanic mixed layer. As a secondary objective, this study examines how chemical model fidelity affects predictions of DIC in the mixed layer. These objectives are addressed using LES of reactive carbonate species at small scales for different strengths of Langmuir turbulence and different chemical models, including time-dependent and equilibrium models. The simulations are enabled by the development of a new reduced mechanism for carbonate chemistry, as well as the implementation of a Runge--Kutta--Chebyshev numerical integrator to handle the stiffness of the governing equations. Without these developments, the computational cost of the LES is prohibitive. In the following, details of the numerical simulations are provided in Section \ref{sec:sims}, including the development of the reduced carbonate chemistry model. Section \ref{results} outlines the simulation results and Section \ref{sec:discuss} discusses the implications of these results for ESMs, as well as how the present observations might vary for different ocean conditions. Conclusions and directions for future research are provided at the end. \section{Description of Numerical Simulations\label{sec:sims}} \subsection{Governing Equations and Solver} The governing equations solved in the simulations are the wave-averaged Boussinesq equations \citep{suzuki2016} with additional transport equations for reactive species concentrations (termed ``tracers'' in the following) \citep{smith2016}, namely \begin{linenomath*} \begin{eqnarray}\label{mom_eq} \frac{D {\bf u}}{Dt} &=& - \nabla p -{\bf f}_\mathrm{c} \times {\bf u}_\mathrm{L} - u_{\mathrm{L},j} \nabla u_{\mathrm{s},j} + b \hat{{\bf z}}+ \textbf{SGS}_u\,,\\ \frac{D b}{Dt} &=& \mathrm{SGS}_b\,,\label{buoy}\\ \frac{D {\bf c}}{Dt} &=& {\bf S} + {\bf SGS}_c\,, \label{concentration_cont} \\ \nabla \cdot {\bf u} &=& 0\,,\label{continuity} \end{eqnarray} \end{linenomath*} where $D/Dt \equiv \partial/\partial t + ({\bf u}_\mathrm{L} \cdot \nabla)$ is the material derivative, ${\bf u}_\mathrm{L} \equiv {\bf u} + {\bf u}_\mathrm{s}$ is the Lagrangian velocity, ${\bf u}$ is the Eulerian velocity averaged over surface gravity waves, ${\bf u}_\mathrm{s}$ is the Stokes drift velocity created by surface gravity waves, $p$ is the pressure normalized by a reference density $\rho_0$, ${\bf f}_\mathrm{c}$ is the Coriolis parameter, and $b$ is the buoyancy. Buoyancy and density $\rho$ are related by $b = -g \rho/\rho_0$, where $g$ is gravitational acceleration. The density is related to the potential temperature $\theta$ by the relation $\rho = \rho_0 [1+\beta_T (\theta_0 - \theta)]$, where $\beta_T$ is the thermal expansion coefficient and $\theta_0$ is a reference temperature. In Eq.~\eqref{concentration_cont}, ${\bf c}$ denotes the vector of Eulerian concentration fields for each of the tracers. The tracers are passive and thus do not impact the dynamics of ${\bf u}$ or $b$. However, they are non-conserved and ${\bf S}$ in Eq.\ (\ref{concentration_cont}) accounts for sources and sinks due to chemical reactions, as outlined in Section \ref{reactions}. Each of the subgrid-scale (SGS) terms in Eqs.~\eqref{mom_eq}--\eqref{concentration_cont} are fluxes from the SGS model used in the LES. Note that the form of Eq.~\eqref{mom_eq} is obtained by \citet{suzukifoxkemper16}, although it is mathematically identical to the form in \citet{mcwilliams1997}. Langmuir turbulence is created in the simulations by the Stokes drift velocity ${\bf u}_\mathrm{s}$, which appears in Eqs.~\eqref{mom_eq}--\eqref{concentration_cont}. This additional forcing term is expressed in the present LES as \begin{linenomath*} \begin{equation}\label{stokes_eq} {\bf u}_\mathrm{s}(z) = u_\mathrm{s} (z) \left[ \cos(\vartheta_\mathrm{s}) \hat{{\bf x}}+\sin(\vartheta_\mathrm{s}) \hat{{\bf y}}\right]\,, \end{equation} \end{linenomath*} where $u_\mathrm{s}(z)$ is the Stokes drift magnitude vertical profile, which decays faster than exponentially from the surface \citep{donelan1985,webb2011}, and $\vartheta_\mathrm{s}$ is the angle of the Stokes drift velocity in the horizontal (i.e., $x$--$y$) plane. Note that, in the present study, ${\bf u}_\mathrm{s}$ is constant in time and the same at all horizontal locations, and thus depends only on $z$. Wind, $\vartheta_\mathrm{w}$, and Stokes drift, $\vartheta_\mathrm{s}$, directions are taken to be the same in all simulations (thereby representing wind, as opposed to crossing swell, waves) and both wave-spreading and breaking wave effects are neglected \citep{webb2015}. Prior studies \citep{mcwilliams1997,vanroekel2012,hamlington2014,smith2016} have shown that the inclusion of the Stokes drift velocity in Eqs.~\eqref{mom_eq}--\eqref{concentration_cont} leads to the creation of small-scale, counter-rotating Langmuir cells throughout the domain, with the strongest cells occurring close to the surface. The numerical code used to perform the simulations is the National Center for Atmospheric Research (NCAR) LES model \citep{moeng1984,mcwilliams1997,sullivan2007}. Horizontal spatial derivatives are calculated pseudo-spectrally, while second- and third-order finite differences are used for vertical derivatives of velocity and tracers, respectively. Third-order Runge--Kutta (RK) time-stepping is used with a constant Courant number. Subgrid-scale viscosity, buoyancy diffusivity, and tracer diffusivity are spatially varying according to the scheme outlined by \citet{sullivan1994}. \subsection{Reduced Carbonate Chemistry Model\label{reactions}} The reduced carbonate chemistry model implemented in the LES is based on the mechanism from \citet{zeebe2001} for carbonate reactions in seawater. This mechanism includes seven species concentrations for $\ce{CO2}$, $\ce{HCO3^-}$, $\ce{CO3^2-}$, $\ce{H^+}$, $\ce{OH^-}$, $\ce{B(OH)3}$, and $\ce{B(OH)4^-}$ (denoted $c_1$--$c_7$; see Table \ref{tracer refs.}), plus $\ce{H2O}$, which is assumed to have a constant concentration. The system of seven reactions describing the mechanism is given as \citep{zeebe2001} \begin{linenomath*} \begin{eqnarray*} \label{eq:r1}\ce{CO2} + \ce{H2O} &\ce{<=>[\alpha_1][\beta_1]}& \ce{HCO3^-} + \ce{H^+} \\ \label{eq:r2}\ce{CO2} + \ce{OH^-} &\ce{<=>[\alpha_2][\beta_2]}& \ce{HCO3^-} \\ \label{eq:r3}\ce{CO3^2-} + \ce{H^+} &\ce{<=>[\alpha_3][\beta_3]}& \ce{HCO3^-} \\ \label{eq:r4}\ce{HCO3^-} + \ce{OH^-} &\ce{<=>[\alpha_4][\beta_4]}& \ce{CO3^2-} + \ce{H2O} \\ \label{eq:r5}\ce{H2O} &\ce{<=>[\alpha_5][\beta_5]}& \ce{H^+} + \ce{OH^-} \\ \label{eq:r6}\ce{B(OH)3} + \ce{OH^-} &\ce{<=>[\alpha_6][\beta_6]}& \ce{B(OH)4^-} \\ \label{eq:r7}\ce{CO3^2-}+ \ce{B(OH)3} + \ce{H2O} &\ce{<=>[\alpha_7][\beta_7]}& \ce{B(OH)4^-} + \ce{HCO3^-} \end{eqnarray*} \end{linenomath*} where $\alpha_i$ and $\beta_i$ are, respectively, forward and backward reaction coefficients. Table\ \ref{rates_zeebe2001} provides temperature- and salinity-dependent equations for each of these coefficients, as well as values for the coefficients at a temperature of \SI{25}{\degreeCelsius} and salinity of 35 ppt \cite{dickson1994,zeebe2001}. \begin{table}[t!] \centering \caption{Definition of tracer concentrations $c_i$, terminology, and equilibrium values of tracer concentrations used to initialize the simulations. The equilibrium values correspond to approximate surface values at a temperature of \SI{25}{\degreeCelsius}, salinity of 35 ppt, alkalinity of \SI{2427.89}{\micro\mole\per\kilo\gram}, and DIC concentration of \SI{1992.28}{\micro\mole\per\kilo\gram}.}\label{tracer refs.} \begin{tabular}{@{}l l l c@{}} \hline Tracer & Species & Name & Equilibrium Value (\si{\micro\mole\per\kilo\gram}) \\ \hline $c_1$ & $\ce{CO2}$ & Carbon dioxide & 7.57 \\ $c_2$ & $\ce{HCO3^-}$ & Bicarbonate & \num{1.67e3} \\ $c_3$ & $\ce{CO3^2-}$ & Carbonate & \num{3.15e2}\\ $c_4$ & $\ce{H^+}$ & Hydrogen ion & \num{6.31e-3} \\ $c_5$ & $\ce{OH^-}$ & Hydroxyl & 9.60 \\ $c_6$ & $\ce{B(OH)3}$ & Boric acid & \num{2.97e2} \\ $c_7$ & $\ce{B(OH)4^-}$ & Tetrahydroxyborate & \num{1.19e2} \\ \hline \end{tabular} \end{table} The source terms $S_i$ on the right-hand-side of Eq.~\eqref{concentration_cont} for the rate equation of each tracer $c_i$ are obtained using the law of mass action as \begin{linenomath*} \begin{eqnarray} \hspace{-0.3in}S_1 &=& -(\alpha_1 + \alpha_2 c_5) c_1 + (\beta_1 c_4 + \beta_2) c_2 \,,\label{s1} \\ \hspace{-0.3in}S_2 &=& (\alpha_1 + \alpha_2 c_5) c_1 - (\beta_1 c_4 + \beta_2 + \beta_3 + \alpha_4 c_5 + \beta_7 c_7 ) c_2 + (\alpha_3 c_4 + \beta_4 +\alpha_7 c_6 ) c_3 \,, \\ \hspace{-0.3in}S_3 &=& (\beta_3 + \alpha_4 c_5 + \beta_7 c_7 ) c_2 - (\alpha_3 c_4 + \beta_4 + \alpha_7 c_6 ) c_3 \,, \\ \hspace{-0.3in}S_4 &=& \alpha_1 c_1 - (\beta_1 c_4 - \beta_3)c_2 - \alpha_3 c_4 c_3 + (\alpha_5 - \beta_5 c_4 c_5) \,, \\ \hspace{-0.3in}S_5 &=& -\alpha_2 c_5 c_1+(\beta_2-\alpha_4 c_5) c_2 + \beta_4 c_3 + (\alpha_5 - \beta_5 c_4 c_5)-(\alpha_6 c_5 c_6 - \beta_6 c_7) \,, \\ \hspace{-0.3in}S_6 &=& \beta_7 c_7 c_2 - \alpha_7 c_6 c_3 - (\alpha_6 c_5 c_6 -\beta_6 c_7)\,, \\ \hspace{-0.3in}S_7 &=& -\beta_7 c_7 c_2 + \alpha_7 c_6 c_3 + (\alpha_6 c_5 c_6 -\beta_6 c_7)\,. \label{s7} \end{eqnarray} \end{linenomath*} The system of rate equations resulting from this reaction mechanism is, however, numerically stiff and requires a prohibitively small time step to accurately and stably integrate within NCAR LES using the native third-order RK scheme. To overcome this difficulty, two measures were taken: (\emph{i}) a computational singular perturbation analysis and subsequent quasi-steady-state approximation were used to reduce the chemical mechanism, and (\emph{ii}) a Runge--Kutta--Chebyshev scheme was used to integrate the resulting system of rate equations. \begin{table}[t!] \centering \caption{Temperature and salinity dependent reaction coefficient equations and values at a temperature of \SI{25}{\degreeCelsius} and salinity of 35 ppt for the carbonate chemistry model used in the present study. All values and expressions are taken from \citet{zeebe2001}. Here $A_1 = \SI{4.70e7}{\kilo\gram\per\mole\per\second}$, $E_1 = \SI{23.2}{\kilo\joule\per\mole}$, $A_6 = \SI{4.58e10}{\kilo\gram\per\mole\per\second}$, $E_6 = \SI{20.8}{\kilo\joule\per\mole}$, $A_7 = \SI{3.05e10}{\kilo\gram\per\mole\per\second}$, and $E_7 = \SI{20.8}{\kilo\joule\per\mole}$. The temperature and salinity dependent equilibrium constant equations for $K_1^*$, $K_2^*$, $K_W^*$, and $K_B^*$ are given by \citet{dickson1994}. }\label{rates_zeebe2001} \begin{tabular}{@{}lccc@{}} \hline Symbol & Equation & Value & Units \\ \hline $\alpha_1$ & $\exp[1246.98 - \num{6.19e4} / \theta - 183.0 \ln(\theta)]$ & 0.037 & \si{\per\second} \\ $\beta_1$ & $\alpha_1/ K_1^*$ & \num{2.66e4} & \si{\kilo\gram\per\mole\per\second}\\ $\alpha_2$ & $A_1 \exp(-E_1/R \theta)$ & \num{4.05e3} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_2$ & $\alpha_2 K_W^*/K_1^*$ & \num{1.76e-4} & \si{\per\second}\\ $\alpha_3$ & constant & \num{5.0e10} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_3$ & $\alpha_3 K_2^*$ & \num{59.4} & \si{\per\second}\\ $\alpha_4$ & constant & \num{6.0e9} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_4$ & $\alpha_4 K_W^* / K_2^*$ & \num{3.06e5} & \si{\per\second}\\ $\alpha_5$ & constant & \num{1.40e-3} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_5$ & $\alpha_5/ K_W^*$ & \num{2.31e10} & \si{\kilo\gram\per\mole\per\second}\\ $\alpha_6$ & $A_6 \exp(-E_7/R \theta)$ & \num{1.04e7} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_6$ & $\alpha_6 K_W^* / K_B^*$ & \num{249} & \si{\per\second}\\ $\alpha_7$ & $A_7 \exp(-E_8/R \theta)$ & \num{6.92e6} & \si{\kilo\gram\per\mole\per\second}\\ $\beta_7$ & $\alpha_7 K_2^* / K_B^*$ & \num{3.26e6} & \si{\kilo\gram\per\mole\per\second}\\ \hline \end{tabular} \label{default} \end{table} \subsubsection{Computational Singular Perturbation Analysis} Computational singular perturbation (CSP) analysis \citep{lam1988,lam1994,goussis1992,lam1993} was applied to the chemical kinetic system represented by Eqs.~\eqref{s1}--\eqref{s7} to identify candidate species for quasi-steady state (QSS) approximations following the approach outlined by \citet{lu2008a,lu2008b} and \citet{niemeyer2015}. To perform the CSP analysis, the reaction rate equations from Eqs.~\eqref{concentration_cont} and \eqref{s1}--\eqref{s7} were first written as a zero-dimensional (i.e., only time dependent) system given by \begin{linenomath*} \begin{equation} \frac{d\mathbf{c}}{dt} = \mathbf{S}(\mathbf{c})\quad \Rightarrow \quad \frac{d\mathbf{S}}{dt}= \mathbf{J} \mathbf{S}\,, \end{equation} \end{linenomath*} where $\mathbf{J} = \partial \mathbf{S}/\partial \mathbf{c}$ is the Jacobian matrix. The CSP analysis decomposes the source terms $\mathbf{S}$ into a vector of modes $\mathbf{f}$ using row basis vectors $\mathbf{B}$ as $\mathbf{f} = \mathbf{B} \mathbf{S}$. The time derivative of $\mathbf{f}$ then gives \begin{linenomath*} \begin{equation}\label{feq} \frac{d \mathbf{f}}{d t} = \left(\frac{d\mathbf{B}}{dt}+\mathbf{B}\cdot\mathbf{J}\right)\mathbf{S} = \mathbf{\Lambda} \mathbf{f}\,, \end{equation} \end{linenomath*} where $\mathbf{\Lambda}$ is given by \begin{linenomath*} \begin{equation} \mathbf{\Lambda}=\left(\frac{d\mathbf{B}}{dt}+\mathbf{B}\mathbf{J}\right)\mathbf{A}\,, \end{equation} \end{linenomath*} and $\mathbf{A}=\mathbf{B}^{-1}$. For simplicity, the Jacobian matrix was assumed to be time independent such that $d\mathbf{B}/dt = 0$, leading to \begin{linenomath*} \begin{equation}\label{eigen} \mathbf{\Lambda} = \mathbf{B} \mathbf{J} \mathbf{A}\,, \end{equation} \end{linenomath*} where $\mathbf{\Lambda}$ contains the eigenvalues of $\mathbf{J}$ on the diagonal. The eigendecomposition of the Jacobian was performed using the NumPy function \texttt{numpy.linalg.eig}~\citep{vanderwalt2011}; the Jacobian itself was evaluated analytically using SymPy~\citep{Meurer2017}. The CSP basis vectors $\mathbf{A}$ and $\mathbf{B}$ are then the right and left eigenvectors of $\mathbf{J}$, respectively. Using the eignevalues from Eq.~\eqref{eigen}, the system dynamics were separated into fast and slow subspaces, where the evolution of the modes $\mathbf{f}$ in each subspace is given from Eq.~\eqref{feq} by \cite{niemeyer2015} \begin{linenomath*} \begin{equation} \frac{d}{dt}\left[\begin{array}{c} \mathbf{f}^\mathrm{fast} \\ \mathbf{f}^\mathrm{slow}\end{array}\right] = \left[\begin{array}{cc} \mathbf{\Lambda}^\mathrm{fast} & {} \\ {} & \mathbf{\Lambda}^\mathrm{slow}\end{array}\right]\left[\begin{array}{c} \mathbf{f}^\mathrm{fast} \\ \mathbf{f}^\mathrm{slow}\end{array}\right]\,. \end{equation} \end{linenomath*} The fast modes $\mathbf{f}^\mathrm{fast}$ decay rapidly and have negative eigenvalues $\mathbf{\Lambda}^\mathrm{fast}$ that are much larger in magnitude than the eigenvalues $\mathbf{\Lambda}^\mathrm{slow}$ associated with the slow subspace $\mathbf{f}^\mathrm{slow}$. Formally, the fast and slow subspaces were identified by defining a cutoff time scale $\tau_\text{c} / \gamma_{\text{CSP}}$, where $\tau_\text{c}$ is a characteristic time scale of the global system dynamics and $\gamma_{\text{CSP}}$ is a safety factor. The two subspaces were then separated by requiring that the time scale associated with the smallest magnitude eigenvalue in the fast subspace (corresponding to the slowest mode in the fast subspace), denoted $\lambda_\mathrm{min}(\mathbf{\Lambda}^\mathrm{fast})$, be less than the cutoff time scale: \begin{linenomath*} \begin{equation} \frac{-1}{\lambda_{\min} \left(\mathbf{\Lambda}^{\text{fast}} \right)} <\frac{\tau_c}{\gamma_{\text{CSP}}}\,. \label{eq:csp_separation} \end{equation} \end{linenomath*} The negative-valued eigenvalues in $\mathbf{\Lambda}$ with magnitudes greater than $\lambda_\mathrm{min}$ are all part of the fast subspace, while the remaining eigenvalues are part of the slow subspace. From $\mathbf{\Lambda}^\mathrm{fast}$ and $\mathbf{\Lambda}^\mathrm{slow}$, it was then possible to identify the fast and slow subspace modes, $\mathbf{f}^\mathrm{fast}$ and $\mathbf{f}^\mathrm{slow}$, respectively. The characteristic time $\tau_\text{c}$ was defined to be the relaxation time for \ce{CO2} to reach 1\% of its equilibrium concentration after a \SI{1}{\micro\mole\per\kilo\gram} increase in \ce{CO2}, a \SI{1}{\micro\mole\per\kilo\gram} decrease in \ce{CO^2-_3}, and a \SI{2}{\micro\mole\per\kilo\gram} increase in \ce{OH-}~\cite{zeebe2001}, giving $\tau_\text{c} = \SI{63.03}{\second}$. The safety factor was set as $\gamma_{\text{CSP}} = 50$. \subsubsection{Quasi-Steady-State Approximation} Projecting $\mathbf{S}$ onto the fast and slow subspaces gives $\mathbf{S} = \mathbf{S}^\mathrm{fast}+\mathbf{S}^\mathrm{slow}$, where $\mathbf{S}^\mathrm{fast}=\mathbf{Q}^{\mathrm{fast}}\mathbf{S}$ and $\mathbf{S}^\mathrm{slow}=\mathbf{Q}^{\mathrm{slow}}\mathbf{S}$. Here, $\mathbf{Q}^{\mathrm{fast}}$ and $\mathbf{Q}^{\mathrm{slow}}$ are, respectively, the fast and slow projection matrices given by \begin{linenomath*} \begin{equation} \mathbf{Q}^{\mathrm{fast}} = \mathbf{A}^{\text{fast}} \mathbf{B}^{\text{fast}} \,, \quad \mathbf{Q}^{\mathrm{slow}} = \mathbf{A}^{\text{slow}} \mathbf{B}^{\text{slow}}\,. \end{equation} \end{linenomath*} The basis vectors were split into fast- and slow-mode vectors by applying Eq.~\eqref{eq:csp_separation} to identify the associated fast and slow eigenvalues, and $\mathbf{Q}^{\mathrm{slow}}$ was constructed. Then, species were identified as good candidates for the QSS assumption if they correlated (or projected) weakly to the slow subspace. For the $i$th species, this was determined using \begin{linenomath*} \begin{equation} \left| \mathbf{Q}_{i,i}^{\text{slow}} \right| < \epsilon_{\text{CSP}} \,, \label{eq:csp_criterion} \end{equation} \end{linenomath*} where $\mathbf{Q}_{i,i}^{\text{slow}}$ is the $i$th diagonal element of $\mathbf{Q}^{\text{slow}}$ and $\epsilon_{\text{CSP}}$ is a small threshold value (0.1 was used here). Practically, species were determined to satisfy the criterion given by Eq.~\eqref{eq:csp_criterion} by calculating the maximum values of $\mathbf{Q}_{i,i}^{\text{slow}}$ for all species over a simulated relaxation back to equilibrium after a 10\% perturbation to the concentration of \ce{CO2}, at a temperature of \SI{25}{\celsius} and salinity of 35 ppt. \begin{figure}[t!] \centering \includegraphics[width=0.45\linewidth]{figure1} \caption{Relaxation times for the concentration of \ce{CO2}, denoted $c_1$, determined using the full kinetic model (solid black line) and using the reduced model with the QSS approximation in Eq.~\eqref{qss} applied to \ce{H^+} (dash-dot red line).} \label{fig:relaxation_compare} \end{figure} The CSP analysis identified two QSS candidates: \ce{H^+} and \ce{OH^-}, with slow-subspace contributions (i.e., $\mathbf{Q}_{i,i}^{\text{slow}}$) of \num{1.81e-5} and \num{2.72e-2} respectively. Both species satisfied the criterion in Eq.~\eqref{eq:csp_criterion}, but it was found that the approximation could only be applied to \ce{H^+} without introducing significant error. Thus, the CSP analysis determined that \ce{H^+} (which is connected to the pH) was a candidate for the QSS approximation by identifying it as a ``radical'' (in the CSP context) because it contributed little to the slow, controlling modes of the system dynamics, below a safety factor. Using the QSS approximation for the concentration of \ce{H^+} (tracer $c_4$, see Table \ref{tracer refs.}), it was assumed that $S_4=0$ and that $c_4$ could be obtained algebraically as \begin{linenomath*} \begin{equation}\label{qss} c_4^* = \frac{\alpha_1 c_1 + \beta_3 c_2 + \alpha_5}{\beta_1 c_2 + \alpha_3 c_3 + \beta_5 c_5}\,, \end{equation} \end{linenomath*} where $c_4^*$ denotes the QSS approximation for $c_4$. Computationally, the resulting reduced chemical mechanism was less stiff due to the use of QSS for one of the three fastest-evolving species, allowing a 50\% increase in the time step required for the simulations, and required the integration of only six, as opposed to seven, coupled differential equations [i.e., no differential equation needed to be integrated for $c_4$, since this tracer concentration was given algebraically by Eq.~\eqref{qss}]. The error due to the QSS assumption was estimated using a zero-dimensional test where the system was perturbed by an increase of \SI{1}{\micro\mole\per\kilo\gram} increase in $c_1$, a \SI{1}{\micro\mole\per\kilo\gram} decrease in $c_3$ (to maintain constant DIC concentration), and a \SI{2}{\micro\mole\per\kilo\gram} increase in $c_5$ (to maintain constant alkalinity)~\cite{zeebe2001}, after which all species relaxed back to their respective equilibrium values. Examining the temporal evolution, the concentrations of all species in the reduced model agreed within \num{1e-5}\% of the full model [except for the concentration of $c_4$, which was analytically provided by Eq.\ (\ref{qss}) resulting from the QSS assumption] over the entire equilibration period ($\sim$\SI{60}{\second}). Figure~\ref{fig:relaxation_compare} shows the results from this test. \subsubsection{Runge--Kutta--Chebyshev Solver} In the simulations, time integration of the advection and chemistry was split \cite{strang1968} such that the advection remained within the pre-existing third-order RK scheme in NCAR LES and the chemistry was integrated in two half steps, before and after the advection step. The chemistry integration used an explicit second-order Runge--Kutta--Chebyshev (RKC) scheme that is robust for moderately stiff equations \cite{sommeijer1997,verwer2004,niemeyer2014}. While explicit, the RKC algorithm is stabilized to handle more stiffness than traditional RK methods. The RKC scheme is explicit and constructed like other multistage explicit RK methods, but uses an increased, variable number of stages and coefficients chosen to increase the stability region rather than accuracy---thus the method is known as a stabilized explicit scheme. The use of the RKC solver provided an additional increase in the time step required for the simulations, from roughly $10^{-5}$ s without the RKC solver (i.e., using the native third-order RK scheme in NCAR LES) to roughly $0.1$ s with the RKC solver. \subsection{Physical Setup \label{phys scenario}} The physical and computational parameters used to setup the simulations are summarized in Table \ref{tab:setup}. All simulations were initialized with a mixed layer depth of \SI{30}{\meter}, with uniform stratification (i.e., linearly increasing density) below. Buoyancy, density, and temperature were all spatially and temporally varying in the simulations, but salinity was assumed fixed at 35 ppt. The physical domain size was $L_x\times L_y \times L_z = 320\times 320\times \SI{96}{\meter\cubed}$ with a horizontal ($x-y$) resolution of \SI{2.5}{\meter} and a vertical ($z$) resolution of \SI{0.75}{\meter}. The initial velocities were motionless. Periodic boundary conditions were used in horizontal directions and a zero vertical velocity condition was applied at the bottom boundary. A surface wind stress of \SI{0.025}{\newton\per\square\meter} was applied to all simulations along the $x$ direction, with a friction velocity of $u_\tau = \SI{5.3e-3}{\meter\per\second}$, corresponding to a \SI{10}{\meter} wind speed of \SI{5.75}{\meter\per\second}. Zero-gradient boundary conditions were used for the temperature at the top and bottom of the domain, and the diurnal cycle was not modeled in the simulations. \begin{table}[b!] \caption{Summary of physical and computational parameters used in the numerical simulations.}\label{tab:setup} \centering \begin{tabular}{@{}l c c c c@{}} \hline Physical size, $L_x \times L_y \times L_z$ & \multicolumn{4}{c}{\SI{320}{\meter}$\times$\SI{320}{\meter}$\times$\SI{-96}{\meter}} \\ Grid size, $N_x \times N_y \times N_z$ & \multicolumn{4}{c}{128$\times$128$\times$128} \\ Grid Resolution, $\triangle_x \times \triangle_y \times \triangle_z$ & \multicolumn{4}{c}{\SI{2.5}{\meter}$\times$\SI{2.5}{\meter}$\times$\SI{0.75}{\meter}} \\ Reference density, $\rho_0$ & \multicolumn{4}{c}{\SI{1000}{\kilo\gram\per\cubic\meter}} \\ Thermal expansion coefficient, $\beta_T$ & \multicolumn{4}{c}{\SI{2e-4}{\per\kelvin}} \\ Coriolis parameter, $\textbf{f}_\mathrm{c}$ & \multicolumn{4}{c}{\SI{0.729e-4}{\per\second}$\hat{\textbf{z}}$} \\ Initial mixed layer depth, $H_{\text{ML},0}$ & \multicolumn{4}{c}{\SI{-30}{\meter}} \\ Wind speed at 10 m, $U_{10}$ & \multicolumn{4}{c}{\SI{5.75}{\meter\per\second}} \\ Stokes drift direction, $\vartheta_\mathrm{s}$ & \multicolumn{4}{c}{\SI{0}{\degree}} \\ Water-side wind friction velocity, $u_\tau=\sqrt{\tau/\rho_o}$ & \multicolumn{4}{c}{\SI{5.3e-3}{\meter\per\second}} \\ Wind stress, $\tau$ & \multicolumn{4}{c}{\SI{0.025}{\newton\per\square\meter}} \\ Wind direction, $\vartheta_\mathrm{w}$ & \multicolumn{4}{c}{\SI{0}{\degree}} \\ \hline Surface stokes drift, $u_\mathrm{s}(0)$ (\si{\meter\per\second}) & 0.000 & 0.032 & 0.080 & 0.132 \\ Langmuir number, $\mathrm{La}_\mathrm{t} \equiv [u_\tau/u_\mathrm{s}(0)]^{1/2}$ & $\infty$ & 0.40 & 0.30 & 0.20 \\ Simulation label & NS & La04 & La03 & La02 \\ \hline \end{tabular} \end{table} Four wave-forcing scenarios were examined by varying the Stokes drift velocity: a single case with no Langmuir turbulence and three cases with increasing strengths of Langmuir turbulence. The Stokes drift velocity profiles, $u_\mathrm{s}(z)$, applied in each of the Langmuir cases are shown in Figure~\ref{figure1}. The strength of the Langmuir turbulence is characterized by its turbulent Langmuir number, $\text{La}_\mathrm{t}^2 = u_\tau / u_\mathrm{s}(0)$, where $u_\mathrm{s}(0)$ is the surface Stokes drift from each of the profiles shown in Figure~\ref{figure1}. The four scenarios examined correspond to $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2, where $\mathrm{La}_\mathrm{t} = \infty$ is the non-Langmuir case. The range of Langmuir numbers explored here is realistic \citep{li2016,li2017}, and 0.3 is the value attained under fully developed seas \citep{webb2011}. The $\mathrm{La}_\mathrm{t} = 0.4$ and $0.2$ cases are intended to reveal the effects of weaker and stronger Langmuir turbulence, respectively, as compared to the baseline value. Additional discussion of the physical setup represented by these simulations, as well as how turbulence-chemistry interactions would vary for different conditions, is provided in Section \ref{time}. \begin{figure}[t!] \centering \includegraphics[width=0.45\linewidth]{figure2} \caption{Stokes drift velocity $u_\mathrm{s}(z)$ as a function of depth $z$ for Langmuir numbers $\text{La}_\mathrm{t} = 0.4$, 0.3, and 0.2 (red, blue, and green lines, respectively), where $\mathrm{La}_\mathrm{t} \equiv [u_\tau/u_\mathrm{s}(0)]^{1/2}$). The main plot shows $u_\mathrm{s}(z)$ on linear axes and the inset shows $u_\mathrm{s}(z)$ on semilog axes.} \label{figure1} \end{figure} According to \citet{callaghan2008}, less than 0.25\% of the global sea surface area is expected to be covered by whitecapping (i.e., breaking waves) for the wind strength considered here. Consequently, no wave-breaking parameterization was used in these simulations, although one has been developed for the NCAR LES model \citep{sullivan2007} and could be explored in future work. Similarly, bubble parameterizations were not included in these simulations \citep{liang2011}, although the effects of bubbles are likely to be significant \citep{woolf1993}, particularly given their connection to Langmuir turbulence \citep{farmer1995,thorpe2003}. Although the present study is specifically focused on the effects of enhanced vertical mixing by Langmuir turbulence, future work will explore the effects of bubbles, for example using the parameterization for bubble-enhanced air-sea fluxes given by \citet{woolf1993}. For each physical scenario, species concentrations were initialized uniformly throughout the domain using equilibrium values for a temperature of \SI{25}{\degreeCelsius}, salinity of 35 ppt, alkalinity of \SI{2427.89}{\micro\mole\per\kilo\gram}, and DIC of \SI{1992.28}{\micro\mole\per\kilo\gram} (see Table~\ref{tracer refs.}). Here, DIC is the sum of all carbon containing species and its concentration, denoted $c_\mathrm{DIC}$, is defined as $c_\mathrm{DIC}\equiv c_1+c_2+c_3$. Each tracer was subject to periodic boundaries in horizontal directions with, initially, no vertical fluxes at the bottom and top boundaries. \subsection{Simulation Procedure} After approximately seven days during which turbulence was allowed to develop and tracers relaxed to their equilibrium values based on the local value of the temperature (salinity is fixed), additional \ce{CO2} was allowed to enter through the top boundary according to Henry's law for gas flux across the air-sea interface \citep{wanninkhof1992}. This flux law is given as \citep{smith2016} \begin{linenomath*} \begin{equation}\label{flux} F_{\ce{CO2}}(x,y,t) = k_{\ce{CO2}} \left[c_{1}^{\text{air}} - c_1(x,y,0,t)\right]\,, \end{equation} \end{linenomath*} where $F_{\ce{CO2}}$ is the downward flux rate across the boundary, which varies over horizontal directions and time as temperature and $c_1$ vary, $k_{\ce{CO2}}$ is the species flux rate (or piston velocity), $c_{1}^{\text{air}}$ is the concentration in air, and $c_1(x,y,0,t)$ is the concentration just below the surface. The value of $c_{1}^{\text{air}}$ was fixed at a 10\% increase above the initial mixed layer average of $c_1$ (namely, $c_{1}^{\text{air}}$ = \SI{8.3}{\micro\mole\per\kilo\gram}, see Table~\ref{tracer refs.}). The piston velocity, $k_{\ce{CO2}}$, is given as a function of the 10 m wind speed $U_{10}$ (see Table \ref{tab:setup}) and Schmidt number $\mathrm{Sc}$ as \begin{linenomath*} \begin{equation}\label{henry} k_{\ce{CO2}} = 0.31 U_{10}^2 \sqrt{\frac{660}{\mathrm{Sc}}}\,, \end{equation} \end{linenomath*} where $\mathrm{Sc}$ is a function of temperature given by \citep{wanninkhof1992} \begin{linenomath*} \begin{equation}\label{schmidt} \mathrm{Sc}= 2073.1-125.62\theta + 3.6276\theta^2-0.043219\theta^3\,. \end{equation} \end{linenomath*} Note that in the above expressions, $k_{\ce{CO2}}$ has units of \SI{}{\centi\meter\per\hour} and $\theta$ in Eq.~\eqref{schmidt} is assumed to have units of \SI{}{\degreeCelsius}. The piston velocity $k_{\ce{CO2}}$ from Eq.~\eqref{henry} does not include a bubble parameterization, but consideration of bubbles, as well as their coupling to Langmuir turbulence, is an important direction for future research. The simulations were run for six additional hours after initiating the air-sea flux of \ce{CO2} and analysis of the data was carried out after this period. Longer simulations were not performed due to the computational expense of integrating the time-dependent chemistry, and also due to the artificiality of neglecting the diurnal cycle over long periods. Six hours was found to be sufficient for identifying trends in the data, but all of the conclusions contained herein should be understood as only strictly valid up to six hours; future work is necessary to determine carbonate chemistry evolution over much longer time periods, including diurnal and seasonal cycles. Two additional sets of simulations were also performed: one in which each of the chemical species concentrations were calculated at carbonate chemical equilibrium \citep{zeebe2001} and one in which there were no chemical reactions, but still including surface fluxes, transport, and mixing. The equilibrium model is implemented by ensuring that, at each location and time, there is no propensity for the concentrations $c_i$ to change due to reactions. This is accomplished by setting $\mathbf{S}$ to zero and solving the system of nonlinear coupled algebraic equations represented by Eqs.~\eqref{s1}--\eqref{s7} to find the equilibrium values of $\mathbf{c}$. The three chemistry models are referred to in the following as the Time-dependent Chemistry (TC), the Equilibrium Chemistry (EC), and the No Chemistry (NC) models, respectively. The physical scenarios for these three sets of simulations were identical and are described in Section \ref{phys scenario}. The EC model is representative of how carbonate chemistry is most commonly calculated within ESMs, where reactions are assumed to be instantaneous (i.e., infinitely fast) with respect to modeled physical processes. The NC model, by contrast, effectively represents reactions that are infinitely slow. These two sets of simulations thus give upper and lower bounds for ocean carbonate chemistry reaction times. \section{Results\label{results}} \subsection{Physical Ocean State\label{physics}} Figure~\ref{fig:figure3} shows fields of vertical velocity and potential temperature fluctuations for the non-Langmuir ($\text{La}_\mathrm{t} = \infty$) and three Langmuir ($\text{La}_\mathrm{t} =$ 0.4, 0.3, 0.2) cases. The $x$--$y$ surface fields of vertical velocity in Figures~\ref{fig:figure3}(b-d) for the three Langmuir cases show the streak-like patterns formed by long counter-rotating Langmuir cells that are characteristic of Langmuir turbulence. Although these streaks are spatially variable in direction and magnitude, they are preferentially aligned with the wind direction along the $x$-axis (since $\vartheta_\mathrm{s}=\vartheta_\mathrm{w}=0^\circ$; see Table \ref{tab:setup}) and generally increase in magnitude as $\mathrm{La}_\mathrm{t}$ decreases (i.e., with increasing strength of Langmuir turbulence). \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{figure3} \caption{Fields of (a-d) vertical velocity $w$ in units of m s$^{-1}$ and (e-h) potential temperature fluctuation $\theta-\langle \theta\rangle_0$ in units of K in horizontal planes at the surface (top subpanels) and in vertical planes in the middle of the domain (bottom subpanels) for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$ (a,e), 0.4 (b,f), 0.3 (c,g), and 0.2 (d,h) using time-dependent chemistry. The background potential temperature $\langle \theta\rangle_0$ is computed as the $x$--$y$ average of $\theta$ at the surface.} \label{fig:figure3} \end{figure} The vertical velocities are smallest in the shear-only non-Langmuir case [Figure~\ref{fig:figure3}(a)] and largest in the $\mathrm{La}_\mathrm{t} = 0.2$ case [Figure~\ref{fig:figure3}(d)]. These enhanced vertical velocities are also evident in the $x$--$y$ averaged depth profiles of vertical velocity variance shown in Figure~\ref{fig:figure4}(a), where the peak magnitude of the vertical velocity variance is greatest for the smallest Langmuir number (i.e., $\mathrm{La}_\mathrm{t} = 0.2$), with a progressive increase in magnitude from the non-Langmuir (i.e., $\mathrm{La}_\mathrm{t} = \infty$) case. Figure~\ref{fig:figure3} further shows that, in addition to the increase in magnitude of vertical mixing, the vertical extent of mixing is greater for the Langmuir cases than for the non-Langmuir case. This enhanced mixing, which increases in strength as Langmuir number decreases, results in a deeper mixed layer. This is indicated by the fluctuating potential temperature fields in Figures~\ref{fig:figure3}(e-h), which show that, as the Langmuir number decreases, greater temperature fluctuations are observed throughout the mixed layer, and the mixed layer extends to slightly greater depths. The deepening of the mixed layer is perhaps more evident in the $x$--$y$ averaged potential temperature profiles shown in Figure~\ref{fig:figure4}(b). All simulations begin with the same temperature profile (the dashed gray line in Figure~\ref{fig:figure4}) and all deviate from this initial profile by the end of the seven day spin-up period, but the deviation becomes increasingly pronounced as the strength of Langmuir turbulence increases. In particular, the increased mixing associated with Langmuir turbulence has deepened the mixed layer by approximately \SIrange{1}{3}{\meter}, depending on the case, over the course of the spin-up. This deepening not only increases the total volume of the mixed layer, thereby increasing the short-term new carbon reservoir size, but also decreases the average temperature of the mixed layer by entraining cooler waters from below. While this decrease in temperature may not seem substantial in the larger context, carbonate chemistry and air-sea gas fluxes are both sensitive to temperature, as indicated by the temperature-dependent reaction rate coefficients in Table \ref{rates_zeebe2001} and the Henry's law gas flux expression in Eq.~(\ref{henry}). As temperatures cool, \ce{CO2} becomes more soluble in water, allowing more \ce{CO2} to enter the domain. However, reaction times also decrease, leaving carbon as \ce{CO2} longer before it is converted into $\ce{HCO3^-}$ and $\ce{CO3^2-}$. This effect of temperature (i.e., the competition between increased solubility and decreased reaction times), is not directly examined in this study, although future studies exploring these effects within the context of Langmuir turbulence and carbonate chemistry are certainly warranted. \begin{figure}[t!] \centering \includegraphics[width=0.8\linewidth]{figure4} \caption{Vertical profiles of (a) vertical velocity variance $\langle w'^2\rangle$ and (b) average potential temperature $\langle \theta\rangle$ for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2 (black, red, blue, and green lines, respectively). Statistics are computed in horizontal $x$--$y$ planes as a function of depth $z$. The dashed gray line in (b) shows the initial temperature profile with uniform temperature above $z = \SI{-30}{\meter}$ and constant stratification below.} \label{fig:figure4} \end{figure} \subsection{Effects of Langmuir Turbulence on Carbonate Chemistry\label{langmuir effect}} As Langmuir turbulence strengthens, additional carbon is brought through the surface and progressively further down into the mixed layer, as shown in Figure~\ref{fig:figure5}. In particular, Figure~\ref{fig:figure5} shows that the vertical extent of $c_\mathrm{DIC}$ distribution in the mixed layer increases as $\mathrm{La}_\mathrm{t}$ decreases, while the peak surface concentrations decrease. This is partially due to the fast mixing and increased vertical flux associated with Langmuir turbulence, but also to the mixed layer deepening effect of Langmuir turbulence \citep{hamlington2014}; both of these physical effects were described in Section \ref{physics}. \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{figure5} \caption{Fields of DIC concentration $c_\mathrm{DIC}$ in units of $\mu$mol kg$^{-1}$ in horizontal planes at the surface (top subpanels) and in vertical planes in the middle of the domain (bottom subpanels) for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2 (a-d) using time-dependent chemistry.} \label{fig:figure5} \end{figure} \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{figure6} \caption{Vertical profiles of (a) average $c_\mathrm{DIC}$, (b) standard deviation of $c_\mathrm{DIC}$, and (c) vertical flux of $c_\mathrm{DIC}$ for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2 (black, red, blue, and green lines, respectively), using time-dependent chemistry. Statistics are computed in horizontal $x$--$y$ planes as a function of depth $z$.} \label{fig:figure6} \end{figure} Figures~\ref{fig:figure6}(a) and (b) show the $x$--$y$ average and standard deviation, respectively, of $c_\mathrm{DIC}$ as a function of depth for $\text{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2. In the non-Langmuir case (i.e., $\text{La}_\mathrm{t} = \infty$), there is a much greater concentration and standard deviation of $c_\mathrm{DIC}$ near the surface and very little near the base of the mixed layer. Conversely, the three Langmuir cases have progressively more uniform concentrations and lower variance throughout the mixed layer. Again, the more uniform vertical distribution and decreased standard deviation of the three Langmuir cases, in comparison to the non-Langmuir case, can largely be attributed to the faster vertical mixing associated with Langmuir turbulence. Figure~\ref{fig:figure6}(c) shows that the Langmuir cases all exhibit increased downward vertical flux near the surface in comparison to the non-Langmuir case. While their magnitudes are quite similar near the surface, the stronger Langmuir cases have sustained increased flux deeper into the domain. As \ce{CO2} is mixed away from the surface, a larger air-sea flux results from Henry's law in Eq.~\eqref{henry}. Thus, an increase in $c_\mathrm{DIC}$ is expected to occur as the strength of Langmuir turbulence increases. Figure~\ref{fig:figure7}(a) shows the total domain-integrated change in $c_\mathrm{DIC}$ after six hours relative to the initial concentration when the air-sea flux of $\ce{CO2}$ begins (defined here to be at $t=0$). This total change, denoted $\Delta c_\mathrm{DIC}$, is expressed as \begin{linenomath*} \begin{equation}\label{cdelta} \Delta c_\mathrm{DIC}(t) = \langle c_\mathrm{DIC} \rangle_V(t)-\langle c_\mathrm{DIC} \rangle_V(t=0)\,, \end{equation} \end{linenomath*} where $\langle\cdot\rangle_V$ is an average over the entire domain in $x$--$y$--$z$ directions at a particular time. Figure~\ref{fig:figure7}(a) shows that $\Delta c_\mathrm{DIC}(t=6\,\mathrm{hours})$ progressively increases as $La_\mathrm{t}$ decreases, indicating that the Langmuir cases have indeed brought additional DIC into the domain as compared to the non-Langmuir case. \begin{figure}[t!] \centering \includegraphics[width=\linewidth]{figure7} \caption{Dependence on Langmuir number of (a) volume-integrated change in $c_\mathrm{DIC}$, where $\Delta c_\mathrm{DIC}$ is given by Eq.\eqref{cdelta}, (b) time series of the enhancement in $c_\mathrm{DIC}$ within the domain, where $E_\mathrm{DIC}$ is defined in Eq.~\eqref{percent_enhance}, and (c) time series of the normalized change in surface flux, where $\widetilde{\Delta} F_{\ce{CO2}}$ is defined in Eq.~\eqref{flux_enhance}. All panels show results for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2 (black, red, blue, and green bars and lines, respectively) using time-dependent chemistry.} \label{fig:figure7} \end{figure} At first glance, the differences in Figure~\ref{fig:figure7}(a) may appear to be small. However, Figure~\ref{fig:figure7}(b) shows that there can be a significant enhancement in the amount of DIC brought into the mixed layer by Langmuir turbulence. This can be quantified by comparing $\Delta c_\mathrm{DIC}$ for the Langmuir and non-Langmuir cases at each time and by defining an enhancement parameter, $E_\mathrm{DIC}$, that expresses the difference relative to the non-Langmuir case. This parameter is calculated as \begin{linenomath*} \begin{equation} E_\mathrm{DIC}(t) =100\times \frac{\Delta c_\mathrm{DIC}(t) - [\Delta c_\mathrm{DIC}]_\mathrm{base}(t)}{[\Delta c_\mathrm{DIC}]_\mathrm{base}(t)}\label{percent_enhance}\,, \end{equation} \end{linenomath*} where $E_\mathrm{DIC}$ is expressed as a percentage and $[\Delta c_\mathrm{DIC}]_\mathrm{base}$ is the baseline change in domain-integrated $c_\mathrm{DIC}$ against which the Langmuir cases are compared. In this section, $[\Delta c_\mathrm{DIC}]_\mathrm{base}$ is taken to be $\Delta c_\mathrm{DIC}$ for the non-Langmuir case with time-dependent chemistry. Figure~\ref{fig:figure7}(b) shows that, for this simulation configuration and after six hours of constant, uniform wind and wave forcing, there is a Langmuir-induced enhancement of 0.09--0.14\% more DIC in the domain as compared to the non-Langmuir case with just wind-driven shear turbulence. Fundamentally, the observed differences in new DIC brought into the domain are due to differences in the flux rate of \ce{CO2} across the air-sea interface, given by $F_{\ce{CO2}}$ in Eq.~\eqref{flux}. Figure~\ref{fig:figure7}(c) shows the change in horizontally ($x$--$y$) averaged $F_{\ce{CO2}}$ as a function of time, denoted $\widetilde{\Delta} F_{\ce{CO2}}$, where the notation $\widetilde{\Delta}$ reflects the fact that the change is normalized by the average $F_{\ce{CO2}}$ at the initial time. This quantity is calculated as \begin{linenomath*} \begin{equation} \widetilde{\Delta} F_{\ce{CO2}} (t) =100\times \frac{\langle F_{\ce{CO2}}\rangle(t) - \langle F_{\ce{CO2}}\rangle(t=0)}{\langle F_{\ce{CO2}}\rangle(t=0)}\label{flux_enhance}\,, \end{equation} \end{linenomath*} where, as with $E_\mathrm{DIC}$ in Eq.~\eqref{percent_enhance}, $\widetilde{\Delta} F_{\ce{CO2}}$ is expressed as a percentage. The time series of $\widetilde{\Delta} F_{\ce{CO2}}$ in Figure~\ref{fig:figure7}(c) show that all cases have a sharp initial decline in air-sea flux rate. However, $\widetilde{\Delta} F_{\ce{CO2}}$ for the non-Langmuir case continues to decrease at a faster rate in comparison to the three Langmuir cases, indicating that Langmuir flux enhancement may persist over diurnal and synoptic time scales if saturation does not occur. If there is a build-up of \ce{CO2} at the surface, the air-sea gradient in \ce{CO2} concentration decreases, thereby decreasing $F_{\ce{CO2}}$ locally, and $\langle F_{\ce{CO2}}\rangle$ over the entire surface. If, instead, the concentration of \ce{CO2} at the surface is maintained at a lower value for a longer period of time, the air-sea gradient in \ce{CO2} is relatively unchanged and $F_{\ce{CO2}}$ is not reduced as dramatically. \subsection{Effects of Chemical Model Fidelity on Carbonate Chemistry\label{chemistry effect}} To isolate the effect of chemical model fidelity on the air-sea flux rate of \ce{CO2} and the resulting enhancement of DIC within the upper ocean, this section compares the shear-only non-Langmuir case and the $\text{La}_\mathrm{t} = 0.3$ case for each of the three chemistry models (i.e., TC, EC, and NC). \begin{figure}[b!] \centering \includegraphics[width=\linewidth]{figure8} \caption{Dependence on chemical model fidelity of (a) volume-integrated change in $c_\mathrm{DIC}$, where $\Delta c_\mathrm{DIC}$ is given by Eq.\eqref{cdelta}, (b) time series of the enhancement in $c_\mathrm{DIC}$ within the domain, where $E_\mathrm{DIC}$ is defined in Eq.~\eqref{percent_enhance} and the non-Langmuir equilibrium chemistry case is used as a baseline, and (c) time series of the normalized change in surface flux, where $\widetilde{\Delta} F_{\ce{CO2}}$ is defined in Eq.~\eqref{flux_enhance}. All panels show results for the time-dependent (TC), equilibrium (EC), and no (NC) chemistry models, for both non-Langmuir (NS) and $\text{La}_\mathrm{t}$ = 0.3 (La03) cases.} \label{fig:figure8} \end{figure} The volume-integrated change in DIC [defined in Eq.~\eqref{cdelta}] shown in Figure~\ref{fig:figure8}(a) indicates that both the TC and EC models bring more carbon into the domain in comparison with the NC model. This is because carbonate chemistry, in either time-dependent or equilibrium forms, provides a sink of \ce{CO2} and preserves the air-sea gradient, resulting in chemistry flux enhancement. The EC case exceeds the TC case in carbon uptake. This occurs because reactions in the EC case are infinitely fast and thus aqueous \ce{CO2} is instantly converted into its respective proportions (based on the local temperature, salinity, DIC, and alkalinity) of \ce{CO2}, \ce{HCO3^-}, and \ce{CO3^2-}. For the TC case, by contrast, \ce{CO2} persists for a finite amount of time before reacting and/or being removed by advection, leaving an increased surface concentration of \ce{CO2}, which slows fluxes. Notably, the same trend between the non-Langmuir case and $\text{La}_\mathrm{t} = 0.3$ is seen for each chemistry model, however, the difference between the two sets varies with the chemistry model (combined effects are discussed in the next section). Figure~\ref{fig:figure8}(b) shows the percent enhancement in the volume-integrated new DIC as a function of time for each of the chemistry models with respect to the non-Langmuir EC case. The enhancement is given by Eq.~\eqref{percent_enhance}, with $[\Delta c_\mathrm{DIC}]_\mathrm{base}$ now defined as the non-Langmuir EC case. This case is chosen as the baseline since this chemistry model and physical configuration resemble those used in ESMs. Once again, the general trends between the non-Langmuir and $\text{La}_\mathrm{t} = 0.3$ cases are consistent across the different chemistry models, with $\text{La}_\mathrm{t} = 0.3$ showing greater enhancement, but the detailed differences between these cases are dependent on the chemistry model (i.e., 0.12\%, 0.16\%, and 2.7\% increases in new DIC between the non-Langmuir and $\text{La}_\mathrm{t} = 0.3$ cases for the TC, EC, and NC models, respectively). Finally, Figure~\ref{fig:figure8}(c) shows time series of the change in air-sea flux rate of \ce{CO2} [defined in Eq.~\eqref{flux_enhance}] for the three chemistry models. The NC case undergoes a dramatic decline in air-sea flux rate as non-reactive aqueous \ce{CO2} builds up and slows fluxes. The two reactive cases (TC and EC), by contrast, have much higher flux rates. Both the TC and EC cases convert \ce{CO2} into \ce{HCO3^-} and \ce{CO3^2-}, which maintains a greater air-sea gradient of \ce{CO2} and allows more \ce{CO2} to enter the ocean. Comparing the two reactive cases, the EC case has an elevated flux rate over the TC case due to its faster reaction sink. \subsection{Combined Effects of Langmuir Turbulence and Chemical Model Fidelity} The previous sections have shown that both the enhanced vertical flux due to Langmuir turbulence and the chemical model fidelity affect the air-sea flux rate and reduce surface concentrations of \ce{CO2}, thereby impacting the DIC content of the oceanic mixed layer. Here, the combined effects of Langmuir turbulence enhancement and chemical model fidelity are considered. \begin{figure}[b!] \centering \includegraphics[width=\linewidth]{figure9} \caption{Percent enhancement of new DIC, denoted $E_\mathrm{DIC}$ and defined in Eq.~\eqref{percent_enhance}, in comparison to the non-Langmuir, time-dependent chemistry case for the three chemical models (a--c) and the four Langmuir cases. Gray dashed lines are least-squares fits for the decrease in percent enhancement as a function of Langmuir number and the gray number is the slope of the fit. All panels show results for Langmuir numbers $\mathrm{La}_\mathrm{t} = \infty$, 0.4, 0.3, and 0.2 (black, red, blue, and green symbols, respectively).} \label{fig:figure9} \end{figure} Figure~\ref{fig:figure9} shows the percent enhancement in the volume-integrated new DIC at six hours for each of the three chemistry models and for each of the four Langmuir cases. The expression in Eq.~(\ref{percent_enhance}) is once again used to compute the percentage of new DIC, and all values are now referenced to the non-Langmuir, time-dependent case, which occupies the zero value in panel (a) of Figure~\ref{fig:figure9}. Figure~\ref{fig:figure9} shows that the Langmuir flux enhancement is a function of the chemistry model, reflecting a complex, non-linear relationship between the chemical model and small-scale turbulence. Langmuir turbulence in the NC case provides a large enhancement over just wind-driven shear turbulence, while TC and EC cases have modest Langmuir enhancement. Consequently, the effect of Langmuir turbulence on air-sea fluxes of a gas varies substantially depending on whether the gas is reactive (e.g., \ce{CO2}), or non-reactive (e.g., oxygen). Smaller differences distinguish the two reactive cases. For the EC case, the Langmuir enhancement is greater than in the TC case, yet successive increases in Langmuir strength do not affect the EC case as much as in the TC case. \section{Discussion\label{sec:discuss}} In the following, the implications of the results described in Section \ref{results} are discussed with respect to ESMs, and variations in these results for different ocean conditions are outlined. The latter discussion is focused, in particular, on how the strength of the interactions between vertical mixing and chemical processes vary over the global ocean, as well as over diurnal and seasonal cycles. \subsection{Implications for Earth System Models} The results in Section \ref{langmuir effect} indicate that over the entire ocean surface, approximately \SIrange{0.07}{0.1}{\peta\gram} of extra carbon per year is brought into the ocean due to the presence of Langmuir turbulence. This estimate is based on an approximate global air-to-sea \ce{CO2} flux for 2000--2009 of \SI{80}{\peta\gram\carbon\per\year}; for reference, the estimated global net air-sea \ce{CO2} flux for this same period is \SI{2.3 \pm 0.7}{\peta\gram\carbon\per\year} \citep{ciais2013}. This assumes that these exact conditions remain constant throughout the year across the entire ocean surface and that the air concentration of \ce{CO2} is always 10\% greater than the mixed layer equilibrium concentration. Regional and seasonal deviations are likely, as imbalances will occur, for example, during upwelling, cooling, and warming events. These deviations are not random and may introduce systematic biases depending on the turbulence forcing mechanism. Nevertheless, the results in Section \ref{langmuir effect} indicate that Langmuir turbulence has a meaningful effect on the uptake of carbon by the ocean. Similarly, the results in Section \ref{chemistry effect} indicate that the finite-time delay in \ce{CO2} conversion due to the use of TC chemistry would result in a roughly \SI{0.1}{\peta\gram} decrease in the global uptake of carbon by the ocean in comparison with models that use the instantaneous EC chemistry. Even larger discrepancies in total domain carbon are found between the NC and other cases. These changes in \ce{CO2} flux rate with chemical model and Langmuir turbulence are on the same order as basin-scale differences in flux rate for different ESMs. Most current ESM simulations make two assumptions: (\emph{i}) that boundary layer turbulence effects on all chemical species can be parameterized in the same way, and (\emph{ii}) that carbonate chemistry is virtually instantaneous in comparison to turbulent processes, and thus can be represented by an equilibrium chemistry model. Results from the present study thus contradict these assumptions and indicate that errors from both assumptions combine in a complex and non-linear way. It should be noted, however, that errors resulting from the neglect of Langmuir turbulence and the use of equilibrium chemistry are likely to be dominated by errors in other physical models within ESMs, particularly for globally integrated annual quantities. The greatest impacts from the inclusion of a Langmuir parameterization and finite-rate chemistry are thus likely to be felt at regional spatial scales and over shorter time scales. \subsection{Dependence on Ocean Conditions\label{time}} The physical parameters chosen for the simulations, as outlined in Section \ref{physics} and summarized in Table \ref{tab:setup}, were selected on the basis of convenience combined with realism. These parameters result in potentially matched timescales for chemical and turbulent processes, and the extent of the matching can be expressed using various non-dimensional timescale ratios. To this end, the Damk\"{o}hler number is given as the ratio of the turbulent advection timescale, $\tau_\mathrm{t}$, to a characteristic timescale of the overall reaction process, $\tau_\mathrm{c}$, namely $\mathrm{Da} = \tau_\mathrm{t}/\tau_\mathrm{c}$. In general terms, $\tau_\mathrm{t}$ can be estimated from the integral or eddy turnover timescale. For a configuration similar to that studied here, \citet{teixeira2010} estimated a near-surface integral timescale of \SI{430}{\second} for a Langmuir turbulence simulation with $u_\tau = \SI{6.1e-3}{\meter \per \second}$. Calculations of the integral timescale in the present simulation yield a similar result, and so $\tau_\mathrm{t}$ can be estimated as $\tau_\mathrm{t} \approx \SI{400}{\second}$. In determining $\tau_\mathrm{c}$, it is common to use a characteristic timescale associated with the global rate of reaction, and the relaxation time after a 10\% perturbation to the concentration of \ce{CO2} (discussed in more detail in Section \ref{reactions}) gives $\tau_\mathrm{c} \approx \SI{60}{\second}$. As a result, Da in the present case can be estimated as $\mathrm{Da} \approx \SI{400}{\second} / \SI{60}{\second} = 6.7$. This value of Da indicates that interactions between Langmuir turbulence and carbonate chemical reactions are important, but that reactions are favored. This is consistent with results outlined in Section \ref{results}, particularly with respect to those in Figure \ref{fig:figure9} where variations in the chemical model fidelity were shown to have a larger impact on the amount of DIC in the mixed layer than the strength of Langmuir turbulence, at least for the present ocean conditions. Variations in the value of Da can be inferred for different ocean conditions from the analysis of \citet{teixeira2010}. In that study, $\tau_\mathrm{t} = \SI{430}{\second}$ was obtained from the turbulent kinetic energy $k$ and the dissipation rate $\varepsilon$ following the $k-\varepsilon$ modeling approach, which in later work was extended to Langmuir turbulence in a variety of settings \citep{grant2009,belcher2012}. This approach, validated against LES, predicts $\tau_\mathrm{t} \sim k/\varepsilon$, where \begin{linenomath*} \begin{equation} k\propto u_\tau^2 \quad \mathrm{or} \quad [u_\tau^2 u_\mathrm{s}(0)]^{2/3} \quad \mathrm{or} \quad (B_0 h_\mathrm{b})^{2/3} \,, \end{equation} \end{linenomath*} for wind-, wave-, and convection-dominated conditions, respectively, and \begin{linenomath*} \begin{equation} \varepsilon\propto \frac{2u_\tau^3[1-\exp(-\mathrm{La}_\mathrm{t}/2)]}{h_\mathrm{b}} + 0.22 \frac{u_\tau^2 u_\mathrm{s}(0)}{h_\mathrm{b}} + 0.3 B_0\,. \end{equation} \end{linenomath*} Here $B_0$ is the buoyancy flux and $h_\mathrm{b}$ is the turbulent boundary layer depth. Thus, for wind-dominated conditions $\mathrm{Da}\propto h_\mathrm{b} / [u_\tau (1 - \exp(-\mathrm{La}_\mathrm{t}/2))]$, for wave-dominated conditions $\mathrm{Da}\propto h_\mathrm{b} / [u_\tau^2 u_\mathrm{s}(0)]^{1/3}$, and for convection-dominated conditions $\mathrm{Da}\propto (h_\mathrm{b}^2/B_0)^{1/3}$. Observations suggest that the real ocean is typically somewhere between these different scalings \citep{belcher2012,li2017}. Assuming that $h_b$ is proportional to the initial mixed layer depth $H_{\text{ML},0}$, the scalings above indicate that $\mathrm{Da} \propto H_{\text{ML},0}$ or $H_{\text{ML},0}^{2/3}$. In the present simulations, $H_{\text{ML},0} = \SI{30}{\meter}$, but variations between \SI{10}{\meter} and \SI{500}{\meter} can occur in the real ocean depending on location and season \citep{li2016}. The depth $h_b$ has a similar range, but a value of \SI{30}{\meter} or less is typical in the tropics and during the summertime. This range roughly corresponds to a decrease in Da by a factor of 3 for the shallowest layers, or 15 times larger for the deepest layers. In the former case, this corresponds to a stronger interaction between Langmuir turbulence and carbonate chemistry, while the latter case corresponds to a weaker interaction. Assuming a fixed reaction rate (i.e., neglecting temperature and salinity effects on reaction rates) and using the \citet{large2009} monthly-mean wind stresses and the updated \citet{de2004} mixed layer depth climatology, $\mathrm{Da}$ based on wind stress scaling was estimated to have 90\% confidence limits of 3 and 20 with a median near 7. As higher-frequency winds tend to induce faster mixing during intermittent events, it is expected that these estimates are biased toward high Da. Surface cooling is not used here, but $B_0$ is an important scaling parameter for the amount of convective mixing in the world oceans, and varies from \SI{1e-9}{\meter^2 \second^{-3}} to \SI{5e-7}{\meter^2 \second^{-3}} by season and time of day. As a result, the dependence on $B_0$ might further decrease Da by a factor of 7 under extreme events, once again corresponding to stronger turbulence-chemistry interactions. Similarly, the typical wind stress over the ocean is \SI{0.1}{\newton \meter^{-2}}, or four times the value used here, which would roughly halve the value of Da (holding other parameters fixed). Using the \citet{large2009} monthly-mean buoyancy fluxes and the updated \citet{de2004} mixed layer depth climatology, the convection-based $\mathrm{Da}$ scaling using the same datasets had a factor of 7 spread for the 90\% confidence limits, although a median value cannot be calculated without the normalization factor from a convectively-forced LES. Note that it is also possible to define and estimate other relevant Damk\"ohler numbers. For example, a method to determine the smallest possible timescale of oceanic turbulence is to use the \citet{kolmogorov1941} time scale, typifying the timescale of the smallest-scale turbulence in the ocean. However, the vast range of energy dissipation rates throughout the world ocean does not make this estimate very precise \citep{pearson2018} beyond a range such as \SI{0.1}{\second} to \SI{1000}{\second}, which spans a wide range enclosing $\mathrm{Da}={\cal O}(1)$. Similarly, the second Damk\"ohler number, ${\rm Da}_2$, is the dimensionless ratio of mass diffusion timescale $\tau_\kappa$ to $\tau_\mathrm{c}$, namely \begin{linenomath*} \begin{equation} {\rm Da}_2=\frac{\tau_\kappa}{\tau_c}\,. \end{equation} \end{linenomath*} In the models used here, the subgrid-scale viscosity, buoyancy diffusivity, and tracer diffusivity are spatially-varying according to the scheme proposed by \citet{sullivan1994}. The diffusivities provided by this scheme are much larger than the molecular values for seawater. However, they are scaled in a flow-aware way through the LES approach so that $\mathrm{Da}\ll {\rm Da}_2$. As the simulations and Langmuir turbulence in the ocean are already in the regime where $\mathrm{Da}\ge 1$, it is expected that the consequences of using the LES diffusivities rather than the molecular diffusivities will be small since ${\rm Da}_2\gg \mathrm{Da}1\ge 1$. Note that the modeled LES diffusivities are not used for the Schmidt number when calculating the piston velocity from Eq.~\eqref{schmidt} \citep{wanninkhof1992}. Instead, the values of seawater Schmidt number consistent with observations are used, which probably includes the effects of turbulence not resolved in these simulations. In summary, the wind stress and surface cooling used here are conservative estimates of conditions typically observed in the ocean, while boundary layer depth is typical of the tropics or midlatitude summers. From the estimate $\mathrm{Da} \approx 6.7$, lower values of Da will result for stronger surface winds and cooling, while larger Da will result from deeper boundary layers. Thus, finite-time chemistry effects will typically be strongest during mixed layer deepening under strong wind and cooling events, such as cold air outbreaks. A full assessment of the climatology of Da for a full range of seasonal and regional conditions is beyond the scope here, but is planned for future work. \section{Conclusions} The interactions between carbonate chemical reactions and turbulent mixing in the upper ocean have been examined using LES for four different strengths of wave forcing and three different carbonate chemistry models, from infinitely slow non-reactive chemistry to infinitely fast equilibrium chemistry. The novel model in between is a time-dependent seven-species carbonate chemistry model that uses a QSS assumption for \ce{H^+} and is integrated using an efficient RKC solver that is robust for stiff problems. The results presented here indicate that enhanced vertical mixing by Langmuir turbulence results in a small, but measurable, increase in DIC in the ocean mixed layer as compared to a case with no Langmuir turbulence. Conversely, the use of an equilibrium chemical model results in a small, but measurable, reduction of DIC in the mixed layer as compared to a more realistic time-dependent model. The combined effects of Langmuir turbulence and chemical model fidelity are complicated and coupled, but the effects of Langmuir turbulence are more pronounced when using time-dependent chemistry than when using equilibrium chemistry. With respect to ESMs, this study has resulted in three major insights. First, compared with shear-only turbulence, Langmuir turbulence increases the flux rate of \ce{CO2} across the air-sea interface by approximately 0.1\%, or \SIrange{0.07}{0.1}{\peta\gram\carbon\per\year} globally. Second, the more accurate finite-time chemistry decreases the flux rate of \ce{CO2} into the domain by approximately 0.1\%, or \SI{0.1}{\peta\gram\carbon\per\year}, in comparison with equilibrium chemistry (and increases versus no chemistry). Third, Langmuir turbulence has a much greater effect on flux rates of a non-reactive gas such as oxygen than on a reactive gas such as \ce{CO2}. The magnitude of these differences is also expected to depend on other aspects of the ocean state, requiring further study in the future. In the future, additional research is required to determine whether Langmuir turbulence and finite-rate chemistry have different impacts at other ocean locations and for different conditions. The effects of wave breaking and bubbles are also likely to be important in air-sea fluxes of \ce{CO2}, and future simulations are planned using a bubble parameterization. Finally, carbonate chemistry evolution was only examined over a relatively short period in the present study, and longer simulations that incorporate diurnal and seasonal cycles will be performed in the future, thus providing more accurate estimates of the annual impacts of Langmuir turbulence and chemical model fidelity on \ce{CO2} flux rates. \acknowledgments KMS, PEH, and NSL were supported by NSF OCE-1258995, KEN was supported by NSF OAC-1535065, and BFK was supported by NSF OCE-1258907 and a grant from The Gulf of Mexico Research Initiative. The data analyzed in this paper are available from Mendeley Data (https://data.mendeley.com). This work utilized the RMACC Summit supercomputer supported by NSF (ACI-1532235, ACI-1532236), CU-Boulder, and CSU, as well as the Yellowstone (ark:/85065/d7wd3xhc) and Cheyenne (doi:10.5065/D6RX99HX) supercomputers provided by NCAR CISL, sponsored by NSF.

e63432e730fb28a2dd26d69db93b0c2f8b53f6e0

\section{Introduction} Recent years have seen the rise of community question answering forums, which allow users to ask questions and to get answers in a collaborative fashion. One issue with such forums is that duplicate questions easily become ubiquitous as users often ask the same question, possibly in a slightly different formulation, making it difficult to find the best (or one correct) answer \cite{cQA:Survey:2018,lai-bui-li:2018:C18-1}. Many forums allow users to signal such duplicates, but this can only be done after the duplicate question has already been posted and has possibly received some answers, which complicates merging the question threads. Discovering possible duplicates at the time of posting is much more valuable from the perspective of both (\emph{i})~the forum, as it could prevent a duplicate from being posted, and (\emph{ii})~the users, as they could get an answer immediately. \noindent Duplicate question detection is a special case of the more general problem of question-question similarity. The latter was addressed using a variety of textual similarity measures, topic modeling \cite{Cao:2008:RQU:1367497.1367509,zhang2014question}, and syntactic structure \cite{wang2009syntactic,SemEval2016:task3:KeLP,DaSanMartino:CIKM:2016,SemEval2016:task3:ConvKN,SemEval-2017:task3:KELP}. Another approach is to use neural networks such as feed-forward \cite{nakov-marquez-guzman:2016:EMNLP2016}, convolutional \cite{dossantos-EtAl:2015:ACL-IJCNLP,bonadiman-uva-moschitti:2017:EACLshort,Wang:2018:CAC:3183892.3151957}, long short-term memory \cite{Romeo:2016coling}, and more complex models \cite{LeiJBJTMM16,DBLP:conf/cikm/NicosiaM17,P18-2046,Joty:2018:multitask,Zhang:Wu:2018}. Translation models have also been popular \cite{zhou2011phrase,Jeon:2005:FSQ:1099554.1099572,ACL2016:MTE-NN-cQA,SemEval2016:task3:MTE-NN}. The above work assumes labeled training data, which exists for question-question similarity, e.g.,~from SemEval-2016/2017 \cite{agirre-EtAl:2016:SemEval1,nakov-EtAl:2016:SemEval,SemEval-2017:task3}, and for duplicate question detection, e.g.,~SemEval-2017 task 3 featured four StackExchange forums, \emph{Android}, \emph{English}, \emph{Gaming}, and \emph{Wordpress}, from CQADupStack \cite{hoogeveen2015cqadupstack,Hoogeveen+:2016a}. Yet, such annotation is not available for many other forums, e.g.,~the \emph{Apple} community on StackExchange. In this paper, we address this lack of annotation using adversarial domain adaptation (ADA) to effectively use labeled data from another forum. Our contributions can be summarized as follows: \vspace{-3pt} \begin{itemize}[leftmargin=10pt] \setlength\itemsep{-5pt} \item we are the first to apply adversarial domain adaptation to the problem of duplicate question detection across different domains;\footnote{The code and the data are available at the following link:\\ \url{http://github.com/darsh10/qra_code}} \item on the StackExchange family of forums, our model outperforms the best baseline with an average relative improvement of 5.6\% (up to 14\%) across all domain pairs. \item we study when transfer learning performs well and what properties of the domains are important in this regard; and \item we show that adversarial domain adaptation can be efficient even for unseen target domains, given some similarity of the target domain with the source one and with the regularizing adversarial domain. \end{itemize} Adversarial domain adaptation (ADA) was proposed by \citet{ganin2015unsupervised}, and was then used for NLP tasks such as sentiment analysis and retrieval-based question answering \cite{chen2016adversarial,Ganin:2016:DTN:2946645.2946704,Li:2017:EAM:3172077.3172199,liu2017adversarial,yu2017modelling,zhang2017aspect}, including cross-language adaptation \cite{joty-EtAl:2017:CoNLL} for question-question similarity.\footnote{Prior work on cross-language adaptation for question-question similarity used cross-language tree kernels \cite{DaSanMartino:2017:CQR:3077136.3080743}.} The rest of this paper is organized as follows: Section 2 presents our model, its components, and the training procedure. Section 3 describes the datasets we used for our experiments, stressing upon their nature and diversity. Section 4 describes our adaptation experiments and discusses the results. Finally, Section 5 concludes with possible directions for future work. \section{Method} \label{qsnn} \begin{figure}[t] \center \includegraphics[width=1.0\linewidth]{model2_001.jpeg} \caption{\small Our cross-domain question-question similarity model. The Question encoder is common for the questions from the source domain and from the target domain. The model and its training procedure are described in Section 2.} \label{fig:model} \end{figure} Our ADA model has three components: (\emph{i})~question encoder, (\emph{ii})~similarity function, and (\emph{iii})~domain adaptation component, as shown in Figure~\ref{fig:model}. The encoder $E$ maps a sequence of word tokens $x=(x_{1},..,x_{n})$ to a dense vector $\vec v = E(x)$. The similarity function $f$ takes two question vectors, $\vec v_1$ and $\vec v_2$, and predicts whether the corresponding questions are duplicates. \noindent The domain classifier $g$ takes a question vector $\vec v$ and predicts whether the question is from the source or from the target domain. We train the encoder not only to do well on the task for the source data, but also to fool the domain classifier, as shown in Algorithm~1. We describe the design choices considered for our domain adaptation model in the following two subsections. \subsection{Question Similarity Function} \label{simmodel} We consider two options for our similarity function $f(\vec v_{1}, \vec v_{2})$: \newline \noindent{(\emph{i})}~a logistic function that computes the probability that two questions are similar/duplicates, which is trained with the cross-entropy loss: \vspace{-0.5em} \begin{align*} \text{sigmoid}\left({\vec W}^\top ({\vec v_1}\odot{\vec v_2}) + {\vec b}\right) \end{align*} where $\odot$ is an element-wise vector product between unit encodings of questions; \newline \noindent (\emph{ii)}~a simple cosine similarity function, i.e.,~$cosine( \vec v_{1}, \vec v_{2})$, trained using the pairwise hinge loss with a margin $m$: \vspace{-0.5em} $$\sum_{i} max(\{(1-y^{i})f(\vec v^{i}_{1}, \vec v^{i}_{2}) + m - y^{i}f(\vec v^{i}_{1}, \vec v^{i}_{2})\} ,0)$$ Our experiments reported in Table~\ref{table:2} show that the $cosine$ similarity function performs far better. \subsection{Domain Adaptation Components} The adversarial component is responsible for reducing the difference between the source and the target domain distributions. There are two common approaches to achieve this: (\emph{i})~classification-based~\cite{ganin2015unsupervised} and (\emph{ii})~Wasserstein~\cite{arjovsky2017wasserstein}. The main difference between them is in the way the domain discrepancy loss is computed. In the classification-based approach, the adversarial component is a classifier trained to correctly predict the domain (source vs.~target) of the input question. In contrast, the question encoder is optimized to confuse the domain classifier, which, as a result, encourages domain invariance. \newcite{arjovsky2017towards} showed that this adversarial optimization process resembles minimizing the Jenson-Shannon (JS) divergence between the source $P_{s}$ and the target distribution $P_{t}$: $$JS(P_{s}, P_{t}) = KL(P_{s},P_{m}) + KL(P_{g},P_{m})$$ \noindent where $P_{m} = (P_{s}+P_{t})/2 $ and KL is the Kullback-Leibler divergence. \noindent In contrast, the Wasserstein method attempts to reduce the approximated Wasserstein distance (also known as \emph{Earth Mover's Distance}) between the distributions for the source and for the target domain as follows: $$W(P_{s}, P_{t}) = \underset{||f||_{L}\leq1}{sup} \E_{x\sim{}P_s}[f(x)] - \E_{x\sim{}P_t}[f(x)]$$ \noindent where $f$ is a Lipchitz-1 continuous function realized by a neural network. \citet{arjovsky2017wasserstein} have shown that the Wasserstein method yields more stable training for computer vision tasks. \subsection{Training Procedure} Algorithm 1 describes the procedure to train the three components of our model. Adversarial training needs two kinds of training data: (\emph{i})~annotated question pairs from the source domain, and (\emph{ii})~unlabeled questions from the source and the target domains. The question encoder is trained to perform well on the source domain using the similarity classification loss $L_{c}$. In order to enforce good performance on the target domain, the question encoder is simultaneously trained to be incapable in discriminating between question pairs from the source vs. the target domain. This is done through the domain classification loss $L_{d}$. \begin{algorithm}[t] \label{alg:training} \footnotesize \caption*{\textbf{Algorithm 1}: Training Procedure } \begin{algorithmic} \State Input: source data $X^{s}$; target data $X^{t}$ \State Hyper-parameters: learning rates $\alpha_{1},\alpha_{2}$; batch size $m$; adversarial importance $\lambda$ \State Parameters to be trained: question encoder $\theta_{e}$, question similarity classifier $\theta_{s}$ and domain classifier $\theta_{d}$ \State Similarity classification loss $L_{c}$ is either the cross-entropy loss or hinge loss, described in Section 2.1 \State Adversarial loss $L_{d}$, described in Section 2.2 \Repeat \For{\texttt{each batch}} \State Construct a sub-batch of similar and dissimilar question pairs from the annotated source data $\{(x_{i_1}^s,x_{i_2}^s),y_i^s\}_{i=1}^m$ \State Calculate the classification loss $L_{c}$ using $\theta_{e}$ and $\theta_{s}$ for this sub-batch \State Construct a sub-batch of questions $\{x_{i}^s,x_{j}^t\}_{i=1}^m$ from the corpora of source and target domains \State Calculate the domain discrepancy loss $L_{d}$ using $\theta_{e}$ and $\theta_{d}$ for this sub-batch \State Total loss $L = L_{c} - \lambda L_{d}$ \State $\theta_{e} = \theta_{e} - \alpha_{1} {\triangledown_{\theta_{e}}{L}}$ \State $\theta_{s} = \theta_{s} - \alpha_{1} {\triangledown_{\theta_{s}}{L}}$ \State $\theta_{d} = \theta_{d} + \alpha_{2} {\triangledown_{\theta_{d}}{L}}$ \EndFor \Until{ $\theta_{e}$, $\theta_{s}$ and $\theta_{d}$ converge} \end{algorithmic} \end{algorithm} \begin{table}[t] \begin{center} \begin{adjustbox}{max width=\linewidth} \begin{tabular}{ m{4.5em} m{3.9em} m{4em} m{2.5em} m{2.5em} m{2.5em} } \toprule \bf Dataset & \bf Questions & \bf Duplicates & \bf Train & \bf Dev & \bf Test\\ \midrule AskUbuntu & 257,173 & 27,289 & 9,106 & 1,000 & 1,000 \\ SuperUser & 343,033 & 11,407 & 9,106 & 1,000 & 1,000 \\ Apple & 80,466 & 2,267 & -- & 1,000 & 1,000 \\ Android & 42,970 & 2,371 & -- & 1,000 & 1,000 \\ \midrule Sprint & 31,768 & 23,826 & 9,100 & 1,000& 1,000 \\ \midrule Quora & 537,211 & 149,306 & 9,100 & -- & -- \\ \bottomrule \end{tabular} \end{adjustbox} \end{center} \caption{\small Statistics about the datasets. The table shows the number of question pairs that have been manually marked as similar/duplicates by the forum users (i.e.,~positive pairs). We further add 100 negative question pairs per duplicate question by randomly sampling from the full corpus of questions.} \label{table:6} \end{table} \section{Datasets} The datasets we use can be grouped as follows: \begin{itemize}[leftmargin=10pt] \item{\bf Stack Exchange} is a family of technical community support forums. We collect questions (composed by title and body) from the XML dumps of four forums: \emph{AskUbuntu}, \emph{SuperUser}, \emph{Apple}, and \emph{Android}. Some pairs of similar/duplicate questions in these forums are marked by community users. \item{\bf Sprint FAQ} is a newly crawled dataset from the Sprint technical forum website. It contains a set of frequently asked questions and their paraphrases, i.e., three similar questions, paraphrased by annotators. \item{\bf Quora} is a dataset of pairs of similar questions asked by people on the Quora website. They cover a broad set of topics touching upon philosophy, entertainment and politics. \end{itemize} Note that these datasets are quite heterogeneous: the \emph{StackExchange} forums focus on specific technologies, where questions are informal and users tend to ramble on about their issues, the \emph{Sprint FAQ} forum is technical, but its questions are concise and shorter, and the \emph{Quora} forum covers many different topics, including non-technical. Statistics about the datasets are shown in Table~\ref{table:6}. Moreover, in order to quantify the differences and the similarities, we calculated the fraction of unigrams, bigrams and trigrams that are shared by pairs of domains. Table~\ref{table:0} shows statistics about the $n$-gram overlap between \emph{AskUbuntu} or \emph{Quora} as the source and all other domains as the target. As one might expect, there is a larger overlap within the \emph{StackExchange} family. \section{Experiments and Evaluation} \subsection{Experimental Setup} \begin{table}[t] \begin{center} \footnotesize \begin{adjustbox}{max width=1.05\linewidth} \begin{tabular}{ m{4.2em} m{4.2em} m{3.4em} m{3.2em} m{3.2em} } \toprule \multicolumn{1}{c}{\bf Source} & \multicolumn{1}{c}{\bf Target} & \multicolumn{1}{c}{\bf Unigrams} & \multicolumn{1}{c}{\bf Bigrams} & \multicolumn{1}{c}{\bf Trigrams} \\ \hline \multirow{4}{*}{AskUbuntu} & Android &0.989 & 0.926 & 0.842\\ & Apple & 0.991&0.926&0.853\\ & SuperUser & 0.990&0.921&0.822\\ \cline{2-5} & Sprint &0.959 &0.724&0.407\\ & Quora & 0.922 &0.696 &0.488\\ \hline \multirow{4}{*}{Quora} & AskUbuntu & 0.949&0.647&0.326\\ & Apple & 0.969&0.721&0.426\\ & Android & 0.973&0.762&0.473\\ & SuperUser & 0.958 & 0.663 & 0.338\\ \cline{2-5} & Sprint & 0.942&0.647 & 0.310 \\ \bottomrule \end{tabular} \end{adjustbox} \end{center} \caption{\footnotesize Proportion of $n$-grams that are shared between the source and the target domains.} \label{table:0} \end{table} \textbf{Baselines} We compare our ADA model to the following baselines: (\emph{a})~\emph{direct transfer}, which directly applies models learned from the source to the target domain without any adaptation; and (\emph{b})~the standard unsupervised \emph{BM25} ~\cite{robertson2009probabilistic} scoring provided in search engines such as Apache Lucene ~\cite{mccandless2010lucene}. \textbf{Models} We use a bi-LSTM~\cite{hochreiter1997long} encoder that operates on 300-dimensional GloVe word embeddings ~\cite{pennington2014glove}, which we train on the combined data from all domains. We keep word embeddings fixed in our experiments. For the adversarial component, we use a multi-layer perceptron. \textbf{Evaluation Metrics} As our datasets may contain some duplicate question pairs, which were not discovered and thus not annotated, we end up having false negatives. Metrics such as MAP and MRR are not suitable in this situation. Instead, we use AUC (area under the curve) to evaluate how well the model ranks positive pairs vs. negative ones. AUC quantifies how well the true positive rate ($tpr$) grows at various false positive rates ($fpr$) by calculating the area under the curve starting from $fpr = 0$ to $fpr = 1$. We compute the area integrating the false positive rate ($x$-axis) from $0$ up to a threshold $t$, and we normalize the area to $[0,1]$. This score is known as AUC$(t)$. It is more stable than MRR and MAP in our case when there could be several false negatives.\footnote{For illustration, say of 100 candidates, 2 false negatives are ranked higher than the correct pair, the AUC score drops by 3 points (linear drop), as compared to the 66.67 point drop for MRR. We can avoid the expensive manually tagging of the negative pairs for experiments by using the AUC score.} \begin{table}[t] \begin{center} \footnotesize \begin{tabular}{cccc} \toprule \bf Adaptation & \bf Similarity & \bf AUC(0.05) & \bf AUC(0.1)\\ \hline --- & Sigmoid & 0.431 & 0.557 \\ --- & Cosine & 0.692 & 0.782 \\ \hline Classification & Cosine & 0.791 & 0.862 \\ Wasserstein & Cosine & 0.795 & 0.869 \\ \bottomrule \end{tabular} \end{center} \vspace{-.5em} \caption{\footnotesize Duplicate question detection: direct transfer vs. adversarial domain adaptation from \emph{AskUbuntu} to \emph{Android}. } \label{table:2} \end{table} \subsection{Choosing the Model Components} \textbf{Model Selection} We select the best components for our domain adaptation model via experimentation on the \emph{AskUbuntu}--\emph{Android} domain pair. Then, we apply the model with the best-performing components across all domain pairs. \textbf{Hyperparameters} We fine-tune the hyper-parameters of all models on the development set for the target domain. \textbf{Similarity Function} Table \ref{table:2} shows the AUC at 0.05 and 0.1 for different models of question similarity, training on \emph{AskUbuntu} and testing on \emph{Android}. The first row shows that using cosine similarity with a hinge loss yields much better results than using a cross-entropy loss. This is likely because (\emph{i})~there are some duplicate question pairs that were not tagged as such and that have come up as negative pairs in our training set, and the hinge loss deals with such outliers better. (\emph{ii})~The cosine similarity is domain-invariant, while the weights of the feed-forward network of the softmax layers capture source-domain features. \textbf{Domain Adaptation Component} We can see that the Wasserstein and the classification-based methods perform very similarly, after proper hyper-parameter tuning. However, Wasserstein yields better stability, achieving an AUC variance 17 times lower than the one for classification across hyper-parameter settings. Thus, we chose it for all experiments in the following subsections. \begin{table}[t] \begin{center} \footnotesize \begin{tabular}{ m{4.5em} m{4.5em} m{2.6em} m{2.5em} m{2.5em} } \toprule \bf Source & \bf Target & \bf Direct & \bf BM25 & \bf Adv. \\ \hline \multirow{4}{*}{AskUbuntu} & Android & 0.692 & 0.681 & \textbf{0.790}\\ & Apple & 0.828 & 0.747 & \textbf{0.855}\\ & SuperUser & 0.908 & 0.765 & \textbf{0.911}\\ & Sprint & 0.917 & \textbf{0.956} & 0.937 \\ \hline \multirow{3}{*}{SuperUser} & AskUbuntu & 0.730 & 0.644 & \textbf{0.796}\\ & Apple & 0.828 & 0.747 & \textbf{0.861}\\ & Android & 0.770 & 0.681 & \textbf{0.790}\\ & Sprint & 0.928 & \textbf{0.956} & 0.932 \\ \bottomrule \end{tabular} \end{center} \vspace{-.5em} \caption{\footnotesize Domain adaptation for the \emph{StackExchange} source-target domain pairs when using the \emph{Direct} approach, \emph{BM25}, and our adaptation model, measured with AUC(0.05).} \label{table:3} \end{table} \subsection{When Does Adaptation Work Well?} Tables~\ref{table:3} and ~\ref{table:4} study the impact of domain adaptation when applied to various source-target domain pairs, using the \emph{Direct} approach, \emph{BM25}, and our adaptation model. We can make the following observations: \begin{itemize} \item For almost all source--target domain pairs from the \emph{StackExchange} family, domain adaptation improves over both baselines, with an average relative improvement of 5.6\%. This improvement goes up to 14\% for the \emph{AskUbuntu}--\emph{Android} source--target domain pair. \item Domain adaptation on the \emph{Sprint} dataset performs better than direct transfer, but it is still worse than \emph{BM25}. \item Domain adaptation from \emph{Quora} performs the worst, with almost no improvement over direct transfer, which is far behind \emph{BM25}. \item The more similar the source and the target domains, the better our adaptation model performs. \end{itemize} Table~\ref{table:0} shows that \emph{AskUbuntu} has high similarity to other \emph{StackExchange} domains, lower similarity to \emph{Sprint}, and even lower similarity to \emph{Quora}. The Pearson coefficient \cite{myers2010research} between the $n$-gram fractions and the domain adaptation effectiveness for unigrams, bigrams and trigrams is 0.57, 0.80 and 0.94, respectively, which corresponds to moderate-to-strong positive correlation. This gives insight into how simple statistics can predict the overall effectiveness of domain adaptation. \begin{table}[t] \begin{center} \footnotesize \begin{tabular}{ m{4.5em} m{4.5em} m{2.6em} m{2.5em} m{2.5em} } \toprule \bf Source & \bf Target & \bf Direct & \bf BM25 & \bf Adv. \\ \hline \multirow{3}{*}{Sprint} & AskUbuntu & 0.615 & \textbf{0.644} & 0.615\\ & Apple & 0.719 & \textbf{0.747} & 0.728\\ & Android & 0.627 & \textbf{0.681} & 0.648\\ & Sprint & 0.977 & 0.956 & -- \\ & SuperUser & 0.795 & 0.765 & \textbf{0.795}\\ \hline \multirow{3}{*}{Quora} & AskUbuntu & 0.446 & \textbf{0.644} & 0.446 \\ & Apple & 0.543 & \textbf{0.747} & 0.543 \\ & Android & 0.443 & \textbf{0.681} & 0.460 \\ & Sprint & 0.786 & \textbf{0.956} & 0.794 \\ & SuperUser & 0.624 & \textbf{0.765} & 0.649 \\ \bottomrule \end{tabular} \end{center} \vspace{-.5em} \caption{\footnotesize Domain adaptation results when using \emph{Sprint} and \emph{Quora} as the source domains with the \emph{Direct} approach, \emph{BM25}, and our adaptation model, measured with AUC(0.05).} \label{table:4} \end{table} \begin{table}[h] \begin{center} \footnotesize \begin{tabular}{ m{5em} m{4.5em} m{2.5em} m{3.5em} } \toprule {\textbf{Pivot}}$\backslash${\textbf{Target}} & SuperUser & Apple & Android \\ \hline SuperUser & 0.911 & 0.827 & 0.678 \\ Apple & 0.900 & 0.855 & \textbf{0.711} \\ Android & 0.904 & \textbf{0.843} & 0.790 \\ Quora & 0.906 & 0.815 & 0.673 \\ \hline Direct & 0.908 & 0.828 & 0.692 \\ \bottomrule \end{tabular} \end{center} \vspace{-.5em} \caption{\footnotesize AUC(0.05) of ADA to unseen domains, with \emph{AskUbuntu} as a source.} \label{table:7} \end{table} \subsection{Adapting to Unseen Domains} We also experiment with domain adaptation to a target domain that was not seen during training (even adversarial training). We do so by training to adapt to a pivot domain different from the target. Table~\ref{table:7} shows that this yields better AUC compared to \emph{direct transfer} when using \emph{Apple} and \emph{Android} as the pivot/target domains. We hypothesize that this is due to \emph{Apple} and \emph{Android} being closely related technical forums for iOS and Android devices. This sheds some light on the generality of adversarial regularization. \section{Conclusion and Future Work} We have applied and analyzed adversarial methods for domain transfer for the task of duplicate question detection; to the best of our knowledge, this is the first such work. Our experiments suggest that (\emph{i})~adversarial adaptation is rather effective between domains that are similar, and (\emph{ii})~the effectiveness of adaptation is positively correlated with the $n$-gram similarity between the domains. In future work, we plan to develop better methods for adversarial adaptation based on these observations. One idea is to try source-pivot-target transfer, similarly to the way this is done for machine translation \cite{Wu2007}. Another promising direction is to have an attention mechanism \cite{luong2015effective} for question similarity which can be adapted across domains.\footnote{In our experiments, we found that using attention was lowering the adaptation performance. Upon analysis, we found that adversarial techniques alone were not enough to make the attention weights domain-invariant.} \section*{Acknowledgments} This research was carried out in collaboration between the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Computing Research Institute (QCRI), HBKU.

55cd8d7f560302f7bfa3ac12fad3107b10c6e8bd

\part{\@startsection{part}{0 \z@{\linespacing\@plus\linespacing}{.5\linespacing {\normalfont\bfseries\centering}} \makeatother \newtheorem{lem}{Lemma} \newtheorem{thm}{Theorem} \newtheorem{prop}{Proposition} \newtheorem{cor}{Corollary} \theoremstyle{remark} \newtheorem{remark}{Remark} \newtheorem{example}{Example} \setlength\parindent{0pt} \newcommand{\rr}{\mathbb{R}} \newcommand{\ff}{\mathbb{F}} \newcommand{\supp}{\operatorname{supp}} \usepackage{setspace} \usepackage{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \DeclareMathOperator{\dist}{dist} \DeclareMathOperator{\adj}{adj} \DeclareMathOperator{\dive}{div} \newcommand{\rn}{\mathbb{R}^{n}} \newcommand{\rk}{\mathbb{R}^{k}} \newcounter{co} \newtheorem{definition}[cor]{Definition} \newtheorem{theorem}[cor]{Theorem} \newtheorem{lemma}[cor]{Lemma} \newtheorem{propn}[cor]{Proposition} \newtheorem{conjecture}[cor]{Conjecture} \newtheorem{question}[cor]{Question} \title[Multilinear Duality and Factorisation with Applications]{Multilinear Duality and Factorisation for Brascamp--Lieb-type Inequalities with applications} \author{Anthony Carbery, Timo S. H\"anninen and Stef\'an Ingi Valdimarsson} \address{Anthony Carbery, School of Mathematics and Maxwell Institute for Mathematical Sciences, University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Road, King's Buildings, Mayfield Road, Edinburgh, EH9 3FD, Scotland.} \email{[email protected]} \address{Timo S. H\"anninen, Department of Mathematics and Statistics, University of Helsinki, P.O. Box 68, FI-00014 Helsinki, Finland.} \email{[email protected]} \address{Stef\'an Ingi Valdimarsson, \'Islandsbanki, Hagasm\'ari 3, 201 K\'opavogur, Iceland, and Science Institute, University of Iceland, Dunhagi 5, 107 Reykjav\'ik, Iceland} \email{[email protected]} \date{4th September 2018} \begin{document} \begin{abstract} We initiate the study of a duality theory which applies to norm inequalities for pointwise weighted geometric means of positive operators. The theory finds its expression in terms of certain pointwise factorisation properties of function spaces which are naturally associated to the norm inequality under consideration. We relate our theory to the Maurey--Nikisin--Stein theory of factorisation of operators, and present a fully multilinear version of Maurey's fundamental theorem on factorisation of operators through $L^1$. The development of the theory involves convex optimisation and minimax theory, functional-analytic considerations concerning the dual of $L^\infty$, and the Yosida--Hewitt theory of finitely additive measures. We explore the connections of the theory with the theory of interpolation of operators. We discuss the ramifications of the theory in the context of concrete families of geometric inequalities, including Loomis--Whitney inequalities, Brascamp--Lieb inequalities and multilinear Kakeya inequalities. \end{abstract} \maketitle \section{Introduction}\label{intro} In this paper we shall be mainly concerned with norm inequalities for pointwise weighted geometric means $$ \prod_{j=1}^d (T_j f_j(x))^{\alpha_j}$$ of positive linear operators $T_j$ defined on suitable spaces, where $\alpha_j \geq 0$ and $\sum_{j=1}^d \alpha_j = 1$. Before we describe the scope of our work in this paper, and to set the scene for our study, we briefly visit the analogous territory in the linear setting ($d=1$) in order to help provide a context for what we are aiming to achieve. Throughout the whole paper we shall be dealing with real-valued rather than complex-valued functions. \subsection{The linear setting}\label{linset} \medskip \noindent Let $X$ and $Y$ be measure spaces and let $T: L^p(Y) \to L^q(X)$ be a bounded linear operator, that is, it satisfies \begin{equation}\label{lin1} \|Tf\|_q \leq A \|f\|_p \end{equation} for all $f \in L^p(Y)$, for some $A > 0$. Here, $1 \leq p \leq \infty$ and $0 < q < \infty$. Since $L^q$ is a Banach space only when $q \geq 1$, it is natural to focus separately on the regimes $q \geq 1$ and $0 < q < 1$. \medskip \noindent (i) {\bf Case $q \geq 1$}. Since $\|h\|_q = \max\{ |\int hg| \, : \, \|g\|_{q'} = 1\}$, inequality \eqref{lin1} holds if and only if for all $ f \in L^p$ and {\em all} $g \in L^{q'}$ we have \begin{equation}\label{lin2} | \int (Tf) g| \leq A \|f\|_p\|g\|_{q'}. \end{equation} Using the relation $\int (Tf) g = \int f (T^* g)$, this is in turn equivalent to the statement $\|T^*g\|_{p'} \leq A \|g\|_{q'}$ -- that is the boundedness of the adjoint operator $T^*$ between the dual spaces of $L^q$ and $L^p$ respectively, (at least when $1 < p, \, q < \infty$). We are therefore firmly in the terrain of classical linear duality theory, a theory whose utility and importance cannot be overstated. Notice that if $q = 1$ and $T$ is also assumed to be positive (that is, $Tf \geq 0$ whenever $f \geq 0$), the equivalence of \eqref{lin1} and \eqref{lin2} is essentially without content since in this case it suffices to check on the function $g \equiv 1$. \medskip \noindent (ii) {\bf Case $0 <q <1$}. Since $\|h\|_q = \min\{ |\int hg| \, : \, \|g\|_{q'} = 1\}$, we have that \eqref{lin1} holds if and only if for all $f \in L^p$ {\em there exists} an (extended real-valued) $g \in L^{q'}$ such that \eqref{lin2} holds. Note that $q' < 0$ in this situation, so it is implicit that such a $g$ satisfies $g(x) \neq 0$ almost everywhere. It is a remarkable result of Maurey, that, under certain conditions -- such as positivity of $T$ -- given inequality \eqref{lin1}, there exists a {\em single} $g\in L^{q'}$ with $\|g\|_{q'} = 1$\footnote{Of course the quantity $\left(\int|g|^{r}\right)^{1/r}$ does not define a norm when $r \in (- \infty, 1)$.} such that \eqref{lin2} holds for all $f \in L^p$. Such a result is an instance of the celebrated theory of factorisation of operators which is developed in \cite{Mau}. Indeed, it is a case of {\em factorisation through $L^1$} since the inequality $$ | \int (Tf) g| \leq A \|f\|_p$$ demonstrates that $T$ may be factorised as $T = M_{g^{-1}} \circ S$ where $S = M_g \circ T$ satisfies $\|S\|_{L^p \to L^1} \leq A$ and $M_{g^{-1}}$, the operator of multiplication by $g^{-1}$, satisfies $\|M_{g^{-1}}\|_{L^1 \to L^{q}} = \|g\|_{q'}^{-1} = 1$. \bigskip \noindent Observe that there is no obvious point of direct contact between the two regimes $q \geq 1$ and $0 < q < 1$ in this linear setting. \medskip \noindent The result of Maurey to which we refer falls within the wider scope of Maurey--Nikisin--Stein theory, which considers factorisation of operators in a broad variety of contexts. This includes consideration of non-positive operators, sublinear operators (for example maximal functions), operators with various domains and codomains, and factorisation through various weak- and strong-type spaces, often under some auxiliary hypotheses. The particular case of positive operators defined on normed lattices, taking values in $L^q$ for $q < 1$, and factorising through (strong-type) $L^1$ was considered by Maurey, however, and for this reason we refer specifically to the Maurey theory rather than the broader Maurey--Nikisin--Stein theory. For an overview of this larger theory see \cite{GCRdeF}, \cite{Gilb}, \cite{Mau}, \cite{Pis}. \subsection{The multilinear setting}\label{multlinset} The purpose of this paper is to develop duality and factorisation theories for certain classes of multilinear operators which are analogous to those that we have set out above in the linear setting. Amusingly, the notion of ``factorisation'' manifests itself in two distinct ways in our development. One of these is as a multilinear analogue of a formulation of a Maurey-type theorem as was briefly outlined in the discussion of the case $0 < q < 1$ above. The other is that our duality theory (corresponding to the case $1 \leq q < \infty$) will be expressed in terms of pointwise factorisation properties of certain spaces of functions. \medskip \noindent We begin by describing the scenario in which we shall work and the classes of operators we shall consider. \medskip \noindent Let $(X,{\rm d}\mu)$ and $(Y_j,{\rm d}\nu_j)$, for $j=1,\dots,d$, be measure spaces,\footnote{Throughout the paper, when we refer to measure spaces $X$, $Y$ or $Y_j$ without explicit mention of the measure, it is implicit that the corresponding measures are $\mu$, $\nu$ and $\nu_j$ respectively, unless the context demands otherwise.} let $\mathcal{S}(Y_j)$ denote the class of real-valued simple functions (i.e. finite linear combinations of characteristic functions of measurable sets of {\em finite} measure) on $Y_j$, and let $\mathcal{M}(X)$ denote the class of real-valued measurable functions on $X$. Let $T_1,\dots,T_d$ be linear maps $$ T_j : \mathcal{S}(Y_j) \to \mathcal{M}(X).$$ We suppose throughout that the $T_j$ are positive in the sense that if $f \geq 0$ almost everywhere on $Y_j$, then $T_j f \geq 0$ almost everywhere on $X$. \medskip In this paper we shall be concerned with ``multilinear'' Lebesgue-space inequalities of the form \begin{equation} \label{premainineq} \left\|\prod_{j=1}^d (T_jF_j)^{\beta_j}\right\|_{L^q(X)} \leq C \prod_{j=1}^d \Big\|F_j\Big\|_{L^{p_j}(Y_j)}^{\beta_j} \end{equation} where $0 < \beta_j < \infty$, $0 < p_j \leq \infty$\footnote{We shall soon focus on the case $p_j \geq 1$ and $\sum_j \beta_j = 1$.}and $0 < q \leq \infty$. \medskip \noindent These inequalities are to be interpreted in an {\em a priori} sense, with the $F_j$ being simple functions defined on $Y_j$. We are especially interested in the case that either the $T_j$ are not bounded operators from $L^{p_j}(Y_j, {\rm d} \nu_j)$ to $L^q(X, {\rm d} \mu)$, or that they are bounded but do not enjoy effective bounds. \medskip \noindent Strictly speaking such inequalities are multilinear only when each $\beta_j = 1$; we shall nevertheless abuse language and will refer to the inequalities under consideration as ``multilinear''. In fact the case when $\sum_{j=1}^d \beta_j =1$ will play a special role in what follows. Of course we may always assume either that $q=1$ or that $\sum_{j=1}^d \beta_j = 1$. \medskip \noindent To fix ideas, we discuss some examples of inequalities falling under the scope of our study. \medskip \noindent \subsection{Examples}\label{exples} \begin{example}\label{Holder}{\rm [H\"older's inequality] The multilinear form of H\"older's inequality for nonnegative functions is simply $$ \int_X F_1(x)\cdots F_d(x) {\rm d} \mu(x) \leq \|F_1 \|_{L^{p_1}(X)} \cdots\|F_d \|_{L^{p_d}(X)}$$ where $p_j > 0$ and $\sum_{j=1}^d p_j^{-1} = 1$. This is of the form \eqref{premainineq}, with $T_j = I$ for all $j$, $q=1$ and each $\beta_j = 1$. But, for any fixed set of positive exponents $\{\beta_j\}$, it is also trivially equivalent to the inequality $$ \| f_1^{\beta_1} \cdots f_d^{\beta_d} \|_q \leq \|f_1 \|_{q_1}^{\beta_1} \cdots\|f_d \|_{q_d}^{\beta_d}$$ for all $ 0 < q_j < \infty$ and $0 <q < \infty$ which satisfy $\sum_{j=1}^d \beta_j q_j^{-1} = q^{-1}$. In particular, there is an equivalent formulation of the multilinear H\"older inequality taking the form \eqref{premainineq} with $\sum_{j=1}^d\beta_j = 1$. In fact there are many such equivalent forms, limited only by the requirement that $\sum_{j=1}^d \beta_j q_j^{-1} = q^{-1}$. Special cases of choices of exponents $\{\beta_j, q_j, q\}$ satisfying this condition are (i) $\beta_j$ arbitrary subject to $\sum_{j=1}^d \beta_j = 1$, $q_j =1$ for all $j$, and $q =1$; and (ii) $\beta_j = d^{-1}$ for all $j$, $q_j$ arbitrary subject to $\sum_{j=1}^d q_j^{-1} = 1$, and $q = d$. This observation demonstrates that we may expect that a given multilinear inequality might have {\em multiple} equivalent manifestations, each of the form \eqref{premainineq}, with $\sum_{j=1}^d \beta_j = 1$. } \end{example} \begin{example}\label{Loomis--Whitney} {\rm [Loomis--Whitney inequality] For $1 \leq j \leq n$ let $\pi_j: \mathbb{R}^n \to \mathbb{R}^{n-1}$ be projection on the coordinate hyperplane perpendicular to the standard unit basis vector $e_j$; that is, $\pi_j x = (x_1, \dots, \widehat{x_j}, \dots, x_n)$. The Loomis--Whitney inequality \cite{LW} is $$ |\int_{\mathbb{R}^n} F_1(\pi_1 x) \cdots F_n(\pi_n x) \, {\rm d} x| \leq \|F_1\|_{L^{n-1}(\mathbb{R}^{n-1})} \cdots \|F_n\|_{L^{n-1}(\mathbb{R}^{n-1})}.$$ For each $0 < p < \infty $, this is equivalent to the inequality $$ \|f_1(\pi_1 x)^{1/n} \cdots f_n(\pi_n x)^{1/n} \|_{L^{np/(n-1)}(\mathbb{R}^n)} \leq \|f_1\|_{L^{p}(\mathbb{R}^{n-1})}^{1/n} \cdots \|f_n\|_{L^{p}(\mathbb{R}^{n-1})}^{1/n}.$$ Each of these inequalities is of the form \eqref{premainineq} with $\sum_{j=1}^n \beta_j = 1$. \medskip \noindent A very special case of the Loomis--Whitney inequality occurs in two dimensions where it becomes the trivial identity $$ \int_{\mathbb{R}^2} F_1(x_2)F_2(x_1) \; {\rm d} x_1 {\rm d} x_2 = \int_\mathbb{R} F_1 \int_\mathbb{R} F_2.$$ In spite of its simplicity, this example will play an important guiding role for us. See Sections~\ref{factinterp}, \ref{Loomis--Whitneyrevisited}, \ref{NLLW} and \ref{BLfactdetails}. \medskip \noindent The Loomis--Whitney inequality has many variants -- for example Finner's inequalities, the affine-invariant Loomis--Whitney inequality and the nonlinear Loomis--Whitney inequaity. See \cite{Fi}, \cite{BCW}, and Sections ~\ref{Loomis--Whitneyrevisited} and \ref{NLLW}. } \end{example} \begin{example}\label{BL} {\rm [Brascamp--Lieb inequalities] The class of Brascamp--Lieb inequalities includes the previous examples. Let $B_j : \mathbb{R}^n \to \mathbb{R}^{n_j}$ be linear surjections, $ 1 \leq j \leq d$. For $0 < p_j < \infty$ and $F_j$ nonnegative we consider the Brascamp--Lieb inequality \begin{equation}\label{BL1} |\int_{\mathbb{R}^n} \prod_{j=1}^d F_j(B_jx)^{p_j}\; {\rm d} x |\leq C \prod_{j=1}^d \left(\int_{\mathbb{R}^{n_j}} F_j \right)^{p_j}. \end{equation} It is not hard to see that in order for this inequality to hold with a finite constant $C$, it is necessary that $\sum_{j=1}^d p_j n_j = n$. It is known that the constant $C$ is finite if and only if, in addition to $\sum_{j=1}^d p_j n_j = n$, it holds that $$ {\rm dim} \, V \leq \sum_{j=1}^d p_j {\rm dim} B_j V$$ for all $V$ in the lattice of subspaces of $\mathbb{R}^n$ generated by $\{ \ker B_j\}_{j=1}^d$. (See \cite{BCCT1}, \cite{BCCT2} and \cite{Vald}.) From this one sees easily that $\cap_{j=1}^d {\rm ker} B_j = \{ 0 \}$, $\sum_{j=1}^d p_j \geq 1$, and $p_j \leq 1$ are also necessary conditions for the finiteness of $C$. A celebrated theorem of Lieb \cite{L} states that the value of the best constant $C$ is obtained by checking the inequality on Gaussian inputs $F_j$. Lieb's theorem generalises Beckner's theorem \cite{Beckner} on extremisers for Young's convolution inequality. \medskip \noindent Suppose that $0 < r_j < \infty$ and $0 < s < \infty$. Setting $F_j = f_j^{r_j}$ in \eqref{BL1} and taking $s$'th roots, we see that \eqref{BL1} is equivalent to $$\| \prod_{j=1}^d f_j(B_jx)^{p_jr_j/s} \|_{L^s(\mathbb{R}^n)} \leq {C}^{1/s} \prod_{j=1}^d \|f_j \|_{L^{r_j}(\mathbb{R}^{n_j})}^{p_jr_j/s}.$$ If $\sum_{j=1}^d p_j r_j = s$ this is an inequality of the form \eqref{premainineq} with $\sum_{j=1}^d \beta_j = 1$. In particular we can take $r_j = n_j$ and $s =n$ to obtain the equivalent form $$ \| \prod_{j=1}^d f_j(B_jx)^{p_jn_j/n} \|_{L^n(\mathbb{R}^n)} \leq {C}^{1/n} \prod_{j=1}^d \| f_j \|_{L^{n_j}(\mathbb{R}^{n_j})}^{p_jn_j/n};$$ or we can take $r_j = 1$ and $s = \sum_{j=1}^d p_j$ (recall that this number is at least $1$ when the inequality is nontrivial) to obtain another equivalent form \begin{equation}\label{BL3} \| \prod_{j=1}^d f_j(B_jx)^{p_j/s} \|_{L^s(\mathbb{R}^n)} \leq {C}^{1/s} \prod_{j=1}^d \| f_j \|_{L^{1}(\mathbb{R}^{n_j})}^{p_j/s}. \end{equation} \medskip \noindent A special case of the class of Brascamp--Lieb inequalities is the class of {\em geometric} Brascamp--Lieb inequalities. Suppose that the linear surjections $B_j: \mathbb{R}^n \to \mathbb{R}^{n_j}$ satisfy $$ \sum_{j=1}^d p_j B_j^\ast B_j = I_n.$$ Then, by a result of Ball and Barthe, (\cite{Ball} and \cite{Barthe}, see also \cite{BCCT1}) we have \begin{equation}\label{BL2} |\int_{\mathbb{R}^n} \prod_{j=1}^d F_j(B_jx)^{p_j}\; {\rm d} x |\leq \prod_{j=1}^d \left(\int_{\mathbb{R}^{n_j}} F_j \right)^{p_j}, \end{equation} and the sharp constant $1$ is achieved by the standard Gaussians $F_j(y) = e^{- \pi |y|^2}$. Correspondingly, in the equivalent variants presented above, the constants are also $1$. The geometric Brascamp--Lieb inequalities include a suitably reformulated version of the sharp Young inequality of Beckner. } \end{example} \begin{example}\label{mgrt} {\rm [Multilinear generalised Radon transforms] There is a vast literature on multilinear generalised Radon transforms into which we do not wish to enter. For us, this term will mean consideration of multilinear inequalities of the form \eqref{premainineq} when the operators $T_j$ take the form $T_jf = f \circ B_j$ for suitable mappings $B_j: X \to Y_j$. In most cases, $X$ and $Y_j$ will be endowed with a topological or smooth structure, and the mappings $B_j$ will respect that structure in such a way that issues of measurability do not arise. \medskip \noindent The class of multilinear generalised Radon transforms includes the Brascamp--Lieb inequalities. The most basic multilinear generalised Radon transform which is not included in the Brascamp--Lieb inequalities is probably the nonlinear Loomis--Whitney inequality. See Section~\ref{NLLW} below. } \end{example} \begin{example}\label{MK} {\rm [Multilinear Kakeya inequalities] The Loomis--Whitney inequality of Example \ref{Loomis--Whitney} is equivalent to $$ \int_{\mathbb{R}^n} \prod_{j=1}^n \left(\sum_{P _j \in \mathcal{P}_j} a_{P_j} \chi_{P_j}(x) \right)^{1/(n-1)} \; {\rm d}x \leq \prod_{j=1}^n \left(\sum_{P_j \in \mathcal{P}_j} a_{P_j}\right)^{1/(n-1)},$$ where $\mathcal{P}_j$ is a finite family of $1$-tubes in $\mathbb{R}^n$ which are parallel to the $j$'th standard basis vector $e_j$, and the $a_{P_j}$ are arbitrary positive numbers. (A $1$-tube is simply a neighbourhood of a doubly infinite line in $\mathbb{R}^n$ which has $(n-1)$-dimensional cross-sectional area equal to $1$.) Multilinear Kakeya inequalities have the same set-up, but now we allow the tubes in the family $\mathcal{P}_j$ to be {\em approximately} parallel to $e_j$, i.e. the direction $e(P) \in \mathbb{S}^{n-1}$ of the central axis of the tube $P \in \mathcal{P}_j$ must satisfy $|e(P) - e_j| \leq c_n$ where $c_n$ is a small dimensional constant. Such inequalities have been studied in \cite{BCT}, \cite{MR2746348}, \cite{BG} and \cite{CV} and have proved to be very important over the last decade with significant applications in partial differential equations and especially in number theory -- see for example \cite{B7}, \cite{B8}, \cite{B9} and \cite{B10}. The multilinear Kakeya inequality is the statement $$ \Big\|\prod_{j=1}^n \left(\sum_{P _j \in \mathcal{P}_j} a_{P_j} \chi_{P_j}(x) \right)^{1/n} \Big\|_{L^{n/(n-1)}(\mathbb{R}^n)} \leq C_n \prod_{j=1}^n \left(\sum_{P_j \in \mathcal{P}_j} a_{P_j}\right)^{1/n}.$$ This inequality is of the form \eqref{premainineq} with $X = \mathbb{R}^n$, $q = n/(n-1)$, $Y_j = \mathcal{P}_j$ with counting measure, $p_j = 1$ for all $j$, $\beta_j = 1/n$ for all $j$, and $T ((a_{P_j}))(x) = \sum_{P _j \in \mathcal{P}_j} a_{P_j} \chi_{P_j}(x)$. It was Guth's approach to such multilinear Kakeya inequalities in \cite{MR2746348} which inspired the present paper. \medskip \noindent The recent multilinear $k_j$-plane Kakeya inequalities, and indeed the even more general perturbed Brascamp--Lieb inequalities, both recently established by Zhang \cite{Z}, also fit into our framework, the latter as a generalisation of inequality \eqref{BL3}. } \end{example} \bigskip \noindent We shall return to consider these examples in some detail later in Part III. In particular we shall discuss the affine-invariant Loomis--Whitney inequality, the nonlinear Loomis--Whitney inequality and certain aspects of Brascamp--Lieb inequalities in the light of the theory we develop. \subsection{The weighted geometric mean operator}\label{wgmo} \noindent As we have just seen, all of our examples fit into the framework of inequality \eqref{premainineq} with $\sum_{j=1}^d \beta_j = 1$, and with the $L^{p_j}$ (and $L^q$) spaces in the Banach regime, i.e. with $p_j \geq 1$ (and $q \geq 1$). We shall therefore be concerned in this paper with norm inequalities for the {\em weighted geometric mean} operator $$ \mathcal{T}_\alpha:(f_1, \dots, f_d) \mapsto (T_1 f_1)^{\alpha_1} \cdots (T_d f_d)^{\alpha_d}$$ where $\alpha = (\alpha_1, \dots, \alpha_d)$ and the $\alpha_j$ are positive numbers satisfying $ \sum_{j=1}^d \alpha_j = 1$. That is, we shall consider inequalities of the form \begin{equation} \label{submainineq} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)}\leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation} for nonnegative simple functions $f_j \in \mathcal{S}(Y_j)$, in the regime $p_j \geq 1$ and $q > 0$. While the case $q \geq 1$ is pertinent to our examples, we also wish to consider the case $0 < q < 1$ because this corresponds to the situation treated by Maurey in the linear setting. Throughout the paper, we use the quantities $\alpha_j$ to represent positive numbers whose sum is $1$. \medskip \noindent We have chosen to present our general theory for the weighted geometric mean operator $\mathcal{T}_\alpha$ -- which is manifestly {\em not} linear in its arguments $f_1, \dots , f_d$ -- mainly because of the extra elegance and simplicity that such a treatment affords. Nevertheless, nearly all of the examples above also have equivalent strictly multilinear formulations. In particular, the multilinear Kakeya inequality of Example~\ref{MK} can be re-cast as the manifestly multilinear $$ \Big\|\prod_{j=1}^n \left(\sum_{P _j \in \mathcal{P}_j} \beta_{P_j} \chi_{P_j}(x) \right) \Big\|_{L^{1/(n-1)}(\mathbb{R}^n)} \leq C \prod_{j=1}^n \left(\sum_{P_j \in \mathcal{P}_j} \beta_{P_j}\right).$$ The one class of examples that does not admit a genuinely multilinear reformulation consists of the perturbed Brascamp--Lieb inequalities which were briefly mentioned in Example~\ref{MK}. \medskip \noindent Our first purpose in this paper is to propose and undertake a systematic study of the duality theory associated to the weighted geometric mean operator $\mathcal{T}_\alpha$ in the context of inequality \eqref{submainineq} in the case $q \geq 1$, and some of its generalisations. It is hoped that the framework for this multilinear duality theory will in time have applications in a wide variety of contexts. Our second purpose is to establish suitable analogues of Maurey's theorems in the context of \eqref{submainineq} in the case $0 < q < 1$. Interestingly, the case $q=1$ will be central to our development of both the regimes $q \geq 1$ and $0 < q < 1$, unlike in the classical linear setting where the case $q=1$ is essentially vacuous, and in which there appears to be no direct link between the two regimes $q \geq 1$ and $0 < q < 1$. \subsection{A theory of multilinear duality -- the regime $q \geq 1$}\label{mdbanach} \medskip \noindent We begin with the Banach regime $q \geq 1$. \medskip \noindent One half of our duality theory -- the `easy' half -- is largely contained in the following simple observation, the content of which is that if we have a certain pointwise factorisation property for the space $L^{q^\prime}$, then the weighted geometric mean norm inequality \eqref{submainineq} will hold. \begin{prop} \label{thmguthbaby} Suppose that $T_j : L^{p_j}(Y_j) \to L^q(X)$ are positive linear operators, that $p_j, q \geq 1$ and that $\sum_{j=1}^d \alpha_j = 1$. Suppose that for every nonnegative $G \in L^{q^\prime}(X)$ there exist nonnegative measurable functions $g_j$ defined on $X$ such that \begin{equation} \label{mainprobbaby} \begin{aligned} &G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\text{a.e. on $X$,}\\ \text{and}\qquad &\Big\|\,T_j^\ast g_j\Big\|_{L^{{p_j}^\prime}(Y_j)} \leq A\Big\|G\Big\|_{L^{q^\prime}(X)}\qquad\text{for all $j$.} \end{aligned} \end{equation} Then, for all $f_j \in \mathcal{Y}_j$, $$ \; \; \; \; \; \qquad\left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \; ;$$ that is, \eqref{submainineq} holds, for all $f_j \in L^{p_j}(Y_j)$. \end{prop} \medskip \noindent For the (easy) proof and some discussion of this result, see the more general Proposition~\ref{thmguth} below. \medskip \noindent Rather surprisingly, the implication in Proposition \ref{thmguthbaby} can be essentially reversed, and one of the main aims of this paper is to show that the factorisation property \eqref{mainprobbaby} enunciated in Proposition \ref{thmguthbaby} is in fact {\em necessary} as well as sufficient for \eqref{submainineq} to hold. This is the second half of the multilinear duality principle referred to in the abstract of the paper. \medskip \noindent Before doing this, however, we note that if there is a subset of $X$ of positive measure upon which $T_jf_j$ vanishes for all $f_j \in L^{p_j}$, then this subset will play no role in the analysis of inequality \eqref{submainineq}. There is therefore no loss of generality in assuming such subsets do not exist. We formalise this notion by introducing the notion of {\bf saturation} below.\footnote{For a related notion, see \cite{Zaanen}.} In order to facilitate what follows later, we at the same time introduce two closely related notions, and also make the definitions in slightly greater generality than what is required by the current discussion. The definitions apply to linear operators $T: \mathcal{Y} \to \mathcal{M}(X)$, with $\mathcal{Y}$ a normed lattice and $(X, {\rm d} \mu)$ a measure space, which are positive in the sense that for every nonnegative $f \in \mathcal{Y}$ we have $Tf \geq 0$. (What is currently relevant is the fact that the space of simple functions defined on a measure space $Y$, together with the $L^p$ norm for $p \geq 1$, forms a normed lattice.) \begin{definition}\label{saturation} \noindent (i) We say that $T$ {\bf saturates} $X$ if for each subset $E \subseteq X$ of positive measure, there exists a subset $E^\prime \subseteq E$ with $\mu(E^\prime) > 0$ and a nonnegative $h \in \mathcal{Y}$ such that $Th > 0$ a.e. on $E^\prime$. \smallskip \noindent (ii) We say that $T$ {\bf strongly saturates} $X$ if there exists a nonnegative $h \in \mathcal{Y}$ such that $Th > 0$ a.e. on $X$. \smallskip \noindent (iii) We say that $T$ {\bf super-strongly saturates} $X$ if there exists a nonnegative $h \in \mathcal{Y}$ such that $Th$ is a.e. bounded away from $0$ on $X$. \end{definition} For further discussion of the relevance of these conditions see Remarks~\ref{bnm} and \ref{mnb} below. \medskip \noindent Now we can state one of the main results of the paper: \begin{thm} \label{thmmainbaby} Suppose that $X$ and $Y_j$, for $j=1,\dots,d$, are $\sigma$-finite measure spaces. Suppose that the linear operators $T_j: \mathcal{S}(Y_j) \to \mathcal{M}(X)$ are positive and that each $T_j$ saturates $X$. Suppose that $p_j \geq 1$, $ 1 \leq q < \infty$,\footnote{See Remark~\ref{mmmmm} for the case $q = \infty$.} $\sum_{j=1}^d \alpha_j = 1$ and that \begin{equation*} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation*} for all simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then for every nonnegative $G \in L^{q^\prime}(X)$ there exist nonnegative measurable functions $g_j$ on $X$ such that \begin{equation}\label{factorisebaby} G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\mbox{a.e. on $X$,} \end{equation} and such that for each $j$, \begin{equation}\label{controlbaby} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|G\|_{L^{q^{\prime}}}\|f_j\|_{p_j} \end{equation} for all nonnegative simple functions $f_j$ on ${Y}_j$. \end{thm} This result will follow from the more general Theorem~\ref{thmmain} below. (Note that we have used the formulation \eqref{controlbaby} instead of one explicitly involving $T_j^\ast$ as we did in \eqref{mainprobbaby} because it is not immediately clear how $T_j^*$ should be defined this context.) \medskip \noindent The special case of Theorem~\ref{thmmainbaby} corresponding to $q=1$ and $G \equiv 1$ can be singled out: \begin{cor} \label{cormainbaby} Suppose that $X$ and $Y_j$, for $j=1,\dots,d$, are $\sigma$-finite measure spaces. Suppose that the operators $T_j: \mathcal{S}(Y_j) \to \mathcal{M}(X)$ are positive and that each $T_j$ saturates $X$. Suppose that $p_j\geq 1$, $\sum_{j=1}^d \alpha_j = 1$ and that \begin{equation*} \int_X \prod_{j=1}^d (T_jf_j)^{\alpha_j} {\rm d} \mu \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation*} for all nonnegative simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then there exist nonnegative measurable functions $g_j$ on $X$ such that \begin{equation*} 1 \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\mbox{a.e. on $X$,} \end{equation*} and such that for each $j$, \begin{equation*} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A \|f_j\|_{p_j} \end{equation*} for all nonnegative simple functions $f_j$ on ${Y}_j$. \end{cor} \medskip \noindent In fact, Theorem~\ref{thmmainbaby} is equivalent to Corollary~\ref{cormainbaby}. Indeed, suppose that \begin{equation*} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation*} for all simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then, for all nonnegative $G \in L^{q'}(X)$ with $\|G\|_{L^{q'}}=1$, we have \begin{equation*} \int_X \prod_{j=1}^d (T_jf_j)^{\alpha_j} G \, {\rm d} \mu \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation*} for all nonnegative simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. It is easy to see that if $T_j$ saturates $X$ with respect to the measure ${\rm d} \mu$, then it also does so with respect to $G \, {\rm d}\mu$. By Corollary~\ref{cormainbaby} applied with the measure $G \, {\rm d} \mu$ in place of ${\rm d} \mu$, there are nonnegative measurable functions $\gamma_j$ such that \begin{equation*} 1 \leq \prod_{j=1}^d \gamma_j(x)^{\alpha_j} \qquad G \, {\rm d} \mu \mbox{-a.e. on $X$,} \end{equation*} and such that for each $j$, \begin{equation*} \int_X \gamma_j(x)T_jf_j(x) G(x) {\rm d}\mu(x) \leq A \|f_j\|_{p_j} \end{equation*} for all nonnegative simple functions $f_j$ on ${Y}_j$. Setting $g_j = \gamma_j G$ gives the desired conclusion of Theorem~\ref{thmmainbaby}. \medskip \noindent \begin{remark}\label{mmmmm} Upon a more careful inspection, we see that the argument presented above in fact gives an upgrade of Theorem~\ref{thmmainbaby} to include the case $q = \infty$, and, when $1 < q \leq \infty$, to remove the hypothesis that $X$ be $\sigma$-finite. Indeed, when $1 < q \leq \infty$ and $G \in L^{q'}(X, {\rm d} \mu)$, it is an easy exercise to show that the measure $G \, {\rm d} \mu$ is $\sigma$-finite, irrespective of whether or not $\mu$ is a $\sigma$-finite measure, and the remainder of the argument is unchanged. Similar observations will generate a corresponding upgrade to Theorem~\ref{thmmain} below, see Remark~\ref{mmmm} below. \end{remark} \subsection{Multilinear Maurey-type factorisation -- the regime $0 < q < 1$.}\label{mmf} We now state a multilinear Maurey-type theorem. \begin{thm} \label{thmmainmaurey} Suppose that $X$ and $Y_j$, for $j=1,\dots,d$, are $\sigma$-finite measure spaces. Suppose that the operators $T_j: \mathcal{S}(Y_j) \to \mathcal{M}(X)$ are positive and that each $T_j$ saturates $X$. Suppose that $p_j \geq 1$, $0 < q <1$, $\sum_{j=1}^d \alpha_j = 1$ and that \begin{equation}\label{987} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{L^{p_j}(Y_j)}^{\alpha_j} \end{equation} for all simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then there exist nonnegative measurable functions $g_j$ on $X$ such that \begin{equation}\label{factorisebabymaurey} \|\prod_{j=1}^d g_j(x)^{\alpha_j}\|_{L^{q'}(X)} = 1 \end{equation} and such that for each $j$, \begin{equation}\label{controlbabymaurey} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|f_j\|_{p_j} \end{equation} for all nonnegative simple functions $f_j$ on ${Y}_j$. \end{thm} We shall give the proof of this result, as a consequence of Corollary~\ref{cormainbaby}, in Section~\ref{maureysec} below. Conversely, it is easy to see using H\"older's inequalty that if there exist $g_j$ such that \eqref{factorisebabymaurey} and \eqref{controlbabymaurey} hold, then so does \eqref{987}. \medskip \noindent This result can be seen as a factorisation result in the spirit of Maurey: if we let $S_j f_j(x) = g_j(x) T_j f_j(x)$, $\mathcal{S}_\alpha(f_1, \dots , f_d) = \prod_{j=1}^d(S_j f_j)^{\alpha_j}$ and $g(x) = \prod_{j=1}^d g_j(x)^{\alpha_j}$, then $$ \mathcal{T}_\alpha = M_{g^{-1}} \circ \mathcal{S}_\alpha$$ where $$ \|S_j\|_{L^{p_j} \to L^1} \leq A $$ for all $j$, and $$ \|M_{g^{-1}}\|_{L^1 \to L^q} = \|g\|_{q'}^{-1} = 1.$$ In fact it has the rather strong conclusion that {\em each} $S_j$ is bounded from $L^{p_j}(Y_j)$ to $L^1( {\rm d} \mu)$ with constant at most $A$ (rather than the much weaker corresponding statement for the geometric mean $\mathcal{S}_\alpha$ alone). \medskip \noindent Other, different, versions of multilinear Maurey-type theorems have been studied. See for example \cite{Schep} and \cite{Diestel}. For some extensions of our work on Maurey-type theorems, see Section~\ref{future} below. \subsection{Structure of the paper} The paper is divided into three parts. \medskip \noindent In Part I (Sections 2--5) we present the theory of multilinear duality and factorisation and prove the main theorems. \medskip \noindent In Section~\ref{sdmr} we state and discuss the main results at some length. The principal result is Theorem~\ref{thmmain}. Taken together with Proposition~\ref{thmguth}, Theorem~\ref{thmmain} forms the statement of the multilinear duality principle referred to in the abstract of the paper. (Theorem~\ref{thmmainbaby} and Proposition~\ref{thmguthbaby} presented in the introduction are more readily digested versions of Theorem~\ref{thmmain} and Proposition~\ref{thmguth} respectively.) The multilinear Maurey factorisation theorem, Theorem~\ref{thmmainmaureyx}, is proved as a consequence of Theorem~\ref{thmmain}. (A more accessible version of Theorem~\ref{thmmainmaureyx} is found as Theorem~\ref{thmmainmaurey} in the introduction.) \medskip \noindent In Section~\ref{outline} we give a proof of a finitistic case of Theorem~\ref{thmmain} which recognises and emphasises its structure as a convex optimisation or minimax problem. This perspective sets the scene for the remainder of the theoretical part of the paper. In this case, none of the functional-analytic and measure-theoretic difficulties that we encounter later are present. However, Section~\ref{outline} is not strictly speaking logically necessary for the development of the theory. \medskip \noindent In Section~\ref{overv} we begin to address the proof of Theorem~\ref{thmmain}. Our strategy will be to first consider the setting of finite measure spaces. Theorem~\ref{thmsimple} is the main result in this case, and while it represents an important step in the proof of Theorem~\ref{thmmain}, it is not subsumed by it, and it has its own independent interest. Already in Theorem~\ref{thmsimple} we are faced with substantial functional-analytic and measure-theoretic difficulties. These derive from the need to establish certain compactness statements necessary for the application of a minimax theorem. Briefly, they involve working with the dual space of $L^\infty(X)$, and dealing with various issues in the theory of finitely additive measures. \medskip \noindent In Section~\ref{dett} we give the details of the proofs of Theorem~\ref{thmsimple} and Theorem~\ref{thmmain}. We begin with a couple of technical but very important lemmas. Next, we pass to the proof of the finite-measure result, Theorem~\ref{thmsimple}, via the minimax theory. Finally, for general $\sigma$-finite $X$, we ``glue together'' factorisations obtained for subsets $X$ of finite measure via Theorem~\ref{thmsimple}, and we obtain the factorisations needed in Theorem~\ref{thmmain}. \medskip \noindent In a much shorter Part II, we begin to explore connections with other topics -- in particular the theory of interpolation in Sections~\ref{factinterp} and \ref{qzf}, and the extent to which the theory might apply in the context of more general multilinear operators in Section~\ref{culture}. \medskip \noindent Finally, in Part III, we revisit the examples discussed earlier in this introduction in the light of the multilinear duality theory which has been developed. In Section~\ref{classical} we give factorisation-based proofs of the affine-invariant Loomis--Whitney inequality (see Section~\ref{Loomis--Whitneyrevisited}) and the sharp nonlinear Loomis--Whitney inequality (see Section~\ref{NLLW}). In Section~\ref{GBL} we pose an interesting question related to the sharp Young convolution inequality and geometric Brascamp--Lieb inequalities, while in Section~\ref{GGBL} we propose an algorithm for factorising the general Brascamp--Lieb inequality. In Section~\ref{MKrevisited} we revisit the multilinear Kakeya inequality which inspired the paper in the light of the findings of Section~\ref{culture}, and make an observation about the size of the constant in the finite-field version of the multilinear Kakeya inequality which is derived from our methods. \subsection{Future work}\label{future} In the forthcoming paper \cite{CHV}, we further develop our theory of multilinear duality and factorisation in various ways, two of which we now briefly outline. \medskip \noindent Firstly, we give an alternate approach to the minimax arguments of Section~\ref{dett}, which, in certain cases, bypasses the need to consider $(L^\infty)^\ast$. This leads to formulations of Theorems~\ref{thmmain} and \ref{thmmainmaureyx} under the more natural hypotheses that each $\mathcal{Y}_j$ is a normed lattice and each $T_j$ merely saturates the measure space $X$. \medskip \noindent Secondly, we refine multilinear factorisation (Theorem~\ref{thmmainmaureyx}) in the presence of $p$-convexity: if each lattice $\mathcal{Y}_j$ is $p_j$-convex, then we can conclude that the $T_j$ factorise through $L^{p_j}$ rather than just through $L^1$. More precisely, if $\sum_{j=1}^d \alpha_j/p_j=1$\footnote{This is the one place where we break our standing convention that $\sum_{j=1}^d \alpha_j = 1$.} and inequality~\eqref{tez} holds, then we can conclude the existence of measurable $g_j$ such that \eqref{factorisebabymaureyx} holds, and such that \eqref{controlbabymaureyx} holds with $L^{p_j}$ norms replacing $L^1$ norms on the left-hand side. \subsection{Acknowledgements} This paper has benefited substiantally from discussions with many individuals. In particular we should like to thank Keith Ball, Jon Bennett, Michael Christ, Michael Cowling, Alastair Gillespie, Gilles Pisier, Sandra Pott, Stuart White and Jim Wright for the insights they have shared with us on various aspects of the material of the paper. The first author is especially grateful to Michael Christ for his ongoing encouragement in the quest to find explicit factorisations. The second author is grateful to Professor I.~Verbitsky for introducing him to the Maurey theory of factorisation in their collaboration \cite{HV} on characterising two-weight norm inequalities via factorisation. The first and third author would like to record their appreciation of the hospitality and support of the Isaac Newton Institute during the programme ``Discrete Analysis'' between March and July 2011, where the preliminary stages of this research were carried out and presented. The second author is supported by the Academy of Finland through funding of his postdoctoral researcher post (Funding Decision No 297929), and he is a member of the Finnish Centre of Excellence in Analysis and Dynamics Research. \part{Statements and proofs of the theorems} \bigskip \section{Statement and discussion of the main results}\label{sdmr} It turns out that the theory we shall develop is most naturally presented in a more general setting. Moreover, limiting ourselves to the Lebesgue spaces $L^{p_j}$ and $L^q$ in the multilinear duality theory is unnecessarily restrictive. For example, one may wish to consider multilinear inequalities of the form~\eqref{submainineq} in which the $L^{p_j}$ and $L^q$ spaces are replaced by certain Lorentz spaces, Orlicz spaces or mixed-norm spaces, especially if the inequality under consideration is an endpoint inequality. We therefore introduce a more general framework in which we consider suitable spaces $\mathcal{X}$ and $\mathcal{Y}_j$ corresponding to $L^q(X)$ and $L^{p_j}(Y_j)$ respectively. Thus, for $\sum_{j=1}^d \alpha_j = 1$, which we recall is a standing convention, we now consider inequalities of the form \begin{equation} \label{mainineq} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{\mathcal{X}} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j}. \end{equation} \medskip \noindent Each $\mathcal{Y}_j$ will be a normed lattice. In many cases, $\mathcal{Y}_j$ will be a normed lattice of measurable functions defined on a $\sigma$-finite measure space $Y_j$ such that $\mathcal{S}(Y_j) \subseteq \mathcal{Y}_j$, but in some parts of the theory we can accommodate abstract normed lattices. \medskip \noindent On the other hand, we will take $\mathcal{X}$ to be Banach space of locally integrable\footnote{That is, $\int_E |f| {\rm d} \mu < \infty$ whenever $\mu(E) < \infty$.} functions defined on $X$, which contains the simple functions, and is such that if $f \in \mathcal{M}(X)$ and $g \in \mathcal{X}$ satisfy $|f(x)| \leq |g(x)|$ a.e., then $f \in \mathcal{X}$ and $\|f\|_{\mathcal{X}} \leq \|g\|_{\mathcal{X}}$. {\em These properties of $\mathcal{X}$ we shall assume throughout without any further mention.} We may as well also assume that $X$ is a complete measure space. (In almost all cases\footnote{The exceptions are in Remarks~\ref{mmmmm} and \ref{mmmm}.} our measure space $X$ will be assumed to the $\sigma$-finite; these properties together then identify $\mathcal{X}$ as a {\bf K\"othe space}, see \cite{LT}, Vol.~II, p.~28.) Natural examples of spaces with these properties include $L^q$ for $1 \leq q \leq \infty$. We shall need a suitable primordial dual of $\mathcal{X}$, denoted by $\mathcal{X}'$, and defined to be $$\mathcal{X}^\prime = \{ g \in \mathcal{M}(X) \, : \, \|g\|_{\mathcal{X}'} = \sup_{ \|f\|_\mathcal{X} \leq 1} \int |fg|\, {\rm d} \mu < \infty \}.$$ If $\mathcal{X} = L^q$ for $1 \leq q \leq \infty$, then $\mathcal{X}^\prime = L^{q'}$ where $1/q + 1/ q^\prime = 1$. It is clear that $\mathcal{X}'$ is a linear space which contains the simple functions (as $\mathcal{X}$ is contained in the class of locally integrable functions) and is contained in the class of locally integrable functions (as $\mathcal{X}$ contains the simple functions). The quantity $ \|g\|_{\mathcal{X}^\prime}$ defines a seminorm on $\mathcal{X}'$ which is a norm (under the usual convention of almost-everywhere equal functions being identified) provided that $X$ is $\sigma$-finite. While by definition we always have the H\"older inequality $$ |\int fg \, {\rm d} \mu| \leq \|f\|_{\mathcal{X}} \|g \|_{\mathcal{X'}},$$ it may or may not be the case that $\mathcal{X}'$ is {\bf norming} (for $\mathcal{X}$), i.e. that \begin{equation}\label{norming} \|f\|_{\mathcal{X}} = \sup \{\, | \int fg \, {\rm d} \mu |\, : \, \|g\|_{\mathcal{X}^\prime} \leq 1 \} \end{equation} holds for all $f \in \mathcal{X}$.\footnote{By a result of Lorentz and Luxemburg (see \cite{LT}, Vol.~II, p.~29), if $X$ is a K\"othe space, $\mathcal{X}'$ is norming if and only if $\mathcal{X}$ has the so-called Fatou property, that is, whenever $f_n \in \mathcal{X}$ are such that $f_n \to f$ a.e., with $f_{n+1} \geq f_n \geq 0$, then $\|f_n \|_\mathcal{X} \to \|f\|_\mathcal{X}$. This is automatic when $\mathcal{X}$ is separable. If $\mathcal{X}$ is $L^\infty$ then \eqref{norming} holds by inspection since $\mathcal{X}^\prime$ is simply $L^1$. We shall need the notion of norming only for Proposition~\ref{thmguth} and Remark~\ref{mmmm}.} Indeed, when $X$ is $\sigma$-finite, $\mathcal{X}'$ is isometrically embedded in the dual $\mathcal{X}^\ast$, but the two spaces may not coincide in general. \medskip {\em From now on, we shall adopt once and for all the convention that all named functions ($f, g, h, F, G$, $H, \beta, G, S, \psi$ etc., often adorned with subscripts) are assumed to be nonnegative. The two exceptions to this are the functions $L$ and $\Lambda$ appearing in the proofs of the main results.} \subsection{Duality theory -- easy half}\label{deh} The easy half of our duality theory is expressed in the following simple observation, the content of which is that if we have a certain factorisation property for the space $\mathcal{X}^\prime$, then the weighted geometric mean norm inequality \eqref{mainineq} will hold. \begin{prop} \label{thmguth} Suppose that $T_j : \mathcal{Y}_j \to \mathcal{X}$ are positive linear operators where $\mathcal{X}'$ is norming for $\mathcal{X}$ and the $\mathcal{Y}_j$ are normed lattices. Suppose furthermore that for every $G \in \mathcal{X}^\prime$ there exist measurable functions $g_j$ on $X$ such that \begin{equation} \label{mainprob} \begin{aligned} &G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\text{a.e. on $X$,}\\ \text{and}\qquad &\Big\|\,T_j^\ast g_j\Big\|_{\mathcal{Y}_j^\ast} \leq A\Big\|G\Big\|_{\mathcal{X}^\prime}\qquad\text{for all $j$.} \end{aligned} \end{equation} Then, for all $f_j \in \mathcal{Y}_j$ $$ \; \; \; \; \; \qquad\left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{\mathcal{X}} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j} \; .$$ That is, \eqref{mainineq} holds, for all $f_j \in \mathcal{Y}_j$. \end{prop} \begin{proof} Take $f_j\in\mathcal{Y}_j$ for $j=1,\dots,d$, and $G\in\mathcal{X}^\prime$ with $\|G\|_{\mathcal{X}^\prime}\leq1$. Then \begin{align*} \int_X G(x) \prod_{j=1}^d (T_jf_j)^{\alpha_j} {\rm d} \mu(x) & \leq\int_X \prod_{j=1}^dg_j(x)^{\alpha_j}\prod_{j=1}^d T_jf_j(x)^{\alpha_j} {\rm d}\mu(x) =\int_X \prod_{j=1}^d(g_j(x) T_jf_j(x))^{\alpha_j} {\rm d} \mu(x)\\ & \leq\prod_{j=1}^d\left(\int_X g_j(x) T_jf_j(x) {\rm d}\mu(x)\right)^{\alpha_j} = \prod_{j=1}^d \,\left((T_j^\ast g_j)(f_j)\right)^{\alpha_j}\\ & \leq \prod_{j=1}^d \left(\|T_j^\ast g_j\|_{\mathcal{Y}_j^\ast} \|f_j\|_{\mathcal{Y}_j}\right)^{\alpha_j} \leq \prod_{j=1}^d\left(A\|G\|_{\mathcal{X}^\prime}\|f_j\|_{\mathcal{Y}_j}\right)^{\alpha_j} \leq A\prod_{j=1}^d\|f_j\|_{\mathcal{Y}_j}^{\alpha_j} \end{align*} where the inequalities follow in order from the first condition of \eqref{mainprob}, H\"older's inequality, the second condition of \eqref{mainprob}, and the assumption that $\|G\|_{\mathcal{X}^\prime}\leq1$. The proposition now follows by taking the supremum over all such $G$, using the fact that $\mathcal{X}'$ is norming for $\mathcal{X}$. \end{proof} \begin{remark} If the spaces $\mathcal{Y}_j$ are complete, the assumption that $T_j: \mathcal{Y}_j \to \mathcal{X}$ is positive automatically implies that $T_j$ is bounded,\footnote{Indeed, if not, we can find nonnegative $f_n$ with $\|f_n\| \leq 2^{-n}$ but $\|T_j f_n \| \geq 2^n$. So for each $n$, $2^n \leq \|T_j f_n \| \leq \|T_j (\sum_{n=1}^\infty f_n)\| \leq C$ for some finite $C$ since $\sum_{n=1}^\infty f_n \in \mathcal{Y}_j$ (because $\mathcal{Y}_j$ is a Banach space). This is a contradiction.} and so the adjoint operator $T_j^\ast$ is well-defined. If not, we interpret the second condition of \eqref{mainprob} as \begin{equation} \label{mainprobfixed2} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|G\|_{\mathcal{X}^\prime}\|f_j\|_{\mathcal{Y}_j} \end{equation} for $f_j \in \mathcal{Y}_j$ and $j=1,\dots,d$, and we can still conclude the validity of \eqref{mainineq} for functions $f_j \in \mathcal{Y}_j$ as the proof clearly demonstrates. \end{remark} \begin{remark} If the $T_j$ are known to be bounded, it is immediate that \eqref{mainineq} holds with $A$ replaced by $\prod_{j=1}^d \|T_j\|^{\alpha_j}$.\footnote{This follows since $\| \prod_{j=1}^d h_j^{\alpha_j} \| \leq \prod_{j=1}^d \| h_j\|^{\alpha_j}$ which in turn follows from the case where each $\|h_j\| = 1$, which itself follows by Young's numerical inequality and the triangle inequality.} However, the best constant $A$ in \eqref{mainprob} will in general be much smaller, and this assertion is the main content of Proposition~\ref{thmguth}. \end{remark} \begin{remark} Observe that Proposition \ref{thmguth} does not require any topological structure of the space $X$, only its nature as a measure space. \end{remark} \begin{remark}\label{bnm} Notice that in order for the proof to go through, we only require that the factorisation property -- i.e. the first condition of \eqref{mainprob} -- holds for those $x$ which contribute to $\int_X G(x) \prod_{j=1}^d (T_jf_j)^{\alpha_j} {\rm d} \mu(x)$ for some functions $f_j \in \mathcal{Y}_j$. In other words, if a set $E \subseteq X$ with $\mu(E) >0$ has the property that for all choices $f_j$ of nonnegative functions in $\mathcal{Y}_j$, $\prod_{j=1}^d (T_jf_j)(x)^{\alpha_j} = 0$ a.e. on $E$, then $E$ will play no role in the analysis. There is therefore no loss of generality in assuming such sets do not exist. So we may assume without loss of generality that for all $E \subseteq X$ with $\mu(E) > 0$, there exist nonnegative $f_j \in \mathcal{Y}_j$ such that ``$\prod_{j=1}^d T_j f_j(x)^\alpha_j = 0$ a.e. on $E$'' fails -- i.e. such that there exists $E' \subseteq E$, with $\mu(E') > 0$ such that $\prod_{j=1}^d T_j f_j(x)^\alpha_j > 0$ on $E'$. That is, we may assume that for all $E \subseteq X$ with $\mu(E) > 0$, there exists $E' \subseteq E$ with $\mu(E') > 0$, and, for each $j$, a nonnegative $f_j \in \mathcal{Y}_j$ such that for all $x \in E'$, $T_jf_j(x) >0$. This condition is equivalent to the formally slightly weaker condition that for each $j$, for all $E \subseteq X$ with $\mu(E) > 0$, there exists $E' \subseteq E$ with $\mu(E') > 0$ and nonnegative $f_j \in \mathcal{Y}_j$ such that for all $x \in E'$, $T_jf_j(x) >0$.\footnote{If the latter condition holds, apply it to each $j$ in turn to obtain the former condition.} But this is simply the statement that each $T_j$ saturates $X$, as in Definition~\ref{saturation}. It is unsurprising that we will require saturation when it comes to formulating and proving the converse statement. \end{remark} \begin{remark}\label{weaksuff} Note that in place of $$\qquad \Big\|\,T_j^\ast g_j\Big\|_{\mathcal{Y}_j^\ast} \leq A\Big\|G\Big\|_{\mathcal{X}^\prime}\qquad\text{for all $j$.}$$ we could have assumed the (formally weaker) condition $$\qquad \prod_{j=1}^d \Big\|\,T_j^\ast g_j\Big\|_{\mathcal{Y}_j^\ast}^{\alpha_j} \leq A\Big\|G\Big\|_{\mathcal{X}^\prime}.$$ (A homogeneity argument shows that the two conditions are indeed equivalent.) \end{remark} \begin{remark} Similarly, it suffices to suppose a formally weaker hypothesis (``weak factorisation''), namely that for every $G \in \mathcal{X}^\prime$, there exist measurable functions $g_{jk}$ on $X$ such that $$G(x) \leq \sum_k \prod_{j=1}^d g_{jk}(x)^{\alpha_j}$$ a.e. on $X$, and $$ \sum_k \Big\|\,T_j^\ast g_{jk}\Big\|_{\mathcal{Y}_j^\ast} \leq A\Big\|G\Big\|_{\mathcal{X}^\prime}$$ for all $j$. But if this holds, and if we define $g_j = \sum_k g_{jk}$, Minkowski's inequality and H\"older's inequality yield \eqref{mainprob}. So this observation does not represent a genuine broadening of the scope of Proposition~\ref{thmguth}. \end{remark} \begin{remark} The argument for Proposition \ref{thmguth} was effectively given by Guth, in a less abstract form, in his proof the endpoint multilinear Kakeya inequality \cite{MR2746348}. However, any strategy which includes an application Proposition \ref{thmguth} to establish an inequality of the form \eqref{mainineq} involves the potentially difficult matter of first finding a suitable factorisation. Indeed, the main work of \cite{MR2746348} consisted precisely in finding such. In this context see also \cite{CV}. \end{remark} \subsection{Duality theory -- difficult half}\label{ddh} As suggested above, the implication in Proposition \ref{thmguth} can be essentially reversed, and a principal aim of this paper is to show that the factorisation property \eqref{mainprob} enunciated in Proposition \ref{thmguth} is in fact {\em necessary} as well as sufficient for \eqref{mainineq} to hold under very mild hypotheses. More precisely we prove: \begin{thm} \label{thmmain} Suppose that $(X, {\rm d} \mu)$ is a $\sigma$-finite measure space, $\mathcal{Y}_j$ are normed lattices, and $T_j : \mathcal{Y}_j \to \mathcal{M}(X)$ are positive linear maps. Suppose that $\mathcal{X}$ has the property that $\|\chi_E \|_{\mathcal{X}} < \infty$ implies $\mu(E) < \infty$.\footnote{See Remarks ~\ref{mmmm} and \ref{plpl}.}Suppose that {\em either} (i) each $T_j$ strongly saturates $X$, {\em or} (ii) each $T_j$ saturates $X$ and each $\mathcal{Y}_j$ is a normed lattice of measurable functions on a $\sigma$-finite measure space $Y_j$ which contains the simple functions. Suppose that \begin{equation*} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{\mathcal{X}}\leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j} \end{equation*} for all $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then there exists a weight function\footnote{i.e. a measurable function $w$ with $w(x) > 0$ a.e.} $w$ on $X$ such that for every $G\in\mathcal{X}^\prime$, there exist functions $g_j\in L^1(X, \, w\, {\rm d}\mu)$ such that \begin{equation}\label{factorise} G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\mbox{a.e. on $X$,} \end{equation} and such that for each $j$, \begin{equation}\label{control} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|G\|_{\mathcal{X}^\prime}\|f_j\|_{\mathcal{Y}_j} \end{equation} for all $f_j \in \mathcal{Y}_j$. \end{thm} \begin{remark} In the case $d=1$ the factorisation is trivial, and \eqref{control} is simply the usual duality relation corresponding to \eqref{lin2}. \end{remark} \begin{remark}\label{smaller} If there exist $g_j$ satisfying \eqref{factorise} and \eqref{control}, then by making one of the $g_j$ smaller if necessary, we can find $g_j$ satisfying \eqref{factorise} with {\em equality} in addition to \eqref{control}. \end{remark} \begin{remark} We emphasise that the constant $A$ appearing in \eqref{control} is {\em precisely} the constant $A$ occuring in the hypothesis. \end{remark} \begin{remark}\label{mnb} The hypothesis that each $T_j$ saturates $X$ is very natural as pointed out in Remark \ref{bnm} above. Indeed, for the reasons set out there, without this hypothesis we cannot expect the conclusion to hold. Needless to say, it will play an important role in the proof of Theorem~\ref{thmmain}. \begin{remark}\label{mmmm} There is no requirement in Theorem~\ref{thmmain} that $\mathcal{X}'$ be norming. For this reason, the general case of Theorem~\ref{thmmain} does not automatically follow from the special case $\mathcal{X} = L^1(X)$. On the other hand, if $\mathcal{X}'$ {\em is} norming, and if $X$ is $\sigma$-finite, we can replace the hypothesis that $\|\chi_E \|_{\mathcal{X}} < \infty$ implies $\mu(E) < \infty$ by the alternative hypothesis that $\|\chi_E \|_{\mathcal{X}'} < \infty$ implies $\mu(E) < \infty$, and conclude the factorisation statement for $G \in \mathcal{X}'$ in terms of measurable functions $g_j$.\footnote{i.e. we can no longer place them in a weighted $L^1$-space which is independent of $G$.} This is done by applying the special case $\mathcal{X} = L^1$ of Theorem~\ref{thmmain}, but with the measure ${\rm d} \mu$ replaced by the $\sigma$-finite measure $G \,{\rm d \mu}$, where $G \in \mathcal{X}'$. (See Remark~\ref{mmmmm} for a little more detail.) This line of argument extends further to the case where $X$ is not assumed to be $\sigma$-finite. Indeed, if $\mathcal{X}'$ is norming, $\mathcal{X}'$ has the property that $\|\chi_E \|_{\mathcal{X}'} < \infty$ implies $\mu(E) < \infty$, and, for all nonnegative $G \in \mathcal{X}'$ we have that $G$ is finite almost everywhere, we then have that $G \,{\rm d \mu}$ is a $\sigma$-finite measure irrespective of whether $\mu$ is a $\sigma$-finite measure or not; in such circumstances we can make the same factorisation statement for $G \in \mathcal{X}'$ in terms of measurable functions $g_j$. \end{remark} \end{remark} \begin{remark}\label{plpl} Except when the measure $\mu$ is finite, the condition that $\|\chi_E \|_{\mathcal{X}} < \infty$ implies $\mu(E) < \infty$ excludes the case $ \mathcal{X} = L^\infty$. This condition is important when it comes to ensuring that there is a sufficiently rich collection of subsets of $X$ of finite measure whose factorisations can then be glued together. See Lemmas~\ref{strsat} and \ref{sat} in Section~\ref{tsjk}. \end{remark} \medskip \noindent The proof of Theorem~\ref{thmmain} is highly nonconstructive and comes about as a result of duality methods in the theory of convex optimisation which ultimately rely upon a form of the minimax principle. Certain parts of the (linear) Maurey theory of factorisation of operators also employ such arguments, see \cite{Mau} and \cite{GCRdeF}, but we will need a more refined minimax principle than what was used in these sources. Other parts of the theory (see \cite{Pis}) instead use the Dunford--Pettis theorem. For the full details of the proof see Sections~\ref{outline}, \ref{overv} and \ref{dett} below. Nevertheless, in some cases, constructive factorisations can be given, and in other cases, the existence of the factorisation raises interesting questions and links with other areas of analysis. See Sections~\ref{factinterp}, \ref{culture}, \ref{Loomis--Whitneyrevisited}, \ref{NLLW} and \ref{BLrevisited}. \subsection{Multilinear Maurey-type theory}\label{maureysec} In this section we state and prove a slight generalisation of Theorem~\ref{thmmainmaurey}, using the case $\mathcal{X} = L^1(X)$ of Theorem~\ref{thmmain}. Interestingly, the {\em linear} case of Maurey's theorem corresponds to a {\em bilinear} case $d=2$ of Theorem~\ref{thmmain}, in which one of the normed lattices is one-dimensional. So, in contrast to the situation for Theorem~\ref{thmmain}, the case $d=1$ of what follows is not trivial. \begin{thm} \label{thmmainmaureyx} Suppose $(X, {\rm d} \mu)$ is a $\sigma$-finite measure space, $\mathcal{Y}_j$ are normed lattices, and $T_j : \mathcal{Y}_j \to \mathcal{M}(X)$ are positive linear maps. Suppose that {\em either} (i) each $T_j$ strongly saturates $X$, {\em or} (ii) each $T_j$ saturates $X$ and each $\mathcal{Y}_j$ is a normed lattice of measurable functions on a $\sigma$-finite measure space $Y_j$ which contains the simple functions. Let $0 < q <1$, and suppose \begin{equation}\label{tez} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{L^q(X)} \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j} \end{equation} for all $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then there exist nonnegative measurable functions $g_j$ on $X$ such that \begin{equation}\label{factorisebabymaureyx} \|\prod_{j=1}^d g_j(x)^{\alpha_j}\|_{L^{q'}(X)} = 1 \end{equation} and such that for each $j$, \begin{equation}\label{controlbabymaureyx} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|f_j\|_{\mathcal{Y}_j} \end{equation} for all $f_j \in \mathcal{Y}_j$. \end{thm} It is an easy exercise using H\"older's inequality to show that if there exist $g_j$ such that \eqref{factorisebabymaureyx} and \eqref{controlbabymaureyx} hold, then \eqref{tez} also holds. As in the case of Theorem~\ref{thmmainmaurey}, Theorem~\ref{thmmainmaureyx} admits an interpretation as a statement about factorisation of operators, see Section~\ref{mmf}. \begin{proof} The main hypothesis is that $$ \int_X \prod_{j=1}^d (T_jf_j)^{\alpha_j q} {\rm d} \mu \leq A^q \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j q}$$ for all $f_j \in \mathcal{Y}_j$. Let $\beta_j = \alpha_j q$ for $1 \leq j \leq d$ and let $\beta_{d+1} = 1 - \sum_{j=1}^d \beta_j = 1 - q > 0$. Let $Y_{d+1} = \{0\}$ and let $\mathcal{Y}_{d+1}$ be the trivial normed lattice $\mathbb{R}$ defined on the singleton measure space $\{0\}$. Let $T_{d+1} : \mathcal{Y}_{d+1} \to \mathcal{M}(X)$ be the linear map $\lambda \mapsto \lambda {\bf 1}$ where ${\bf 1}$ denotes the constant function taking the value $1$ on $X$. Then we have $$ \int_X \prod_{j=1}^{d+1} (T_jf_j)^{\beta_j} {\rm d} \mu \leq A^q \prod_{j=1}^{d+1} \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\beta_j}$$ for all $f_j \in \mathcal{Y}_j$. So by Theorem~\ref{thmmain} in the case $\mathcal{X} = L^1(X)$ (and with $d+1$ in place of $d$, see also Remark~\ref{smaller} above), we conclude that there exist measurable functions $G_1, \dots, G_{d+1}$ such that \begin{equation}\label{factorisez} \prod_{j=1}^{d+1} G_j(x)^{\beta_j} = 1 \qquad\mbox{a.e. on $X$,} \end{equation} and such that for each $1 \leq j \leq d+1$, \begin{equation}\label{controlz} \int_X G_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A^q \|f_j\|_{\mathcal{Y}_j} \end{equation} for all $f_j \in \mathcal{Y}_j$. \medskip \noindent For $1 \leq j \leq d$, set $g_j(x) = A^{1-q}G_j(x)$; then \eqref{controlz} immediately gives \eqref{controlbabymaureyx} for $1 \leq j \leq d$. By \eqref{factorisez} we have $$ \prod_{j=1}^d g_j(x)^{\alpha_j} = A^{1-q}G_{d+1}(x)^{(q-1)/q}$$ while \eqref{controlz} for $j = d+1$ gives $$ \int_X G_{d+1}(x) {\rm d} \mu(x) \leq A^q.$$ Combining these last two relations gives \eqref{factorisebabymaureyx} as desired. \end{proof} \noindent \begin{remark} It is not hard to verify that Theorem~\ref{thmmainmaureyx} also implies Theorem~\ref{thmmainbaby}. \end{remark} \noindent \begin{remark} As a simple application of Theorem~\ref{thmmainmaureyx}, suppose that $X = Y_1 = \dots = Y_d$ is a compact group with Haar measure $ {\rm} \mu$, and $T_j$ are as in the statement of the theorem and are also (left- or right-) translation-invariant. So, with $g_j$ as in the conclusion of the theorem, for all simple $f_j$ and for all $t \in X$ we have $$\int_X g_j(x)\tau_t (T_jf_j)(x) {\rm d}\mu(x) = \int_X g_j(x)T_j (\tau_t f_j)(x) {\rm d}\mu(x) \leq A\|\tau_t f_j\|_{\mathcal{Y}_j} = A \|f_j\|_{\mathcal{Y}_j}$$ where $\tau_t$ denotes translation by $t$. Therefore we have, for all $t \in X$, $$\int_X (\tau_{t^{-1}}g_j)(x)T_j f_j(x) {\rm d}\mu(x) \leq A\|f_j\|_{\mathcal{Y}_j}$$ and integrating with resepect to $t$ gives that for each $j$, $T_j: \mathcal{Y}_j \to L^1(X)$. Similar statements, which we leave the reader to formulate, can be made for amenable groups. \end{remark} \section{Proof of Theorem~\ref{thmmain} -- a special case}\label{outline} \subsection{Basic set-up} The idea behind the proof of Theorem \ref{thmmain} is to view problem \eqref{factorise} and \eqref{control} as a convex optimisation problem. That is, we replace the number $A$ in \eqref{control} by a variable $K$ and seek to minimise over $K$. To illustrate how this works, we first prove the theorem in a model case when $X$ and $Y_j$ are finite sets endowed with counting measure and $\mathcal{Y}_j=L^1(Y_j)$ for $j=1,\dots,d$. One reason for doing this case first is that there are no measure-theoretic or functional-analytical difficulties to be dealt with in this setting, and indeed $\mathcal{X}^\prime = \mathcal{X}^\ast$ is simply the class of all functions defined on $X$ with the norm dual to that of $\mathcal{X}$. It therefore allows us to emphasise the nature of the problem as one concerning convex optimisation. \medskip \noindent The minimisation problem we propose to examine now reads as follows. Fix $G: X \to [0, \infty)$ and consider \begin{equation} \label{finiteminim} \begin{aligned} \gamma =&\inf_{K,g_j} K\\ \text{such that}\quad&G(x)\leq\prod_{j=1}^d g_j(x)^{\alpha_j}\quad\text{for all $x\in X$ and}\\ &\max_{y_j\in Y_j}T_j^\ast g_j(y_j)\leq K\|G\|_{\mathcal{X}^\ast}\quad\text{for all $j=1,\dots,d$.} \end{aligned} \end{equation} We note that this is a convex optimisation problem since we are minimising a convex, in fact linear, function on the convex domain consisting of the $(d+1)$-tuples $(K, g_j)$ satisfying the constraints in \eqref{finiteminim}. The convexity of this domain follows from the fact that the second set of inequalities is linear in the arguments $K$ and $g_j$, and the operation of taking the geometric mean on the right hand side of the first set of inequalities is a concave function. We note that the set of $(K, g_j)$ satisfying the constraints in \eqref{finiteminim} is not empty and that we can in fact find $(K, g_j)$ satisfying these constraints with strict inequality by taking each $g_j$ to be $2G + 1$ and letting $K$ be sufficiently large. Thus problem \eqref{finiteminim} satisfies what is known as {\em Slater's condition}. (We do not give full details here as the discussion will eventually be subsumed into that of the next section.) In particular we certainly have $\gamma < + \infty$. \medskip We therefore follow a standard approach to convex optimisation problems, see for example \cite{MR2061575}. We introduce Lagrange multipliers $\psi$ and $h_j$, where $\psi: X \to \mathbb{R}_+$ (for the first set of constraints), and $h_j: Y_j \to \mathbb{R}_+$ (for the second set). Note that we are only interested case where these functions take nonnegative values since each of the constraints is an inequality constraint. We then introduce the Lagrange functional \begin{equation} \label{finitelagrange} L=K+\sum_{x\in X}\psi(x)\left(G(x)-\prod_{j=1}^d g_j(x)^{\alpha_j}\right) +\sum_{j=1}^d\sum_{y_j\in Y_j}h_j(y_j)(T_j^\ast g_j(y_j)-K\|G\|_{\mathcal{X}^\ast}). \end{equation} We emphasise that this function and the corresponding one defined in the proof of the general case are the only functions which we allow to take negative values. \medskip For nonnegative $K$, $g_j$, $\psi$, and $h_j$ we now consider the two problems\footnote{The subscript $\mathcal{L}$ in $\gamma_\mathcal{L}$ indicates that we are looking at the Lagrangian version of the problem as opposed to the original version which has $\gamma$ without a subscript.} \[ \gamma_\mathcal{L} =\inf_{K,g_j}\sup_{\psi,h_j}L \qquad\text{and}\qquad \eta=\sup_{\psi,h_j}\inf_{K,g_j}L \] called the primal problem and the dual problem respectively. We shall show that (i) the problem for $\gamma$ is identical to the problem for $\gamma_\mathcal{L}$, (ii) $\eta\leq A$ where $A$ is any number such that inequality \eqref{mainineq} holds, and (iii) $\eta=\gamma_\mathcal{L}$ (it is obvious that $\eta \leq \gamma_\mathcal{L}$). Finally, we show that the infimum in the definition of $\gamma$ is attained, and this will complete the proof of the theorem in the special case. \subsection{Identification of the problems for $\gamma$ and $\gamma_\mathcal{L}$.} We begin by studying $\gamma_\mathcal{L}$. Fix $K \geq 0$ and $g_j$ and consider $\sup_{\psi,h_j}L$. Suppose that any of the conditions in \eqref{finiteminim} is not satisfied at some point. Then take the relevant function $\psi$ or $h_j$ for some $j$ to have value $t>0$ at a point where an inequality fails and let all of the functions be zero everywhere else. Then let $t\to\infty$ and notice that $\sup_{\psi,h_j}L$ goes to $+\infty$ since $t$ is multiplied by a positive number. So if $\sup_{\psi,h_j}L<+\infty$ we must have that the conditions of \eqref{finiteminim} are satisfied. Conversely, if these conditions are satisfied then all factors multiplying $\psi(x)$ and $h_j(y_j)$ for any $x$ and $y_j$ are non-positive so the supremum is attained by taking them all to equal $0$. So, for each fixed $(K, g_j)$, we have that $\sup_{\psi,h_j}L < + \infty $ if and only if the conditions in \eqref{finiteminim} hold, in which case, $\sup_{\psi,h_j}L = K$. Thus we see that the problem for $\gamma_\mathcal{L}$ is identical to problem \eqref{finiteminim}, yielding $\gamma_\mathcal{L}=\gamma$. Moreover the infimum in the definition of $\gamma$ is attained if and only if the infimum in the definition of $\gamma_\mathcal{L}$ is attained \subsection{Proof that $\eta\leq A$.} We rearrange $L$ as follows: \begin{equation} \label{finitelagrangere} \begin{aligned} L=& \sum_{x\in X}\psi(x)G(x)+ K\left(1- \|G\|_{\mathcal{X}^\ast}\sum_{j=1}^d\sum_{y_j\in Y_j}h_j(y_j)\right)\\ &+ \sum_{x\in X}\left( \sum_{j=1}^dg_j(x)T_jh_j(x) -\prod_{j=1}^d g_j(x)^{\alpha_j}\psi(x)\right) \end{aligned} \end{equation} Let us fix $\psi$ and $h_j$ and consider $\inf_{K,g_j}L$. First of all, note that $\inf_{K,g_j}L=-\infty$ unless \begin{equation} \label{condKfinite} \|G\|_{\mathcal{X}^\ast}\sum_{j=1}^d\sum_{y_j\in Y_j}h_j(y_j)\leq 1 \end{equation} since if this inequality fails then the term multiplying $K$ in $L$ is negative and so by taking $g_j=0$ and letting $K$ go to infinity we get that $\inf_{K,g_j}L=-\infty$. Also note that $\inf_{K,g_j}L=-\infty$ unless \begin{equation} \label{condbetafinite} \psi(x)\leq\prod_{j=1}^d (\alpha_j^{-1}T_jh_j(x))^{\alpha_j} \end{equation} for all $x\in X$. Seeing this is a matter of choosing $g_j(x)$ to balance the arithmetic-geometric mean inequality. Specifically, suppose that this condition \eqref{condbetafinite} fails at a point $x_0$. Then we let $K=0$ and $g_j(x)=0$ for all $x\neq x_0$ and all $j=1,\dots,d$. There are now two cases to consider. Firstly, if there exists an index $j_0$ such that $T_{j_0}h_{j_0}(x_0)=0$ then we take $g_j(x_0)=1$ for all $j\neq j_0$ and $g_{j_0}(x_0)=t>1$. Then \[L= \sum_{x\in X}\psi(x)G(x) + \sum_{\substack{j=1\\j\neq j_0}}^d T_{j}h_{j}(x_0)-t^{\alpha_{j_0}}\psi(x_0).\] Since $\alpha_{j_0}>0$ we can let $t$ go to infinity and see that $\inf_{K,g_j}L=-\infty$. In the other case we have that $T_{j}h_{j}(x_0)>0$ for all $j=1,\dots,d$. Then we let \[g_{j}(x_0)=t \alpha_j ((T_jh_j)(x_0))^{-1}\prod_{j'=1}^d (\alpha_{j'}^{-1}T_{j'}h_{j'})(x_0)^{\alpha_{j'}}\] and note that \[L= \sum_{x\in X}\psi(x)G(x)+ t\left(\prod_{j=1}^d (\alpha_j^{-1}T_jh_j(x_0))^{\alpha_j} - \psi(x_0)\right).\] So by the assumption of the failure of \eqref{condbetafinite} at $x_0$ we see that letting $t \to \infty$ yields $\inf_{K,g_j}L=-\infty$. \medskip Conversely, if conditions \eqref{condKfinite} and \eqref{condbetafinite} hold then the factor multiplying $K$ is nonnegative and an application of the arithmetic-geometric mean inequality gives that for any choice of $g_j$ then for each $x\in X$ the term in the second bracket of \eqref{finitelagrangere} is nonnegative, so we attain $\inf_{K,g_j}L$ by letting $K=0$ and $g_j(x)=0$ for all $x\in X$ and $j=1,\dots,d$. Hence, for each fixed $(\psi, h_j)$, $\inf_{K,g_j}L > - \infty$ if and only if $\psi$ and $h_j$ satisfy conditions \eqref{condKfinite} and \eqref{condbetafinite}, in which case $\inf_{K,g_j}L = \sum_{x \in X} \psi(x) G(x)$. Noting that there always exist $\psi$ and $h_j$ satisfying conditions \eqref{condKfinite} and \eqref{condbetafinite}, we see that $\eta$ is the solution to \begin{equation} \label{finitemaxim} \begin{aligned} \eta=&\sup_{\psi,h_j} \sum_{x\in X}\psi(x) G(x)\\ \text{such that}\quad&\psi(x)\leq\prod_{j=1}^d (\alpha_j^{-1}T_jh_j(x))^{\alpha_j}\quad\text{for all $x\in X$ and}\\ &\|G\|_{\mathcal{X}^\ast}\sum_{j=1}^d\sum_{y_j\in Y_j}h_j(y_j)\leq 1. \end{aligned} \end{equation} For any $\psi$ and $h_j$ satisfying the conditions in \eqref{finitemaxim} we can calculate \begin{align*} \sum_{x\in X}\psi(x) G(x) &\leq \sum_{x\in X}\prod_{j=1}^d (\alpha_j^{-1}T_jh_j(x))^{\alpha_j} G(x) \leq \left\|\prod_{j=1}^d (\alpha_j^{-1}T_jh_j)^{\alpha_j}\right\|_{\mathcal{X}} \Big\|G\Big\|_{\mathcal{X}^\ast}\\ &\leq A\prod_{j=1}^d \|\alpha_j^{-1}h_j\|_{\mathcal{Y_j}}^{\alpha_j} \|G\|_{\mathcal{X}^\ast} \leq A\sum_{j=1}^d \|h_j\|_1 \|G\|_{\mathcal{X}^\ast}\leq A \end{align*} where the inequalities follow in order from the first condition of \eqref{finitemaxim}, the definition of the norm on $\mathcal{X}^\ast$, the inequality \eqref{mainineq}, the arithmetic-geometric mean inequality, and the second condition of \eqref{finitemaxim}. Taking the supremum now yields $\eta\leq A$. \subsection{Proof that $\gamma_\mathcal{L}=\eta$ and existence of minimisers.} This is a minimax argument. As we have noted above, it is immediate that $\eta\leq\gamma_\mathcal{L}$ and this is referred to as weak duality. The other direction, giving $\gamma_\mathcal{L}=\eta$, is called strong duality and does not hold in general. However there are various conditions which guarantee strong duality, such as Slater's condition which is the condition that the original problem \eqref{finiteminim} is convex and there exists a point satisfying all of the constraints with strict inequality. See \cite{MR2061575}, p.226. We have noted above that Slater's condition holds in our setting. Moreover, Slater's condition guarantees the existence of a maximiser for the dual problem. However, we need optimisers for the primal problem. If for all $x\in X$ we have $T_j {\bf 1}(x)>0$ -- which is simply the saturation hypothesis in our present case -- then the set of $g_j$'s which satisfy the constraints of \eqref{finiteminim} with $K=2A$ will be compact, and therefore a minimiser will exist. \section{Theorem~\ref{thmmain} -- Overview of the general case}\label{overv} Let us now turn to the argument for Theorem~\ref{thmmain} in the general case. It will entail substantial measure-theoretic and functional-analytic considerations not present in the case when $X$ and $Y_j$ are finite sets. While it is an attractive idea to try to establish Theorem~\ref{thmmain} by approximating the general case by the discrete case, this does not seem a feasible route, even when $\mathcal{X}$ and $\mathcal{Y}_j$ are $L^q$ and $L^{p_j}$ spaces respectively, and a direct approach is therefore required. The bulk of the proof of Theorem~\ref{thmmain} will be devoted to establishing the special case in which $X$ is a finite measure space.\footnote{To be clear, a measure space $(X, {\rm d} \mu)$ with $\mu (X) < \infty$, not a finite set $X$ with counting measure.} In this case we can take the weight function $w$ to be $1$ and we can work with arbitrary normed lattices $\mathcal{Y}_j$, but we need to impose super-strong saturation instead of saturation. The result reads as follows: \begin{thm} \label{thmsimple} Suppose $X$ is a finite measure space, $\mathcal{Y}_j$ are normed lattices, and that the linear operators $T_j: \mathcal{Y}_j \to \mathcal{M}(X)$ are positive. Suppose that each $T_j$ super-strongly saturates $X$. Suppose that \begin{equation}\label{again} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{\mathcal{X}}\leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j} \end{equation} holds for all $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$.\footnote{Notice that a hypothesis of super-strong saturation is unrealistic in the presence of inequality~\eqref{again} unless $X$ has finite measure.} Then for any $G\in\mathcal{X}^\prime$ there exist functions $g_j\in L^1(X, {\rm d}\mu)$ such that \begin{equation}\label{factorise1} G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\mbox{a.e. on $X$,} \end{equation} and such that for each $j$, \begin{equation}\label{control1a} \int_X g_j(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|G\|_{\mathcal{X}^\prime}\|f_j\|_{\mathcal{Y}_j} \end{equation} for all $f_j \in \mathcal{Y}_j$. \end{thm} In the proof of this theorem we introduce an extended real-valued Lagrangian function $L(\Phi,\Psi)$ (where $\Phi$ corresponds to the variables $(K, g_j)$ and $\Psi$ corresponds to the variables $(\psi, h_j)$ of the discrete model case discussed above). See Section~\ref{turn} below for precise details of the definition of $L$. As in the model case, we relate $\sup_\Psi\inf_\Phi L$ (which we had previously called $\eta$) to problem \eqref{factorise1} and \eqref{control1a} and $\inf_\Phi\sup_\Psi L$ (which we had previously called $\gamma$) to inequality \eqref{again}. We then need to show that $$ \min_\Phi\sup_\Psi L(\Phi, \Psi) = \sup_\Psi\inf_\Phi L(\Phi, \Psi),$$ and for this we need to use the Lopsided Minimax Theorem (which can be found as Theorem 7 from Chapter 6.2 of \cite{MR749753}): \begin{thm}\label{minimax} Suppose $C$ and $D$ are convex subsets of vector spaces and that $C$ is endowed with a topology for which the vector space operations are continuous. Further, suppose that $L : C \times D \to \mathbb{R}$ satisfies \begin{enumerate} \item [(i)]$\Phi\mapsto L(\Phi,\Psi)$ is convex for all $\Psi\in D$; \item [(ii)]$\Psi\mapsto L(\Phi,\Psi)$ is concave for all $\Phi\in C$; \item [(iii)]$\Phi\mapsto L(\Phi,\Psi)$ is lower semicontinuous for all $\Psi\in D$; and \item [(iv)]there exists a $\Psi_0\in D$ such that the sublevel sets $\{\Phi \in C \, : \, L(\Phi,\Psi_0)\leq\lambda\}$ are compact for all sufficiently large $\lambda\in\mathbb{R}$ \end{enumerate} Then \begin{equation}\label{minimaxconcl} \min_{\Phi\in C}\sup_{\Psi\in D} L(\Phi, \Psi) = \sup_{\Psi\in D}\inf_{\Phi\in C} L(\Phi, \Psi). \end{equation} \end{thm} \begin{remark} The existence of the minimum on the left-hand side of \eqref{minimaxconcl} is part of the conclusion: there exists a $\bar{\Phi}\in C$ such that $ \sup_{\Psi\in D} L(\bar{\Phi},\Psi) = \inf_{\Phi\in C}\sup_{\Psi\in D} L(\Phi, \Psi)$. Once we know that $ \inf \sup = \sup \inf$ this is easy because (iii) tells us that the map $\Phi \mapsto \sup_{\Psi\in D} L(\Phi, \Psi)$ is lower semicontinuous, and (iv) then tells us that the sublevel sets $$ \{ \Phi \in C \, : \, \sup_{\Psi\in D} L(\Phi, \Psi) \leq \lambda\} \subseteq \{ \Phi \in C \, : \, L(\Phi, \Psi_0) \leq \lambda\} $$ are closed and compact, and hence $\Phi \mapsto \sup_{\Psi\in D} L(\Phi, \Psi)$ achieves its minimum on any such set. The fact that $\sup \inf \leq \inf \sup$ is trivial, so the main content of the theorem is that $ \inf \sup \leq \sup \inf$. \end{remark} \begin{remark} There is nothing to stop both sides of \eqref{minimaxconcl} from being $+\infty$. Indeed, a nontrivial conclusion of the theorem is that if the right-hand side is finite, so is the left-hand side. \end{remark} \begin{remark} Traditional versions of minimax theorems assume that $C$ itself is compact, rather than compactness of certain sublevel sets as condition (iv). However, in our case, we cannot, for the reasons set out below, expect $C$ to be compact. {\em It is a remarkable feature of our analysis that the saturation hypothesis we must impose corresponds {\bf precisely} to condition (iv) of the minimax theorem.} \end{remark} \begin{remark}\label{urgh} The observant reader will have noticed that we have indicated our intention to introduce an extended real-valued Lagrangian $L$, but the minimax theorem applies only to real-valued Lagrangians. This mismatch necessitates a small detour which we wish to suppress here. For details see Section~\ref{detourdetails} below. \end{remark} \bigskip \noindent It is a somewhat delicate matter to choose the vector space where we will locate the variables $\Phi$ featuring in the Lagrangian which we will use. Corresponding to the variables $g_j$ occuring in the discrete model case of Section~\ref{outline}, we will now have variables $S_j$, which we would like to take to be elements of $L^1(X)_+$. (It is the weighted geometric mean of a particular collection of these which will ultimately furnish the desired factorisation.) However, it turns out to be helpful to instead allow, in the first instance, the $S_j$ be elements of the larger space $L^\infty(X)^\ast_+$, that is, the positive cone of the dual of $L^\infty(X)$.\footnote{Had we instead opted to work from the outset with $S_j \in L^1(X)_+$, we would have been forced to place unnatural topological conditions on $X$ in order to identify $L^1(X)_+$ with a subspace of a dual space, and in any case we would have to work in the larger space of finite regular Borel measures on $X$ in order to exploit weak-star compactness.} Thus we consider the vector space $\mathbb{R} \times L^\infty(X)^\ast \times \dots \times L^\infty(X)^\ast$ and take $C$ to be a suitable subset of the positive cone in this space. Ideally we would like to take $C$ to be a {\em norm-bounded} convex subset and then use the Banach--Alaoglu theorem to assert compactness of $C$; but since we are not expecting any quantitative $L^1$ bounds on the functions $S_j$ appearing in the factorisation, there is no natural norm-bounded set with which to work. Instead, we take $C$ to be the whole positive cone $\mathbb{R}_+ \times L^\infty(X)^\ast_+ \times \dots \times L^\infty(X)^\ast_+$, endowed with the weak-star topology. The price for this is the need to verify hypothesis (iv) of Theorem~\ref{minimax}. Fortunately this turns out to be not so difficult, and in fact is rather natural in our setting. Carrying out this process will yield some distinguished members of $L^\infty(X)^\ast_+$. However, working with the dual of $L^\infty(X)$ presents its own difficulties since some elements of $L^\infty(X)^\ast_+$ are quite exotic. Fortunately the theory of finitely additive measures comes to the rescue, and we will be able to show that elements satisfying the properties we require can be in fact be found in the smaller space $L^1(X)_+$. See Section~\ref{turn} below for more details. \medskip \noindent To set the scene for this, we recall three results of Yosida and Hewitt which can be found in \cite{MR0045194}. The setting for each of these results is a $\sigma$-finite measure space $(X, {\rm d \mu})$. \begin{thm} \label{thmdual} There is an isometric isomprphism between the space of finitely additive measures on $X$ of finite total variation which are $\mu$-absolutely continuous\footnote{This means that the finitely additive measure $\tau$ satisfies $\tau(E) = 0$ whenever $\mu(E) = 0$.} and the space of bounded linear functionals on $L^\infty(X, {\rm d}\mu)$. For a finitely additive measure $\tau$ with these properties the corresponding element of $L^\infty(X, {\rm d}\mu)^\ast$ is given by $\tau(\psi)=\int_X \psi \, {\rm d} \tau $ (where the integral is the so-called Radon integral). Furthermore $L^1(X,{\rm d}\mu)$ embeds isometrically into $L^\infty(X, {\rm d}\mu)^\ast$ in such a way that the application of $g\in L^1(X, {\rm d}\mu)$ to an element of $\psi\in L^\infty(X, {\rm d}\mu)$ is given by $\int_X \psi g {\, \rm d}\mu $ where the integral is now the Lebesgue integral. \end{thm} \begin{thm} \label{thmYH1} Any element $S\in L^\infty(X, {\rm d}\mu)^\ast$ can be written uniquely as $S=S_{\mathrm{ca}}+S_{\mathrm{pfa}}$ where $S_{\mathrm{ca}}$ is countably additive (and hence is given by integration against a function in $L^1(X, {\rm d} \mu)$) and $S_{\mathrm{pfa}}$ is purely finitely additive. Furthermore $S \geq 0$ if and only if $S_{\mathrm{ca}}\geq 0$ and $S_{\mathrm{pfa}}\geq 0$. \end{thm} \medskip We need not concern ourselves here with the definition of purely finitely additive measures, but, in order to be able to use these results, we do need a useful characterisation of which measures are purely finitely additive. \begin{thm} \label{thmYH2} A nonnegative finitely additive measure $\tau$ which is $\mu$-absolutely continuous is purely finitely additive if and only if for every nonnegative countably additive measure $\sigma$ which is $\mu$-absolutely continuous, every measurable set $E$ and every pair of positive numbers $\delta_1$ and $\delta_2$, there is a measurable subset $E'$ of $E$ such that $\sigma(E')<\delta_1$ and $\tau(E\setminus E')<\delta_2$. \end{thm} \medskip \noindent We wish to remark that analysis related to the dual space of $L^\infty$ has also been employed in a number of other contexts recently. See for example \cite{BGM}, \cite{SS}, \cite{Toland}, and, in the financial mathematics literature, \cite{MMS}. \section{Theorems~\ref{thmmain} and ~\ref{thmsimple} -- details of the proofs}\label{dett} \subsection{Preliminaries}\label{Preliminaries} We shall first need two lemmas which will be useful for the proof of Theorem~\ref{thmsimple} and also that of Theorem~\ref{thmmain} itself. The first one is they key technical tool which, in the context of Theorem~\ref{thmsimple}, will allow us do induce existence of suitable integrable functions from existence of corresponding members of the dual of $L^\infty$. We shall continue to assume that $\alpha_j > 0$ and that $\sum_{j=1}^d \alpha_j = 1$. \begin{lem}\label{abscont} Let $(X, {\rm d} \mu)$ be a $\sigma$-finite measure space and suppose that $S_j\in L^\infty(X)_+^\ast$. Suppose that $G$ is a measurable function such that \begin{equation}\label{opl} \int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x)\leq\sum_{j=1}^d \alpha_j S_j(\beta_j) \qquad\text{for all simple functions $\beta_j$ on $X$.} \end{equation} If $({S}_{j\mathrm{rn}})$ denotes the Radon--Nikodym derivative with respect to $\mu$ of the component of $S_j$ which is countably additive, then \begin{equation}\label{lpo} G(x)\leq \prod_{j=1}^d {S}_{j\mathrm{rn}}(x)^{\alpha_j}\quad\text{a.e. on $X$. } \end{equation} Conversely, if $G$ is such that \eqref{lpo} holds, then \eqref{opl} holds. \end{lem} \begin{remark} This result extends the special case $d=1$ which is implicit in Theorem \ref{thmYH1}. \end{remark} \begin{proof} The converse statement follows immediately from the arithmetic-geometric mean inequality, so we turn to the forward assertion. Suppose not. Then there exists a set $E_0$ with $\mu(E_0)>0$ such that inequality \eqref{lpo} fails on $E_0$ and we can find an $\varepsilon>0$ and $E_1\subseteq E_0$ with $\mu(E_1)>0$ such that \begin{displaymath} G(x)-\varepsilon> \prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \end{displaymath} for all $x\in E_1$. Now take $\beta_j$ to be simple functions supported on $E_1$. Then we get \begin{equation} \label{bigexpr} \begin{gathered} \int_{E_1} \prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x) + \varepsilon \int_{E_1} \prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d} \mu(x)\\ \leq \int_{E_1} G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x) \leq \sum_{j=1}^d \alpha_j S_j(\beta_j)\\ = \sum_{j=1}^d \alpha_j \int_{E_1} S_{j\mathrm{rn}}(x) \beta_j(x) \, {\rm d}\mu(x) + \sum_{j=1}^d \alpha_j \int_{E_1} \beta_j(x) \, {\rm d}\tau_{j\mathrm{pfa}} \end{gathered} \end{equation} where $\tau_{j\mathrm{pfa}}$ is the purely finitely additive measure associated to the purely finitely additive component $S_{j \mathrm{pfa}}$ of $S_j$. \medskip We can find a subset $E_2 \subseteq E_1$ with $\mu(E_2)>0$ and a $C>0$ such that $S_{j\mathrm{rn}}(x)\leq C$ for all $x\in E_2$ and all $j$. There are now two cases to consider. \medskip First, assume that there exists a subset $E_3 \subseteq E_2$ such that $\mu(E_3)>0$ and an index $j_0$ such that $S_{j_0\mathrm{rn}}(x)=0$ for all $x\in E_3$. Now let $\delta$ be small and positive (to be specified later) and $E_4$ a subset of $E_3$ with $0< \mu(E_4) < \infty$ (also to be specified later\footnote{We are using $\sigma$-finiteness of $\mu$ to ensure that we can find such an $E_4$ with $\mu(E_4)$ finite.}), and take $\beta_j=\delta\chi_{E_4}$ for $j\neq j_0$ and $\beta_{j_0}=\delta^{1-\alpha^{-1}_{j_0}}\chi_{E_4}$. This implies that $\prod_{j=1}^d \beta_j^{\alpha_j}(x)=1$ for all $x\in E_4$ and so the top line of~\eqref{bigexpr} equals $0+\varepsilon\mu(E_4)$. The first term on the bottom line of~\eqref{bigexpr} can be bounded by $C\delta\mu(E_4)$ since there is no contribution from the term with index $j_0$. The second term we can bound by $\delta^{1-\alpha^{-1}_{j_0}} \left(\sum_j\tau_{j\mathrm{pfa}}\right)(E_4)$. We have not chosen $E_4$ precisely yet. To do this we use Theorem \ref{thmYH2} above. \medskip Indeed, applying Theorem \ref{thmYH2} with $\tau := \sum_j\tau_{j\mathrm{pfa}}$, $\sigma := \mu$, $E := E_3$ and $\delta_1 := \mu(E_3)/2$ gives that for all $\delta_2 > 0$ there is an $E_4 \subseteq E_3$ such that $\mu(E_3 \setminus E_4) < \mu(E_3)/2$ and $\tau(E_4) < \delta_2$. So ~\eqref{bigexpr} implies $$ \varepsilon \mu(E_4) \leq C\delta\mu(E_4) + \delta^{1-\alpha^{-1}_{j_0}}\tau(E_4) \leq C\delta\mu(E_4) + \delta^{1-\alpha^{-1}_{j_0}} \delta_2.$$ Now choose $\delta_2 = \varepsilon \mu(E_3)/4 \delta^{1-\alpha^{-1}_{j_0}}$, so that for some $E_4 \subseteq E_3$ we have $$ \varepsilon \mu(E_4) \leq C\delta\mu(E_4) + \varepsilon \mu(E_3)/4 \leq C\delta\mu(E_4) + \varepsilon \mu(E_4)/2.$$ Finally, choosing $\delta < \varepsilon/(2C)$ yields a contradiction, since by construction $\mu(E_4) > 0$. \medskip In the other case, we have that $S_{j\mathrm{rn}}(x)>0$ for a.e. $x\in E_2$ and all $j$. Then we can find a subset $E_3 \subseteq E_2$ with $0<\mu(E_3)<\infty$ and a number $c>0$ such that $S_{j\mathrm{rn}}(x)\geq c$ for all $x\in E_3$ and all $j$. We define $u_{j}$ on this set as \begin{displaymath} u_{j}(x)=S_{j\mathrm{rn}}(x)^{-1}\prod_{k=1}^d S_{k\mathrm{rn}}^{\alpha_k}(x) \end{displaymath} and note that $u_{j}(x)\leq c^{-1}C$ and that $\prod_j u_{j}^{\alpha_j}(x)=1$. Since these functions are bounded then if we are given $\delta>0$ we can find simple functions $\tilde{\beta}_{j}$ such that $u_j(x)-\delta\leq \tilde{\beta}_j(x)\leq u_j(x)$ for all $x\in E_3$. We may assume that $\tilde{\beta}_j(x)\geq c C^{-1}$ for all $x\in E_3$. Let us take $\beta_j=\tilde{\beta}_j\chi_{E_4}$ where $E_4$ is a subset of $E_3$ to be chosen. Then for $x\in E_4$ we have that \begin{align*} 0\leq& \prod_{j=1}^du_j^{\alpha_j}(x)- \prod_{j=1}^d\beta_j^{\alpha_j}(x) = \sum_{k=1}^d \left(\prod_{j=1}^{k} u_j^{\alpha_j}(x) \left(u_{k}^{\alpha_{k}}(x)-\beta_{k}^{\alpha_{k}}(x)\right) \prod_{j=k+1}^d\beta_{j}^{\alpha_j}(x)\right)\\ =& \sum_{k=1}^d \left(\prod_{j=1}^{k} u_j^{\alpha_j}(x) \alpha_k \frac{\xi_{k}^{\alpha_k}(x)}{\xi_{k}(x)} \left(u_{k}(x)-\beta_{k}(x)\right) \prod_{j=k+1}^d\beta_{j}^{\alpha_j}(x)\right) \leq d C^2 c^{-2}\delta. \end{align*} Here $\xi_{j_0}(x)$ lies between $\beta_{j_0}(x)$ and $u_{j_0}(x)$ and we have used that $c/C \leq \beta_j(x),\xi_j(x),u_j(x)\leq C/c $. Now we can estimate the first term on the top line of \eqref{bigexpr} from below by \begin{displaymath} \int_{E_4}\prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \left(\prod_{j=1}^d u_j^{\alpha_j}(x)- d C^2c^{-2}\delta \right) {\rm d}\mu(x) \geq \int_{E_4}\prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \, {\rm d}\mu(x) - \mu(E_4) d C^3c^{-2}\delta, \end{displaymath} and the second term on the top line of \eqref{bigexpr} we can estimate from below by $\varepsilon\mu(E_4)(1-dC^2c^{-2}\delta)$ since $\prod_j u_{j}^{\alpha_j}(x)=1$ on $E_4$. \medskip The first term on the bottom line of \eqref{bigexpr} we can estimate from above using $\beta_j \leq u_j$ on $E_4$ and the definition of $u_j$ by \begin{displaymath} \int_{E_4}\prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \, {\rm d}\mu(x). \end{displaymath} The second term we can bound by $Cc^{-1}\left(\sum_j\tau_{j\mathrm{fa}}\right)(E_4)$. Collecting this we have that \begin{gather*} \int_{E_4}\prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \, {\rm d} \mu(x)-\mu(E_4)dC^3c^{-2}\delta +\varepsilon\mu(E_4)(1-d C^2c^{-2}\delta)\\\leq \int_{E_4}\prod_{j=1}^d S_{j\mathrm{rn}}^{\alpha_j}(x) \, {\rm d} \mu(x) +Cc^{-1}\left(\sum_j\tau_{j\mathrm{pfa}}\right)(E_4). \end{gather*} The integrals cancel\footnote{Since $S_j \in (L^\infty)^*$ we have $S_{j\mathrm{rn}} \in L^1$, and the terms we are cancelling are indeed finite.} and we get $$ \varepsilon\mu(E_4) \leq \mu(E_4)(\varepsilon d C^2c^{-2}\delta + dC^3c^{-2}\delta) + Cc^{-1}\tau(E_4)$$ where $\tau = \sum_j\tau_{j\mathrm{pfa}}$. We can then choose $E_4$ and $\delta$ in much the same way as before to yield a contradiction. Note that $\delta$ will only depend on $d$, $C$, $c$ and $\varepsilon$. \medskip So in both cases we have a contradiction to the existence of $E_0$, and so \eqref{lpo} must hold. \end{proof} \medskip \noindent It turns out that we shall need to consider the action of $S \in L^\infty(X, {\rm d}\mu)^\ast_+$ not just on $L^\infty(X)$, but on general nonnegative measurable functions in $\mathcal{M}(X)$. The reasons for this are explained in Section~\ref{turn} below. To this end, we extend $S$ to $\mathcal{M}(X)_+$ by declaring, for $F \in \mathcal{M}(X)_+$, $$ S(F) := \sup\{ S(f) \, : \, 0 \leq f \leq F, f \in L^\infty(X)\} = \sup\{ S(\phi) \, : \, 0 \leq \phi \leq F, \; \phi \mbox{ simple}\}.$$ Of course $S(F)$ will often now take the value $+ \infty$. \medskip \noindent The second lemma concerns continuity properties of this extension. Consider the map $S \mapsto S(F)$ for fixed $F \in \mathcal{M}(X)_+$ as $S$ ranges over $L^\infty(X)_+^*$. If $F \in L^\infty(X)$ this map is norm continuous and hence weak-star continuous. For $F \in \mathcal{M}(X)_+$ we can assert less. \begin{lem}\label{lsc} Fix $F \in \mathcal{M}(X)_+$. Then the map $S \mapsto S(F)$ from $L^\infty(X)_+^*$ to $\mathbb{R} \cup \{+ \infty\}$ is weak-star lower semicontinuous. \end{lem} \noindent We remark that we have to be cautious here since $L^\infty(X)_+^*$ with the weak-star topology is not a metric space; so we cannot simply concern ourselves with sequential lower semicontinuity. \begin{proof} Let $S \in L^\infty(X)_+^*$. Either $S(F) = + \infty$ or $S(F) < + \infty$. Let us first deal with the latter case. We need to show that for every $\epsilon > 0$ there is a weak-star open neighbourhood $U$ of $S$ such that for $R \in U$ we have $R(F) \geq S(F) - \epsilon$. \medskip \noindent Since $S(F) < + \infty$ there is an $f \in L^\infty(X)$ with $ 0 \leq f \leq F$ such that $S(f) > S(F) - \epsilon$. Let $$U = \{R \in L^\infty(X)_+^* \, : \, R(f) > S(F) - \epsilon\}.$$ Then $S \in U$, and $U$ is weak-star open since for each $f \in L^\infty(X)$ the functional $R \mapsto R(f)$ is weak-star continuous. So for $R \in U$ we have $$ R(F) \geq R(f) > S(F) - \epsilon$$ which is what we needed. \medskip \noindent Now we look at the case $S(F) = +\infty$. We now need to show that for every $N \in \mathbb{N}$ there is a weak-star open neighbourhood $U$ of $S$ such that for $R \in U$ we have $R(F) \geq N$. \medskip \noindent Since $S(F) = + \infty$ there is an $f \in L^\infty(X)$ with $ 0 \leq f \leq F$ such that $S(f) > N$. Let $$U = \{R \in L^\infty(X)_+^* \, : \, R(f) > N \}.$$ Then $S \in U$, and $U$ is weak-star open since for each $f \in L^\infty(X)$ the functional $R \mapsto R(f)$ is weak-star continuous. So for $R \in U$ we have $$ R(F) \geq R(f) > N$$ which is what we needed. \end{proof} \subsection{Proof of Theorem~\ref{thmsimple}}\label{turn} Suppose we are in the situation in the statement of Theorem~\ref{thmsimple}. In particular, we assume that $G \in \mathcal{X}^\prime$, and we may clearly assume that $\|G\|_{\mathcal{X}'} \neq 0$. \medskip We recall from Section~\ref{overv} that we take $C$ (in which we locate the variables $\Phi = (K, S_j)$) to be the positive cone $\mathbb{R}_+ \times (L^\infty(X)^\ast_+)^d$ in the vector space $\mathbb{R} \times (L^\infty(X)^\ast)^d$, and $C$ is given the topology inherited from the product topology of the corresponding weak-star topologies. We take $D$ (in which we locate the variables $\Psi=(\beta_j,h_j)$) to be the positive cone in the vector space is $\mathcal{S}(X)^d \times \mathcal{Y}_1 \times \dots \times \mathcal{Y}_d$. \medskip \noindent Therefore, for $K\in\mathbb{R}_+$, $S_j\in L^\infty(X)^\ast_+$, $\beta_j$ simple functions on $X$ and $h_j \in \mathcal{Y}_j$ we consider the functional \begin{align*} L = \, &K+\left(\int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x)-\sum_{j=1}^d \alpha_j S_j(\beta_j) \right)\\ &+\sum_{j=1}^d\left(S_j(T_jh_j)- K\|G\|_{\mathcal{X}^\prime}\|h_j\|_{\mathcal{Y}_j}\right). \end{align*} \medskip Note that the integral term is well-defined since $G \in \mathcal{X}^\prime$ (which is contained in $L^1(X, {\rm d} \mu)$ when $\mu$ is a finite measure, as we have previously observed) and the $\beta_j$ are simple functions, and that the terms $S_j(\beta_j)$ are also well-defined since the $\beta_j$ are bounded functions. The terms $S_j(T_j h_j)$ are well-defined via the extension of $S_j$ to $\mathcal{M}(X)_+$ as discussed in Section~\ref{Preliminaries} above. Thus $L: C \times D \to \mathbb{R} \cup \{+ \infty\}$ well-defined and takes values in $\mathbb{R} \cup \{+ \infty\}$, with the possible value $+\infty$ arising when $T_jh_j$ is not a {\em bounded} measurable function. \medskip \noindent We next want to see how we can apply the minimax theorem, Theorem~\ref{minimax} to this Lagrangian. Recall from Remark~\ref{urgh} above that we have a problem in so doing, since our Lagrangian may take the value $+ \infty$, while Theorem~\ref{minimax} requires that the Lagrangian be real-valued. To be clear, what we desire -- and what we shall indeed obtain -- is the conclusion $$\min_{\Phi\in C}\sup_{\Psi\in D} L(\Phi, \Psi) = \sup_{\Psi\in D}\inf_{\Phi\in C} L(\Phi, \Psi)$$ of Theorem~\ref{minimax} in our case, but in order to achieve this we need to make a detour. \subsubsection{A detour}\label{detourdetails} We now describe the necessary detour. This involves modifying the Lagrangian we have defined in order to make it real-valued, but without altering its essential purpose. The main technical difference is that instead of allowing $S_j$ to act on the possibly unbounded $T_jh_j$, we have it act on an arbitrary nonnegative simple function $\psi_j$ satisfying $\psi_j \leq T_j h_j$, \medskip \noindent We therefore introduce a new Lagrangian $\Lambda: C \times \tilde{D} \to \mathbb{R}$, where $C$ is as before, and where $$\tilde{D} = \{(\beta_j, h_j, \psi_j) \in \mathcal{S}(X)^d \times \mathcal{Y}_1 \times \dots \times \mathcal{Y}_d \times \mathcal{S}(X)^d \, : \, \beta_j \geq 0,\, h_j \geq 0,\, 0 \leq \psi_j \leq T_j h_j \}.$$ Note that $\tilde{D}$ is convex. \medskip \noindent For $(K, S_j) \in C$ and $(\beta_j, h_j, \psi_j) \in \tilde{D}$ we define \begin{align*} \Lambda = \, &K+\left(\int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x)-\sum_{j=1}^d \alpha_j S_j(\beta_j) \right)\\ &+\sum_{j=1}^d\left(S_j\psi_j- K\|G\|_{\mathcal{X}^\prime}\|h_j\|_{\mathcal{Y}_j}\right). \end{align*} Note that $\Lambda$ is real-valued since $S_j \psi_j$ is real-valued. Moreover, note that by the definition of the extension of $S_j$ to $\mathcal{M}(X)_+$, we have \begin{equation}\label{lah} L((K, S_j),(\beta_j, h_j)) = \sup_{\{\psi_j \, : \, \psi_j \leq T_jh_j\}} \Lambda((K, S_j),(\beta_j, h_j,\psi_j)). \end{equation} We will momentarily check that the Lagrangian $\Lambda$ satisfies the hypotheses of Theorem~\ref{minimax}, but taking this as read for now, we deduce using \eqref{lah} that $$ \min_{(K,S_j)\in C}\sup_{(\beta_j, h_j)\in {D}} L = \min_{(K,S_j)\in C}\sup_{(\beta_j, h_j, \psi_j)\in \tilde{D}} \Lambda = \sup_{(\beta_j, h_j, \psi_j)\in \tilde{D}}\inf_{(K,S_j)\in C} \Lambda.$$ \medskip \noindent But since trivially $\sup \inf \leq \inf \sup$, we have, using \eqref{lah} once more, $$ \sup_{(\beta_j, h_j, \psi_j)\in \tilde{D}}\inf_{(K,S_j)\in C} \Lambda \leq \sup_{(\beta_j, h_j)\in {D}}\inf_{(K,S_j)\in C} \sup_{\psi_j\leq T_jh_j} \Lambda = \sup_{(\beta_j, h_j)\in {D}}\inf_{(K,S_j)\in C}L.$$ Combining the last two displays we obtain $$ \min_{(K,S_j)\in C}\sup_{(\beta_j, h_j)\in {D}} L \leq \sup_{(\beta_j, h_j)\in {D}}\inf_{(K,S_j)\in C}L.$$ Since the reverse inequality is once again trivial we conclude that $$\min_{\Phi\in C}\sup_{\Psi\in D} L(\Phi, \Psi) = \sup_{\Psi\in D}\inf_{\Phi\in C} L(\Phi, \Psi)$$ as we needed. \medskip Now we need to look at conditions (i) -- (iv) of Theorem~\ref{minimax} in our case where $\Lambda$ replaces $L$. Concerning (i), the map $S \mapsto S(F)$ is linear on $L^\infty(X)^\ast$ for each fixed $F \in \mathcal{S}(X)$. Therefore, for each fixed $\tilde\Psi \in \tilde{D}$, the map $\Phi \mapsto \Lambda(\Phi, \tilde\Psi)$ is affine, thus convex on $C$. Concerning (ii), the map $F \mapsto S(F)$ is linear and hence concave on $\mathcal{S}(X)$ for each fixed $S \in L^\infty(X)^\ast_+$. Moreover the geometric mean is a concave operation and the map $h \mapsto \|h\|_{\mathcal{Y}_j}$ is convex. Therefore, for each fixed $\Phi \in C$, the map $\tilde\Psi \mapsto \Lambda(\Phi, \tilde\Psi)$ is concave on $\tilde{D}$. Concerning (iii), this follows directly from the norm-continuity of $S \mapsto S(F)$ on $L^\infty(X)^\ast$ for each fixed $F \in \mathcal{S}(X)$. \medskip \noindent Condition (iv) is more interesting, and it is in verification of this condition that we use the crucial saturation hypothesis of Theorem~\ref{thmsimple}. We need to see that for some $\tilde\Psi_0 \in \tilde{D}$ the sublevel sets $\{\Phi \in C \, : \, L(\Phi,\tilde\Psi_0)\leq\lambda\}$ are compact for all sufficiently large $\lambda$. We will show that for a suitable choice of $\tilde\Psi_0$ these sets are norm-bounded, and from this the Banach--Alaoglu theorem will give us compactness. \medskip \noindent We take $\tilde\Psi_0 = (\beta_j, h_j, \psi_j)$ to have $\beta_j=0$ for all $j$. We take $h_j \in \mathcal{Y}_j$ such that $T_j h_j \geq c_0 > 0$ a.e. on $X$, as guaranteed by the hypothesis of Theorem~\ref{thmsimple}. By multiplying by a suitable positive constant if necessary, we can certainly assume that $ \sum_{j=1}^d\|h_j\|_{\mathcal{Y}_j} < (2 \|G\|_{\mathcal{X}'})^{-1}$. Finally, we take $\psi_j = c_0 {\bf 1}$ which satisfies $0 \leq \psi_j \leq T_j h_j$. \medskip \noindent For such a choice of $\tilde\Psi_0$ we have \begin{displaymath} \Lambda((K,S_j), \tilde \Psi_0)= K\left(1-\|G\|_{\mathcal{X}^\prime}\sum_{j=1}^d\|h_j\|_{\mathcal{Y}_j}\right) +\sum_{j=1}^dS_j \psi_j \geq K/2 + c_0 \sum_{j=1}^dS_j {\bf 1}. \end{displaymath} Therefore, for given $\lambda > 0$, $$ \{(K, S_j) \in C \, : \, \Lambda((K,S_j), \tilde \Psi_0) \leq \lambda\} \subseteq [0, 2 \lambda] \times \{S \in L^\infty(X)^\ast_+ \, : S ({\bf 1}) \leq c_0^{-1}\lambda\}^d.$$ \medskip \noindent But it is easy to see that for $S \geq 0$, $S({\bf 1}) = \|S\|_{L^\infty(X)^\ast}$. Indeed, by definition we have $$\|S\|_{L^\infty(X)^\ast}=\sup\{|S(u)| \, : \, u\in L^\infty(X),\|u\|_\infty\leq 1\}.$$ So let us take such a function $u$ with $\|u\|_\infty \leq 1$. Since $S(-u)=-S(u)$ we may by choosing either $u$ or $-u$ assume that $S(u)\geq 0$. We have that $u\leq 1$ a.e. and therefore the non-negativity of $S$ gives us that $S(u)\leq S({\bf 1})$, as needed. \medskip \noindent Therefore $$ \{(K, S_j) \in C \, : \, \Lambda((K,S_j), \tilde \Psi_0) \leq \lambda\} \subseteq [0, 2 \lambda] \times \{S \, : \|S\|\leq c_0^{-1}\lambda\}^d$$ is a norm-bounded, weak-star closed, hence weak-star compact subset of $\mathbb{R} \times (L^\infty(X)^\ast)^d$, by the Banach--Alaoglu theorem. This completes the verification of condition (iv) of Theorem~\ref{minimax} in our case, and we conclude that $$\min_{\Phi\in C}\sup_{\Psi\in D} L(\Phi, \Psi) = \sup_{\Psi\in D}\inf_{\Phi\in C} L(\Phi, \Psi).$$ \subsubsection{Return to the main argument} We may therefore conclude, by Theorem~\ref{minimax}, that if for non-zero $G \in \mathcal{X}^\prime $ fixed we define\footnote{The notation here is perhaps confusing. We shall consider four problems: $\gamma$, $\gamma^\ast$, $\gamma_\mathcal{L}$ and $\gamma_\mathcal{L}^\ast$. When there is no superscript we are dealing with the variant of the problem pertaining to $L^1$, and presence of the superscript $\ast$ denotes that we are dealing with the variant of the problem which pertains to $(L^\infty)^\ast$; when there is no subscript we are dealing with the original version of the problem, and presence of the subscript $\mathcal{L}$ denotes that we are dealing with the Lagrangian formulation. This is consistent with the notation we adopted in the treatment of the finite discrete case above; in that case there was no distinction between $L^1$ and $(L^\infty)^\ast$. We do not adorn $\eta$ with either a superscript $*$ nor a subscript $\mathcal{L}$ since there is only one $\eta$-problem. Nevertheless we emphasise that the $\eta$-problem does indeed deal with the Langrangian formulation in the form pertaining to $(L^\infty)^\ast$.} $$\gamma_\mathcal{L}^\ast = \inf_{(K,S_j) \in C}\sup_{(\beta_j,h_j) \in D} L \qquad\text{and}\qquad \eta = \sup_{(\beta_j,h_j) \in D}\inf_{(K,S_j) \in C} L$$ then $\eta = {\gamma}_\mathcal{L}^\ast$ and the infimum in the problem for $\gamma_\mathcal{L}^\ast$ is achieved as a minimum. (It should be noted that we are not yet in a position to assert the finiteness of either of these numbers.) \bigskip In order to progress further, we shall also consider the problem \begin{equation} \label{convprob} \begin{aligned} \gamma =& \inf K\\ \text{such that}\quad & G(x)\leq \prod_{j=1}^d S_j(x)^{\alpha_j}\quad\text{a.e. on $X$, and}\\ & \int_X S_j(x) T_j h_j(x) \, {\rm d}\mu(x) \leq K \|G\|_{\mathcal{X}^\prime} \|h_j\|_{\mathcal{Y}_j} \quad\text{for all $j$ and all $h_j \in \mathcal{Y}_j$}\\ \end{aligned} \end{equation} where the $S_j$ are taken to be in $L^1(X, {\rm d \mu)}$. We emphasise that this is the problem we really want to solve: if we can prove that $\gamma \leq A$ and that minimisers exist, we will have our desired factorisation. Nevertheless, we should point out that it is not yet even clear that there exist $(K, S_j)$ satisfying the constraints of \eqref{convprob}. We shall be able to infer the existence of such $(K, S_j)$ and hence the finiteness of $\gamma$ only from the conclusion of Theorem~\ref{thmsimple}. \medskip Our strategy is to show that (i) $\gamma_\mathcal{L}^\ast = \gamma$ and that if the problem for $\gamma_\mathcal{L}^\ast$ admits minimisers $\Phi$, then the problem for $\gamma$ also admits minimisers; and (ii) $0 \leq \eta \leq A$. Combining these with the minimax result $\gamma_\mathcal{L}^\ast=\eta$ and existence of minimsiers for $\gamma_\mathcal{L}^\ast$, we can conclude that the problem for $\gamma$ admits minimisers and that $\gamma \leq A$, which will conclude the proof of Theorem~\ref{thmsimple}. \medskip \textbf{Proof that $\gamma = \gamma_\mathcal{L}^\ast$ and that existence of minimisers for $\gamma_\mathcal{L}^\ast$ implies existence of minimisers for $\gamma$.} We begin by studying $\gamma_\mathcal{L}^\ast$ so let us consider for which $(K,S_j) \in C$ we have $\sup_{\beta_j,h_j} L < \infty$. Fix $(K, S_j)$. First of all, suppose that $S_j$ are such that there exists a tuple $(\beta_j)$ such that \begin{displaymath} \int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x)\, {\rm d}\mu(x)-\sum_{j=1}^d\alpha_jS_j(\beta_j)>0. \end{displaymath} Then by setting $h_j=0$ and substituting $\beta_j \mapsto t \beta_j$ and letting $t \to \infty$ we see that the supremum is infinite. Therefore, if $\sup_{\beta_j,h_j} L < \infty$, we must have \begin{equation} \label{condbeta} \int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x)\leq\sum_{j=1}^d\alpha_jS_j(\beta_j) \qquad\text{for all simple functions $\beta_j$,} \end{equation} which, by Lemma~\ref{abscont}, is equivalent to \begin{equation}\label{eqnabscont} G(x)\leq \prod_{j=1}^d {S}_{j\mathrm{rn}}(x)^{\alpha_j}\quad\text{a.e. on $X$. } \end{equation} \medskip Now assume there exists a $j_0$ and an $h_{j_0} \in \mathcal{Y}_{j_0}$, such that $S_{j_0}(T_{j_0}h_{j_0})>K\|G\|_{\mathcal{X}^\prime}\|h_{j_0}\|_{\mathcal{Y}_{j_0}}$. Taking $\beta_j = 0$, and $h_j= 0$ for $j\neq j_0$ and multiplying $h_{j_0}$ by a factor $t$ which we send to infinity we again see that the supremum is infinite. Therefore, if $\sup_{\beta_j,h_j} L < \infty$, then we must also have \begin{equation} \label{condScircT} S_{j}(T_{j}h_{j})\leq K\|G\|_{\mathcal{X}^\prime}\|h_j\|_{\mathcal{Y}_j} \end{equation} for all nonnegative $h_j$ and all $j$. From the positivity of $S_{j\mathrm{pfa}}$ we see that this implies \begin{equation}\label{condTastS} \int_X S_{j\mathrm{rn}}(x)T_jh_j(x)\, {\rm d}\mu(x)\leq K\|G\|_{\mathcal{X}^\prime}\|h_j\|_{\mathcal{Y}_j} \end{equation} for all nonnegative $h_j$ and all $j$. \medskip On the other hand, if for fixed $(K, S_j)$ conditions~\eqref{eqnabscont} and~\eqref{condScircT} are satisfied, then when we are looking for $\sup_{\beta_j,h_j} L$, we can do no better than taking $\beta_j =0$ and $h_j=0$ for all $j$. So for fixed $(K, S_j)$, we have $\sup_{\beta_j,h_j} L < \infty$ if and only if conditions ~\eqref{eqnabscont} and~\eqref{condScircT} hold, in which case $\sup_{\beta_j,h_j} L = K$. So the problem for $\gamma_\mathcal{L}^\ast$ is identical with the problem \begin{align*} \gamma^\ast =& \inf K\\ \text{such that}\quad & G(x)\leq \prod_j S_{j\mathrm{rn}}(x)^{\alpha_j}\qquad\text{a.e.},\\ & S_{j}(T_{j}h_{j})\leq K\|G\|_{\mathcal{X}^\prime}\|h_j\|_{\mathcal{Y}_j}\quad \text{for all $j$ and all $h_j \in \mathcal{Y}_j$,} \end{align*} where we emphasise that the $\inf$ is taken over $(K, S_j)$ with $S_j \in L^\infty(X)_+^\ast$. \medskip Likewise, the problem $$\gamma_\mathcal{L} := \inf_{(K,S_j) \in \mathbb{R}_+ \times (L^1(X)_+)^d}\sup_{\beta_j, h_j}L$$ is identical with problem~\eqref{convprob} for $\gamma$. \medskip It is clear that $\gamma_\mathcal{L}^\ast \leq \gamma_\mathcal{L}$ as the infimum for the left hand side is over a larger set than for the right hand side. \medskip {\bf Claim:} $\gamma_\mathcal{L} \leq \gamma_\mathcal{L}^\ast$, and if minimisers $\Phi = (K, S_j)$ exist for problem $\gamma_\mathcal{L}^\ast$, they also exist for problem $\gamma_\mathcal{L}$. \medskip Indeed, assume that $\gamma_\mathcal{L}^\ast < \infty$, let $\varepsilon > 0$ and let $(K, S_j)$ with $S_j \in L^\infty(X)^\ast$ and satisfying conditions~\eqref{eqnabscont} and~\eqref{condScircT} be such that $K < \gamma_\mathcal{L}^\ast + \varepsilon$. Then the absolutely continuous component $S_{j\mathrm{rn}}$ satisfies~\eqref{eqnabscont} and~\eqref{condTastS}, and so $(K,S_{j\mathrm{rn}})$ contributes to the infimum in the problem for $\gamma_\mathcal{L}$. Thus $\gamma_\mathcal{L} \leq \gamma_\mathcal{L}^\ast + \varepsilon$. Letting $\varepsilon \to 0$ establishes the first part of the claim. Now suppose that minimisers $\Phi = (K, S_j)$ exist for problem $\gamma_\mathcal{L}^\ast$. In particular this supposes that $\gamma_\mathcal{L}^\ast < \infty$. Let $(K, S_j)$ with $S_j \in L^\infty(X)^\ast$ and satisfying conditions~\eqref{eqnabscont} and~\eqref{condScircT} be such that $K = \gamma_\mathcal{L}^\ast$. Then the absolutely continuous component $S_{j\mathrm{rn}}$ satisfies~\eqref{eqnabscont} and~\eqref{condTastS}, and so $(K,S_{j\mathrm{rn}})$ contributes to and indeed achieves the infimum in the problem for $\gamma_\mathcal{L}$ (otherwise $\gamma_\mathcal{L}$ would be strictly less than $\gamma_\mathcal{L}^\ast$). \medskip Summarising, the problems for $\gamma$ and $\gamma_\mathcal{L}$ are equivalent; the problems for $\gamma^\ast$ and $\gamma_\mathcal{L}^\ast$ are equivalent; $\gamma_\mathcal{L} = \gamma_\mathcal{L}^\ast$, and if extremisers exist for $\gamma_\mathcal{L}^\ast$, they also exist for $\gamma_\mathcal{L}$, and hence too for $\gamma$. \medskip \textbf {Proof that $0 \leq \eta \leq A$.} We wish to carry out a similar analysis for $\inf_{K,S_j} L$, and for that we first of all rewrite $L$ as \begin{align*} L=&\int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) \, {\rm d}\mu(x) +K\left(1-\|G\|_{\mathcal{X}^\prime}\sum_{j=1}^d\|h_j\|_{\mathcal{Y}_j}\right)\\ &+\sum_{j=1}^dS_j\left(T_jh_j-\alpha_j\beta_j \right). \end{align*} We consider for which $(\beta_j,h_j) \in D$ we have $\inf_{K,S_j} L >-\infty$. \medskip First, by taking $S_j=0$ for all $j=1,\dots,d$ and letting $K$ to go infinity we see that if $\inf_{K,S_j} L > -\infty$ then we must have \begin{equation} \label{condsumhj} \|G\|_{\mathcal{X}^\prime}\sum_{j=1}^d \|h_j\|_{\mathcal{Y}_j}\leq1. \end{equation} Secondly, assume that there exists an index $j_0$ and a set $E\subseteq X$ with $\mu(E)>0$ such that $T_{j_0}h_{j_0}(x)< \alpha_{j_0} \beta_{j_0}(x)$ for a.e. $x\in E$. Then by taking $S_{j_0}= t \chi_E\in L^1$, $S_j=0$ for $j\neq j_0$ and $K=0$ and letting $t \to \infty$, then we see that $\inf_{K,S_j}L =-\infty$. Thus if $\inf_{K,S_j} L > -\infty$, we must also have \begin{equation} \alpha_j\beta_j(x)\leq T_jh_j(x)\quad\text{a.e. on $X$ for all $j$.} \label{condThbeta} \end{equation} If conditions \eqref{condsumhj} and \eqref{condThbeta} are both satisfied we can do no better than take $K=0$ and $S_j=0$ for all $j$. So, for fixed $(\beta_j,h_j)$, $\inf_{K,S_j} L >-\infty$ if and only if conditions \eqref{condsumhj} and \eqref{condThbeta} hold, in which case $\inf_{K,S_j} L = \int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) {\rm d}\mu(x)$. We can always find $(\beta_j,h_j)$ such that conditions \eqref{condsumhj} and \eqref{condThbeta} hold, so $\eta = \sup_{\beta_j, h_j} \int_X G(x)\prod_{j=1}^d \beta_j^{\alpha_j}(x) {\rm d}\mu(x)$ subject to conditions \eqref{condsumhj} and \eqref{condThbeta}. In particular this tells us that $\eta \geq 0$. \medskip Let us now derive an upper bound for $\eta$. Examining the condition \eqref{condThbeta} on $\beta_j$ we see that \begin{align*} \eta \leq& \sup_{h_j} \int_{X} G(x) \prod_{j=1}^d \left(\alpha_j^{-1}T_{j}h_j(x)\right)^{\alpha_j} \, {\rm d}\mu(x)\\ \text{such that}\quad & \|G\|_{\mathcal{X}^\prime }\sum_{j=1}^d \|h_j\|_{\mathcal{Y}_j}\leq1. \end{align*} Clearly there exist functions $h_j \in \mathcal{Y}_j$ such that $\|G\|_{\mathcal{X}^\prime}\sum_j \|h_j\|_{\mathcal{Y}_j}\leq1$, and for any such we have \begin{align*} \int_X G(x) \prod_{j=1}^d \left(\alpha_j^{-1} T_jh_j(x)\right)^{\alpha_j} \, {\rm d}\mu(x) &\leq \|G\|_{\mathcal{X}^\prime} \| \prod_{j=1}^d \left(T_j(\alpha_j^{-1} h_j)(x)\right)^{\alpha_j } \|_\mathcal{X} \\ &\leq \|G\|_{\mathcal{X}^\prime} A\prod_{j=1}^d\|\alpha_j^{-1}h_j\|_{\mathcal{Y}_j}^{\alpha_j} \\ &\leq \|G\|_{\mathcal{X}^\prime} A\sum_{j=1}^d\alpha_j\| \alpha_j^{-1}h_j\|_{\mathcal{Y}_j} \\ &= \|G\|_{\mathcal{X}^\prime} A\sum_{j=1}^d\|h_j\|_{\mathcal{Y}_j}\leq A \end{align*} by H\"older's inequality in the form $\int Gf \leq \|G\|_{\mathcal{X}'} \|f\|_{\mathcal{X}}$, the multilinear inequality \eqref{mainineq} which is our main hypothesis, the arithmetic-geometric mean inequality and finally the assumption on the $h_j$. This clearly implies $\eta \leq A$, and thus concludes the proof of Theorem~\ref{thmsimple}. \qed \subsection{Proof of Theorem~\ref{thmmain}}\label{tsjk} We prove Theorem~\ref{thmmain} by reducing it to Theorem~\ref{thmsimple}. In order to do this, it is convenient to begin with two lemmas whose purpose is to allow us to construct suitable exhausting sequences of sets in $X$, each member of which has finite measure. The two lemmas correspond to the two alternate hypotheses of Theorem~\ref{thmmain}. \begin{lem}\label{strsat} Suppose that $(X, {\rm d \mu)}$ is a $\sigma$-finite measure space, $\mathcal{Y}_j$ are normed lattices and $T_j:\mathcal{Y}_j \to \mathcal{M}(X)$ are positive linear operators. Suppose that $\mathcal{X}$ is such that $\| \chi_E \|_\mathcal{X} < \infty$ implies $\mu(E) < \infty$. Suppose $h_j \in \mathcal{Y}_j$ is such that $\|h_j\|_{\mathcal{Y}_j} \leq 1$ and $T_j h_j > 0$ a.e. on $X$. Suppose furthermore that for some $A < \infty$, $$ \| \prod_{j=1}^d (T_j h_j)^{\alpha_j} \|_\mathcal{X} \leq A \prod_{j=1}^d \|h_j\|_{\mathcal{Y}_j}^{\alpha_j}.$$ Let $E_n = \{x \in X \, : \, T_jh_j(x) > 1/n \mbox{ for all } j\}.$ Then $(E_n)$ is an increasing and exhausting\footnote{Here and below, we take ``exhausting'' to mean up to a set of measure zero.} sequence of sets, each of which has finite measure. \end{lem} \begin{proof} Clearly the sequence $(E_n)$ is increasing and exhausting, and each $E_n$ is of finite measure since $$ \frac{1}{n} \| \chi_{E_n}\|_{\mathcal{X}} \leq \| \prod_{j=1}^d (T_j h_j)^{\alpha_j} \|_\mathcal{X} \leq A \prod_{j=1}^d \|h_j\|_{\mathcal{Y}_j}^{\alpha_j} < \infty.$$ \end{proof} \begin{lem}\label{sat} Suppose that $X$ and $Y_j$, for $j=1,\dots,d$, are $\sigma$-finite measure spaces. Suppose that $T_j: \mathcal{S}(Y_j) \to \mathcal{M}(X)$ are positive linear operators. Suppose $\mathcal{X}$ is such that $\| \chi_E \|_\mathcal{X} < \infty$ implies $\mu(E) < \infty$. Let $\mathcal{Y}_j$ be normed lattices of measurable functions defined on $Y_j$ which contain the simple functions. Suppose that each $T_j$ saturates $X$. Suppose that for some $A < \infty$, \begin{equation*} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_{\mathcal{X}}\leq A \prod_{j=1}^d \Big\|f_j\Big\|_{\mathcal{Y}_j}^{\alpha_j} \end{equation*} for all simple functions $f_j \in \mathcal{Y}_j$, $1 \leq j \leq d$. Then there exist $h_{j,n} \in \mathcal{Y}_j$ such that $\|h_{j,n}\|_{\mathcal{Y}_j} \leq 1$ for all $j$ and $n$, such that $h_{j,m} \geq h_{j,n}$ for $m \geq n$ and such that if we let $$ E_n = \{ x \in X \, : \, T_j h_{j,n}(x) > 1/n \mbox{ for all }j\},$$ then $(E_n)$ is an increasing and exhausting sequence of subsets of $X$, each of which has finite measure. \end{lem} \begin{proof} By $\sigma$-finiteness of $Y_j$ there exists an increasing and exhausting sequence of subsets $(Y_{j,n})_{n=1}^\infty$ of $Y_j$, each of which has finite measure. We take $Y_{j,0}=\emptyset$ for each $j$ and define an increasing sequence $(h_{j,n})_{n=1}^\infty$ of simple functions on $Y_j$ by \[h_{j,n} = \sum_{m=1}^n \gamma_{j,m}{\chi_{(Y_{j,m}\setminus Y_{j,m-1})}}\] where the positive sequence $(\gamma_{j,m})_{m=0}^\infty$ decreases sufficiently rapidly in order to ensure that $\|h_{j,n}\|_{\mathcal{Y}_j} \leq 1$ for all $j$ and $n$. \medskip Let $G_n:=\{x\in X \, : \, T_jh_{j,n}(x)>0 \text{ for all $j$}\}$ and note that $(G_n)_{n=0}^\infty$ is an increasing sequence of subsets of $X$. We show that this sequence is exhausting. Suppose not, then there exists an $A\subseteq X$ with $\mu(A)>0$ and an index $j$ such that $T_jh_{j,n}=0$ a.e. on $A$ for all $n$. By the assumption that $T_j$ saturates $X$, we can find a simple function $\tilde{h}_j$ on $Y_j$ and $A^\prime\subseteq A$ with $\mu(A^\prime)>0$ such that $T_j\tilde{h}_j(x)>0$ for a.e. $x \in A^\prime$. Let $\tilde{h}_{j,m}=\chi_{(Y_{j,m}\setminus Y_{j,m-1})}\tilde{h}_j$. We claim that there exists an $m$ and a subset $A^{\prime\prime} \subseteq A^\prime$ with $\mu(A^{\prime\prime})>0$ such that $T_j\tilde{h}_{j,m}>0$ a.e. on $A^{\prime\prime}$. This follows from the following weak continuity lemma for positive operators which uses transfinite induction: \begin{lem} Suppose that $Y$ is a $\sigma$-finite measure space and that $Y_n$ is an increasing sequence of sets of finite measure exhausting $Y$. Let $H_n = Y_n \setminus Y_{n-1}$. Suppose $T$ is a positive linear operator mapping simple functions on $Y$ to measurable functions on a measure space $(X, {\rm d} \mu)$. Suppose there is a simple function $h$ on $Y$ such that $Th(x)>0$ for $x \in A \subseteq X$ where $\mu(A)>0$. Let $h_n := h \chi_{H_n}$. Then there exists an $n \in \mathbb{N}$ and a subset $A^\prime \subseteq A$ with $\mu(A^\prime)>0$ such that $Th_n(x)>0$ for $x \in A^\prime$. \end{lem} \begin{proof} If not, then for all $n$, $Th_n=0$ a.e. on $A$. For $M \subseteq \mathbb{N}$ let $$ \Delta_M := \bigcup_{m \in M} H_m \mbox{ and } h_M := h \chi_{\Delta_M}.$$ Note that $h_{\{m\}} = h_m$, $h_{\mathbb{N}} = h$ since $\{Y_n\}$ is exhausting, and if $M \cap N = \emptyset$ then $h_{M \cup N} = h_M + h_N$. Let $$\mathcal{M} = \{ M \subseteq \mathbb{N} \, : \, Th_M = 0 \mbox{ a.e. on } A\}.$$ Then $\mathcal{M} \neq \emptyset$ as each singleton $\{m\}$ is in $\mathcal{M}$. Moreover, since the power set of $\mathbb{N}$ is partially ordered by inclusion and has maximal element $\mathbb{N}$, Zorn's lemma implies that $\mathcal{M}$ has a maximal element $\overline{M}$. Now $\overline{M}$ is a {\em proper} subset of $\mathbb{N}$ as if $\overline{M} = \mathbb{N}$ we would have $Th_{\overline{M}} = Th_\mathbb{N} = Th = 0$ a.e. on $A$, which is false by hypothesis. So there is an $m_0 \in \mathbb{N} \setminus M$. But then $Th_{\overline{M} \cup \{m_0\}} = T(h_{\overline{M}} + h_{\{m_0\}}) = Th_{\overline{M}} + Th_{\{m_0\}}= Th_{\overline{M}} + Th_{m_0}$, and both terms are zero a.e. on $A$, contradicting the maximality of $\overline{M}$. \end{proof} By shrinking $A^{\prime\prime}$ if necessary (while maintaining $\mu(A^{\prime\prime})>0$), we may assume that $T_j\tilde{h}_{j,m}>c$ a.e. on $A^{\prime\prime}$. Since $\tilde{h}_{j,m}$ is bounded above on $Y_{j,m} \setminus Y_{j,m-1}$ and since $h_{j,m}$ is bounded away from $0$ on this set, we conclude by the positivity of $T_j$ that $T_jh_{j,m}$ is bounded below on $A^{\prime \prime}$, which, since $T_jh_{j,n}=0$ a.e. on $A$ for all $n$, contradicts the statement that $A^{\prime \prime}$ is a subset of $A$ of {\em positive} measure. This contradiction establishes the claim that the sets $G_n$ exhaust $X$. \medskip Now let $G_{n,m}=\{x\in G_n \, : \, T_jh_{j,n}(x)>1/m \text{ for all $j$}\}$ and note that $(G_{n,m})_{m=1}^\infty$ is an increasing and exhausting sequence for $G_n$. Let ${E}_{m}=\cup_{n = 1}^m G_{n,m}$. Then for $m \geq n$, $G_{n,m}\subseteq {E}_m\cap G_{n}$, and so the sequence $({E}_m\cap G_{n})_{m=n}^\infty$ exhausts $G_{n}$, hence $({E}_m)_{m=1}^\infty$ exhausts $X$. Note that if $x\in {E}_m$ then $x\in G_{n,m}$ for some $n\leq m$, and so for all $j$, $T_jh_{j,m}(x)\geq T_jh_{j,n}(x)> 1/m$. \medskip Finally, we note that $E_n$ has finite measure since $$ \frac{1}{n} \| \chi_{E_n}\|_{\mathcal{X}} \leq \| \prod_{j=1}^d (T_j h_{j,n})^{\alpha_j} \|_\mathcal{X} \leq A \prod_{j=1}^d \|h_{j,n}\|_{\mathcal{Y}_j}^{\alpha_j} < \infty.$$ \end{proof} \medskip \noindent We shall use Lemmas~\ref{strsat} and \ref{sat} in the following form: \begin{cor}\label{corp} Assume the hypotheses of Theorem~\ref{thmmain} and that $A < \infty$. Then for all $n$ there exist $h_{j,n} \in \mathcal{Y}_j$ such that $\|h_{j,n}\|_{\mathcal{Y}_j} \leq 1$, $h_{j,n} \leq h_{j,m}$ for $m \geq n$, and an increasing and exhausting sequence of subsets $E_n \subseteq X$, each of finite measure, such that for each $j$ and $n$, $T_j h_{j,n}(x) > 1/n$ for $x \in E_n$. \end{cor} \begin{proof} Under the hypotheses of Lemma~\ref{sat} this is immediate. Under the hypotheses of Lemma~\ref{strsat} we simply take $h_{j,n} = h_j$ for all $n$. \end{proof} \medskip \noindent We will also need the following lemma whose proof is an easy exercise in elementary point-set topology, and which is therefore omitted. \begin{lem}\label{top} Let $Z$ be a compact topological space and suppose $(z_n)$ is an infinite sequence of distinct points in $Z$. Then there exists a point $z \in Z$ such that every open neighbourhood of $z$ contains infinitely many $z_n$'s. \end{lem} \medskip {\em Proof of Theorem~\ref{thmmain}}. We may assume that $A < \infty$ otherwise there is nothing to prove. Take a nonzero $G \in \mathcal{X}^\prime$, and take $E_n$ as in Corollary~\ref{corp}. For each $m$ we can apply Theorem~\ref{thmsimple}, with $X$ replaced by $E_m$, to conclude that there exist $g_{j,m} \in L^1(E_m, {\rm d} \mu)$ such that \begin{equation}\label{kjh} G(x) \leq \prod_{j=1}^d g_{j,m}(x)^{\alpha_j}\qquad\mbox{a.e. on $E_m$,} \end{equation} and such that for each $j$, \begin{equation}\label{control1b} \int_{E_m} g_{j,m}(x)T_jf_j(x) {\rm d}\mu(x) \leq A\|G\|_{\mathcal{X}^\prime}\|f_j\|_{\mathcal{Y}_j} \end{equation} for all functions $f_j \in \mathcal{Y}_j$. \medskip \noindent If the sets $E_m$ stabilise in the sense that for some $M \in \mathbb{N}$, $E_M = X$ (up to sets of zero measure), we simply take $g_j = g_{j,M}$ and we are finished. So we may assume that the sets $E_m$ do not stabilise and therefore that there are infinitely many distinct $g_{j,m}$ for each $j$. \medskip Define $w_j(x)$ for $x \in E_m \setminus E_{m-1}$ to be $T_j h_{j,m}(x)$, where we take $E_0 = \emptyset$. Let $w(x) = \min_j w_j(x)$. Note that $w$ is a.e. positive. Let us now calculate \begin{align*} \|g_{j,m}\|_{L^1(w \,{\rm d}\mu)} &\leq \int_{{E}_m} g_{j,m}(x)\,w_j (x)\, {\rm d}\mu(x) =\sum_{n=0}^m\int_{{E}_n\setminus {E}_{n-1}} g_{j,m}(x)\,T_j h_{j,n}(x) \, {\rm d}\mu(x)\\ &\leq \int_{{E}_m} g_{j,m}(x)\, T_j h_{j,m}(x) \, {\rm d}\mu(x) \leq A \|G\|_{\mathcal{X}^\prime}\|h_{j,m}\|_{\mathcal{Y}_j}\leq A\|G\|_{\mathcal{X}^\prime}. \end{align*} Thus the functions $g_{j,m}$ all lie in a ball in $L^\infty(X, \, w\, {\rm d}\mu)^\ast$ which, by the Banach--Alaoglu theorem, is weak-star compact. It is therefore tempting to extract a weak-star convergent subsequence. However, we must resist this temptation since $L^\infty(X, \, w\, {\rm d}\mu)$ is not separable, and thus $L^\infty(X, \, w\, {\rm d}\mu)^\ast$ is not metrisable. We must therefore proceed with some caution. We will use Lemma~\ref{top} as a substitute for the existence of weak-star convergent subsequences. \medskip \noindent It is convenient to consider the vectors $${\bf g}_n = (g_{1,n}, \dots , g_{d,n}) \in L^1(X, w \,{\rm d}\mu) \times \dots \times L^1(X, w \,{\rm d}\mu)$$ $$ \subseteq L^\infty(X, \, w\, {\rm d}\mu)^\ast \times \dots \times L^\infty(X, \, w\, {\rm d}\mu)^\ast = (L^\infty(X, \, w\, {\rm d}\mu) \times \dots \times L^\infty(X, \, w\, {\rm d}\mu))^\ast.$$ By Lemma~\ref{top} there is a point ${\bf S} = (S_1, \dots , S_d) \in (L^\infty(X, \, w\, {\rm d}\mu) \times \dots \times L^\infty(X, \, w\, {\rm d}\mu))^\ast$ such that every weak-star open neighbourhood of ${\bf S}$ contains infinitely many of the ${\bf g}_n$. \begin{lem}\label{vector} Suppose $({\bf g}_n)$ and ${\bf S}$ are as above. \medskip \noindent (a) If for some ${\bf q} \in \mathcal{M}(X, w \, {\rm d} \mu)_+^d$ we have $$ {\bf g}_n ({\bf q}) = \sum_{j=1}^d \int_X g_{j,n} q_j \, w \, {\rm d} \mu = \int_X {\bf g}_n \cdot {\bf q} \, w \, {\rm d} \mu \leq K$$ for all sufficiently large $n$, then $$ {\bf S}({\bf q}) \leq K.$$ \medskip \noindent (b) If for some ${\bf q} \in L^\infty(X, w \, {\rm d} \mu)^d$ we have $$ {\bf g}_n ({\bf q}) = \sum_{j=1}^d \int_X g_{j,n} q_j w \, {\rm d} \mu = \int_X {\bf g}_n \cdot {\bf q} \, w \, {\rm d} \mu \geq L$$ for all sufficiently large $n$, then $$ {\bf S}({\bf q}) \geq L.$$ \end{lem} \begin{proof} (a) Suppose for a contradiction that ${\bf S}({\bf q}) \geq K'> K$ for some finite $K'$. Let $$U = \{{\bf R} \in ((L^\infty(w \, {\rm d} \mu))^d)^* \, : \; {\bf R} ({\bf q}) > (K + K')/2 \}.$$ Then $ {\bf S} \in U$, and $U$ is weak-star open since for each ${\bf q} \in \mathcal{M}(w \,{\rm d} \mu)_+^d$ the functional $ {\bf R} \mapsto {\bf R} ({\bf q})$ is weak-star lower semicontinuous, by (a vector-valued version of) Lemma~\ref{lsc}. Thus $U$ is an open neighbourhood of ${\bf S}$ in the weak-star topology. By the above remarks, $U$ must contain infinitely many of the $({\bf g}_n)$. But for all $n$ sufficiently large, $$ {\bf g}_n ({\bf q}) = \int_X {\bf g}_n \cdot {\bf q} \, w \, {\rm d}\mu \leq K < (K + K')/2$$ and so none of these ${\bf g}_n$ can be in $U$. This is a contradiction, and therefore ${\bf S}({\bf q}) \leq K$. \medskip \noindent (b) Suppose for a contradiction that ${\bf S}({\bf q})= L' < L$. Let $$U = \{{\bf R} \in ((L^\infty(w \, {\rm d} \mu))^d)^* \, : \; {\bf R} ({\bf q}) < (L + L')/2 \}.$$ Then $ {\bf S} \in U$, and $U$ is weak-star open since for each ${\bf q} \in L^\infty(w \,{\rm d} \mu)_+^d$ the functional $ {\bf R} \mapsto {\bf R} ({\bf q})$ is weak-star continuous. Thus $U$ is an open neighbourhood of ${\bf S}$ in the weak-star topology. By the above remarks, $U$ must contain infinitely many of the $({\bf g}_n)$. But for all $n$ sufficiently large, $$ {\bf g}_n ({\bf q}) = \int_X {\bf g}_n \cdot {\bf q} \, w \, {\rm d}\mu \geq L > (L + L')/2$$ and so none of these ${\bf g}_n$ can be in $U$. This is a contradiction, and therefore ${\bf S}({\bf q}) \geq L$. \end{proof} \medskip We now wish to verify that the absolutely continuous components $({S}_{j\mathrm{rn}})$ (where the Radon--Nikodym derivative is with respect to the measure $w \, {\rm d} \mu$) of $(S_j)$ satisfy \begin{equation}\label{rew} \begin{aligned} & G(x)\leq \prod_{j=1}^d {S}_{j\mathrm{rn}}(x)^{\alpha_j}\quad\text{a.e. on $X$, and}\\ & \int_X {S}_{j\mathrm{rn}}(x) T_j f_j(x) \, {\rm d}\mu(x) \leq A \|G\|_{\mathcal{X}^\prime} \|f_j\|_{\mathcal{Y}_j}\\ \end{aligned} \end{equation} for all $j$ and for all $f_j \in \mathcal{Y}_j$. Since we know that $S_j \in (L^\infty(X, w \, {\rm d} \mu))^\ast$, we will therefore have ${S}_{j\mathrm{rn}} \in L^1( w \, {\rm d} \mu)$, and this will conclude the proof of Theorem~\ref{thmmain}. \medskip \noindent We may suppose that $\|G\|_{\mathcal{X}'} = 1$. \medskip \noindent We look at the second inequality from \eqref{rew} first. Fix $m$ and consider $$ \int_{E_m} S_{jrn} (x) T_j f_j(x) {\rm d} \mu (x) = \int_X S_{jrn} (x) w(x)^{-1} \chi_{E_m}(x) T_j f_j(x) w(x) {\rm d} \mu (x)$$ $$ \leq S_j( w^{-1} \chi_{E_m} T_j f_j)$$ by positivity of each component in the Yosida--Hewitt decomposition of $S_j$, (recall Theorem~\ref{thmYH1}). Now, for $n \geq m$, $$ \int g_{jn}(x) [w(x)^{-1} T_j f_j(x) \chi_{E_m}(x)] w(x) {\rm d} \mu (x) \leq \int g_{jn}(x) T_j f_j(x) {\rm d} \mu (x) \leq A \|f_j \|_{\mathcal{Y}_j},$$ so that by Lemma~\ref{vector}(a) (in the scalar case), $$ S_j(w^{-1} T_j f_j \chi_{E_m}) \leq A \|f_j \|_{\mathcal{Y}_j}.$$ Thus $$ \int_{E_m} S_{jrn} (x) T_j f_j(x) {\rm d} \mu (x) \leq A \|f_j \|_{\mathcal{Y}_j}$$ and we now let $m \to \infty$ to get the second inequality of \eqref{rew}. \medskip \noindent Now we look at the first inequality from \eqref{rew}. By Lemma~\ref{abscont} (using the measure $w \, {\rm d} \mu)$ and the fact that the $E_m$ exhaust $X$, it suffices to show that for each fixed $m$, and all simple $\beta_j$, $$ \int_{E_m} G(x) \prod_{j=1}^d \beta_j(x)^{\alpha_j} w(x) {\rm d} \mu(x) \leq \sum_{j=1}^d \alpha_j S_j(\beta_j).$$ Take $n \geq m$. By \eqref{kjh} and the arithmetic-geometric mean inequality, the left-hand side is at most $$ \int_{E_m} \prod_{j=1}^d [g_{jn}(x)\beta_j(x)]^{\alpha_j} w(x) {\rm d} \mu(x) \leq \sum_{j=1}^d \alpha_j \int_{E_m} g_{jn}(x)\beta_j(x) w(x) {\rm d} \mu(x)$$ $$ \leq \sum_{j=1}^d \alpha_j \int_X g_{jn}(x)\beta_j(x) w(x) {\rm d} \mu(x) = \sum_{j=1}^d \int_X g_{jn}(x) [\alpha_j \beta_j(x)] w(x) {\rm d} \mu(x). $$ Thus for all $n \geq m$, $$ \sum_{j=1}^d \int_X g_{jn}(x) [\alpha_j \beta_j(x)] w(x) {\rm d} \mu(x) \geq \int_{E_m} G(x) \prod_{j=1}^d \beta_j(x)^{\alpha_j} w(x) {\rm d} \mu(x).$$ Since the simple functions $\beta_j$ are bounded, Lemma~\ref{vector}(b) gives us that $$ \sum_{j=1}^d S_j (\alpha_j \beta_j) \geq \int_{E_m} G(x) \prod_{j=1}^d \beta_j(x)^{\alpha_j} w(x) {\rm d} \mu(x),$$ which is what we want. \medskip \noindent This completes the proof of Theorem~\ref{thmmain}. \qed \part{Connections with other topics} \bigskip \section{Complex interpolation and factorisation}\label{factinterp} We begin by observing that the trivial identity of Example \ref{Loomis--Whitney}, $$ \int_{\mathbb{R}^2} f_1(x_2) f_2(x_1) \, {\rm d}x_1 {\rm d}x_2 = \int_\mathbb{R} f_1 \int_\mathbb{R} f_2,$$ immediately implies via Theorem \ref{thmmainbaby} that, for every nonnegative $G \in L^2(\mathbb{R}^2)$, there exist nonnegative $g_1$ and $g_2$ such that $$G(x) \leq \sqrt{g_1(x) g_2(x)} \;\; \mbox{ for almost every }\; x \in \mathbb{R}^2$$ and $$ \sup_{x_2} \int g_1(x_1, x_2) {\rm d} x_1 \leq \|G\|_2 \; \mbox{ and } \; \sup_{x_1} \int g_2(x_1, x_2) {\rm d} x_2 \leq \|G\|_2.$$ While it is not perhaps entirely obvious how to do this explicity (a point to which we return in Sections~\ref{Loomis--Whitneyrevisited}, \ref{NLLW} and \ref{BLfactdetails} below), for now we want to point out that this example highlights the connection between our multilinear duality theory and the theory of interpolation of Banach spaces. In particular, we consider the {\em upper} method of complex interpolation of A.~P.~Calder\'on, \cite{Calderon}. \medskip \noindent Suppose that $Z_0$ and $Z_1$ are Banach lattices of measurable functions defined on some measure space. We define $$Z_0^{1-\theta}Z_1^{\theta} = \{f \; : \, \mbox{ there exist } f_j \in Z_j \mbox{ such that } |f| \leq |f_0|^{1-\theta}|f_1|^{\theta}\}$$ with $$\|f\|_{Z_0^{1-\theta}Z_1^{\theta}} = \inf\{ \|f_0\|_{Z_0}^{1-\theta} \|f_1\|_{Z_1}^{\theta}\},$$ the inf being taken over all possible decompositions of $f$. Under the assumption that the unit ball of $ Z_0^{1-\theta}Z_1^{\theta}$ is closed in $Z_0 + Z_1$, Calder\'on showed that $$ Z_0^{1-\theta}Z_1^{\theta} = [Z_0, Z_1]^\theta$$ where $[Z_0, Z_1]^\theta$ is the interpolation space between $Z_0$ and $Z_1$ obtained by the upper complex method. \medskip \noindent With this in mind, the factorisation statement in our example is tantamount to the statement $$ L^2(\mathbb{R}^2) \hookrightarrow [L^\infty_{x_1} (L^1_{x_2}),L^\infty_{x_2} (L^1_{x_1})]^{1/2}.$$ Many other special cases of our theory can be similarly expressed in the language of interpolation. We leave it to the interested reader to pursue this point of view more systematically. \medskip \noindent In this particular example, the there is further structure, see for example Pisier \cite{Pis2}, (in whch some of the ideas are attributed to Lust-Piquard). There it is established that we have $$ L^2(\mathbb{R}^2) = HS(L^2(\mathbb{R})) \hookrightarrow \mathcal{L}_{{\rm reg}}(L^2) = [L^\infty_{x_1} (L^1_{x_2}),L^\infty_{x_2} (L^1_{x_1})]^{1/2}$$ where $HS$ denotes the class of Hilbert--Schmidt operators and $\mathcal{L}_{{\rm reg}}(L^2)$ is the space of {\em regular} bounded linear operators on $L^2$. In rough terms, a regular bounded linear operator on $L^2$ is one such that if its kernel is $K(s,t)$, then $|K(s,t)|$ is also the kernel of a bounded linear operator. \medskip \noindent The implicit factorisation arguments involved in establishing results of this type rely on the Hahn--Banach theorem or the Perron--Froebenius theorem, and are thus related to minimax theory; they are similarly non-constructive. \section{Factorisation and convexity}\label{qzf} It is also natural to enquire about how factorisation and interpolation interact at the level of particular families of inequalities. For the sake of concreteness, suppose we are in the setting of multilinear generalised Radon transforms on euclidean spaces -- so that $T_jF_j = F_j \circ B_j$ for suitable $B_j$. We shall suppress consideration of any of the technical hypotheses of Theorem~\ref{thmmain} in what follows. Suppose that we have the pair of inequalities \begin{equation}\label{int1} \left\|\prod_{j=1}^d T_jF_j\right\|_{L^{q_k}} \leq A_k \prod_{j=1}^d \left\|F_j\right\|_{L^{p_{jk}}} \end{equation} for $k = 0,1$, where $q_k, p_{jk} \geq 1$. \medskip \noindent Each of these has a family of corresponding equivalent factorisation statements, according to Theorem~\ref{thmmain} and the remarks in Section~\ref{exples}. See also Section~\ref{rrr} below. After some changes of notation, one such equivalent pair of statements is as follows. For $k = 0,1$, let $s_k : = q_k \sum_j p_{jk}^{-1}$. Then for all nonnegative $G_k$ ($k = 0,1$) such that $\int G_k^{s_k'} = 1$, there are nonnegative $M_{10}, \dots , M_{d0}$ and $M_{11}, \dots , M_{d1}$ such that \begin{equation}\label{6571} G_k(x) \leq \prod_{j=1}^d M_{jk}(x)^{q_k/p_{jk}s_k}\mbox{ a.e.} \end{equation} and such that for all $f_j$ with $\int f_j \leq 1$, \begin{equation}\label{5671} \int f_j(B_jx) M_{jk}(x) {\rm d} x \leq A_k^{q_k/s_k} \end{equation} for $k = 0,1$. \medskip \noindent From \eqref{6571} and \eqref{5671} we shall deduce a factorisation statement which implies the natural interpolation statement \begin{equation}\label{int2} \left\|\prod_{j=1}^d T_jF_j\right\|_{L^{q_\theta}} \leq A_0^{1-\theta}A_1^\theta \prod_{j=1}^d \left\|F_j\right\|_{L^{p_{j\theta}}} \end{equation} for $0 < \theta < 1$, where, as usual, $1/q_\theta = (1-\theta)/q_0 + \theta/q_1$, and similarly for $1/p_{j\theta}$. \medskip \noindent Indeed, given a nonnegative $G$ with $\int G = 1$, let $G_k = G^{1/s_k'}$. Taking convex combinations in \eqref{6571} gives us \begin{equation}\label{765} G(x) \leq \prod_{j=1}^d M_{j0}(x)^{q_0 s_0'(1-\theta)/p_{j0}s_0} M_{j1}(x)^{q_1 s_1'\theta/p_{j1}s_1} \mbox{ a.e.} \end{equation} \medskip \noindent Next, we define $$ \gamma_j(\theta) := \frac{q_0s_0'}{p_{j0} s_0}(1-\theta) + \frac{q_1s_1'}{p_{j1} s_1} \theta$$ and define $M_{j\theta}$ by $$ M_{j\theta}^{\gamma_j(\theta)} := M_{j0}(x)^{q_0 s_0'(1-\theta)/p_{j0}s_0} M_{j1}(x)^{q_1 s_1'\theta/p_{j1}s_1}.$$ Then, by \eqref{5671} we have $$ \int f_j(B_jx) M_{j\theta}(x) {\rm d} x = \int f_j(B_jx) M_{j0}(x)^{q_0 s_0'(1-\theta)/p_{j0}s_0 \gamma_j(\theta)} M_{j1}(x)^{q_1 s_1'\theta/p_{j1}s_1\gamma_j(\theta)} {\rm d} x$$ $$ \leq \left(\int f_j(B_jx) M_{j0}(x) {\rm d} x\right)^{q_0 s_0'(1-\theta)/p_{j0}s_0 \gamma_j(\theta)} \left(\int f_j(B_jx) M_{j1}(x) {\rm d} x\right)^{q_1 s_1'\theta/p_{j1}s_1 \gamma_j(\theta)}$$ by H\"older's inequality, since $\gamma_j(\theta)$ is defined precisely to ensure the two exponents on the right hand side here add to $1$. \medskip \noindent Therefore, if $\int f_j \leq 1$, $$ \int f_j(B_jx) M_{j\theta}(x) {\rm d} x \leq \left[A_0^{q_0/s_0}\right]^{q_0 s_0'(1-\theta)/p_{j0}s_0 \gamma_j(\theta)} \left[A_1^{q_1/s_1}\right]^{q_1 s_1'\theta/p_{j1}s_1 \gamma_j(\theta)}.$$ \medskip \noindent Now let $\beta_j(\theta) := \lambda(\theta) \gamma_j(\theta)$ where $\lambda(\theta)$ is defined so that $\sum_{j=1}^d \beta_j(\theta) = 1$. By the definition of $s_0$ and $s_1$ we have $$\sum_{j=1}^d \gamma_j(\theta) = \sum_{j=1}^d\left(\frac{q_0s_0'}{p_{j0} s_0}(1-\theta) + \frac{q_1s_1'}{p_{j1} s_1} \theta\right) = (1-\theta) s_0' + \theta s_1'.$$ So, we take $$ \lambda(\theta) := \frac{1}{(1-\theta) s_0' + \theta s_1'}.$$ \medskip \noindent Now, bearing in mind Remark~\ref{weaksuff}, we conclude that $$\prod_{j=1}^d\left(\int f_j(B_jx) M_{j\theta}(x) {\rm d} x\right)^{\beta_j(\theta)} \leq \left[A_0^{q_0/s_0}\right]^{\sum_jq_0 s_0'(1-\theta)\beta_j(\theta)/p_{j0}s_0 \gamma_j(\theta)} \left[A_1^{q_1/s_1}\right]^{\sum_jq_1 s_1'\theta \beta_j(\theta)/p_{j1}s_1 \gamma_j(\theta)}$$ $$ = \left[A_0^{q_0/s_0}\right]^{\lambda(\theta)\sum_jq_0 s_0'(1-\theta)/p_{j0}s_0} \left[A_1^{q_1/s_1}\right]^{\lambda(\theta)\sum_jq_1 s_1'\theta/p_{j1}s_1} = \left[A_0^{q_0/s_0}\right]^{\lambda(\theta)s_0'(1-\theta)} \left[A_1^{q_1/s_1}\right]^{\lambda(\theta)s_1'\theta}$$ $$ = A_0^{\frac{s_0'q_0 \lambda(\theta)(1-\theta)}{s_0}} A_1^{\frac{s_1'q_1 \lambda(\theta)\theta}{s_1}} = \left\{\left[A_0^{\frac{s_0'q_0 \lambda(\theta)(1-\theta)}{s_0}} A_1^{\frac{s_1'q_1 \lambda(\theta)\theta}{s_1}}\right]^{S(\theta)/Q(\theta)}\right\}^{Q(\theta)/S(\theta)}$$ for a certain quantity $S(\theta)/Q(\theta)$ to which we turn our attention next. Indeed we define this quantity ({\em not} $S(\theta)$, $Q(\theta)$ separately), so that the exponents on $A_0$ and $A_1$ inside the curly brackets sum to $1$. That is, $$ \frac{Q(\theta)}{S(\theta)} := \lambda(\theta) \left(\frac{s_0'q_0 (1-\theta)}{s_0} + \frac{s_1'q_1 \theta}{s_1}\right). $$ Let us define these exponents of $A_0$ and $A_1$ as $1 - \alpha(\theta)$ and $\alpha(\theta)$ respectively; that is, we define $\alpha(\theta)$ by $$ \alpha(\theta) := \frac{S(\theta)}{Q(\theta)} \lambda(\theta) \frac{s_1'q_1 \theta}{s_1}.$$ \medskip \noindent Next, we want the $\beta_j = \lambda \gamma_j$ to be of the form $\beta_j(\theta) = \frac{Q(\theta)}{P_j(\theta)S(\theta)}$ for certain $P_j(\theta)$; that is, $\frac{1}{P_j(\theta)} = \frac{S(\theta) \beta_j(\theta)}{Q(\theta)} = \frac{S(\theta) \lambda(\theta) \gamma_j(\theta)}{Q(\theta)}$. So, bearing in mind the definitions of $\gamma_j$ and $S/Q$, we define $P_j(\theta)$ by $$\frac{1}{P_j(\theta)} := \frac{\frac{q_0s_0'}{p_{j0} s_0}(1-\theta) + \frac{q_1s_1'}{p_{j1} s_1} \theta} {\frac{s_0'q_0 (1-\theta)}{s_0} + \frac{s_1'q_1 \theta}{s_1}}.$$ \medskip \noindent Finally, we define $Q(\theta)$ by $$ \frac{1}{Q(\theta)} := (1 - \alpha(\theta))\frac{1}{q_0} + \alpha(\theta) \frac{1}{q_1}.$$ It is not hard to check that with all these definitions in place, we have, for each $j$, $$ \frac{1}{P_j(\theta)} = (1 - \alpha(\theta))\frac{1}{p_{j0}} + \alpha(\theta) \frac{1}{p_{j1}}.$$ \medskip \noindent We therefore have that for each $ 0 \leq \theta \leq 1$, for all $G_\theta = G^{1/S'(\theta)}$ such that $\int G_\theta^{S'(\theta)} = 1$, there exist $M_{j\theta}$ such that $$G_\theta(x) \leq \prod_{j=1}^d M_{j\theta}(x)^{Q(\theta)/P_{j}(\theta)S(\theta)}$$ and, for $f_j$ such that $\int f_j \leq 1$, $$ \prod_{j=1}^d \left(\int f_j(B_jx) M_{j \theta}(x) {\rm d}x\right)^{Q(\theta)/P_j(\theta)S(\theta)} \leq \left(A_0^{1 - \alpha(\theta)} A_1^{\alpha(\theta)}\right)^{Q(\theta)/S(\theta)}.$$ Note particularly that the exponents $Q(\theta)/P_{j}(\theta)S(\theta)$ sum to $1$ since $\sum_{j=1}^d \beta_j = 1$. \medskip \noindent Consequently, using the flexibility that Remark~\ref{weaksuff} affords us, \begin{equation*} \left\|\prod_{j=1}^d T_jF_j\right\|_{L^{Q(\theta)}} \leq A_0^{1-\alpha(\theta)}A_1^{\alpha(\theta)} \prod_{j=1}^d \left\|F_j\right\|_{L^{P_{j}(\theta)}} \end{equation*} for $0 < \theta < 1$. Noting that the map $\alpha: [0,1] \to [0,1]$ is a surjection completes the argument proving \eqref{int2}. \medskip \noindent The argument given here provides no insight into cases in which \eqref{int2} might hold with a smaller constant than $A_0^{1 - \theta} A_1^\theta$. \subsection{Factorisation and multiple manifestations of generalised Radon transforms}\label{rrr} As we have observed in Section~\ref{exples} there may be multiple equivalent manifestations of the same multilinear inequality. For concreteness, suppose that we are once again considering multilinear generalised Radon transforms on euclidean spaces so that $T_j f = f \circ B_j$ for suitable $B_j$. Then the two inequalities \begin{equation*} \left\|\prod_{j=1}^d (T_jf_j)^{\alpha_j}\right\|_q \leq A \prod_{j=1}^d \Big\|f_j\Big\|_{p_j}^{\alpha_j} \end{equation*} and \begin{equation*} \left\|\prod_{j=1}^d (T_j\tilde{f}_j)^{\tilde{\alpha}_j}\right\|_{\tilde{q}} \leq \tilde{A} \prod_{j=1}^d \Big\|\tilde{f}_j\Big\|_{\tilde{p}_j}^{\tilde{\alpha}_j} \end{equation*} (where we are imposing $\sum_{j=1}^d \alpha_j = 1 = \sum_{j=1}^d \tilde{\alpha}_j$) are clearly equivalent provided that $A^q = \tilde{A}^{\tilde{q}}$ and $\alpha_j \tilde {p}_j/\tilde{\alpha}_j p_j = \tilde{q}/q$ for all $j$. The corresponding factorisation statements \bigskip {\em For all nonnegative $G \in L^{q^\prime}$ there exist nonnegative locally integrable functions $g_j$ such that} \begin{equation*} G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j}\qquad\mbox{a.e.} \end{equation*} {\em and such that for each $j$, for all $f_j \in L^{p_j}$,} \begin{equation*} \int g_j(x)f_j(B_jx) {\rm d} x \leq A\|G\|_{q^{\prime}}\|f_j\|_{p_j}. \end{equation*} and \bigskip \noindent {\em For all nonnegative $\tilde{G} \in L^{\tilde{q}^\prime}$ there exist nonnegative locally integrable functions $\tilde{g}_j$ such that} \begin{equation*} \tilde{G}(x) \leq \prod_{j=1}^d \tilde{g}_j(x)^{\tilde{\alpha}_j}\qquad\mbox{a.e.} \end{equation*} {\em and such that for each $j$, for all $\tilde{f}_j \in L^{\tilde{p}_j}$,} \begin{equation*} \int \tilde{g}_j(x)\tilde{f}_j(B_jx) {\rm d}x \leq \tilde{A}\|\tilde{G}\|_{L^{\tilde{q}^{\prime}}}\|\tilde{f}_j\|_{\tilde{p}_j}. \end{equation*} \bigskip are therefore also equivalent (subject to suitable hypotheses), by Proposition~\ref{thmguthbaby} and Theorem~\ref{thmmainbaby}. However it is not immediately apparent whether this equivalence can be seen directly via changes of notation coupled with simple convexity arguments. In this connection the remarks in Section~5.7 of \cite{MR2061575} may be helpful. \section{Factorisation and more general multilinear operators}\label{culture}\label{fdmgs} The multilinear operators we have considered have a rather special form in so far as they are built out of a collection of positive linear operators by taking a pointwise geometric mean. One may ask to what extent the theory we have developed is valid for more general multilinear operators $T : \mathcal{Y}_1 \times \dots \times \mathcal{Y}_d \to \mathcal{X}$. In such a setting we will no longer be able to attribute different ``weights'' $\alpha_j$ to the different components $\mathcal{Y}_j$, and all of them will need to be treated on an equal footing. \medskip \noindent For a nonnegative kernel $K$, let us therefore consider multilinear operators of the form $$ T(f_1, \dots , f_d)(x) = \int_{Y_d} \dots \int_{Y_1} K(x, y_1, \dots , y_d) f_1(y_1) \dots f_d(y_d) {\rm d} \nu_1(y_1) \dots {\rm d} \nu_d(y_d)$$ and inequalities of the form \begin{equation}\label{newmultilinear} \left\|T(f_1,\dots,f_d)^{1/d}\right\|_{\mathcal{X}}\leq A \prod_{j=1}^d \|f_j\|_{\mathcal{Y}_j}^{1/d}. \end{equation} \medskip \noindent When $K$ is of the form $K(x, y_1, \dots , y_d) = K_1(x, y_1) \dots K_d(x, y_d)$, these are the special cases of the inequalities \eqref{mainineq} given by $\alpha_j=1/d$ for $j=1,\dots,d$. It is very natural to ask whether there is a general duality/factorisation result along the same lines as Proposition~\ref{thmguth} and Theorem~\ref{thmmain} which yields a necessary and sufficient condition for the validity of inequality \eqref{newmultilinear}. \medskip \noindent As the reader will readily verify by following the proof of Proposition~\ref{thmguth}, inequality \eqref{newmultilinear} does indeed hold (under hypotheses on $\mathcal{X}$ and $\mathcal{Y}_j$ similar to those of Proposition~\ref{thmguth}), if, for all $G\in \mathcal{X}^\prime$ such that $\|G\|_{\mathcal{X}^\prime}\leq 1$, we have that there there exist nonnegative functions $S_j$ on $X\times Y_j$ such that \begin{equation}\label{newfactorisation} \begin{aligned} K(x,y_1,\dots,y_d)^{1/d}G(x)&\leq \prod_{j=1}^d S_j(x,y_j)^{1/d}\quad\text{a.e.}\\ \mbox{and }\left\|\int_X S_j(x,\cdot) d\mu(x)\right\|_{\mathcal{Y}_j^\ast}&\leq A. \end{aligned} \end{equation} \medskip \noindent This observation has proved very useful in multilinear Kakeya theory, see Section~\ref{MKrevisited} below. \medskip \noindent However, the converse is not true, namely inequality \eqref{newmultilinear} does not in general imply the existence of $S_j$ such that \eqref{newfactorisation} holds. To see this, consider the following example. \medskip \noindent Let $d=2$, $X=Y_1=Y_2$ be two-point spaces $\{1,2\}$ with counting measure, $\mathcal{X}=L^4(X)$, and $\mathcal{Y}_1=\mathcal{Y}_2=L^2(Y_1)$. Furthermore, let \begin{displaymath} K(1,1,1)=K(2,1,1)=K(2,2,2)=1 \end{displaymath} and let $K$ equal zero otherwise. Let $f_1=(a_1,a_2)$ and $f_2=(b_1,b_2)$ and we assume $a_1^2+a_2^2=b_1^2+b_2^2=1$. Then \begin{displaymath} T(f_1,f_2)(1) = a_1b_1\quad\text{and}\quad T(f_1,f_2)(2) = a_1b_1+a_2b_2 \end{displaymath} so \begin{displaymath} \int T(f_1,f_2)(x)^2 {\rm d}x = (a_1b_1)^2+(a_1b_1+a_2b_2)^2. \end{displaymath} This is clearly maximised, subject to the normalisation constraints, by taking $a_1=b_1=1$, $a_2=b_2=0$ and the maximum is $2$. So we see that the multilinear inequality \eqref{newmultilinear} holds for this operator with $A=2^{1/4}$. \medskip For problem \eqref{newfactorisation}, consider $G=(0,1)$. Then the non-trivial constraints are \begin{displaymath} 1\leq \sqrt{S_1(2,1)S_2(2,1)} \quad\text{and}\quad 1\leq \sqrt{S_1(2,2)S_2(2,2)} \end{displaymath} and \begin{gather*} \sqrt{(S_1(1,1)+S_1(2,1))^2+(S_1(1,2)+S_1(2,2))^2}\leq A\\ \sqrt{(S_2(1,1)+S_2(2,1))^2+(S_2(1,2)+S_2(2,2))^2}\leq A. \end{gather*} Using $uv\leq (u^2+v^2)/2$ on the lower bounds gives \begin{displaymath} 1\leq (S_1(2,1)^2+S_2(2,1)^2)/2 \quad\text{and}\quad 1\leq (S_1(2,2)^2+S_2(2,2)^2)/2, \end{displaymath} so \begin{displaymath} {2}\leq S_1(2,1)^2+S_2(2,1)^2 \quad\text{and}\quad {2}\leq S_1(2,2)^2+S_2(2,2)^2, \end{displaymath} so \begin{displaymath} 4\leq S_1(2,1)^2+S_2(2,1)^2+ S_1(2,2)^2+S_2(2,2)^2, \end{displaymath} and thus \begin{displaymath} 2\leq\max\{S_1(2,1)^2+S_1(2,2)^2,S_2(2,1)^2+S_2(2,2)^2\} \end{displaymath} giving $A \geq 2^{1/2}$, which is strictly larger than $2^{1/4}$. So while inequality \eqref{newmultilinear} holds in this case, there are $G$ for which there are {\em no} $S_j$ satisfying \eqref{newfactorisation}. \medskip \noindent We invite the reader to use this idea to construct examples where \eqref{newmultilinear} holds with $A=1$ but for which \eqref{newfactorisation} holds for no finite $A$. \medskip \noindent See \cite{GT} for a different approach to inequalities of the form \eqref{newmultilinear}, based upon considerations related to Schur's lemma rather than duality. \part{Examples and applications of the theory} \bigskip \noindent In this part we revisit the examples in the introduction which motivated our study. We examime what insights our duality--factorisation results bring to, and have gained from, each of them. In some cases we reap the benefits of more direct and streamlined factorisation-based proofs of known inequalities. In others, an interesting challenge is posed -- it can be argued that we cannot claim to have a full understanding of an inequality until we can exhibit its equivalent factorisation statement. \section{Classical inequalities revisited}\label{classical} \subsection{H\"older's inequality}\label{Holderrevisited} We observed above that the multilinear form of H\"older's inequality for nonnegative functions is equivalent, for any fixed set of exponents $\alpha_j > 0$ with $\sum_{j=1}^d \alpha_j = 1$, to $$ \| f_1^{\alpha_1} \cdots f_d^{\alpha_d} \|_q \leq \|f_1 \|_{q_1}^{\alpha_1} \cdots\|f_d \|_{q_d}^{\alpha_d}$$ for any choice of indices $ 1 \leq q_j < \infty$ and $1 \leq q < \infty$ which satisfies $\sum_{j=1}^d \alpha_j q_j^{-1} = q^{-1}$. \medskip \noindent By Theorem~\ref{thmmain}, each instance of this inequality is equivalent to the existence of a subfactorisation of any $ G \in L^{q'}$ as $$ G(x) \leq \prod_{j=1}^d g_j(x)^{\alpha_j} \; \; \mbox{ a.e.}$$ where $$\|g_j \|_{{q_j}^\prime} \leq \|G\|_{q'}.$$ \medskip \noindent Taking $g_j = \lambda_j G^{\gamma_j}$ for appropriate $\lambda_j$ and $\gamma_j$ verifies this. In particular, if we take $q_j = q \geq 1$ for all $j$, then we can simply take $g_j = G$ for all $j$. \subsection{Affine-invariant Loomis--Whitney inequality}\label{Loomis--Whitneyrevisited} Recall that the Loomis--Whitney inequality is $$ |\int_{\mathbb{R}^n} F_1(\pi_1 x) \cdots F_n(\pi_n x) \, {\rm d} x| \leq \|F_1\|_{L^{n-1}(\mathbb{R}^{n-1})} \cdots \|F_n\|_{L^{n-1}(\mathbb{R}^{n-1})},$$ where $\pi_j x = (x_1, \dots, \widehat{x_j}, \dots, x_n)$ is projection onto the hyperplane perpendicular to the $j$'th standard basis vector $e_j$. For every $0 < p < \infty $ this is equivalent to the inequality $$ \|f_1(\pi_1 x)^{1/n} \cdots f_n(\pi_n x)^{1/n} \|_{L^{np/(n-1)}(\mathbb{R}^n)} \leq \|f_1\|_{L^{p}(\mathbb{R}^{n-1})}^{1/n} \cdots \|f_n\|_{L^{p}(\mathbb{R}^{n-1})}^{1/n}.$$ Each of these inequalities with $p \geq 1$ falls under the scope of our theory. \medskip \noindent For example when $p =1$ we have the equivalent formulation \begin{equation*} \int_{\mathbb{R}^n} \prod_{j=1}^n f_j(\pi_{j} x)^{1/(n-1)} {\rm d}x \leq \prod_{j=1}^n \left(\int_{\mathbb{R}^{n-1}} f_j\right)^{1/(n-1)}. \end{equation*} \medskip \noindent More generally, if $\pi_{\omega^\perp}$ represents orthogonal projection onto the hyperplane perpendicular to $\omega \in \mathbb{S}^{n-1}$, we have the affine-invariant Loomis--Whitney inequality \begin{equation}\label{afflw} \int_{\mathbb{R}^n} \prod_{j=1}^n f_j(\pi_{\omega_j^\perp} x)^{1/(n-1)} {\rm d}x \leq (\omega_1 \wedge \dots \wedge \omega_n)^{-1/(n-1)} \prod_{j=1}^n \left(\int_{\mathbb{R}^{n-1}} f_j\right)^{1/(n-1)}, \end{equation} where $(\omega_1 \wedge \dots \wedge \omega_n)^{-1/(n-1)}$ is the best constant in the inequality. Here, $\omega_1 \wedge \dots \wedge \omega_n$ is the modulus of the determinant of the matrix whose columns are $\omega_1, \dots, \omega_n$, and it is the volume of the parallepiped whose sides are given by the vectors $\omega_j$. (Clearly if we choose all the $\omega_j$ to be the same we cannot expect a finite constant, and the constant in general should reflect ``quantitative linear independence'' of the $\omega_j$.) \medskip \noindent We give a direct and elegant proof of \eqref{afflw} by explicitly establishing a suitable factorisation. Indeed, according to Proposition~\ref{thmguthbaby}, it is sufficient that for every nonnegative $M \in L^n(\mathbb{R}^n)$ we can find $S_1, \dots, S_n$ such that $$ M(x) = S_1(x)^{1/n} \cdots S_n(x)^{1/n} \mbox{ a.e.}$$ and, for all $j$ and almost every $x$, $$ \int S_j(x + t \omega_j) {\rm d}t = (\omega_1 \wedge \dots \wedge \omega_n)^{-1/n} \|M\|_n.$$ This is because for any $f : \mathbb{R}^{n-1} \to \mathbb{R}$ and $S: \mathbb{R}^n \to \mathbb{R}$, writing $x \in \mathbb{R}^n$ as $x = u + t \omega_j$ with $u \in \omega_j^\perp$, we have $$\int f(\pi_{\omega_j^\perp} x) S(x) {\rm d}x = \int_{\mathbb{R}^{n-1}} \int_\mathbb{R} f(\pi_{\omega_j^\perp}(u + t \omega_j))S(u + t \omega_j) {\rm d}t {\rm d}u $$ $$ = \int_{\mathbb{R}^{n-1}} f(u) \left(\int_{\mathbb{R}} S(u + t \omega_j) {\rm d}t \right){\rm d}u.$$ \medskip \noindent Let $M :\mathbb{R}^n \to \mathbb{R}$ be a nonnegative function which satisfies $\int_{\mathbb{R}^n} M(x)^n {\rm d}x = 1$. For $\omega_1, \dots, \omega_n \in \mathbb{S}^{n-1}$ and $\xi \in \mathbb{R}^n$ let us first note that if we set, for $s = (s_1, \dots , s_n) \in \mathbb{R}^n$, $$y(s) = \xi + s_1 \omega_1 + \dots + s_{n-1} \omega_{n-1} + s_n \omega_n,$$ then we have that the Jacobian map $\partial y / \partial s $ satisfies $$ |\det \left(\partial y / \partial s \right)| = \omega_1 \wedge \dots \wedge \omega_n.$$ \medskip \noindent Therefore, for every $\xi \in \mathbb{R}^n$, $$\int M(\xi + s_1 \omega_1 + \dots + s_{n-1} \omega_{n-1} + s_n \omega_n)^n {\rm d}s_1 {\rm d}s_2 \dots {\rm d}s_{n} = \int M(y(s))^n {\rm d}s $$ $$ = \int M(y)^n \frac{1}{ |\det \left(\partial y / \partial s \right)|} {\rm d}y = (\omega_1 \wedge \dots \wedge \omega_n)^{-1}.$$ \medskip \noindent Secondly, $M(x)^n$ can be written (for a.e. $x$) as a telescoping product \begin{eqnarray*} \begin{aligned} & \frac{M(x)^n}{\int M(x + s_1 \omega_1)^n {\rm d}s_1} \times \frac{\int M(x + s_1 \omega_1)^n {\rm d}s_1}{\int M(x + s_1 \omega_1 + s_2 \omega_2)^n {\rm d}s_1 {\rm d}s_2} \times \dots \\ & \times \frac{\int M(x + s_1 \omega_1 + \dots + s_{n-1} \omega_{n-1})^n{\rm d} s_1 {\rm d}s_2 \dots {\rm d}s_{n-1}}{\int M(x + s_1 \omega_1 + \dots + s_{n-1} \omega_{n-1} + s_n \omega_n)^n {\rm d}s_1 {\rm d}s_2 \dots {\rm d}s_{n}} \times \left(\omega_1 \wedge \dots \wedge \omega_n\right)^{-1} \end{aligned} \end{eqnarray*} \medskip $$ := S_1(x) \dots S_n(x)$$ where $$ S_j(x) = \frac{\int M(x + s_1 \omega_1 + \dots + s_{j-1} \omega_{j-1})^n {\rm d}s_1 {\rm d}s_2 \dots {\rm d}s_{j-1}}{\int M(x + s_1 \omega_1 + \dots + s_{j-1} \omega_{j-1} + s_j \omega_j)^n{\rm d}s_1 {\rm d}s_2 \dots {\rm d}s_{j}} \times \left(\omega_1 \wedge \dots \wedge \omega_n\right)^{-1/n}.$$ \medskip \noindent If we replace $x$ by $x + t \omega_j$ in this formula, the denominator is unchanged, and so if we then integrate with respect to $t$ we immediately see that $$\int S_j(x + t \omega_j) {\rm d} t = \left(\omega_1 \wedge \dots \wedge \omega_n\right)^{-1/n}$$ identically for $x \in \mathbb{R}^n$, as we needed. \medskip \noindent A similar approach works when we instead consider projections onto subspaces whose codimensions sum to $n$. Indeed, suppose that we have subspaces $E_j$ of $\mathbb{R}^n$ with $\dim E_j = k_j$ and $\sum_{j=1}^d k_j = n$ and assume that $\mathbb{R}^n = E_1 + \dots + E_d$ as an algebraic direct sum. \medskip \noindent We identify a quantity which measures lack of orthogonality of these subspaces in the same way that the wedge product $\omega_1 \wedge \dots \wedge \omega_n$ measures the degeneracy in the directions $\omega_1, \dots , \omega_n \in \mathbb{S}^{n-1}$. Let $\{e_{j1}, e_{j2}, \dots , e_{jk_j}\}$ be an orthonormal basis for $E_j$ and define $$E_1 \wedge \dots \wedge E_d := \wedge_{j=1}^d \wedge_{k=1}^{k_j} e_{jk};$$ that is, $E_1 \wedge \dots \wedge E_d$ is the absolute value of the determinant of the $n \times n$ matrix whose $j$'th block of $k_j$ columns comprises an orthonormal basis for $E_j$. It is easily checked that this quantity is independent of the particular orthonormal bases chosen, and it can of course be defined in a more canonical and invariant way. \medskip \noindent For $E_j$ as above, let $\pi_j$ be the projection whose kernel is $E_j$. Then we have the affine-invariant $k_j$-plane Loomis--Whitney inequality: \begin{equation}\label{aikjplw} \begin{aligned} &\int_{\mathbb{R}^n} f_1(\pi_1 x)^{1/(d-1)} \dots f_d(\pi_d x)^{1/(d-1)} {\rm d}x \\ &\leq \left(E_1 \wedge \dots \wedge E_d\right)^{-1/(d-1)} \left(\int f_1 \right)^{1/(d-1)} \dots \left(\int f_d \right)^{1/(d-1)}. \end{aligned} \end{equation} \medskip \noindent The proof via factorisation is formally the same as the case when $k_j = 1$ for all $j$, where now the roles of the variables $s_j \in \mathbb{R}^1$ are replaced by copies of $\mathbb{R}^{k_j}$. We leave the details to the reader. \medskip \noindent In the special case of the trivial identity, $$ \int_{\mathbb{R}^2} F_1(x_2)F_2(x_1) \; {\rm d} x_1 {\rm d} x_2 = \int_\mathbb{R} F_1 \int_\mathbb{R} F_2,$$ (see Section~\ref{factinterp}), a suitable factorisation of $M \in L^2(\mathbb{R}^2)$ with $\|M\|_2 = 1$ is given by $M(x)^2 = S_1(x) S_2(x)$ a.e., where $$S_1(x_1, x_2) = \frac{M(x_1, x_2)^2}{\int_\mathbb{R} M(s, x_2)^2 {\rm d} s}$$ and $$S_2(x_1, x_2) = \int_\mathbb{R} M(s, x_2)^2 {\rm d} s.$$ \medskip \noindent Note that this factorisation depends upon the order we have assigned to $\{1,2\}$. On the other hand, given this ordering, the essentially unique way to write $$ M(x)^2 = S_1(x_1, x_2)S_2(x_2)$$ where $\|S_1(\cdot, x_2)\|_1 = 1$ for all $x_2$ and $\|S_2\|_1 = 1$ is as we have given. See Section \ref{BLfactdetails}, where this observation drives related issues. \medskip \noindent There are many variants of the Loomis--Whitney inequality -- for example Finner's inequalities \cite{Fi} -- which can likewise be established by the same factorisation method. \subsection{The nonlinear Loomis--Whitney inequality}\label{mgrtrevisited}\label{NLLW} Nonlinear Loomis--Whitney inequalities (and some multilinear generalised Radon transforms) can likewise be established by similar methods. In fact the first proof of the nonlinear Loomis--Whitney inequality with essentially the sharp constant was obtained via an explicit factorisation technique. We give the details. \medskip \noindent Let $V$ be an open neighbourhood of $0$ in $\mathbb{R}^n$ and $U$ an open neighbourhood of $0$ in $\mathbb{R}^{n-1}$. Let $\pi: V \to U$ be a $C^1$ submersion onto $U$, and for $x \in V$ let $\omega(x)$ be the wedge product of the rows of ${\rm d} \pi(x)$. We assume that the fibres $\pi^{-1}(u)$ for $u \in U$ can be parametrised by $C^1$ curves $t \mapsto \gamma(t,x)$ in such a way that \begin{itemize} \item for all $x \in V $, $\gamma(0,x) = x$ \item for all $x \in V $, for all $t$, $\pi \gamma(t,x) = \pi x$ \item (semigroup property) for all $x \in V $, for all $t$ and $s$, $$\gamma(t, \gamma(s,x)) = \gamma(s+t, x)$$ \item for all $x$ and $t$, $\frac{\rm d}{{\rm d}t} \gamma(t,x) = \omega(\gamma(t,x))$. \end{itemize} \medskip \noindent The domain of each curve $\gamma(\cdot, x)$ will be an open interval $I_x$ containing $0$ which we largely suppress in what follows, but we stress that $\gamma(I_x,x)$ is the entire fibre containing $x$. In all the $t$-integrals below it is assumed that we are integrating over such maximal domains. \medskip \noindent We note that under these assumptions, especially the last one, the co-area formula gives $$ \int_V f(\pi x) S(x) {\rm d}x = \int_U f(u) \left(\int S(\gamma(t, \tilde{u})) {\rm d} t \right){\rm d}u$$ for any reasonable functions $f$ and $S$. \medskip \noindent We now assume that we have $n$ submersions $\pi_1, \dots, \pi_n$ as above, and we assume that $\omega_1(0) \wedge \dots \wedge \omega_n(0) \neq 0$. For each $x \in V$ we define the maps ${\bf t} \mapsto \Phi_x({\bf t})$ by $$ \Phi_x : (t_1, \dots, t_n) \mapsto \gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_n(t_n,x)) \dots )$$ which satisfy $\Phi_x(0) = x$ and also $$ |\det( D \Phi_x) (0)| = (\omega_1 \wedge \dots \wedge \omega_n)(x) \neq 0$$ provided $x$ is sufficiently close to $0$. \medskip \noindent We shall assume that $V$ is sufficiently small so that for each $x \in V$, the map $ \Phi_x $ is injective -- as was pointed out in \cite{BCW}, even in two dimensions some global hypothesis of this sort is needed. \medskip \noindent With the set-up above, for $x \in V$ let $$W(x) := \inf_{\xi \in V} \det |(D \Phi_\xi)(\Phi_\xi^{-1}(x))|.$$ Note that $W(x) \leq \omega_1(x) \wedge \dots \wedge \omega_n(x)$, (take $\xi = x$), and that $W(\Phi_x({\bf t})) \leq |\det(D \Phi_x)({\bf t})|$ for all $x$ and ${\bf t}$. \medskip \noindent For $1 \leq j \leq n$ and suitable $F$ let $$ S_j(x) = \frac{\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots , \gamma_{j-1}(t_{j-1},x)) \dots )) {\rm d}t_{j-1} \dots {\rm d}t_{1}} {\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots , \gamma_{j}(t_{j},x)) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1}}$$ (so that $S_1$ has no integrals in the numerator). \medskip \noindent Then we have $$ S_j( \gamma_j (\tau, x)) = \frac{\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j-1}(t_{j-1},\gamma_j (\tau, x))) \dots )) {\rm d}t_{j-1} \dots {\rm d}t_{1}} {\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j}(t_{j},\gamma_j (\tau, x))) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1}}.$$ \medskip \noindent We claim that for each $j$ and each $x$, $$\int S_j( \gamma_j (\tau, x)) {\rm d} \tau =1.$$ \medskip \noindent Indeed, notice that the denominator in the previous expression, $$\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j}(t_{j},\gamma_j (\tau, x))) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1},$$ equals $$ \int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j}(t_{j} + \tau , x )) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1}$$ $$ = \int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j}(t_{j}, x )) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1}$$ by the semigroup property, and is therefore independent of $\tau$. So $$ \int S_j( \gamma_j (\tau, x)) {\rm d} \tau = \frac{\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j-1}(t_{j-1},\gamma_j (\tau, x))) \dots )) {\rm d}t_{j-1} \dots {\rm d}t_{1} {\rm d} \tau} {\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{j}(t_{j}, x )) \dots )) {\rm d}t_{j} \dots {\rm d}t_{1}}$$ which equals $1$ by Fubini's theorem. \medskip \noindent On the other hand, $$ \prod_{j=1}^n S_j(x) = \frac{F(x)}{\int \dots \int F(\gamma_1(t_1, \gamma_2(t_2, \dots ,\gamma_{n}(t_{n},x)) \dots )) {\rm d}t_{n} \dots {\rm d}t_{1}},$$ so that $$F(x) = \prod_{j=1}^n S_j(x) \int F(\Phi_x({\bf t})){\rm d} {\bf t}.$$ \medskip \noindent Taking $F(x) = M(x)^n W(x)$, we therefore have $$ M(x)^n W(x) = \prod_{j=1}^n S_j(x) \int M(\Phi_x({\bf t}))^n W(\Phi_x({\bf t})) {\rm d} {\bf t}$$ $$ \leq \prod_{j=1}^n S_j(x) \int M(\Phi_x({\bf t}))^n |\det (D \Phi_x)({\bf t})| {\rm d} {\bf t} = \prod_{j=1}^n S_j(x) \int_V M(y)^n {\rm d} y $$ since $W(\Phi_x({\bf t})) \leq |\det (D \Phi_x)({\bf t})|$ for all ${\bf t}$ and since each $\Phi_x$ is injective. We also have that for each $j$ and each $x$, $$ \int_V f(\pi_j x) S_j(x) {\rm d}x = \int_{U_j} f(u)\left(\int S_j( \gamma_j (\tau, \tilde{u})) {\rm d} \tau \right) {\rm d} u = \int_{U_j} f.$$ \medskip \noindent By the easy half of the duality argument, this shows that for all nonnegative $f_j \in L^1(U_j)$ we have $$\| \prod_{j=1}^n f_j(\pi_j x)^{1/n} W(x)^{1/n} \|_{L^{n/(n-1)}(V)} \leq \prod_{j=1}^n \left(\int_{U_j} f_j \right)^{1/n}$$ which is the same as $$ \int_V \prod_{j=1}^n f_j(\pi_j x)^{1/(n-1)} W(x)^{1/(n-1)} \; {\rm d} x \leq \prod_{j=1}^n \left(\int_{U_j} f_j \right)^{1/(n-1)}.$$ Now given $\epsilon >0$ we can choose $V$ sufficiently small that $\omega_1(0) \wedge \dots \wedge \omega_n(0) \leq (1+ \epsilon)^{n-1}W(x)$ for all $x \in V$, and therefore, for such $V$, $$ \int_V \prod_{j=1}^n f_j(\pi_j x)^{1/(n-1)} \; {\rm d} x \leq (1+\epsilon) \omega_1(0) \wedge \dots \wedge \omega_n(0)^{-1/(n-1)} \prod_{j=1}^n \left(\int_{U_j} f_j \right)^{1/(n-1)}$$ which is the sharp form of the nonlinear Loomis--Whitney inequality of \cite{BCW}. Since the work in this section was presented in various public fora, Bennett et al \cite{Betal} have shown, using methods based on induction on scales, that any Brascamp--Lieb inequality with so-called simple data has a corresponding nonlinear counterpart with loss in the constant of at most $1 + \epsilon$. \medskip \noindent Noting that $W(x) \leq \omega_1(x) \wedge \dots \wedge \omega_n(x)$, one might ask whether $$ \int_V \prod_{j=1}^n f_j(\pi_j x)^{1/(n-1)} \omega_1(x) \wedge \dots \wedge \omega_n(x)^{1/(n-1)} \; {\rm d} x \leq \prod_{j=1}^n \left(\int_{U_j} f_j \right)^{1/(n-1)}$$ holds for sufficiently small $V$. \section{Brascamp--Lieb inequalities revisited}\label{BLrevisited} We shall discuss the Brascamp--Lieb inequalities under two headings. Firstly we shall address geometric Brascamp--Lieb inequalities (where in particular we can identify the sharp constant and existence of gaussian extremisers), and secondly we will examine general Brascamp--Lieb inequalities with a finite (but unquantified) constant. \subsection{Geometric Brascamp--Lieb inequalities}\label{GBL} The next result is a direct application of Theorem~\ref{thmmainbaby} to the geometric Brascamp--Lieb inequalities of Example~\ref{BL}. \begin{thm} For $1 \leq j \leq d$ let $V_j$ be a subspace of $\mathbb{R}^n$. Let $B_j: \mathbb{R}^n \to V_j$ be orthogonal projection. Suppose there exist $p_j$ with $0 < p_j < \infty$ such that $$ \sum_{j=1}^d p_j B_j^* B_j = I_n.$$ Let $1 \leq q_j < \infty$, and define $q = \sum_{j=1}^d p_j q_j$. Then for all $G \in L^{q'}(\mathbb{R}^n)$ there exist $g_1, \dots , g_d$ such that $$ G(x) \leq g_1(x)^{p_1q_1/q} \dots g_1(x)^{p_dq_d/q} \; \; {\mbox{ a.e.}}$$ and, for all $j$, $$ \left\| \int_{V_j^\perp} g_j \right\|_{L^{{q_j}^{\prime}}(V_j)} \leq \left\|G\right\|_{q'}.$$ \end{thm} One simply needs to note (see the discussion in Example~\ref{BL}) that under the hypothesis of this theorem, $$ \left\| \prod_{j=1}^d (f_j\circ B_j)^{p_jq_j/q} \right\|_{L^q(\mathbb{R}^n)} \leq \prod_{j=1}^d \left\|f_j\right\|_{L^{q_j}(V_j)}^{p_jq_j/q}$$ and $\sum_{j=1}^d p_j \geq 1$, and thus $q \geq 1$. Therefore Theorem~\ref{thmmainbaby} applies. \medskip \noindent Except in some rather trivial cases\footnote{For example when the $V_j$ are mutually orthogonal and $p_j = 1$ for all $j$.} we do not know any such explicit factorisations with the sharp constant $1$. For example, let $v_1, v_2$ and $v_3$ be unit vectors in $\mathbb{R}^2$ with angle $2\pi/3$ between each pair. Then, with $B_j$ being orthogonal projection onto the span of $v_j$, we have $$ \frac{2}{3} (B_1^*B_1 + B_2^*B_2 + B_3^*B_3) = I_2.$$ Take $q_j = 1$ for each $j$ so that $q = 2$. Consequently, for all $G \in L^2(\mathbb{R}^2)$, there exist $g_1, g_2, g_3$ such that $$ G(x) \leq g_1(x)^{1/3} g_2(x)^{1/3} g_3(x)^{1/3} \; \; \mbox{ a.e.}$$ and, for each $j$, $$ \sup_s \int_{\mathbb{R}} g_j(sv_j + tv_j^\perp) {\rm d}t \leq \|G\|_2.$$ Even in such simple cases as this the factorisation is not yet understood explicitly. \subsection{General Brascamp--Lieb inequalities}\label{GGBL} On the other hand, under the conditions \begin{equation}\label{scalingxx} \sum_{j=1}^d p_j \, {\rm dim \; im} B_j = n \end{equation} and \begin{equation}\label{dimensionxx} {\rm dim} \, V \leq \sum_{j=1}^d p_j \, {\rm dim} B_j V \end{equation} for all $V$ in the lattice of subspaces of $\mathbb{R}^n$ generated by $\{ \ker B_j\}_{j=1}^d$, we now indicate how to construct semi-explicit factorisations yielding the finiteness of the constant $C$ in \eqref{BL1}. We use the term ``semi-explicit'' because the construction is algorithmic in nature. Notwithstanding, we give an informal discursive treatment rather than a collection of flow-charts. We assume throughout the discussion that the $B_j$ are nonzero mappings, that is, $n_j = {\rm rank}(B_j) \geq 1$ for each $j$. (If some $B_j =0$ it plays no role in inequality \eqref{BL1}, nor in \eqref{scalingxx} or \eqref{dimensionxx}, and it can simply be dropped.) When $n=1$ matters quickly reduce to consideration of H\"older's inequality, which is treated in Section~\ref{Holderrevisited} above, so we shall focus on what happens when $n \geq 2$. \medskip \noindent We now sketch how this is done, and we begin with a couple of definitions from \cite{BCCT1} and \cite{BCCT2}. Given a collection of linear surjections $\{B_j\}$, its {\em Brascamp--Lieb polytope} is defined by $$\mathcal{P}(\{B_j\}) = \{(p_1, \dots, p_d) \in [0, \infty)^d \, : \, {\rm dim} \, V \leq \sum_{j=1}^d p_j \, {\rm dim} B_j V \mbox{ for all subspaces }V \}.$$ This is manifestly a closed convex set, and, as has been previously noted, is contained in $[0,1]^d$, and is therefore the convex hull of its extreme points. Given data $\{B_j\}$ and $\{p_j\}$, a {\em critical subspace} is a nontrivial proper subspace $V$ of $\mathbb{R}^n$ for which \begin{equation}\label{crit} {\rm dim} \, V = \sum_{j=1}^d p_j \, {\rm dim} B_j V. \end{equation} The construction of the factorisations hinges on the question of existence or non-existence of critical subspaces for the problem with data $\{B_j, p_j\}$. Indeed, if there is a critical subspace $V$ for $\{B_j, p_j\}$, then the problem of factorising a function on $\mathbb{R}^n$ decomposes into two factorisation subproblems on the spaces $V$ and $V^\perp$, each of which has positive but {\em strictly smaller} dimension than $n$.\footnote{To facilitate the discussion which follows we should strictly speaking replace the roles of $\mathbb{R}^n$ and $\mathbb{R}^{n_j}$ by those of abstract $n$- and $n_j$-dimensional real Hilbert spaces respectively.} This will allow us in effect to induct on the parameter $n$. These subproblems inherit the same $\{p_j\}$ and have ``new'' $B_j$ which are related to the ``old'' $B_j$ in a precise way. The two subproblems inherit the conditions corresponding to \eqref{scalingxx} and \eqref{dimensionxx}: indeed, \eqref{scalingxx} for each of the two subproblems holds precisely because the subspace $V$ is critical, and we shall make crucial use of this fact. We isolate the details of how this works -- in particular how factorisations for the two subproblems combine to give a factorisation for the original problem -- in Section~\ref{BLfactdetails} below. \medskip \noindent On the other hand, if there is no critical subspace for the problem $\{B_j, p_j\}$, then $(p_1, \dots , p_d)$ lies in the interior of $\mathcal{P}(\{B_j\})$. To establish a factorisation for the problem in this case, it therefore suffices to (i) establish factorisations for the extreme points of $\mathcal{P}(\{B_j\})$ and (ii) to show, given factorisations at the extreme points, how to establish factorisations at all interior points of $\mathcal{P}(\{B_j\})$. Point (ii) is tantamount to showing that factorisations behave well under multilinear interpolation, and this we have already successfully addressed separately in Section~\ref{qzf}. \medskip \noindent To deal with point (i), we consider the Brascamp--Lieb problems at the extreme points $(\tilde{p}_1, \dots, \tilde{p}_d)$ of $\mathcal{P}(\{B_j\})$,\footnote{An algorithm for locating these extreme points can be found in \cite{Vald}.} and, at each of them, ask the same question -- does there exist a critical subspace? Since $(\tilde{p}_1, \dots, \tilde{p}_d)$ is an extreme point of $\mathcal{P}(\{B_j\})$, there will certainly be subspaces $V$ of $\mathbb{R}^n$ satisfying \begin{equation}\label{crit'} {\rm dim} \, V = \sum_{j=1}^d \tilde{p}_j \, {\rm dim} B_j V. \end{equation} If there is a {\em nontrivial and proper} such subspace, we have a critical subspace for the problem $\{B_j, \tilde{p}_j\}$, and we can proceed as above, in effect going around the loop. The only remaining possibility is that the only subspaces $V$ of $\mathbb{R}^n$ satisfying \eqref{crit'} are $\{0\}$ and $\mathbb{R}^n$ itself. \medskip \noindent We are thus left to deal with the special case of our original problem in which $(p_1, \dots, p_d)$ is an extreme point of $\mathcal{P}(\{B_j\})$, but for which the only subspaces of $V$ of $\mathbb{R}^n$ satisfying $\eqref{crit}$ are $\{0\}$ and $\mathbb{R}^n$ itself. Matters quickly reduce to rather trivial considerations. Indeed, in this situation, $\mathcal{P}(\{B_j\})$ consists precisely of those $(p_1, \dots, p_d) \in [0,\infty)^d$ lying on the hyperplane $\sum_{j=1}^d p_j n_j = n$, and its extreme points are precisely those of the form $(0, \dots, 0, n/n_j , 0 , \dots, 0)$. Since $n_j \leq n$ always, and since $\mathcal{P}(\{B_j\}) \subseteq [0,1]^d$, the only circumstances in which this case arises is when $n_j = n$ for all $j$. In this case, our Brascamp--Lieb problem at an extreme point of $\mathcal{P}(\{B_j\})$ is necessarily of the form (modulo permutations of the coordinate axes) $$ \int_{\mathbb{R}^n} f_1(B_1 x)^1f_2(B_2 x)^0 \dots f_d(B_d x)^0 {\rm d}x \leq C \left(\int_{\mathbb{R}^n} f_1\right)^1 \left(\int_{\mathbb{R}^n} f_2\right)^0 \dots \left(\int_{\mathbb{R}^n} f_d\right)^0$$ or, equivalently, $$ \int_{\mathbb{R}^n} f_1(B_1 x){\rm d}x \leq C \left(\int_{\mathbb{R}^n} f_1\right)$$ where $B_1$ is invertible. This of course holds with equality with $C = (\det B_1)^{-1}$, and a trivial factorisation applies. \medskip \noindent Running the machine described above in reverse will thus eventually furnish a factorisation in the general case, and, indeed, the only possible loss in terms of sharp constants occurs at steps where interpolation is employed. \subsubsection{Factorisation in the presence of a critical subspace}\label{BLfactdetails} We give the details needed to close the argument set out above in the presence of a critical subspace. The only place we use criticality is that it implies that \eqref{scalingxx} and \eqref{dimensionxx} hold for the two subproblems which arise -- see \cite{BCCT1}. Since these are the necessary and sufficient conditions for finiteness of the constant, we may assume that factorisations for the two subproblems exist. (Formally we proceed by induction on $n$, and the case $n=1$ is trivial.) \medskip \noindent Let $B_j : \mathbb{R}^n \to \mathbb{R}^{n_j}$ be linear surjections. Suppose that $U$ is a nontrivial proper subspace of $\mathbb{R}^n$. (As indicated above, we do not assume that it is a critical subspace.) Define $ \tilde{B}_j :\, U \to B_jU$ and $\tilde{\tilde{B}}_j :\, U^\perp \to (B_jU)^\perp$ by $$ \tilde{B}_j(x) = B_j x$$ and $$ \tilde{\tilde{B}}_j(y) = \Pi_{(B_jU)^\perp} B_j y.$$ If some $\tilde{B}_j$ or $ \tilde{\tilde{B}}_j$ is zero we can simply discard it. (It cannot be the case that every $\tilde{B}_j$ is zero, for if this happened, we would have $U \subseteq \cap_{j=1}^d \ker B_j$, and, as we have noted previously, a necessary condition for finiteness of the Brascamp--Lieb constant is that $\cap_{j=1}^d \ker B_j = \{0\}$. For similar reasons it cannot be the case that every $\tilde{\tilde{B}}_j$ is zero.) \medskip \noindent Also define $\Gamma_j : \, U^\perp \to B_jU$ by $$ \Gamma_j(y) = \Pi_{(B_jU)} B_j y.$$ Here, $\Pi_W$ denotes orthogonal projection onto a subspace $W$. So for $x \in U$ and $y \in U^\perp$, $$B_j(x + y) = \tilde{B}_jx + \tilde{\tilde{B}}_j y + \Gamma_j y = \left(\tilde{B}_jx + \Gamma_j y\right) + \tilde{\tilde{B}}_j y \in B_jU \oplus (B_jU)^\perp.$$ \medskip \noindent The two Brascamp--Lieb subproblems arising can be written in the form $$ \int_U \prod_{j=1}^d f_j(\tilde{B}_j x)^{p_j/p} {\rm d} x \leq C \prod_{j=1}^d \left(\int f_j\right)^{p_j/p}$$ and $$ \int_{U^\perp} \prod_{j=1}^d f_j(\tilde{\tilde{B}}_j x)^{p_j/p} {\rm d} x \leq C \prod_{j=1}^d \left(\int f_j\right)^{p_j/p}$$ where $p = \sum_j p_j \geq 1$. With $\alpha_j = p_j/p$, we are entitled to suppose that the following two corresponding factorisation statments hold: \medskip \noindent For all $H \in L^{p'}(U)$ of norm $1$ there exist $H_1, \dots, H_d$ such that $$ H(x) \leq \prod_{j=1}^d H_j(x)^{\alpha_j}$$ and, for all $\phi \in L^1(B_jU)$ of norm at most $1$, $$ \int_U \phi(\tilde{B}_j x) H_j(x) {\rm d}x \leq K_1;$$ \medskip \noindent For all $M \in L^{p'}(U^\perp)$ of norm $1$ there exist $M_1, \dots, M_d$ such that $$ M(y) \leq \prod_{j=1}^d M_j(y)^{\alpha_j}$$ and, for all $\psi \in L^1((B_jU)^\perp)$ of norm at most $1$, $$ \int_{U^\perp} \psi(\tilde{\tilde{B}}_j y) M_j(y) {\rm d}y \leq K_2.$$ \medskip \noindent Given $G \in L^{p'}(\mathbb{R}^n)$ of norm $1$, we want to subfactorise it as $$ G(z) \leq \prod_{j=1}^d G_j(z)^{\alpha_j}$$ such that for all $f \in L^1(\mathbb{R}^{n_j})$ of norm at most $1$, $$ \int_{\mathbb{R}^n} f({B}_j z) G_j(z) {\rm d}z \leq K_1 K_2.$$ \medskip \noindent This is a factorisation statement corresponding to the problem $$ \int_{\mathbb{R}^n} \prod_{j=1}^d f_j({B}_j x)^{p_j/p} {\rm d} x \leq C \prod_{j=1}^d \left(\int f_j\right)^{p_j/p}.$$ If we can do this, then the procedure described above for factorising Brascamp--Lieb problems closes. \medskip \noindent We begin by writing $G \in L^{p'}$ of norm $1$ as \begin{equation}\label{qfj} G(x,y) = H_y(x) M(y) \end{equation} where $\|H_y\|_{p'} = 1$ for all $y$ and $\|M\|_{p'} = 1$. We will then factorise $M$ and each $H_y$ as above, and combine the factorisations to obtain a suitable factorisation for $G$. \medskip \noindent Indeed, defining $H_y$ and $M$ by $$ G(x,y) = \frac{G(x,y)}{\left(\int G(x,y)^{p'} {\rm d} x\right)^{1/p'}} \left(\int G(x,y)^{p'} {\rm d} x\right)^{1/p'} := H_y(x) M(y)$$ is essentially the unique way to achieve \eqref{qfj} with the desired conditions.\footnote{Indeed, suppose $G$ is in the mixed-norm space $L^r_{{\rm d}y}(L^s_{{\rm d}x})$ and we want to write $G(x,y) = H(x,y)M(y)$ where $\|M\|_r = \|G\|_{L^r(L^s)}$ and where $\|H(\cdot, y)\|_s = 1$ for all $y$. Integrating $G(x,y)^s = H(x,y)^sM(y)^s$ with respect to $x$ shows that the only way to do this is to take $M(y) = \|G(\cdot, y)\|_s$ and $H(x,y) = G(x,y)/ \|G(\cdot, y)\|_s$. See the remarks at the end of Section~\ref{Loomis--Whitneyrevisited}.} \medskip \noindent Therefore, $$ G(x,y) \leq \prod_{j=1}^d [H_{jy}(x) M_j(y)]^{\alpha_j} := \prod_{j=1}^d G_j(x,y)^{\alpha_j}$$ where for all $y \in U^\perp$, for all $\phi \in L^1(B_jU)$ of norm at most $1$, $$ \int_U \phi(\tilde{B}_j x) H_{jy}(x) {\rm d}x \leq K_1$$ and where for all $\psi \in L^1((B_jU)^\perp)$ of norm at most $1$, $$ \int_{U^\perp} \psi(\tilde{\tilde{B}}_j y) M_j(y) {\rm d}y \leq K_2.$$ \medskip \noindent We want to show that for all $f \in L^1(\mathbb{R}^{n_j})$ of norm at most $1$, $$ \int_{\mathbb{R}^n} f({B}_j z) G_j(z) {\rm d}z = \int_{U^\perp} \int_U f({B}_j(x,y))H_{jy}(x) M_j(y) {\rm d}x {\rm d}y \leq K_1 K_2.$$ \medskip \noindent Fix $y \in U^\perp$ and write the inner integral over $U$ as $$\int_U f({B}_j(x,y))H_{jy}(x) {\rm d}x = \int_U f({B}_jx + B_jy) H_{jy}(x) {\rm d}x.$$ Now $f({B}_jx + B_jy) = f((\tilde{B}_j x +\Gamma_j y)+ \tilde{\tilde{B}}_jy)$. For $w \in B_jU$ and $\xi \in (B_jU)^\perp$ let $\phi_\xi(w) := f( w + \xi)$. Therefore $f({B}_jx + B_jy) = \phi_{\tilde{\tilde{B}}_j y}(\tilde{B}_j x + \Gamma_j y) =( \tau_{(\Gamma_j y)} \phi_{\tilde{\tilde{B}}_j y})(\tilde{B}_j x)$, where $(\tau_\eta \chi)( \cdot) = \chi( \cdot + \eta)$ denotes translation by $\eta$. So, $$\int_U f({B}_j(x,y))H_{jy}(x) {\rm d}x = \int_U ( \tau_{(\Gamma_j y)} \phi_{\tilde{\tilde{B}}_j y})(\tilde{B}_j x) H_{jy}(x) {\rm d}x \leq K_1 \| \tau_{(\Gamma_j y)} \phi_{\tilde{\tilde{B}}_j y}\|_1$$ by what we are assuming. \medskip \noindent Now, by translation invariance, $\| \tau_{(\Gamma_j y)} \phi_{\tilde{\tilde{B}}_j y}\|_1 = \|\phi_{\tilde{\tilde{B}}_j y}\|_1$. Therefore, letting $\psi(\xi) := \|\phi_{\xi}\|_1$ for $\xi \in (B_jU)^\perp$, $$ \int_{\mathbb{R}^n} f({B}_j z) G_j(z) {\rm d}z \leq K_1 \int_{U^\perp} \psi(\tilde{\tilde{B}}_j y) M_j(y){\rm d}y \leq K_1 K_2 \|\psi\|_1.$$ \medskip \noindent Finally, $$ \|\psi\|_1 = \int_{(B_jU)^\perp} \left(\int_{B_jU} \phi_\xi(w) {\rm d} w \right) {\rm d} \xi = \int_{\mathbb{R}^{n_j}} f(z) {\rm d} z = 1,$$ and this gives what we wanted. \section{Multilinear Kakeya inequalities revisited}\label{MKrevisited} Recall that we have families $\mathcal{P}_j$ of $1$-tubes in $\mathbb{R}^n$, and for $P \in \mathcal{P}_j$, its direction $e(P) \in \mathbb{S}^{n-1}$ satisfies $|e(P) - e_j| \leq c_n$ where $c_n$ is a small dimensional constant. The multilinear Kakeya theorem of Guth \cite{MR2746348} (see also \cite{CV}) states that $$ \Big\|\prod_{j=1}^n \left(\sum_{P _j \in \mathcal{P}_j} a_{P_j} \chi_{P_j}(x) \right)^{1/n} \Big\|_{L^{n/(n-1)}(\mathbb{R}^n)} \leq C_n \prod_{j=1}^n \left(\sum_{P_j \in \mathcal{P}_j} a_{P_j}\right)^{1/n}.$$ This inequality is of the form \eqref{submainineq} with $X = \mathbb{R}^n$, $q = n/(n-1)$, $Y_j = \mathcal{P}_j$ with counting measure, $p_j = 1$ for all $j$, $\alpha_j = 1/n$ for all $j$, and $T ((a_{P_j})(x) = \sum_{P _j \in \mathcal{P}_j} a_{P_j} \chi_{P_j}(x)$. \medskip \noindent Guth proved this result essentially by establishing a suitable subfactorisation for each nonnegative $M \in L^n(\mathbb{R}^n)$, and then applying Proposition~\ref{thmguthbaby}. His subfactorisation is described in terms of an auxiliary polynomial $p$ of `low' degree dominated by $\|M\|_n$, whose zero set $Z_p$ has `large' {\bf visibility} on each unit cube $Q$ of $\mathbb{R}^n$ in the sense that ${\rm vis}(Z_p \cap Q) \gtrsim \int_Q M$. We do not enter into the details of the definition of visibility, nor into how this gives the desired subfactorisation, but instead refer the reader to \cite{MR2746348} and \cite{CV}. (In the latter paper the approach using Proposition~\ref{thmguthbaby} is explicit while in the former it is implicit. And one should note that the definition of visibility used in \cite{CV} is a power of the original one used in \cite{MR2746348}.) It was the shock of seeing such an unlikely functional-analytic method succeed which inspired us to study the general question of necessity of factorisation as taken up in this paper. In hindsight, our linkage of subfactorisation of functions with Maurey's theory of factorisation of operators helps place Guth's method in perspective. \medskip \noindent Bourgain and Guth in \cite{BG} established an affine invariant form of the multilinear Kakeya inequality, removing the hypothesis that $|e(P) - e_j| \leq c_n$ for $P \in \mathcal{P}_j$, at the price of inserting a damping factor on the left-hand side which is consistent with the affine-invariant Loomis--Whitney inequality of Section~ \ref{Loomis--Whitneyrevisited}. That is, they proved that for $\mathcal{P}_j$ arbitrary families of $1$-tubes, $$ \int_{\mathbb{R}^n}\left(\sum_{P_1 \in \mathcal{P}_1} a_{P_1} \chi_{P_1}(x) \dots \sum_{P_n \in \mathcal{P}_n} a_{P_n} \chi_{P_n}(x) e(P_1) \wedge \dots \wedge e(P_n) \right)^{1/(n-1)} {\rm d}x $$ $$ \leq C_n \prod_{j=1}^n \left(\sum_{P_j \in \mathcal{P}_j} a_{P_j}\right)^{1/(n-1)}.$$ \medskip \noindent As the reader will readily verify (using the same argument as in the proof of Proposition~\ref{thmguthbaby}, see also Section~\ref{culture} above), in order to establish this, it suffices to show that for every nonnegative $M \in L^n(\mathbb{R}^n)$ which is constant on unit cubes in a standard lattice $\mathcal{Q}$, there exist nonnegative functions $S_j : \mathcal{Q} \times \mathcal{P}_j \to \mathbb{R}$ such that $$ M(Q) \lesssim \frac{S_1(Q, P_1)^{1/n} \cdots S_n(Q, P_n)^{1/n}}{ e(P_1) \wedge \dots \wedge e(P_n)^{1/n}}$$ whenever the $1$-tubes $P_j$ meet at $Q$, and, for all $j$, for all $P_j \in \mathcal{P}_j$, $$ \sum_{Q \in \mathcal{Q}, \, Q \cap P_j \neq \emptyset} S_j(Q, P_j) \lesssim \left(\sum_{Q \in \mathcal{Q}} M(Q)^n\right)^{1/n}.$$ And indeed this is what Bourgain and Guth essentially did (see also \cite{CV}). It is therefore very tempting to ask whether, in analogy with the situation of Theorem~\ref{thmmainbaby}, this method is {\em guaranteed} to work in so far as the statement of the affine-invariant multilinear Kakeya inequality automatically implies the existence of a subfactorisation as in the last two displayed inequalities. Unfortunately, as we have established above in Section~\ref{fdmgs}, there is no such general functional-analytic principle which guarantees this. \medskip \noindent The recent Multilinear Kakeya $k_j$-plane inequalities, and indeed the even more general perturbed Brascamp--Lieb inequalities, both recently established by Zhang \cite{Z}, also fit into the framework we consider, the latter as a generalisation of inequality \eqref{BL3}. \subsection{The finite field multilinear Kakeya inequality}\label{ffmk} Zhang \cite{Zh2} has recently solved the discrete analogue of the multilinear Kakeya problem. Let $\mathbb{F}$ be a field and let $\mathcal{L}_j$ be arbitrary families of lines in $\mathbb{F}^n$. For $l_j \in \mathcal{L}_j$ declare $e(l_1) \wedge \dots \wedge e(l_n)$ to be $1$ if the vectors $\{e(l_j)\}$ are linearly independent and to be $0$ otherwise. Zhang has proved that for a certain $C_n$ depending only on $n$, \begin{equation} \label{ffkakeya} \begin{aligned} \sum_{ x \in \mathbb{F}^n}&\left(\sum_{l_1 \in \mathcal{L}_1} a_{l_1} \chi_{l_1}(x) \dots \sum_{l_n \in \mathcal{L}_n} a_{l_n} \chi_{l_n}(x) e(l_1) \wedge \dots \wedge e(l_n) \right)^{1/(n-1)} \\ & \leq C_n \prod_{j=1}^n \left(\sum_{l_j \in \mathcal{L}_j} a_{l_j}\right)^{1/(n-1)}. \end{aligned} \end{equation} \medskip \noindent When $n=2$ the constant $C_2 = 1$, as is readily verified using $1/(n-1) = 1$ and changing the order of summation on the left-hand side. Moreover, for general $n$, if all the lines in $\mathcal{L}_j$ are parallel to some fixed vector $y_j$ with $\{y_j\}_{j=1}^n$ linearly independent, the constant is likewise $1$, since matters can then be reduced to the classical Loomis--Whitney inequality via an invertible linear transformation of $\mathbb{F}^n$, (or one can write down a suitable factorisation as in Example~\ref{Loomis--Whitneyrevisited}). \medskip \noindent The presence of the factor $e(l_1) \wedge \dots \wedge e(l_n)$ in \eqref{ffkakeya} precludes any assertion that \eqref{ffkakeya} is equivalent to a factorisation statement: see Section~\ref{culture} above. If however $\mathcal{L}_j$ are presumed to satisfy the property that if $(l_1, \dots , l_n) \in \mathcal{L}_1 \times \dots \times \mathcal{L}_n$, then the directions $\{e(l_1), \dots , e(l_n)\}$ are linearly independent, we have that the term $e(l_1) \wedge \dots \wedge e(l_n)$ is identically $1$, and the result then falls under the scope of Theorem~\ref{thmmain}. \medskip \noindent In particular, when $n = 2$ and we have two finite families of lines $\mathcal{L}_1$ and $\mathcal{L}_2$ in $\mathbb{F}^2$ such that no line in $\mathcal{L}_1$ is parallel to any line in $\mathcal{L}_2$, this holds. Let $J \subseteq \mathbb{F}^2$ be the set of points where some $l_1 \in \mathcal{L}_1$ meets an $l_2 \in \mathcal{L}_2$. Suppose $\sum_{x \in J} G(x)^2 = 1$. Then there exist $g_1, g_2 : J \to \mathbb{R}_+$ such that for all $x \in J$ $$ G(x) = \sqrt{g_1(x) g_2(x)}$$ and also, for all $l_j \in \mathcal{L}_j$, $j=1,2$, $$ \sum_{x \in J \cap l_j} g_j(x) \leq 1.$$ In spite of the extreme simplicity of the original problem, an explicit such factorisation is not known. \medskip \noindent There had been some speculation that even in higher dimensions, the constant $C_n$ might still be $1$. However, we shall show that this turns out to have been over-optimistic, since if we take $n=3$, by considering the field $\mathbb{F}_3$ with three elements,\footnote{In the case of $\mathbb{F}_2^3$ the constant is $1$.} we can see that $C_3 \geq 1.04$. \medskip In this example, for each $j=1,2,3$, we nominate two directions, and the family $\mathcal{L}_j$ will consist of all lines with one of these directions. The two directions for each $j$ will be chosen so that each of the eight choices of one direction from each of the three families results in a linearly independent set of directions, so that the terms $e(l_1) \wedge e(l_2) \wedge e(l_3)$ are all $1$. Each family of coefficients $a$ -- as a function defined on $\mathcal{L}_j$ and more properly denoted by $a_j$ -- is defined to be supported on three lines from $\mathcal{L}_j$ in such a way that the $x$-summand on the left-hand side of \eqref{ffkakeya} is non-zero at five points. Each $a$ will take nonzero values in $\{1,2\}$ and thus each line under consideration will have a {\em weight} equal to $1$ or $2$. For each $j$ we shall have that two of the three lines pass through two of these five points and the remaining line passes through the remaining point. \medskip More concretely, let \begin{itemize} \item $\mathcal{L}_1$ be the lines with direction $(1,1,0)$ or $(2,1,1)$; \item $\mathcal{L}_2$ be the lines with direction $(0,1,0)$ or $(0,1,1)$; and \item $\mathcal{L}_3$ be the lines with direction $(1,0,1)$ or $(0,0,1)$. \end{itemize} It is straightforward to verify that the directions of any three lines, one from each collection, span $\mathbb{F}_3^3$. \medskip \noindent We now proceed to properly define the coefficients $a$. We denote by $a_j$ the function whose domain is $\mathcal{L}_j$, and which is defined as follows: \begin{itemize} \item Let $a_1$ be \begin{itemize} \item $2$ on the line with direction $(1,1,0)$ passing through $(0,2,2)$ and $(2,1,2)$, \item $2$ on the line with direction $(2,1,1)$ passing through $(0,2,1)$ and $(2,0,2)$, \item $1$ on the line with direction $(1,1,0)$ through $(0,0,0)$, and \item $0$ on other lines of $\mathcal{L}_1$. \end{itemize} \item Let $a_2$ be \begin{itemize} \item $2$ on the line with direction $(0,1,0)$ passing through $(2,0,2)$ and $(2,1,2)$, \item $2$ on the line with direction $(0,1,1)$ passing through $(0,0,0)$ and $(0,2,2)$, \item $1$ on the line with direction $(0,1,0)$ through $(0,2,1)$, and \item $0$ on other lines of $\mathcal{L}_2$. \end{itemize} \item Let $a_3$ be \begin{itemize} \item $2$ on the line with direction $(0,0,1)$ passing through $(0,2,1)$ and $(0,2,2)$, \item $2$ on the line with direction $(1,0,1)$ passing through $(0,0,0)$ and $(2,0,2)$, \item $1$ on the line with direction $(0,0,1)$ through $(2,1,2)$, and \item $0$ for other lines of of $\mathcal{L}_3$. \end{itemize} \end{itemize} \noindent Each $\mathcal{L}_j$ has two lines of $a$-value or weight $2$ and one of weight $1$. \medskip We can see that the only points where lines from all three families intersect are the five points mentioned, namely $(0,0,0)$, $(0,2,1)$, $(0,2,2)$, $(2,0,2)$ and $(2,1,2)$. At the three points $(0,0,0)$, $(0,2,1)$ and $(2,1,2)$ we have two lines of weight $2$ and one of weight $1$ meeting; at the two points $(0,2,2)$ and $(2,0,2)$ we have three lines of weight $2$ meeting. So the value of the $x$-summand on the left-hand side of \eqref{ffkakeya} is $2$ at the three points $(0,0,0)$, $(0,2,1)$ and $(2,1,2)$, and is $2^{3/2}$ at the two points $(0,2,2)$ and $(2,0,2)$. The left-hand side adds up to $3\cdot2+2\cdot2^{3/2} > 11.65$. The value of the right-hand side of \eqref{ffkakeya} is $C_3 \cdot 5^{3/2}$ $\leq C_3 \cdot 11.19$. This shows that $C_3 \geq \frac{6 + 2^{3/2}}{5^{3/2}} > 11.65/11.19 > 1.04 > 1$. \medskip A counterexample to the conjecture that \eqref{ffkakeya} holds with $C_n=1$ was first found by use of the duality theory developed above, which is valid under the assumption that any $n$-tuple of lines taken from $\mathcal{L}_1 \times \dots \times \mathcal{L}_n$ has linearly independent directions. To explain why this route was taken, let us assume that we are considering a finite field of size $q$. If we let $\mathcal{L}_j$ consist of all lines with directions in some given set of size $r$ then the input to \eqref{ffkakeya}, namely the tuple $(a_1,\dots,a_n)$ belongs to a real vector space of dimension $nq^{n-1}r$. The input to problem \eqref{mainprob} is the function $G$ which belongs to a real vector space of dimension $q^{n}$. In our case we have $n=q=3$ and $r=2$ so the input to the problem \eqref{mainprob} belongs to a smaller vector space than the input to \eqref{ffkakeya}. The additional cost of solving the comvex optimisation problem compared with the cost of simply evaluating each side of \eqref{ffkakeya} does not significantly alter the balance of cost. \medskip The solution to the convex optimisation problem was found using the software package CVXOPT \cite{CVXOPT}, which yields the solution for both the primal and dual problems. The solution to the dual problem was then slightly simplified by hand for neater exposition and this is what is presented here. \bibliographystyle{plain}

8e086ac596908037b76f154e530518928fe7d4a7

\section{Introduction} \label{intro_section} In this paper we shall consider a vectorial, asymptotically free gauge theory (in four spacetime dimensions, at zero temperature) with gauge group $G$ with massless fermions transforming according to multiple different representations of $G$, which has an exact infrared (IR) fixed point (IRFP) of the renormalization group \cite{fm}. For technical simplicity, we will restrict ourselves to two different representations. We thus take the theory to contain $N_f$ copies (flavors) of Dirac fermions, denoted $f$, in the representation $R$ of $G$, and $N_{f'}$ copies of fermions, denoted $f'$, in a different representation $R'$ of $G$. In the case in which $f'$ transforms according to a self-conjugate representation, the number $N_{f'}$ refers equivalently to a theory with $N_{f'}$ Dirac fermions or $2N_{f'}$ Majorana fermions and hence in this case $N_{f'}$ may take on half-integral as well as integral values. One motivation for such theories is a possible direction for ultraviolet completions of the Standard Model (e.g., \cite{rs09,bv}). In \cite{bv} we studied the infrared evolution and phase structure of this type of theory. Here we go beyond Refs. \cite{rs09,bv} in presenting (scheme-independent) calculations of anomalous dimensions of gauge-invariant operators. We denote the running gauge coupling as $g=g(\mu)$, where $\mu$ is the Euclidean energy/momentum scale at which this coupling is measured. We define $\alpha(\mu) = g(\mu)^2/(4\pi)$. Since the theory is asymptotically free, its properties can be computed reliably in the deep ultraviolet (UV) region at large $\mu$, where the coupling approaches zero. The dependence of $\alpha(\mu)$ on $\mu$ is described by the renormalization-group (RG) beta function, $\beta = d\alpha(\mu)/dt$, where $dt=d\ln\mu$ (the argument $\mu$ will often be suppressed in the notation). We will consider a theory in which the fermion content is such that the RG flow from the UV to the IR ends in an exact IR fixed point, as determined by the zero in the beta function nearest to the origin for physical coupling, denoted $\alpha_{IR}$. Since $\beta=0$ at $\alpha=\alpha_{IR}$, the resultant theory in this IR limit is scale-invariant, and is deduced also to be conformally invariant \cite{scalecon}. The properties of the resultant conformal field theory at this IRFP are of considerable importance. Physical quantities defined at the IRFP obviously cannot depend on the scheme used for the regularization and renormalization of the theory. In conventional computations of these quantities, one first writes them as series expansions in powers of the coupling, and then evaluates these series expansions with $\alpha$ set equal to $\alpha_{IR}$, calculated to a given loop order. These calculations have been performed for anomalous dimensions of gauge-invariant fermion bilinears in a theory with a single fermion representation up to four-loop level \cite{bvh}-\cite{bc} and to five-loop level \cite{flir}. However, as is well known, these conventional (finite-order) series expansions are scheme-dependent beyond the leading terms. Indeed, this is a generic property of higher-order calculations in quantum field theory, such as computations in quantum chromodynamics (QCD) used to compare with data from the Fermilab Tevatron and CERN Large Hadron Collider (LHC). There is thus strong motivation to calculate and analyze series expansions for physical properties at the IRFP which are scheme-independent at each finite order. The fact that makes this possible is simple but powerful. To review this, we first specialize to a theory with $N_f$ fermions in a single representation, $R$, of the gauge group $G$. The constraint of asymptotic freedom means that $N_f$ must be less than a certain upper ($u$) bound, denoted $N_{f,u}$. Here and below, we will often formally generalize the number(s) of fermions in one or multiple representations from non-negative integers to non-negative real numbers, with the understanding that for a physical quantity one restricts to integral values. Furthermore, as noted above, if an $f'$ fermion transforms according to a self-conjugate representation, then the number $N_{f'}$ refers equivalently to a theory with $N_{f'}$ Dirac fermions or $2N_{f'}$ Majorana fermions, so that in this case, $N_{f'}$ may take on half-integral physical values. As $N_f$ approaches $N_{f,u}$ from below, the value of the IRFP, $\alpha_{IR}$, approaches zero. This means that one can re-express series expansions for physical quantities at this IRFP in powers of the manifestly scheme-independent variable \cite{bz,gkgg} \beq \Delta_f = N_{f,u}-N_f \ . \label{deltaf} \eeq In recent work, for theories with $N_f$ fermions in a single representation of the gauge group $G$, we have calculated scheme-independent series expansions for the anomalous dimensions of gauge-invariant fermion bilinears and the derivative $d\beta/d\alpha$, both evaluated at the IRFP, to the respective orders $O(\Delta_f^4)$ and $O(\Delta_f^5)$ \cite{gtr}-\cite{bpcgt}. These are the highest orders to which these quantities have been calculated. We gave explicit expressions for the case $G={\rm SU}(N_c)$ and $R$ equal to the fundamental, adjoint, and rank-2 symmetric and antisymmetric tensor representations, and for other Lie groups, including orthogonal, symplectic, and exceptional groups. In this paper we shall generalize our previous scheme-independent series calculations of physical quantities at an IRFP from the case of an asymptotically free gauge theory with $N_f$ fermions in a single representation of the gauge group $G$ to the case of fermions in multiple different representations. Specifically, we consider a theory with $N_f$ fermions in a representation $R$ of $G$ and $N_{f'}$ fermions in a different representation, $R'$, of $G$. We present scheme-independent calculations of the anomalous dimensions of gauge-invariant fermion bilinear operators to cubic order in the respective expansion variable ($\Delta_f$ in Eq. (\ref{deltaf}) for $\bar f f $ and $\Delta_{f'}$ in Eq. (\ref{deltaf2}) for $\bar f' f'$) and to quartic order in $\Delta_f$ and $\Delta_{f'}$ for the derivative of the beta function, evaluated at the infrared fixed point. The condition of asymptotic freedom requires that the value of a certain linear combination of $N_{f'}$ and $N_f$ must be less than an upper bound given below by Eq. (\ref{asymptotic_freedom}). For a fixed $N_{f'}$, this implies an upper bound denoted as $N_f < N_{f,u}$, and for a fixed $N_f$, this implies the upper bound $N_{f'} < N_{f',u}$ given respectively in Eqs. (\ref{Nfu}) and (\ref{Nf2u}) below. For fixed $N_{f'}$, as $N_f$ approaches $N_{f,u}$ from below, $\alpha_{IR}$ approaches zero. Therefore, one can rewrite the series expansions for physical quantities as power series in the variable $\Delta_f$. The coefficients in these series expansions depend on $N_{f'}$. If $\Delta_f$ is small, the value of $\alpha_{IR}$ is also small, so that the resultant IR theory may be inferred to be in a (deconfined) non-Abelian Coulomb phase (NACP), often called the conformal window. Strong evidence for this in the single-representation case comes from fully nonperturbative lattice simulations \cite{lgtreviews,simons,lnncomment}. In the same way, for fixed $N_f$, one can rewrite the series expansions for physical quantities as power series in the variable \beq \Delta_{f'} = N_{f',u} - N_{f'} \ . \label{deltaf2} \eeq For a general operator ${\cal O}$, we denote the full scaling dimension as $D_{\cal O}$ and its free-field value as $D_{{\cal O},free}$. The anomalous dimension of this operator, denoted $\gamma_{\cal O}$, is defined via the relation \cite{gammaconvention} \beq D_{\cal O} = D_{{\cal O},free} - \gamma_{\cal O} \ . \label{anomdim} \eeq Let us denote the fermions of type $f$ as $\psi_i$, $i=1,...,N_f$ and the fermions of type $f'$ as $\chi_j$, $j=1,...,N_{f'}$. We shall calculate scheme-independent series expansions for the anomalous dimensions, denoted $\gamma_{\bar\psi\psi,IR}$ and $\gamma_{\bar\chi\chi,IR}$ of the respective (gauge-invariant) fermion bilinears \beq \bar\psi\psi = \sum_{j=1}^{N_f} \bar\psi_j \psi_j \label{psibarpsi} \eeq and \beq \bar\chi\chi = \sum_{j=1}^{N_{f'}} \bar\chi_j \chi_j \ . \label{chibarchi} \eeq The anomalous dimension of $\bar\psi\psi$ is the same as that of the (gauge-invariant) bilinear $\sum_{j,k=1}^{N_f} \bar\psi_j {\cal T}_a \psi_k$, where ${\cal T}_a$ is a generator of the Lie algebra of SU($N_f$) \cite{gracey_gammatensor}, and we shall use the symbol $\gamma_{\bar\psi\psi,IR}$ to refer to both. An analogous comment applies to $\gamma_{\bar\chi\chi,IR}$. We write the scheme-independent series expansions of $\gamma_{\bar f f,IR}$ as \beq \gamma_{\bar f f ,IR} = \sum_{j=1}^\infty \kappa^{(f)}_j \, \Delta_f^j \label{gamma_ir_Deltaseries} \eeq and \beq \gamma_{\bar f' f' ,IR} = \sum_{j=1}^\infty \kappa^{(f')}_j \, \Delta_{f'}^j \label{gamma_ir_Delta2series} \eeq We shall illustrate our general results in an SU($N_c$) gauge theory with $N_F$ fermions of type $f$ in the fundamental ($F$) representation and $N_{Adj}$ fermions of type $f'$ in the adjoint ($Adj$) representation. For this theory we will also use an explicit notation with coefficients $\kappa^{(f)} = \kappa^{(F)}$ and $\kappa^{(f')}=\kappa^{(Adj)}$. We shall calculate two equivalent scheme-independent series expansions of the derivative $\beta'_{IR}$. With $N_{f'}$ fixed, and $N_f$ variable, one may write the series as an expansion in powers of $\Delta_f$: \beq \beta'_{IR} = \sum_{j=2}^\infty d_j \, \Delta_f^j \ . \label{betaprime_ir_Deltaseries} \eeq Alternately, one may take $N_f$ to be fixed and write $\beta'_{IR}$ as a series expansion in powers of $\Delta_{f'}$, as \beq \beta'_{IR} = \sum_{j=2}^\infty \tilde d_j \, \Delta_{f'}^j \ . \label{betaprime_ir_Deltaseries2} \eeq Note that $d_1 = \tilde d_1 = 0$ for all $G$ and fermion representations. This paper is organized as follows. In Section \ref{methods_section} we discuss the methodology for our calculations. In Sections \ref{kappa_section} and \ref{betaprime_section} we present our new results for scheme-independent expansions of the anomalous dimensions of fermion bilinears and $d\beta/d\alpha$, both evaluated at the infrared fixed point. We discuss the special cases of the anomalous dimension and $\beta'_{IR}$ results for an illustrative theory with gauge group SU($N_c$) containing fermions in the fundamental and adjoint representations in Sections \ref{kappa_fadj_section} and \ref{betaprime_fadj_section}, respectively. Our conclusions are given in Section \ref{conclusion_section}, and some relevant group-theoretic results are reviewed in Appendix \ref{groupinvariants}. \section{Calculational Methods} \label{methods_section} \subsection{Beta Function and Series Expansions for Physical Quantities} In this section we discuss some background and the calculational methods that are relevant for our present work. The series expansion of $\beta$ in powers of the squared gauge coupling is \beq \beta = -2\alpha \sum_{\ell=1}^\infty b_\ell \, a^\ell \ , \label{beta} \eeq where $a=g^2/(16\pi^2) = \alpha/(4\pi)$ and $b_\ell$ is the $\ell$-loop coefficient. With an overall minus sign extracted, as in Eq. (\ref{beta}), the condition of asymptotic freedom is that $b_1 > 0$. The one-loop coefficient, $b_1$, is independent of the scheme used for regularization and renormalization. Mass-independent schemes include minimal subtraction \cite{ms} and modified minimal subtraction, denoted $\overline{\rm MS}$ \cite{msbar}. For mass-independent schemes, the two-loop coefficient, $b_2$, is also independent of the specific scheme used \cite{gross75}. For a theory with a general gauge group $G$ and $N_f$ fermions in a single representation, $R$, the coefficients $b_1$ and $b_2$ were calculated in \cite{b1} and \cite{b2}, while $b_3$, $b_4$, and $b_5$ were calculated in the commonly used $\overline{\rm MS}$ scheme in \cite{b3}, \cite{b4}, and \cite{b5}, respectively (see also \cite{b5su3}). For the analysis of a theory with fermions in multiple different representations, one needs generalizations of these results. These are straightforward to derive in the case of $b_1$ and $b_2$, but new calculations are required for higher-loop coefficients. These have recently been performed in \cite{zoller} (again in the $\overline{\rm MS}$ scheme) up to four-loop order, and we use the results of Ref. \cite{zoller} here. The expansion of the anomalous dimension of the fermion bilinear $\gamma_{\bar\psi\psi}$ in powers of the squared gauge coupling is \beq \gamma_{\bar\psi\psi} = \sum_{\ell=1}^\infty c^{(f)}_\ell a^\ell \ , \label{gamma} \eeq where $c^{(f)}_\ell$ is the $\ell$-loop coefficient. The analogous expansion applies for $\gamma_{\bar\chi\chi}$ with the replacement $c^{(f)}_\ell \to c^{(f')}_\ell$. The one-loop coefficient $c^{(f)}_1$ is scheme-independent, while the $c^{(f)}_\ell$ with $\ell \ge 2$ are scheme-dependent, and similarly with the $c^{(f')}_\ell$. For a general gauge group $G$ and $N_f$ fermions in a single representation $R$ of $G$, the $c^{(f)}_\ell$ have been calculated up to loop order $\ell=4$ in \cite{c4} and $\ell=5$ in \cite{c5}. For the case of multiple fermion representations, the anomalous dimension coefficients for the fermion bilinears have been calculated up to four-loop order in \cite{chetzol}. We use the results of \cite{chetzol} up to three-loop order here. Concerning scheme-independent series expansions, the calculation of the coefficient $\kappa^{(f)}_j$ in Eq. (\ref{gamma_ir_Deltaseries}) requires, as inputs, the values of the $b_\ell$ for $1 \le \ell \le j+1$ and the $c^{(f)}_\ell$ for $1 \le \ell \le j$, and similarly for $\kappa^{(f')}_j$, with the replacement $c^{(f)}_\ell \to c^{(f')}_\ell$. The calculation of the coefficients $d_j$ and $\tilde d_j$ in Eqs. (\ref{betaprime_ir_Deltaseries}) and (\ref{betaprime_ir_Deltaseries2}) requires, as inputs, the values of the $b_\ell$ for $1 \le \ell \le j$. Thus, using the calculation of the beta function for multiple fermion representation to four-loop order in \cite{zoller}, together with the calculation of the anomalous dimensions of the fermion bilinears in \cite{chetzol} up to three-loop order, we can calculate $\gamma_{\bar\psi\psi,IR}$ to order $O(\Delta_f^3)$ and $\gamma_{\bar\chi\chi,IR}$ to $O(\Delta_{f'}^3)$ for the case of multiple fermion representations. (Note that we cannot make use of the four-loop calculation of the anomalous dimensions of fermion bilinears in \cite{chetzol} to compute $\gamma_{\bar\psi\psi,IR}$ to order $O(\Delta_f^4)$ and $\gamma_{\bar\chi\chi,IR}$ to $O(\Delta_{f'}^4)$, because this would require, as an input, the five-loop coefficient $b_5$ in the beta function for this case of multiple fermion representations, and, to our knowledge, this has not been calculated.) Similarly, using the four-loop beta function from \cite{zoller}, we can calculate the $d_j$ and $\tilde d_j$ for $\beta'_{IR}$ to order $j=4$. We denote the truncation of these series to maximal power $j=p$ as $\gamma_{\bar\psi\psi,IR,\Delta_f^p}$, $\gamma_{\bar\chi\chi,IR,\Delta_f^p}$, $\beta'_{IR,\Delta_f^p}$, and $\beta'_{IR,\Delta_{f'}^p}$, respectively. Although we use these coefficients as calculated in the $\overline{\rm MS}$ scheme below, we emphasize that our results are scheme-independent, so the specific scheme used for their calculation does not matter. An explicit illustration of this using several schemes is given in \cite{cgs}. We refer the reader to our previous work for detailed discussions of the procedure for calculating the coefficients $\kappa_j$ and $d_j$ in the case of a theory with $N_f$ fermions in a single representation of $G$. Our procedure for calculating scheme-independent series expansions requires that the IRFP be exact, and hence we restrict our consideration to the non-Abelian Coulomb phase, where this condition is satisfied. For sufficiently smaller values of $N_f$ and/or $N_{f'}$, there is spontaneous chiral symmetry breaking (S$\chi$SB), giving rise to dynamical masses for the $f$ and/or $f'$ fermions \cite{phasenote}. Most-attractive channel arguments suggest that as $N_f$ and/or $N_{f'}$ decrease(s) and $\alpha_{IR}$ increases, the fermion with the largest value of $C_f$ would be the first to form bilinear fermion condensates and hence obtain dynamical masses and be integrated out of the low-energy effective field theory (EFT). Assuming that this happens and, say, the $f'$ fermions condense out, then one would proceed to examine the resultant EFT with the remaining massless $f$ fermions to determine the further evolution of this theory into the infrared. The details of the construction of the EFT will not be relevant here, since we restrict our analysis to the (chirally symmetric) non-Abelian Coulomb phase. \subsection{Relevant Range of $(N_f,N_{f'})$} Since we require that the theory should be asymptotically free and since our scheme-independent calculational method requires an exact IR fixed point, which is satisfied in the non-Abelian Coulomb phase, a first step is to discuss the corresponding values of the pair $(N_f,N_{f'})$ that satisfy these conditions. We denote this set of values, or more generally, the region in the first quadrant of the ${\mathbb R}^2$ plane defined by the generalization of $(N_f,N_{f'})$ from non-negative integers (or half-integers in the case of a Majorana fermion in a self-conjugate representation) to non-negative real numbers, where the theory has an IRFP in the non-Abelian Coulomb phase as the region ${\cal R}_{NACP}$. We next discuss the boundaries of this region. For a specified gauge group $G$ and fermion representations $R$ and $R'$, the numbers $N_f$ and $N_{f'}$ are bounded above by the asymptotic freedom (AF) condition that $b_1 > 0$. This condition is expressed as the inequality on the linear combination \beq N_fT_f + N_{f'}T_{f'} < \frac{11C_A}{4} \ , \label{asymptotic_freedom} \eeq where $C_A$ and $T_f$ are group invariants defined in Appendix \ref{groupinvariants}. Thus, for fixed $N_{f'}$, the AF property implies that $N_f$ is bounded above as $N_f < N_{f,u}$, where \beq N_{f,u} = \frac{11C_A - 4N_{f'}T_{f'}}{4T_f} \ , \label{Nfu} \eeq and similarly, for fixed $N_f$, the AF condition implies that $N_{f'}$ is bounded above as $N_{f'} < N_{f',u}$, where \beq N_{f',u} = \frac{11C_A - 4N_fT_f}{4T_{f'}} \ . \label{Nf2u} \eeq The upper boundary of this asymptotically free region, which is also the upper boundary of the region ${\cal R}_{NACP}$, in $N_f$ and $N_{f'}$ is the locus of solutions to the condition $b_1=0$. This is a finite segment of the line $N_fT_f + N_{f'}T_{f'} = 11C_A/4$. We may picture the first quadrant in the ${\mathbb R}^2$ space defined by non-negative $(N_f,N_{f'})$ to be such that $N_f$ is the horizontal axis and $N_{f'}$ is the vertical axis. Then the line segment bounding the asymptotically free region is an oblique line segment running from the upper left to the lower right, with slope \beq \frac{\partial N_{f'}}{\partial N_f}\bigg |_{b_1=0} = -\frac{T_f}{T_{f'}} \ . \label{slope_b1zeroline} \eeq This line segment intersects the horizontal axis at the point $(N_f,N_{f'})=(11C_A/(4T_f),0)$ and the vertical axis at the point $(N_f,N_{f'})=(0,11C_A/(4T_{f'}))$. Without loss of generality, we take $f$ to be the (nonsinglet) fermion representation of smaller dimension. The respective scheme-independent expansions in powers of $\Delta_f$ and $\Delta_{f'}$ amount to moving into the interior of the non-Abelian Coulomb phase from the upper boundary line horizontally (moving leftward) and vertically (moving downward). In our earlier work on theories with $N_f$ fermions in a single fermion representation of the gauge group, we denoted the lower boundary of the NACP as $N_{f,cr}$. In that case, we assumed that $N_f$ was in the NACP interval $I_{NACP}: \ N_{f,cr} < N_f < N_{f,u}$. Here the generalization of this is the set of physical values of $N_f$ and $N_{f'}$ in the region ${\cal R}_{NACP}$. Even in the case of a single fermion representation, the value of $N_{f,cr}$ is not known precisely. This question of the value of $N_{f,cr}$ for various specific theories has been investigated in a number of lattice studies \cite{lgtreviews,simons}, which continue at present. As noted above, we have previously presented approximate analytic results relevant for this study in \cite{rs09,bv} Corresponding lattice studies could be carried out for theories with multiple different fermion representations to study properties of the respective theories. An example is a recent lattice study of an SU(4) gauge theory with $N_f=2$ Dirac fermions in the fundamental representation and $N_{f'}=2$ Dirac fermions in the (self-conjugate) antisymmetric rank-2 tensor representation \cite{su4lgt1,su4lgt2}, which finds that the (zero-temperature) theory is in the phase with chiral symmetry breaking for both types of fermions. Since our results are restricted to an exact infrared fixed point in the (conformally invariant) non-Abelian Coulomb phase, they are not directly applicable to this theory. For the present study, with the axes of the first-quadrant quarter plane in $(N_f,N_{f'}) \in {\mathbb R}^2$ as defined above, the upper boundary of the NACP is the line segment resulting from the $b_1=0$ condition. The analogue of the lower boundary of the NACP at $N_{f,cr}$ for the present study with two fermion representations is a line segment or nonlinear curve displaced in the direction to the lower left relative to the oblique $b_1=0$ line, so that the resultant NACP forms a region in which physical values of $N_f$ and $N_{f'}$ define possible IR theories. This lower boundary of the NACP intersects the horizontal axis at the point $(N_f,N_{f'})=(N_{f,cr},0)$ and intersects the vertical axis at the point $(N_f,N_{f'})=(0,N_{f',cr})$. Although this lower boundary of the NACP is not known, one can get a rough idea of where it lies by generalizing the analysis that we gave in our previous work for theories with a single fermion representation \cite{gsi,dex,dexl}. This analysis was based on the observation that the two-loop beta function has an IR zero if $N_f$ is sufficiently large that $b_2$ is negative (with $b_1 > 0$). In this case of a single fermion representation, for small $N_f$, $b_2$ is positive, and turns negative when $N_f$ exceeds a certain lower ($\ell$) value $N_{f,\ell} < N_{f,u}$ where $b_2=0$, namely \beq N_{f,\ell} = \frac{17C_A^2}{2T_f(5C_A+3C_f)} \quad ({\rm for} \ N_{f'}=0). \label{nell} \eeq Thus, in this single-representation case, if and only if $N_f$ lies in an interval that we have denoted previously as $I_{IRZ}$, the two-loop beta function has an IR zero (IRZ). This interval $I_{IRZ}$ is \beq I_{IRZ}: \quad N_{f,\ell} < N_f < N_{f,u} \quad ({\rm for} \ N_{f'}=0). \label{interval} \eeq Although $N_{f,\ell}$ is not, in general, equal to $N_{f,cr}$, it is moderately close to the latter in theories that have been studied. As an example, in the case of an SU($N_c$) gauge theory with $N_f$ fermions in the fundamental ($F$) representation, \beq {\rm SU}(N_c), \ R=F: \quad N_{f,\ell} = \frac{34N_c^3}{13N_c^2-3} \ . \label{nell_sun} \eeq In the intensively studied case $N_c=3$ theory, $N_\ell = 153/19 \simeq 8.05$. This is close to the estimates of $N_{f,cr}$ for this theory from our previous studies and from a number of lattice simulations \cite{gsi,dexl,lgtreviews,simons}. In our present asymptotically free theory with two fermion representations, the two-loop beta function has an IR zero if and only if $b_2 < 0$, which is the inequality \beqs &&N_fT_f(5C_A+3C_f) + N_{f'}T_{f'}(5C_A+3C_{f'}) > \frac{17C_A^2}{2} \ . \cr\cr && \label{b2_negative} \eeqs This IR zero of the two-loop ($2\ell$) beta function occurs at $\alpha=\alpha_{IR,2\ell}$, where \begin{widetext} \beq \alpha_{IR,2\ell}=-\frac{4\pi b_1}{b_2} =\frac{2\pi\Big [11C_A - 4(N_fT_f + N_{f'}T_{f'})\Big ] } {\Big [2N_fT_f(5C_A+3C_f) + 2N_{f'}T_{f'}(5C_A+3C_{f'}) - 17C_A^2\Big ]} \ . \label{alfir_2loop_ff2} \eeq \end{widetext} We thus define the two-dimensional region in the first quadrant of the ${\mathbb R}^2$ plane defined by non-negative real values of $(N_f,N_{f'})$ where the theory is asymptotically free and the two-loop beta function has an IR zero as the region ${\cal R}_{IRZ}$, given by the conditions (\ref{asymptotic_freedom}) and (\ref{b2_negative}). The upper boundary of ${\cal R}_{IRZ}$ is the same as the upper boundary of ${\cal R}_{NACP}$, while the lower boundary of ${\cal R}_{IRZ}$ can provide a rough guide to the lower boundary of ${\cal R}_{NACP}$ and has the advantage that it is exactly calculable. This lower boundary of the region ${\cal R}_{IRZ}$ is given by the solution of the condition that $b_2=0$ in the first quadrant of the ${\mathbb R}^2$ plane. This condition is obtained from Eq. (\ref{b2_negative}) by replacing the inequality by an equality. The corresponding line defining the lower boundary of ${\cal R}_{IRZ}$ has the slope \beq \frac{\partial N_{f'}}{\partial N_f}{}\bigg |_{b_2=0} = -\frac{T_f(5C_A+3C_f)}{T_{f'}(5C_A+3C_{f'})} \ . \label{slope_b2zeroline} \eeq This lower boundary of the region ${\cal R}_{IRZ}$ crosses the horizontal axis in the $(N_f,N_{f'})$ space at the point $(N_{f,\ell},0)$, where $N_{f,\ell}$ was given above in Eq. (\ref{nell}), and it crosses the vertical axis at the corresponding value $(0,N_{f',\ell})$, where \beq N_{f',\ell} = \frac{17C_A^2}{2T_{f'}(5C_A+3C_{f'})} \ . \label{nf2ell} \eeq As noted, the lower boundary of this ${\cal R}_{IRZ}$ region provides a rough guide to the actual lower boundary of the NACP region ${\cal R}_{NACP}$. The determination of the true lower boundary of ${\cal R}_{NACP}$ would require a fully nonperturbative analysis, e.g., via lattice simulations. Although our calculational methods require the IRFP to be exact and hence, strictly speaking, apply only in the non-Abelian Coulomb phase, they could also be useful for the investigation of quasi-conformal gauge theories. In turn, the latter have been of interest as possible ultraviolet completions of the Standard Model. Specifically, (a) if the transition from the lower part of the non-Abelian Coulomb phase to the quasi-conformal regime in the variables $(N_f,N_{f'})$ is continuous, and (b) if our series calculations are sufficiently accurate in this region, our results for $\gamma_{\bar\psi\psi,IR}$, $\gamma_{\bar\chi\chi,IR}$, and $\beta'_{IR}$ could provide approximate estimates for the values of these quantities in the quasi-conformal regime just below the lower boundary with the NACP. \subsection{Example with Fermions in the Fundamental and Adjoint Representations} As an illustrative example, we consider a theory with the gauge group SU($N_c$) that contains $N_f \equiv N_F$ fermions in the fundamental ($F$) representation and $N_{f'} \equiv N_{Adj}$ fermions in the adjoint representation, $Adj$. We denote this as the FA theory. Here the upper boundary of the NACP region ${\cal R}_{NACP}$, which is also the upper boundary of the region ${\cal R}_{IRZ}$, is given by the line \beq {\rm FA \ theory}: \quad N_F + 2N_c N_{Adj} = \frac{11N_c}{2} \ . \label{nacp_upper_fund_adj} \eeq Thus, $N_F < (11/2)N_c$ if $N_{Adj}=0$ and $N_{Adj} < 11/4 = 2.75$ if $N_f=0$. The lower boundary of ${\cal R}_{IRZ}$, which can provide an approximate estimate to the lower boundary of ${\cal R}_{NACP}$, is given by the line $b_2=0$, namely \beqs && {\rm FA \ theory}: \ \ (13N_c - 3N_c^{-1})N_F + 32N_c^2 N_{Adj} = 34N_c^2 \ . \cr\cr && \label{alfir_2loop_lower_fund_adj} \eeqs Thus, in ${\cal R}_{IRZ}$, it follows that $N_F > 34N_c^3/(13N_c^2-3)$ if $N_{Adj}=0$ and $N_{Adj} > 17/16 = 1.0625$ if $N_F=0$. In this FA theory, the line $b_1=0$ has slope \beq {\rm FA \ theory}: \quad \frac{\partial N_{Adj}}{\partial N_F}{}\Bigg |_{b_1=0} = -\frac{1}{2N_c} \ , \label{b1zslope_fadj} \eeq while the line $b_2=0$ has slope \beq {\rm FA \ theory}: \quad \frac{\partial N_{Adj}}{\partial N_F}{}\Bigg |_{b_2=0} = -\frac{(13N_c^2-3)}{32N_c^3} \ . \label{b2zslope_fadj} \eeq For example, in the FA theory with $N_c=3$, so $G={\rm SU}(3)$, these slopes (\ref{b1zslope_fadj}) and (\ref{b2zslope_fadj}) are $-1/6=-0.16667$ and $-19/144=-0.13194$, respectively, where the floating-point values are given to the indicated accuracy. The $b_1=0$ line crosses the horizontal and vertical axes at $(N_f,N_{f'})=(16.5,0)$ and $(0,2.75)$, respectively, while the $b_2=0$ line crosses the horizontal and vertical axes at $(N_f,N_{f'})=(8.0526,0)$ and $(0,1.0625)$, respectively. In Table \ref{fadj_su3_theories_table} we list the physical integral values of $N_F$ and integral and half-integral (Majorana) values of $N_{Adj}$ in the region ${\cal R}_{IRZ}$ in this SU(3) theory. Considering $(N_F,N_{Adj})$ as a point in the first quadrant of an ${\mathbb R}^2$ space, we list in the second column the distance $d_u$ of this point from the line $b_1=0$ that forms the upper boundary of the regions ${\cal R}_{IRZ}$ and ${\cal R}_{NACP}$, and in the third column the distance $d_\ell$ of this point from the line $b_2=0$ that forms the lower boundary of the region ${\cal R}_{IRZ}$. (By distance of a point $P$ from a line $L$, we mean the length of the line segment perpendicular to the line $L$ that passes through the point $P$.) Thus, Ttable \ref{fadj_su3_theories_table} provides a guide to the position of a theory with a given set of values of $(N_F,N_{Adj})$ in the region ${\cal R}_{IRZ}$. In general, theories with small values of $d_u$ are close to the upper boundary of the region ${\cal R}_{NACP}$ and have correspondingly small values of $\alpha_{IR}$. In order for our perturbative analysis to be self-consistent, it is necessary that $\alpha_{IR}$ should not be excessively large, and so one may require, say, that $\alpha_{IR,2\ell} < 1$. Our perturbative analysis is expected to be most accurate for the $(N_F,N_{Adj})$ FA theories with small $d_u$ and hence small $\alpha_{IR,2\ell}$ in the upper part of the NACP. We will discuss this illustrative two-representation FA theory further below. \section{Scheme-Independent Calculation of Anomalous Dimensions of Fermion Bilinear Operators} \label{kappa_section} In this section, for a theory with a general gauge group $G$ containing $N_f$ fermions in a representation $R$ and $N_{f'}$ fermions in a representation $R'$, we present our new calculations of the coefficients $\kappa^{(f)}_j$ and $\kappa^{(f')}_j$ in the scheme-independent expansions of the anomalous dimensions $\gamma_{\bar\psi\psi,IR}$ and $\gamma_{\bar\chi\chi,IR}$ in Eqs. (\ref{gamma_ir_Deltaseries}) and the analogue for $\gamma_{\bar\chi\chi,IR}$ with $1 \le j \le 3$. It will be useful to define a factor that occurs repeatedly in the denominators of various expressions, namely \beq {\cal D}_f = C_A(7C_A+11C_f) + 4N_{f'}T_{f'}(C_{f'}-C_f) \ . \label{dcal} \eeq In the previously studied theory with a single fermion representation, i.e., $N_{f'}=0$, this factor ${\cal D}$ reduces as \beq {\cal D}_f = C_AD \quad {\rm if} \ \ N_{f'}=0 \ , \label{dcad} \eeq where \beq D=7C_A+11C_f \ , \label{dfac} \eeq as defined in Eq. (2.13) of our earlier work \cite{dex,dexl}. For the first two coefficients we calculate \beq \kappa^{(f)}_1 = \frac{8C_fT_f}{{\cal D}_f} \label{kappa1} \eeq and \begin{widetext} \beq \kappa^{(f)}_2 = \frac{4C_fT_f^2}{3{\cal D}_f^3} \, \Bigg [ C_A(7C_A+4C_f)(5C_A+88C_f) + 2^4 N_{f'}T_{f'}(C_{f'}-C_f) \Big (10C_A+8C_f+C_{f'}\Big ) \Bigg ] \ . \label{kappa2} \eeq For the third coefficient, we write \beq \kappa^{(f)}_3 = \frac{4C_fT_f}{3^4 {\cal D}_f^5} \Bigg [ A^{(f)}_0 + A^{(f)}_1 N_{f'} + A^{(f)}_2 N_{f'}^2 + A^{(f)}_3 N_{f'}^3 \Bigg ] \ . \label{kappa3} \eeq It follows that the $A^{(f)}_0$ term is independent of $N_{f'}$ and hence, taking into account the difference in the prefactor, it is equal to $C_A$ times the terms in the square bracket of Eq. (6.7) in our earlier Ref. \cite{dex} or equivalently Eq. (3.4) in our Ref. \cite{dexl}. We have \beqs A^{(f)}_0 &=& C_A \Bigg [ 3C_AT_f^2 \bigg ( -18473C_A^4 + 144004 C_A^3C_f +650896C_A^2C_f^2 +356928C_AC_f^3+569184C_f^4 \bigg ) \cr\cr &+&2^7D\bigg (-20T_f^2\frac{d_A^{abcd}d_A^{abcd}}{d_A} +352C_AT_f\frac{d_f^{abcd}d_A^{abcd}}{d_A} -1331C_A^2\frac{d_f^{abcd}d_f^{abcd}}{d_A} \bigg ) \cr\cr &+& 33 \cdot 2^{10}D \zeta_3 \, \bigg (2T_f^2 \frac{d_A^{abcd}d_A^{abcd}}{d_A} -13 C_A T_f \frac{d_f^{abcd}d_A^{abcd}}{d_A} +11 C_A^2 \frac{d_f^{abcd}d_f^{abcd}}{d_A} \bigg ) \ \Bigg ] \ \ , \label{kappa3_A0} \eeqs where $\zeta_s = \sum_{n=1}^\infty n^{-s}$ is the Riemann zeta function. Here, the group invariants $C_A$, $C_f$, $T_f$, $d_A^{abcd}d_A^{abcd}$, $d_f^{abcd}d_A^{abcd}$, $d_f^{abcd}d_f^{abcd}$, and $d_f^{abcd}d_{f'}^{abcd}$ are defined in Appendix \ref{groupinvariants}, and $d_A$ is the dimension of the adjoint representation of $G$. For the other $A_s^{(f)}$ with $1 \le s \le 3$, we calculate \beqs A^{(f)}_1 &=& C_AT_f^2 T_{f'}\Bigg [ 273840C_A^3(C_{f'}-C_f) +C_A^2 \bigg ( -1511040C_f^2 + 1916256C_f C_{f'} -405216 C_{f'}^2 \bigg ) \cr\cr &&+C_A\bigg ( -129600C_f^3+522432C_f^2 C_{f'}-485568C_f C_{f'}^2 + 92736C_{f'}^3 \bigg ) \cr\cr &&+C_f \bigg (-1241856C_f^3 +1020096C_f^2C_{f'} + 76032C_f C_{f'}^2 +145728C_{f'}^3 \bigg ) \ \Bigg ] \cr\cr &&+10240 T_f^2 T_{f'}(C_f-C_{f'}) \frac{d_A^{abcd}d_A^{abdc}}{d_A} +C_AT_f T_{f'} \bigg ( -114688C_A -360448C_f +180224C_{f'} \bigg ) \frac{d_f^{abcd}d_A^{abcd}}{d_A} \cr\cr &&+C_AT_f^2 \Big ( 114688C_A + 180224 C_f \Big ) \frac{d_{f'}^{abcd}d_A^{abcd}}{d_A} +C_A^2 T_{f'}\bigg ( 867328 C_A + 2044416 C_f -681472 C_{f'} \bigg ) \frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &&+C_A^2T_f\Big (-867328C_A-1362944C_f\Big ) \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} \cr\cr &&+\zeta_3 \Bigg [ 270336 T_f^2 T_{f'}(C_{f'} - C_f) \frac{d_A^{abcd}d_A^{abcd}}{d_A} + C_AT_fT_{f'} \bigg ( 1118208 C_A + 3514368 C_f -1757184C_{f'} \bigg ) \frac{d_f^{abcd}d_A^{abcd}}{d_A} \cr\cr &&+C_A T_f^2 \Big ( -1118208C_A-1757184C_f \Big ) \frac{d_{f'}^{abcd}d_A^{abcd}}{d_A} +C_A^2T_{f'} \bigg (-1892352 C_A -4460544 C_f +1486848C_{f'} \bigg ) \frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &&+C_A^2T_f \Big ( 1892352C_A + 2973696C_f \Big ) \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} \Bigg ] \label{kappa3_A1} \eeqs \beqs A^{(f)}_2 &=& T_f^2 T_{f'}^2 \Bigg [ 350976 C_A^2(C_f-C_{f'})^2 + C_A \bigg (-94464 C_f^3 -2304C_f^2C_{f'} +288000C_f C_{f'}^2 - 191232C_{f'}^3 \bigg ) \cr\cr && +225792C_f^4 -370944C_f^3C_{f'} + 119808C_f^2C_{f'}^2 -29952C_f C_{f'}^3 + 55296C_{f'}^4 \Bigg ] \cr\cr &&+ 2^{16}T_fT_{f'}(C_f-C_{f'})( T_{f'} \frac{d_f^{abcd}d_A^{abcd}}{d_A} - T_f \frac{d_{f'}^{abcd}d_A^{abcd}}{d_A}) \cr\cr &&+C_A T_{f'}^2 \bigg ( -157696C_A+495616C_{f'}-743424C_f \bigg ) \frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &&+C_AT_fT_{f'}\bigg ( 315392C_A+991232C_f-495616C_{f'} \bigg ) \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} +C_A T_f^2 ( -157696C_A -247808C_f ) \frac{d_{f'}^{abcd}d_{f'}^{abcd}}{d_A} \cr\cr &&+ \zeta_3 \Bigg [ 638976T_f T_{f'}(C_f-C_{f'})\bigg ( T_f \frac{d_{f'}^{abcd}d_A^{abcd}}{d_A} -T_{f'} \frac{d_f^{abcd}d_A^{abcd}}{d_A} \bigg ) \cr\cr &&+C_AT_{f'}^2\bigg ( 344064C_A +1622016C_f - 1081344C_{f'} \bigg ) \frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &&C_AT_f T_{f'}\bigg ( -688128C_A -2162688C_f +1081344C_{f'} \bigg ) \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} + C_A T_f^2 \Big ( 344064C_A + 540672C_f \Big ) \frac{d_{f'}^{abcd}d_{f'}^{abcd}}{d_A} \ \Bigg ] \cr\cr && \eeqs and \beqs A^{(f)}_3 &=& 2^{13} \, T_{f'}(C_f-C_{f'})(11-24\zeta_3) \bigg ( T_{f'}^2 \frac{d_f^{abcd}d_f^{abcd}}{d_A} - 2 T_f T_{f'} \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} + T_f^2 \frac{d_{f'}^{abcd}d_{f'}^{abcd}}{d_A} \bigg ) \ . \cr\cr && \label{A3A3bar} \eeqs \end{widetext} The coefficients $\kappa^{(f')}_j$ are obtained from these $\kappa^{(f)}_j$ by interchanging $f$ and $f'$ in all expressions. For example, \beq {\cal D}_{f'} = C_A(7C_A+11C_{f'}) + 4N_fT_f(C_f-C_{f'}) \ , \label{dcalf2} \eeq \beq \kappa^{(f')}_1 = \frac{8C_{f'}T_{f'}}{{\cal D}_{f'}} \ , \label{kappa1f2} \eeq and so forth for the other expressions. An important result that we found in our previous work \cite{gsi}-\cite{dexo} was that for a theory with a single representation, $\kappa^{(f)}_1$ and $\kappa^{(f)}_2$ are manifestly positive, and for all of the specific gauge groups and fermion representations that we considered, $\kappa^{(f)}_3$ and $\kappa^{(f)}_4$ are also positive. This property implied several monotonicity relations for our calculation of $\gamma_{\bar\psi\psi}$ to maximal power $\Delta_f^p$, denoted $\gamma_{\bar\psi\psi,\Delta_f^p}$, namely that (i) for fixed $p$, $\gamma_{\bar\psi\psi,\Delta_f^p}$ is a monotonically increasing function of $\Delta_f$, i.e., a monotonically increasing function of decreasing $N_f$, and (ii) for fixed $N_f$, $\gamma_{\bar\psi\psi,\Delta_f^p}$ is a monotonically increasing function of the maximal power $p$. A basic question that we may ask concerning these results is how a coefficient $\kappa^{(f)}$ changes as one goes from the single-representation theory with ${N_f'}=0$ to theories with an increasing number $N_{f'}$ of fermions in a different representation, and vice versa for the dependence of $\kappa^{(f')}$ on $N_f$. For the purpose of this discussion, we recall that, by convention, we take $f$ to be the fermion in the representation with a smaller dimension. In the cases with which we deal, this also means that $C_f < C_{f'}$. The question is readily answered in the case of $\kappa_1^{(f)}$ and $\kappa_1^{(f')}$. As a lemma, we observe that ${\cal D}_f$ is a monotonically increasing function of $N_{f'}$, while ${\cal D}_{f'}$ is a monotonically decreasing function of $N_f$. Hence, $\kappa_1^{(f)}$ is a monotonically decreasing function of $N_{f'}$, while $\kappa_1^{(f')}$ is a monotonically increasing function of $N_f$. The dependence of $\kappa^{(f)}_j$ on $N_{f'}$ and of $\kappa^{(f')}_j$ on $N_f$ for indices $j=2,3$ will be analyzed below for particular theories. Concerning the question of the positivity of $\kappa^{(f)}_j$ and $\kappa^{(f')}_j$, in a theory with fermions in multiple different representations, there are terms of both signs in the expressions for the coefficients $\kappa^{(f)}_j$. Nevertheless, anticipating our results below, in the specific FA theories that we have studied in detail, both $\kappa^{(F)}_j$ and $\kappa^{(Adj)}_j$ are positive for all of the orders that we have calculated, namely $j=1,2,3$. In our earlier work \cite{gtr}-\cite{pgb} on scheme-independent series calculations for theories with a $N_f$ fermions transforming according to a single type of representation, we carried out detailed studies of the reliability of these expansions using a variety of methods. One of the simplest procedures is to analyze the fractional change in a quantity, calculated to a given order $O(\Delta_f^p)$, as one increases the maximal power $p$ of the expansion. Here we shall apply this method in our illustrative theory discussed in the next section. \section{Anomalous Dimensions in a Theory with Fermions in the Fundamental and Adjoint Representations of SU($N_c$)} \label{kappa_fadj_section} In this section we discuss our scheme-independent calculations of $\gamma_{\bar\psi\psi,IR}$ and $\gamma_{\bar\chi\chi,IR}$ for the illustrative case of a theory with gauge group SU($N_c$) containing $N_f \equiv N_F$ fermions in the fundamental representation and $N_{f'} \equiv N_{Adj}$ fermions in the adjoint representation. As before, we call this the FA theory. In this case, the denominator factor ${\cal D}_f$ takes the form \beq {\rm FA \ theory}: \quad {\cal D}_f = \frac{1}{2}\bigg [ 25N_c^2-11+ 4N_{Adj}(N_c^2+1) \bigg ] \ . \label{dcal_fa} \eeq We have given the values of $(N_F,N_{Adj})$ in Table \ref{fadj_su3_theories_table} for the region ${\cal R}_{IRZ}$. For the first-order coefficients we calculate \beq \kappa_1^{(F)} = \frac{4(N_c^2-1)} {N_c \Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \Big ]} \label{kappa1_fund_fa} \eeq and \beq \kappa_1^{(Adj)} = \frac{8N_c^3}{18N_c^3 - N_F(N_c^2+1)} \ . \label{kappa1_adj_fa} \eeq If $N_{Adj}=0$, then the coefficient $\kappa_1^{(F)}$ reduces to the expression $4(N_c^2-1)/[N_c(25N_c^2-11)]$, as given in Eq. (6.8) of our earlier work \cite{dex}. Similarly, if $N_F=0$, then $\kappa_1^{(Adj)}$ reduces to the value 4/9, as given in Eq. (6.18)) of \cite{dex}. For the second-order coefficients, we find \begin{widetext} \beq \kappa_2^{(F)} = \frac{4(N_c^2-1)\Big [ (9N_c^2-2)(49N_c^2-44) + 8N_{Adj}(N_c^2+1)(15N_c^2-4) \, \Big ]} {3N_c^2\Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \Big ]^3} \label{kappa2_fund_fa} \eeq and \beq \kappa_2^{(Adj)} = \frac{4N_c^4\Big [ 1023N_c^5 -2N_F(N_c^2+1)(37N_c^2-1) \, \Big ] } {3\Big [ 18N_c^3-N_F(N_c^2+1) \Big ]^3} \ . \label{kappa2_adj_fa} \eeq \end{widetext} If $N_{Adj}=0$, then $\kappa_2^{(F)}$ reduces to the expression given in Eq. (6.9) of \cite{dex}, and if $N_F=0$, then $\kappa_2^{(Adj)}$ reduces to the value $341/1458 = 341/(2 \cdot 3^6)$ as given in Eq. (6.19) of \cite{dex}. Our results for the third-order coefficients are as follows: \begin{widetext} \beq \kappa_3^{(F)} = \frac{2^3(N_c^2-1)\Big [ \kappa^{(F)}_{3,0} + \kappa^{(F)}_{3,1}N_{Adj} + \kappa^{(F)}_{3,2}N_{Adj}^2 + \kappa^{(F)}_{3,3}N_{Adj}^3 \ \Big ]} {3^3 N_c^3\Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \, \Big ]^5 } \ , \label{kappa3_fund_fa} \eeq where \beqs \kappa^{(F)}_{3,0} &=& 274243N_c^8-455426N_c^6-114080N_c^4+47344N_c^2+35574 \cr\cr &-& 2^7 \cdot 33 \zeta_3 N_c^2(4N_c^2-11)(25N_c^2-11) \label{kappabar_fund_fa_Nadjoint0} \eeqs \beqs \kappa^{(F)}_{3,1} &=& 135848N_c^8 - 215832N_c^6 + 291424N_c^4 -189168N_c^2 -25872 \cr\cr &-& 2^9 \cdot 3^2 \zeta_3 N_c^2 \Big ( 73N_c^4+132N_c^2-121 \Big ) \label{kappabar_fund_fa_Nadjlinear} \eeqs \beqs \kappa^{(F)}_{3,2} &=& 2^5(N_c^2+1)\Big [ 689N_c^6-2651N_c^4+2775N_c^2 +147 + 2^6 \cdot 3^2 \zeta_3 N_c^2(6N_c^2 -11) \, \Big ] \label{kappabar_fund_fa_Nadjoint_quadratic} \eeqs and \beq \kappa^{(F)}_{3,3} = 2^{10}N_c^2(N_c^2+1)^2(-11+24\zeta_3) \ . \label{kappabar_fund_fa_Nadjcubic} \eeq Further, \beq \kappa_3^{(Adj)} = \frac{4N_c^5\Big [ \kappa^{(Adj)}_{3,0} + \kappa^{(Adj)}_{3,1}N_F + \kappa^{(Adj)}_{3,2}N_F^2 + \kappa^{(Adj)}_{3,3}N_F^3 \ \Big ]} {3^3\Big [18N_c^3-N_F(N_c^2+1) \Big ]^5} \ , \label{kappa3_adj_fa} \eeq where \beq \kappa^{(Adj)}_{3,0} = 3^3 N_c^8(61873N_c^2-42624) \label{kappabar_adj_fa_Nfund0} \eeq \beq \kappa^{(Adj)}_{3,1} = -36N_c^3\Big ( 6728N_c^6 -5857N_c^4-1247N_c^2+138 + 11520 \zeta_3N_c^4 \Big ) \label{kappabar_adj_fa_Nfundlinear} \eeq \beqs \kappa^{(Adj)}_{3,2} &=& 2^5(N_c^2+1)\Big ( 287N_c^6-1187N_c^4 + 27N_c^2+9 + 2448\zeta_3 N_c^4 \Big ) \label{kappabar_adj_fa_Nfundquadratic} \eeqs and \beq \kappa^{(Adj)}_{3,3} = -2^7N_c(N_c^2+1)^2(-11+24\zeta_3) \ . \label{kappabar_adj_fa_Nfundcubic} \eeq \end{widetext} If $N_{Adj}=0$, then the coefficient $\kappa_3^{(F)}$ reduces to the expression in Eq. (6.10) of our earlier work \cite{dex}, while if $N_F=0$, then $\kappa_3^{(Adj)}$ reduces to Eq. (6.20) of \cite{dex}. The agreement of these reductions of $\kappa_j^{(F)}$ for $N_{Adj}=0$ and of $\kappa_j^{(Adj)}$ for $N_F=0$ with our earlier calculations in \cite{dex} for $j=1,2,3$ serves as a check on our present results. As was discussed in \cite{dex,dexl}, these coefficients have the leading large-$N_c$ dependence \beq \kappa_j^{(F)} \sim N_c^{-j} \quad {\rm as} \ N_c \to \infty \label{kappa_fund_largeNc} \eeq and \beq \kappa_j^{(Adj)} \sim N_c^0 \quad {\rm as} \ N_c \to \infty \ . \label{kappa_adj_largeNc} \eeq As specific examples of these FA theories, we consider the following sets of SU(3) gauge theories in ${\cal R}_{IRZ}$ with the indicated fermion content: \beqs (N_F,N_{Adj}) &=& (8,\frac{1}{2}), \quad (8,1), \quad (10,0), \quad (10,\frac{1}{2}), \cr\cr && (10,1), \quad (12,0), \quad (12,\frac{1}{2}) \ . \label{fa_theories} \eeqs The respective positions of these theories in the regions ${\cal R}_{IRZ}$ and ${\cal R}_{NACP}$ can be ascertained by referring to Table \ref{fadj_su3_theories_table}. The corresponding values of the coefficients $\kappa^{(F)}_j$ with $j=1,2,3$, as functions of $N_{Adj}$, are listed in Table \ref{kappa_fund_su3fa_table}, and the values of $\kappa^{(Adj)}_j$ with $j=1,2,3$, as functions of $N_F$, are listed in Table \ref{kappa_adj_su3fa_table}. We observe that all of these coefficients are positive, and so the generalizations of the monotonicity relations that we found in our earlier work for the theory with fermions in a single representation also hold for this FA theory, namely (i) for fixed $N_{Adj}$, $\gamma_{\bar\psi\psi,IR}$ is a monotonically increasing function of $\Delta_F$, i.e., a monotonically increasing function of decreasing $N_F$; (ii) for fixed $N_F$, $\gamma_{\bar\chi\chi,IR}$ is a monotonically increasing function of $\Delta_{Adj}$, i.e., a monotonically increasing function of decreasing $N_{Adj}$; (iii) for fixed $N_{f'}$, $\gamma_{\bar\psi\psi,IR,\Delta_F^p}$ is a monotonically increasing function of $p$; and (iv) for fixed $N_f$, $\gamma_{\bar\chi\chi,IR,\Delta_{Adj}^p}$ is a monotonically increasing function of $p$. Separately, we also note a generalization of the monotonicity relation that we proved for $\kappa^{(F)}_1$ and proved for $\kappa^{(Adj)}_1$, namely that for these FA theories, the $\kappa^{(F)}_j$ coefficients with $j=1,2,3$ are monotonically decreasing functions of $N_{Adj}$, and the $\kappa^{(Adj)}_j$ coefficients with $j=1,2,3$ are monotonically increasing functions of $N_F$. Having calculated these coefficients $\kappa^{(F)}_j$ and $\kappa^{(Adj)}_j$ with $j=1,2,3$ for this FA theory, we next proceed to substitute them in the general scheme-independent expansions (\ref{gamma_ir_Deltaseries}) for $f=F$ and the analogue for $f'=Adj$. Explicitly, with $f = \psi$ and $f'=\chi$, \beq \gamma_{\bar\psi\psi,IR} = \sum_{j=1}^\infty \kappa^{(F)}_j \Delta_F^j \label{gamma_fund_Deltaseries_fa} \eeq and \beq \gamma_{\bar\chi\chi,IR} = \sum_{j=1}^\infty \kappa^{(Adj)}_j \Delta_{Adj}^j \ , \label{gamma_adj_Deltaseries_fa} \eeq where \beq \Delta_F = (N_{F,u}-N_F) \ , \label{Delta_fund} \eeq with \beq N_{F,u} = \frac{N_c(11-4N_{Adj})}{2} \label{Nfundup} \eeq and \beq \Delta_{Adj} = (N_{Adj,u}-N_{Adj}) \ , \label{Delta_Adj} \eeq with \beq N_{Adj,u} = \frac{11}{4} - \frac{N_F}{2N_c} \ . \label{Nadjup} \eeq For reference, we list the values of $N_{F,u}$ and $N_{Adj,u}$ from Eqs. (\ref{Nfundup}) and (\ref{Nadjup}) for these $(N_F,N_{Adj})$ FA SU(3) theories in Table \ref{nup_values_su3_fa}. In Table \ref{gamma_fund_fa_su3_values} we list the values of $\gamma_{\bar\psi\psi,IR}$ calculated to $O(\Delta_F^p)$ for $p=1,2,3$, denoted as $\gamma_{\bar\psi\psi,IR,\Delta_F^p}$. Similarly, in Table \ref{gamma_adj_fa_su3_values} we list the values of $\gamma_{\bar\chi\chi,IR}$ calculated to $O(\Delta_{Adj}^p)$ for $p=1,2,3$, denoted as $\gamma_{\bar\chi\chi,IR,\Delta_{Adj}^p}$. The monotonicity relations noted above are evident in these tables. From an examination of the fractional changes in the anomalous dimensions as one increases the order of calculation, one may infer that these scheme-independent expansions should be reasonably reliable. For example, in the SU(3) FA theory with $(N_F,N_{Adj})=(12,1/2)$ theory, the fractional change in the $\gamma_{\bar\psi\psi,IR}$ anomalous dimension is \begin{widetext} \beq {\rm SU}(3), \ (N_F,N_{Adj}) = (12,1/2) \Longrightarrow \quad \frac{\gamma_{\bar\psi\psi,IR,\Delta_F^3} - \gamma_{\bar\psi\psi,IR,\Delta_F^2}}{\gamma_{\bar\psi\psi,IR,\Delta_F^2}} = 0.81 \times 10^{-2} \ . \eeq In the SU(3) FA theory with $(N_F,N_{Adj})=(10,1)$ theory, the fractional change in $\gamma_{\bar\psi\psi,IR}$ is even smaller: \beq {\rm SU}(3), \ (N_F,N_{Adj}) = (10,1) \Longrightarrow \quad \frac{\gamma_{\bar\psi\psi,IR,\Delta_F^3} - \gamma_{\bar\psi\psi,IR,\Delta_F^2}}{\gamma_{\bar\psi\psi,IR,\Delta_F^2}} = 0.87 \times 10^{-3} \ , \eeq yielding identical entries listed to three significant figures in Table \ref{gamma_fund_fa_su3_values}. Similar comments apply to the calculations of $\gamma_{\bar\chi\chi,IR,\Delta_{Adj}^p}$. \end{widetext} \section{Scheme-Independent Calculation of $\beta'_{IR}$} \label{betaprime_section} In this section we return to the general asymptotically free gauge theory with gauge group $G$ containing $N_f$ and $N_{f'}$ fermions in the respective representations $R$ and $R'$ and present our calculations of the coefficients $d_j$ and $\tilde d_j$ in the scheme-independent expansions of the derivative of the beta function evaluated at the IR fixed point, $\beta'_{IR}$, in powers of $\Delta_f$ in Eqs. (\ref{betaprime_ir_Deltaseries}) and in powers of $\Delta_{f'}$ in Eq. (\ref{betaprime_ir_Deltaseries2}), respectively. As before in this paper, this IR fixed point is taken to be in the non-Abelian Coulomb phase. Part of the physical interest in the quantity $\beta'_{IR}$ stems from the fact that, owing to the trace anomaly relation \cite{traceanomaly}, it is equivalent to the anomalous dimension of the field-strength tensor term ${\rm Tr}(F^a_{\mu\nu}F^{a \mu\nu})$ in the Lagrangian \cite{tracerel,dex}. As noted above, generalizing our result for the single-representation case, $d_1 = \tilde d_1=0$ for arbitrary $G$, $R$, and $R'$. For the higher coefficients we find \beq d_2 = \frac{2^5 T_f^2}{3^2 {\cal D}_f} \label{d2} \eeq \beq d_3 = \frac{2^7 T_f^3 (5C_A+3C_f)}{3^3 {\cal D}_f^2} \label{d3} \eeq and \beq d_4 = -\frac{2^3T_f^2}{3^6{\cal D}_f^5} \, \Big [ B^{(f)}_0 + B^{(f)}_1 N_{f'} + B^{(f)}_2 N_{f'}^2 + B^{(f)}_3 N_{f'}^3 \Big ] \ , \label{d4} \eeq where we explicitly indicate the dependence on $f$ in the $B^{(f)}_s$, $s=0,1,2,3$. (We extract a minus sign in Eq. (\ref{d4}) to maintain the same notation as in our earlier works \cite{dex,dexl}, where we found that in the case of fermions in a single representation $R=F$, $d_4$ is negative.) As was the case with $A^{(f)}_0$ in $\kappa^{(f)}_3$, the $B^{(f)}_0$ term in $d_4$ is independent of $N_{f'}$ and hence, taking into account the difference in the prefactor, it is equal to $C_A$ times the terms in the square bracket of Eq. (5.11) in our earlier Ref. \cite{dex} or equivalently, Eq. (4.8) of our Ref. \cite{dexl}. We have \begin{widetext} \beqs B^{(f)}_0 &=& C_A \Bigg [ -3C_AT_f^2 \bigg ( 137445 C_A^4 + 103600C_A^3C_f+72616C_A^2C_f^2 +951808C_AC_f^3 - 63888C_f^4 \bigg ) \cr\cr &+&2^8D \bigg ( -20T_f^2 \frac{d_A^{abcd}d_A^{abcd}}{d_A} +352C_AT_f\frac{d_f^{abcd}d_A^{abcd}}{d_A} -1331C_A^2\frac{d_f^{abcd}d_f^{abcd}}{d_A} \bigg ) \cr\cr &+& 8448D \zeta_3 \, \bigg \{ C_A^2T_f^2\Big ( 21C_A^2+12C_AC_f-33C_f^2\Big ) + 16T_f^2 \frac{d_A^{abcd}d_A^{abcd}}{d_A} - 104 C_AT_f \frac{d_f^{abcd}d_A^{abcd}}{d_A} + 88C_A^2\frac{d_f^{abcd}d_f^{abcd}}{d_A} \bigg \} \ \Bigg ] \ . \cr\cr & & \label{d4_B0} \eeqs For the $B^{(f)}_j$ with $j=1,2,3$, we calculate \beqs B^{(f)}_1 &=& 194880C_A^4T_f^2T_{f'}(C_{f'}-C_f) + C_A^3T_f^2T_{f'}\bigg ( -1854816C_f^2+2715648C_fC_{f'}-860832C_{f'}^2\bigg ) \cr\cr &+& C_A^2T_f^2T_{f'}\bigg ( 903168C_f^3+153216C_f^2C_{f'} -1241856C_fC_{f'}^2+185472C_{f'}^3 \bigg ) \cr\cr &+&C_AT_f^2T_{f'}C_f \bigg ( -139392C_f^3-164736C_f^2C_{f'} +12672C_fC_{f'}^2+291456C_{f'}^3 \bigg ) \cr\cr &+& C_A^2T_f^2T_{f'} (C_f-C_{f'}) \zeta_3 \, \bigg ( -967680C_A^2 +608256C_AC_f + 3345408C_f^2 \bigg ) \cr\cr &+& T_f^2T_{f'}(C_f-C_{f'})\Big (20480-540672\zeta_3\Big ) \frac{d_A^{abcd}d_A^{abcd}}{d_A} \cr\cr &+& C_AT_fT_{f'}\bigg [ -229376C_A-720896C_f+360448C_{f'} + \zeta_3\Big ( 2236416C_A+7028736C_f-3514368C_{f'} \Big ) \, \bigg ]\, \frac{d_f^{abcd}d_A^{abcd}}{d_A} \cr\cr &+&C_AT_f^2 \bigg [ 229376C_A+360448C_f + \zeta_3\bigg (-2236416C_A-3514368C_f\bigg ) \, \bigg ]\frac{d_{f'}^{abcd}d_A^{abcd}}{d_A} \cr\cr &+&C_A^2T_{f'}\bigg [ 1734656C_A+4088832C_f-1362944C_{f'} + \zeta_3\bigg ( -3784704C_A-8921088C_f+2973696C_{f'}\bigg ) \, \bigg ] \frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &+&C_A^2T_f \bigg [ C_A\Big ( -1734656+3784704\zeta_3\Big ) + C_f\Big ( -2725888+5947392\zeta_3\Big ) \, \bigg ] \frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} \label{d4_B1} \eeqs \beqs B^{(f)}_2 &=& T_f^2T_{f'}^2 \Bigg [ 669696 C_A^2(C_f-C_{f'})^2 +C_A\bigg (437760C_f^3-1327104C_f^2C_{f'}+1340928C_fC_{f'}^2-451584C_{f'}^3\bigg) \cr\cr &+& 25344C_f^4+59904C_f^3C_{f'}-87552C_f^2C_{f'}^2-105984C_fC_{f'}^3+108288C_{f'}^4 \cr\cr &+& C_A(C_f-C_{f'})^2\, \zeta_3 \, \Big (-110592C_A-1216512C_f \Big ) \, \Bigg ] \cr\cr &+& T_fT_{f'}(C_f-C_{f'})\Big (131072-1277952\zeta_3\Big )\bigg [ T_{f'}\frac{d_f^{abcd}d_A^{abcd}}{d_A} -T_f\frac{d_{f'}^{abcd}d_A^{abcd}}{d_A} \bigg ] \cr\cr &+&C_AT_{f'}^2\bigg [ -315392C_A-1486848C_f+991232C_{f'} + \zeta_3\bigg (688128C_A+3244032C_f-2162688C_{f'}\bigg ) \, \bigg ]\frac{d_f^{abcd}d_f^{abcd}}{d_A} \cr\cr &+&C_AT_fT_{f'}\bigg [ 630784C_A+1982464C_f-991232C_{f'} + \zeta_3\bigg ( -1376256C_A - 4325376C_f+2162688C_{f'}\bigg ) \, \bigg ]\frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} \cr\cr &+&C_AT_f^2\bigg [-315392C_A-495616C_f + \zeta_3\Big (688128C_A+1081344C_f\Big ) \, \bigg ]\frac{d_{f'}^{abcd}d_{f'}^{abcd}}{d_A} \label{d4_B2} \eeqs and \beqs B^{(f)}_3 &=& T_{f'}(C_f-C_{f'})\Bigg [ 2^{11} \cdot 3 T_f^2T_{f'}^2(C_f-C_{f'})^2\Big (-23 + 24\zeta_3\Big ) \cr\cr &&+2^{14}(11-24\zeta_3)\bigg (T_{f'}^2 \frac{d_f^{abcd}d_f^{abcd}}{d_A} -2T_fT_{f'}\frac{d_f^{abcd}d_{f'}^{abcd}}{d_A} + T_f^2\frac{d_{f'}^{abcd}d_{f'}^{abcd}}{d_A} \bigg ) \, \Bigg ] \ . \label{d4_B3} \eeqs In passing, we note that $B^{(f)}_3$ has the same prefactor as $A^{(f)}_3$ in Eq. (\ref{A3A3bar}), namely $T_{f'}(C_f-C_{f'})$. \end{widetext} The corresponding coefficients for the expansion (\ref{betaprime_ir_Deltaseries2}) are obtained from these by interchanging $f$ and $f'$. Thus, for example, \beq \tilde d_2 = \frac{2^5 T_{f'}^2}{3^2 {\cal D}_{f'}} \label{d2f2} \eeq \beq \tilde d_3 = \frac{2^7 T_{f'}^3 (5C_A+3C_{f'})}{3^3 {\cal D}_{f'}^2} \ , \label{d3f3} \eeq and similarly for $\tilde d_4$. \section{Results for $\beta'_{IR}$ in a Theory with Fermions in the Fundamental and Adjoint Representations of SU($N_c$)} \label{betaprime_fadj_section} In this section we discuss the special case of our general calculation of $\beta'_{IR}$ for an SU($N_c$) theory with $N_f$ fermions in the fundamental representation and $N_{Adj}$ fermions in the adjoint representation (i.e., the FA theory). We write Eqs. (\ref{betaprime_ir_Deltaseries}) and (\ref{betaprime_ir_Deltaseries2}) as \beq \beta'_{IR} = \sum_{j=2}^\infty d_j \Delta_F^j \label{betaprime_ir_Deltaseries_fa} \eeq and \beq \beta'_{IR} = \sum_{j=2}^\infty \tilde d_j \Delta_{Adj}^j \ , \label{betaprime_ir_Deltaseries2_fa} \eeq where $\Delta_F$ and $\Delta_{Adj}$ were given explicitly in Eqs. (\ref{Delta_fund})-(\ref{Nadjup}). We calculate \beq d_2 = \frac{2^4}{3^2\Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \, \Big ]} \label{d2_fa} \eeq \beq \tilde d_2 = \frac{2^5N_c^3}{3^2\Big [18N_c^3-N_F(N_c^2+1) \, \Big ]} \label{d2tilde_fa} \eeq \beq d_3 = \frac{2^5(13N_c^2-3)}{3^3N_c\Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \, \Big ]^2 } \label{d3_fa} \eeq \beq \tilde d_3 = \frac{2^{10}N_c^6}{3^3 \Big [18N_c^3-N_F(N_c^2+1) \, \Big ]^2} \label{d3tilde_fa} \eeq \beq d_4 = \frac{2^4\Big [ d_{4,0} + d_{4,1} N_{Adj} + d_{4,2} N_{Adj}^2 + d_{4,3} N_{Adj}^3 \ \Big ]} {3^5 N_c^2 \Big [ 25N_c^2-11 + 4N_{Adj}(N_c^2+1) \, \Big ]^5 } \ . \label{d4_fa} \eeq where \begin{widetext} \beqs d_{4,0} &=& 366782 N_c^8 -865400N_c^6+1599316N_c^4-571516N_c^2-3993 \cr\cr &+&\zeta_3 N_c^2\Big [ -660000N_c^6 + 765600N_c^4-2241888N_c^2+894432 \Big ] \label{d40} \eeqs \beqs d_{4,1} &=& 18416N_c^8 + 346944N_c^6 - 756920N_c^4+530256N_c^2+2904 \cr\cr &+& \zeta_3 N_c^2 \Big [ 28800N_c^6+372096N_c^4+1026432N_c^2-975744 \Big ] \label{d4_Nadj_linear} \eeqs \beqs d_{4,2} &=& 2^4(N_c^2+1)\Big [ -3161N_c^6+10589N_c^4-10155N_c^2-33 \cr\cr &+&\zeta_3 N_c^2 ( 3744N_c^4-13248N_c^2 + 22176 ) \, \Big ] \label{d4_Nadj_quadratic} \eeqs \beqs d_{4,3} &=& 2^8 N_c^2(N_c^2+1)^2 \Big [ -23N_c^2+65+\zeta_3(24N_c^2-168) \ , \, \Big ] \label{d4_Nadj_cubic} \eeqs and \beq \tilde d_4 = \frac{2^3 N_c^5 \Big [ \hat d_{4,0} + \hat d_{4,1} N_F + \hat d_{4,2} N_F^2 + \hat d_{4,3} N_F^3 \ \Big ]} {3^5\Big [ 18N_c^3 - N_F(N_c^2+1) \, \Big ]^5 } \ , \label{dtilde4_fa} \eeq where \beqs \hat d_{4,0} &=& 3^3N_c^8(46871N_c^2 + 85248) \label{d4hat0} \eeqs \beq \hat d_{4,1} = 36 N_c^3 \Big [ 1287N_c^6-23350N_c^4-1961N_c^2+276 + \zeta_3 N_c^4 \Big (-6912N_c^2+16128 \Big ) \, \Big ] \label{d4hat_NF_linear} \eeq \beq \hat d_{4,2} = 4(N_c^2+1)\Big [ -5153N_c^6+18113N_c^4-747N_c^2-141 +\zeta_3 N_c^4 \Big (6912N_c^2-32256\Big ) \, \Big ] \label{d4hat_NF_quadratic} \eeq \beq \hat d_{4,3} = 2^5 N_c(N_c^2+1)^2 \Big [ 23N_c^2 -65+\zeta_3(-24N_c^2+168) \, \Big ] \ . \label{d4hat_NF_cubic} \eeq \end{widetext} If $N_{Adj}=0$, then $d_2$, $d_3$, and $d_4$ reduce to our previous results in, respectively, Eqs. (5.14), (5.15), and (5.16) of \cite{dex}. Similarly, if $N_F=0$, then $\tilde d_2$, $\tilde d_3$, and $\tilde d_4$ reduce to our previous results in, respectively, Eqs. (5.59), (5.60), and (5.61) of \cite{dex}. The agreement of these reductions of $d_j$ for $N_{Adj}=0$ and of $\tilde d_j$ for $N_F=0$ with our results in \cite{dex} for $j=1,2,3$ serves as a check on our present calculations. As was discussed in \cite{dex,dexl}, these coefficients have the leading large-$N_c$ dependence \beq d_j \sim N_c^{-j} \quad {\rm as} \ N_c \to \infty \label{dj_largeNc} \eeq and \beq \tilde d_j \sim N_c^0 \quad {\rm as} \ N_c \to \infty \ . \label{dtildej_largeNc} \eeq In Table \ref{betaprime_fa_su3_values} we present our scheme-independent calculations of $\beta'_{IR}$ to order $O(\Delta_F^p)$ via the expansion (\ref{betaprime_ir_Deltaseries_fa}) and to $O(\Delta_{Adj}^p)$ via the expansion (\ref{betaprime_ir_Deltaseries2_fa}), with $p=1,2,3$, where $\Delta_F$ and $\Delta_{Adj}$ were defined in Eqs. (\ref{Delta_fund})-(\ref{Nadjup}). These are denoted $\beta'_{IR,\Delta_F^p}$ and $\beta'_{IR,\Delta_{Adj}^p}$, respectively. Graphically, in the first quadrant of ${\mathbb R}^2$ defined by $(N_F,N_{adj})$ (formally generalized to non-negative real numbers), the series (\ref{betaprime_ir_Deltaseries_fa}) is an expansion in a leftward horizontal direction from the $b_1=0$ line toward a given point $(N_F,N_{Adj})$ in the NACP, while the series (\ref{betaprime_ir_Deltaseries_fa}) is an expansion inward in a downward vertical direction from the $b_1=0$ line toward this point $(N_F,N_{Adj})$. Since these are two alternate expansions for the same quantity, one expects that as the maximal power $p$ in the series increases, they should yield similar values, and we see that this expectation is satisfied by our results at the highest order, $p=3$, as listed in Table \ref{betaprime_fa_su3_values}. The agreement between the two series is best when the $(N_F,N_{Adj})$ theory is near to the upper end of the non-Abelian Coulomb phase, since in this case the expansion parameters $\Delta_F$ and $\Delta_{Adj}$ are the smallest. Some explicit examples that demonstrate this accuracy are provided by the following fractional differences: \begin{widetext} \beq {\rm SU}(3), \ (N_F,N_{Adj})=(10,1) \ \Longrightarrow \quad \frac{|\beta'_{IR,\Delta_F^4}-\beta'_{IR,\Delta_{Adj}^4}|} {\beta'_{IR,\Delta_F^4}} = 2.2 \times 10^{-5} \eeq and \beq {\rm SU}(3), \ (N_F,N_{Adj})=(12,\frac{1}{2}) \ \Longrightarrow \quad \frac{|\beta'_{IR,\Delta_F^4}-\beta'_{IR,\Delta_{Adj}^4}|} {\beta'_{IR,\Delta_F^4}} = 1.0 \times 10^{-3} \ . \eeq \end{widetext} \section{Conclusions} \label{conclusion_section} In this paper, generalizing our previous work, we have considered an asymptotically free gauge theory with gauge group $G$ and two different fermion representations, with the property that it exhibits an infrared fixed point such that the infrared theory is in a non-Abelian Coulomb phase. Specifically, we have considered a theory with $N_f$ fermions transforming according to a representation $R$ of $G$ and $N_{f'}$ fermions transforming according to a different representation, $R'$. We have calculated scheme-independent series expansions of the anomalous dimensions of gauge invariant fermion bilinears and the derivative $\beta'_{IR}$ evaluated at the IR fixed point in the respective expansion parameters $\Delta_f$ and $\Delta_{f'}$. As an explicit application, we have presented calculations for an SU($N_c$) theory with $N_F$ fermions in the fundamental representation and $N_{Adj}$ fermions in the adjoint representation. Our results for scheme-independent expansions of gauge-invariant fermion bilinears extend up to $O(\Delta_F^3)$ and $O(\Delta_{Adj}^3)$, while our results for $\beta'_{IR}$ extend up to $O(\Delta_F^4)$ and $O(\Delta_{Adj}^4)$. These results provide further information about the properties of these conformal field theories. To the extent that the transition from the lower part of the non-Abelian Coulomb phase to the quasi-conformal regime in the variables $(N_f,N_{f'})$ is continuous and our finite-order perturbative calculations in the lower part of the non-Abelian Coulomb phase are sufficiently accurate, our present results can also be useful for the investigation of quasi-conformal theories with possible relevance to ultraviolet completions of the Standard Model. \begin{acknowledgments} This research was supported in part by the Danish National Research Foundation grant DNRF90 to CP$^3$-Origins at SDU (T.A.R.) and by the U.S. NSF Grant NSF-PHY-16-1620628 (R.S.) \end{acknowledgments} \begin{appendix} \section{Group Invariants} \label{groupinvariants} In this appendix we discuss some relevant group-theoretic quantities. Let us denote the generators of the Lie algebra of the gauge group $G$, in the representation $R$, as $T^a_R$, with $1 \le a \le d_A$, where $d_A$ is the order of the group. These generators satisfy the commutation relations \beq [T^a_R,T^b_R]=if^{abc} T^c_R \ , \label{algebra} \eeq where the $f^{abc}$ are the associated structure constants of this Lie algebra. Here and elsewhere, a sum over repeated indices is understood. We denote the dimension of a given representation $R$ as $d_R = {\rm dim}(R)$. In particular, we denote the adjoint representation by $A$, with the dimension $d_A$ equal to the number of generators of the group, i.e., the order of the group. The trace invariant is given by \beq {\rm Tr}_R(T^a_R T^b_R) = T(R)\delta_{ab} \ . \label{trace} \eeq The quadratic Casimir invariant $C_2(R)$ is defined by \beq T^a_RT^a_R = C_2(R) I \ , \label{c2} \eeq where $I$ is the $d_R \times d_R$ identity matrix. For a fermion $f$ transforming according to a representation $R$, we often use the equivalent compact notation $T_f \equiv T(R)$ and $C_f \equiv C_2(R)$. We also use the notation $C_A \equiv C_2(A) \equiv C_2(G)$. The invariants $T(R)$ and $C_2(R)$ satisfy the relation $C_2(R)d_R = T(R)d_A$. For $G={\rm SU}(N_c)$, $C_A = N_c$ and for $R$ equal to the fundamental representation, $T(R)=1/2$ and $C_2(R)=(N_c^2-1)/(2N_c)$. At the four-loop and five-loop level, one encounters traces of quartic products of the Lie algebra generators. For a given representation $R$ of $G$, \beqs d^{abcd}_R &=& \frac{1}{3!} {\rm Tr}_R\Big [ T_a(T_bT_cT_d + T_bT_dT_c + T_cT_bT_d \cr\cr &+& T_cT_dT_b + T_dT_bT_c + T_dT_cT_b) \Big ] \ . \label{d_abcd} \eeqs As with the quadratic invariants, for a fermion $f$ in the represenation $R$ of $G$, we often use the notation $d_R^{abcd} \equiv d_f^{abcd}$. In this context, for $R=Adj$, we use $d_R^{abcd}=d_A^{abcd}$. The quantities that appear in the anomalous dimensions and derivative of the beta function $\beta'_{IR}$ that we calculate are products of these $d_R^{abcd}$ of the form $d_R^{abcd}d_{R'}^{abcd} \equiv d_f^{abcd}d_{f'}^{abcd}$, summed over the group indices $a, \ b, \ c, \ d$. For further discussion of these, with references to the literature, see \cite{rsv,dexo} and references therein. \end{appendix}

90042df422b56ca096a853d4f5b72edeba6a25a7

\section{Introduction} Let $d\in \NN_+:=\{1,2,\cdots\}$ and $\alpha\in(0,d\wedge 2)$. We consider the following Schr\"{o}dinger operator \begin{equation}\label{e:op1} \mathcal{L} := \Delta^{\alpha/2} +q \end{equation} on $\RR^d$, where $\Delta^{\alpha/2}:=-(-\Delta)^{\alpha/2}$ and $q(x) = \kappa|x|^{-\alpha}$ with \begin{equation}\label{e:constant}\kappa=\kappa_\delta:= \frac{2^\alpha\Gamma(\frac{\alpha-\delta}{2}) \Gamma(\frac{d+\delta}{2})} {\Gamma(\frac{-\delta}{2})\Gamma(\frac{d+\delta-\alpha}{2})} \end{equation} for any $\delta \in(0,\alpha).$ Here, $\Gamma(-\delta/2)=\int_0^\infty r^{-1-\delta/2}(e^{-r}-1)\,dr<0$ for $\delta\in (0,\alpha)$, and $\Gamma(z)=\int_0^\infty r^{z-1}e^{-r}\,dr$ for all $z>0$. We note that $\kappa_\delta<0$, and so $q(x) <0$ on $\RR^d$; we also note that the radial function $|x|^{-\alpha}$ comes from the Hardy inequality for fractional Laplacian $\Delta^{\alpha/2}$ (see \cite{2016-KB-BD-PK-pa, FLS} and the references therein for more details). Thus, the operator $\mathcal{L}$ given by \eqref{e:op1} is the fractional Schr\"{o}dinger operator with negative Hardy potential. Denote by $\tp(t,x,y)$ the heat kernel associated with the operator $\mathcal{L}=\Delta^{\alpha/2} +q$; see Subsection \ref{e:sub2-1} below for more details. Our main result is as follows. \begin{theorem}\label{th:main}For any $\delta\in (0,\alpha)$, the Schr\"{o}dinger operator $\mathcal{L}$ given by \eqref{e:op1} has the heat kernel $\tp(t,x,y)$, which is jointly continuous on $(0,\infty)\times \RR^d\times \RR^d$, and satisfies two-sided estimates as follows \begin{align}\label{EQ:main} \tp(t,x,y) \approx \left(1\wedge \frac{|x|}{t^{1/\alpha}}\right)^\delta \left(1\wedge \frac{|y|}{t^{1/\alpha}}\right)^\delta \left(t^{-d/\alpha}\wedge\frac{t}{|x-y|^{d+\alpha}}\right),\quad x,y\in \RR^d,\; t>0. \end{align} \end{theorem} We note that the last expression in \eqref{EQ:main} may be replaced by the heat kernel $p(t,x,y)$ of $\Delta^{\alpha/2}$ (see Subsection 2.1 and \eqref{eq:pest}). As pointed out before Lemma \ref{L:2.3} below, the function $\delta\mapsto \kappa_\delta$ is strictly decreasing on $(0,\alpha)$ with $\lim_{\delta\to0}\kappa_\delta=0$ and $\lim_{\delta\to \alpha}\kappa_\delta=-\infty.$ Hence, Theorem \ref{th:main} essentially gives us two-sided estimates and the joint continuity of heat kernel associated with the operator $\Delta^{\alpha/2} -\lambda |x|^{-\alpha}$ for all $\lambda>0$. It is well known that the fractional Laplacian $\Delta^{\alpha/2}$ is the infinitesimal generator of the rotationally symmetric $\alpha$-stable process, which now has been attracted a lot of interests in the field of probability and potential theory (see \cite{BBK} and references therein). Recently there are also a few works concerning on gradient perturbations and Schr\"{o}dinger perturbations of fractional Laplacian (see e.g. \cite{2008-KB-WH-TJ-sm, BJ, BJ2012, 2013-KB-TJ-SS-jee, CKS3, CKS2, 2009-TJ-pa, KK, 2013-TK-pa, Song1, Take}). In particular, according to \cite[Theorem 3.4]{Song1}, when the potential belongs to the so-called Kato class, heat kernel estimates for Schr\"{o}dinger perturbations of fractional Laplacian are comparable with these for fractional Laplacian (at least for any fixed finite time). Note that $q(x)=-\lambda|x|^{-\alpha}$ does not belong to the Kato class. As shown in Theorem \ref{th:main}, the heat kernel $\tp(t,x,y)$ associated with the Schr\"{o}dinger operator $\mathcal{L}$ given by \eqref{e:op1} exhibits behaviour which is different from that of the case that $q(x)=-\lambda|x|^{-\gamma}$ with $\gamma\in (0,\alpha),$ which is in the Kato class. The study of heat kernel estimates for Schr\"odinger-type perturbations by the Hardy potential of fractional Laplacian is much more delicate. In the classical case $\alpha=2$, the Schr\"odinger-type perturbations by the Hardy potential were considered for the first time by Baras and Goldstein \cite{MR742415}. They proved the existence of nontrivial nonnegative solutions of the classical heat equation $\partial_t=\Delta+\kappa|x|^{-2}$ in $\Rd$ for $0\le \kappa\le (d-2)^2/4$, and nonexistence of such solutions, that is explosion, for bigger constants $\kappa$. Sharp upper and lower bounds for the heat kernel of the Schr\"odinger operator $\Delta+\kappa|x|^{-2}$ were obtained by Liskevich and Sobol \cite[p.~365 and Examples 3.8, 4.5 and 4.10]{2003-VL-ZS-pa} for $0<\kappa< (d-2)^2/4$. Milman and Semenov proved the upper and lower bounds for $\kappa\le (d-2)^2/4$, see \cite[Theorem 1]{2004-PM-YS-jfa} and \cite{MR2180080}. In this paper, they also allowed $\kappa <0$ and obtained the sharp upper and lower bounds for the perturbed kernel (see \cite[Theorem 2 and Corollary 4]{2004-PM-YS-jfa}). See \cite{IKO} and the references therein for the recent works of this topic. For $\alpha \in (0, d \land 2)$ the Schr\"odinger operator $\mathcal{L}$ with $\kappa \ge 0$ attains recently more and more interest. In \cite{AMPP1,AMPP2} for $ \kappa > \kappa^* := \frac{2^{\alpha} \Gamma((d+\alpha)/4)^2 }{\Gamma((d-\alpha)/4)^{2}}$ the phenomenon of instantaneous blow up of heat kernel was proven. In \cite{BA}, the author gives the upper bound for the heat kernel of $\mathcal{L}$ with the Dirichlet conditions on bounded open subsets of $\RR^d$. In the recent paper \cite{2017-KB-TG-TJ-DP}, the following sharp estimates for the heat kernel $\tilde{p}(t,x,y)$ of $\mathcal{L}$ were obtained. For $0 \le \kappa \le \kappa^*$, there is a unique constant $\delta \in [0,(d-\alpha)/2]$ such that for all $t>0$ and $x,y \in \RR^d_0:=\RR^d\backslash\{0\}$, \begin{equation} \label{EQ:positEst} \tilde{p}(t,x,y)\approx \left(1+t^{\delta/\alpha}|x|^{-\delta} \right)\left(1+t^{\delta/\alpha}|y|^{-\delta} \right) p(t,x,y). \end{equation} Note that since the singularity of the function $\Rd\ni x\mapsto\kappa |x|^{-\alpha}$ at the origin is critical, $\tilde{p}(t,x,y)$ is not comparable with the unperturbed kernel $p(t,x,y)$. Like in Theorem \ref{th:main}, the choice of $\kappa$ influences the growth rate or the decay rate of the heat kernel at the origin. This rate is represented by the function $|x|^{-\delta}$, where $\delta$ is connected with $\kappa$ via the formula $\kappa=\frac{2^\alpha\Gamma(\frac{\alpha+\delta}{2}) \Gamma(\frac{d-\delta}{2})}{\Gamma(\frac{\delta}{2})\Gamma(\frac{d-\delta-\alpha}{2})}$ (compared with \eqref{e:constant}). \begin{remark}Theorem \ref{th:main} and \cite[Theorem 1.1]{2017-KB-TG-TJ-DP} (the main result of \cite{2017-KB-TG-TJ-DP}) can be stated together as follows. \emph{ For any $-\infty < \kappa \le \kappa^*=\frac{2^{\alpha} \Gamma((d+\alpha)/4)^2 }{\Gamma((d-\alpha)/4)^{2}}$, the heat kernel $\tp(t,x,y)$ corresponding to the Schr\"odinger operator $\Delta^{\alpha/2}+\kappa |x|^{-\alpha}$ satisfies \begin{align*} \tilde{p}(t,x,y)\approx \left(1+\frac{t^{1/\alpha}}{|x|^{\alpha}} \right)^\delta \left(1+\frac{t^{1/\alpha}}{|y|^{\alpha}} \right)^\delta p(t,x,y), \qquad t>0,\; x,y \in \RdO, \end{align*} where $\delta \in (-\alpha,\frac{d-\alpha}{2}]$ is uniquely determined by \begin{align*} \kappa= \frac{2^\alpha\Gamma(\frac{\alpha+\delta}{2}) \Gamma(\frac{d-\delta}{2})} {\Gamma(\frac{\delta}{2})\Gamma(\frac{d-\delta-\alpha}{2})}. \end{align*}} \end{remark} In this setting, Theorem \ref{th:main} may be treated as both a fractional counterpart of the result obtained in \cite{2004-PM-YS-jfa} and the extension of \eqref{EQ:positEst} to negative values of $\kappa$. Here, we would like to stretch out one difference between the cases $\alpha=2$ and $\alpha<2$ for $\kappa<0$. The general form of the estimate in both cases is similar, i.e., the perturbed kernel $\tp(t,x,y)$ is comparable with the unperturbed kernel $p(t,x,y)$ multiplied by some weighted functions. However, in \cite[Theorem 2 and Corollary 4]{2004-PM-YS-jfa}, for $\alpha=2$, the exponent of the weighted function is equal to $\delta= \frac{ \sqrt{(d-2)^2 - 4\kappa} - (d-2)}{2} $ and converges to infinity as $\kappa \to -\infty$. In our case $\alpha<2$, as it was mentioned below the statement of Theorem \ref{th:main}, $\delta \to \alpha$ for $\kappa \to -\infty$. Since $q(x)=\kappa |x|^{-\alpha}$ is negative and does not belong to any Kato class on $\RR^d$, the construction and proofs of the estimates of $\tp(t,x,y)$ are very delicate. In particular, we cannot use the perturbation series (at least for large values of $-\kappa$) to construct $\tp(t,x,y)$ as used in \cite{2016-KB-BD-PK-pa,2017-KB-TG-TJ-DP,2008-KB-WH-TJ-sm}. That is why we will consider the Dirichlet fractional Laplacian operator $\Delta^{\alpha/2}$ on $\RdO = \RR^d \setminus \{0\}$ and via the Feyman-Kac formula, we construct $\tp(t,x,y)$ on $(0,\infty)\times \RdO \times \RdO$. Hence, the operator $\mathcal{L}$ with negative values of $\kappa$ also enjoys some probabilistic meaning. Roughly speaking, it is connected with a symmetric $\alpha$-stable process with the killing rate $e^{-\kappa|x|^{-\alpha}}$, which strongly affects the behaviour of $\tp(t,x,y)$ for $x$ and $y$ near $0$. It turns out that due to the strong singularity of $q(x)$ at $0$, the heat kernel (or the transition density function) $\tp(t,x,y)$ is equal to $0$ when $x=0$ or $y=0$. In consequence, the kernel $\tp(t,x,y)$ defined on $(0,\infty) \times \RdO \times \RdO$ may be continuously extended to $(0,\infty) \times \RR^d \times \RR^d$. We note that Theorem \ref{th:main} was proved independently in a very recent paper \cite{CKSV}. In the proofs, the authors use generally probabilistic tools. In our paper we propose a different method. Although the perturbed kernel $\tp(t,x,y)$ is defined by the Feyman-Kac formula, in the proofs we apply only analytical tools. For upper bounds, we generally use the Chapman-Kolmogorov equation and the method of \lq\lq \emph{self-improving estimates}'' (see the proofs of Proposition \ref{lem:mass_est} and Theorem \ref{thm:UB}, see also the proof of \cite[Theorem 1.1]{2016-TJ-GS-natmaa}). Roughly speaking, to show the inequality $f(x) \le C F(x)$, we first show that $f(x) \le g_1(x) + c_1 F(x)$, where $g_1(x)$ is in some sense small. Next, by plugging this estimate to the proper functional inequality on $f$, we get the improved estimate of the form $f(x) \le g_n(x) + c_n F(x)$, where $g_n(x) \to 0$ as $n \to \infty$ and $\sup_{n \in \NN_+}c_n <\infty$. By passing with $n$ to infinity we obtain the desired estimate. To obtain lower bounds we use the generally well known estimate from Lemma \ref{lem:tpexp} and upper bound estimates. Although the estimate from Lemma \ref{lem:tpexp} is generally well known, we couldn't find the proper reference with the assumptions on the potential satisfied by $q(x)$. We note that the setting of \cite{CKSV} is more general than the present paper. From the other side, we give more details about the kernel $\tilde{p}(t,x,y)$, see e.g. Theorem \ref{thm:0}. We also note that in our paper we show the straightforward dependence between the exponent $\delta$ and the potential $q(x)$, while in \cite[Theorem 3.9]{CKSV} this dependence, given by double integral, is much more complicated. The paper is organized as follows. In Section 2, we construct $\tp(t,x,y)$ and prove some basic properties of this kernel. In Section 3, we give the proof of Theorem \ref{th:main}. First, we prove upper bounds in Theorem \ref{thm:UB}. Next, we show lower bounds in Theorem \ref{Thm:LB} and joint continuity (Theorem \ref{thm:cont}). We end this section with short discussion on Dirichlet forms associated with the Schr\"{o}dinger operator $\mathcal{L}$ given by \eqref{e:op1}. Finally, in the Appendix, we present the proof Lemma \ref{lem:tpexp}. Throughout the paper, we write $f\approx g$ for $f,g\ge0$, if there is a constant $c\ge1$ such that $c^{-1}f\le g\le c f$ on their common domain. The constants $c, C, c_i$, whose exact values are unimportant, are changed in each statement and proof. Let $B(x,r)$ be the open ball with center $x\in \RR^d$ and radius $r>0$. As usual we write $a \land b := \min(a,b)$ and $a \vee b := \max(a,b)$. \section{Preliminary estimates} \subsection{Fractional Laplacian and rotationally symmetric $\alpha$-stable L\'evy process} Let \begin{equation}\label{e:levymea} \nu(z) =\frac{ \alpha 2^{\alpha-1}\Gamma\big((d+\alpha)/2\big)}{\pi^{d/2}\Gamma(1-\alpha/2)}|z|^{-d-\alpha}\,,\quad z\in \Rd. \end{equation} For (smooth and compactly supported) test function $\varphi\in C^\infty_c(\Rd)$, we define the fractional Laplacian by \begin{equation*} \Delta^{\alpha/2}\varphi(x)=-(-\Delta)^{\alpha/2}\varphi(x):= \lim_{\varepsilon \downarrow 0}\int_{\{|z|>\varepsilon\}} \left[\varphi(x+z)-\varphi(x)\right]\nu(z)\,dz\,, \quad x\in \Rd\,. \end{equation*} In terms of the Fourier transform (see \cite[Section~1.1.2]{FLS}), $\widehat{\Delta^{\alpha/2}\varphi}(\xi)=-|\xi|^{\alpha}\hat{\varphi}(\xi)$. Denote by $p(t,x,y)$ the heat kernel (or the fundamental function) of $\Delta^{\alpha/2}$ (or equivalently, the transition density function of a (rotationally) symmetric $\alpha$-stable L\'evy process $(X_t)_{t\ge0}$). It is well known that $p(t,x,y)$ is symmetric in the sense that $p(t,x,y)=p(t,y,x)$ for any $t>0$ and $x,y\in \RR^d$, and enjoys the following scaling property $$p(t,x,y)=t^{-d/\alpha}p(1,t^{-1/\alpha} x, t^{-1/\alpha}y),\quad t>0,\; x,y\in \RR^d.$$ Moreover, \begin{align}\label{eq:pest} p(t,x,y) \approx t^{-d/\alpha} \land \frac{t}{|x-y|^{d+\alpha}},\quad t>0,\;x,y\in \RR^d. \end{align} We also note that $p(t,x,y)$ is a function of $t$ and $x-y$, so sometimes we also write it as $p(t,x-y)$, i.e. $p(t,x,y) = p(t,x-y)$. See \cite{BBK} for more details. \subsection{Fractional Laplacian Schr\"{o}dinger operator and Feynman-Kac formula}\label{e:sub2-1} In this part, we apply some results from \cite[Chapter 2]{DC} to the operator $\mathcal{L}=\Delta^{\alpha/2} +q$ given by \eqref{e:op1}, where $q(x)=\kappa_\delta |x|^{-\alpha}<0$. Let $\RR_0^d:=\RR^d\backslash\{0\}$. We first recall \cite[Chapter 2, Definition 2.1]{DC}. A nonnegative Borel measurable function $V$ on $\RR^d_0$ is said to belong to \emph{the Kato class} $\mathcal K_\alpha$, if $$\lim_{t\to0}\sup_{x\in \RR^d_0} \int_0^t\int_{\RR^d_0} p(s,x,y)V(y)\,dy\,ds=0.$$ A nonnegative Borel measurable function $V$ on $\RR^d_0$ is said to belong to \emph{the local Kato class} $\mathcal K_{\alpha, {\rm loc}}$, if $V\I_D\in \mathcal K_{\alpha}$ for all compact subsets $D$ of $\RR^d_0$. A Borel measurable function $V$ on $\RR^d_0$ is said to belong to \emph{the Kato-Feller class}, if its positive part $V_+:=\max\{V,0\}\in \mathcal K_\alpha$ and its negative part $V_-:=\max\{-V,0\}\in \mathcal K_{\alpha, {\rm loc}}.$ (Different from \cite{DC}, in the present setting we start from the nonpositive definite operator $\Delta^{\alpha/2} +q$, and so we make the corresponding changes in the definition of the Kato-Feller class.) It is easily seen from \cite[Lemma 2.3]{2017-KB-TG-TJ-DP} that $-q\notin \mathcal K_\alpha$, but always we have $-q\in \mathcal K_{\alpha, {\rm loc}}$. In particular, $q$ belongs to the Kato-Feller class. In the following, we will restrict ourselves on the killed subprocess of the symmetric $\alpha$-stable L\'evy process $(X_t)_{t\ge0}$ upon exiting $\RR_0^d$ (or hitting the origin), i.e., \begin{equation*} X^{\RR_0^d}_t:=\begin{cases} X_t,\quad\text{if}\ t<\tau_{\RR_0^d},\\ \,0,\,\,\quad \text{if}\ t\ge \tau_{\RR_0^d}, \end{cases} \end{equation*} where $\tau_{\RR_0^d}:=\inf \{t \ge 0: X_t \notin \RR_0^d\}=\inf\{t\ge0:X_t=0\}$. By the strong Markov property of the process $(X_t)_{t\ge0}$, it is easy to see that the process $(X_t^{\RR_0^d})_{t\ge0}$ has a transition density (or Dirichlet heat kernel) $p^{\RR_0^d}(t,x,y)$, which enjoys the following relation with $p(t,x,y)$: \begin{equation*} \begin{split} p^{\RR_0^d}(t,x,y)&=p(t,x,y)-\Ee^x\big[p(t-\tau_{\RR_0^d},X_{\tau_{\RR_0^d}},y)\I_{\{t\ge \tau_{\RR_0^d}\}}\big],\quad x, y \in {\RR_0^d};\\ p^{\RR_0^d}(t,x,y)&=0,\quad x=0 \text{ or } y =0. \end{split} \end{equation*} Since the process $(X_t)_{t\ge0}$ is transient due to $\alpha<d$, $\Pp^x(\tau_{\RR_0^d} < \infty)=0$ for all $x\in \RR_0^d$, and consequently \begin{equation}\label{e:hk1}p^{\RR_0^d}(t,x,y)=p(t,x,y),\quad t>0,\;x,y\in \RR_0^d.\end{equation} It is well known that for every $t>0$ the function $p(t,\cdot,\cdot)$ is continuous on ${\RR^d}\times {\RR^d}$, and $p(t,x,y)$ satisfies the following Chapman-Kolmogorov equation \begin{align}\label{eq:CKp} p(t+s,x,y)=\int_{\RR^d} p(t,x,z)p(s,z,y)\,dz,\quad t,s>0,\; x,y\in \RR^d. \end{align} Regard $\mathcal{L}=\Delta^{\alpha/2} +q$ as the operator defined on $C_\infty(\RR_0^d)$; that is, we consider a negative perturbation of the fractional Laplacian on $\RR^d_0$ (with the Dirichlet boundary condition at $\{0\}$). Therefore, according to \cite[Theorem 2.5]{DC}, the operator $\mathcal{L}=\Delta^{\alpha/2} +q$ can generate a strongly continuous and positivity preserving semigroup $(\tilde P_t)_{t\ge0}$ on $C_\infty(\RR_0^d)$, which is given by $$ \tilde P_t f(x)=\int \tp(t,x,y)f(y)\,dy,\quad f\in C_\infty (\RR^d_0),\; x\in \RR_0^d, $$ where the kernel $\tp(t,x,y)$ satisfies the Chapman-Kolmogorov equation too, i.e., \begin{align}\label{eq:CKpt} \tp(t+s,x,y)=\int \tp(t,x,z)\tp(s,z,y)\,dz,\quad t,s>0,\; x,y\in \RR_0^d. \end{align} Additionally, for $t>0$, we put $\tp(t,x,y) =0$, whenever $x=0$ or $y=0$. Moreover, $(\tilde P_t)_{t\ge0}$ also acts as a strongly continuous semigroup in $L^p(\RR_0^d;dx)$ for all $1\le p<\infty$, and, in $L^2(\RR_0^d;dx)$ the semigroup $(\tilde P_t)_{t\ge0}$ is self-adjoint. Meanwhile, $(\tilde P_t)_{t\ge0}$ is given via the Feynman-Kac formula: \begin{equation}\label{e:hk2} \begin{split} \tilde P_t f(x)=&\Ee^x \left( f\big(X^{\RR^d_0}_t\big)e^{\int_0^t q\big(X^{\RR^d_0}_s\big)\,ds}\right)\\ =&\Ee^x \left( f(X_t)e^{\int_0^t q(X_s)\,ds}\I_{\{t<\tau_{\RR_0^d}\}}\right),\quad f\in C_\infty(\RR_0^d),\; t>0,\; x\in \RR^d_0.\end{split} \end{equation} As mentioned above, $\Pp^x(\tau_{\RR_0^d} < \infty)=0$ for all $x\in \RR_0^d$, thus we have \begin{equation}\label{e:hk3}\tilde P_t f(x)=\Ee^x \left( f(X_t)e^{\int_0^t q(X_s)\,ds}\right),\quad f\in C_\infty(\RR_0^d),\; t>0,\;x\in \RR_0^d.\end{equation} Due to $q(x)<0$ on $\RR^d_0$ again, it follows from \eqref{e:hk1} and \eqref{e:hk2} that $$\tp(t,x,y) \le p(t,x,y),\quad t>0,\; x,y\in \RR^d. $$ Since $(\tilde P_t)_{t\ge0}$ is self-adjoint in $L^2(\RR_0^d;dx)$, $$\tp(t,x,y)=\tp(t,y,x),\quad t>0,\; x,y\in \RR^d.$$ According to \cite[Propositions 5.2 and 5.3]{DC} and their proofs, $\tp(t,x,y)$ will satisfy the following Duhamel's formula: \begin{equation}\label{e:for} \begin{split}\tp(t,x,y) &= p(t,x,y) + \int_0^t \int_{\RR^d} p(t-s,x,z)q(z)\tp(s,z,y)\, dz\,ds\\ &=p(t,x,y)+\int_0^t \int_{\RR^d} \tp(t-s,x,z)q(z)p(s,z,y)\, dz\,ds \end{split} \end{equation} for all $t>0$ and $x,y\in \RR_0^d$. Next, we show that $\tp(t,x,y)$ enjoys the same scaling property as $p(t,x,y)$. \begin{lemma}\label{L:sca} For any $t>0$ and $x,y\in \RR^d$, $$\tp (t,x,y)=t^{-d/\alpha} \tp(1,xt^{-1/\alpha},yt^{-1/\alpha}).$$ \end{lemma} \begin{proof} We only consider the case that $x,y\in \RR_0^d$; otherwise, the statement holds trivially. Recall that for the symmetric $\alpha$-stable process $(X_t)_{t\ge0}$, the processes $(X_{ut})_{t\ge0}$ and $(t^{1/\alpha}X_u)_{t\ge0}$ enjoy the same law for any fixed $u>0$. For fixed $t>0$, set $\hat X_u=X_{ut}$ for $u\ge0$. Then, by \eqref{e:hk3}, for any $f\in C_\infty(\RR^d_0)$, $t>0$ and $x\in \RR_0^d$, \begin{align*}\tilde P_t f(x)=&\Ee^x \left( f(X_t)e^{\int_0^t q(X_s)\,ds}\right)=\Ee^x \left( f(\hat X_1)e^{\int_0^t q(\hat X_{s/t})\,ds}\right)= \Ee^x \left( f(\hat X_1)e^{t\int_0^1 q(\hat X_{u})\,du}\right)\\ =&\Ee^{t^{-1/\alpha}x} \left( f(t^{1/\alpha} X_1)e^{t\int_0^1 q( t^{1/\alpha}X_{u})\,du}\right)=\Ee^{t^{-1/\alpha}x} \left( f(t^{1/\alpha} X_1)e^{\int_0^1 q(X_{u})\,du}\right),\end{align*} where in the last equality we used the fact that $q(x)=\kappa_\delta |x|^{-\alpha}.$ Hence, the desired assertion follows from the equality above. \end{proof} \subsection{Integral analysis for fractional Laplacian Schr\"{o}dinger operator} \begin{lemma} Let $\beta \in (0,2)$. Then, \begin{align}\label{e:pos} |x|^\beta = 2^{d+\beta} \pi^{d/2} \frac{\Gamma\left(\frac{d+\beta}{2}\right)}{\Gamma\left(\frac{d+\beta}{\alpha}\right) \left|\Gamma\left(\frac{-\beta}{2}\right)\right|} \int_0^\infty [p(t,0) - p(t,x)]t^{\frac{d-\alpha+\beta}{\alpha}} dt\, \end{align} where $\Gamma(-\beta/2)=\int_0^\infty t^{-1-\beta/2}(e^{-t}-1)\,dt$. \end{lemma} \begin{proof} We follow the method used in the proof of \cite[Proposition 5]{2016-KB-BD-PK-pa}. First, let $\eta_t(s)$ be the density function of the distribution of the $\alpha/2$-stable subordinator at time $t$. Let $g_t(x) = (4\pi t)^{-d/2} e^{-|x|^2/(4t)}$. Then, \begin{align*} p(t,x) = \int_0^\infty g_s(x) \eta_t(s)\, ds. \end{align*} By \cite[(24)]{2016-KB-BD-PK-pa}, for $\gamma < d/2-1$, \begin{align*} \int_0^\infty g_s(x) s^{\gamma} \,ds = 4^{-\gamma-1}\pi^{-d/2} \Gamma(d/2-\gamma-1) |x|^{2\gamma-d+2}. \end{align*} Then, by integrating by parts, for $d/2-1 < \gamma < d/2$, we get \begin{align*} \int_0^\infty (g_s(0)-g_s(x)) s^{\gamma} \,ds &= (4\pi)^{-d/2} \int_0^\infty (1-e^{-\frac{|x|^2}{4s}}) s^{\gamma-d/2} \,ds \\ &= \frac{(4\pi)^{-d/2}}{\gamma+1-d/2} \int_0^\infty \frac{|x|^2}{4s^2} e^{-\frac{|x|^2}{4s}} s^{\gamma+1-d/2} \,ds \\ &= \frac{|x|^2}{4(\gamma+1-d/2)} \int_0^\infty g_s(x) s^{\gamma-1} \,ds \\ &= 4^{-\gamma-1}\pi^{-d/2} \frac{\Gamma(d/2-\gamma)}{\gamma+1-d/2} |x|^{2\gamma-d+2}. \end{align*} Note that, for any $\gamma>-1$, \begin{align*} \int_0^\infty t^\gamma \eta_t(s)\, dt = \frac{\Gamma(\gamma+1)}{\Gamma(\frac{\alpha(\gamma+1)}{2})} s^{\frac{\alpha(\gamma+1)}{2}-1}, \end{align*} see \cite[(23)]{2016-KB-BD-PK-pa}. (Note that the condition that $\gamma<d/\alpha -1$ is not required in the proof of \cite[(23)]{2016-KB-BD-PK-pa}.) We further obtain \begin{align*} \int_0^\infty [p(t,0) - p(t,x)] t^{\frac{d-\alpha+\beta}{\alpha}} \,dt &=\int_0^\infty \int_0^\infty[g_s(0) - g_s(x)] \eta_t(s) t^{\frac{d-\alpha+\beta}{\alpha}}\, dt\, ds \\ &=\frac{\Gamma(\frac{d+\beta}{\alpha})}{\Gamma(\frac{d+\beta}{2})}\int_0^\infty[g_s(0) - g_s(x)] s^{\frac{d+\beta-2}{2}}\, ds \\ &= 2^{-d-\beta}\pi^{-d/2}\frac{\Gamma(\frac{d+\beta}{\alpha})}{\Gamma(\frac{d+\beta}{2})} \frac{\Gamma(\frac{2-\beta}{2})}{\frac{\beta}{2}} |x|^\beta. \end{align*} Since $\Gamma(\frac{2-\beta}{2}) = -\frac{\beta}{2} \Gamma(-\frac{\beta}{2})$, we get the assertion of the lemma. \end{proof} We recall from \cite[(25)]{2016-KB-BD-PK-pa} that for any $\beta \in (0,d)$, \begin{align}\label{e:neg} |x|^{-\beta} = 2^{d-\beta} \pi^{d/2} \frac{\Gamma\left(\frac{d-\beta}{2}\right)}{\Gamma\left(\frac{d-\beta}{\alpha}\right) \Gamma\left(\frac{\beta}{2}\right)} \int_0^\infty p(t,x)t^{\frac{d-\alpha-\beta}{\alpha}} \,dt, \quad x\in \RR^d. \end{align} Thus, \eqref{e:pos} may be treated as an extension of the formula \eqref{e:neg} to negative $\beta$. Note that in the proof of \eqref{e:pos} we have to use a compensated kernel $p(t,0)$ to ensure convergence of the integral involved. Now, let $\beta \in (0,\alpha)$. By \eqref{e:neg}, \begin{align*} |x|^{\beta-\alpha} = 2^{d+\beta-\alpha} \pi^{d/2} \frac{\Gamma\left(\frac{d+\beta-\alpha}{2}\right)}{\Gamma\left(\frac{d+\beta-\alpha}{\alpha}\right) \Gamma\left(\frac{\alpha-\beta}{2}\right)} \int_0^\infty p(t,x)t^{\frac{d-2\alpha+\beta}{\alpha}}\,dt. \end{align*} On the other hand, let $f(r) = c r^{(d-\alpha+\beta)/\alpha}$ with \begin{equation}\label{e:constant00}c=2^{d+\beta} \pi^{d/2} \frac{\Gamma\left(\frac{d+\beta}{2}\right)}{\Gamma\left(\frac{d+\beta}{\alpha}\right) \left|\Gamma\left(\frac{-\beta}{2}\right)\right|}.\end{equation} Then, according to \eqref{e:pos}, \begin{equation}\label{e:01} |x|^{\beta} = \int_0^\infty [p(t,0) -p(t,x)] f(t) \,dt. \end{equation} Combining two equations above together, we will find that \begin{align*} -\kappa_\beta |x|^{-\alpha} = \frac{\int_0^\infty p(r,x) f'(r) \,dr}{\int_0^\infty [p(r,0) -p(r,x)] f(r) \,dr}, \end{align*} where $$ \kappa_\beta = \frac{2^\alpha\Gamma(\frac{\alpha-\beta}{2}) \Gamma(\frac{d+\beta}{2})}{\Gamma(\frac{-\beta}{2})\Gamma(\frac{d+\beta-\alpha}{2})}. $$ In particular, \begin{align}\label{e:02} -\kappa_\beta |x|^{\beta-\alpha} = \int_0^\infty p(r,x) f'(r) \,dr. \end{align} We note that $\kappa_\beta <0$ for any $\beta\in(0,\alpha)$ and $\lim_{\beta \to \alpha} \kappa_\beta = -\infty$. For convenience, let $\kappa_0=0.$ Moreover, write $$\kappa_\beta=-\frac{2^\alpha\Gamma(\frac{\alpha-\beta}{2}) \Gamma(\frac{d+\beta}{2})\frac{\beta}{2}}{\Gamma(\frac{2-\beta}{2})\Gamma(\frac{d+\beta-\alpha}{2})},$$ and let $$r(t)=\frac{\Gamma(\frac{\alpha}{2}-t)\Gamma(\frac{d}{2}+t)}{\Gamma(\frac{d-\alpha}{2}+t)\Gamma(1-t)},\quad 0<t<\alpha/2.$$ Then, using the formula $$ \frac{\Gamma'(x)}{\Gamma(x)}=-\gamma-\sum_{k=0}^\infty\left(\frac{1}{x+k}-\frac{1}{1+k}\right),\quad x>0$$ with the Euler-Mascheroni constant $\gamma$ (see \cite[(1.2.13)]{AAR}) and following the argument in the end of the proof for \cite[Proposition 5]{2016-KB-BD-PK-pa}, one can check that $r(t)$ is strictly increasing on $(0,\alpha/2)$, and so $\beta\mapsto \kappa_\beta$ is strictly decreasing on $(0,\alpha)$. \begin{lemma}\label{L:2.3} For $\beta \in (0,\alpha)$, $t>0$ and $x\in\RR^d$, we have \begin{equation}\label{e:03} \int_{\RR^d} p(t,x,y)|y|^{\beta}\, dy = |x|^{\beta} - \kappa_\beta\int_0^t \int_{\RR^d} p(s,x,y) |y|^{\beta-\alpha} \,dy \,ds \end{equation} \end{lemma} \begin{proof}Let $f(r) = c r^{(d-\alpha+\beta)/\alpha}$ with the constant $c$ given by \eqref{e:constant00}. By \eqref{e:02} and \eqref{e:01}, for any $t>0$ and $x\in \RR^d$, \begin{align*} -\kappa_\beta\int_0^t \int_{\RR^d} p(s,x,y) |y|^{\beta-\alpha} \,dy \,ds &= \int_0^t \int_0^\infty p(s+r,x) f'(r)\, dr\, ds\\ &= -\int_0^t \int_0^\infty \frac{\partial}{\partial s} p(s+r,x) f(r)\, dr\, ds\\ &= \int_0^\infty [ p(r,x)-p(t+r,x)] f(r) \,dr \\ &= \int_0^\infty [ p(r,x)- p(r,0) + p(r,0)\! -\!p(t+r,x)] f(r)\, dr \\ &= -|x|^{\beta} \!+\!\! \int_0^\infty\! \!\int_{\RR^d} \!p(t,x,y)[p(r,0)\!-\!p(r,y)]f(r) \,dy\, dr\\ &= -|x|^{\beta} + \int_{\RR^d} p(t,x,y) |y|^{\beta}\,dy, \end{align*} where in the second equality we used the fact that $$ \lim_{r \to \infty}p(s+r,x) f(r)\le c_1\lim_{r\to\infty} (s+r)^{-d/\alpha} f(r)=0.$$ This completes the proof. \end{proof} Set $h_\beta(x)=|x|^\beta$. Letting $t\to0$ in \eqref{e:03}, informally it holds that \begin{equation}\label{e:eee}(\Delta^{\alpha/2}+\kappa_\beta|x|^{-\alpha}) h_\beta(x)=0\end{equation} for all $x\in \RR^d$. That is, the function $h_\beta$ is harmonic with respect to the operator $\Delta^{\alpha/2}+\kappa_\beta|x|^{-\alpha}.$ \emph{From now, we will fix $\delta\in (0,\alpha)$, and write $\kappa_\delta$ as $\kappa$ for simplicity.} The following theorem is an analog of \cite[Theorem 3.1]{2017-KB-TG-TJ-DP}. Since there is no problem with convergence of the integrals involved, the proof is much simpler than that of \cite[Theorem 3.1]{2017-KB-TG-TJ-DP}. \begin{theorem}\label{thm:0} For $\beta \in (0, \alpha)$, $t>0$ and $x\in \RR^d$, we have \begin{align}\label{eq:0} \int_{\RR^d} \tp(t,x,y) |y|^{\beta}\, dy = |x|^{\beta} + (\kappa - \kappa_\beta) \int_0^t \int_{\RR^d} \tp(s,x,y) |y|^{\beta-\alpha}\,dy\,ds. \end{align} In particular, for any $t>0$ and $x\in \RR^d$, \begin{align}\label{eq:00} \int_{\RR^d} \tp(t,x,y) |y|^{\delta}\, dy = |x|^{\delta}. \end{align} \end{theorem} \begin{proof} When $x=0$, both sides of \eqref{eq:0} and \eqref{eq:00} are equal to zero, since $\tp(t,0,y)=0$ for all $t>0$ and $y\in \RR^d$. Below, we consider the case that $x\in \RR_0^d$. By \eqref{e:for} and \eqref{e:03}, for any $t>0$ and $x\in \RR_0^d$, \begin{align*} & -\kappa_\beta\int_0^t \int_{\RR^d} \tp(s,x,y) |y|^{\beta-\alpha}\, dy \,ds \\ &= -\kappa_\beta\int_0^t \int_{\RR^d} p(s,x,y)|y|^{\beta-\alpha}\, dy\, ds \\ &\quad- \kappa_\beta\int_0^t \int_{\RR^d} \int_u^t \int_{\RR^d} \tp(u,x,z) q(z) p(s-u,z,y) |y|^{\beta-\alpha}\,dz\, du\, dy\, ds \\ &= -|x|^{\beta} + \int_{\RR^d} p(t,x,y) |y|^{\beta} \,dy\\ & \quad-\int_0^t \int_{\RR^d} \tp(u,x,z) q(z) \left(|z|^\beta -\int_{\RR^d} p(t-u,z,y) |y|^{ \beta} dy \right) \,dz\, du. \end{align*} Hence, according to \eqref{e:for} again, for any $t>0$ and $x\in \RR_0^d$, \begin{align*} -(\kappa_\beta - \kappa)\int_0^t \int_{\RR^d} \tp(s,x,y) |y|^{\beta-\alpha}\, dy\, ds &= -|x|^{\beta} + \int_{\RR^d} p(t,x,y) |y|^{\beta} \,dy \\ &\quad + \int_{\RR^d} (\tp(t,x,y)- p(t,x,y)) |y|^{\beta}\, dy \\ &= -|x|^{\beta} + \int_{\RR^d} \tp(t,x,y) |y|^{\beta} \,dy, \end{align*} which proves \eqref{eq:0}. Now, \eqref{eq:00} follows by taking $\beta=\delta$. The proof is complete. \end{proof} Note that \eqref{e:eee} implies that $$[\Delta^{\alpha/2}+\kappa |x|^{-\alpha}]h_\beta(x)=(\kappa-\kappa_\beta)|x|^{\beta-\alpha},\quad x\in \RR^d.$$ For this it is easy to obtain \eqref{eq:0} via the Feynman-Kac semigroup associated with $\Delta^{\alpha/2}+\kappa |x|^{-\alpha}.$ Although the following lemma is not used in the proofs, we state it as one of the results. From this lemma we see that the right-hand side of \eqref{eq:logx} behaves near $0$ as $-\log|x|$. \begin{lemma} For any $t>0$ and $x\in \RR^d$, it holds that \begin{align}\label{eq:logx} C\int_{\RR^d} \tp(t,x,y) |y|^{\delta} (\ln|y| - \ln|x|)\, dy = \int_0^t \int_{\RR^d} \tp(s,x,y) |y|^{\delta-\alpha}\,dy\,ds, \end{align} where $$C:=\lim_{\beta \to \delta} \frac{\delta-\beta}{\kappa_\beta-\kappa}>0.$$ \end{lemma} \begin{proof} By \eqref{eq:0}, \eqref{eq:00} and the dominated convergence theorem, \begin{align*} \int_0^t \int_{\RR^d} \tp(s,x,y) |y|^{\delta-\alpha}\,dy\,ds &= \lim_{\beta \to \delta} \frac{1}{\kappa_\beta-\kappa} \int_{\RR^d} \tp(t,x,y) |y|^{\delta} (|x|^{\beta-\delta} - |y|^{\beta-\delta})\, dy \\ &= C \int_{\RR^d} \tp(t,x,y) |y|^{\delta} (\ln|y| - \ln|x|) \,dy, \end{align*} proving the desired assertion. \end{proof} \section{Two-sided estimates and joint continuity of $\tp(t,x,y)$} \subsection{Upper bounds of $\tp(1,x,y)$} For any $t>0$ and $x\in \RR^d$, define $$ H(t,x) = \int_{\RR^d} \tp(t,x,y)\,dy. $$ Note that, by Lemma \ref{L:sca}, for all $t>0$ and $x\in \RR^d$, we have \begin{align}\label {eq:H_1} H(t,x) = \int_{\RR^d} t^{-d/\alpha}\tp(1,t^{-1/\alpha}x,t^{-1/\alpha}y)\, dy = H(1,t^{-1/\alpha}x). \end{align} On the other hand, by the fact $0 \le \tp(t,x,y)\le p(t,x,y)$ for any $t>0$ and $x,y\in \RR^d$, it also holds that \begin{align}\label{eq:H_2} 0 \le H(t,x) \le \int_{\RR^d} p(t,x,y)\, dy = 1,\quad t>0,\; x\in \RR^d. \end{align} \begin{prop}\label{lem:mass_est} There is a constant $C>0$ such that for all $x\in \RR^d$, $$ H(1,x) \le C(1 \land |x|^\delta). $$ \end{prop} \begin{proof} By the Chapman-Kolmogorov equation \eqref{eq:CKpt} (which holds true for all $x,y\in \RR^d$) and \eqref{eq:H_1}, for any $x,y\in \RR^d $, \begin{equation}\label{eq:tp13}\begin{split} \tp(1,x,y) &= \int_{\RR^d} \int_{\RR^d} \tp(1/3,x,z) \tp(1/3,z,w) \tp(1/3,w,y)\, dw \,dz \\ &\le \int_{\RR^d} \int_{\RR^d} \tp(1/3,x,z)\cdot c \cdot \tp(1/3,w,y)\, dw\, dz \\ &= c H(1/3,x) H(1/3,y) = c H(1,3^{1/\alpha}x) H(1,3^{1/\alpha}y), \end{split}\end{equation} where the constant $c$ comes from the estimate $\tp(1/3,x,y) \le p(1/3,x,y) \le c$. Denote by $|B(0,r)|$ the Lebesgue measure of $B(0,r)$. Fix $r>0$ small enough such that $\eta := c|B(0,r)|<3^{-\delta/\alpha}$. According to \eqref{eq:0}, \eqref{eq:tp13} and \eqref{eq:H_2}, for any $x\in \RR^d$, we have \begin{equation}\label{eq2:mass_est}\begin{split} H(1,x) & \le \int_{B(0,r)} \tp(1,x,y) \,dy + \frac{1}{r^{\delta}} \int_{B(0,r)^c} \tp(1,x,y) |y|^\delta \,dy \\ &\le \int_{B(0,r)} \tp(1,x,y) \, dy +M|x|^\delta\le \int_{B(0,r)} cH(1,3^{1/\alpha}x) \,dy + M|x|^\delta,\\ &= \eta H(1,3^{1/\alpha}x) + M|x|^\delta, \end{split}\end{equation} where $M ={r^{-\delta}}$. Now, we can iterate the inequality \eqref{eq2:mass_est} to obtain that for all $x\in \RR^d$, \begin{align*} H(1,x) &\le \eta H(1,3^{1/\alpha}x) + M|x|^\delta \\ &\le \eta \left[\eta H(1,3^{2/\alpha}x) + M|3^{1/\alpha}x|^\delta\right] + M|x|^\delta \\ &\le \eta^2 \left[ \eta H(1,3^{3/\alpha}x)+M |3^{2/\alpha}x|^\delta\right] + M(1+\eta 3^{\delta/\alpha})|x|^\delta\\ &\le \cdots\\ &\le \eta^n H(1,3^{n/\alpha}x) + M[1+\eta 3^{\delta/\alpha} + \cdots + (\eta 3^{\delta/\alpha})^{n-1}]|x|^\delta. \end{align*} By \eqref{eq:H_2}, taking $n \to \infty$ in the inequality above, we get that for any $x\in \RR^d$, \begin{align*} H(1,x) \le \frac{M}{1-\eta 3^{\delta/\alpha}} |x|^\delta, \end{align*} yielding the desired assertion. \end{proof} Applying Proposition \ref{lem:mass_est} to \eqref{eq:tp13}, we immediately get \begin{corollary}\label{eq:ub1} There is a constant $C>0$ such that $$ \tp(1,x,y) \le C (1 \land |x|^\delta) (1 \land |y|^\delta), \quad x,y \in \RR^d. $$ \end{corollary} Next, we further refine upper bounds for $\tp(t,x,y)$. \begin{lemma}\label{eq:1} For any $t >0$ and $x,y \in \RR^d$, we have $$ \int_{B(y,|x-y|/2)} p(t,x,z)p(t,z,y) \,dz \le \frac{p(2t,x,y)}{2}. $$ \end{lemma} \begin{proof} Fix $t>0$ and $x,y\in \RR^d$. By symmetry, \begin{align*} \int_{B(y,|x-y|/2)} p(t,x,z)p(t,z,y) \,dz = \int_{B(x,|x-y|/2)} p(t,x,z)p(t,z,y)\, dz. \end{align*} Hence, by \eqref{eq:CKp}, \begin{align*} &2\int_{B(y,|x-y|/2)} p(t,x,z)p(t,z,y)\, dz\\ &= \int_{B(y,|x-y|/2)} p(t,x,z)p(t,z,y)\, dz + \int_{B(x,|x-y|/2)} p(t,x,z)p(t,z,y) \,dz\\ &\le \int_{\RR^d} p(t,x,z)p(t,z,y)\, dz = p(2t,x,y). \end{align*} This completes the proof. \end{proof} \begin{lemma}\label{lem:intsplit} There exists a constant $M>0$ such that for any $t > 0$ and $x,y \in \RR^d$, we have $$ \tp(t,x,y) \le \int_{B(y,|x-y|/2)} \tp(t/2,x,z)\tp(t/2,z,y) \,dz + M h(t,x) p(t,x,y), $$ where $h(t,x) = t^{-\delta/\alpha}|x|^\delta$. \end{lemma} \begin{proof} By \eqref{eq:CKpt}, for any $t>0$ and $x,y\in \RR^d$, we have \begin{align*} \tp(t,x,y) &= \int_{B(y,|x-y|/2)} \tp(t/2,x,z)\tp(t/2,z,y) \,dz \\ &\quad+ \int_{B(y,|x-y|/2)^c} \tp(t/2,x,z)\tp(t/2,z,y) \,dz. \end{align*} Note that, according to \eqref{eq:H_1} and Proposition \ref{lem:mass_est}, \begin{align*} \int_{\RR^d} \tp(t,x,y) \,dy =H(t,x)=H(1,t^{-1/\alpha}x)\le c h(t,x),\quad t>0,\;x\in \RR^d. \end{align*} For $t>0$ and $x,y,z\in \RR^d$ with $z \in B(y,|x-y|/2)^c$, we have $$\tp(t/2,z,y) \le p(t/2,z,y) \le c_1 p(t,x,y).$$ Hence, for any $t>0$ and $x,y\in \RR^d$, \begin{align*} \int_{B(y,|x-y|/2)^c} \tp(t/2,x,z)\tp(t/2,z,y)\, dz &\le c_1 p(t,x,y)\int_{B(y,|x-y|/2)^c} \tp(t/2,x,z)\, dz \\ &\le c c_1 h(t/2,x) p(t,x,y)\\ & \le M h(t,x) p(t,x,y), \end{align*} thus we get the assertion of the lemma. \end{proof} \begin{theorem}\label{thm:UB}{\bf (Upper bounds)} There is a constant $C>0$ such that for all $x, y \in \RR^d$, \begin{align}\label{EQ:UB} \tp(1,x,y) \le C (1\land |x|^\delta)(1\land |y|^\delta) p(1,x,y). \end{align} \end{theorem} \begin{proof} Let $\eta = 1/2$ and $\nu = 2^{(\delta-\alpha)/\alpha} <1$. As in Lemma \ref{lem:intsplit}, denote $h(t,x) = t^{-\delta/\alpha} |x|^\delta$. Note that \begin{align}\label{eq0:UB} \eta h(t/2,x) = \tfrac{1}{2} |(t/2)^{-1/\alpha}x|^\delta = 2^{(\delta-\alpha)/\alpha} |t^{-1/\alpha}x|^\delta = \nu h(t,x), \quad t>0,\, x\in\RR^d. \end{align} Let $M$ be the constant from Lemma \ref{lem:intsplit}. We will claim that for $n \ge 0$, \begin{align}\label{e:proof} \tp(t,x,y) \le [\eta^{n+1} + (1+\nu+\ldots+\nu^n)M h(t,x)] p(t,x,y), \quad t>0, \; x,y\in \RR^d. \end{align} Indeed, for $t\in(0,1]$ and $x,y\in \RR^d$, by Lemmas \ref{eq:1} and \ref{lem:intsplit}, \begin{align*} \tp(t,x,y) \le [\eta + M h(t,x)]p(t,x,y), \end{align*} where we used the fact $\tp(t,x,y)\le p(t,x,y)$ for any $t>0$ and $x,y\in \RR^d$. Next, we use induction. Suppose that \begin{align*} \tp(t,x,y) \le [\eta^{n} + (1+\nu+\ldots+\nu^{n-1})M h(t,x)] p(t,x,y),\quad t>0, \; x,y\in \RR^d. \end{align*} Then, for any $t>0$ and $x,y\in \RR^d$, by Lemma \ref{lem:intsplit}, Lemma \ref{eq:1} and \eqref{eq0:UB}, \begin{align*} \tp(t,x,y) &\le \int_{B(y,|x-y|/2)} \tp(t/2,x,z)p(t/2,z,y)\, dz + M h(t,x)p(t,x,y) \\ &\le \int_{B(y,|x-y|/2)} \!\![\eta^{n} + (1+\nu+\ldots+\nu^{n-1})M h(t/2,x)] p(t/2,x,z)p(t/2,z,y) \,dz \\ &\quad + M h(t,x)p(t,x,y) \\ &\le [\eta^{n} + (1+\nu+\ldots+\nu^{n-1})M h(t/2,x)] \eta p(t,x,y) + M h(t,x)p(t,x,y) \\ &\le [\eta^{n+1} + (\nu+\ldots+\nu^{n})M h(t,x)] p(t,x,y) + M h(t,x)p(t,x,y) \\ &= [\eta^{n+1} + (1+\nu+\ldots+\nu^{n})M h(t,x)] p(t,x,y), \end{align*} and \eqref{e:proof} follows. Since $h(1,x) = |x|^\delta$, by letting $n$ to infinity in \eqref{e:proof}, we get \begin{align}\label{e:proof1} \tp(1,x,y) \le \frac{M}{1-\nu} |x|^\delta p(1,x,y), \quad x,y\in \RR^d. \end{align} In the following, we pass to the proof of \eqref{EQ:UB}. By symmetry, we may and do assume that $|x| \le |y|$. For $x,y\in \RR^d$ with $|y|\ge |x| \ge 1$, \eqref{EQ:UB} follows by the estimate $\tp(1,x,y) \le p(1,x,y)$. For $|x|\le |y|\le1$, we use Corollary \ref{eq:ub1} and the estimate that $p(1,x,y) \ge c$. Finally, for $|x| <1 \le |y|$, \eqref{EQ:UB} follows by \eqref{e:proof1}. \end{proof} \subsection{Lower bounds of $\tp(1,x,y)$} We first begin with the following lemma, which is a consequence of Theorem \ref{thm:UB}. \begin{lemma} There is a constant $C>0$ such that \begin{align*} H(1,x) \ge C(1 \land |x|^\delta), \quad x \in \RR^d. \end{align*} \end{lemma} \begin{proof} Let $R>0$ and $x \in B(0,R/2)$. Then \begin{align*} \int_{B(0,R)^c} |y|^\delta p(1,x,y) \,dy \le c_1\int_{B(0,R)^c} \frac{|y|^\delta}{|y|^{d+\alpha}}\, dy = c_2 R^{\delta-\alpha} \to 0 \quad \mbox{as } R \to \infty. \end{align*} Choose $R\ge 1$ be such that $c_2 C R^{\delta-\alpha} \le 1/2$, where $C$ is the constant given in \eqref{EQ:UB}. Then, by \eqref{eq:0}, for $r \ge R\ge1$ and $x \in B(0,r/2)$, we have \begin{equation}\label{e:lower} \begin{split} \int_{\RR^d} \tp(1,x,y) \,dy &\ge r^{-\delta} \int_{B(0,r)} \tp(1,x,y) |y|^\delta \,dy \\ & = r^{-\delta} \left(|x|^\delta - \int_{B(0,r)^c} \tp(1,x,y) |y|^\delta \,dy \right) \\ & \ge r^{-\delta} \left(|x|^\delta - C\int_{B(0,r)^c} |x|^\delta p(1,x,y) |y|^\delta \,dy \right) \ge \frac{|x|^\delta}{2r^\delta}. \end{split}\end{equation} Hence, for $x \in B(0,R/2)$, by \eqref{e:lower}, \begin{align*} \int_{\RR^d} \tp(1,x,y)\, dy \ge \frac{|x|^\delta}{2R^\delta} \ge \frac{|x|^\delta \land 1}{2R^\delta}; \end{align*} for $x\in B(0,R/2)^c$, taking $r = 2|x|+1$ in \eqref{e:lower}, we can get that \begin{align*} \int_{\RR^d} \tp(1,x,y) dy \ge \frac{|x|^\delta}{2(2|x|+1)^\delta}\ge \frac{|x|^\delta}{2(4|x|)^\delta}\ge \frac{1}{4^{\delta+1}}. \end{align*} Combining both estimates above, we can prove the desired assertion. \end{proof} To obtain lower bounds of $\tp(t,x,y)$, we need to consider the difference between $p(t,x,y)$ and $\tp(t,x,y)$. Motivated by Duhamel's formula \eqref{e:for}, we define \begin{align*} p_1(t,x,y) = \int_0^t\int_{\RR^d}p(t-s,x,z) |z|^{-\alpha} p(s,z,y) \,dz\,ds,\quad t>0,\; x,y\in \RR_0^d. \end{align*} It is easy to see that $p_1(t,x,y)$ also enjoys the same scaling property as $p(t,x,y)$, i.e., \begin{align}\label{eq:scalp1} p_1(t,x,y)=t^{-d/\alpha} p_1(1,xt^{-1/\alpha},yt^{-1/\alpha}),\quad t>0,\; x,y\in \RR_0^d. \end{align} Let \begin{align*} G(t,x) = \int_0^t\int_{\RR^d}p(s,x,z) |z|^{-\alpha}\,dz\,ds, \quad t>0,\, x \in \RR_0^d. \end{align*} By \cite[Lemma 2.3]{2017-KB-TG-TJ-DP}, \begin{align}\label{eq:Gest} G(t,x) \approx \log(1 + t|x|^{-\alpha}), \quad t>0, \, x \in \RR_0^d. \end{align} \begin{lemma} \label{lem:p1est} For all $x,y \in \RR_0^d$, we have $$p_1(1,x,y) \approx[G(1,x) + G(1,y)]\, p(1,x,y).$$ \end{lemma} \begin{proof} By the 3P inequality (see \cite[(9)]{BJ} or \cite[(2.11)]{CKS2}), for any $x,y\in \RR_0^d$, \begin{align*}&\int_0^1\int_{\RR^d} p(1-s,x,z)|z|^{-\alpha}p(s,z,y)\,dz\,ds\\ &\le c_1p(1,x,y) \int_0^1\int_{\RR^d}(p(1-s,x,z)+p(s,z,y)) |z|^{-\alpha}\,dz\,ds \\ & = c_1 (G(1,x) + G(1,y) ) p(1,x,y)\,, \end{align*} thus we get the upper bound. Now, we pass to the lower bound. Since the function $|x|\mapsto\log(1 + |x|^{-\alpha})$ is decreasing in $|x|$, by \eqref{eq:Gest} and the symmetry of $p(t,x,y)$ and $p_1(t,x,y)$, it suffices to prove \begin{align*} p_1(1,x,y) \ge c_2 G(1,x) p(1,x,y), \quad 0 < |x| \le |y|. \end{align*} First, let $|y|>4$ and $|x|<2$. Then, by \eqref{eq:pest}, \begin{align*} \int_0^{1/2} \int_{B(0,3)} p(s,x,z) \frac{1}{|z|^\alpha}\, dz \, ds &\ge \int_0^{1/2} \int_{B(x,s^{1/\alpha})} p(s,x,z) \frac{1}{|z|^\alpha}\, dz \, ds \\ &\ge c_0 \ge c_0 \int_0^{1/2} \int_{B(0,3)^c} p(s,x,z) \frac{1}{|z|^\alpha}\, dz\, ds \end{align*} and so $$\int_0^{1/2} \int_{\RR^d} p(s,x,z) \frac{1}{|z|^\alpha}\, dz \, ds \approx \int_0^{1/2} \int_{B(0,3)} p(s,x,z) \frac{1}{|z|^\alpha}\, dz \, ds .$$ Hence, \begin{equation}\label{eq1:p1est}\begin{split} p_1(1,x,y) &\ge \int_0^{1/2} \int_{B(0,3)} p(s,x,z) \frac{1}{|z|^\alpha} p(1-s,z,y)\, dz\, ds\\ &\approx \int_0^{1/2} \int_{B(0,3)} p(s,x,z) \frac{1}{|z|^\alpha} p(1,x,y) \,dz \,ds\\ &\approx \int_0^{1/2} \int_{\RR^d} p(s,x,z) \frac{1}{|z|^\alpha} p(1,x,y)\, dz \,ds\\ & = G(1/2,x) p(1,x,y)\approx G(1,x) p(1,x,y), \end{split}\end{equation} where in the second step we used the fact that $$p(1-s,z,y)\approx p(1,x,y),\quad 0<s\le 1/2, |x|<2, |z|\le 3, |y|>4.$$ Next, suppose that $|x| \le 1$ and $|x| \le |y| \le 4$. Then, $p(1,x,y) \approx c$. Note that \begin{align*} \int_{\RR^d} p_1(t,x,z) p(r,z,y)\,dz &= \int_{\RR^d} \int_0^t\int_{\RR^d}p(s,x,w) |w|^{-\alpha} p(t-s,w,z) p(r,z,y) \,dw\,ds\, dz \\ &= \int_0^t\int_{\RR^d}p(s,x,w) |w|^{-\alpha} p(t+r-s,w,y) \,dw\,ds \\ & \le p_1(t+r,x,y)\,. \end{align*} Hence, by the scaling property of $p_1(t,x,y)$ and \eqref{eq1:p1est}, \begin{align*} p_1(1,x,y) &\ge \int_{B(0,8)^c} p_1(1/2^\alpha,x,z) p(1-1/2^\alpha,z,y)\,dz\\ &= \int_{B(0,8)^c} 2^d p_1(1,2x,2z) p(1-1/2^\alpha,z,y)\,dz\\ & \ge c_3 G(1,2x) \int_{B(0,8)^c} p(1,2x,2z) p(1-1/2^\alpha,z,y)\,dz\\ &\ge c_4 G(1,x) \ge c_5 G(1,x) p(1,x,y), \end{align*} where in the third inequality we used the fact that \begin{align*}\int_{B(0,8)^c} p(1,2x,2z) p(1-1/2^\alpha,z,y)\,dz\ge &c_6\int_{B(0,8)^c} \frac{1}{|z-x|^{d+\alpha} |z-y|^{d+\alpha}}\,dz\\ \ge &c_7\int_{B(0,8)^c}\frac{1}{|z|^{2d+2\alpha}}\,dz \ge c_8.\end{align*} At last, suppose that $1 \le |x| \le |y|$. Then $G(1,x) \approx |x|^{-\alpha}$. Hence, \begin{align*} p_1(1,x,y) &\ge \int_0^{1/2} \int_{B(|x|,1/2)} p(s,x,z) \frac{1}{|z|^\alpha}p(1-s,z,y)\,dz\,ds\\ &\ge c_9\int_0^{1/2} \int_{B(|x|,1/2)} p(s,x,z) \frac{1}{|x|^\alpha}p(1-s,z,y)\,dz \,ds\\ &\ge c_{10} G(1,x) p(1,x,y), \end{align*} where the last inequality follows from the facts that $$p(1-s,z,y)\approx p(1,x,y),\quad 0<s<1/2, z\in B(x,1/2), 1 \le |x| \le |y|$$ and $$\int_0^{1/2} \int_{B(|x|,1/2)} p(s,x,z) \,dz\,ds \ge c_{11}.$$ The proof is complete. \end{proof} The following estimate is generally well known (see e.g. \cite[Section 6]{2008-KB-WH-TJ-sm} for further background). \begin{lemma}\label{lem:tpexp} For all $t>0$ and $x,y \in \RR^d_0$, we have \begin{align*} \tp(t,x,y) \ge p(t,x,y) \exp\left[\kappa\, \frac{p_1(t,x,y)}{p(t,x,y)}\right]. \end{align*} \end{lemma} \begin{proof} Since the proof is a little long, we will postpone it to the appendix. \end{proof} We note that the estimate in Lemma \ref{lem:tpexp} is not sharp. More precisely, one may show that $\lim\limits_{s \to 0^+} \frac{p_1(1,sx,y)}{-\ln (s|x|) p(1,sx,y)} = \frac{\Gamma(\frac{d-\alpha}{2})}{2^{\alpha-1}\Gamma(\frac{\alpha}{2})\Gamma(\frac{d}{2})}$. Hence, by \eqref{EQ:main}, for fixed $y \not=0$, $\tp(1,x,y) e^{-\kappa\, \frac{p_1(1,x,y)}{p(1,x,y)}} \to \infty$ as $x \to 0$. However, we still can get the following useful estimate. \begin{corollary}\label{cor:lowerb} There are constants $c,\gamma>0$ such that for all $t>0$ and $x,y \in \RR^d_0$, we have \begin{align}\label{eq:lowerb} \tp(t,x,y) \ge c \left[1 \land (t^{-1/\alpha}|x|)^\gamma\right]\left[1 \land (t^{-1/\alpha}|y|)^\gamma\right] p(t,x,y) \end{align} \end{corollary} \begin{proof} Lemmas \ref{lem:tpexp} and \ref{lem:p1est} along with \eqref{eq:Gest} yield that for any $x,y\in \RR_0^d$, \begin{align*} \tp(t,x,y) &\ge p(t,x,y) \exp\left[\kappa\, \frac{p_1(t,x,y)}{p(t,x,y)}\right]\\ &\ge p(t,x,y) \exp\left[-c (G(t,x) + G(t,y))\right]\\ & \ge p(t,x,y) \exp\left[- C (\log(1 + t|x|^{-\alpha}) + \log(1 + t|y|^{-\alpha})) \right]\\ & = p(t,x,y) \left[(1 + t|x|^{-\alpha})^{-C}\right] \left[(1 + t|y|^{-\alpha})^{-C} \right]\\ & \ge p(t,x,y) \left[2^{-C}(1 \vee t|x|^{-\alpha})^{-C}\right] \left[2^{-C}(1 \vee t|y|^{-\alpha})^{-C} \right], \end{align*} thus we get \eqref{eq:lowerb} with $\gamma = \alpha C$ and $c = 4^{-C}$. \end{proof} \begin{lemma}\label{lem:lowerCase1} For any $r>0$, there is a constant $C_r>0$ such that for all $x,y\in \RR^d$ with $|x|\land |y|\ge r$, \begin{align*} \tp(1,x,y) \ge C_r p(1,x,y). \end{align*} \end{lemma} \begin{proof} For $r>0$ and $x,y\in \RR^d$ with $|x| \land |y| \ge r$, by \eqref{eq:lowerb}, we get \begin{align*} \tp(1,x,y) \ge c (1 \land r)^{2\gamma} p(1,x,y), \end{align*} where $c$ and $\gamma$ are the constants from Corollary \ref{cor:lowerb}. \end{proof} \begin{lemma}\label{lem:lowerCase2} For any $R>0$, there is a constant $C_R>0$ such that for any $x,y\in \RR^d$ with $|x| \vee |y| \le R$, $$ \tp(1,x,y)\ge C_R |x|^\delta |y|^\delta. $$ \end{lemma} \begin{proof} By \eqref{e:lower}, there exists a constant $R_0\ge 2 \cdot 3^{1/\alpha}$ large enough such that for all $x\in B(0,R_0/2)$, $$ \int_{B(0,R_0)} \tp(1,x,y)\,dy\ge R_0^{-\delta}\int_{B(0,R_0)}\tp(1,x,y) |y|^\delta\,dy \ge \frac{|x|^\delta}{2 R_0^\delta}. $$ On the other hand, by \eqref{EQ:UB}, for all $r_0>0$ and $x\in \RR^d$, $$ \int_{B(0,r_0)}\tp (1,x,y)\,dy\le C_1 |x|^\delta \int_{B(0,r_0)}p(1,x,y)|y|^\delta\,dy\le C_1|x|^\delta r_0^\delta. $$ We take $r_0=\frac{1}{(4C_1)^{1/\delta} R_0}$. For $0<a<b<\infty$, let $D(a,b) = B(0,b) \setminus B(0,a)$. Then, for $|x| < R_0/2$, \begin{align}\label{eq:lowerCase2} \int_{D(r_0,R_0)} \tp(1,x,z) dz \ge |x|^\delta\left(\frac{1}{2R_0^\delta} - C_1 r_0^\delta\right) = \frac{|x|^{\delta}}{4R_0^\delta}. \end{align} Therefore, by \eqref{eq:CKpt} and \eqref{eq:lowerCase2}, for all $x,y\in \RR^d$ with $|x|\vee |y|\le R_0/2$, \begin{align*}\tp(3,x,y)&\ge \int_{D(r_0,R_0)}\int_{D(r_0,R_0)}\tp (1, x,z) \tp(1,z,w)\tp(1,w,y)\,dz\,dw\\ &\ge \frac{|x|^\delta |y|^\delta}{16 R_0^{2\delta}} \inf_{z,w\in D(r_0, R_0)}\tp(1,z,w). \end{align*} Next, by Lemma \ref{lem:lowerCase1}, \begin{equation}\label{e:rem} \inf_{z,w\in D(r_0,R_0)}\tp(1,z,w) \ge C_{r_0}\inf_{z,w\in D(r_0,R_0)} p(1,z,w) \ge \frac{c C_{r_0}}{(2R_0)^{d+\alpha}} >0, \end{equation} where $C_{r_0}>0$ is a constant given in Lemma \ref{lem:lowerCase1}. Hence, $$\tp(3,x,y) \ge c_0 |x|^\delta |y|^\delta,\quad |x| \vee |y| < R_0/2.$$ Now, by the scaling property of $\tp$, we obtain \begin{align*} \tp(1,x,y) = 3^{-d/\alpha} \tp(3,3^{-1/\alpha}x, 3^{-1/\alpha}y) \ge c_0 3^{-(d+2\delta)/\alpha} |x|^\delta |y|^\delta, \quad |x| \vee |y| \le \frac{R_0}{2 \cdot 3^{1/\alpha}}. \end{align*} This completes the proof. \end{proof} \begin{remark}Instead of applying Lemma \ref{lem:lowerCase1}, we can make use of the Feynman-Kac formula \eqref{e:hk2} for the semigroup $(\tilde P_t)_{t\ge0}$ and Dirichlet heat kernel estimates for fractional Laplacian obtained in \cite{CKS} to achieve \eqref{e:rem}. \end{remark} \begin{theorem}\label{Thm:LB}{\bf (Lower bounds)}\,\, There is a constant $C>0$ such that for all $x,y \in \RR^d$, \begin{align*} \tp(1,x,y) \ge C(1\land |x|^\delta)(1\land |y|^\delta) p(1,x,y). \end{align*} \end{theorem} \begin{proof} By symmetry, we will consider only $|x| \le |y|$. For $|w_1| \le 1/4$, $|w_2| > 1$ and $1/4 \le |z| \le 1/2$, by Lemmas \ref{lem:lowerCase2} and \ref{lem:lowerCase1}, we have $$ \tp(1,w_1,z) \ge c_1 |w_1|^\delta $$ and $$\tp(1,w_2,z) \ge c_2 p(1,w_2,z)\approx p(2,w_2,z)\approx p(2,w_2,w_1).$$ Hence, for any $|w_1|\le 1/4$ and $|w_2|>1$, \begin{equation}\label{EQ:lower1}\begin{split} \tp(2,w_1,w_2) &\ge \int_{B(0,1/2) \setminus B(0,1/4)} \tp(1,w_1,z) \tp(1,z,w_2) \,dz \ge c |w_1|^\delta p(2,w_1,w_2). \end{split}\end{equation} Therefore, for $|x| \le 2^{-1/\alpha}/4$ and $|y| > 2^{-1/\alpha}$, by \eqref{EQ:lower1}, \begin{align*} \tp(1,x,y) = &2^{d/\alpha} \tp(2,2^{1/\alpha}x, 2^{1/\alpha}y) \ge c 2^{d/\alpha} |2^{1/\alpha}x|^\delta p(2,2^{1/\alpha}x,2^{1/\alpha}y)\\ = & c 2^{\delta/\alpha} |x|^\delta p(1,x,y). \end{align*} Next, for $|x| \land |y| \ge 2^{-1/\alpha}/4$, we use Lemma \ref{lem:lowerCase1}. Finally, for $|x| \vee |y| \le 2^{-1/\alpha}$, we apply Lemma \ref{lem:lowerCase2}. \end{proof} Two-sided estimates for $\tp(t,x,y)$ stated in Theorem \ref{th:main} is a direct consequence of the scaling property of $\tp(t,x,y)$ and Theorems \ref{thm:UB} and \ref{Thm:LB}. \subsection{Joint continuity of $\tp(t,x,y)$} To prove the joint continuity of $\tp(t,x,y)$, we just follow the same argument of \cite[Subsection 4.3]{2016-KB-BD-PK-pa}. For the sake of completeness, we present the proof here. \begin{lemma}\label{L:con1}For any fixed $x\in \RR_0^d$, the function $\RR^d_0\ni y \mapsto \tp(t,x,y)$ is continuous. \end{lemma} \begin{proof} Fix $x,y,z\in \RR_0^d$ with $z\to y$. Then, by \eqref{e:for}, \begin{align*}\tp (1,x,y)-\tp (1,x,z)=&p(1,x,y)-p(1,x,z)\\ &+\int_0^1\int_{\RR^d} \tp(1-s,x,w)q(w)(p(s,w,y)-p(s,w,z))\,dw\,ds. \end{align*} For any $\varepsilon>0$ small enough, by \eqref{eq:Gest}, \begin{equation}\label{e:con1}\begin{split}&-\int_0^\varepsilon \int_{\RR^d} \tp(1-s,x,w)q(w)p(s,w,y)\,dw\,ds\\ &\le -\int_0^\varepsilon \int_{\RR^d} p(1-s,x,w)q(w)p(s,w,y)\,dw\,ds\\ &\le-c_1\int_0^\varepsilon \int_{\RR^d} p(s,w,y)q(w)\,dw\,ds\\ &= -\kappa c_1 G(\varepsilon,y) \le c_2 \varepsilon |y|^{-\alpha}. \end{split}\end{equation} Similarly, we have \begin{align*}-\int_0^\varepsilon \int_{\RR^d} \tp(1-s,x,w)q(w)p(s,w,z)\,dw\,ds\le c_2\varepsilon |z|^{-\alpha}.\end{align*} For any $\varepsilon\le s \le 1$ and $w,y,z\in \RR^d$ with $z\to y$, we have $p(s,w,y)\asymp p(s,w,z)$. By the dominated convergence theorem, it holds that $$\int_\varepsilon^1\int_{\RR^d} \tp(1-s,x,w)q(w)(p(s,w,y)-p(s,w,z))\,dw\,ds \to0,\quad z\to y.$$ Combining with all the estimates above, we prove the desired assertion. \end{proof} \begin{prop}\label{pro-con} The function $\tp(t,x,y)$ is jointly continuous with respect to $t>0$ and $x,y \in \RR^d_0.$\end{prop} \begin{proof} By the scaling property of $\tp(t,x,y)$, it suffices to show the continuity of $\tp(1,x,y)$ with respect to $x,y\in \RR^d_0$. As indicated in the proof of Lemma \ref{L:con1}, we only need to verify that $$-\int_0^1\int_{\RR^d} |\tp(1-s,\tilde x,w)p(s,w,\tilde y)-\tp(1-s,x,w)p(s,w,y)|q(w)\,dw\,ds\to 0$$ for any $x,y,\tilde x, \tilde y\in \RR_0^d$ with $\tilde x\to x$ and $\tilde y\to y$. In addition to \eqref{e:con1}, we have \begin{align*}&-\int_{1-\varepsilon}^1\int_{\RR^d} \tp(1-s,x,w)q(w)p(s,w,y)\,dw\,ds\\ &=-\int_0^\varepsilon \int_{\RR^d} \tp(s,x,w)q(w)p(1-s,w,y)\,dw\,ds\\ &\le - \int_0^\varepsilon \int_{\RR^d} p(s,x,w)q(w)p(1-s,w,y)\,dw\,ds \\ &\le c_1\varepsilon |x|^{-\alpha}. \end{align*} For any $\varepsilon<s<1-\varepsilon$ and $x,y,z,\tilde x,\tilde y\in \RR_0^d$ with $x\to \tilde x$ and $y\to \tilde y$, $p(s,z,\tilde y)\approx p(s,z,y),$ and $ \tilde p(1-s,\tilde x,z)\approx p(1-s,x,z),$ thanks to Lemma \ref{L:con1}. Then, by the dominated convergence theorem, it holds that $$-\int_\varepsilon^{1-\varepsilon}\int_{\RR^d} |\tp(1-s,\tilde x,w)p(s,w,\tilde y)-\tp(1-s,x,w)p(s,w,y)|q(w)\,dw\,ds\to 0$$ for any $x,y,\tilde x, \tilde y\in \RR_0^d$ with $\tilde x\to x$ and $\tilde y\to y$. Hence, according to all the estimates above, we prove the desired assertion. \end{proof} \begin{theorem}\label{thm:cont} {\bf (Joint continuity)}\,\, The function $\tp(t,x,y)$ is jointly continuous with respect to $t>0$ and $x,y \in \RR^d.$\end{theorem} \begin{proof}According to Proposition \ref{pro-con} and the scaling property of $\tp(t,x,y)$, we only need to verify that $p(1,x,y)$ is jointly continuous with respect to $x,y\in \RR^d$ when $x=0$ or $y=0$. Since $\tp(1,x,y)=0$ when $x=0$ or $y=0$, the desired assertion for the joint continuity is a direct consequence of the fact that $\tp(1,x,y)\ge0$ and two-sided estimates for $\tp(1,x,y)$ on $\RR_0^d\times \RR_0^d$. \end{proof} \subsection{Dirichlet forms} Finally, we discuss the Dirichlet form associated with the Schr\"{o}dinger operator $\mathcal{L}$ given by \eqref{e:op1}; see \cite{FOT} for the theory of Dirichlet forms. According to \cite[Theorem 2.5]{DC}, the Feynman-Kac semigroup $(\tilde P_t)_{t\ge0}$ in $L^2(\RR^d_0;dx)$ coincides with the semigroup corresponding to $\tilde\E$ with the domain $$ \mathscr{D}(\tilde \E)=\Big\{f\in L^2(\RR_0^d;dx): \iint_{\RR^d_0\times \RR_0^d} \frac{(f(x)-f(y))^2}{|x-y|^{d+\alpha}}\,dx\,dy+\int_{\RR^d_0} f^2(x)|q(x)|\,dx<\infty\Big\} $$ and defined by \begin{align*} \tilde \E (f,g)=&\frac{1}{2}\iint_{\RR^d_0\times \RR_0^d} {(f(x)-f(y))(g(x)-g(y))}\nu(x-y)\,dx\,dy\\ &+\int_{\RR^d_0} f(x)g(x)|q(x)|\,dx \end{align*} for any $f,g\in \mathscr{D}(\tilde \E),$ where $\nu$ is defined by \eqref{e:levymea}. Clearly, the quadratic form $(\tilde \E, \mathscr{D}(\tilde \E) )$ is equivalently given by \begin{align*} \tilde \E (f,g)=&\frac{1}{2}\iint_{\RR^d\times \RR^d} {(f(x)-f(y))(g(x)-g(y))}\nu(x-y)\,dx\,dy\\ &+\int_{\RR^d} f(x)g(x)|q(x)|\,dx,\\ \mathscr{D}(\tilde \E)=&\Big\{f\in L^2(\RR^d;dx): \tilde \E(f,f)<\infty\Big\}, \end{align*}which are extended to be defined on $L^2(\RR^d;dx).$ \begin{prop} $(\tilde \E, \mathscr{D}(\tilde \E) )$ is a symmetric regular Dirichlet form on $L^2(\RR^d;dx)$ with core $C_c^\infty(\RR^d).$ \end{prop} \begin{proof} Define $$\E(f,g)=\frac{1}{2}\iint_{\RR^d\times \RR^d} (f(x)-f(y))(g(x)-g(y))\nu(x-y)\,dx\,dy,\quad f,g \in \mathscr{D}(\E)$$ and $$\mathscr{D}(\E)=\Big\{f\in L^2(\RR^d;dx):\E(f,f)<\infty\Big\}.$$ Then, $(\E,\mathscr{D}(\E))$ is a symmetric Dirichlet form on $L^2(\RR^d;dx)$ associated with fractional Laplacian; moreover, $C_c^\infty(\RR^d)\subset \mathscr{D}(\Delta^{\alpha/2})$ (here $\mathscr{D}(\Delta^{\alpha/2})$ denotes the $L^2$-domain of $\Delta^{\alpha/2}$ on $L^2(\RR^d;dx)$) and $(\E,\mathscr{D}(\E))$ is regular with core $C_c^\infty(\RR^d)$; see \cite[Section 2.2.2]{CF} for more details. On the other hand, due to $\alpha<d$, we can verify that $$\int_{\RR^d} f(x)^2|q(x)|\,dx<\infty,\quad f\in C_c^\infty(\RR^d).$$ In particular, $C_c^\infty(\RR^d)\subset \mathscr{D}(\Delta^{\alpha/2})\cap L^2(\RR^d;|q(x)|\,dx)\subset \mathscr{D}(\tilde \E)$. It is easy to prove that $(\tilde \E, \mathscr{D}(\tilde \E) )$ is a symmetric Dirichlet form on $L^2(\RR^d;dx)$. Next, we claim that $C_c^\infty(\RR^d)$ is dense in $\mathscr{D}(\tilde\E)$ with the norm $\sqrt{\tilde\E} +\|\cdot\|_{L^2(\RR^d;dx)}.$ According to the Hardy inequality for fractional Laplacian (see \cite[Proposition 5]{2016-KB-BD-PK-pa}), there is a constant $C_0>0$ such that for all $f\in L^2(\RR^d;dx)$, $$\int_{\RR^d} f^2(x)|q(x)|\,dx\le C_0 \E(f,f).$$ Thus, the norms $\sqrt{\tilde\E} +\|\cdot\|_{L^2(\RR^d;dx)}$ and $\sqrt{\E} +\|\cdot\|_{L^2(\RR^d;dx)}$ are equivalent. Therefore, the desired assertion above immediately follows from the fact that $(\E,\mathscr{D}(\E))$ is regular with core $C_c^\infty(\RR^d)$. \end{proof} Let $h(x) = |x|^\delta$, and define \begin{align*} \bar{\cE}(f,f) = \frac{1}{2}\iint_{\Rd\times\Rd} \left(\frac{f(x)}{h(x)} - \frac{f(y)}{h(y)}\right)^2 h(x)h(y) \nu(x-y)\, dx\, dy, \quad f\in\mathscr{D}(\bar\cE), \end{align*} where $\mathscr{D}(\bar\cE) = \{f \in L^2(\RR^d; dx) \colon \bar{\cE}(f,f)< \infty\}$. \begin{prop} We have $\mathscr{D}(\tilde\cE) = \mathscr{D}(\bar\cE)$ and$$ \tilde\cE(f,f) = \bar{\cE}(f,f), \quad f \in \mathscr{D}(\tilde\cE). $$ \end{prop} \begin{proof} Denote by $\langle\cdot,\cdot\rangle$ the inner product of $L^2(\Rd;dx)$. Recall that $(\tilde P_t)_{t\ge0}$ is well defined on $L^2(\Rd;dx)$ by setting $\tilde P_tf(0)=0$ for any $f\in L^2(\Rd;dx)$. According to \eqref{eq:00}, \begin{align*} \langle f-\tilde{P}_t f,f \rangle &= \int_{\Rd} \left(f(x) - \int_{\Rd} \tp(t,x,y) f(y)\, dy\right) f(x)\, dx \\ &= \int_{\Rd} \left( \frac{f(x)}{h(x)}\int_\Rd \tp(t,x,y) h(y) \,dy - \int_{\Rd} \tp(t,x,y)h(y) \frac{f(y)}{h(y)}\, dy\right) f(x)\, dx \\ &= \iint_{\Rd\times\Rd} \tp(t,x,y)\left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right) \frac{f(x)}{h(x)} h(x) h(y) \,dx \,dy. \end{align*} Hence, by the symmetry, \begin{align*} \langle f-\tilde{P}_t f,f \rangle &= \frac{1}{2} \iint_{\Rd\times\Rd} \tp(t,x,y)\left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y)\,dx \,dy \end{align*} Note that \begin{align}\label{eq1:DF} \lim_{t \to 0^+} \frac{\tp(t,x,y)}{t} = \nu(x-y),\quad x,y\in \RR_0^d. \end{align} Indeed, by the Duhamel formula \eqref{e:for}, for any $x,y\in \RR_0^d$, \begin{align*} \lim_{t \to 0^+} \frac{\tp(t,x,y)}{t} &= \lim_{t \to 0^+} \frac{p(t,x,y)}{t} + \lim_{t \to 0^+} \frac{1}{t} \int_0^t \int_\Rd \tp(t-s,x,z) q(z) p(s,z,y)\, dz\, ds\\ &= \nu(x-y) + \lim_{t \to 0^+} \frac{1}{t} \int_0^t \int_\Rd \tp(t-s,x,z) q(z) p(s,z,y)\, dz \, ds. \end{align*} Next, by the fact that $\tp(t,x,y)\le p(t,x,y)$ for all $t>0$ and $x,y\in \RR^d$, Lemma \ref{lem:p1est}, \eqref{eq:scalp1} and \eqref{eq:Gest}, for any $x,y\in \RR_0^d$, \begin{align*} \frac{1}{t}\int_0^t \int_\Rd \tp(t-s,x,z) |z|^{-\alpha} p(s,z,y) \,dz \,ds \le c_1t^{-1}{p(t,x,y)}[G(t,x)+G(t,y)] \stackrel{t\to0}{\longrightarrow} 0, \end{align*} and so we get \eqref{eq1:DF}. Now, for $f \in \mathscr{D}(\tilde\cE)$, by Fatou's Lemma and \eqref{eq1:DF}, we have \begin{align*} \tilde\cE(f,f) &= \lim_{t \to 0} \frac{1}{t}\langle f-\tilde{P}_t f,f \rangle \\ &= \lim_{t \to 0}\frac{1}{2} \iint_{\Rd\times\Rd} \frac{\tp(t,x,y)}{t}\left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y)\,dy\, dx\\ &\ge \frac{1}{2} \iint_{\RR^d_0\times\RR^d_0} \liminf_{t \to 0}\frac{\tp(t,x,y)}{t}\left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y)\,dy \, dx\\ &= \frac{1}{2} \iint_{\RR^d_0\times \RR^d_0} \left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y) \nu(x-y) \, dy\, dx = \bar\cE(f,f). \end{align*} Hence, $\mathscr{D}(\tilde\cE) \subset \mathscr{D}(\bar\cE)$. On the other hand, we take $f \in \mathscr{D}(\bar\cE)$. Since $\tp(t,x,y) \le p(t,x,y) \le c_2t\nu(x-y)$ for all $t>0$ and $x,y\in \RR^d$, according to the dominated convergence theorem and \eqref{eq1:DF} again, we have \begin{align*} \tilde\cE(f,f) &= \lim_{t \to 0} \frac{1}{t}\langle f-\tilde{P}_t f,f \rangle \\ &= \lim_{t \to 0}\frac{1}{2} \iint_{\Rd\times\Rd} \frac{\tp(t,x,y)}{t}\left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y)\,dy\, dx\\ &= \frac{1}{2} \iint_{\Rd\times\Rd} \left( \frac{f(x)}{h(x)} - \frac{f(y)}{h(y)} \right)^2 h(x) h(y) \nu(x-y) \, dy\, dx=\bar\cE(f,f). \end{align*} Combining with both inequalities, we prove the desired assertion. \end{proof} \begin{remark}The construction of $(\bar \E, \mathscr{D}(\bar \E))$ can be deduced from Doob's theory of $h$-transformations; see \cite[Chapter 11]{CW} for more details. Indeed, as shown by \eqref{e:eee}, the function $h=|x|^\delta$ is harmonic with respect to the operator $\mathcal{L}$ given by \eqref{e:op1}. Define $\mathcal{L}_hf(x):=h(x)^{-1} \mathcal{L}(fh)(x)$ for all $f\in L^2(\RR^d;h(x)^2\,dx)$. It is easy to see that the operator $\mathcal{L}_h$ is symmetric on $L^2(\RR^d;h(x)^2\,dx)$, and the associated symmetric regular Dirichlet form $(\E_h, \mathscr{D}(\E_h))$ on $L^2(\RR^d;h(x)^2\,dx)$ is given by \begin{align*}\E_h(f,f)=&-\langle \mathcal{L}_h f,f\rangle_{L^2(\RR^d;h(x)^2\,dx)}\\ =&-\iint_{\RR^d\times \RR^d} \left(f(y)h(y)-f(x)h(x)\right)f(x)h(x)\nu(x-y)\,dy\,dx\\ &-\kappa\int_{\RR^d} |x|^{-\alpha} f(x)^2h^2(x)\,dx \\ =&-\iint_{\RR^d\times \RR^d} \left(f(y)-f(x)\right)f(x)h(y)h(x)\nu(x-y)\,dy\,dx\\ &-\iint_{\RR^d\times \RR^d}\left(h(y)-h(x)\right) f(x)^2h(x)\nu(x-y)\,dy\,dx\\ &-\kappa\int_{\RR^d} |x|^{-\alpha} f(x)^2h^2(x)\,dx\\ =&-\iint_{\RR^d\times \RR^d} \left(f(y)-f(x)\right)f(x)h(y)h(x)\nu(x-y)\,dy\,dx\\ =&\frac{1}{2}\iint_{\RR^d\times \RR^d} \left(f(y)-f(x)\right)^2h(y)h(x)\nu(x-y)\,dy\,dx\end{align*} for all $f\in \mathscr{D}(\E_h)$, where in the fourth equality we used \eqref{e:eee} and the last equality follows form the property that $\nu(x-y)=\nu(y-x)$. Note that $$\E_h(f,f)=-\langle \mathcal{L}_h f,f\rangle_{L^2(\RR^d;h(x)^2\,dx)}=- \langle \mathcal{L} (h f),hf\rangle_{L^2(\RR^d;dx)}=\tilde \E(hf,hf).$$ Combining both equalities above together, we arrive at $$\tilde \E(f,f)=\E_h(fh^{-1},fh^{-1})=\frac{1}{2}\iint_{\RR^d\times \RR^d} \left(\frac{f(y)}{h(y)}-\frac{f(x)}{h(x)}\right)^2h(y)h(x)\nu(x-y)\,dy\,dx.$$ The right side of the equality above coincides with the expression of $(\bar\cE, \mathscr{D}(\bar\cE)).$ \end{remark} \section{Appendix: Proof of Lemma \ref{lem:tpexp}} Let $q_0(x)=-|x|^{-\alpha}$. For any $\lm\ge 0$, denote by $p^{\lm}(t,x,y)$ the heat kernel associated with the generator $\Delta^{\alpha/2} + \lambda q_0(x)$. Hence, by Duhamel's formula (see \cite[Propositions 5.2 and 5.3]{DC} and their proofs), we have $$ p^{\lambda}(t,x,y) = p^\nu(t,x,y) + (\lm-\nu) \int_0^t \int_{\RR_0^d} p^\nu(t-s,x,z) q_0(z) p^{\lm}(s,z,y)\, dz\, ds $$ for any $t>0$ and $x,y\in \RR_0^d$. Noting that $q_0(x)<0$ for all $x\in \RR_0^d$, we can rewrite the equality above as $$ p^\nu(t,x,y) = p^{\lambda}(t,x,y) + \int_0^t \int_{\RR_0^d} p^\nu(t-s,x,z) (\lm-\nu) |q_0(z)| p^{\lm}(s,z,y)\, dz\, ds.$$ For any $\lm \ge0$, $t>0$ and $x,y\in \RR_0^d$, we set \begin{align*}p_0^{\lm}(t,x,y) &= p^{\lm}(t,x,y),\\ p_n^{\lm}(t,x,y) &= \int_0^t \int_{\RR_0^d} p_{n-1}^{\lm}(t-s,x,z) |q_0(z)| p^{\lm}(s,z,y)\, dz\,ds\,, \quad n \ge 1.\end{align*} Then, by \cite[Lemma 1 and the proof of Theorem 2]{2008-KB-WH-TJ-sm}, $$ p_{n+m+1}^{\lm}(t,x,y) = \int_0^t \int_{\RR_0^d} p_{n}^{\lm}(t-s,x,z) |q_0(z)| p_m^{\lm}(s,z,y)\, dz\,ds\,, \quad m,n \ge 0 $$ and \begin{align*} p^{\eta-\lm}(t,x,y) = \sum_{n=0}^\infty \lm^n p_n^\eta(t,x,y),\quad \eta > \lm >0. \end{align*} Furthermore, we have \begin{lemma} Let $0 < \lm < \eta< \infty$. For all $x,y\in \RR_0^d$ and $t>0$, \begin{equation}\label{eq:pkv_rel} \sum_{n=k}^\infty \binom{n}{k}\lm^{n-k} p_n^{\eta}(t,x,y) = p_k^{\eta - \lm}(t,x,y)\,, \quad k =0,1,2,\cdots. \end{equation} \end{lemma} \begin{proof} We use induction. When $k=0$, \eqref{eq:pkv_rel} holds trivially. For $k=1$, \begin{align*} p_1^{\eta - \lm}(t,x,y) &= \int_{0}^t \int_{\RR_0^d} p^{\eta-\lm}(t-s,x,z) |q_0(z)| p^{\eta-\lm}(s,z,y)\, dz\,ds \\ & = \int_{0}^t \int_{\RR_0^d} \sum_{i=0}^\infty \lm^i p_i^{\eta}(t-s,x,z) |q_0(z)| \sum_{j=0}^\infty \lm^j p_j^{\eta}(s,z,y)\, dz\, ds\\ & = \sum_{i=0}^\infty \sum_{j=0}^\infty \lm^{i+j} p_{i+j+1}^{\eta}(t,x,y) = \sum_{n=1}^\infty n \lm^{n-1} p_n^{\eta}(t,x,y)\,. \end{align*} Next, we assume that (\ref{eq:pkv_rel}) holds for some $k \in \NN_+$. We get \begin{align*} p_{k+1}^{\eta-\lm}(t,x,y) &= \int_{0}^t \int_{\RR_0^d} p_k ^{\eta-\lm}(t-s,x,z) |q_0(z)| p^{\eta-\lm}(s,z,y)\, dz\,ds \\ & = \int_{0}^t \int_{\RR_0^d} \sum_{i=k}^\infty \binom{i}{k}\lm^{i-k} p_i^{\eta}(t-s,x,z) |q_0(z)| \sum_{j=0}^\infty \lm^j p_j^{\eta}(s,z,y) \,dz\, ds\\ & = \sum_{i=k}^\infty \sum_{j=0}^\infty \binom{i}{k} \lm^{i+j-k} p_{i+j+1}^{\eta}(t,x,y) \\ & = \sum_{n=k+1}^\infty \sum_{j=0}^{n-k-1} \binom{n-j-1}{k} \lm^{n-k-1} p_n^{\eta}(t,x,y)\\ & = \sum_{n=k+1}^\infty \binom{n}{k+1} \lm^{n-k-1} p_n^{\eta}(t,x,y), \end{align*} where in the last equality we used the fact proved in the proof of \cite[Lemma 6]{2008-KB-WH-TJ-sm} (cf. \cite[(5.26)]{GKP}). The proof is complete. \end{proof} Next, we consider some properties of the function $\lambda\mapsto p^\lm(t,x,y)$. \begin{lemma} For fixed $x,y\in \RR_0^d$ and $t>0$, the function $$ h(\lm) = p^\lm(t,x,y), \quad \lm >0 $$ is completely monotone, i.e., $(-1)^kh^{(k)}(\lm)\ge0$ for all $\lm>0$ and $k=0,1,2,\cdots$. \end{lemma} \begin{proof} For $\lambda> 0$, we take $\eta>\lambda$. Choosing $k=0$ in (\ref{eq:pkv_rel}), we get $$ h(\lm) = \sum_{n=0}^\infty (\eta-\lm)^n p_n^{\eta}(t,x,y)\,. $$ By (\ref{eq:pkv_rel}), we get \begin{align}\label{eq:deriv} \frac{d^k }{d \lm^k} h(\lm)= {k!} \sum_{n=k}^\infty (-1)^k \binom{n}{k} (\eta-\lm)^{n-k} p^{\eta}_n(t,x,y) = (-1)^k \,{k!} \, p_k^{\lm}(t,x,y). \end{align} Since $p_k^{\lm}(t,x,y) \ge 0$, we conclude that $h$ is completely monotone on $(0,\infty)$. \end{proof} By the Bernstein theorem (see \cite[Theorem 1.4]{Schilling}), we get \begin{corollary}\label{cor:meas_exists} For fixed $x,y \in \RR_0^d$ and $t>0$, there exists a nonnegative Borel measure $\mu_{t,x,y}(du)$ on $[0,\infty)$ such that $$ p^{\lm}(t,x,y) = \int_0^\infty e^{-\lm u}\, \mu_{t,x,y}(du). $$ \end{corollary} The next lemma will yield the monotonicity of the function $\lm \mapsto \frac{p_1^{\lm}(t,x,y)}{p^{\lm}(t,x,y)}$. \begin{lemma}\label{lem:5} For every $n \ge 1$, $\lambda\ge0$, $t>0$ and $x,y \in \RR_0^d$, we have $$ (n+1)p_{n+1}^\lambda(t,x,y) p_{n-1}^\lambda(t,x,y) \ge n p_n^\lambda(t,x,y)^2. $$ \end{lemma} \begin{proof} Fix $x,y\in \RR_0^d$ and $t>0$, and let $\mu = \mu_{t,x,y}$ be the nonnegative measure from Corollary \ref{cor:meas_exists}. Then, by (\ref{eq:deriv}), $$ n!\, p^\lambda_n(t,x,y) = (-1)^n\,\frac{d^n}{d\lm^n} p^{\lm}(t,x,y) = \int_0^\infty e^{-\lambda u} u^n\, \mu(du)\,. $$ According to the Cauchy-Schwarz inequality, \begin{align*} \left[\int_0^\infty e^{-\lambda u} u^n\, \mu(du)\right]^2 &= \left[\int_0^\infty e^{-\lambda u/2} u^{(n+1)/2} \cdot e^{-\lambda u/2}u^{(n-1)/2} \,\mu(du)\right]^2 \\ &\le \left(\int_0^\infty e^{-\lambda u} u^{n+1} \,\mu(du)\right)\left(\int_0^\infty e^{-\lambda u} u^{n-1} \,\mu(du)\right). \end{align*} Hence, $$ (n! \,p_n^\lambda(t,x,y) )^2 \le (n+1)! \,p_{n+1}^\lambda(t,x,y) \cdot(n-1)! \,p_{n-1}^\lambda(t,x,y), $$ and so the desired assertion follows. \end{proof} \begin{lemma}\label{lem:decr} For fixed $x,y \in \RR_0^d$ and $t>0$, the function $\lm \mapsto \dfrac{p_1^{\lm}(t,x,y)}{p^{\lm}(t,x,y)}$ is decreasing on $(0,\infty)$. \end{lemma} \begin{proof} Let $H(\lambda)=\frac{-h'(\lambda)}{h(\lambda)}=\dfrac{p_1^{\lm}(t,x,y)}{p^{\lm}(t,x,y)}$ for $\lambda>0$, where in the second equality we used \eqref{eq:deriv}. Combining \eqref{eq:deriv} again with Lemma \ref{lem:5}, we find that $$H'(\lambda)=\frac{-h''(\lambda)h(\lambda)+h'(\lambda)^2}{h^2(\lambda)}=\frac{-2p^\lambda_2(t,x,y)p^\lambda(t,x,y)+p_1^\lambda(t,x,y)^2}{p^\lambda(t,x,y)^2}\le 0,$$ which yields the desired assertion. \end{proof} We now present the main result in this appendix, which immediately gives us Lemma \ref{lem:tpexp}. \begin{theorem}\label{thm:wlb} For every $\lm>0$, $t >0$ and $x,y \in \RR_0^d$, we have $$ p^{\lm} (t,x,y) \ge p(t,x,y) \exp \left[\dfrac{\lm p_1(t,x,y)}{p(t,x,y)}\right] $$ \end{theorem} \begin{proof} Fix $x,y \in \RR^d_0$ and $t>0$, and let $h(\lm) = p^{\lm}(t,x,y)$. Since $h \ge 0$, $$ h(\lm) = h(0) \exp\left[\int_0^\lm (\ln h(u))' \,du \right] = h(0) \exp\left[\int_0^\lm \dfrac{h'(u)}{h(u)} \,du \right]. $$ By \eqref{eq:deriv} and Lemma \ref{lem:decr}, we get $$ p^{\lm}(t,x,y) = p(t,x,y) \exp\left[-\int_0^\lm \dfrac{p_1^{u}(t,x,y)}{p^{u}(t,x,y)}\, du\right] \ge p(t,x,y) \exp \left[-\dfrac{\lm p_1(t,x,y)}{p(t,x,y)}\right]. $$ The proof is complete. \end{proof} \medskip \noindent \textbf{Acknowledgements.} We would like to thank Krzysztof Bogdan and Kamil Kaleta for interesting discussions and helpful comments.

c321ab353de4f40c58dae0a07139a4b674443004

\section{Quantum amplification and noise}The fundamental relations between quantum noise and quantum amplification are most straightforwardly derived in the Heisenberg picture. Thus, the standard way \cite{caves1982} to describe linear phase-preserving quantum amplification of a bosonic mode amplitude $a$ is through Caves' relation for the annihilation operator $\hat{a}$, \begin{eqnarray}\label{caves} \hat{a}_{{\rm out}}=\sqrt{G}\hat{a}_{{\rm in}}+\sqrt{G-1}\hat{b}^+_{{\rm in}}, \end{eqnarray} where $\hat{b}^+$ is the creation operator corresponding to an independent auxiliary bosonic mode $b$. Here the input field amplitude of mode $a$ is amplified by a factor of $\sqrt{G}$, but there is a cost: extra noise arising from the additional mode $b$ \footnote{This could be a mode internal to the detector.}. If this mode contains (thermal) excitations, mode $a$ after amplification will contain excitations, too, and their number is multiplied by $G-1$. Even if mode $b$ is in the vacuum state, it still adds noise \cite{caves1982}. It is clear that this extra noise is due to the additional creation operator term proportional to $\sqrt{G-1}$ in (\ref{caves}), but since that term is necessary so as to preserve the standard bosonic commutation relation $[\hat{a}_{{\rm out}},\hat{a}^+_{{\rm out}}]=\openone$ this tradeoff between linear amplification and added noise is fundamental. Indeed, phase-preserving linear amplification in proposed number resolving platforms using superconducting qubits have noise bounded by the Caves limit \cite{metelmann2014,clerk2010}. Recently, there has been some effort to describe all parts of the photo detection process, including amplification \cite{yang2019}, fully quantum mechanically \cite{young2018b,young2018,vanenk2017,dowling2018}. One conclusion that may be drawn from that research is that there is no severe amplification-driven tradeoff between efficiency and (thermally induced) dark counts. In particular, even though a few-photon signal must be amplified to a macroscopic level [forcing us to consider $G\gg 1$], thermal fluctuations in internal detector modes do not get amplified by the same factor of $G$. Experiments \cite{marsili2013,wollman2017uv} on superconducting nanowires demonstrate that over a wide range of detected wavelengths dark count rates can indeed be extremely low (on the order of one dark count per day). How can we reconcile these results with that of the previous paragraph? The answer, as we will show, is that amplification is not necessarily linear. That is, in the Heisenberg picture, the transformation of the bosonic annihilation operator can be nonlinear while still preserving the bosonic commutation relation. And, perhaps surprisingly, that way of amplifying can decrease the amount of noise added. \section{Nonlinear amplification}The idea is that for detecting single photons it is sufficient to have an output field whose total number of excitations is given by $N_{\rm out}=N_{\rm in}+G n_a$ with $n_a$ the number of input photons we would like to detect, and $N_{\rm in}$ the (fluctuating) number of excitations initially present in the output mode, which is {\em not} amplified. A physically allowed but highly idealized unitary transformation that accomplishes this is easiest written down in the Schr\"odinger picture (valid for any $n$ \footnote{The transformation (\ref{simple}) can be realized only when $M\geq Gn$. There is always such a restriction on amplification relations; the energy transferred to reservoir 2 must come from somewhere.}, even though in practice we will be interested mainly in small values of $n$, say, $n=0,1,2$) as \begin{eqnarray}\label{simple} \ket{n}_a \ket{M}_1\ket{N}_2 \longmapsto \ket{n}_a\ket{M-Gn}_1\ket{N+Gn}_2. \end{eqnarray} All states here are number (Fock) states of bosonic modes. The transformation involves two energy reservoirs: energy is transferred from the first reservoir to the second with the amount of energy transferred determined by the number $n$ of input photons in mode $a$ (with nothing happening at all when $n=0$). The assumption is that excitations of the two reservoirs have identical energies, $\hbar\omega'$, such that energy is conserved. The input mode can have any frequency $\omega$. The second reservoir ideally starts out with $N=0$ excitations---corresponding to the zero temperature limit---such that in the end it would contain exactly $Gn$ excitations if the input field contained $n$ photons. Clearly, this ideal transformation would represent perfect (noiseless) amplification of a photon number state (and $G$ will have to be an integer for this to work). \begin{figure}[t] \includegraphics[width=.8\linewidth]{idealamp2.jpeg} \caption{An input photon with frequency $\omega$ undergoes amplification into a macroscopic signal via electron-shelving \cite{dehmelt1975,wineland1980,bergquist1986}: when an on-resonance photon is absorbed, an atom (modeled here as a three-level system) enters the first excited state and a laser tuned to the second transition frequency $\omega_L\approx \omega'$ induces fluorescence. If there are multiple input photons, they are absorbed by multiple atoms and the fluorescence signal is increased proportionally. The number of fluorescence modes may be reduced by using a high-Q cavity so that amplification moves towards the ideal transformation given in Eq.~\ref{simple}.} \label{idealamp} \end{figure} Now we wish to describe this ideal process in the Heisenberg picture so as to make a direct comparison with Eq.~(\ref{caves}). In that picture, the ideal transformation that is linear in the number operator for the excitations in the second reservoir should be \footnote{\label{construction} The construction of the transformation (\ref{simpleb}) (and transformations including higher powers of the input photon number operator) is highly constrained by two conditions: that the spectrum of the operator-representation be $\mathbb{N}^0$ (the natural numbers including zero) and that the commutator be preserved. From these, we need \emph{at least} one term with a number operator on the right in (\ref{simpleb}) with a prefactor of one. Additional reservoirs with arbitrary prefactors are allowed but they will carry additional noise and decrease the SNR.} \begin{equation}\label{simpleb} \hat{b}^+_{\rm out}\hat{b}_{\rm out}=\hat{b}^+_{\rm in}\hat{b}_{\rm in}+G\hat{a}^+_{\rm in}\hat{a}_{\rm in}. \end{equation} We are now going to do three things: (A) we will construct expressions for $\hat{b}^+_{\rm out}$ and $\hat{b}_{\rm out}$ such that (\ref{simpleb}) is reproduced and such that their commutator $[\hat{b}^+_{\rm out},\hat{b}_{\rm out}]=\openone$; (B) we will add non-ideal features that make the model more realistic, and (C) we will include fluctuations in the initial number of excitations in the $b$ mode and calculate the signal-to-noise ratio (SNR) for the final number of excitations in the $b$ mode, both for the ideal limit and the more realistic models. A comparison with linear amplification will then show how nonlinear amplification improves upon the former. To start with part of task (B), we adjust the idealized Schr\"{o}dinger picture to get rid of two features that make the process (\ref{simple}) obviously inapplicable to real detectors, but such that the Heisenberg picture (\ref{simpleb}) is still valid. First note that the $n$ photons in the process (\ref{simple}) are not destroyed, whereas in a standard detector they are. We fix that by introducing another quantum system $S$ with a continuum of energies $E$ that can absorb the energy $n\hbar\omega$ of the $n$ photons. This modifies (\ref{simple}) by adding a step \begin{equation} \ket{n}_a\ket{E}_S \longmapsto\ket{0}_a\ket{E+n\hbar\omega}_S. \end{equation} Since this extra step does not affect the state of reservoir 2, the crucial equation (\ref{simpleb}) stays the same. The second change concerns phase: in the Schr\"{o}dinger picture we can insert random phase factors $\exp(i\phi^{\mbox{\fontsize{5}{0}$\clubsuit$}})$ on the right-hand side of (\ref{simple}). This makes the amplification process irreversible (as any amplification process in a real detector is) and it destroys superpositions of different number states (e.g., coherent states will not be coherently amplified). It destroys any entanglement between the different modes as well \footnote{Phase randomization is necessary for optimal amplification and measurement of photon number due to number-phase uncertainty.}. For task (A) we would like to use the polar decompositions of the creation and annihilation operators. That is, in analogy to the polar decomposition of a complex number, $z=\exp(i\phi)\sqrt{|z|^2}$, we would like to write \begin{eqnarray}\label{bb} \hat{b}_{{\rm out}}&= \hat{S}\,\sqrt{(\hat{b}^+\hat{b})_{{\rm in}}+G(\hat{a}^+\hat{a})_{{\rm in}}}, \end{eqnarray} where $\hat{S}$ is a unitary operator written in the suggestive form $\exp(i\hat{\phi})$ for some hermitian operator $\hat{\phi}$. In a finite-dimensional Hilbert space of dimension $s+1$ there is no problem defining $\hat{S}$: it is a shift operator that acts on number states $\ket{N}$ of the bosonic mode as \begin{eqnarray}\label{S} \hat{S}\ket{N}=e^{i\phi^{\mbox{\fontsize{5}{0}$\clubsuit$}}}\ket{N-1}\,\,\,{\rm for}\,\,s \geq N>0, \end{eqnarray} with $\hat{S}\ket{0}=\ket{s}$ and $\phi^{\mbox{\fontsize{5}{0}$\clubsuit$}}$ the random phase we introduced earlier. Since Fock space is infinite-dimensional, we use the Pegg-Barnett trick \cite{pegg1989} of truncating the Hilbert space at a high excitation number $s$ and only in the end (when calculating physical quantities) taking the limit $s\rightarrow\infty$. It is easy to verify that the relation (\ref{bb}) yields the commutator $[\hat{b}_{{\rm out}},\hat{b}^+_{{\rm out}}]=\openone_{{\rm in}} - (s+1) \ket{s}\bra{s}$, in which the extra Pegg-Barnett term won't contribute to any physical quantity, while ensuring a traceless commutator, necessary in finite dimensions \footnote{Note the dimension of both input and output mode Hilbert spaces is $s+1$; they necessarily match in the Heisenberg picture.}. The nonlinear equation (\ref{bb}) does not seem to have appeared in the large literature on bosonic amplification (for a review, see, e.g., \cite{clerk2010}). Refs.~\cite{yuen1986,ho1994,yuen1996} did discuss photon-number amplifiers (especially in the high-photon number limit) decades ago, but no attempt was made there to find commutator-preserving operator equations. \section{More realistic models for amplification}\label{Models} Continuing with task (B), in a more realistic description the reservoirs consist of many modes. So, instead of having just one bosonic output mode $b$ we really should describe many output reservoir modes. For example, we may have $G$ modes $b_k$ [recall $G$ is an integer now] each one of which satisfies \begin{eqnarray}\label{bk} \hat{b}_{k\,{\rm out}}&=\hat{S}_k\,\sqrt{(\hat{b}^+\hat{b})_{k\,{\rm in}}+(\hat{a}^+\hat{a})_{{\rm in}}},\,\,\,\,\,k=1\ldots G. \end{eqnarray} Here the macroscopic signal monitored and analyzed consists of the {\em sum} of all detected excitations (since each mode by itself contains just a microscopic number of excitations we cannot simply assume to be able to count those individual numbers: then we would not need amplification at all!). That is, we consider as our macroscopic output signal \begin{eqnarray}\label{Iout} \hat{I}_{{\rm out}}= \sum_{k=1}^G (\hat{b}^+\hat{b})_{k\,{\rm out}} =\sum_{k=1}^G (\hat{b}^+\hat{b})_{k\,{\rm in}} + G (\hat{a}^+\hat{a})_{{\rm in}}. \end{eqnarray} Another extension is to ``avalanche" photodetection where one small-scale amplification event triggers the next and the process repeats, giving rise to a macroscopic signal. Iterating the transformation (\ref{simpleb}) of single mode amplification $N$ times with a gain factor $g$ in each step gives a total gain factor $G=g^N$ and an input-output relation \begin{equation}\label{simplebn} (\hat{b}^+\hat{b})_{\rm N \,out}=\sum_{k=1}^{N}g^{N-k} (\hat{b}^+\hat{b})_{k\,{\rm in}} + G (\hat{a}^+\hat{a})_{{\rm in}} \end{equation} where mode $\hat{b}_k$ here contains the output of the $k$th amplification step, and the last mode $b_N$ contains the signal. Another extension, relevant for $n>1$, describes multiplexing: the idea is that $n$ photons are most conveniently detected by $n$ detectors that each detect one (and only one) photon, along the lines of \cite{nehra2017,yu2018}. We will not describe this model in any detail, except to state that amplification would in that case be described by $Gn$ modes, each containing exactly one extra excitation. Lastly, we combine both multi-mode and multi-step extensions above by repeating the process in (\ref{bk}) and (\ref{Iout}) of amplification into several ($g$) modes $N$ times, again with a total gain factor defined $G=g^n$ and an input-output relation for the macroscopic signal \begin{eqnarray}\label{Ioutn} \\ \hat{I}_{{\rm out}} = \sum_{k_N=1}^G (\hat{b}^+\hat{b})_{k_N\,{\rm out}} = \sum\limits_{n=1}^{N}\sum\limits_{k_n=1}^{g^n} (\hat{b}^+\hat{b})_{k_n\,{\rm in}} + G(\hat{a}^+\hat{a})_{{\rm in}}\nonumber \end{eqnarray} where the mode $\hat{b}_{k_n}$ is the $k_n$th mode in the $n$th step. Note that in our nonlinear amplification models the amplified signal ends up in different bosonic mode(s): indeed, a photodetector typically converts the input signal (light) to an output signal of a physically different type, e.g., electron-hole pairs (which may sometimes be approximated as composite bosons; see also \cite{keldysh1968,devoret2000,laikhtman2007,combescot2007}). \section{Number fluctuations}We turn to task (C) and calculate the noise in photon number introduced by the amplification process and by the coupling to reservoirs. For the reservoir we monitor, we write \begin{eqnarray}\label{bvar} \expect{(\hat{b}^+ \hat{b})_{\rm{in}}}=\overline{n}_b;\,\expect{(\hat{b}^+ \hat{b})^2_{\rm{in}}}=\overline{n}^2_b + \Delta n_b^2 \end{eqnarray} and make no further assumptions about its initial state. We assume that there is some (unknown) number of photons in the input mode $a$ that we want to measure. We thus consider input states that are diagonal in the photon number basis, with some nonzero photon number fluctuations $\Delta n_a$. (Thanks to the randomized phase assumption we can use this assumption without loss of generality for our nonlinear models.) So, we write \begin{eqnarray}\label{avar} \expect{(\hat{a}^+\hat{a})_{\rm{in}}}=\overline{n}_a;\,\expect{(\hat{a}^+\hat{a})^2_{\rm{in}}}=\overline{n}^2_a +\Delta n_a^2. \end{eqnarray} In the following we always assume the initial states of modes $a$ and $b$ to be independent, such that \begin{equation} \expect{f(\hat{a},\hat{a}^+)g(\hat{b},\hat{b}^+)}=\expect{f(\hat{a},\hat{a}^+)} \expect{g(\hat{b},\hat{b}^+)} \end{equation} for any functions $f$ and $g$. For linear phase-insensitive amplification (\ref{caves}), we find the following variance in the number of excitations in the amplified signal: \begin{eqnarray} \label{cavesvar} \sigma_{(\hat{a}^+\hat{a})_{\rm out}}^2 &=&G^2\Delta n_a^2 + (G-1)^2 \Delta n_b^2 +\nonumber\\ && G(G-1) (2 \overline{n}_a\,\overline{n}_b + \overline{n}_a+\overline{n}_b + 1). \end{eqnarray} Not only are the fluctuations in the auxiliary mode $b$ amplified [second term in (\ref{cavesvar})], there is inherent noise from the amplification process itself even if $\Delta n_b^2=0$ [the second line is strictly positive for $G>1$]. We should also consider linear phase-sensitive amplification \cite{caves1982}, described by \begin{eqnarray}\label{phase} \hat{a}_{{\rm out}}=\sqrt{G}\hat{a}_{{\rm in}}+\sqrt{G-1}\hat{a}^+_{{\rm in}}. \end{eqnarray} Here, compared to (\ref{caves}) the $\hat{b}^+$ term is replaced by the $\hat{a}^+$ term, such that the commutator $[\hat{a}_{{\rm out}},\hat{a}^+_{{\rm out}}]$ is still preserved. This gives a variance \begin{eqnarray} \label{phasevar} \!\!\sigma_{(\hat{a}^+\hat{a})_{\rm out}}^2 &=& (6G(G-1) + 1) \Delta n_a^2 +\nonumber\\ && 2G(G-1) (\overline{n}^2_a+\overline{n}_a+1). \end{eqnarray} There is again extra amplification noise for $G>1$ [the second line], much like what we found for phase-insensitive amplification. We compare these two results for linear amplification to the result for the nonlinear amplification process described by (\ref{bb}). The variance in excitation number is \begin{eqnarray}\label{bbvar} \sigma_{(\hat{b}^+\hat{b})_{\rm out}}^2 &=\Delta n_b^2 + G^2 \Delta n^2_a. \end{eqnarray} Here the number fluctuations in the auxiliary mode are {\em not} amplified and there is no additional amplification noise either. Already we are able to see that the scheme of amplification into a single mode is optimal; any transformation that would reduce the prefactor of $\Delta n_b^2$ in (\ref{bbvar}) below unity would fail to realize a well-behaved annihilation operator (for details, see again \footnotemark[3])! For nonlinear amplification into many modes described by the more realistic model equations (\ref{bk}) and (\ref{Iout}), we find \begin{eqnarray}\label{MNvar} \sigma_{\hat{I}_{\rm out}}^2 &=G\Delta n_b^2 + G^2 \Delta n^2_a, \end{eqnarray} where for simplicity we assumed all reservoir modes to be independent with the same number fluctuations. This shows amplifying according to (\ref{bk}) is suboptimal; even though it still beats both linear amplification limits (\ref{cavesvar}) and (\ref{phasevar}) the noise in the reservoir modes is still amplified. Similarly, amplification using multiple fermionic degrees of freedom will be sub-optimal; a similar multi-mode description will be necessary \footnote{One way around this limitation is for the incident photons to only interact with a single symmetrized collective degree of freedom of many fermions, on which a measurement is then made. In this idealized case, this collective degree of freedom plays the role of a single bosonic mode and amplification could still be described by (\ref{bb}) and photon number amplification is improved past the limit for linear fermionic amplification \cite{yurke2004}.}. Defining the total gain $g^N=G$ with $N$ the number of steps, we find for our multi-step models that \begin{eqnarray}\label{iter} \sigma_{(\hat{b}^+\hat{b})_{\rm out}}^2=\frac{G^2-1}{g^2-1}\Delta n^2_b+G^2\Delta n^2_a \end{eqnarray} for amplification of $g$ excitations into a single mode and \begin{eqnarray}\label{itermult} \sigma_{\hat{I}_{\rm out}}^2=G\frac{G-1}{g-1}\Delta n^2_b+G^2\Delta n^2_a \end{eqnarray} for amplification of a single excitation into $g$ modes. \section{Signal-to-noise ratios}We can now write down explicit tradeoff relations between amplification and number fluctuations in terms of signal-to-noise ratios for all types of amplification discussed here, for the case where the number of input photons is {\em fixed} to be $n_a$ (and so $\Delta n_a=0$). Using the standard signal-to-noise ratio as the number of excitations in the amplified mode minus the background, divided by the standard deviation in the number of excitations, we get \begin{eqnarray}\label{all} \textnormal{SNR}_{{\rm PhaseInsensitive}} &\leq&\frac{G}{G-1}\frac{n_a}{\Delta n_b}\label{PhaseInsensitive}\\ \textnormal{SNR}_{{\rm PhaseSensitive}} &\leq&\frac{2G - 1}{\sqrt{2G(G-1)}} n_a\label{PhaseSensitive}\\ \textnormal{SNR}_{{\rm SingleMode}}&=& \frac{Gn_a}{\Delta n_b}\label{SM}\\ \textnormal{SNR}_{{\rm GModes}}&=&\frac{Gn_a}{\sqrt{G}\Delta n_b}=\frac{\sqrt{G}n_a}{\Delta n_b}.\label{GM} \end{eqnarray} The linear amplification mechanisms have increasingly worse signal-to-noise ratios as $G$ increases \footnote{The signal-to-noise ratios (\ref{PhaseInsensitive}) and (\ref{PhaseSensitive}) for linear amplification become infinite at $G=1$ simply because there is no noise when both $G=1$ and $\Delta n_a = 0$.}, albeit saturating in the limit $G\rightarrow\infty$. In contrast, the signal-to-noise ratios for the nonlinear amplification mechanisms improve with increasing $G$, with amplification into a single-mode performing best \footnote{We find the linear dependence on $G$ resulting from single-shot single-mode amplification holds for transformations describing higher-order amplification of photon number operator, again subject to the constraints of [13].}. Similarly, we consider multi-step amplification models \begin{eqnarray} \textnormal{SNR}_{{\rm MultiStepSingleMode}}&=&\frac{G\sqrt{g^2-1}n_a}{\sqrt{G^2-1}\Delta n_b}\label{ISM}\\ \textnormal{SNR}_{{\rm MultiStepMultiMode}}&=& \frac{\sqrt{G(g-1)}n_a}{\sqrt{G-1}\Delta n_b}.\label{IMM} \end{eqnarray} These intermediate noise limits fill in the space between the optimal SNR (\ref{SM}) and linear amplification. (Indeed, multi-step multi-mode amplification performs slightly \emph{worse} than both linear mechanisms for $g=2$!) \section{Single-photon pre-amplification}While our paper focuses on the amplification part of the photo detection process, we very briefly consider the pre-amplification process now. We certainly cannot treat that part in full generality here and we adopt several simplifications in order to arrive at an important result concerning the suppression of thermal noise. First, we assume that we can decouple the amplification stage from the pre-amplification filtering [by having an irreversible step in between the two] such that filtering does not interfere negatively with the absorption/transduction part \cite{young2018}. We then focus on just the time/frequency degree of one incoming photon \footnote{ Photon-number resolved photo detection can be achieved by multiplexing an $n$-photon signal to many ($N\gg n$) single photon detectors \cite{nehra2017}, each satisfying (\ref{transfer}) independently. However, this means an additional noise mode will be added with each splitting of the signal, decreasing the integrated signal-to-noise ratio. To avoid added noise a non-linear multi-photon filtering process could be used, but for this a full S-matrix treatment must be used, see Refs. ~\cite{fan2010,caneva2015,xu2015}.}. A single absorber with some resonance frequency $\omega_0$ able to absorb that single photon will act as a frequency filter. If the pre-amplification filtering is passive (easy to implement, but we certainly can go beyond this \footnote{See, for example, Ref.~\cite{raymer2010}. The result is that, instead of certain frequencies, it is certain spectral ``Schmidt modes'' that are detected perfectly.}) and unitary (i.e., lossless: we consider this because we are interested in the fundamental limits of photo detection. Internal losses only degrade performance.), then frequency filtering is described by the linear transformation \begin{equation}\label{transfer} \hat{a}_{{\rm out}}(\omega)=T(\omega) \hat{a}_{{\rm in}}(\omega)+R(\omega)\hat{c}_{{\rm in}}(\omega) \end{equation} where $c_{{\rm in}}(\omega)$ is yet another internal bosonic detector mode at the same frequency as the input mode \cite{Propp}. Here $T(\omega)$ and $R(\omega)$ are ``transmission'' and ``reflection'' coefficients which satisfy $|T(\omega)|^2+|R(\omega)|^2=1$ and which are determined by the resonance structures internal to the photodetector. The amplification process that follows the initial absorption of the photon energy is applied to the operator $\hat{a}_{\rm out}(\omega)$ of Eq. (\ref{transfer}), so that (explicitly displaying the different frequencies of the modes now) ideal amplification (single-mode and single-shot) is described \begin{equation}\label{simplec} \!\hat{b}^+_{\rm out}(\omega')\hat{b}_{\rm out}(\omega')=\hat{b}^+_{\rm in}(\omega')\hat{b}_{\rm in}(\omega')+G\hat{a}^+_{\rm out}(\omega)\hat{a}_{\rm out}(\omega). \end{equation} This makes rigorous the idea that one {\em can} amplify at any frequency, enabling the mantra that one {\em should} amplify at high (optical) frequencies \cite{Dowling}. Namely, thermal fluctuations at a frequency $\omega'$ may be suppressed by choosing the reservoir mode frequency $\omega'$ such that $\hbar\omega'\gg kT$. This suppression is exponential: $\Delta n_b^2\propto\bar{n}_b\propto \exp(-\hbar\omega'/kT)$. Note that number fluctuations in the internal mode $c_{{\rm in}}(\omega)$ at the input frequency will be amplified by the subsequent amplification process. However, one can in principle construct ideal detectors for light with a particular frequency $\omega_0$ \cite{young2018}, such that $|T(\omega_0)|=1$ and hence $R(\omega_0)=0$ \cite{Propp}, avoiding internally generated dark counts at that particular frequency. If instead of a single spectral mode a small range of frequencies is amplified with differing probabilities, matching the amplification spectrum to the filtering spectrum is sufficient for reducing internally generated dark counts, as we will discuss in more detail in work in preparation \cite{Propp2}. \section{Further applications}The models for amplification considered here apply to other types of quantum measurement as well. For example, electron-shelving \cite{dehmelt1975,wineland1980,bergquist1986} is a well-known method to perform atomic state measurements. Here one particular atomic state (e.g., one of the hyperfine ground states of an ion) is coupled resonantly to a higher-lying excited state which can then decay back by fluorescence only to that same ground state. A laser tuned to that transition can then induce the atom to emit a macroscopic amount (visible by eye) of fluorescent light. In the language accompanying Eq.~(\ref{simple}), the laser beam forms the first reservoir, while the second reservoir consists of vacuum modes that are filled with fluorescent light as described by (\ref{bk}). The gain factor $G$ (the number of fluorescence photons) is determined by the ratio of Einstein's coefficients for spontaneous and stimulated emission and the total integration time. By placing the atom/ion inside a high-Q optical resonator (with resonant frequency $\omega'$) we would reduce the number of output modes and thereby get closer to the optimum. The idea of placing a detector inside a resonant cavity is, of course, not new \cite{unlu1995}, but that idea is usually associated with increasing the coupling to light. Although we do have that effect as well, the main purpose here is to reduce the number of output modes, and thereby increase the SNR (Fig.~\ref{idealamp}). In Ref.~\cite{yuen1986} a transformation similar to (\ref{simple}) is given, with $n$ photons being converted to $Gn$ photons in a single mode directly. Though this transformation is unphysical (there is no way to preserve the commutator), a SNR is calculated that increases linearly with $G$ like our Eq. (\ref{SM}). However, the SNR found in \cite{yuen1986} diverges for a photo detector with unit efficiency, which is not the case once fluctuations in the reservoir mode are properly taken into account as our results clarify. In Ref.~\cite{yang2019} an interesting model for amplification is constructed that makes use of a first-order phase transition for a collection of $N$ interacting spin-1/2 particles. These spins are coupled both to an input photon and to an output bosonic mode. The SNR (as we define it here) for that model scales as $\sqrt{N}$ while the gain $G$ of that model is linear in $N$. Thus, the SNR scales with $\sqrt{G}$ just as our Eq.~(\ref{GM}): the number of spins in Ref.~\cite{yang2019}'s model plays a similar role as our number of amplification modes. \section{Conclusions}We discussed various linear and nonlinear amplification schemes for bosonic modes. For detecting few photons, we found that the latter add considerably less noise, leading to better signal-to-noise ratios, as exemplified in Eqs.~(\ref{all})--(\ref{IMM}). Unlike for linear amplification, number fluctuations in internal detector modes are not amplified, while the number of photons that we want to detect {\em is} amplified. All amplification schemes explicitly preserve the bosonic commutation relations. While amplification into a single-mode may not be feasible in practice, it provides the fundamental lower limit to noise in photon-number measurements across amplification mechanisms. In practice, one may have many output modes and thus may find a SNR closer to Eq. (\ref{GM}), which is worse by a factor of $\sqrt{G}$ than the fundamental limit (but still better by a factor of $\sqrt{G}$ than linear amplification), or one may have multiple amplification steps (\ref{ISM}), or both (\ref{IMM}). To test this, we suggest that measurement of the gain dependence of the SNR for a given photo detector should provide a rough but useful indication of the underlying amplification mechanism. \vspace{1em} This work is supported by funding from DARPA under Contract No. W911NF-17-1-0267. We thank Joseph Altepeter and Sae Woo Nam for their useful comments on this project.

8c93dab593d1a4349fa1bc2098634a5d7695f580

\section{Introduction} Traditional attention-based encoder-decoder neural machine translation (NMT) models learn \emph{word-level} embeddings, with a continuous representation for each unique word type~\cite{bahdanau2014neural}. However, this results in a long tail of rare words for which we do not learn good representations. More recently, it has become standard practice to mitigate the vocabulary size problem with Byte-Pair Encoding (BPE) \cite{Gage1994,sennrich2016subword}. BPE iteratively merges consecutive characters into larger chunks based on their frequency, which results in the breaking up of less common words into ``subword units.'' While BPE addresses the vocabulary size problem, the spellings of the subword units are still ignored. On the other hand, purely \emph{character-level} NMT translates one character at a time and can implicitly learn about morphological patterns within words as well as generalize to unseen vocabulary. Recently,~\newcite{cherry2018revisiting} show that very deep character-level models can outperform BPE, however, the smallest data size evaluated was 2 million sentences, so it is unclear if the results hold for low-resource settings and when translating into a range of different morphologically rich languages. Furthermore, tuning deep character-level models is expensive, even for low-resource settings. Deep character-level models are sensitive to the dropout rate and tuning takes much longer due to longer sequence lengths~\cite{cherry2018revisiting}. A middle-ground alternative is character-\emph{aware} word-level modeling. Here, the NMT system operates over words but uses word embeddings that are sensitive to spellings and thereby has the ability to learn morphological patterns in the language. Such character-aware approaches have been applied successfully in NMT to the \emph{source-side} word embedding layer \cite{costa2016character}, but surprisingly, similar gains have not been achieved on the target side~\cite{belinkov2017neural}. While source-side character-aware models only need to make the \emph{source embedding layer} character-aware, on the target-side we require both the \emph{target embedding layer} and the \emph{softmax layer} to be character-aware, which presents additional challenges. We find that the trivial application of methods from~\newcite{costa2016character} to these target-side embeddings results in significant drop in performance. Instead, we propose mixing compositional and standard word embeddings via a gating function. While simple, we find it is critical to successful target-side character awareness. It is worth noting that unlike some purely character-level methods our aim is not to generate novel words, though this method can function on top of subword methods which do so~\cite{shapiro2018bpe}. Rather, the character-aware representations decrease the sparsity of embeddings for rare words or subwords, which are a problem in low-resource morphologically rich settings. We summarize our contribution as follows: \begin{enumerate} \item We propose a method for utilizing character-aware embeddings in an NMT decoder that can be used over word or subword sequences. \item We explore how our method interacts with BPE over a range of merge operations (including word-level and purely character-level) and highlight that there is no ``typical BPE'' setting for low-resource NMT. \item We evaluate our model on $14$ target languages and observe consistent improvements over baselines. Furthermore, we analyze to what extent the success of our method corresponds to improved handling of target language morphology. \end{enumerate} \section{Related Work} NMT has benefited from character-aware word representations on the source side~\cite{costa2016character}, which follows language modeling work by \newcite{kim2016character} and generate source-side input embeddings using a CNN over the character sequence of each word. Further analysis revealed that hidden states of such character-aware models have increased knowledge of morphology ~\cite{belinkov2017neural}. They additionally try using character-aware representations in the target side embedding layer, leaving the softmax matrix with standard word representations, and found no improvements. Our work is also aligned with the character-aware models proposed in \cite{kim2016character}, but we additionally employ a gating mechanism between character-aware representations and standard word representations similar to language modeling work by \cite{miyamoto2016gated}. However, our gating is a learned type-specific vector rather than a fixed hyperparameter. There is additionally a line of work on purely character-level NMT, which generates words one character at a time \cite{ling2015character,chung2016character,passban2018improving}. While initial results here were not strong, \newcite{cherry2018revisiting} revisit this with deeper architectures and sweeping dropout parameters and find that they outperform BPE across settings of the merge hyperparameter. They examine different data sizes and observe improvements in the smaller data size settings---however, the smallest size is about 2 million sentence pairs. In contrast, we look at a smaller order of magnitude data size and present an alternate approach which doesn't require substantial tuning of parameters across different languages. Finally, Byte-Pair Encoding (BPE) \cite{sennrich2016subword} has become a standard preprocessing step in NMT pipelines and provides an easy way to generate sequences with a mixture of full words and word fragments. Note that BPE splits are agnostic to any morphological pattern present in the language, for example the token \texttt{politely} in our dataset is split into \texttt{pol+itely}, instead of the linguistically plausible split \texttt{polite+ly}.\footnote{We observe this split when merge parameter was $15$k.} Our approach can be applied to word-level sequences and sequences at any BPE merge hyperparameter greater than $0$. Increasing the hyperparameter results in more words and longer subwords that can exhibit morphological patterns. Our goal is to exploit these morphological patterns and enrich the word (or subword) representations with character-awareness. \section{Encoder-Decoder NMT} \label{method} An attention-based encoder-decoder network \cite{bahdanau2014neural,luong2015effective} models the probability of a target sentence $\vec{y}$ of length $J$ given a source sentence $\vec{x}$ as: \begin{align} p(\vec{y} \mid \vec{x}) &= \prod_{j=1}^{J}p(y_j \mid \vec{y}_{0:j-1}, \vec{x}; \vec{\theta}) \label{eq:objective} \end{align} where $\vec{\theta}$ represents all the parameters of the network. At each time-step the $j'$th output token is generated by: \begin{align} p(y_j \mid \vec{y}_{0:j-1}, \vec{x})=\text{softmax}(\vec{W_o} \vec{s}_j) \label{eq:outputdist} \end{align} where $\vec{s}_j \in \mathbb{R}^{D\times1}$ is the decoder hidden state at time $j$ and $\vec{W_o} \in \mathbb{R}^{\lvert\mathcal{V}\rvert \times D}$ is the weight matrix of the softmax layer, which provides a continuous representation for target words. $\vec{s}_j$ is computed using the following recurrence: \begin{align} \vec{s}_j &= \text{tanh}(\vec{W_c}\;[\vec{c}_j;\vec{\tilde{s}}_j])\\ \vec{\tilde{s}}_j &= f([\vec{s}_{j-1}; \vec{w_s}^{y_{j-1}}; \vec{\tilde{s}}_{j-1}]) \end{align} where $f$ is an LSTM cell.\footnote{Note that our notation diverges from \newcite{luong2015effective} so that $\vec{s}_j$ refers to the state used to make the final predictions.} $\vec{W_s} \in \mathbb{R}^{\lvert\mathcal{V}\rvert \times E}$ is the target-side embedding matrix, which provides continuous representations for the previous target word when used as input to the RNN. Here, $\vec{w_s}^{y_{j-1}} \in \mathbb{R}^{1 \times E}$ is a row vector from the embedding matrix $\vec{W_s}$ corresponding to the value of $y_{j-1}$. $\mathcal{V}$ is the target vocabulary set, $D$ is the is the RNN size and $E$ is embedding size. Often these matrices $\vec{W_o}$ and $\vec{W_s}$ are tied. The context vector $\vec{c}_j$ is obtained by taking a weighted average over the concatenation of a bidirectional RNN encoder's hidden states. \begin{align} \vec{c}_j &= \sum_{i=1}^{{I}}\alpha_i,\vec{h}_i \\ \alpha_i & = \frac{\exp{(\vec{s}_j^T \vec{W_a} \vec{h}_i)}}{\sum_l \exp{(\vec{s}_j^T \vec{W_a} \vec{h}_l})} \end{align} The attention matrix $\vec{W_a} \in \mathbb{R}^{D\times H}$ is learned jointly with the model, multiplying with the previous decoder state and bidirectional encoder state $\vec{h}_i \in \mathbb{R}^{H\times1}$, normalized over encoder hidden states via the softmax operation. \section{Character-Aware Extension In this section we detail the incorporation of character-awareness into the two decoder embedding matrices $\vec{W_o}$ and $\vec{W_s}$. To begin, we consider an example target side word (or subword in the case of preprocessing with BPE), \texttt{cat}. In both $\vec{W_o}$ and $\vec{W_s}$, there exist row vectors, $\vec{w_o}^{\texttt{cat}}$ and $\vec{w_s}^{\texttt{cat}}$ that contain the continuous vector representation for the word \texttt{cat}. In a traditional NMT system, these vectors are learned as the entire network tries to maximize the objective in Equation \ref{eq:objective}. The objective does not require the vectors $\vec{w_o}^{\texttt{cat}}$ and $\vec{w_s}^{\texttt{cat}}$ to model any aspect of the spelling of the word. Figure~\ref{fig:nocomposition} illustrates a simple non-compositional word embedding. At a high level, we can view our notion of character-awareness as a composition function $\text{comp}(.;\vec{\omega})$, parameterized by $\vec{\omega}$, that takes the character sequence that makes up a word (i.e. its spelling) as input and then produces a continuous vector representation: \begin{align} \vec{w}_{\text{comp}}^{\texttt{cat}} &= \text{comp}(\langle s\rangle, \texttt{c},\texttt{a}, \texttt{t}, \langle/s\rangle; \vec{\omega}) \label{eq:comptemplate} \end{align} $\vec{\omega}$ is learned jointly with the overall objective. Special characters $\langle s \rangle$ and $\langle /s \rangle$ denote the beginning and end of sequence respectively. Figure~\ref{fig:cnn} illustrates our compositional approach to generating embeddings~\cite{kim2016character}. First, a character-embedding layer converts the spelling of a word into a sequence of character embeddings. Next, we apply $4$ convolution operations, with kernel sizes $3,4,5$ and $6$, over the character sequence and the resulting output matrix is max-pooled. We set the output channel size of each convolution to $\frac{1}{4}$ of the final desired embedding size. The max-pooled vector from each convolution is concatenated to create the composed word representation. Finally, we add highway layers to obtain the final embeddings. \begin{figure} \centering \resizebox{\columnwidth}{!}{ \subfloat[]{ \tikzset{every picture/.style={line width=1.00pt}} \begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1] \draw [color={rgb, 255:red, 92; green, 92; blue, 92 } ,draw opacity=1 ][fill={rgb, 255:red, 155; green, 155; blue, 155 } ,fill opacity=0.7 ] (21,7.61) -- (30.33,7.61) -- (30.33,67.33) -- (21,67.33) -- cycle ; \draw (28,94) node {$\vec{w}^{v}_{\text{std}}$}; \draw (28,119) node {$v=\texttt{cat}$}; \end{tikzpicture} \label{fig:nocomposition} } \subfloat[]{ \tikzset{every picture/.style={line width=1.00pt}} \begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1] \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ] (67.4,86.58) -- (137.52,86.98) ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ] (97.8,41.2) -- (67.4,86.58) ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ] (107.4,41.6) -- (137.52,86.98) ; \draw [color={rgb, 255:red, 245; green, 166; blue, 35 } ,draw opacity=1 ][fill={rgb, 255:red, 245; green, 166; blue, 35 } ,fill opacity=0.4 ] (7.8,91.78) -- (17.92,91.78) -- (17.92,101.67) -- (7.8,101.67) -- cycle ; \draw [color={rgb, 255:red, 245; green, 166; blue, 35 } ,draw opacity=1 ][fill={rgb, 255:red, 245; green, 166; blue, 35 } ,fill opacity=0.4 ] (37.6,91.98) -- (47.72,91.98) -- (47.72,101.87) -- (37.6,101.87) -- cycle ; \draw [color={rgb, 255:red, 245; green, 166; blue, 35 } ,draw opacity=1 ][fill={rgb, 255:red, 245; green, 166; blue, 35 } ,fill opacity=0.4 ] (67.4,91.58) -- (77.52,91.58) -- (77.52,101.47) -- (67.4,101.47) -- cycle ; \draw [color={rgb, 255:red, 245; green, 166; blue, 35 } ,draw opacity=1 ][fill={rgb, 255:red, 245; green, 166; blue, 35 } ,fill opacity=0.4 ] (97,91.98) -- (107.12,91.98) -- (107.12,101.87) -- (97,101.87) -- cycle ; \draw [color={rgb, 255:red, 245; green, 166; blue, 35 } ,draw opacity=1 ][fill={rgb, 255:red, 245; green, 166; blue, 35 } ,fill opacity=0.4 ] (127.4,91.98) -- (137.52,91.98) -- (137.52,101.87) -- (127.4,101.87) -- cycle ; \draw [color={rgb, 255:red, 74; green, 144; blue, 226 } ,draw opacity=1 ][fill={rgb, 255:red, 74; green, 144; blue, 226 } ,fill opacity=0.4 ] (142.6,13.08) -- (151.93,13.08) -- (151.93,72.8) -- (142.6,72.8) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ][fill={rgb, 255:red, 184; green, 233; blue, 134 } ,fill opacity=0.7 ] (23,42) -- (32,42) -- (32,71.6) -- (23,71.6) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ][fill={rgb, 255:red, 184; green, 233; blue, 134 } ,fill opacity=0.7 ] (53,42) -- (62,42) -- (62,71.6) -- (53,71.6) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ][fill={rgb, 255:red, 184; green, 233; blue, 134 } ,fill opacity=0.7 ] (83,42) -- (92,42) -- (92,71.6) -- (83,71.6) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ][fill={rgb, 255:red, 184; green, 233; blue, 134 } ,fill opacity=0.7 ] (113,43) -- (122,43) -- (122,72.6) -- (113,72.6) -- cycle ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ][fill={rgb, 255:red, 65; green, 117; blue, 5 } ,fill opacity=0.4 ] (38,12) -- (47,12) -- (47,41.6) -- (38,41.6) -- cycle ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ][fill={rgb, 255:red, 65; green, 117; blue, 5 } ,fill opacity=0.4 ] (68,12) -- (77,12) -- (77,41.6) -- (68,41.6) -- cycle ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ][fill={rgb, 255:red, 65; green, 117; blue, 5 } ,fill opacity=0.4 ] (98,12) -- (107,12) -- (107,41.6) -- (98,41.6) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ] (32,56.8) -- (141.2,57.59) ; \draw [shift={(143.2,57.6)}, rotate = 180.41] [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ][line width=1.00] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ; \draw [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ] (47,26.8) -- (141.25,26.75) ; \draw [shift={(143.25,26.75)}, rotate = 539.97] [color={rgb, 255:red, 65; green, 117; blue, 5 } ,draw opacity=1 ][line width=1.00] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ; \draw [color={rgb, 255:red, 74; green, 144; blue, 226 } ,draw opacity=1 ] (142.6,42.94) -- (151.93,42.94) ; \draw [color={rgb, 255:red, 74; green, 144; blue, 226 } ,draw opacity=1 ][fill={rgb, 255:red, 74; green, 144; blue, 226 } ,fill opacity=0.4 ] (215.6,13.08) -- (224.93,13.08) -- (224.93,72.8) -- (215.6,72.8) -- cycle ; \draw [color={rgb, 255:red, 74; green, 144; blue, 226 } ,draw opacity=1 ][fill={rgb, 255:red, 255; green, 255; blue, 255 } ,fill opacity=1 ] (190.6,13.08) -- (199.93,13.08) -- (199.93,72.8) -- (190.6,72.8) -- cycle ; \draw [color={rgb, 255:red, 74; green, 144; blue, 226 } ,draw opacity=1 ][fill={rgb, 255:red, 255; green, 255; blue, 255 } ,fill opacity=1 ] (166.6,13.08) -- (175.93,13.08) -- (175.93,72.8) -- (166.6,72.8) -- cycle ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ] (32,71.6) -- (47.72,86.98) ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ] (23,71.6) -- (7.8,86.78) ; \draw [color={rgb, 255:red, 126; green, 211; blue, 33 } ,draw opacity=1 ] (7.8,86.78) -- (47.72,86.98) ; \draw [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ] (151.93,42.94) -- (164.67,43.06) ; \draw [shift={(166.67,43.08)}, rotate = 180.56] [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ][line width=1.00] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ; \draw [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ] (175.93,42.94) -- (188.67,43.06) ; \draw [shift={(190.67,43.08)}, rotate = 180.56] [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ][line width=1.00] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ; \draw [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ] (199.93,42.94) -- (212.67,43.06) ; \draw [shift={(214.67,43.08)}, rotate = 180.56] [color={rgb, 255:red, 0; green, 0; blue, 0 } ,draw opacity=1 ][line width=1.00] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ; \draw (43.47,115.93) node [align=left] {c}; \draw (104.67,115.53) node [align=left] {t}; \draw (74.07,115.73) node [align=left] {a}; \draw (172,43) node [rotate=-270] [align=left] {{\tiny highway 1}}; \draw (196,43) node [rotate=-270] [align=left] {{\tiny highway 2}}; \draw (134,114) node {$\langle /s\rangle $}; \draw (13,114) node {$\langle s\rangle $}; \draw (220,88) node {$\vec{w}^{v}_{\text{comp}}$}; \draw (217,116) node {$v=\texttt{cat}$}; \end{tikzpicture} \label{fig:cnn} } } \caption{Different approaches to generating embeddings. (a) standard word embedding that treats words as a single symbol. (b) CNN-based composition function. We use multiple CNNs with different kernel sizes over the character embeddings. The resulting hidden states are combined into a single word embedding via max pooling. Note that (b) shows only $2$ convolution filters for clarity, in practice we use 4.} \label{fig:compositions} \end{figure} \subsection{Composed \& Standard Gating}\label{sec:CG} The composition is applied to every type in the vocabulary and thus generates a complete embedding matrix (and softmax matrix). In doing so, we assume that \emph{every} word in the vocabulary has a vector representation that can be composed from its spelling sequence. This is a strong assumption as many words, in particular high frequency words, are not normally compositional, e.g.\ the substring \texttt{ing} in \texttt{thing} is not compositional in the way that it is in \texttt{running}. Thus, we mix the compositional and standard embedding vectors. We expect standard embeddings to better represent the meaning of certain words, such has function words and other high-frequency words. For each word $v$ in the vocabulary we also learn a gating vector $\vec{g}^{v} \in [0,1]^{1\times D}$. \begin{align} \vec{g}^{v} &= \sigma(\vec{w}_{\text{gate}}^{v}) \label{eq:gating} \end{align} Where, $\sigma$ is a sigmoid operation and type-specific parameters $\vec{w}_{\text{gate}}^{v}$ are jointly learned along with all the other parameters of the composition function. These parameters are regularized to remain close to $\vec{0}$ using dropout. \footnote{However, in practice we found that this regularization did not affect performance noticeably in this setting.} Our final mixed word representation for each word $v \in \mathcal{V}$ is given by: \begin{align} \vec{w}^{v}_{\text{mix}} = \vec{g}^{v} \odot \vec{w}^{v}_{\text{std}} + (\vec{1.} - \vec{g}^{v}) \odot \vec{w}^{v}_{\text{comp}} \label{eq:mixed} \end{align} Where $\vec{w}^{v}_{\text{mix}}$ is the final word embedding, $ \vec{w}^{v}_{\text{std}}$ is the standard word embedding, $\vec{w}^{v}_{\text{comp}}$ is the embedding by the composition function and $\vec{g}^{v}$ is the type-specific gating vector for the $v$'th word. The weight matrix is obtained by stacking the word vectors for each word $v \in \mathcal{V}$. The same representation is used for the target embedding layer and the softmax layer i.e. we set $\vec{w_o}^{\texttt{cat}} = \vec{w_s}^{\texttt{cat}} = \vec{w}_{\text{mix}}^{\texttt{cat}}$, when $v = \texttt{cat}$. Thus, tying the composition function parameters for the softmax weight matrix and the target-side embedding matrix. Experiments comparing the standard embedding model and the compositional embedding model with and without gating are summarized in Table~\ref{tab:prelim}. Row ``C'' shows the performance of naively using the composition function (which works in the source-side) on the target-side. We observe a catastrophic drop in BLEU ($-14.62$) compared to a standard NMT encoder-decoder. The Character-aware gated model(CG), however, outperforms the baseline by $0.91$ BLEU points suggesting that the CNN composition function and standard embeddings work in a complementary fashion. \begin{table}[] \small \begin{center} \begin{tabular}{@{}ll@{}} \toprule \textbf{Composition Method} & \textbf{BLEU} \\ \midrule Std.\phantom{-} (no composition) & 26.84 \\ C\phantom{GS} (without gating)& 12.22 \\ CG (target embedding only) & 26.61 \\ CG (softmax embedding only) & 27.16 \\ CG (both)& \textbf{27.75} \\ \bottomrule \end{tabular} \end{center} \caption{Experiments to determine the effectiveness of composition based embeddings and gated embeddings. We used en-de language pair from the TED multi-target dataset. Std. is our baseline with standard word embeddings, model C is the composition only model and CG combines the character-aware (composed) embedding and standard embedding via a gating function. } \label{tab:prelim} \end{table} \subsection{Large Vocabulary Approximation}\label{sec:approx} In Equation~\ref{eq:outputdist} of the general NMT framework, the softmax operation generates a distribution over the output vocabulary. Our character-aware model requires a much larger computation graph as we apply convolutions (and highway layers) over the spellings (character embeddings) of entire target vocabulary, placing a limitation on the target vocabulary size for our model. Which is problematic for word-level modeling (without BPE). To make our character-aware model accommodate large target vocabulary sizes, we incorporate an approximation mechanism based on~\cite{chousing}. Instead of computing the softmax over the entire vocabulary, we uniformly sample $20$k vocabulary types and the vocabulary types that are present in the training batch. During decoding, we compute the forward pass $\vec{W_o}\vec{s}_j$ in Equation~\ref{eq:outputdist} in several splits of the target vocabulary. As no backward pass is required we clear the memory (i.e.\ delete the computation graph) after each split is computed. \section{Experiments} \label{sec:exp} \begin{table*}[t] \small \centering \begin{tabular}{@{}cccccccccc@{}} \toprule \multicolumn{1}{l}{\textbf{Language}} & \multicolumn{3}{c|}{\textbf{BPE Sweep}} & \multicolumn{3}{c|}{\textbf{@ $30$k BPE}}& \multicolumn{3}{c}{\textbf{@ Word-level}}\\ \multicolumn{1}{c}{} & \textbf{Std}(Best BPE) & \textbf{CG}(Best BPE) & \multicolumn{1}{c|}{\textbf{$\Delta$}} & \textbf{Std} & \textbf{CG} & \multicolumn{1}{c|}{\textbf{$\Delta$}}& \textbf{Std} & \textbf{CG} & \textbf{$\Delta$}\\ \midrule cs & 20.57 (7.5k) & 21.41 (7.5k) & \multicolumn{1}{l|}{+0.84}& 18.73 & 21.28 & \multicolumn{1}{l|}{+2.55}& 18.44 & 21.49 & +3.05 \\ uk & 15.79 (7.5k) & 16.60 (30k) & \multicolumn{1}{l|}{+0.81}& 14.27& 16.60&\multicolumn{1}{l|}{ +2.33}& 12.94 & 15.30 & +2.36\\ pl & 16.76 (15k) & 18.00 (30k) & \multicolumn{1}{l|}{+1.24}& 15.98& 18.00 & \multicolumn{1}{l|}{+2.02}& 15.49 & 17.20& +1.71\\ tr & 15.11 (7.5k) & 15.83 (30k) & \multicolumn{1}{l|}{+0.72}& 13.82& 15.83& \multicolumn{1}{l|}{+2.01}& 12.58 & 14.75& +2.17\\ hu & 16.61 (3.2k) & 17.23 (15k) & \multicolumn{1}{l|}{+0.62}& 15.45& 17.21& \multicolumn{1}{l|}{+1.76}& 14.18& 16.52 & +2.34\\ he & 23.36 (3.2k) & 23.86 (30k) & \multicolumn{1}{l|}{+0.50}& 22.47 & 23.86 & \multicolumn{1}{l|}{+1.39} & 21.26 & 23.01 & +1.75\\ pt & 37.85 (15k) & 38.35 (30k) & \multicolumn{1}{l|}{+0.50} & 37.05 & 38.35 & \multicolumn{1}{l|}{+1.30} & 37.13 & 38.36& +1.23\\ ar & 16.22 (7.5k) & 16.28 (30k) & \multicolumn{1}{l|}{+0.06}& 15.05& 16.28& \multicolumn{1}{l|}{+1.23}& 14.45& 16.05&+1.60 \\ de & 27.37 (7.5k) & 28.12 (30k) & \multicolumn{1}{l|}{+0.75}& 26.94& 28.12& \multicolumn{1}{l|}{+1.21} & 26.84 & 27.75 & +0.91\\ ro & 24.02 (3.2k) & 24.20 (15k) & \multicolumn{1}{l|}{+0.18}& 22.88& 24.00& \multicolumn{1}{l|}{+1.12} &22.39 & 23.27& +0.88\\ bg & 31.63 (7.5k) & 32.20 (15k) & \multicolumn{1}{l|}{+0.57}& 30.92& 31.90& \multicolumn{1}{l|}{+0.98}& 30.18& 31.43& +1.25\\ fr & 35.97 (1.6k) & 36.17 (7.5k) & \multicolumn{1}{l|}{+0.20}& 35.31& 35.92& \multicolumn{1}{l|}{+0.61}& 35.28 & 36.01 & +0.73\\ fa & 12.94 (30k) & 13.52 (30k) & \multicolumn{1}{l|}{+0.58}& 12.94 & 13.52& \multicolumn{1}{l|}{+0.58}& 12.85 & 12.79& -0.06\\ ru & 19.28 (30k) & 19.61 (30k) & \multicolumn{1}{l|}{+0.33}& 19.28 & 19.61& \multicolumn{1}{l|}{+0.33}& 17.60& 19.04& +1.44\\ \bottomrule \end{tabular} \caption{Best BLEU scores swept over $6$ different BPE merge setting ($1.6$k, $3.2$k, $7.5$k, $15$k, $30$k, $60$k), and at a standard setting of $30$k. We notice a consistent improvement across languages and settings of the merge operation parameter. } \label{tab:avg_bpe} \end{table*} We evaluate our character aware model on $14$ different languages in a low-resource setting. Additionally, we sweep over several BPE merge hyperparameter settings from character-level to fully word-level for both our model and the baseline and find consistent gains in the character-aware model over the baseline. These gains are stable across all BPE merge hyperparameters all the way up to word-level where they are the highest. \subsection{Datasets} We use a collection of TED talk transcripts \cite{duh18multitarget,mauro2012wit3}. This dataset has languages with a variety of morphological typologies, which allows us to observe how the success of our character-aware decoder relates to morphological complexity. We keep the source language fixed as English and translate into $14$ different languages, since our focus is on the decoder. The training sets for each vary from 74k sentences pairs for Ukrainian to around 174k sentences pairs for Russian (provided in Appendix A), but the validation and test sets are ``multi-way parallel'', meaning the English sentences (the source side in our experiments) are the same across all $14$ languages, and are about $2$k sentences each. We filter out training pairs where the source sentence was longer that $50$ tokens (before applying BPE). For word-level results, we used a vocabulary size of $100$k (keeping the most frequent types) and replaced rare words by an \texttt{<UNK>} token. \subsection{NMT Setup} We work with OpenNMT-py~\cite{klein2017OpenNMT}, and modify the target-side embedding layer and softmax layer to use our proposed character-aware composition function. A $2$ layer encoder and decoder, with $1000$ recurrent units were used in all experiments The embeddings sizes were made to match the RNN recurrent size. We set the character embedding size to $50$ and use four CNNs with kernel widths $3, 4, 5$ and $6$. The four CNN outputs are concatenated into a compositional embeddings and gated with a standard word embedding. The same composition function (with shared parameters) was used for the target embedding layer and the softmax layer. We optimize the NMT objective (Equation~\ref{eq:objective}) using SGD.\footnote{SGD outperformed both Adam and Adadelta. Others have found similar trends, see \newcite{bahar2017empirical} and \newcite{maruf2017document}.} An initial learning rate of 1.0 was used for the first $8$ epochs and then decayed with a decay rate of $0.5$ until the learning rate reached a minimum threshold of $0.001$. We use a batch size of 80 for our main experiments. At the end of each epoch we checkpoint and evaluate our model on a validation datset and used validation accuracy as our model selection criteria for test time. During decoding, a beam size of $5$ was chosen for all the experiments. \subsection{Results} \begin{table}[t] \small \centering \begin{tabular}{@{}lcccc@{}} \toprule \multicolumn{1}{@{}l@{}}{\multirow{2}{*}{\textbf{Lang}}} & \multicolumn{1}{c}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Char-\\ Shallow\end{tabular}}}} & \multicolumn{1}{c}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Char-\\ Deep\end{tabular}}}} & \multicolumn{1}{c}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\textbf{CG}\\ \textbf{($30$k BPE)}\end{tabular}}} & \multicolumn{1}{@{}c@{}}{\multirow{2}{*}{\textbf{$\Delta$}}}\\ \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} \\ \midrule uk & \phantom{1}4.77& 13.34 & 16.60 & +3.26\\ cs & 11.16& 18.45 & 21.28 & +2.83\\ de &23.89&25.93& 28.12 & +2.19 \\ bg & 26.40&29.81& 31.90 & +2.09\\ tr & \phantom{1}5.29 & 13.94 & 15.83 & +1.89\\ pl & 10.65 & 16.31& 18.00 & +1.69\\ ru &14.63&18.01& 19.61 & +1.60\\ ro &21.58&22.45&24.00 & +1.55\\ pt &35.00& 37.06& 38.35 & +1.29\\ hu & \phantom{1}2.51& 16.02 & 17.21 & +1.19\\ fr & 32.71& 34.76 &35.92 & +1.16\\ fa & \phantom{1}7.44&12.73 & 13.52 & +0.79\\ ar & \phantom{1}3.58&15.89& 16.28 & +0.39\\ he & 22.28&23.87& 23.86 & -0.01\\ \bottomrule \end{tabular} \caption{BLEU scores (lowercased) comparing character-level models against CG when used on $30$k BPE sequences. We show that without sweeping BPE, CG generally outperforms purely character-level methods, even when the purely character-level networks are deepened as was shown to help in~\newcite{cherry2018revisiting}. } \label{tab:chars} \end{table} We provide case insensitive BLEU scores for our main experiments, comparing our character-aware model (CG) against a baseline model that uses only standard word (and subword) embeddings. We divide the results of our model's performance into three parts: \begin{enumerate*}[label=(\roman*)] \item over a sweep of BPE merge operations, including a commonly used setting of $30$k merge operations \item with word-level source and target sequences and finally, \item against a purely character-level model. \end{enumerate*} \subsubsection{BPE Results} Part $1$ of Table~\ref{tab:avg_bpe} compares the best BLEU score obtained by the baseline model, after performing a BPE sweep from $1.6$k to $60$k, to the best BLEU obtained by CG after sweeping over the same BPE range. While our study focuses on the target side, BPE (with the same number of merge operations) was applied to both source and target for our experiments. We find that after this sweep, CG outperforms the baseline in all 14 languages. The exhaustive table of results for these experiments is presented in Appendix \ref{detailed_results}. \subsubsection*{No Typical BPE Setting} Additionally, we see that the BPE setting that achieves best BLEU in the baseline model varies considerably from $1.6$k to $30$k depending on the target language, indicating that \emph{there is no ``typical'' BPE for low-resource settings}. In the CG model, however, performance was usually best at $30$k. Part $2$ of Table~\ref{tab:avg_bpe} compares the baseline and CG at BPE of $30$k where CG performs optimally. We find that our CG model consistently outperforms the baseline for almost all BPE merge hyperparameters across all 14 languages. Figure~\ref{fig:deltableu} shows the gains observed by the CG model as we sweep over BPE merge operations. While the baseline model does slightly better than CG at small BPE settings for a few languages (all points below the $0$ value), a majority of the points show positive gains. \subsubsection{Word-Level Results} In Part $3$ of Table \ref{tab:avg_bpe} we show results with our approximation for word level. While our best results are generally with BPE, we note that we get the biggest relative gains using our method at the word level, which we expect is due to always having the whole word to learn character patterns over. For the CG model, in $60$k BPE and word-level settings we used the large vocabulary approximation discussed in Section~\ref{sec:approx}. \subsubsection{Character-Level Results} Finally, in Table \ref{tab:chars}, we compare two character-level models against our CG model at $30$k BPE. The shallow character-level model used $2$ encoder and decoder layers with $1000$ recurrent units, while the deep model used $6$ encoder and decoder layers with $512$ recurrent units .\footnote{Increasing the recurrent size for deep models resulted in significant drop in BLEU scores. We set the dropout rate to $0.1$.} Furthermore, the improved results from the deep model were only attainable using the Fairseq toolkit with Noam optimization and $100$ warmup steps~\cite{gehring2017convs2s}. As Table~\ref{tab:chars} shows, our CG model with $30$k BPE compares favorably to even deep character-level models for this low-resource setting. \definecolor{col1}{RGB}{76, 114, 176} \definecolor{col2}{RGB}{221, 132, 82} \definecolor{col3}{RGB}{85, 167, 104} \definecolor{col4}{RGB}{196, 78, 82} \definecolor{col5}{RGB}{129, 114, 178} \definecolor{col6}{RGB}{147, 120, 96} \definecolor{col7}{RGB}{218, 139, 195} \definecolor{col8}{RGB}{140, 140, 140} \definecolor{col9}{RGB}{204, 185, 116} \definecolor{col10}{RGB}{100, 181, 205} \definecolor{col11}{RGB}{198, 219, 239} \definecolor{col12}{RGB}{158, 202, 225} \definecolor{col13}{RGB}{107, 174, 214} \definecolor{col14}{RGB}{49, 130, 189} \begin{figure} \centering \small \resizebox {\columnwidth} {!} { \begin{tikzpicture} \begin{axis}[ ylabel style={at={(axis description cs:+0.15,.5)},anchor=south,align=center,font=\small}, xlabel style={font=\small}, legend pos=outer north east, xmin=-0.35, xmax=6.35, xlabel=BPE Merge Operations, ylabel=$\Delta$ BLEU\\(CG - Std.), xticklabels={0, 1.62k, 3.2k, 7.5k, 15k, 30k, 60k$^*$, W$^*$}] \addplot[smooth,color=col1, opacity=0.4,thin ] plot coordinates { (0, 0.43) (1, 0.52) (2, 0.84) (3, 1.54) (4, 2.55) (5, 3.37) (6, 3.05) }; \addplot[clip mode=individual, only marks, color=col1, mark=text, text mark=cs, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.43) (1, 0.52) (2, 0.84) (3, 1.54) (4, 2.55) (5, 3.37) (6, 3.05) }; \addplot[smooth,color=col2, opacity=0.4] plot coordinates { (0, 0.45) (1, 0.65) (2, -0.31) (3, 0.92) (4, 2.33) (5, 3.04) (6, 2.36) }; \addplot[clip mode=individual, only marks, color=col2, mark=text, text mark=uk, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.45) (1, 0.65) (2, -0.31) (3, 0.92) (4, 2.33) (5, 3.04) (6, 2.36) }; \addplot[smooth,color=col3, opacity=0.4,thin ] plot coordinates { (0, 0.81) (1, 0.28) (2, 1.26) (3, 0.62) (4, 1.76) (5, 2.24) (6, 2.34) }; \addplot[clip mode=individual, only marks, color=col3, mark=text, text mark=hu, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.81) (1, 0.28) (2, 1.26) (3, 0.62) (4, 1.76) (5, 2.24) (6, 2.34) }; \addplot[smooth,color=col4, opacity=0.4,thin] plot coordinates { (0, 0.74) (1, 0.72) (2, 0.50) (3, 0.87) (4, 2.02) (5, 1.85) (6, 1.71) }; \addplot[clip mode=individual, only marks, color=col4, mark=text, text mark=pl, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.74) (1, 0.72) (2, 0.50) (3, 0.87) (4, 2.02) (5, 1.85) (6, 1.71) }; \addplot[smooth,color=col5, opacity=0.4 ,thin] plot coordinates { (0, 0.45) (1, 0.02) (2, 0.33) (3, 0.57) (4, 1.39) (5, 0.94) (6, 1.75) }; \addplot[clip mode=individual, only marks, color=col5, mark=text, text mark=he, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.45) (1, 0.02) (2, 0.33) (3, 0.57) (4, 1.39) (5, 0.94) (6, 1.75) }; \addplot[smooth,color=col6, opacity=0.4 ,thin] plot coordinates { (0, -0.50) (1, 0.67) (2, 0.40) (3, 0.79) (4, 2.01) (5, 1.36) (6, 2.17) }; \addplot[clip mode=individual, only marks, color=col6, mark=text, text mark=tr, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, -0.50) (1, 0.67) (2, 0.40) (3, 0.79) (4, 2.01) (5, 1.36) (6, 2.17) }; \addplot[smooth,color=col7, opacity=0.4,thin ] plot coordinates { (0, 0.30) (1, -0.12) (2, -0.05) (3, 0.29) (4, 1.23) (5, 0.67) (6, 1.60) }; \addplot[clip mode=individual, only marks, color=col7, mark=text, text mark=ar, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.30) (1, -0.12) (2, -0.05) (3, 0.29) (4, 1.23) (5, 0.67) (6, 1.60) }; \addplot[smooth,color=col8, opacity=0.4 ,thin] plot coordinates { (0, 0.47) (1, 0.45) (2, 0.16) (3, 0.43) (4, 1.30) (5, 1.00) (6, 1.23) }; \addplot[clip mode=individual, only marks, color=col8, mark=text, text mark=pt, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.47) (1, 0.45) (2, 0.16) (3, 0.43) (4, 1.30) (5, 1.00) (6, 1.23) }; \addplot[smooth,color=col9, opacity=0.4 ,thin] plot coordinates { (0, 0.07) (1, -0.60) (2, -0.11) (3, 0.42) (4, 1.12) (5, 0.65) (6, 0.88) }; \addplot[clip mode=individual, only marks, color=col9, mark=text, text mark=ro, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.07) (1, -0.60) (2, -0.11) (3, 0.42) (4, 1.12) (5, 0.65) (6, 0.88) }; \addplot[smooth,color=col10, opacity=0.4 ,thin] plot coordinates { (0, 0.26) (1, 0.30) (2, 0.18) (3, 1.11) (4, 0.98) (5, 1.14) (6, 1.25) }; \addplot[clip mode=individual, only marks, color=col10, mark=text, text mark=bg, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.26) (1, 0.30) (2, 0.18) (3, 1.11) (4, 0.98) (5, 1.14) (6, 1.25) }; \addplot[smooth,color=col1, opacity=0.4,thin] plot coordinates { (0, 0.51) (1, 0.55) (2, 0.35) (3, 0.50) (4, 0.33) (5, 0.95) (6, 1.44) }; \addplot[clip mode=individual, only marks, color=col1, mark=text, text mark=ru, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, 0.51) (1, 0.55) (2, 0.35) (3, 0.50) (4, 0.33) (5, 0.95) (6, 1.44) }; \addplot[smooth,color=col2, opacity=0.4 ,thin] plot coordinates { (0, -0.04) (1, 0.21) (2, 0.09) (3, 0.66) (4, 1.18) (5, 0.16) (6, 0.91) }; \addplot[clip mode=individual, only marks, color=col2, mark=text, text mark=de, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, -0.04) (1, 0.21) (2, 0.09) (3, 0.66) (4, 1.18) (5, 0.16) (6, 0.91) }; \addplot[smooth,color=col3, opacity=0.4 ,thin] plot coordinates { (0, -0.52) (1, 0.27) (2, 0.52) (3, 0.42) (4, 0.58) (5, 0.11) (6, -0.06) }; \addplot[clip mode=individual, only marks, color=col3, mark=text, text mark=fa, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, -0.52) (1, 0.27) (2, 0.52) (3, 0.42) (4, 0.58) (5, 0.11) (6, -0.06) }; \addplot[smooth,color=col4, opacity=0.4,thin ] plot coordinates { (0, -0.08) (1, -0.07) (2, 0.35) (3, 0.20) (4, 0.61) (5, 0.75) (6, 0.73) }; \addplot[clip mode=individual, only marks, color=col4, mark=text, text mark=fr, text mark as node, text mark style={font=\tiny}] plot coordinates { (0, -0.08) (1, -0.07) (2, 0.35) (3, 0.20) (4, 0.61) (5, 0.75) (6, 0.73) }; \addplot[smooth,color=Black, opacity=1.0, ultra thick,mark=*,mark size=1pt] plot coordinates { (0, 0.24) (1, 0.28) (2, 0.32) (3, 0.67) (4, 1.39) (5, 1.3) (6, 1.53) }; \addplot[mark=none, Black, samples=2, domain=-1:7, thin] {0.0}; \end{axis} \end{tikzpicture} } \caption{Plot of the difference between the BLEU scores from CG model and baseline model at various BPE settings for each of the $14$ languages (shown in color, with language identifier). The bold black line shows the average difference across the languages for each BPE setting.} \label{fig:deltableu} \end{figure} \section{Analysis}\label{sec:analysis} \begin{table}[] \resizebox{\columnwidth}{!}{ \small \centering \begin{tabular}{@{}p{1.75cm}lccccc@{}} \toprule \multirow{2}{*}{\textbf{Features}} & \multicolumn{3}{c}{\textbf{\begin{tabular}[c]{@{}c@{}}Corpus-\\ dependent\end{tabular}}} & \multicolumn{2}{c}{\textbf{\begin{tabular}[c]{@{}c@{}}Corpus-\\ independent\end{tabular}}} \\ & \textbf{TT} & \textbf{A} & \textbf{H} & \textbf{UT} & \textbf{UTC} \\ \midrule \textbf{Correlation} & 0.04 & 0.59 & 0.67 & 0.80 & 0.49 \\ \bottomrule \end{tabular}} \caption{The Pearsons correlation between the features and the relative gain in BLEU obtained by the CG model. See Section~\ref{sec:analysis} for details regarding features.} \label{tab:correlation} \end{table} We are interested in understanding whether our character-aware model is exploiting morphological patterns in the target language. We investigate this by inspecting the relationship between a set of hand-picked features and improvements obtained by our model over the baseline at word-level inputs. These features fall into two categories, \emph{corpus-dependent} and \emph{corpus-independent}. We following~\newcite{bentz2016comparison}, and extract features known to correlate with human judgments of morphological complexity. The following corpus-dependent features were used: \begin{enumerate}[label=(\roman*)] \item Type-Token Ratio (TT): the ratio of the number of word types to the total number of word tokens in the target side. We note that a large corpus tends to have a smaller type-token ratio compared to small corpus. \item Word-Alignment Score (A): computed as $A = \frac{\mid\text{many-to-one} \mid - \mid \text{one-to-many} \mid} { \mid \text{all-alignments} \mid}$. One-to-one, one-to-many and many-to-one alignment types are illustrated in Figure~\ref{fig:alignments}.\footnote{We use FastAlign~\cite{dyer2013simple} for word alignments with the grow-diag-final-and heuristic from~\cite{och2003systematic} for symmetrization.} We intuit that a morphologically poor source language (like English) paired with a richer target language should exhibit more many-to-one alignments---a single word in the target will contain more information (via morphological phenomena) that can only be translated using multiple words in the source. \item Word-Level Entropy (H): computed as $H = \sum_{v \in \mathcal{V}} p(v) \log p(v)$ where $v$ is a word type. This metric reflects the average information content of the words in a corpus. Languages with more dependence on having a large number of word types rather than word order or phrase structure will score higher. \end{enumerate} \begin{figure}[h] \centering \tikzset{every picture/.style={line width=1.00pt}} \begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1] \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ][line width=1.00] (90,21.5) -- (70,47) ; \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ][line width=1.00] (70,21.5) -- (70,47) ; \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ][line width=1.00] (51,22) -- (70,47) ; \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ] (46.5,45.5) -- (31,22) ; \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ] (26.5,46) -- (8,21) ; \draw [color={rgb, 255:red, 155; green, 155; blue, 155 } ,draw opacity=1 ] (10,45.5) -- (8,21) ; \draw (30,12) node {$s_{1}$}; \draw (10,12) node {$s_{0}$}; \draw (50,12) node {$s_{2}$}; \draw (70,12) node {$s_{3}$}; \draw (90,12) node {$s_{4}$}; \draw (30,52) node {$t_{1}$}; \draw (10,52) node {$t_{0}$}; \draw (50,52) node {$t_{2}$}; \draw (70,52) node {$t_{3}$}; \end{tikzpicture} \caption{Example of one-to-many ($s_0$ to $t_0,t_1$), one-to-one ($s_1$ to $t_2$) and many-to-one ($s_2,s_3,s_4$ to $t_3$) alignments. For this example $A = (3-2)/6$.} \label{fig:alignments} \end{figure} For the corpus-independent features we used a morphological annotation corpus called UniMorph~\cite{sylak2015universal}. The UniMorph corpus contains a large list of inflected words (in several languages) along with the word's lemma and a set of morphological tags. For example, the French UniMorph corpus contains the word \texttt{marchai} (walked), which is associated with its lemma, \texttt{marcher} and a set of morphological tags \textbf{$\{$\texttt{V,IND,PST,1,SG,PFV}$\}$}. There are $19$ such tags in the French UniMorph corpus. A morphologically richer language like Hungarian, for example, has $36$ distinct tags. We used the number of distinct tags (UT) and the number of different tag combinations (UTC) that appear in the UniMorph corpus for each language. Note that we do not filter out words (and its associated tags) from the UniMorph corpus that are absent in our parallel data. This ensures that the UT and UTC features are completely corpus independent. The Pearson's correlation between these hand-picked features and relative gain observed by our model is shown in Table~\ref{tab:correlation}. For this analysis we used the relative gain obtained from the word-level experiments. Concretely, the relative gain for Czech was computed as $\frac{21.49 - 18.44}{18.44}$ We see a strong correlation between the corpus-independent feature (UT) and our model's gain. Alignment score and Word Entropy are also moderately correlated. Surprisingly, we see no correlation to type-token ratio. As the correlation analysis only examines the relation between BLEU gains and an \emph{individual} feature, we further analyzed how the features \emph{jointly} relate to BLEU gains. We fitted a linear regression model, setting the relative gains as the predicted variable $y$ and the feature values as the input variables $\vec{x}$, with the goal of studying the linear regression weights $\vec{\phi}$.\footnote{The input features were min-max normalized for the regression analysis.} We used feature-augmented domain adaptation where we consider each language as a domain~\cite{daume2007frustratingly}, allowing the model to find a set of ``general'' weights as well language-specific weights that best fit the data (Equation~\ref{eq:lrobj}). The general feature weights can be interpreted as being indicative of the overall trends in the dataset across all the languages, while the language-specific weights indicate language deviation from the overall trend. \begin{align} \vec{\mathcal{L}}(\vec{\phi}) &= \sum_{i \in \mathcal{I}} \mid y_i - \tilde{y}_i \mid^2 - \lambda \mid \vec{\phi} \mid^2\\ \label{eq:lrobj} \tilde{y}_i&= \vec{\phi}_{\text{ALL}}^T\vec{x}_i + \vec{\phi}_i^T\vec{x}_i \end{align} Where, $y$ is the true relative gain in BLEU, $\tilde{y}$ is the predicted gain, $\vec{x}$ is a vector of input feature values, $\vec{\phi}_{\text{ALL}}$ and $\vec{\phi}_i$ are the general and language-specific weights, and $i$ indexes into the set of languages in our analysis. We set $\lambda$ to $0.05$. \begin{figure} \centering \includegraphics[width=0.9\columnwidth]{nobiasdata_nobpenosrc_weights.png} \caption{Feature weights of the feature-augmented language adapted linear regression model. The first row represents the ``general'' set of weights used for all of the languages. Each row below are the language-adapted weights that only ``fire'' for that specific language.} \label{fig:lrweights} \end{figure} The matrix of learned weights $\vec{\phi}$ is visualized in Figure~\ref{fig:lrweights}. The first row of weights correspond to the ``general'' weights that are used for all the languages, followed by language-specific weights sorted by relative gain. While the general weights align with the correlation results (Table~\ref{tab:correlation}), this analysis also shows that the UTC weight for Czech and Turkish are much larger than any of the other languages' and indeed we can verify that these languages have $194$ and $300$ different tag combinations while the average tag combinations is $\approx110$. From the corpus-dependent features, word alignment score strongly predicts the gain in BLEU scores. For Czech, Ukrainian, Turkish, Hungarian, and Polish we see additional weight placed on this feature. A similar trend can be seen for the word-entropy feature. While type-token ratio does not exhibit a strong overall trend, we see that Ukrainian and Farsi are outliers. Our correlation and regression analysis strongly suggest that CG character-aware modeling helps the most when the target language has inherent morphological complexity and that it does indeed have the ability to handle morphological patterns present in the target languages. \subsection{Qualitative Examples} \begin{table} \small \begin{center} \begin{tabular}{ |l|l| } \hline Src & here he is : leonardo \textbf{da vinci} . \\ Ref & h*A hw -- lywnArdw \textbf{dA fyn\$y} .\\ Std & hnA hw : lywnArdw \textbf{dA dA} .\\ CG & hnA hw : lywnArdw \textbf{dA fy+n\$y} .\\ \hline Src & i 'm the \textbf{mexican} in the family . \\ Ref & AnA \textbf{Almksyky} fy AlEA\}lp .\\ Std & AnA \textbf{mksy+Any} fy AlEA\}lp . \\ CG & AnA \textbf{Almksy+ky} fy AlEA\}lp . \\ \hline Src & there was going to be a national \textbf{referendum} . \\ Ref & wtm AlAEdAd lAHrA' \textbf{AstftA'} \$Eby .\\ Std & sykwn hnAk \textbf{f+tA'} wTny .\\ CG & sykwn hnAk \textbf{Ast+f+tA'} wTny . \\ \hline Src & there are ordinary \textbf{heroes} . \\ Ref & fhnAk \textbf{AbTAl} TbyEywn .\\ Std & hnAk \textbf{ASdqA'} EAdy .\\ CG & hnAk \textbf{AbTAl} EAdyyn .\\ \hline \end{tabular} \end{center} \caption{Examples from En-Ar, transliterated with the Buckwalter schema. We show the version of our model and the English using `+' to denote where BPE splits words up, while BPE has not been applied to the target reference. \Adi{room for one or two more example?}} \label{tab:examples} \end{table} We additionally look at specific examples of where our model is outperforming the baseline in the case of $30$k BPE in En-Ar. We see a few trends, which we show examples of in Table \ref{tab:examples}. The first trend, corresponding to the first example, is that it gets names better. This might be because Arabic is not written in the Latin alphabet, and the spelling-aware model may be able to transliterate better. Another trend is that CG gets the endings of rare words correct, in particular when the BPE segmentation is \emph{not} according to morpheme boundaries. The second example illustrates this, where the word for ``Mexican'' appears in the training data broken up by BPE with various morphological endings, all of which are spelled beginning with ``ky'' in the second subword. The morpheme boundaries here would be ``Al+mksyk+y.'' Note that CG also gets the definite article ``Al'' correct while the baseline does not. Finally, we see a pattern where our model does better for words which are rare and appear both with and without the definite article ``Al.'' Our third example in Table \ref{tab:examples} illustrates this with an infrequent word, the word for ``referendum'', which gets broken up into subwords. In particular, the first subword sometimes has an ``Al'' attached in the training data. Our model is able to translate this subword, while the baseline skips the subword altogether, outputting two subwords that alone are not a valid word. Again, the word is not broken up along morpheme boundaries by BPE. Here there would be no way to break this word up into morphological segments---it consists of non-concatenative derivational morphology. This occurs again in the fourth example in the word for ``heroes,'' where the baseline predicts the word for ``friends.'' In this case the word was not split up by BPE, but similarly it is rare but occurs with the definite article attached in the training data as well. \section{Conclusion} We extend character-aware word-level modeling to the decoder for translation into morphologically rich languages. Our improvements were attained by augmenting the softmax and the target embedding layers with character-awareness. We also find it critical to add a gating function to balance compositional embeddings with standard embeddings. We evaluate our method on a low-resource dataset translating from English into 14 languages, and on top of a spectrum of BPE merge operations. Furthermore, for word-level and higher merge hyperparameter settings, we introduced an approximation to the softmax layer. We achieve consistent performance gains across languages and subword granularities, and perform an analysis indicating that the gains for each language correspond to morphological complexity. For future work, we would like to explore how our methods might be of use in higher-resource settings. Furthermore, it would be interesting to see how these methods might interact with multilingual systems and if they might be able to improve what information is shared between related languages. \Pamela{I added a sentence to the conclusion to help us hit 8 pages and to end on a bit more of a positive note (and not the one the sassy reviewer commented on)} \section*{Acknowledgements} This project originated at the Machine Translation Marathon 2018. We thank the organizers and attendees for their support, feedback and helpful discussions during the event. This work is supported in part by the Office of the Director of National Intelligence, IARPA. The views contained herein are those of the authors and do not necessarily reflect the position of the sponsors.

e923896a3a244104a61e8ef4c38ba125026fe29b

\section{Introduction} Spin and orbital angular momentum has been a subject of fundamental interest since long ago. Recall that plane waves carry longitudinal orbital angular momentum (OAM) given by its wave vector, and also longitudinal spin angular momentum (SAM) arising from its intrinsic helicity due to circular polarization; they can be associated to, respectively, spatial and polarization degrees of freedom that can be straightforwardly separated in paraxial wavefields. Nonetheless, recent work carried out on SAM and OAM in a variety of optical fields beyond paraxial fields \cite{Aiello2015a,Bekshaev2015,Bliokh2015,Bliokh2015b,Gong2018} reveals a wealth of spin-orbit interactions (SOI) of light that are attracting a great deal of attention nowadays \cite{Bliokh2015a}. In this regard, non-paraxial, subwavelength-structured wavefields appear naturally in Plasmonics and Nano-Optics. Therein the vector nature of electromagnetic waves has to be fully accounted for, so that spatial and polarization properties can no longer be decoupled, leading to a variety of novel phenomenology and functionalities at the nanoscale where spin-orbit interactions play a crucial role \cite{Bliokh2015a}. Particularly relevant and widespread in nano-optics is the emergence of transverse SAM, theoretically described in connection with evanescent waves: such polarization-independent, transverse spin component stems from the spatially decaying field of the evanescent wave away from the interface, and carries in turn a longitudinal (Belinfante's) SAM \cite{Bliokh2014a}. Transverse SAM has been later on studied in several geometries involving well known evanescent waves \cite{Bliokh2014,OConnor2014,Aiello2015a,Bliokh2015,Sayrin2015,VanMechelen2016,Antognozzi2016,Picardi2017,Picardi2018,Neugebauer2018} such as plasmons, total internal reflection, etc. In this regard, it is worth mentioning that the following experimental works have exploited spin-orbit interaction also in waveguides \cite{Petersen2014,Mitsch2014,LeFeber2015,Alizadeh2016,Coles2016}, in most cases enforcing through the spin-orbit locking the coupling of a circularly polarized exciting field (spin locked in the evanescent tail) into propagating guided waves with ad-hoc directionality (orbit thus driven). Nonetheless, the peculiar structured wavefield arising \textit{inside} waveguides has not been addressed in detail up to now. Recall that cylindrical waveguides support a wealth of transverse and hybrid leaky/guided modes recently shown to be crucial in determining the optical properties of semiconductor nanowires in relevant problems such as photoluminescence \cite{Grzela2014,VanDam2015b,Abujetas2017} and absorption \cite{Paniagua-Dominguez2013,Abujetas2015}. Nanowires and fibers are no doubt extremely interesting optical platforms that hold potential of novel nanophotonic devices, wherein SOI and locking inside could be exploited to manipulate i.e. quantum well and dot emission and photoluminescence. Subwavelength waveguides in lower-frequency regimes are amenable to SOI phenomena; moreover, optical forces and torques in the IR to GHz domain could be manipulated inside i.e. water-filled waveguides \cite{Andryieuski2015}. In this work, we study theoretically the spin and orbital angular momenta of confined light inside waveguides. In Sec. II, we show analytically that guided modes carry transverse spin density and longitudinal (so-called) Belinfante's spin momentum inside the waveguide, connecting them to relevant magnitudes stemming from light confinement such as energy density and transverse wave vector components. This is done for two cases: simplest (planar) waveguide geometry with transverse electric/magnetic modes; and a cylindrical (nanowire) geometry (much more involved), which supports not only pure transverse modes but also hybrid modes with intrinsic helicity. Section III exploits the analytical formulation to explore spin-orbit interactions through the resulting SAM and OAM in nanophotonic waveguide geometries (semiconductor nanoslabs and nanowires). Water-filled waveguides are considered in Sec. IV to illustrate optical forces and torques in the microwave regime. Finally, conclusions are summarized in Sec. V. \section{Spin and orbital angular momenta in planar and cylindrical waveguides: Formulation} \subsection{Transverse guided modes in planar waveguides} Let us start with a simple planar waveguide (see inset in Fig.~\ref{fig_PW_geo}a) consisting of a dielectric slab (core, medium $1$) of thickness $2d$ surrounded by another dielectric material (cladding, medium $2$) that has a lower refractive index, $n_1=c\sqrt(\epsilon_1\mu_1)=\sqrt(\epsilon_{r1}\mu_{r1})>n_2$: we define both relative and absolute dielectric permittivities and magnetic permeabilities ($\epsilon_{r},\mu_{r}$ and $\epsilon,\mu$, respectively) for they will be needed below ($c=1/\sqrt{\epsilon_0\mu_0}$ is the speed of light in vacuum). We assume propagation along the $z$ direction and translational invariance along the $y$ direction. Transverse electric (TE) and magnetic (TM) modes stand for modes with only electric/magnetic field component along the $y$ axis, propagating along the $z$ axis with propagation constant $k_{z}$ and transversal wavevector $k_{t}$. For TE modes, the corresponding fields (omitting the time harmonic factor $e^{-\imath\omega t}$) are \cite{Snyder1983}: \begin{subequations}\begin{eqnarray} \mathbf{E}&&=\hat{y} \sqrt{\mu_i} A_i f(k_{t}x)e^{\imath k_{z}z}, \\ \mathbf{H}&&= A_i\frac{1}{\omega\sqrt{\mu_i}} [k_{z}f(k_{t}x)\hat{x}+\imath k_{t}f'(k_{t}x)\hat{z}] e^{\imath k_{z}z}, \end{eqnarray}\label{eq_EHTE}\end{subequations} with $i=1,2$ and \begin{subequations} \begin{eqnarray} f(x)&&=e^{-\alpha x}, x > d, \\ f(x)&&=\begin{Bmatrix} \sin(k_x x) \\ \cos(k_x x) \end{Bmatrix}, |x|< d, \\ &&=\mp e^{\alpha x}, x < -d. \end{eqnarray}\label{eq_fTE}\end{subequations} The terms in braces denoting antisymmetric (top) and symmetric (bottom) modes, respectively; $f'(k_{t}x)$ denotes derivative with respect to its argument. The electromagnetic fields for TM modes can be straightforwardly obtained from the above Eqs.~(\ref{eq_EHTE}) by replacing $\mathbf{E},\mathbf{H}\Rightarrow\mathbf{H},-\mathbf{E}$ and $\sqrt{\mu}\Rightarrow\sqrt{\epsilon}$. Field amplitudes $A_i$ outside/inside the waveguide are connected through corresponding boundary conditions. The wavevector $\mathbf{k}$ components in Cartesian coordinates of such fields depend also on the medium, and are given by: \begin{align} \mathbf{k} = (k_{t},0,k_{z}), \nonumber \\ \mathrm{with} \;\;\; k_t=\pm\imath\alpha \;(|x|> d) \;\;\; \mathrm{and} \;\;\; k_t=k_{x} \;(|x|< d), \label{eq_kt} \end{align} where we have used the same notation for the transversal component of the wavevector inside and outside the waveguide, bearing in mind that outside is complex, its sign $\imath \alpha$ depending on the considered semi-region. The wavevector components are related by: \begin{subequations}\begin{eqnarray} k_{z}^2+k_t^2=\epsilon_{r1}\mu_{r1}\left(\frac{\omega}{c}\right)^2=(n_1\frac{\omega}{c})^2, |x|> d,\\ k_{z}^2-\alpha^2=\epsilon_{r2}\mu_{r2}\left(\frac{\omega}{c}\right)^2=(n_2\frac{\omega}{c})^2, |x|< d; \end{eqnarray}\label{eq_k}\end{subequations} where $\omega$ is the angular frequency. Since the dielectric waveguide is intended to guide the light, the propagation constant $k_z$ has to be in the range $n_2/n_1<ck_{z}/\omega<n_1$ and will also depend on mode number and polarization. Indeed, upon imposing boundary conditions, we can obtain the corresponding dispersion relation (cf. e.g. Ref.~\cite{Snyder1983}) that determines the wavevectors of symmetric and antisymmetric, transverse electric and magnetic (TE and TM) guided modes (see Fig.\ref{fig_PW_geo}a). \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-wk.pdf} \caption{ (a,b) Dispersion relations for the lowest-order guided modes inside waveguides with refractive index $n=3.43$ (vacuum outside): (a) (anti-)symmetric transverse electric and magnetic (TE and TM) guided modes (slab); (b) TE, TM, and hybrid (HE) guided modes (cylinder). Circles indicate the specific modes and corresponding wavevectors considered below. Insets: Schematic of the dielectric waveguides considered hereafter: (a) slab of width $2d$ and (b) cylinder of radius $R$.} \label{fig_PW_geo} \label{fig_NW_geo} \end{figure} The energy density, defined as: \begin{equation} W=\frac{\epsilon}{4}|\mathbf{E}|^2+\frac{\mu}{4}|\mathbf{H}|^2, \label{eq_W} \end{equation} is given for both TE and TM modes by: \begin{subequations}\begin{eqnarray} W=&&|A_2|^2 \dfrac{k_{z}^2}{2\omega^2}|f(x)|^2, |x| > d, \\ =&& |A_1|^2 \dfrac{1}{2\omega^2} \left[k_{z}^2|f(x)|^2 +\frac{k_x^2}{2}\right], |x| < d, \end{eqnarray}\label{eq_WT}\end{subequations} Next, the Poynting vector density, \begin{equation} \mathbf{P}=\frac{1}{2c^2}\Re\left[\mathbf{E}^*\times\mathbf{H}\right]=\frac{1}{2c^2}\Re\left[\mathbf{E}\times\mathbf{H}^*\right], \label{eq_P} \end{equation} yields: \begin{eqnarray} \mathbf{P}=&& \hat{z}|A|^2\dfrac{k_{z}}{2c^2\omega}|f(x)|^2, \label{eq_PT} \end{eqnarray} for both TE and TM modes. Let us plot the expected electric and magnetic fields along the waveguide from Eqs.~(\ref{eq_EHTE}) and ~(\ref{eq_fTE}), for a typical guided mode (asymmetric TE) at given positions inside the waveguide yielding relevant phenomenology. In Fig.~\ref{fig_PW_EH}a (center of the waveguide for the asymmetric mode), the electric field does vanish, and the magnetic field has only a nonzero longitudinal component, the fields thus appearing as a longitudinal wave. In Fig.~\ref{fig_PW_EH}b (corresponding approx. to the center of the lobe of maximum electric field of the lowest-order asymmetric TE mode), both the electric and magnetic fields have large contributions: the electric one is perfectly transverse, while the magnetic has a strong longitudinal (out of phase) component, apart from its transverse component. This is no doubt expected to yield a strong spin. Finally, a case is shown in Fig.~\ref{fig_PW_EH}c where the wave appears purely transverse (no spin expected). All such varied behavior will certainly give rise to a rich spin phenomenology, as we will now show. \begin{figure} \includegraphics[width=1\columnwidth]{SAM-Fig-EH-Fields.pdf} \caption{ Schematic of the electric and magnetic fields along the waveguide from Eqs.~(\ref{eq_EHTE}) and ~(\ref{eq_fTE}) for an asymmetric TE mode at transverse positions inside the waveguide satisfying: (a) $k_xx=0$, (b) $k_xx=\pi/4$, and (c) $k_xx=\pi/2$. TM modes will be identical upon replacing $\mathbf{E},\mathbf{H}\Rightarrow\mathbf{H},-\mathbf{E}$ . It follows that, as will be shown below, for TE (respectively, TM) modes, the spin inside the waveguide is $S_e\equiv 0$ (respectively, $S_m\equiv 0$) everywhere, and $S_m\not = 0$ (respectively, $S_e \not = 0$) [as in (b)] except for specific $x$-points where (a) or (c) is satisfied. } \label{fig_PW_EH} \end{figure} \subsection{Spin density and momentum in planar waveguides} The spin density is defined as: \begin{equation} \mathbf{S}=\frac{\epsilon_0\mu_r^{-1}}{4\omega}\Im\left[\mathbf{E^*}\times\mathbf{E}\right]+\frac{\mu_0\epsilon_r^{-1}} {4\omega}\Im\left[\mathbf{H^*}\times\mathbf{H}\right]=\mathbf{S}_e+\mathbf{S}_m. \label{eq_s} \end{equation} Recall that the helicity, \begin{eqnarray} h=&& - \frac{\sqrt{\epsilon_{0}\mu_{0}}}{2\omega}\Im[\mathbf{E^*}\cdot\mathbf{H}], \label{eq_h} \end{eqnarray} of all guided modes in this planar waveguide is $h=0$. From the electric and magnetic field components, Eqs. (\ref{eq_EHTE}) and related, it follows that $\mathbf{S}_e=0$ ($\mathbf{S}_m=0$) for TE (respectively, TM) guided modes. The only nonzero components of the spin densities are $\mathbf{S}_m=\hat{y}S_m$ for TE modes and $\mathbf{S}_e=\hat{y}S_e$ for TM modes. Upon introducing them into Eq. (\ref{eq_s}), we end up with the resulting total spin density: \begin{subequations}\begin{eqnarray} \mathbf{S}=\hat{y}&& |A_2|^2\dfrac{k_{z}\alpha}{2\omega^3n_2^2} e^{-2\alpha x}, x > d, \\ =\hat{y}&& \begin{Bmatrix} - \\ + \end{Bmatrix} |A_1|^2 \dfrac{k_{z}k_x}{4\omega^3n_{1}^2} \sin(2k_x x), |x| < d,\label{eq_sTEM_in} \\ =-\hat{y}&&|A_2|^2\dfrac{k_{z}\alpha}{2\omega^3n_2^2} e^{2\alpha x}, x < -d; \end{eqnarray}\label{eq_sTEM}\end{subequations} for both TE and TM modes, where we have made use of $\epsilon_r\mu_r=n^2$. Note that it can be written in a compact form as follows: \begin{eqnarray} \mathbf{S}=\hat{y}&& |A_2|^2\dfrac{k_{z}k_{t}}{4\omega^3n_2^2} \dfrac{d}{d(k_{t}x)}|f(k_{t}x)|^2. \end{eqnarray} Three features should be emphasized at this point regarding the spin density above (\ref{eq_sTEM}): (i) unlike that of the evanescent field, it can be locally positive or negative inside the waveguide [cf. (\ref{eq_sTEM})]; (ii) it is either purely magnetic ($S_e=0$) or electric ($S_m=0$) for, respectively, TE or TM guided modes, and (iii) it is proportional to the transverse component of the wavevector $k_{t}$. We now turn to calculate the extraordinary spin momentum, obeying: \begin{equation} \mathbf{P}^S=\frac{1}{2}\nabla\times\mathbf{S}. \label{eq_ps} \end{equation} Since the spin density has only a nonzero component along the $y$ axis [cf. Eqs. (\ref{eq_sTEM})], which in turn does not depend on the propagation direction $z$, the spin momentum has only a nonzero component along $z$, namely: \begin{eqnarray} \mathbf{P}^S =\hat{z}&& |A|^2\dfrac{k_{z}k_{t}^{2}}{8\omega^3n^2} \dfrac{d^{2}}{d(k_{t}x)^{2}}|f(k_{t}x)|^2. \label{eq_psTEM} \end{eqnarray} for both TE and TM modes. As the spin density, the spin momentum is either purely magnetic (TE modes) or electric (TM modes). Finally, the canonical (orbital) part $\mathbf{P}^O$ of the momentum density can be simply obtained from: \begin{equation} \mathbf{P}=\mathbf{P}^O+\mathbf{P}^S. \label{eq_po} \end{equation} Therefore, from Eqs. (\ref{eq_PT}) and (\ref{eq_psTEM}), it follows that: \begin{subequations}\begin{eqnarray} \mathbf{P}^O= \hat{z}&& |A_2|^2\dfrac{k_{z}^3}{2n_2^2\omega^3} e^{\mp 2\alpha x}, \;\;|x| > d, \\ =\hat{z}&& |A_1|^2 \dfrac{k_{z}}{2n_{2}^2\omega^3} \left[k_{z}^2|f(x)|^2+\frac{k_x^2}{2}\right], |x| < d, \end{eqnarray}\label{eq_PO}\end{subequations} for both TE and TM modes. Incidentally, we have verified that the same $\mathbf{P}^O$ is obtained if directly calculated from its canonical expression \cite{Bliokh2015b}. Nonetheless, the orbital momentum presents both contributions, electric and magnetic. It follows from the latter equations that the orbital momentum in both polarizations is proportional to the wavevector along the propagation direction in the expected manner: \begin{eqnarray} \frac{P^O_z}{W}=\frac{k_{z}}{\omega n^2}. \label{eq_poW} \end{eqnarray}In addition, it can be shown that the spin momentum can be expressed as: \begin{align} \frac{P^S_z}{W} = \dfrac{k_{t}^2}{\omega n^2k_{z}} [1-\tilde{P}_{NP}(\mathbf{r})], \label{eq_PSW_NP} \end{align} where $\tilde{P}_{NP}(\mathbf{r})$ is a non-paraxial term (vanishes for paraxial waves), \begin{align} \tilde{P}_{NP}(\mathbf{r}) = \dfrac{1}{8k_{t}^{2}} \left\lbrace (\mathbf{k}\cdot\mathbf{E})(\mathbf{k}\cdot\mathbf{E}^{*})\epsilon + (\mathbf{k}\cdot\mathbf{H})(\mathbf{k}\cdot\mathbf{H}^{*})\mu \right. \nonumber \\ + \left[ \mathbf{k}\times \mathbf{E} - \mu\omega \mathbf{H} \right]\left[ \mathbf{k}\times \mathbf{E}^{*} - \mu\omega \mathbf{H}^{*} \right]\epsilon \nonumber \\ \left. + \left[ \mathbf{k}\times \mathbf{H} + \epsilon\omega \mathbf{E} \right]\left[ \mathbf{k}\times \mathbf{H}^{*} + \epsilon\omega \mathbf{E}^{*} \right]\mu \right\rbrace. \label{eq_nP} \end{align} Therefore, it is evident from Eqs.~(\ref{eq_sTEM_in}) and~(\ref{eq_psTEM}), the main results of this subsection, that a transverse SAM arises inside planar waveguides, which in turn yields an extraordinary longitudinal spin momentum, both proportional to the guided mode transverse wavevector component. The contribution from the longitudinal spin momentum, which, unlike in the evanescent region, can be positive or negative inside the waveguide (as we will show below), is indeed crucial to retrieve the proper dependence of the canonical momentum on mode wavevector~(\ref{eq_poW}). Before discussing in detail these terms, we will show next that similar transverse SAM arise in a more complex waveguide geometry in order to assess the universal character in connection to guided light. \subsection{Transverse and hybrid guided modes in cylindrical waveguides} Let us now study cylindrical waveguides, which support a wealth of guided modes exhibiting a rich spin phenomenology. The fields of the waveguide can be expressed in cylindrical coordinates [see Fig.~\ref{fig_NW_geo}(b)] as follows: \begin{subequations}\begin{align} & E_{r} = \sum{\left[ \frac{\imath k_{z}}{k_{t}} Z_{m}^{'} \left( k_{t} r \right) a_{m} - \frac{\mu \omega m}{k_{t}^{2} r} Z_{m} \left( k_{t} r \right) b_{m} \right] F_{m}}, \\ & E_{\theta} = - \sum{\left[ \frac{mk_{z}}{k_{t}^{2} r} Z_{m} \left( k_{t} r \right) a_{m} + \frac{\imath \mu \omega}{k_{t}} Z_{m}^{'} \left( k_{t} r \right) b_{m} \right] F_{m}}, \\ & E_{z} = \sum{\left[ Z_{m} \left( k_{t} r \right) a_{m} \right] F_{m}}, \\ & H_{r} = \sum{\left[ \frac{m k^{2}}{\mu \omega k_{t}^{2} r} Z_{m} \left( k_{t} r \right) a_{m} + \frac{\imath k_{z}}{k_{t}} Z_{m}^{'} \left( k_{t} r \right) b_{m} \right] F_{m}}, \\ & H_{\theta} = \sum{\left[ \frac{\imath k^{2}}{\mu \omega k_{t}} Z_{m}^{'} \left( k_{t} r \right) a_{m} - \frac{mk_{z}}{k_{t}^{2} r} Z_{m} \left( k_{t} r \right) b_{m} \right] F_{m}}, \\ & H_{z} = \sum{\left[ Z_{m} \left( k_{t} r \right) b_{m} \right] F_{m}}, \end{align}\label{eq_EHNW}\end{subequations} where $k_{z}$ and $k_{t}$ are again the longitudinal and transverse component component of the wavevector, respectively, and $k = |\mathbf{k}|$ is the modulus of the wavevector. Recall that $k_{t}$ is imaginary outside the waveguide. The function $Z_m\left( k_{t} r \right)$ is the appropriate Bessel function that matches the boundary conditions \cite{Paniagua-Dominguez2013,Snyder1983} and $Z'_m\left( k_{t} r \right) $ denotes the derivative with respect to the argument. Upon using functions $Z_m$, we collapse EM fields inside and outside the waveguide into formally identical equations. In this regard, recall that $a_m,b_m$, ..., obey different expressions inside/outside. The function $F_{m}= e^{\imath \left( m\theta +k_{z}-\omega t \right)}$ is the phase of the wave, the subscript $m$ being an integer related to its azimuthal order. Bear in mind that the longitudinal and transverse components of the wavevector depend on $m$, but we dropped the subscript for simplicity. For guided modes, the wavevector components are defined as in Eq.~(\ref{eq_k}) (assuming vacuum outside $\epsilon_r=\mu_r=1$), replacing the transverse components $k_x$ and $\imath\alpha$ in Eq.~(\ref{eq_kt}) by its transversal (radial) $k_t$ and $\imath\alpha$ counterparts associated with the cylindrical geometry. The fields of the waveguide written in Eq~(\ref{eq_EHNW}) are expressed as the sum over different guided modes, encoded in the subindex $m$. The guided modes satisfy the corresponding dispersion relation (not shown here, cf. Refs. \cite{Paniagua-Dominguez2013,Abujetas2015,Snyder1983}); solutions can be associated to each guided mode, labeled by a pair index $ml$, where $m=0,\pm 1,\pm 2,\ldots$ is the azimuthal index and $l=1,2,3,\ldots$ the radial index. For guided modes with $m=0$, the field is symmetric about the axis, exhibiting a pure transverse character, either electric (TE$_{0l}$, $a_m=0$ so that $E_r = E_z = H_{\phi}=0$) or magnetic (TM$_{0l}$, $b_m=0$ so that $H_r = H_z = E_{\phi}=0$). Hybrid modes arise for $m\neq 0$ (HE$_{ml}$). We show in Fig.~\ref{fig_NW_geo}(b) the dispersion relation of the lowest-order modes: TM$_{01}$, TE$_{01}$, and the hybrid HE$_{11}$ (no cutoff, lowest-order mode), which are the guided modes we will considered explicitly below. We now proceed to calculate the generic expressions for the energy density, Poynting vector, and helicity, from Eqs.~(\ref{eq_W}), (\ref{eq_P}), and~(\ref{eq_h}), for arbitrary guided modes in a cylindrical lossless waveguide with electromagnetic fields given by Eqs.~(\ref{eq_EHNW}). This is done in the Supplemental Material \cite{Supp}. In addition, we also show therein the generic expressions for all spin-related magnitudes, namely: spin density and momentum, Eqs.~(\ref{eq_s}) and~(\ref{eq_ps}), and resulting orbital momentum~(\ref{eq_po}). Interestingly, we should emphasize that the orbital momentum $P^O$ satisfies the expected dependence on longitudinal wavevector $k_{z}$ as follows: \begin{equation} \dfrac{P^O}{W}=\dfrac{k_{z}}{\omega n^2}. \label{eq_POWNW} \end{equation} Let us now discuss all these relevant magnitudes for given lowest-order guided modes with most relevant symmetries. \subsection{Transverse guided modes: Confinement-induced SAM} In the case of transverse guided modes ($m=0$), it is evident from Eq.~(S6) that the helicity vanishes ($h=0$). Thus the spin density should vanish except for the evanescent component of the EM fields outside the waveguide. However, it follows from Eq.~(S8), for the specific case of TE modes, that the spin density vanishes neither outside (as expected) nor inside the waveguide: \begin{subequations}\begin{eqnarray} && S_{r} =S_{z} = 0, \\ && S_{\theta} = \dfrac{\mu k_{z}}{2n^{2}\omega k_{t}} \left|b_{0}\right|^{2} Z_{0}^{*} \left( k_{t} r \right) Z_{0}^{'}\left( k_{t} r \right); \end{eqnarray}\label{eq_SNW_TX}\end{subequations} which leads to an extraordinary spin momentum with a longitudinal component, as follows: \begin{equation} P_{z}^{S} = \dfrac{\mu k_{z}}{4n^{2}\omega} |b_{0}|^{2} \left[ \dfrac{k_{t}^{2}}{|k_{t} |^{2}}\left|Z_{0}^{'} \left( k_{t} r \right)\right|^{2} - \left|Z_{0} \left( k_{t} r \right)\right|^{2}\right]. \label{eq_PSNW_TX}\end{equation} The electromagnetic properties for the TM modes follow the same expressions as for TE waves after replacing $|b_{0}|^{2}\mu \rightarrow |a_{0}|^{2}\epsilon$. Importantly, the contribution to the spin density and momentum for TE (TM) waves is fully magnetic (electric). Actually, if we revisit Eqs.~(\ref{eq_EHNW}) for TE guided modes, we realize that, at given transverse positions within the cylindrical waveguide, the electric and magnetic fields exhibit a behavior analogue to that shown in Fig.~\ref{fig_PW_EH} for planar waveguides, with an imaginary longitudinal component of $\mathrm{H}$ leading to the magnetic spin density (for TM modes, it is the complex $\mathrm{E}$ which yields the electric spin density). Moreover, after rewriting the energy density, Eq.~(S2), for pure transverse modes as: \begin{equation} W = \dfrac{1}{4}|b_{0}|^{2}\mu \left[\dfrac{ k_{z}^{2} + k^{2} }{|k_{t} |^{2}}\left|Z_{0}^{'} \left( k_{t} r \right)\right|^{2} + \left|Z_{n} \left( k_{t} r \right)\right|^{2} \right], \label{eq_WNW_TX} \end{equation} Remarkable, the spin momentum satisfies the same expression as for planar waveguides: \begin{align} \frac{P^S_z}{W} = \dfrac{k_{t}^2}{\omega n^2k_{z}} [1-\tilde{P}_{NP}(\mathbf{r})], \label{eq_PSWNW_NP} \end{align} with the non-paraxial term $\tilde{P}_{NP}$ given by Eq.~(\ref{eq_PSW_NP}). Recall that, for weakly guided waveguides, the electromagnetic field inside becomes paraxial \cite{Snyder1983}; this confirms the crucial role of confinement in the emergence of large spin density and momentum inside waveguides. Therefore, the transverse spin density and longitudinal spin momentum inside both planar and cylindrical waveguides do not vanish (despite not being evanescent) and depend on the transverse momentum $k_{t}$; whereas the orbital momentum, Eq.~(S13), has been shown above to depend as expected on the guided mode wavevector along the propagation direction $k_{z}$. Incidentally, transverse momentum has been associated to an effective mass $k_t\sim m$ if we rewrite Eq.~(\ref{eq_k}) as \cite{Wang2007b,Zang2016} $\omega= \sqrt{k_{t}^2+k_{z}^2}\propto\sqrt{m^2+p^2}$, which thus formally underlies the emergence of extraordinary SAM for guided modes. \subsection{Hybrid guided modes: Intrinsic-helicity-induced spin} We now turn to study the spin angular momentum of the HE$_{11}$ hybrid guided mode, which in fact exhibit strong intrinsic helicity given by: \begin{align} & h = \dfrac{1}{2 \omega n} \left\lbrace \dfrac{2k_{z} k}{|k_{t} |^{2}k_{t} r} Z_{1} \left( k_{t} r \right) Z_{1}^{'*}\left( k_{t} r \right)\left[|a_{1}|^{2}\epsilon + |b_{1}|^{2}\mu \right] \right. \nonumber \\ &\left. + \left[ \left|Z_{1} \left( k_{t} r \right)\right|^{2} + \left( \left|Z_{1}^{'} \left( k_{t} r \right)\right|^{2} + \dfrac{ \left|Z_{1} \left( k_{t} r \right)\right|^{2}}{|k_{t} |^{2} r^{2}} \right) \dfrac{k_{z}^{2} + k^{2}}{|k_{t} |^{2}} \right] \right. \nonumber \\ &\hspace{1in}\left. \times \Im\left[a_{1}b_{1}^{*} \right]\sqrt{\epsilon\mu} \right\rbrace ; \label{eq_hNW_HE} \end{align} the corresponding energy density from Eq.~(\ref{eq_W}) is: \begin{eqnarray} && W = \dfrac{1}{4} \left\lbrace\left[ \left|Z_{1} \left( k_{t} r \right)\right|^{2} + \left( \left|Z_{1}^{'} \left( k_{t} r \right)\right|^{2} + \dfrac{ \left|Z_{1} \left( k_{t} r \right)\right|^{2}}{|k_{t} |^{2} r^{2}} \right) \right.\right. \nonumber \\ && \hspace{1in} \left.\times \dfrac{k_{z}^{2} + k^{2}}{|k_{t} |^{2}} \right] \left[|a_{1}|^{2}\epsilon + |b_{1}|^{2}\mu \right] \nonumber \\ && \hspace{.5in}\left.+ \dfrac{8\sqrt{\epsilon \mu} }{|k_{t} |^{2}} \dfrac{k_{z}k}{k_{t} r} Z_{1} \left( k_{t} r \right) Z_{1}^{'*}\left( k_{t} r \right) \Im\left[a_{1}b_{1}^{*} \right]\right\rbrace. \label{eq_WNW_HE} \end{eqnarray} We can write Eq.~(S8), which becomes in this case very involved, as follows: \begin{align} &S_{r} = 0, \\ &S_{\theta}= \dfrac{1}{2\omega\epsilon_{r}\mu_{r}} \left[ \dfrac{k_{z}}{k_{t}} Z_{1}^{*} \left( k_{t} r \right) Z_{1}^{'}\left( k_{t} r \right)\left[|a_{1}|^{2}\epsilon + |b_{1}|^{2}\mu \right] \right. \nonumber \\ &\hspace{1cm} \left. + \dfrac{2k}{k_{t}^{2}r} \left|Z_{1} \left( k_{t} r \right)\right|^{2} \Im\left[a_{1}b_{1}^{*} \right] \sqrt{\epsilon \mu} \right], \\ & S_{z} = \dfrac{|k_{t} |^{-2}}{2\omega\epsilon_{r}\mu_{r}}\left\lbrace \dfrac{k_{z}^{2} + k^{2} }{k_{t}r}Z_{1} \left( k_{t} r \right) Z_{1}^{'*}\left( k_{t} r\right) \left[|a_{1}|^{2}\epsilon + |b_{1}|^{2}\mu \right] \right. \nonumber \\ & \hspace{1cm}\left. +2k_{z}k \left[ \left|Z_{1}^{'} \left( k_{t} r \right)\right|^{2} + \dfrac{ \left|Z_{1} \left( k_{t} r \right)\right|^{2}}{|k_{t} |^{2} r^{2}}\right] \Im\left[a_{1}b_{1}^{*} \right]\sqrt{\epsilon \mu} \right\rbrace. \label{eq_SNW_HE}\end{align} Now it is actually not trivial to relate the spin density from Eqs.~(S8) either to the helicity~(\ref{eq_hNW_HE}) or to the energy density (\ref{eq_WNW_HE}). However, it is evident from Eqs. ~(\ref{eq_SNW_HE}), that: (i) the transverse spin density $S_{\theta}$ includes both electric and magnetic contributions (first term) along with a new hybrid contribution (second term); (ii) the spin density yields a longitudinal contribution $S_z$ stemming from intrinsic helicity. The corresponding expressions for the orbital and spin momenta ($P_{z}^{O}$, $P_{\theta}^{O}$, $P_{z}^{S}$ and $P_{\theta}^{S}$) are given in the Supplemental Material \cite{Supp}. Interestingly, the longitudinal components of the orbital and the spin angular momentum can be related to the energy as follows: \begin{equation} \dfrac{P_{z}^{O}}{W} = \dfrac{k_{z}}{\omega n^{2}}, \quad \dfrac{P_{z}^{S}}{W} = \dfrac{k_{t}^{2}}{n^{2}k_{z}\omega } \left[1-\tilde{P}_{NP}\right], \label{eq_POWNW_HE}\end{equation} as in the case of pure transverse modes shown above. \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-PW-TEasym.pdf} \caption{(a) Schematic depicting the orientation of orbital $\mathbf{P}^O$ (red arrows) and spin $\mathbf{P}^S$ (green arrows) momenta, inside the planar waveguides, with loops inside illustrating the confinement-induced spin density $\mathbf{s}$. (b-e) The anti-symmetric lowest-order TE mode is considered for a planar waveguide of width $2d$ and $\epsilon_r=11.76$ for normalized half-width $\omega d/c$=0.54 (which corresponds to an InP nanoslab with thickness $2d=150$ nm at $\lambda=870$ nm). (b,d) Contour map and radial dependence of the energy density $W$, the latter including also the orbital $P_z^O$ momentum. (c,e) Contour map and radial dependence of the only non-zero component of the spin density $S_y$, the latter (e) multiplied by the transverse wavevector component and including also the only non-zero component of the spin angular momentum $P_z^S$ (normalized by $P_z^O$).} \label{fig_PW} \end{figure} \section{Spin-orbit interactions inside nanophotonic waveguides} To shed light onto the emergence of confinement-induced spin angular momentum, let us plot all relevant magnitudes for planar and cylindrical waveguides with $\epsilon_r=11.76$. Incidentally, this choice of refractive index makes our results applicable throughout the visible and near-infrared to many semiconductors with similar refractive index \cite{Grzela2014,Abujetas2017,Paniagua-Dominguez2013,Sell2016}, such as crystalline Si, GaP, GaAs, InP, etc.; indeed, it corresponds approximately to the photoluminescence band of InP at $\lambda\sim 870$ nm, which could be illustrative of SOI in emission processes inside InP nanostructures, such as slabs or nanowires \cite{Grzela2014}, especially if circularly polarized (or even more complex) dipole sources could be engineered \cite{Picardi2018}. In fact, circularly polarized luminescence has been recently reported e.g. from coupled InGaN/GaN quantum well and quantum dots structure \cite{Yu2016}, from perovskite nanocrystals suitable to deploy flexible devices \cite {Shi2018a}, also from colloidal CdS quantum dots \cite{Naito2010,Huo2017}. The latter colloidal quantum dots could be exploited in micron-sized, liquid-filled waveguides, in which the liquid itself might play the role of the denser optical medium (its refractive index needn't be very high indeed). On the other hand, recall that magnetic dipole emission is also available through lanthanide-doped nanostructures and nanoparticles \cite{Karaveli2011,Taminiau2012,Liu2013k}. First, we show in Fig.~\ref{fig_PW} the simplest case: a semiconductor nanoslab supporting the first-order anti-symmetric TE mode (see the mode dispersion relation in Fig.~\ref{fig_PW_geo}). The transverse energy density $W$ is plotted for the sake of comparison in Figs.~\ref{fig_PW}b (transverse color map) and~\ref{fig_PW}d; in the latter the orbital momentum is also included to explicitly show that the expected direct proportionality $P^O_z/W= k_z/(\omega n^2)$ is satisfied [Eq.~(\ref{eq_poW}), note that the factor $n$ is different inside/outside]. Transverse spin momentum, Eq.~(\ref{eq_sTEM_in}), arises inside the waveguide with a rich phenomenology depending on the mode symmetry, rotating differently on each half waveguide (evident from the sign of $S_y$ in Fig.~\ref{fig_PW}c,e and depicted also by loops). In turn, it results in a corresponding longitudinal spin momentum [Eq.~(\ref{eq_psTEM})] which can point along (or opposite) to the guided mode propagation direction, and thus to the orbital angular momentum (evident from the sign of $P^S_z$ and depicted by arrows). Moreover, it is simultaneously parallel and antiparallel on each half waveguide. In fact, the latter behavior can be understood in light of the complex electric and magnetic fields plotted in Fig.~\ref{fig_PW_EH}. The spin density maxima occur at symmetric positions inside the waveguide where electric and magnetic fields rotate as in Fig.~\ref{fig_PW_EH}b. At the waveguide center, the spin density vanishes as a result of the vanishing of the electric field (see Fig.~\ref{fig_PW_EH}b); the spin density also vanishes near the waveguide boundaries, at positions where the electric and magnetic fields are locally paraxial as in Fig.~\ref{fig_PW_EH}c. Finally, note that outside the waveguide, the transverse spin density, Eq.~(\ref{eq_sTEM}), stems from the evanescent character of the guided modes outside, as expected from the case of pure evanescent waves \cite{Bliokh2014a}; however, its associated longitudinal spin momentum [Eq.~(\ref{eq_psTEM})] outside the waveguide (unlike inside) always points opposite to the orbital momentum direction. Analogous results for TM modes are included in the Supplemental Material (see Fig. S1 \cite{Supp}) that confirm the emergence of transverse spin density with extraordinary longitudinal momentum; as expected, since the TM mode shown in Fig. S1 is more confined than the TE mode shown in Fig.~\ref{fig_PW}, the evanescent SAM reaches larger values but decays more abruptly away from the waveguide. The behavior for symmetric modes (not shown here) is very similar, except for the fact that spin density/momentum is symmetric/antisymmetric instead. Finally, higher-order guided modes do preserve the symmetry and polarization of the spin-related magnitudes discussed above, introducing additional number of alternating-sign layers (see e.g. Fig.~\ref{fig_GHz} below). \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-NWTE-D270.pdf} \caption{(a) Schematic depicting the orientation of orbital (red arrows) and spin (green arrows) momenta, along with spin density (loops) inside the cylindrical waveguides; as in Fig.\protect{\ref{fig_PW}}. (b-e) The lowest-order TE$_{01}$ guided mode for a cylindrical waveguide of radius $R$ and $\epsilon_r=11.76$ is considered for a normalized radius $\omega R/c=0.83$ (which corresponds to an InP nanowire with $2R=230$ nm at $\lambda=870$ nm): (b,d) Contour map and radial dependence of the energy density $W$, the latter including also the orbital $P_z^O$ momentum. (c,e) Contour map and radial dependence of the only non-zero component of the spin density $S_{\theta}$, the latter (e) multiplied by the transverse wavevector component and including also the only non-zero component of the spin angular momentum $P_z^S$ (normalized by $P_z^O$). } \label{fig_NW_TX} \end{figure} Next, let us plot all spin-related magnitudes for the lowest-order TE guided mode (zero helicity) in a cross section of the cylindrical waveguide with $\epsilon_r=11.76$ in Fig.~\ref{fig_NW_TX}. First of all, a 3D schematic illustrating the orientation of the corresponding spin (loops) and momenta (arrows) is depicted in Fig.~\ref{fig_NW_TX}a, clearly revealing its transverse character with axial symmetry. The energy density $W$ is included for the sake of completeness in Fig.~\ref{fig_NW_TX}b, revealing the strong confinement of this TE$_{01}$ guided mode for the choice of parameters. Its radial dependence is explicitly plotted in Fig.~\ref{fig_NW_TX}d, along with the orbital angular momentum, to confirm again its linear dependence $P^O_z/W\sim k_z/(\omega n^2)$, Eq.~(\ref{eq_POWNW}). The emergence of the transverse confinement-induced SAM becomes evident in the right column. First, a strong contribution within the cylinder to the azimuthal spin density [electric/magnetic for TE/TM modes, cf. Eq.~(\ref{eq_SNW_TX})], is observed in the color map in Fig.~\ref{fig_NW_TX}c, which rotates about the cylinder axis. Its radial dependence is shown explicitly in Fig.~\ref{fig_NW_TX}e: note that it is zero in the center and close to the boundary, achieving its maximum within a ring inside the cylinder. Such behavior can be fully understand in light of the complex electric and magnetic fields represented in Fig.~\ref{fig_PW_EH}, also discussed in connection to Fig.~\ref{fig_PW}, bearing in mind that the translational invariance along the transverse direction in the planar waveguide is replaced by the axial symmetry of this cylindrical waveguide. The corresponding extraordinary longitudinal spin momentum [Eq.~(\ref{eq_PSNW_TX})] is also strong inside the waveguide, concentrated at the center and pointing along the cylinder axis (see Fig.~\ref{fig_NW_TX}e), opposite to the guided mode propagation direction, except for a thin corona near the exterior boundary where it is parallel. A weaker transverse spin density is also observed in Fig.~\ref{fig_NW_TX}c and.~\ref{fig_NW_TX}e, which stems from the evanescent character of the guided modes outside [Eq.~(\ref{eq_SNW_TX})]. Indeed, the spin rotation and momentum direction outside (antiparallel with respect to orbital momentum), unlike those inside, is also fixed, similarly to what was observed in planar waveguides (see Fig.~\ref{fig_PW}). \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-NWHE-D203.pdf} \caption{(a) Schematic depicting the orientation of orbital (red arrows) and spin (green arrows) momenta inside the cylindrical waveguides; loops illustrate the transverse (respectively, longitudinal) spin densities induced inside the waveguide by the confined nature (respectively, helicity) of the hybrid guided modes. The lowest-order hybrid HE$_{11}$ guided mode for a cylindrical waveguide of radius $R$ and $\epsilon_r=11.76$ is considered for a normalized radius $\omega R/c=0.73$ (which corresponds to an InP nanowire with $D=2R=230$ nm at $\lambda=870$ nm): (b,d) Contour map and radial dependence of the energy density $W$, the latter including also the orbital $P_z^O$ momentum. (c,e) Contour map and radial dependence of the only non-zero component of the spin density $S_{\theta}$, the latter (e) multiplied by the transverse wavevector component and including also the only non-zero component of the spin angular momentum $P_z^S$ (normalized by $P_z^O$). (f,h) Radial dependence and contour map and of the helicity density $h$, including also in (f) the helicity-induced transverse orbital momentum $P^O_{\theta}$. (g,i) Radial dependence and contour map in the transverse plane of the helicity-induced longitudinal spin density $S_z$, the former (g) multiplied by the transverse wavevector component, and including also the related transverse spin angular momentum $P_{\theta}^S$ (normalized by $P_{\theta}^O$); dark loops and cyan/magenta arrows indicate spin rotation and transverse spin/orbital momenta. } \label{fig_NW_HE} \end{figure} Slight differences arise depending on whether the guided modes is weakly or strongly confined. What can we expect for the transverse spin from Eq.~(\ref{eq_SNW_TX})? Actually, it depends linearly on the transverse component $k_t$ of the wavevector inside. Recall that this component is smaller the weaker the confinement is. Nonetheless, the energy density $W$, which increases inside the waveguide with increasing confinement, compensates such decrease in a non trivial manner, so that a compromise between transverse wavevector and energy confinement yields the optimum transverse spin; as an example, we show in Fig. S2 the same results as in Fig.~\ref{fig_NW_TX}, but for a TE$_{01}$ guided mode that is weakly confined. Transverse spin density inside the waveguide in Fig. S2 is indeed comparable to that in Fig.~\ref{fig_NW_TX}; differences are in turn more obvious in the transverse spin in the evanescent region, wherein the expected behavior is observed \cite{Bliokh2014a}: more spread outside the cylinder in the latter case (Fig. S2), but larger close to the waveguide boundary in the former (strongly confined case, see Fig.~\ref{fig_NW_TX}). For the sake of completeness, a TM$_{01}$ guided mode is also shown in Fig. S3; apart from the (relevant) fact that the spin density is entirely electric ($S=S_e$), the qualitative behavior is very similar to that of the TE$_{01}$ guided mode shown in Fig.~\ref{fig_NW_TX}. Higher-order transverse modes (not explicitly shown here), as for planar guided modes, essentially preserve the (in this case, rotational) symmetry and polarization of the spin-related magnitudes, introducing additional lobes in the radial dependence in accordance with the radial mode order $l$ (see e.g. Fig.~\ref{fig_GHz} below). Nonetheless, this could lead to various rings alternating spin rotation inside the cylindrical waveguide, preserving for all the modes the continuity across the cylinder boundary of the outermost ring (even if very thin) with respect to the evanescent-region spin, which remains unaltered. In order to shed light on the different contributions to the spin density and angular momentum for hybrid modes, we now plot them making special emphasis on separating intrinsic-helicity terms from transverse confinement effects. We show all of them in Fig.~\ref{fig_NW_HE} for a HE$_{11}$ guided mode in a cross section of the cylindrical waveguide with $\epsilon_r=11.76$, along with 3D schematics illustrating the orientation of the corresponding vectors; the transverse confinement-induced spin (loops) and related momenta (green arrows) are shown in Fig.~\ref{fig_NW_HE}c,e, whereas the longitudinal helicity-induced spin and momenta are shown in Fig.~\ref{fig_NW_HE}g,i. A color map of the energy density $W$ revealing mode confinement is included in Fig.~\ref{fig_NW_HE}b, with its radial dependence and that of the orbital momentum explicitly shown in Fig.~\ref{fig_NW_HE}d; the corresponding intrinsic helicity $h$ is included in a contour map in Fig.~\ref{fig_NW_HE}h for the sake of comparison, showing also explicitly its radial dependence in Fig.~\ref{fig_NW_HE}c, along with that of the helicity-induced transverse orbital momentum $P^O_{\theta}$. First, a strong transverse (respectively, longitudinal) contribution to the spin density (respectively, extraordinary longitudinal momentum) is observed inside, which stems from the transverse confinement ($k_t$-related) as above. Note that the corresponding SAM inside is also concentrated inside an inner ring; however, unlike for transverse modes, it rotates similarly to the evanescent-region SAM, and its longitudinal spin momentum points towards the OAM inside (being opposite outside, as expected for evanescent waves). This behavior reveals a richer phenomenology for the confinement-induced transverse SAM governed by guided mode symmetry. Second, it is clear that the intrinsic helicity governs the emergence of a strong longitudinal SAM, which leads to an azimuthal spin momentum density, both resembling the spatial pattern of the helicity density (see Fig.~\ref{fig_NW_HE}h). Its behavior does not differ much from the helicity-induced spin exhibited by circularly polarized plane and/or evanescent waves, although again guided mode symmetry renders its phenomenology much richer. Higher-order hybrid guided modes (HE$_{ml}$) make the analysis more complex. The rotational symmetry of the all spin-related magnitudes (not shown here) is preserved, introducing however alternating spin-sign rings in the radial dependence when the radial index is $l>1$, similarly to higher-order transverse modes mentioned above. The azimuthal index $m$ is directly connected to the helicity [cf. Eq.~(S6)], as expected; in addition, for $m\geq 2$ the dependence on Bessel functions of the electromagnetic fields inside imposes $W=0$ at the waveguide center. \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-PW-GHz.pdf} \caption{(a,c) Energy densities $W$ (left column) and (b,d) spin densities (right column) for anti-symmetric TE$_{2}$ (top) and TM$_{2}$ (bottom) modes in a planar waveguide of width $2d$ and $\epsilon_r=80$ for normalized half-width $\omega d/c=0.36$ (which corresponds to water-filled slabs with thicknesses $2d=3.4$ cm at $\nu=\omega/(2\pi)=1$ GHz). Arrows in (a,c) indicate the radiation pressure force felt by electric ($p$) and magnetic ($m$) dipole particles due to longitudinal momenta, whereas loops in (b,d) reveal the corresponding torque induced inside the waveguides by transverse confinement-induced spin densities. } \label{fig_GHz} \end{figure} \section{Optical radiation forces and torques in water-filled waveguides} Finally, let us briefly explore the impact on optical forces of the transverse SAM inside waveguides shown above. As an example, we consider water-filled planar channels (inside e.g. metal plates) in the GHz domain; recall that water in this regime exhibits a large refractive index with relatively low absorption, which makes it suitable for tunable high-index-dielectric metamaterials \cite{Andryieuski2015}. Actually, this scenario could also be realized for higher frequencies up to the near-IR regime, which is in turn suitable for optical trapping and manipulation of microscopic particles through optical tweezers \cite{Parkin2007,Nieto-Vesperinas2010,Rodrigues2017}. The lower-refractive index of water (or other liquids) in this electromagnetic regime comes only at the expense of requiring thicker waveguides. By way of example, we consider a planar metallic waveguide filled with water ($n=\sqrt{80}$ at 1 GHz): particularly, we focus on the asymmetric TE$_{2}$ and TM$_{2}$ guided modes in planar waveguides with $2d=3.4$ cm in Fig.~\ref{fig_GHz}. Assuming that the metallic boundaries behave as a perfect electric conductor, the dispersion relation reduces to (upon imposing that the transverse electric field at the interfaces vanishes): \begin{align} k_xd=(2m+1)\frac{\pi}{2} \;\; \mathrm{or} \;\; k_xd=m\pi \;\; (m=0,1,2\ldots). \end{align} We would like to stress the fact that, as pointed out above, qualitatively (and nearly quantitatively) similar results would be obtained for a metallic, water-filled waveguide with widths of the order of a micron operating in the IR ($\lambda\sim 1.55\mu$m). Along with energy and spin densities, radiation forces and torques felt by electric ($p$) and magnetic ($m$) dipole particles stemming from, respectively, longitudinal momenta and transverse SAM, are explicitly revealed through arrows and loops. First, the rich phenomenology inside with alternating layers/rings with opposite torques is evident in Fig.~\ref{fig_GHz}b,d. Moreover, recall that the SAM may stem from either the electric ($S_e$, Fig.~\ref{fig_GHz}d) or magnetic ($S_m$, Fig.~\ref{fig_GHz}b) contribution depending on the nature of the transverse mode (TM and TE, respectively). This implies that $p$ and $m$ dipoles will respond differently to such SAM, exhibiting non-negligible torques only if matching character \cite{Nieto-Vesperinas2010,Picardi2017}: Thus TE (respectively, TM) guided modes exert torque only to magnetic (respectively, electric) dipolar particles, see Fig.~\ref{fig_GHz}b (respectively, Fig.~\ref{fig_GHz}d). On the other hand, both electric/magnetic dipolar particles would suffer a radiation force along the guided mode propagation direction, being locally higher at layers with larger energy density (see Fig.~\ref{fig_GHz}a,c). Such radiation force $F\sim P^O$ can be anomalously large (respectively, small), as compared to the Poynting vector ($\mathbf{P}=\mathbf{P^O}+\mathbf{P^S})$, in alternating layers within the waveguide where the resulting spin momentum $P^S$ points opposite, $P^O=P+P^S$, (respectively, along, $P^O=P-P^S$), as observed in Fig.~\ref{fig_GHz}a,c. Lastly, we analyze the role of the optical forces on hybrid modes that carried intrinsic helicity. We focus on the HE$_{12}$ hybrid mode of a cylindrical metallic waveguide, again, filled by water. In cylindrical waveguides with perfectly conducting walls, guided modes dispersion relation reduces to the values of $k_tR$ for which the corresponding Bessel functions vanish \cite{Snyder1983}. In particular, the HE$_{12}$ hybrid mode exhibits a varied phenomenology. First, radiation forces are shown as expected for both probe particles in Fig.~\ref{fig_NW_GHz}a; as above, recall that radiation forces might be anomalously larger/shorter than those expected from the Poynting vector density for longitudinal spin momentum opposite/along the canonical momentum. Next, the transverse spin density is plotted in Fig.~\ref{fig_NW_GHz}b, revealing three rings with alternating spin sign inside with vanishing spin density at the center. Nonetheless, despite being hybrid, such mode becomes transverse electric so that the spin density contribution is purely magnetic ($S^{(e)}_{\theta}\equiv 0$), leading to radiation torque acting only on the magnetic dipole particle $m$. Finally, the helicity density and related spin density are shown in Fig.~\ref{fig_NW_GHz}c,d, exhibiting three rings alternating sign. It should be mentioned that such helicity-dependent spin density has both electric and magnetic contribution; however, the electric contribution is much larger $S_z^{(e)}\gg S_z^{(m)}$, and is thus indicated in Fig.~\ref{fig_NW_GHz}d with torque only exerted on the $p$ probe particle. Interestingly, this implies that an electric (respectively, magnetic) dipole particle would undergo in such a waveguide a longitudinal (respectively, transverse) torque. Incidentally, we have omitted in Fig.~\ref{fig_NW_GHz} the transverse force induced by the longitudinal spin momentum, which actually exerts no radiation force in the dipole approximation, but does produce a helicity-dependent transverse force in multipolar interactions with larger particles \cite{Bliokh2014a}. \begin{figure} \includegraphics[width=0.9\columnwidth]{SAM-Fig-NW-GHz.pdf} \caption{(a,c) Energy and helicity densities $W$ (left column) and (b,d) spin densities (right column) for the hybrid HE$_{12}$ guided mode for a cylindrical waveguide of radius $R$ and $\epsilon^{(r)}=80$ is considered for a normalized radius $R/\lambda=0.097$ (which corresponds to a water-filled cylinder with thickness $2R=5.8$ cm at $\nu=1$ GHz). Arrows in (a) indicate the radiation pressure force felt by electric ($p$) and magnetic ($m$) dipole particles due to longitudinal momenta, whereas loops in (b,d) reveal the corresponding torque induced inside the waveguides by spin densities: (b) transverse confinement-induced and (d) longitudinal helicity-induced. } \label{fig_NW_GHz} \end{figure} It should be emphasized that sub-wavelength particles with high-refractive index have been shown to yield strong magnetic dipole resonances \cite{Garcia-Etxarri2011}, as experimentally demonstrated in the optical and lower-frequency domains \cite{Evlyukhin2012,Geffrin2012}, leading to a wealth of phenomenology associated to the high-refractive-index dielectric resonant nanostructures \cite{Kuznetsov2016} and sub-wavelength structures in lower frequency regimes in general \cite{Geffrin2012}, where very large refractive indices are indeed ubiquitous. Recall also that, apart from its impact on induced torques, it has been recently demonstrated that such electric/magnetic spin contributions can be experimentally discerned through high-dielectric-index nanoparticles exhibiting both electric and magnetic dipole resonances \cite{Neugebauer2018}. Therefore, multi-resonant sub-wavelength particles would feel electric or magnetic torque depending on the resonant wavelength and guided mode involved (the latter influencing also the spatial dependence of such torque), which overall allows for a rich phenomenology. Finally, recall also a related spin-orbit locking is evidently expected if suitable electric or magnetic dipole sources are located inside, which could be exploited all along the electromagnetic spectrum. \section{Concluding remarks} To summarize, we have analytically investigated the spin and orbital angular momenta of light guided in planar and cylindrical waveguides supporting transverse electric/magnetic and hybrid (only in cylinders) modes. Leaving aside the well known transverse spin associated to the evanescent component of the guided modes outside the waveguides, we put the emphasis on the impact of mode confinement inside. We show that all (transverse an hybrid) modes, despite not being evanescent inside, exhibit a transverse SAM $S_t\propto (W/k_z\omega) k_t$ proportional to the transverse momentum $k_t$ for planar and cylindrical waveguides; the latter has been attributed the role of an effective mass for guided light (dispersion relation such that $\omega/c\sim\sqrt{k_z^2+k_t^2}\sim\sqrt{p^2+m^2}$). Such transverse spin density is shown to carry so called (Belinfante's) longitudinal spin angular momentum, parallel or anti-parallel to the proper orbital angular momentum, governed by the mode (spatial) transverse dependence imposed by its order and polarization. Moreover, it is demonstrated analytically that the spin momentum $P^S$ is crucial through this relationship $P=P^O+ P^S$ ($P$ being related to the Poynting vector momentum density) to retrieve the proper dependence of orbital momentum on propagation wavevector $P^O\sim k_z$. This confinement-induced spin is shown in various specific cases to be comparable or larger to that stemming from evanescent fields. Furthermore, it exhibits a much richer phenomenology, with either electric or magnetic spin rotating in different directions inside waveguides, and even with layers/rings of alternating rotation inside, depending basically on guided-mode spatial and polarization symmetry. Indeed, the extraordinary longitudinal spin momentum $P^S$ may also point along or opposite to the orbital momentum inside the waveguide. Finally, apart from the former transverse spin, hybrid modes in cylindrical waveguides are shown to carry longitudinal spin arising from its intrinsic helicity; such longitudinal SAM concentrates especially inside (strong confinement) the cylindrical waveguide. A discussion is also included on the impact of such transverse SAM inside waveguides on optical forces on dipolar particles inside e.g. water-filled waveguides in the IR to microwave regimes. All this phenomenology has been discussed for lowest-order guided modes in simple planar and cylindrical waveguides for the sake of clarity, but can be straightforwardly extrapolated to other geometries preserving guided mode confinement. Overall, the rich phenomenology demonstrated for such transverse SAM and longitudinal extraordinary momentum associated to guided light opens up a wealth of phenomenology for photon spin-orbit coupling inside waveguides beyond what is known for evanescent waves outside. Among the variety of configurations where such phenomenology could be exploited and tested, let us mention: emitters (quantum dots or wells) located inside semiconductor waveguides/fibers; colloidal quantum-dots in liquid channels; THz or microwave electric/magnetic dipole antennas located inside waveguides; or electric/magnetic dipolar particles for manipulating optical forces inside e.g. tubular, liquid-filled micro- or macro-fluidic cavities. \acknowledgments This work has been supported by the Spanish Ministerio de Ciencia, Innovaci\'on y Universidades (LENSBEAM FIS2015-69295-C3-2-P and FPU PhD Fellowship FPU15/03566). We are also grateful to M. Nieto-Vesperinas for helpful discussions at the early stages of this work.

d3758022b091667707526aa871e47683bb1b887f

\section{Introduction} Since the seminal work of Chandrasekhar, the axiomatization problem of astrophysics has been neglected. In \cite{TOV2} the authors reintroduced the problem, showing that arbitrary clusters of stellar systems experience fundamental constraints. In this paper we continue this work, focusing on a specific class of stellar systems: the TOV systems. In order to state precisely the question being considered and our main results, we recall some definitions presented in \cite{TOV2}. A \emph{stellar system} of degree $(k,l)$ is a pair $(p,\rho)$ of real functions, with $p$ piecewise $C^k$-differentiable and $\rho$ piecewise $C^l$-differentiable, both defined in some union of intervals $I\subset \mathbb{R}$, possibly unbounded. We usually work with systems that are endowed with an additional piecewise $C^m$-differentiable function $M:I\rightarrow\mathbb{R}$, called \emph{mass function}. In these cases we say that we have a system of degree $(k,l,m)$. The space of all these systems is then $$C_{pw}^{k}(I)\times C_{pw}^{l}(I)\times C_{pw}^{m}(I) \equiv C_{pw}^{k,l,m}(I), $$ where $C_{pw}^{r}(I)$ denotes the vector space of piecewise $C^r$-differentiable functions on $I$. A vector subspace $\mathrm{Stellar}^{k,l,m}_I$ of $C_{pw}^{k,l,m}(I)$ is called a \emph{cluster} of systems of degree $(k,l,m)$. Let $\mathrm{Stellar}^{l,m}_I$ denote the subspace obtained from the cluster by forgetting the variable $p$. We say that a system with mass function $(p,\rho,M)$ is a \emph{TOV system} if the Tolman--Oppenheimer--Volkoff (TOV) equation holds \begin{eqnarray} \label{tov1} p'(r) & =- &G \frac{\Bigl(\rho(r)+\frac{p(r)}{c}\Bigr)\Bigl(M(r)+4\pi r^{3}\frac{p(r)}{c^{2}}\Bigr)}{r^{2}\bigl(1-\frac{2M(r)G}{c^{2}r}\bigr)} \end{eqnarray} where $G$ and $c$ are respectively Newton's constant and the speed of light, which from now on we will normalize as $G=1=c$ (see \cite{weinberg}). In a general cluster $\mathrm{Stellar}^{k,l,m}_I$ we define a \emph{continuity equation} as an initial value problem $M'=F(M,\rho,r)$ in $\mathrm{Stellar}^{l,m}_I\times I$ with initial condition $M(a)=M_a$. Since the vector space of piecewise differentiable functions is locally convex but generally neither Banach nor Fr\'{e}chet \cite{TOV2,generalized_norm}, the problem of general existence of solutions for continuity equations is much more delicate \cite{ODE_1,ODE_2}. Therefore one generally works with continuity equations which are integrable. The classical example is \begin{equation}\label{continuity_equation} M'=4\pi r^2\rho. \end{equation} Notice that if a system of degree $(k,l,m)$ is endowed with (\ref{continuity_equation}), then $m=l+1$. We define a \emph{classical TOV system} as a TOV system equipped with the classical continuity equation (for some initial condition). Let $\mathrm{TOV}^{k,l}_I$ be the cluster of stellar systems of degree $(k,l,l+1)$ generated by (i.e, the linear span of) the classical TOV systems. We can now state the main problem: \begin{problem}[classification of TOV clusters] Given $I$, $k$ and $l$, determine the structure of $\mathrm{TOV}^{k,l} _I$ as a subspace of $C_{pw} ^{k,l,l+1}(I)$ for some fixed locally convex topology. \end{problem} Since $(0,0)\in \mathrm{TOV}^{k,l}_I$, this set is non-empty. Thus, the simplest thing one can ask about it is if it is nontrivial. This means asking if for every given $k,l$ there exists at least one nonzero pair $(p,\rho)$ for which equation (\ref{tov1}) is satisfied when assuming (\ref{continuity_equation}). Before answering this question, notice that by means of isolating $\rho$ in (\ref{continuity_equation}) and substituting the expression found in (\ref{tov1}) we see that the TOV equation is of Riccati type \cite{generating_TOV,generating_TOV_2}: \begin{equation} p'(r)=A(r)+B(r)p(r)+C(r)p(r)^{2} \label{riccati} \end{equation} where \begin{eqnarray} \label{a1} A(r) & = & -\frac{M(r)M'(r)}{4\pi r^{4}\bigl(1-2\frac{M(r)}{r}\bigr)},\\ \label{b1} B(r) & = & -\frac{\bigl(\frac{M'(r)}{r}+\frac{M(r)}{r^{2}}\bigr)}{1-2\frac{M(r)}{r}},\\ \label{c1} C(r) & = & -\frac{4\pi r}{1-2\frac{M(r)}{r}}. \end{eqnarray} Therefore, it is a nonlinear and non autonomous equation which, added to the fact that $C_{pw} ^{k,l+1} (I)$ is only a locally convex space (so that the general existence theorems of ordinary differential equations do not apply), makes it hard to believe that $\mathrm{TOV}^{k,l} _I$ has a nontrivial element. Even so, if $\rho$ is a constant function, then (\ref{riccati}) becomes integrable, showing that $\mathrm{TOV}^{k,l} _I$ is at least one-dimensional \cite{weinberg}. There are some results \cite{generating_TOV,generating_TOV_2} allowing that, under certain conditions, a solution of a TOV equation can be deformed into another solution, which may imply a higher dimension. \begin{remark} \emph{We point out that the generating theorem $(P1)$ proved in \cite{generating_TOV,generating_TOV_2} implies that the TOV cluster is invariant under} specific perturbations\emph{. So, a natural question is if there exist $I,k,l$ such that $\mathrm{TOV}^{k,l} _I$ is invariant under} arbitrary \emph{perturbations. This is clearly false, because this would imply that any stellar system is TOV. Instead, we can ask about invariance by arbitrary} small \emph{perturbations. Again, we assert that there are good reasons to believe that the answer is negative. Indeed, when we say that a subset of a topological space is ``invariant by small perturbations'' we are saying that it is actually an open subset. Thus, assuming invariance, we are saying that $\mathrm{TOV}^{k,l} _I$ is an open subset of $C_{pw} ^{k,l+1} (I)$ in our previously fixed locally convex topology. Now assume that the locally convex topology is Hausdorff and that $\mathrm{TOV}^{k,l} _I$ is finite-dimensional. Then the TOV cluster is also a closed subset \cite{top_vec_space_1,top_vec_space_2}. Since topological vector spaces are contractible \cite{top_vec_space_1,top_vec_space_2}, they are connected and therefore a subset which is both open and closed must be empty or the whole space. But we know that $\mathrm{TOV}^{k,l} _I$ is at least one dimensional and does not coincide with $C_{pw} ^{k,l+1} (I)$. In sum, if there exists some locally convex topology in which the TOV cluster is invariant by small perturbations, then the following things cannot hold simultaneously:} \begin{itemize} \item \emph{$\mathrm{TOV}^{k,l} _I$ is finite-dimensional;} \item \emph{the locally convex topology is Hausdorff.} \end{itemize} \emph{However, both conditions are largely expected to hold simultaneously, leading us to doubt the existence of a topology making $\mathrm{TOV}^{k,l} _I$ an open set}\footnote{Actually, this is not a special property of the TOV equation, but a general behavior of the solution space of elliptic differential equations \cite{elliptic_1}. The fact that TOV systems are modeled by an elliptic equation will be explored in a work in preparation.}. \end{remark} Another way to get information about the TOV cluster is not to look at the TOV cluster directly, but to analyze its behavior in some regime. For instance, if in (\ref{tov1}) we expand $1/c$ in a power series and discard all terms of higher order in $c$ (which formally corresponds to taking the limit $c\rightarrow \infty$) we get a new equation approximately describing the TOV equation. Therefore, studying the cluster $\mathrm{Newt}^{k,l}_I$ of stellar systems satisfying this new equation (called \emph{Newtonian systems}) we are getting some information about the original TOV cluster. In fact, if we consider the subspace $\mathrm{Poly}^{k,l}_I \subset \mathrm{Newt}^{k,l}_I$ of Newtonian systems satisfying an additional equation $p=\gamma^{q}\rho+a$, with $\gamma,c\in\mathbb{R}$ and $q\in\mathbb{Q}$, then the Newtonian equation becomes a Lane-Emden equation, which has at least three independent solutions (besides the constant ones), showing that $\mathrm{Newt}^{k,l}_I$ is at least four-dimensional \cite{weinberg,biblia_politropes}. In this article we analyze the structure of $\mathrm{TOV}^{k,l}_I$ in a limit other than the Newtonian one: we work in the \emph{pseudo-asymptotic limit}. Before saying what this limit is, let us first say what it is not. We could think of defining a ``genuine'' asymptotic limit of TOV by taking the limit $r\rightarrow \infty$ in the TOV equation (\ref{tov1}) in a similar way as done for getting the Newtonian limit, trying to obtain a new equation. In doing this we run into two obstacles: \begin{enumerate} \item differently of $c$ (which is a parameter), $r$ is a variable. Therefore, when taking the limit $r\rightarrow \infty$ we have to take into account the $r\rightarrow \infty$ behavior of all functions depending on $r$; \item in order to get a new equation we have to fix boundary conditions for the functions, loosing part of the generality. \end{enumerate} Thus, for us, \emph{pseudo-asymptotic limit is not the same as a genuine asymptotic limit}. Another approach would be to work with an additional differential equation in $C^{l+1}_{pw}(I)\times I$, called a \emph{coupling equation} and given by \begin{equation}{\label{coupling_equation} \Lambda(M,M',M^{(2)},...,M^{(s)},r)=0, } \end{equation} where $M^{(i)}$ denotes the $i$-th derivative of $M$ and $1\leq s \leq l+1$ is the order of the equation. The function $\Lambda$ itself is called the \emph{coupling function} that generates equation (\ref{coupling_equation}). Let us consider the space $\mathrm{Ind}_{\Lambda}^{l+1}(I)$ of all $C^{l+1}$-differentiable mass functions $M$ such that, if they are solutions of the coupling equation defined by $\Lambda$, then the corresponding TOV equation (\ref{riccati}) is integrable. So, for every coupling function $\Lambda$ we have $\mathrm{Ind}_{\Lambda}^{l+1}(I)\subset \mathrm{TOV}^{k,l} _I$ motivating us to define some kind of ``indirect'' asymptotic limit in TOV by taking the genuine asymptotic limit in (\ref{coupling_equation}). Again, we will have the two problems described above, but now the lack of generality is much less problematic, since we only need to consider boundary conditions for the single function $M$ and its derivatives. Even so, \emph{the pseudo-asymptotic limit is not the same as the indirect asymptotic limit}. Let us now explain what we mean by pseudo-asymptotic limit. A \emph{decomposition} for the coupling equation (\ref{coupling_equation}) generated by $\Lambda$ is given by two other coupling functions $\Lambda_0$ and $\Lambda_1$ such that $\Lambda=\Lambda_0 + \Lambda_1$. We say that a decomposition is \emph{nontrivial} if $\Lambda_i\neq0$ for $i=0,1$. A \emph{split decomposition} of $\Lambda$ is a nontrivial decomposition such that $\Lambda_0$ generates a linear equation and $\Lambda_1$ generates a nonlinear equation. We say that a split decomposition is \emph{maximal} when both $\Lambda_0$ and $\Lambda_1$ do not admit split decompositions. Not all coupling functions admit a maximal split decomposition, e.g, when $\Lambda$ is linear. When $\Lambda$ admits such a decomposition we will say that it is \emph{maximally split}. So, let $\Lambda$ be a maximally split coupling function. The \emph{pseudo-asymptotic limit} of (\ref{riccati}) relative to $\Lambda$ is obtained by taking the genuine asymptotic limit in the nonlinear part $\Lambda_1$, added of the boundary condition $\lim _{r\rightarrow \infty} \Lambda_1 = 0$, and maintaining unchanged the linear part $\Lambda_0$. In this case, the equation replacing (\ref{tov1}) is that generated by $\Lambda_0$. This new equation can be understood as a formal ``pseudo-limit'' of $\Lambda$, defined by $$\underset{r\rightarrow \infty}{\operatorname{psdlim}}\,\Lambda=\Lambda_0 + \lim_{r\rightarrow \infty}\Lambda_1.$$ For a given maximally split coupling function $\Lambda$, let $\mathrm{Psd}_{\Lambda}^{l+1}(I)$ denote the subspace of all mass-functions $M$ such that $$\lim_{r\rightarrow \infty} \Lambda_1 (M(r), M'(r),...,M^{(s)},r)=0,$$ i.e, which belong to the boundary conditions for $\Lambda$, and such that they are solutions of the pseudo-limit equation $\operatorname{psdlim}_{r \rightarrow \infty}\,\Lambda=0$. We can now state our main result. \begin{theorem}\label{thm_introduction} There exists $I$ such that for every $l$ there exists at least one maximally split coupling function such that $\mathrm{Psd}_{\Lambda}^{l+1}(I)$ is at least eleven-dimensional. \end{theorem} Generally, when we take a limit in a equation, the solutions of the newer equation are not solutions of the older one. This is why we cannot use the existence of Newtonian systems to directly infer the existence of new TOV systems. But the pseudo-asymptotic limit is different, precisely because it is not a formal limit. Indeed, suppose $M\in \mathrm{Psd}_{\Lambda}^{l+1}(I)$. Then by hypothesis it satisfies $\Lambda_0$. If in addition it satisfies $\Lambda_1$ (instead of obeying only the boundary condition), then it satisfies $\Lambda$ and, therefore, it belong to $\mathrm{TOV}^{k,l} _I$. We will show that by means of modifying $I$ in Theorem \ref{thm_introduction}, many of the mass functions will actually satisfy $\Lambda_1$, ensuring the existence of new integrable TOV systems. More precisely, we will show that if a pseudo-asymptotic system has a well-behaved extension to the extended real line $\overline{R}$, then it is actually a \emph{extended TOV system}, i.e, a TOV system which instead of being defined in a union $I$ of intervals of $\mathbb{R}$, its mass-function, density and pressure are all defined in $\overline{R}$. In sum, we have the following corollary: \begin{corollary}\label{corollary_introduction} The space $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ of piecewise $C^\infty$-differentiable extended TOV systems is at least eleven-dimensional. \end{corollary} The proof of Theorem \ref{thm_introduction} and its extrapolation to extended TOV systems will be done in Section \ref{proof} and in Section \ref{extending}, respectively, by making use of purely analytic arguments. In Section \ref{beyond} we argue that $\mathrm{Psd}_{\Lambda}^{\infty}(I)$ must have a higher dimension and we give two strategies that can be used to verify this. In Section \ref{sec_definitions} we present a physical classification for the densities associated with mass functions in Theorem \ref{thm_introduction} and, therefore, for the new extended TOV systems. We classify them according to if they possess or not cavities/singularities and to if they are composed of ordinary or exotic matter. In the process of classifying them we prove that they generally admit a topology change phenomenon (similar to a phase transition), allowing us to improve Theorem \ref{thm_introduction} and Corollary \ref{corollary_introduction}, showing that $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ is at least fifteen-dimensional. We finish the paper in Section \ref{concluding_remarks} with a summary of the results. \section{Proof of Theorem \ref{thm_introduction}} \label{proof_theorem} Before giving the proof of Theorem \ref{thm_introduction}, let us emphasize that finding integrable TOV systems is a nontrivial problem. Notice that the simplest way to find explicit solutions of the Riccati equation is by choosing coefficients satisfying one of the following conditions: \begin{enumerate} \item $C$ identically zero. In this case the equation becomes a linear homogeneous first order ODE, which is separable. \item $A$ identically zero. Then the equation is of Bernoulli type and, therefore, integrable by quadrature. \item There are constants $a, b, c \in \mathbb{R}$ such that $A(r) = a$, $B(r) = b$ and $C(r) = c$ simultaneously. In this case, the equation is separable. \end{enumerate} \noindent Thus, recalling that TOV systems are described by a Riccati equation, we could think of applying some of these conditions in (\ref{riccati}). It happens that the coefficients of this Riccati equation are not independent, but rather satisfy the conditions \begin{equation}\label{coefrel} \frac{4\pi r^4}{M(r)M'(r)} A(r) = \frac{r^2}{rM'(r) + M(r)} B(r) = \frac{1}{4\pi r} C(r). \end{equation} \noindent Therefore, the first two possibilities are ruled out for making (\ref{riccati}) trivial. Also, the third condition along with (\ref{c1}) implies that \begin{equation} M(r) = r/2 - 2\pi r^2/c \end{equation} \noindent and $M'(r)$ is the corresponding linear polynomial. However, in (\ref{coefrel}) the above $M(r)$ makes the $A$ term a quadratic polynomial, and the $B$ term a rational function of degree 1, so the equality can never hold. \subsection{The proof}\label{proof} Now we will prove Theorem \ref{thm_introduction}. \begin{proof} We start by noticing that in \cite{Tiberiu} it was shown that if the coefficients of any Riccati equation satisfy additional differential or integral conditions, then the nonlinearity of the starting equation can be eliminated, making it fully integrable. Each class of conditions is parametrized by functions $f:I \rightarrow \mathbb{R}$ and real constants. Because the authors of \cite{Tiberiu} worked only with smooth Riccati equations, they assumed $f:I \rightarrow \mathbb{R}$ smooth. However, it should be noticed that in the general situation we may have $f\in C^m_{pw}(I;\mathbb{R})$ or $f\in C^{m+1}_{pw}(I;\mathbb{R})$, where $m$ is the least order of differentiability of the coefficients of the Riccati equation. Furthermore, from (\ref{a1}), (\ref{b1}) and (\ref{c1}) we see that $m=l$, where $l$ is the class of $M$. In the following, we will use these additional equations in order to build maximally split coupling functions. Precisely, one of the integral conditions presented in \cite{Tiberiu} is \begin{equation} A(r) = \frac{f(r) - \left[ B(r) + C(r) \left[ \int \frac{f(s)-B^{2}(s)}{2C(s)}\, ds - c_1 \right] \right]^{2}}{4C(r)}, \label{case1} \end{equation} under which a explicit solution for Riccati equation (\ref{riccati}) is given by \begin{equation} p(r) = \frac{e^{\int^r (B(s)+C(s)h(s)) \, ds}}{c_0-\int^r C(s)e^{\int^s (B(\phi)+C(\phi)h(\phi)) \, d\phi} \, ds} + \frac{h(r)}{2} \label{0P} \end{equation} \noindent where $c_0$ is a constant of integration and \begin{equation} h(r) = \int^{r} \frac{f(s)-B^{2}(s)}{2C(s)}\, ds - c_1. \end{equation} Let us fix $f(r) = B^{2}(r)+2g(r)C(r)$, where $g:I \rightarrow \mathbb{R}$ is some integrable function. We will also take $c_1 = 0$. Let $h(r) = \int g(s)\, ds$. Using the coefficients of TOV equation (\ref{riccati}), (\ref{case1}) becomes \begin{equation} M'(r) + \frac{M(r)M'(r)}{2\pi r^{3}h(r)} - \left( \frac{2h'(r)}{h(r)} - \frac{1}{r} \right) M(r) + 2\pi r^{2}h(r) + \frac{rh'(r)}{h(r)} = 0, \label{nolinear} \end{equation} which is a coupling equation of order 1. Notice that it is maximally split, with maximal splitting decomposition given by \begin{align} \Lambda _0 (M,M',r)&= M'-\left( \frac{2h'(r)}{h(r)} - \frac{1}{r} \right) M + 2\pi r^{2}h(r) + \frac{rh'(r)}{h(r)} \label{decomposition_0}\\ \Lambda _1 (M,M',r)&= \frac{MM'}{2\pi r^{3}h(r)} \label{decomposition_1}. \end{align} Therefore, giving $M$ and $h$ such that $\displaystyle\lim_{r \to \infty} \Lambda_1 (M,M',r) = 0 $, then $M$ will automatically belong to $\mathrm{Psd}_{\Lambda}^{k}(I)$. We found ten such pairs. They are obtained by regarding $h$ as solutions of the differential equation \begin{equation} F(r) = 2\pi r^3h^2(r) + r^2h'(r), \label{int} \end{equation} for different $F:I\rightarrow \mathbb{R}$, as organized in Table \ref{table_appendix} of Appendix \ref{appendix}. Only to illustrate the method, we will show how the first row is obtained. The other rows are direct analogues, only involving more calculations. Recall that we are trying to find pairs $(M,h)$ such that $\operatorname{psdlimit}_{r\rightarrow \infty}\, \Lambda =0$, i.e, such that $\lim _{r\rightarrow \infty} \Lambda _1 =0$ and such that the coupling equation induced by $\Lambda _0$ is satisfied. Suppose such a pair was found. Then, by the linearity of $\Lambda _0$, they become related (up to addition of a constant) by \begin{equation} M(r) = \frac{h^2(r)}{r} \left(\int_1^r -\frac{s^2 \left(h'(s)+2 \pi s h^2(s)\right)}{h^3(s)} \, ds+c_2\right) \label{solutionlinear} \end{equation} where $c_2$ is a integration constant. Taking $F=0$ in (\ref{int}) we see that it is solved for \begin{equation} h(r) = \frac{1}{\pi (r^2 + c_1 / \pi)}, \label{0h} \end{equation} so that (\ref{solutionlinear}) becomes \begin{equation}\label{0M} M(r) = \frac{c_2h^2(r)}{r}, \end{equation} whose derivative is \[ M'(r) = \frac{c_2\left(2rh'(r)h(r)-h^2(r)\right)}{r^2}. \] Thus $\Lambda _1$ is given by the following expression, which clearly goes to zero as $r\to\infty$. \begin{equation*}\label{lambda_1_0} \Lambda_1(M,M',r) = \Lambda _1 (r)= \frac{c_2^2h^2(r)\left(2rh'(r)-h(r)\right)}{2\pi r^6} = -\frac{c_2^2 \left(5 \pi r^2-c_1\right)}{2 \pi r^6 \left(c_1-\pi r^2\right){}^4}. \end{equation*} Defining $(M,h)$ by (\ref{0M}) and (\ref{0h}), they will clearly satisfy the desired conditions, showing that $M\in \mathrm{Psd}_{\Lambda}^{l}(I)$. Now, notice that all involved functions are actually piecewise $C^{\infty}$, so that we can take $l$ arbitrary. On the other hand, the domains of the functions in Table \ref{table_appendix} are different, but we can restrict them to the intersection of the domains and then (since we are working with piecewise differentiable functions) extend all of them trivially to the starting $I$. This finishes the proof of Theorem \ref{thm_introduction}, except by the fact that Table \ref{table_appendix} contains \emph{ten} linearly independent mass functions instead of the \emph{eleven} ones stated in Theorem \ref{thm_introduction}. The one missing is just the well known constant density solution. For completeness, let us show that it can also be directly obtained from our method. Indeed, by taking $f(r) = B^2(r)$ in (\ref{case1}), writing $c= 8\pi ^2 c_1 /3 $, and using the coefficients of the TOV system, the coupling equation (\ref{nolinear}) becomes \begin{equation*} M'(r) + \frac{3M(r)M'(r)}{4\pi cr^{3}} + \frac{M(r)}{r} +\frac{4 \pi cr^{2}}{3} = 0, \end{equation*} which has solution $M(r) = 4\pi cr^3 /3 $, whose associated density is $\rho(r) = c$. \end{proof} \begin{remark} \emph{In the construction, the motivation of the definition of $h(r)$ in (\ref{0P}) is the control of the integral term. By control, we mean that the $h$ function is a integral of a integrable function $g$, freely chosen.} \end{remark} \begin{remark} A note on the differentiability of the pressure. \emph{From (\ref{0P}) we see that $p\in C^{k'}_{pw}(I)$, where $k'$ is the minimum between the differentiability class of $h$ and the class of $M$. The latter is $\infty$, as obtained above, so that $k'$ is the class of $h$. But looking at Table \ref{table_appendix} we see that the class of $h$ is $\infty$ in each of the cases. Thus, $p\in C^{\infty}_{pw}(I)$.} \end{remark} \subsection{Extending}\label{extending} In the last section we proved that there exists a subset $I\subset \mathbb{R}$, which can be regarded as a disjoint union of intervals, such that we have a maximally split coupling function $\Lambda$ of order 1, whose corresponding space of pseudo-asymptotic TOV systems $\mathrm{Psd}_{\Lambda}^{\infty}(I)$ is at least eleven-dimensional. As discussed in the Introduction, a pseudo-asymptotic TOV system does not need to be a TOV system. Here we will show that by means of modifying $I$ properly, i.e, by working on the extended real line, we can assume that some of the pseudo-asymptotic systems that we have obtained really define TOV systems. Recall that if there exists a coupling function $\Lambda$ such that a pseudo-asymptotic mass function $M$ satisfies not only the linear part $\Lambda _0$ but also the nonlinear one $\Lambda _1$, then $M$ actually defines an integrable TOV system. So, our problem is to analyze when the pseudo-asymptotic $M$ obtained in the last section satisfies the differential equation $\Lambda_1 (M,M',r)=0$ for $\Lambda _1$ given by (\ref{decomposition_1}). We will give sufficient conditions on the general pseudo-asymptotic mass functions in order for this to happen. That these conditions are satisfied for our mass functions will be a consequence of their classification. The fundamental step is the following result from real analysis. \begin{lemma}\label{lemma_extension} Let $f: I \rightarrow \mathbb{R}$ be continuous at a point $a_0\in I$ and such that $f(x)\rightarrow 0$ when $x\rightarrow a_0$. Assume that there exists $\varepsilon>0$ such that one of the following conditions is satisfied: \begin{enumerate} \item[\emph{c1)}] $f$ is non-negative and non-decreasing in $(a_0-\varepsilon,a_0]$; \item[\emph{c2)}] $f$ is non-positive and non-increasing in $[a_0,a_0+\varepsilon)$. \end{enumerate} Then there exists $0< \varepsilon' \leq \varepsilon$ such that $f$ is constant and equal to zero in $(a_0-\varepsilon',a_0]$ in the first case, and in $[a_0,a_0+\varepsilon')$ in the second case. \end{lemma} \begin{proof} Because $f$ is continuous in $a_0$ and $f(x)\rightarrow 0$ when $x\rightarrow a_0$, we have $f(a_0)=0$. If $f$ satisfies the first condition, since $f$ is non-decreasing, it follows that if $x \leq y$ in $(a_0-\varepsilon,a_0]$, then $f(x)\leq f(y)$. This means that for every $x$ we have $f(x)\leq f(a_0)=0$. But $f$ is non-negative in $(a_0-\varepsilon,a_0]$, hence we must have $f(x)=0$ in this interval. For the second condition, an analogous argument will give the result. \end{proof} Now, recall that we can extend the real line $\mathbb{R}$ in two different ways: by adding a point at infinity $\infty$ or by adding both $+\infty$ and $-\infty$. In the first case we have the \emph{projectively extended real line} $\hat{\mathbb{R}}$, while in the second one we have the \emph{extended real line} $\overline{\mathbb{R}}$. For the arithmetic construction of these objects, see \cite{extended_line_1}. Topologically, both spaces acquire natural Hausdorff compact topologies: $\hat{\mathbb{R}}$ is the one-point compactification of $\mathbb{R}$ and, therefore, is homeomorphic to the circle $\mathbb{S}^1$, while $\overline{\mathbb{R}}$ is the two-point compactification of $\mathbb{R}$ and has an order topology homeomorphic to $[-1,1]$ \cite{extended_line_2}. In this article we will use only $\overline{\mathbb{R}}$. One can think of the homeomorphism $\phi: \overline{\mathbb{R}}\rightarrow [-1,1]$ as a rescaling of $\overline{\mathbb{R}}$ that brings the infinities closer together. This is similar to the idea of compactification of spacetime used in the Penrose diagrams, with the difference that there the underlining topological space does not become actually compact. Notice that any piecewise continuous function $f:I\rightarrow \mathbb{R}$ which is not oscillating in $\pm \infty$ admits an extension $\overline{f}$ to $\overline{\mathbb{R}}$, as follows. We first extend it to $\mathbb{R}$ by defining $\overline{f}(x)=0$ when $x\notin I$ and then take $\overline{f}(\pm \infty)=\lim_{x\rightarrow \pm \infty}f(x)$. By definition, this extension is continuous at $\pm \infty$ \footnote{This clearly does not mean that $\overline{f}$ is continuous in the whole domain $\overline{\mathbb{R}}$, only piecewise continuous.}. Now, recall that the domain of any function is in one-to-one correspondence with its graph. Furthermore, if the function is piecewise continuous, then this correspondence is actually a piecewise homeomorphism. Thus, if $f:I\rightarrow \mathbb{R}$ is any function as above, we have the commutative diagram below, where $\overline{\overline{f}}=\phi \circ \overline{f} \circ \phi ^{-1}$ is the rescaling of the extension $\overline{f}$. \begin{equation}{\label{diagram} \xymatrix{\mathbb{R} \ \ar@{^(->}[r] & \overline{\mathbb{R}} \ar[r]_{\simeq}^{\phi} & [-1,1] \\ I \ \ar[u]^f \ar@{^(->}[r] & \overline{\mathbb{R}} \ar@{-->}[u]^{\overline{f}} \ar[r]_{\simeq}^{\phi} & [-1,1] \ar@{-->}[u]_{\overline{\overline{f}}} \\ \operatorname{graph}(f) \ar[u]^{\simeq} \ar@{-->}[r] & \operatorname{graph}(\overline{f}) \ar[u]^{\simeq} \ar@{..>}[r]_{\simeq}^{\varphi} & \operatorname{graph}(\overline{\overline{f}}) \ar[u]_{\simeq}}} \end{equation} From the above remarks and from Lemma \ref{lemma_extension} we get the following corollary: \begin{corollary}\label{corollary_extending} Let $\Lambda$ be a maximally split coupling function, $M$ a pseudo-asymptotic mass function for the coupling function $\Lambda$ and suppose that $\Lambda _1$ is not oscillating in $\pm \infty$. If there exists $0< \varepsilon \leq 2$ such that $\overline{\overline{\Lambda_1}}$ satisfies condition \emph{(c1)} (resp. \emph{(c2)}) of Lemma \ref{lemma_extension}, then there exists $\varepsilon '$ such that $\overline{\overline{\Lambda_1}}$ is zero in $(1-\varepsilon,1]$ (resp. in $[-1,-1+\varepsilon ))$. \end{corollary} \begin{proof} By hypothesis $\Lambda _1$ is not oscillating in $\pm \infty$, so that the extension $\overline{\Lambda _1}$ exists and it is continuous at $\pm \infty$. The points $\pm \infty$ are mapped onto $\pm 1$ by $\phi$, so that $\overline{\overline{\Lambda_1}}$ is continuous at $\pm 1$. The result then follows from Lemma \ref{lemma_extension}. \end{proof} Notice that starting with the TOV equation we can extend it to $\overline{\mathbb{R}}$ and then rescale the infinities by working at $[-1,1]$ via $\phi$. All we have done in the previous section will work in the same way. In particular, when finding situations in which $\overline{\overline{\Lambda_1}}=0$ we are finding cases in which the pseudo-asymptotic solutions of the \textbf{extended} TOV equation is a genuine solution of that \textbf{extended} equation. Let $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ denote the space of such solutions which are piecewise smooth, i.e, the space of piecewise $C^\infty$-differentiable extended TOV systems. \begin{proposition}\label{corollary_extending_2} $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ is at least eleven-dimensional. \end{proposition} \begin{proof} Let us consider the pseudo-asymptotic mass functions $M$ obtained in Section \ref{proof}, whose underlying density functions are in Table \ref{table_appendix}, so that $\Lambda_1$ is given by (\ref{decomposition_1}), which depends on a function $h$, also listed in Table \ref{table_appendix}. Writing (\ref{decomposition_1}) explicitly for each $(M,h)$, as presented in Appendix \ref{appendix_2}, we see that $\Lambda_1$ is not oscillating in $\pm \infty$, so that $\overline{\Lambda_1}$ and $\overline{\overline{\Lambda_1}}$ are well-defined. From the commutativity of diagram (\ref{diagram}) we can analyze the graph of $\overline{\overline{\Lambda_1}}$ looking at the asymptotic behavior of the graph of $\Lambda_1$. As we see in Appendix \ref{appendix_2}, each $\Lambda_1$ either becomes non-negative and non-decreasing when $r\rightarrow +\infty$ or non-positive and non-increasing when $r\rightarrow -\infty$, which means that the corresponding $\overline{\overline{\Lambda_1}}$ have the same behavior in a neighborhood sufficiently small of $\pm 1$. The result follows from Corollary \ref{corollary_extending}. \end{proof} \begin{remark}[important remark] \emph{Let $\mathrm{TOV}^{\infty,\infty}_{I,0}\subset \mathrm{TOV}^{\infty,\infty}_I$ be the space of piecewise smooth TOV systems which are not oscillating in $\pm \infty$. We have an inclusion } $$\overline{\overline{(\cdot)}}:\mathrm{TOV}^{\infty,\infty}_{I,0} \hookrightarrow \mathrm{\overline{\overline{TOV}}}^{\infty,\infty},\quad \mathrm{given \;by} \quad \overline{\overline{(p,M}})=(\overline{\overline{p}},\overline{\overline{M}}).$$ \emph{In Proposition \ref{corollary_extending_2} we got ten extended solutions. The remaining one is, again, the classic constant solution, now regarded as an extended solution via the inclusion above. The important fact to have in mind is that the reciprocal does not hold: an extended TOV solution when restricted to some interval $I$ of $\mathbb{R}$ is not necessarily a solution of the actual TOV equation in $I$. Indeed, if an extended TOV system $(p,M)$ depends explicitly on its behavior at $\pm$, then when restricting to $I$ the equation will not be preserved, so that $(p\vert _I,M\vert _I)$ will not belong to $\mathrm{TOV}^{\infty,\infty}_I$. Notice that this is exactly the situation of Corollary \ref{corollary_extending}, so that we cannot use Proposition \ref{corollary_extending_2} in order to get solutions of the actual TOV equation.} \end{remark} \subsection{Beyond}\label{beyond} We proved Theorem \ref{thm_introduction} ensuring the existence of a coupling function $\Lambda$ whose space $\mathrm{Psd}_{\Lambda}^{\infty}(I)$ of pseudo-asymptotic solutions is \emph{at least} eleven-dimensional. In this section we show that it is at least fifteen-dimensional and discuss why it is natural to believe that it has an even higher dimension. The first assertion is due to the following reason: \begin{enumerate} \item \emph{Existence of critical configurations exhibiting phase transitions}. Notice that the integrability conditions of \cite{Tiberiu}, such as (\ref{case1}), (\ref{case6}) and (\ref{case8}), depends on two constants $c_1$ and $c_2$. Consequently, the pseudo-asymptotic solutions of the corresponding coupling equations also depend on such constants. Let us write $M_{1,2}$ to emphasize that $M$ is a pseudo-asymptotic mass function depending on $c_1$ and $c_2$. Now, recall that two mass functions $M,N$ are linearly dependent in $\mathrm{Psd}_{\Lambda}^{l}(I)$ if there exists a real number $c$ such that $M(r)=cN(r)$ for every $r\in I$. This means that if the dependence of $M_{1,2}$ on $c_1$ (resp. $c_2$) is \textbf{not} in the form $M_{1,2}=c_1 M_{2}$ (resp. $M_{1,2}=c_2 M_{1}$), then when varying $c_1$ (resp. $c_2$) we get at least two linearly independent pseudo-asymptotic mass functions. These linearly independent mass functions can be obtained by defining a new piecewise differentiable map $\mathcal{M}:I\times \mathbb{R}\times \mathbb{R} \rightarrow \mathbb{R}$, such that $\mathcal{M}(r,c_1,c_2)=M_{1,2}(r)$, and then by studying its critical points. In typical cases (as for those obtained in Section \ref{proof}, i.e, for those whose density function is in Table \ref{table_appendix}) the function $\mathcal{M}$ is piecewise a submersion. Therefore, the pre-images $\mathcal{M}^{-1}(c)$ are submanifolds $S_c$ of $\mathbb{R}^3$ and the linearly independent mass functions can be obtained by searching for topology changes in $S_c$ when $c$ vary (in a similar way we search for phase transitions in a statistical mechanics system). In the next section we will analyze the topology of the surface $S_c$ corresponding to rows 1, 2 and 7 of Table (\ref{table_appendix}), showing that they admit one, two and one topology changes, respectively. This means that Theorem \ref{thm_introduction} and Corollary \ref{corollary_extending_2} can be improved, giving the theorem below. \begin{theorem} There exists a maximally split coupling function $\Lambda$ such that $\mathrm{Psd}_{\Lambda}^{\infty}(I)$ is at least fifteen-dimensional. Furthermore, $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ is at least fifteen-dimensional. \end{theorem} The second assertion is suggested by the following reason: \item \emph{Existence of other maximally split coupling functions}. Recall that our starting point to get $\Lambda$ in (\ref{nolinear}) was the integral equation (\ref{case1}) obtained in \cite{Tiberiu} which when satisfied induces a solution for the TOV equation. In \cite{Tiberiu}, besides (\ref{case1}), other nine integral/differential equations playing the same role are presented. Applying to these other nine equations a strategy analogous to that used in (\ref{case1}) to obtain (\ref{nolinear}), allows us to obtain new coupling equations. We recall that each integrability equation in \cite{Tiberiu} becomes parametrized by certain constants and by an arbitrary function $f$. By making a suitable choice of $f$ in the sixth and eighth cases of \cite{Tiberiu}, we see that the induced coupling equations coincide with (\ref{nolinear}). Explicitly, the sixth case is given by \begin{equation}\label{case6} B(r) = \frac{f_3(r) -A(r)C(r) - C^2(r)\left[\int \frac{f_3(r)}{2C(r)}\, dr - c_7\right]^{2}}{2C(r)\left[\int \frac{f_3(r)}{2C(r)}\, dr - c_7\right]} \end{equation} where $f_3\in C^{k+1}_{pw}(I)$ and $c_7$ is a constant of integration. For the choice $f_3(s) = 2C(s)g(s)$, where $g \in C^{k+1}_{pw}(I)$, the induced coupling equation reduces to (\ref{nolinear}). Furthermore, the eighth case is \begin{equation}\label{case8} B(r) = \frac{1}{f_4(r)}\left[C(r)\frac{d}{dr}\left(\frac{f_4(r)}{C(r)}\right)-\frac{f_{4}^{2}(r)}{2}-2A(r)C(r)\right] \end{equation} where $f_4\in C^{k+1}_{pw}(I)$. If we choice $f_4(r) = h(r)C(r)$ again we get (\ref{nolinear}). Therefore, in view of the methods developed in Section (\ref{proof}), equations (\ref{case6}) and (\ref{case8}) do not differ from (\ref{case1}). On the other hand, we could not find $f$ which makes the coupling equation induced by each of the other seven cases equals to (\ref{nolinear}). This does not means that they will produce new pseudo-asymptotic mass functions which will eventually define (via Corollary \ref{corollary_extending}) new extended TOV systems\footnote{Indeed, many of the induced maximally split coupling equation have a nonlinear part which is really high nonlinear.}. Even so, it suggests the possibility. \end{enumerate} \section{Classification}\label{classification} \label{sec_definitions} So far we have focused on getting new integrable extended TOV systems. In this section we will work to give physical meaning to discovered systems. In order to do this, we propose a simple classification of general stellar systems in which we will consider the new extended TOV systems. Indeed, let $(p,\rho)\in C^k_{pw}(I)\times C^l_{pw}(I)$ be a stellar system of degree $(k,l)$. We say that it is \begin{itemize} \item \emph{ordinary (resp. exotic)} in an open interval $J \subset I$ if the density $\rho$ is positive (resp. negative) in each point of $J$. That is, $\rho(r)>0$ (resp. $\rho(r)<0$) for $r$ in $J$; \item \emph{without cavities} if $\rho$ has no zeros, i.e, $\rho(r)\neq 0$ for every $r$ in $I$; \item \emph{without singularities} if its domain is an open interval $(0,R)$, where $R$ can be $\infty$; \item \emph{smooth} if it is without singularities and $\rho \in C^{\infty}(I)$; \item \emph{realistic} if it is smooth and ordinary in $I$. \end{itemize} A \emph{cavity radius} of $(p,\rho)$ is a zero of $\rho$. Similarly, a \emph{singularity} of $(p,\rho)$ is a discontinuity point of $\rho$. So, $(p,\rho)$ is without cavities (resp. without singularities) iff it has no cavity radius (resp. singularity). When looking at Table \ref{table_appendix} it is difficult to believe that some of the stellar systems there described are realistic. In fact, as can be rapidly checked, when defined in their maximal domain, these systems are in fact unrealistic. But as we will see, when restricted to a small region, many of them becomes realistic. This becomes more clear looking at Figure \ref{drawfigures} below, which describe the classification of certain rows of Table \ref{table_appendix}. In the schematic drawings, the filling by the grid is associated to exotic matter, whereas the filling by the hexagons to ordinary matter. The dashed circles represents a singularity radius and the dot circles a cavity radius. \begin{figure}[h] \centering \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw1.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph1}} \label{draw1} \end{subfigure} \quad\quad \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw2.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph2}} \label{draw2} \end{subfigure} \\ \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw3.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph6}} \label{draw3} \end{subfigure} \quad\quad \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw4.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph7}} \label{draw4} \end{subfigure} \\ \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw5.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph8}} \label{draw5} \end{subfigure} \quad\quad \begin{subfigure}[h]{0.25\textwidth} \centering \includegraphics[scale=0.25]{draw6.pdf} \caption{\centering Schematic Drawing of Figure \ref{graph9}} \label{draw6} \end{subfigure} \caption{Classification of some new pseudo-asymptotic TOV systems.} \label{drawfigures} \end{figure} Notice that in order to do this classification we need to search for the zeros of the density function, which will give the radii in which there is no matter inside the star. Suppose that we found one of them, say $r_o$. If $\rho$ is continuous in that radius, then it is a cavity; otherwise, it is a singularity. The fundamental difference between them is that continuity implies that $\rho$ cannot change sign in neighborhoods of $r_o$. This means that a star containing only cavities is composed either of ordinary matter or of exotic matter. On the other hand, stellar systems with singularities may contain both ordinary and exotic matter. The stellar systems considered in Table \ref{table_appendix} have densities of the form $\rho (r)=\frac{p(r)}{q(r)}\log(o(r))$, where $p$, $q$ and $o$ are polynomials with integer or fractional powers. Singularities are identified with zeros of $q$, while cavities are given by zeros of $p$ and $o-1$. The existence of real roots for a given polynomial is strongly determined by its coefficients and, in the present situation, the coefficients of $p$, $q$ and $o$ depend on two real parameters $c_1$ and $c_2$. We write $p_{12}$, $q_{12}$ and $o_{12}$ in order to emphasize this fact. We can then search for \emph{critical configurations}, in which a small change of $c_1$ and $c_2$ produces a complete modification of the system, as in a phase transition in statistical mechanics. The critical configurations can be captured by defining new functions $$\mathcal{P},\mathcal{Q},\mathcal{O}:[0,\infty)\times \mathbb{R}\times \mathbb{R} \rightarrow \mathbb{R}$$ by $\mathcal{P}(r,c_1,c_2)=p_{12}(r)$, and so on, which are piecewise submersions. The solution of $\mathcal{Q}$ is then an algebraic submanifold $S_p \subset \mathbb{R}^3$, possibly with boundary, that completely determines the behavior of the singular set of $\rho _{12}$ when we vary $c_1$ and $c_2$. Similarly, the disjoint union of the solutions sets of $\mathcal{P}$ and $\mathcal{O}-1$ also defines an algebraic submanifold $S_{po}$ of $\mathbb{R}^3$ which determines the behavior of the cavities when we vary $c_1$ and $c_2$. Notice that a point $(r,c_1,c_2)\in S_q$ is a critical singularity of $\rho_{12}$ iff it admits topologically distinct neighborhoods. Analogously, the critical cavities are given by points in $S_{po}$ with non-homeomorphic neighborhoods. Finally, the critical configurations of the stellar system with density $\rho_{12}$ are the points of $S_q \sqcup S_{po}$ which are critical singularities or critical cavities. A manifold defined by the inverse image of a function is locally homeomorphic to the graph of the defining function. This means that a neighborhood for $(r,c_1,c_2)$ is just a piece of the graph of $\mathcal{P}$, $\mathcal{Q}$ or $\mathcal{O}-1$. Since the topology of a graph changes only at an asymptote, zeros of $Q$ with fixed $(c_1,c_2)$ give the singularities, while zeros with fixed $(r,c_2)$ and $(r,c_1)$ will give the critical singularities, and similarly for critical cavities. Having obtained singularities, cavities and critical configurations, the classification is completed by determining the kind of matter, which can be done from graphical analysis. In next sections we will apply this strategy to rows 1, 2 and 7 of Table \ref{table_appendix}. The row 1 will produce the schematic drawings (\ref{draw1}) and (\ref{draw2}), while row 2 will produce (\ref{draw3}) and (\ref{draw4}), and row 7 will give (\ref{draw5}) and (\ref{draw6}). \subsection{Row 1 of Table \ref{table_appendix}} In this case the density function in given by \begin{equation}\label{rho_1_2} \rho(r,c_1,c_2)=\frac{c_2 \left(c_1-5 \pi r^2\right)}{4 \pi r^4 \left(\pi r^2-c_1\right){}^3}. \end{equation} Notice that $c_2$ is a multiplicative constant, so that it will give linearly dependent solutions, leading us to fix $c_2=1$. This means that the singular set $S_q$ is a submanifold of $\mathbb{R}^2$, while the set of cavities $S_{po}$ is a submanifold of $\mathbb{R}^3$ of the form $S\times \mathbb{R}$, with $S\subset \mathbb{R}^2$. More explicitly, singularities are given by the solutions of the algebraic equation \begin{equation} 4\pi r^4(\pi r^2 - c_1)^3 = 0, \label{singularities_1} \end{equation} while cavities are determined by the solutions of \begin{equation} c_1 - 5\pi r^2 = 0 \label{cavities_1} \end{equation} Both equations depend explicitly of $c_1$. If $c_1 \leq 0$, the only singularity radius is the origin $r = 0$ and there are no cavities. An example of this behavior is given by Figure \ref{graph1}, for $c_1=0$: \begin{figure}[h] \centering \includegraphics[scale=0.5]{graph1.jpg} \caption{Plot of (\ref{rho_1_2}) for $c_1 = 0$ and $c_2 = 1$.} \label{graph1} \end{figure} It then follows that, for our choice $c_2=1$, the stellar system is composed only by exotic matter. On the other hand, if $c_1 > 0$, the singularities happen at $r = 0$ and at $r_0 =\sqrt{c_1/\pi}$, which is the non-negative solution of $\pi r^2 - c_1 = 0$. The single cavity is given by the single positive root $r_1 = \sqrt{c_1/5\pi}$ of (\ref{cavities_1}), which lies in the interval $(0,r_0)$ bounded by the singularities. Thus, the type of matter inside $(0,r_0)$ may change, but it remains the same after crossing $r_0$. In order to capture this change of matter we analyze the sign of the derivative of $\rho _{12}$ at $r_1$. The derivative is \begin{equation} \rho _{12}'(r) = \frac{c_2 \left(c_1^2-5 c_1\pi r^2 + 10\pi^2 r^4\right)}{\pi r^5 \left(\pi r^2-c_1\right)^{4}} \end{equation} and we see that $\rho _{12}'(r_1) > 0$. So, in that point, the star matter stops being exotic and becomes ordinary. Moreover, we have $\lim_{r\to r_{0+}} \rho _{12}(r) = +\infty$ and $\lim_{r\to r_{0-}} \rho _{12}(r) = -\infty$. It follows that in $(r_0,\infty)$ the star is composed of exotic matter. An example of this behavior is given by Figure \ref{graph2}, where $c_1=7$: \begin{figure}[h] \centering \includegraphics[scale=0.5]{graph2.jpg} \caption{Plot of \ref{rho_1_2} for $c_1 = 7$ and $c_2 = 1$.} \label{graph2} \end{figure} Now, solving (\ref{cavities_1}) and (\ref{singularities_1}) we see that for each given $r_0$ there exists a single $c_1$ making $r_0$ a critical configuration. \subsection{Row 2 of Table \ref{table_appendix}} We start by noticing that in the present case the density function can be written as \begin{equation} \rho_{12}(r) = -\frac{\Gamma_{12}(r)}{8 \pi r^4 \left(r^2-1\right) \left(-c_1+\pi r^2+\pi \log \left(r^2-1\right)\right){}^3} \label{h'rho} \end{equation} where \begin{align} \Gamma_{12}(r) = &-2 c_1 c_2 \left(r^2-1\right)+3 \pi ^2 \left(r^2-1\right) \left(\pi r^2-c_1\right) \notag \\ &\times \left[\log \left(r^2-1\right)\right]^2 - 2\pi [c_2 \left(1-5 r^2\right) r^2 \notag \\ &+ c_1^2 \left(r^4+2 r^2-1\right)] + \pi \left(r^2-1\right) \notag \\ &\times \left[2 \pi c_1 \left(1-3 r^2\right)+2 \left(c_1^2+c_2\right) +\pi ^2 \left(3 r^4-1\right)\right] \notag \\ &\times \log \left(r^2-1\right)-\pi ^2 c_1 \left(3 r^6-13 r^4+r^2+1\right) \notag \\ &+ \pi ^3 \left(r^2-1\right) \left[\log \left(r^2-1\right)\right]^3 \notag \\ &+\pi ^3 \left(r^8-r^6-5 r^4+r^2\right). \label{h'gamma} \end{align} The function $\rho_{12}$ depends non-trivially on both variables $c_1$ and $c_2$ only in $\Gamma_{12}$, so that a priori its singular set and its set of zeros are arbitrary submanifolds of $\mathbb{R}^3$ and $\mathbb{R}^2$, respectively, and therefore we may expect many critical configurations. The interval $[0,1]$ is clearly singular, while the singularities for $r>1$ are determined by solutions of the equation \[ -c_1+\pi r^2+\pi \log \left(r^2-1\right) = 0. \] We assert that for each fixed $c_1$ there is precisely one solution. In other words, we assert that the singular set is diffeomorphic to $([0,1]\times \mathbb{R})\sqcup \mathbb{R}$. Indeed, let $Y:[0,\infty) \times \mathbb{R} \rightarrow \mathbb{R}$ be the piecewise differentiable function \begin{equation} Y(r,c_1) = -c_1+\pi r^2+\pi \log \left(r^2-1\right). \end{equation} Clearly, there are $r_{0}$ and $r_{1}$ such that $Y(r_{0},c_1) < 0$ and $Y(r_{1},c_1) > 0$. But, the derivative of $Y(r,c_1)$ in the direction of $r$ is always positive for $r>1$. If fact, if $r>1$, then \[ \frac{\partial Y}{\partial r}(r,c_1) = \frac{2r(r^2+\pi -1)}{r^2-1} > 0. \] Consequently, $Y$ has a single zero in $r>1$ corresponding to the unique singularity of $\rho _{12}$ in this interval. Moreover, for any $r_0 > 1$, we can set the constant $c_1$ appropriately so that $r_0$ corresponds the singularity. In fact, just set \[ c_1 = \pi r_0^2+\pi \log \left(r_0^2-1\right). \] The cavities of $\rho_{12}$ in (\ref{h'rho}) are given by zeros of $\Gamma_{12}$. Differently of the previous case, we cannot isolate any of the variables, so that $S_{po}\subset \mathbb{R}^3$ will have complicated topology. We assert that the manifold of critical configurations is $\mathbb{R}^2$. Indeed, if we fix $c_1$ and vary $c_2$ we do not find any change of topology, while if we fix $c_2$, say $c_2=1$, we find two critical points, approximately at $c_1 = 1,7$ and at $c_1=4,6$, as shown in Figure \ref{h'figures}. \begin{figure}[h] \centering \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.3in]{graph3.jpg} \caption{$c_1 = 1$} \label{graph3} \end{subfigure} \quad \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.3in]{graph4.jpg} \caption{$c_1 = 1.6$} \label{graph4} \end{subfigure} \\ \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.3in]{graph5.jpg} \caption{$c_1 = 1.8$} \label{graph5} \end{subfigure} \quad \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.3in]{graph6.jpg} \caption{$c_1 = 4.5$} \label{graph6} \end{subfigure} \quad \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.3in]{graph7.jpg} \caption{$c_1 = 4.7$} \label{graph7} \end{subfigure} \caption{Plot of (\ref{h'rho}) for different values of $c_1$ and constant $c_2 = 1$.} \label{h'figures} \end{figure} \subsection{Row 7 of Table \ref{table_appendix}} Previously we studied a stellar system with discrete critical configurations and another with a continuum of critical configurations. In both cases, graphical analysis were used to identify the critical configurations. Here we will discuss a third system, whose special feature is to show that even when we have an analytic expression for cavities and singularities, a graphical analysis is fundamental to complete the classification. The density function is now of the form \begin{equation} \rho(r,c_1,c_2) = f(r,c_1,c_2)/q(r,c_1), \label{h2rho2} \end{equation} where $f$ is a linear combination of polynomials and logarithmic functions, depending on both parameters $c_1$ and $c_2$. We will focus on singularities, so that only the expression of $q$ matters. It is the polynomial \[ q(r,c_1) = 12\pi r^2(-c_1r + \pi r^3 +1)^3. \] Because $c_2$ does not affect $q$ we see that the singular set is a submanifold of $\mathbb{R}^2$. For each fixed $c_1$ the complex roots of $q$ can be written analytically with help of some software (we used Mathematica$\circledR$). They are $r_0=0$, with multiplicity two, and \begin{equation} r_1 = \frac{\sqrt[3]{2} \left(\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }\right){}^{2/3}+2 \sqrt[3]{3} c_1}{6^{2/3} \sqrt{\pi } \sqrt[3]{\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }}} \label{root1} \end{equation} \begin{equation} r_2 = \frac{\sqrt[3]{2} \sqrt[6]{3} \left(-1+i \sqrt{3}\right) \left(\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }\right){}^{2/3}-2 \left(\sqrt{3}+3 i\right) c_1}{2\ 2^{2/3} 3^{5/6} \sqrt{\pi } \sqrt[3]{\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }}} \label{root2} \end{equation} \begin{equation} r_3 = \frac{\sqrt[3]{2} \sqrt[6]{3} \left(-1-i \sqrt{3}\right) \left(\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }\right){}^{2/3}-2 \left(\sqrt{3}-3 i\right)c_1}{2\ 2^{2/3} 3^{5/6} \sqrt{\pi } \sqrt[3]{\sqrt{81 \pi -12 c_1^3}-9 \sqrt{\pi }}}, \label{root3} \end{equation} each of them with multiplicity three. Our stellar system then have singularities only for the values $c_1$ such that some of the radii above are real and non-negative. One can be misled to think that $r_1$ corresponds to a real zero iff $c_1 \leq \sqrt[3]{\frac{81\pi}{12}} \approx 2.77$. Although $c_1 \approx 2.77$ is clearly a critical configuration, a graphical and numerical analysis shows that $r_1$ is always real. In order to see this, we compare in Figure \ref{h2figures} the density function for $c_1 = 5 > 2.77$ and $c_1 = 1 < 2.77$. \begin{figure}[h] \centering \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.5in]{graph8.jpg} \caption{$c_1 = 1$} \label{graph8} \end{subfigure} \quad \begin{subfigure}[h]{0.3\textwidth} \centering \includegraphics[width=1.5in]{graph9.jpg} \caption{$c_1 = 5$} \label{graph9} \end{subfigure} \caption{Plot of (\ref{h2rho2}) of $c_1$ and constant $c_2 = 1$.} \label{h2figures} \end{figure} Although Figure (\ref{graph8}) seems physically interesting, it is not: a numerical evaluation shows that $r_1(c_1) \approx -0.84$, so that despite being real, it is negative. In turn, Figure (\ref{graph9}) should appear strange: it contains two nonzero real roots while the analytic expressions above suggest that there is at most one non-null real root. Again, a numerical evaluation gives $r_1(c_1 = 5) \approx 1.15$, $r_2(c_1 = 5) \approx -1.35 + 1.69\times 10^{-21}i$ and $r_3(c_1 = 5) \approx 0.21 - 4.24\times 10^{-22}i$. So, we see that when $c_1$ grows, the imaginary parts of the roots $r_2$ and $r_3$ become increasingly smaller, allowing us to discard them. Thus, $c_1\approx 2.77$ determines a critical configuration such that for $c_1<2.77$ there are no singularities and for $c_1>2.77$ there are two of them. \section{Summary} \label{concluding_remarks} In this paper we considered the problem of classifying the stellar systems modeled by the piecewise differential TOV equation (\ref{tov1}). We began by introducing the problem in the general context presented in \cite{TOV2}, allowing us to formalize the problem as the determination of the structure of $\mathrm{TOV}^{k,l}_I$ as a subspace of $C^{k,l,l+1}_{pw}(I)$ relative to some locally convex topology. We showed that this subspace is generally not open if the topology is Hausdorff. We introduced another subspace $\mathrm{Psd}^{l+1}(I) \subset C^{k,l,l+1}_{pw}(I)$ of pseudo-asymptotic systems and we showed that for $l=\infty$ this space is at least fifteen-dimensional and that there are good reasons to believe that its dimension is larger. We then considered extended TOV systems, which are TOV systems defined on the extended real line $\mathbb{R}$, and we showed that if a pseudo-asymptotic systems has a nice behavior when extended to $\mathbb{R}$, then it is actually a extended TOV system, leading us to conclude that the space $\mathrm{\overline{\overline{TOV}}}^{\infty,\infty}$ of all of them is at least fifteen-dimensional. We presented a method of classification of these solutions, applying it to some of them, allowing us to show that there are new pseudo-asymptotic TOV systems consisting only of ordinary matter and that contain no cavities or singularities.

a13e2ce30da2960f26a08f3d9582b83e3644de01

\section{Introduction} Networks (or graphs) are used to model data arising from a wide range of applications -- ranging from social network mining, to biology and neuroscience. Of particular interest here are applications that concern learning tasks over networks. For example, in social networks, to answer whether two given users belong in the same social community or not, examining their direct relationship does not suffice -- the users can have many further characteristics in common, such as friendships and interests. Similarly, in friendship recommendations, in order to determine whether two unlinked users are similar, we need to obtain an informative representation of the users and their proximity -- that potentially is not fully captured by handcrafted features extracted from the graph. Towards this direction, representation (or feature) learning algorithms are useful tools that can help us addressing the above tasks. Given a network, those algorithms embed it into a new space (usually a compact vector space), in such a way that both the original network structure and other ``implicit'' features of the network are captured. \par In this paper, we are interested in unsupervised (non-task-specific) feature learning methods; once the vector representations are obtained, they can be used to deal with various data mining tasks on the network, including node classification, community detection and link prediction by existing machine learning tools. Nevertheless, it is quite challenging to design an ``ideal" algorithm for network embeddings. The main reason is that, in order to learn task-independent embeddings, we need to preserve as many important properties of the network as possible in the embedding feature space. But such properties might be not apparent and even after they are discovered, preserving some of them is intractable. This can explain the fact that most of the traditional methods only aim at retaining the first and the second-order proximity between nodes \cite{isomap00}, \cite{lle00}, \cite{laplacian02}. In addition to that, network representation learning methods are expected to be efficient so that they can scale well on large networks, as well as are able to handle different types of rich network structures, including (un)weighted, (un)directed and labeled networks. \par In this paper, we propose BiasedWalk, an unsupervised and Skip-gram-based \cite{skipgram13} network representation learning algorithm which can preserve higher-order proximity information, as well as is able to capture both the homophily and role equivalence relationships between nodes. BiasedWalk relies on a novel node sampling procedure based on biased random walks, that can behave as actual depth-first-search (DFS) and breath-first-search (BFS) explorations -- thus, forcing the sampling scheme to capture both role equivalence and homophily relations between nodes. Furthermore, BiasedWalk is scalable on large scale graphs, and is able to handle different types of network structures, including (un)weighted and (un)directed ones. Our extensive experimental evaluation indicates that BiasedWalk outperforms the state-of-the-art methods on various network datasets and learning tasks. \par Code and datasets for the paper are available at: \href{https://goo.gl/Easwk4}{{\color{blue}{\textbf{https://goo.gl/Easwk4}}}}. \section{Related Work} \label{sec:related} Network embedding methods have been well studied since the 2000s. The early works on the topic consider the embedding task as a dimensionality reduction one, and are often based on factorizing some matrix associated to pairwise distances between all nodes (or data points) \cite{isomap00,lle00,laplacian02}. Those methods rely on the eigen-decomposition of such a matrix are often sensitive to noise (i.e.\ missing or incorrect edges). Some recently proposed matrix factorization-based methods focus on the formal problem of representation learning on networks, and are able to overcome limitations of the traditional manifold learning approaches. For example, while traditional ones only focus on preserving the first-order proximity, GraRep \cite{grarep15} aims to preserve a higher-order proximity by using many matrices, each of them is for $k$-step transition probabilities between nodes. The HOPE algorithm \cite{hope16} defines some similarity measures between nodes which are helpful for preserving higher-order proximities as well and formulates those measures as a product of sparse matrices to efficiently find the latent representations. \par Recently, there is a category of network embedding methods that rely on representation learning techniques in natural language processing (NLP) \cite{deepwalk14}, \cite{tadw2015}, \cite{node2vec16}, \cite{revisiting2016}, \cite{mmdeepwalk2016}, \cite{dynamic2018}. The two best-known methods here are DeepWalk \cite{deepwalk14} and node2vec \cite{node2vec16}. Which both are based on the Skip-gram model \cite{skipgram13}. The model has been proposed to learn vector representations for words in a corpus by maximizing the conditional probabilities of predicting contexts given the vector representations of those words. DeepWalk \cite{deepwalk14} is the first method to leverage the Skip-gram model for learning representations on networks, by extracting truncated random walks in the network and considering them as sentences. The sentences are then fed into the Skip-gram model to learn node representations. The node2vec algorithm \cite{node2vec16} can essentially be considered as an extension of Deepwalk. In particular, the difference between those two methods concerns the node sampling approach: instead of sampling nodes by uniform random walks, node2vec uses biased (and second-order) random walks to better capture the network structure. Because the Skip-gram based methods use random walks to approximate high order proximities, their main advantages include both scalability and the ability to learn latent features that only appear when we observe a high order proximity (\textit{e.g.,} the community relation between nodes). \par As we will present later on, the proposed method which is called BiasedWalk, is also based on the Skip-gram model. The core of BiasedWalk is a node sampling procedure which is able to generate biased (and first-order) random walks that behave as actual depth-first-search (DFS) and breath-first-search (BFS) explorations -- thus, helping towards capturing important structural properties, such as community information and structural roles of nodes. It should be noted that node2vec \cite{node2vec16} also aims to sample nodes based on DFS and BFS random walks, but it cannot control the distances between sampling nodes and the source node. Controlling the distances is crucial as in many specific tasks, such as the one of link prediction, nodes close to the source should have higher probability to form a link from it. Similarly, in the community detection task, given a fixed sampling budget we prefer to visit as many community-wide nodes as possible. Another shortcoming of node2vec is that its second-order random walks require storing the interconnections between the neighbors of every node. More specifically, node2vec's space complexity is $\mathcal{O}(\tilde{D} \cdot |E|)$, where $\tilde{D}$ is the average degree of the input network $G = (V, E)$ \cite{node2vec16}. This can be an obstacle towards embedding dense networks, where the node degrees can be very high. Moreover, we prefer random-walk based sampling rather than pursuing pure DFS and BFS sampling because of the Markovian property of random walks. Given a random walk of length $L$, we can immediately obtain $L-k$ context sets for nodes in the random walk by sliding a window of size $k$ along the walk. \par The LINE algorithm \cite{line15}, though not belonging to any of the previous categories, is also one of the well-known network embedding methods. LINE derives two optimization functions for preserving the first and the second order proximity, and performs the optimizations by stochastic gradient descent with edge sampling. Nevertheless, in general, the performance tends to be inferior compared to Skip-gram based methods. The interesting reader may refer to the following review articles on representation learning on networks \cite{DBLP:journals/corr/abs-1709-05584}, \cite{gem2018}. Lastly, we should mention here the recent progress on deep learning techniques that have also been adopted to learn network embeddings \cite{sdne16, dngr16, deeplearning4, ANRL2018, deeplearning5, deeplearning6}. \section{Problem Formulation} \label{sec:formulation} A network can be represented by a graph $G = (V,E)$, where $V$ is the set of nodes and $E$ is the set of links in the network. Depending on the nature of the relationships, $G$ can be (un)weighted and (un)directed. Table \ref{tab:TableOfNotation} includes notation used throughout the paper. \begin{table}[t]\caption{Table of notation.} \begin{center \begin{tabular}{l p{6cm} } \toprule $G = (V, E)$ & An input network with set $V$ of nodes and set $E$ of links \\ $\Phi(u)$ & Vector representation of $u$ \\ $C(u)$ & Set of neighborhood nodes of $u$ \\ $\Phi^\prime(w)$ & Vector representation of $w$ as it is in the role of a context node \\ $p_v$ & Probability of $v$ being the next node of the (uncompleted) random walk \\ $\tilde{D}$ & Average degree of the input network \\ $\alpha$ & Parameter for controlling the distances from the source node to sampled nodes \\ $L$ & Maximum length of random walks \\ $\tau_v$ & Proximity score of node $v$ \\ $\gamma$ & The number of sampled random walks per node \\ $N(u)$ & Set of neighbors of node $u$ \\ $c$ & Window size of training context of the Skip-gram \\ $\Psi(u, v)$ & Hadamard operator on vector representations of $u$ and $v$ \\ \bottomrule \end{tabular} \end{center} \label{tab:TableOfNotation} \end{table} An \textit{embedding} of $G = (V, E)$ is a mapping from the node set to a low-dimensional space, $\Phi: V \rightarrow \mathbb{R}^d$, where $d \ll |V|$. Network embedding methods are expected to be able to preserve the structure of the original network, as well as to learn latent features of nodes. In this work, inspired by the Skip-gram model, we aim to find embeddings towards predicting neighborhoods (or ``context'') of nodes: \begin{equation}\label{eq1} \argmax\limits_{\Phi} {\sum\limits_{u \in V} \log p(C(u) \mid \Phi(u))} \end{equation} \noindent where $C(u)$ denotes the set of neighborhood nodes of $u$. (Note that, $C(u)$ is not necessarily a set of immediate neighbors of $u$, but it can be any set of sampled nodes for $u$. The insight is that nodes sharing the same context nodes are intended to be close enough in the graph or to be similar). For example, $C(u)$ could be a set of nodes within $k$-hop distance from $u$, or nodes that co-appear with $u$ in a short random walk over the graph. Some related studies have shown that the consequent embedding result is strongly affected by the adopted strategy for sampling the context nodes \cite{deepwalk14}, \cite{node2vec16}. For the purpose of making the problem tractable, we also assume that predicting nodes in a context set is independent of each other, thus Eq. \eqref{eq1} can be expressed as: \begin{equation}\label{eq2} \argmax\limits_{\Phi} {\sum\limits_{u \in V} \sum\limits_{v \in C(u)} \log p(v \mid \Phi(u))} \end{equation} \noindent Lastly, we use the softmax function to define the probability of predicting a context node $v$ given vector representation $\Phi(u)$ of $u$, as $p(v \mid \Phi(u)) = \frac{ \exp({\Phi^\prime(v)}^\intercal \Phi(u)) }{ \sum\limits_{w \in V} \exp({\Phi^\prime(w)}^\intercal \Phi(u)) }$, where $\Phi^\prime(w)$ is the context vector representation of $w$. Note that, there are two representation vectors for each $w \in V$: $\Phi(w)$ as $w$ is in the role of a target node, and $\Phi^\prime(w)$ with $w$ considered as a context node \cite{skipgram13}. \section{Proposed Methodology} \label{sec:methodology} In this section, we introduce BiasedWalk, a method for learning representation for nodes in a network based on the Skip-gram model. To achieve the goal for both capturing the network structure effectively and leveraging the efficient Skip-gram model for learning vector representation, we first propose a novel sampling procedure based on biased random-walks that are able to behave as actual depth-first-search (DFS) and breath-first-search (BFS) sampling schemes. Figure \ref{fig:walks} (a) shows an example of our biased random walks. On a limited budget of samples, random walks staying in the local neighborhood of the source node can be an alternative for BFS sampling \cite{node2vec16}, and such BFS random walks help in discovering the role of source nodes. For example, hub, leaf and bridge nodes can be determined by their neighborhood nodes only. On the other hand, random walks which move further away from the source node are equivalent to DFS sampling since such DFS walks could discover nodes in the community-wide area of the source, and this is helpful to understand the homophily effect between the nodes. \par Given a random walk $u_1, u_2,..., u_L$, the contexts of $u_i$ are nodes in a window of size $c$ centered at that node, i.e., $C(u_i) = \{u_j \mid -c \leq j-i \leq c, j \neq i\}$. Then, we consider each of generated random walks as a sentence in a corpus to learn vector representations for words (\textit{e.g.,} nodes in the random walks) which maximize Eq. \eqref{eq2}. It should be noted here that the efficiency of each Skip-gram based method depends on its own sampling strategy. Our method, instead of uniform random walks (as performed by Deepwalk), uses biased ones for better capturing the network structure. Furthermore, it does not adopt the second-order random walks like node2vec since they are not able to control the distances between the source node and sampled nodes. \begin{figure}[t] \centering \begin{tabular}{ccc} \includegraphics[width=.2\textwidth]{g1.pdf} &~& \includegraphics[width=.2\textwidth]{g2.pdf} \\ (a) &~& (b) \end{tabular} \caption{(a) An example of BFS- and DFS-behavior random walks. \textmd{The two walks in this example both start from node 1. The BFS walk 1-2-3-1-4-7 (red) consists of many nodes in the local neighborhood of the source node, while nodes in the DFS walk 1-2-5-6-10-11 (blue) belong to a much wider portion of the network.} (b) An illustration of BiasedWalk's random-walk based sampling strategy. \textmd{The (uncompleted) walk sequence here is 1-5-7. Every node that has at least one neighbor visited, owns a proximity score that is shown near that node. For each term in a score sum, the node ``giving'' the amount of score is indicated by the same color. The next node for the walk will be one of the neighbors of node 7. In case of DFS-style sampling, nodes 8 and 9 should have higher probability of being selected as the next node (as their proximity scores are lower) compared to nodes 4 and 5.}} \label{fig:walks} \end{figure} To be able to simulate both DFS and BFS explorations, BiasedWalk uses additional information $\tau_v$, called \textit{proximity score,} for each candidate node $v$ of the ongoing random walk, in order to estimate how far (not by the exact number of hops) a candidate is from the source node. More specifically, the nodes whose all neighbor nodes have never been visited by the random walk, should have a proximity score of zero. After the $i$-th node in the walk is discovered, the proximity score of every node adjacent to that node will be increased by $\alpha^{i-1}$, where $0 < \alpha \leq 1$ is a parameter \footnote{The parameter is to control the distances from the source to sampled nodes. Note that, in case $G$ is directed, we increase the proximity scores for all of the in- and out-neighbors of that node.}. Then, the probability distribution of selecting the next node for the current walk is calculated based on the proximity scores of the neighbor nodes of the most recently visited node, and on which type of samplings (DFS or BFS) we desire. In the case of BFS, the probability of a node being the next node should be proportional to its proximity score, i.e., $p_v = \frac{\tau_v}{\sum\limits_{w \in N(u)}{\tau_w}}$. In the case of DFS, the probability should be inversely proportional to that score, i.e., $p_v = \frac{\sfrac{1}{\tau_v}} {\sum\limits_{w \in N(u)} {\sfrac{1}{\tau_w}}}$, where $u$ is the most recently visited node and $N(u)$ defines the set of neighbor nodes of $u$. An illustration of our random-walk based sampling procedure is given in Figure \ref{fig:walks} (b), and its main steps are presented in Algorithm \ref{alg1}. \par The reason for using an exponential function of the current walk length for increasing the proximity scores, is that it helps to clearly distinguish candidates that belong to the local neighborhood of the source node and others far away from the source. Since $0 < \alpha < 1$, those on the local neighborhood should have much higher scores than the ones outside. In addition, our exponential function guarantees that candidates of the same level of distance from the source, should have comparable proximity scores. Thus, such proximity scores are a good estimate of distances from the source node to candidates, and therefore help selecting the next node for our desired (DFS or BFS) random walks. \par Algorithm \ref{alg2} depicts the complete procedure of the proposed BiasedWalk feature learning method. Given a desired sampling type (DFS or BFS), the procedure first generates a set of random walks from every node in the network based on Algorithm \ref{alg1}. In order to adopt the Skip-gram model for learning representations for the nodes, it considers each of these walks (a sequence of nodes) as a sentence (a sequence of words) in a corpus. Finally, the Skip-gram model is used to learn vector representations for words in the ``network'' corpus, which aims to maximize the probability of predicting context nodes as mentioned in Eq. \eqref{alg2}. \begin{algorithm}[t] \KwIn{A network $G = (V, E)$, source node $s, sampling\_type, \alpha, L$} \KwOut{A random walk of maximum length $L$ starting from node $s$} $walk \leftarrow [s]$ \\ $l \leftarrow 1$ {\textbackslash* The current walk length *\textbackslash} \\ \While {$(l < L \And N(walk[l]) \neq \emptyset)$} { $\tau \leftarrow \{ \}$ \textbackslash* A map from node IDs to their proximity scores to keep track nodes which have at least one neighbor visited by the current walk *\textbackslash \\ $u \leftarrow walk[l]$ \\ \ForEach{$v \in N(u)$} { \If ({\textbackslash* In case none of the neighbors of $v$ has been visited *\textbackslash}){$v \not\in \tau.keys()$} { $\tau[v] \leftarrow \alpha^{l-1}$ } \Else { $\tau[v] \leftarrow \tau[v] + \alpha^{l-1}$ } } \ForEach{$v \in N(u)$} { \If {$sampling\_type = BFS$} { $p_v \leftarrow \frac{\tau[v]}{\sum\limits_{w \in N(u)}{\tau[w]}}$ } \ElseIf {$sampling\_type = DFS$} { $p_v \leftarrow \frac{\sfrac{1}{\tau[v]}} {\sum\limits_{w \in N(u)} {\sfrac{1}{\tau[w]}}}$ } } $z \leftarrow $ randomly select a node in $\{v | v \in N(u)\}$ with the probabilities $p_v$ calculated \\ Append node $z$ to $walk$ \\ $l \leftarrow l + 1$ \\ } \Return $walk$ \caption{Global Random Walk}\label{alg1} \end{algorithm} \begin{algorithm}[h] \KwIn{A network $G = (V, E), sampling\_type, \alpha$, number of walks per node $\gamma$, maximum walk length $L$} \KwOut{Vector representations of nodes in $G$} $walk\_set \leftarrow [\ ]$ \\ \For ({\textbackslash* generate $\gamma$ random walks from each node *\textbackslash}){$loop = 1$ \bf{to} $\gamma$} { \ForEach{$s \in V$} { $walk \leftarrow Global\_Random\_Walk(G, s, sampling\_type, \alpha, L)$ {\textbackslash* Alg. \ref{alg1} *\textbackslash} \\ Append $walk$ to $walk\_set$ \\ } } Construct a corpus $T$ consisting of $\gamma.|V|$ sentences where each sentence corresponds to a walk in $walk\_set$. The vocabulary $W$ of this corpus contains $|V|$ words, each corresponding to a node in $G$ \\ Use the Skip-gram to learn representations of words in corpus $T$ with vocabulary $W$ \\ \Return Corresponding vector representations for nodes in $G$ \caption{BiasedWalk}\label{alg2} \end{algorithm} \paragraph{\textbf{Space and time-complexity}} Let $\tilde{D}$ be the average degree of the input network, and assume that every node in the network has a degree bounded by $\mathcal{O}(\tilde{D})$. For each node visited in a random walk of maximum length $L$, Algorithm \ref{alg1} needs to consider all neighbors of that node to update (or initialize) their proximity scores, and then calculate the transition probabilities. The algorithm uses a map $\tau$ to store proximity scores of such neighbor nodes, so the number of keys (node IDs) in the map is $\mathcal{O}(\tilde{D} \cdot L)$. Thus, the time for both accessing and updating a proximity score is in $\mathcal{O}(\log{\tilde{D}} + \log{L})$, as the map can be implemented by a balanced binary search tree. The algorithm needs to select at most $L$ nodes whose degree is $\mathcal{O}(\tilde{D})$ to construct a walk. Therefore, the number of updating and accessing operations is $\mathcal{O}(L \cdot \tilde{D})$, and thus the time complexity of Algorithm \ref{alg1} is $\mathcal{O}(L \cdot \tilde{D}(\log{\tilde{D}} + \log{L}))$. With respect to memory requirement, BiasedWalk requires only $\mathcal{O}(|E|)$ space complexity since it adopts the first-order random walks. \section{Experimental Evaluation} \label{sec:experimental} For the experimental evaluation, we use vector representations of nodes learned by BiasedWalk and compare with four baseline methods, including DeepWalk \cite{deepwalk14}, node2vec \cite{node2vec16}, LINE \cite{line15} and HOPE \cite{hope16} in the tasks of multilabel node classification and link prediction. For each type of our random walks (DFS and BFS), the value $\alpha$ of BiasedWalk varies in the range of $\{0.125, 0.25, 0.5, 1.0\}$. Then, we use 10-fold cross-validation on labeled data to choose the best parameters (including the walk type and value $\alpha$) for each graph dataset. \subsection{Network datasets} Table \ref{table:1} provides a summary of network datasets used in our experiments. More specifically, network datasets for the multilabel classification task include the following: \begin{itemize} \item \textsc{BlogCatalog} \cite{blogcatalog2009}: A network of social relationships between the bloggers listed on the BlogCatalog website. There are $10,312$ bloggers, $333,983$ friendship pairs between them, and the labels of each node is a subset of 39 different labels that represent blogger interests (\textit{e.g.,} political, educational). \item Protein-Protein Interaction (PPI) \cite{ppi08}: The network contains $3,890$ nodes, $38,739$ unweighted edges and $50$ different labels. Each of the labels corresponds to a biological function of the proteins. \item \textsc{IMDb} \cite{WS_KDD13}: A network of $11,746$ movies and TV shows, extracted from the Internet Movie Database (IMDb). A pair of nodes in the network is connected by an edge if the corresponding movies (or TV shows) are directed by the same director. Each node can be labeled by a subset of $27$ different movie genres in the database, such as drama, comedy, documentary and action. \end{itemize} For the link prediction task, we use the following network datasets (we focus on the largest (strongly) connected components instead of the whole networks): \begin{itemize} \item \textsc{AstroPhy} collaboration \cite{snap14}: The network represents co-author relationships between 17,903 scientists in AstroPhysics. \item \textsc{Election-Blogs} \cite{AG05}: This is a directed network of front-page hyperlinks between blogs in the context of the 2004 US election. In the network, each node represents a blog and each edge represents a hyperlink between two blogs. \item Protein-Protein Interaction (PPI): The same dataset used in the node-classification task. \item \textsc{Epinions}\cite{epinions2003}: The network represents who-trust-whom relationships between users of the \textit{epinions.com} product review website. \end{itemize} \begin{table*}[t] \centering \caption{Network datasets. \label{table:1}} \begin{tabular}{l rc rc cc c} \toprule Network & $|V|$ &~~~~& $|E|$ &~~~~& Type &~~~~& {\#}Labels \\ \midrule \textsc{BlogCatalog} & 10,312 &~~~~& 333,983 &~~~~& Undirected &~~~~& 39 \\ \textsc{PPI} & 3,890 &~~~~& 38,739 &~~~~& Undirected &~~~~& 50 \\ \textsc{IMDb} & 11,746 &~~~~& 323,892 &~~~~& Undirected &~~~~& 27 \\ \midrule \textsc{AstroPhy} & 17,903 &~~~~& 197,031 &~~~~& Undirected &~~~~& \\ \textsc{Election-Blogs} & 1,222 &~~~~& 19,021 &~~~~& Directed &~~~~& \\ \textsc{PPI} (for link prediction) & 3,852 &~~~~& 21,121 &~~~~& Undirected &~~~~& \\ \textsc{Epinions} & 75,877 &~~~~& 508,836 &~~~~& Directed &~~~~& \\ \bottomrule \end{tabular} \end{table*} \subsection{Baseline methods} We use DeepWalk \cite{deepwalk14}, node2vec \cite{node2vec16}, LINE \cite{line15} and HOPE \cite{hope16} as baselines for BiasedWalk. The number of dimensions of output vector representations $d$ is set to 128 for all the methods. To get the best embeddings for LINE, the final representation for each node is created by concatenating the first-order and the second-order representations each of 64 dimensions \cite{line15}. The Katz index with decay parameter $\beta = 0.1$ is selected for HOPE's high-order proximity measurement, since this setting gave the best performance in the original article \cite{hope16}. Similar to BiasedWalk, DeepWalk and node2vec belong to the same category of Skip-gram based methods. The parameters for both the node sampling and optimization steps of the three methods are set exactly the same: number of walks per node $\gamma = 10$; maximum walk length $L = 80$; Skip-gram's window size of training context $c = 10$. Since node2vec requires the in-out and return hyperparameters $p, q$ for its second-order random walks, we have performed a grid search over $p, q \in \{0.25, 0.5, 1, 2, 4\}$ and 10-fold cross-validation on labeled data to select the best embedding -- as suggested by the experiment in \cite{node2vec16}. For a fair comparison, the total number of training samples for LINE is set equally to the number of nodes sampled by the three Skip-gram based methods ($\#Sample = |V| \cdot \gamma \cdot L$). \subsection{Experiments on multilabel classification} Multilabel node classification is a challenging task, especially for networks with a large number of possible labels. To perform this task, we have used the learned vector representations of nodes and an one-vs-rest logistic regression classifier using the \textsc{LibLinear} library with $L2$ regularization \cite{liblinear08}. The Micro-F1 and Macro-F1 scores in the 50\%-50\% train-test split of labeled data are reported in Table \ref{table:2}. The best scores gained for each network across the methods are indicated in bold, and the best parameter settings including the walk type and the corresponding value $\alpha$ for BiasedWalk have been shown in the last row. In our experiment (including the link prediction task), each Micro-F1 or Macro-F1 score reported was calculated as the average score on $10$ random instances of a given train-test split ratio. \par The experimental results show that the Skip-gram based methods outperform the rest ones in the multilabel classification task. The main reason is that, the rest methods mainly aim to capture low order proximities for networks. But this is not enough for node classification as nodes which are not in the same neighborhood can be classified by the same labels. More precisely, BiasedWalk gives the best results for \textsc{BlogCatalog} and PPI networks; in \textsc{BlogCatalog}, it improves Macro-F1 score by 15\% and Micro-F1 score by 6\%. In \textsc{IMDb} network, node2vec and BiasedWalk are comparable, with the former being slightly better. Figure \ref{fig:multi_label} depicts the performance of all the methods in different percentages of training data. Given a percentage of training data, in most cases, BiasedWalk has better performance than the rest baselines. An interesting observation from the experiment is that, preserving homophily between nodes is important for the node classification task. This explains the reason why the best performance results obtained by BiasedWalk come from its DFS random-walk based sampling scheme. \begin{table*}[t] \centering \caption{Macro-F1 and Micro-F1 scores (\%) for 50\%-50\% train-test split in the multi-label classification task. \label{table:2}} \begin{tabular}{cccccccccc} \toprule \multirow{3}{*}{Algorithm} && \multicolumn{6}{c}{Networks} \\ & \multicolumn{3}{c}{\textsc{BlogCatalog}} && \multicolumn{2}{c}{\textsc{PPI}} && \multicolumn{2}{c}{\textsc{IMDb}} \\ \cmidrule{3-4} \cmidrule{6-7} \cmidrule{9-10} &~~& Macro & Micro &~~& Macro & Micro &~~& Macro & Micro \\ HOPE &~~& 5.04 & 15.00 &~~& 10.19 & 11.50 &~~& 12.11 & 41.64 \\ LINE &~~& 16.03 & 30.00 &~~& 15.75 & 18.19 &~~& 28.59 & 57.33 \\ \midrule DeepWalk &~~& 22.53 & 37.03 &~~& 15.71 & 17.94 &~~& 42.91 & 65.09 \\ node2vec &~~& 23.82 & 37.52 &~~& 15.76 & 18.22 &~~& \textbf{43.10} & \textbf{66.77} \\ BiasedWalk &~~& \textbf{27.36} & \textbf{39.69} &~~& \textbf{16.30} & \textbf{18.81} &~~& 42.91 & 66.12 \\ (Best combination of: $walk\_type, \alpha$) &~~& \multicolumn{2}{c}{DFS, 1.0} &~~& \multicolumn{2}{c}{DFS, 0.5} &~~& \multicolumn{2}{c}{DFS, 0.25} \\ \bottomrule \end{tabular} \end{table*} \begin{figure*}[t] \centering \includegraphics[width=.98\textwidth]{classification.pdf} \caption{Performance evaluation on various datasets for the node classification task.} \label{fig:multi_label} \end{figure*} \subsection{Experiments on link prediction} To perform link prediction in a network dataset, we first randomly remove half of its edges, so that after the removal the network is not disconnected. The node representations are then learned from the remaining part of the network. To create the negative labels for the prediction task, we randomly select pairs of nodes that are not connected in the original network. The number of such pairs is equal to the number of removed edges. The ``negative'' pairs and the pairs from edges that have been removed, are used together to form the labeled data for this task. Because the prediction task is for edges, we need some way to combine a pair of vector representations of nodes to form a single vector for the corresponding edge. Some operators have been recommended in \cite{node2vec16} for handling that. We have selected the Hadamard operator, as it gave the best overall results. Given vector representations $\Phi(u)$ and $\Phi(v)$ of two nodes $u$ and $v$, the Hadamard operator defines the vector representation of link $(u, v)$ as $\Psi(u, v)$, where $\Psi_i(u, v) = \Phi_i(u) \cdot \Phi_i(v), \forall i =1, \ldots, d$. Then, we use a Linear Support Vector Classifier with the $L2$ penalty to predict whether links exist or not. The experimental results are reported in Table \ref{table:3}. Each score value shows the average over $10$ random instances of a 50\%-50\% train-test split of the labeled data. In this task, HOPE is still inferior to the other methods, except in the \textsc{Election-Blogs} network. Since HOPE is able to preserve the asymmetric transitivity property of networks, it can work well in directed networks, such as the \textsc{Election-Blogs}. LINE's performance is comparable to that of the Skip-gram based methods and even the best one in the PPI network. The reason is that LINE is proposed to preserve the first and the second-order proximity and that is really helpful for the link prediction task. It is totally possible to infer the existence of an edge if we know the role of its nodes and meanwhile, roles of nodes can be discovered by examining just their local neighbors. BiasedWalk, again, gains the best results for this task in most of the networks. Finally, the best results of BiasedWalk in this task are almost based on its BFS sampling. This supports the fact that, discovering role equivalence between nodes is crucial for the link prediction task. \begin{table*}[t] \centering \caption{Macro-F1 and Micro-F1 scores (\%) of the 50\%-50\% train-test split in the link prediction task. HOPE could not handle the large \textsc{Epinions} graph because of its high time and space complexity.} \label{table:3} \begin{tabular}{c c ccccc c c c cc c c} \hline \multirow{4}{*}{Algorithm} && \multicolumn{8}{c}{Network} \\ & \multicolumn{3}{c}{\textsc{AstroPhy}} && \multicolumn{2}{c}{\textsc{Election-Blogs}} && \multicolumn{2}{c}{\textsc{PPI}} && \multicolumn{2}{c}{\textsc{Epinions}} \\ \cmidrule{3-4} \cmidrule{6-7} \cmidrule{9-10} \cmidrule{12-13} &~~& Macro~~ & Micro &~~& Macro & Micro &~~& Macro & Micro &~~& Macro~~ & Micro \\ HOPE &~~& 59.97~~ & 65.61 &~~& 76.34 & 76.99 &~~& 53.15~~ & 61.45 &~~& *~~ & * \\ LINE &~~& 93.56~~ & 93.59 &~~& 74.79 & 74.80 &~~& \textbf{80.29}~~ & \textbf{80.34} &~~& 81.94 & 81.95 \\ \midrule DeepWalk &~~& 92.29~~ & 92.32 &~~& 79.06 & 79.12 &~~& 67.84~~ & 67.99 &~~& 89.14 & 89.18 \\ node2vec &~~& 93.08~~ & 93.10 &~~& 81.04 & 81.15 &~~& 71.24~~ & 71.29 &~~& 89.29 & 89.33 \\ BiasedWalk &~~& \textbf{95.08} ~~& \textbf{95.10} &~~& \textbf{81.54} & \textbf{81.65} &~~& 75.23~~ & 75.34 &~~& \textbf{90.08} & \textbf{90.11} \\ (The best $walk\_type, \alpha$) &~~& \multicolumn{2}{c}{BFS, 0.125} &~~& \multicolumn{2}{c}{BFS, 0.125} &~~& \multicolumn{2}{c}{BFS, 1.0} &~~& \multicolumn{2}{c}{BFS, 0.25} \\ \hline \end{tabular} \end{table*} \subsection{Parameter sensitivity and scalability} We also evaluate how BiasedWalk's performance is changing under different parameter settings. Figure \ref{fig:parameter} shows Macro-F1 scores gained by BiasedWalk in the multilabel classification task in \textsc{BlogCatalog} (the Micro-F1 scores follow a similar trend then it is not necessary to show them here). Except the parameter is being considered, all other parameters in the experiment are set to their default value. Obviously, since the number of walks per node or the walk length is increased, there are more nodes sampled by the random walks and then BiasedWalk should get a better score. The effectiveness of BiasedWalk also depends on the dimension number of output vector representations $d$. Since the number of dimensions is too small, embeddings in the representation space may not be able to preserve the structure information of input networks, and as this parameter is set so high it could negatively affect consequent classification tasks. Finally, we can notice the dependence of Macro-F1 score and the tendency of parameter $\alpha$, this can support us in inferring the best value for the parameter on each network dataset. \begin{figure*}[t] \centering \includegraphics[width=.73\textwidth]{parameter_all.pdf} \caption{Parameter sensitivity of BiasedWalk in \textsc{BlogCatalog}. \label{fig:parameter}} \end{figure*} \par We have also examined the efficiency of the proposed BiasedWalk algorithm. Figure \ref{fig:scalability} depicts the running time required for sampling (blue curve) and both sampling and optimization (orange curve) on the Erd\H{o}s-R\'{e}nyi graphs of various sizes ranging from $100$ to $1M$ nodes with the average degree of 10. As we can observe, BiasedWalk is able to learn embeddings for graphs of millions of nodes in dozens of hours and scales linearly with respect to the size of graphs. Nearly total learning time belongs to the step of sampling nodes that means the Skip-gram is very efficient at solving Eq. \eqref{eq2}. \begin{figure}[h] \begin{center} \includegraphics[width=0.4\textwidth]{running_time.pdf} \end{center} \caption{Scalability of BiasedWalk on Erd\H{o}s-R\'{e}nyi graphs. \label{fig:scalability}} \end{figure} \section{Conclusions and Future Work} \label{sec:conclusions} In this work, we have proposed BiasedWalk, a Skip-gram based method for learning node representations on graphs. The core of BiasedWalk is a node sampling procedure using biased random-walks, that can behave as actual depth-first-search and breath-first-search explorations -- thus, forcing BiasedWalk to efficiently capture role equivalence and homophily between nodes. We have compared BiasedWalk to several state-of-the-art baseline methods, demonstrating its good performance on the link prediction and multilabel node classification tasks. As future work, we plan to theoretically analyze the properties of the proposed biased random-walk scheme and to investigate how to adapt the scheme for networks with specific properties, such as signed networks and ego networks, in order to obtain better embedding results. \balance \bibliographystyle{splncs_srt}

c32fcee59f89ec99e1da34d99072003801d1ffd7

\subsection*{#1}\verbatiminput{#1}} \newcommand{\includematlab}[1]{\subsection*{#1}\lstinputlisting{#1}} \newcommand{\figref}[1]{\ref{#1}} \newcommand{\pder}[2]{\frac{{\partial}#1}{{\partial}#2}} \newcommand{\der}[2]{\frac{{\mathrm d}#1}{{\mathrm d}#2}} \newcommand{\dersqr}[2]{\frac{{\mathrm d^2}#1}{{{\mathrm d}#2}^2}} \newcommand{\trace}{{\mathrm{Tr }}} \newcommand{\KK}{\mathcal{K}} \newcommand{\E}[1]{\cdot 10^{#1}} \newcommand{\RR}{\mathbb{R}} \newcommand{\DD}{\mathbb{D}} \newcommand{\BB}{\mathbb{B}} \newcommand{\CC}{\mathbb{C}} \newcommand{\LL}{\mathbb{L}} \newcommand{\LLL}{\mathcal{L}} \newcommand{\MM}{\mathbb{M}} \newcommand{\ZZ}{\mathbb{Z}} \newcommand{\PP}{\mathbb{P}} \newcommand{\QQ}{\mathbb{Q}} \newcommand{\NN}{\mathbb{N}} \newcommand{\cnj}[1]{\bar{#1}} \newcommand{\diag}{\operatorname{diag}} \newcommand{\kron}{\otimes} \newcommand{\norm}[1]{||#1||} \newcommand{\Arg}[1]{{\mathrm{Arg}~{#1} }} \newcommand{\LLp}{\LL} \newcommand{\LLm}{\LL_-} \newcommand{\hp}{\bar{h}} \newcommand{\matlabfun}[1]{\texttt{#1}} \newcommand{\cover}[1]{{\operatorname{cov}(#1)}} \newcommand{\optcov}{{\operatorname{optcov}}} \newcommand{\Nregions}{{N_{regions}}} \newcommand{\Trep}{{T_{REP}}} \newcommand{\sdiv}{{\operatorname{div}}} \newcommand{\eigvec}{{\operatorname{eigvec}}} \newcommand{\Tau}{\mathrm{T}} \newcommand{\atantwo}[2]{\atan \binom{#1}{#2}} \newcommand{\infvecset}{\vect(\CC^{n\times \infty})} \newcommand{\ej}[1]{{\color{blue}#1}} \newtheorem{theorem}{Theorem}[section] \newtheorem{corollary}{Corollary} \newtheorem*{main}{Main Theorem} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}{Proposition} \newtheorem{conjecture}{Conjecture} \newtheorem*{problem}{Problem} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}{Remark} \newtheorem*{notation}{Notation} \newtheorem{thm}{Theorem} \newtheorem{prop}[thm]{Proposition} \newtheorem{lem}[thm]{Lemma} \newtheorem{cor}[thm]{Corollary} \newtheorem{example}[thm]{Example} \newcommand{\ep}{\varepsilon} \newcommand{\eps}[1]{{#1}_{\varepsilon}} \title[Convergence of SCF using a density matrix approach] { A density matrix approach to the convergence of the self-consistent field iteration} \author[Parikshit Upadhyaya, Elias Jarlebring, Emanuel Rubensson]{} \keywords{self-consistent field iteration, convergence analysis, nonlinear eigenvalue problems, eigenvector nonlinearity, electronic structure calculations, iterative methods} \email{[email protected]} \email{[email protected]} \email{[email protected]} \begin{document} \maketitle \centerline{\scshape Parikshit Upadhyaya} \medskip {\footnotesize \centerline{Lindstedtsv\"agen 25,} \centerline{Department of Mathematics,} \centerline{SeRC - Swedish e-Science research center,} \centerline{Royal Institute of Technology, SE-11428 Stockholm, Sweden} } \medskip \centerline{\scshape Elias Jarlebring} \medskip {\footnotesize \centerline{Lindstedtsv\"agen 25,} \centerline{Department of Mathematics,} \centerline{SeRC - Swedish e-Science research center,} \centerline{Royal Institute of Technology, SE-11428 Stockholm, Sweden} } \medskip \centerline{\scshape Emanuel H. Rubensson} \medskip {\footnotesize \centerline{Division of Scientific Computing, Department of Information Technology,} \centerline{Uppsala University, Box 337, SE-75105 Uppsala, Sweden} } \begin{abstract} In this paper, we present a local convergence analysis of the self-consistent field (SCF) iteration using the density matrix as the state of a fixed-point iteration. Sufficient and almost necessary conditions for local convergence are formulated in terms of the spectral radius of the Jacobian of a fixed-point map. The relationship between convergence and certain properties of the problem is explored by deriving upper bounds expressed in terms of higher gaps. This gives more information regarding how the gaps between eigenvalues of the problem affect the convergence, and hence these bounds are more insightful on the convergence behaviour than standard convergence results. We also provide a detailed analysis to describe the difference between the bounds and the exact convergence factor for an illustrative example. Finally we present numerical examples and compare the exact value of the convergence factor with the observed behaviour of SCF, along with our new bounds and the characterization using the higher gaps. We provide heuristic convergence factor estimates in situations where the bounds fail to well capture the convergence. \end{abstract} \section{Introduction}\label{sect:intro} Let $A:M\to M$, where $M\subset\CC^{n\times n}$ denotes the set of Hermitian matrices. In this work we consider the associated nonlinear eigenvalue problem consisting of determining $(X_1,\Lambda_1)\in\CC^{n\times p}\times \RR^{p\times p}$ such that $(X_1,\Lambda_1)$ is an invariant pair of $A(X_1X_1^H)$, i.e., \begin{subequations}\label{eq:prob} \begin{eqnarray} A(X_1X_1^H)X_1&=&X_1\Lambda_1 ,\\ X_1^HX_1&=& I, \end{eqnarray} \end{subequations} where $\Lambda_1=\diag(\lambda_1,\ldots,\lambda_p)$ and $\lambda_1,\ldots,\lambda_n$ are the eigenvalues of $A(X_1X_1^H)$, numbered in ascending order. This is one of the fundamental computational challenges in quantum chemistry and related fields (Hartree-Fock and Kohn-Sham density functional theory).\footnote{In these settings, the columns of $X_1$ contain the basis-expansion coefficients of the molecular orbitals.} However, this problem also arises as a trace ratio maximization problem in linear discriminant analysis for dimension reduction. See \cite{ngosaadtrp},\cite{zhangliaotrp} and \cite{zhanganotetrp} for more on this application. The self-consistent field (SCF) iteration consists of computing iterates satisfying the linear eigenvalue problem \begin{equation} \label{eq:scf1} A(V_kV_k^H)V_{k+1}=V_{k+1}S_{k+1}, \end{equation} where $V_{k+1}^HV_{k+1}=I$ and $S_{k+1}\in\RR^{p\times p}$ is diagonal. In the case of convergence, $V_{k} \to X_1$ and $S_k \to \Lambda_1$. In this paper we provide a local convergence analysis of this algorithm. SCF is rarely used on its own as a solution method and most of the state-of-the-art procedures are based on its enhancements and improvements, for example, Pulay's DIIS (Direct Inversion in the Iterative Subspace) acceleration in \cite{Pulay:1979}. See also standard references \cite{Szabo:1996:QC},\cite{Helgaker:Jorgensen:Olsen} and further literature discussion below. There is an extensive amount of literature on the convergence of the SCF iteration and its variants. We mention some main approaches to the convergence theory, without an ambition of a complete description. A number of recent works are based on the optimization viewpoint, e.g., \cite{Cances:2000:SCF,Levitt:2012:CONVERGENCE,liu2014convergence,liu2015analysis}. This is natural, since the problem in \eqref{eq:prob} often stems from the first order optimality condition of an energy minimization problem, as in \cite[Section~2.1]{Saad:2010:ELECSTRUCT}. In particular, the Roothaan algorithm with level-shifting and damping are studied in \cite{Cances:2000:SCF}. This analysis was used as a basis for the gradient analysis in \cite{Levitt:2012:CONVERGENCE}, which provided explicit estimates of the convergence rate for the algorithms applied to the Hartree-Fock equations. The convergence of the DIIS acceleration scheme has been studied separately in \cite{ROHWEDDER:2011}. Various approaches are based on measuring the subspace angle and other using chordal norms, e.g., \cite{liu2014convergence,Cai:2017:EIGVEC}, leading to local convergence as well as a global convergence analysis. In contrast to these approaches, we use a density matrix based analysis and derive bounds involving higher gaps (as we explain below). The analysis in \cite{liu2015analysis} provides precise conditions for local convergence (and some global convergence conditions), under the assumption that $A$ only depends on the diagonal of the density matrix $X_1X_1^H$, which in many discretization settings corresponds to the charge density. Not all problems are nonlinear only in the diagonal of the density matrix, as e.g., the example in Section~\ref{subsec:watermolecule}. The work in \cite{Yang:2009:SCF} also provides a convergence analysis, mostly based on a non-zero temperature filter function approach. We note also that a precise local convergence criterion for the classical version was presented in \cite{Stanton:1981:CLOSEDSHELL}, not involving a density matrix analysis. A model of interacting bosons which has received considerable attention is the Gross-Pitaevskii equation. This corresponds to \eqref{eq:prob} with $p = 1$. Convergence results for the SCF iteration for this case can be found for example in \cite{Bai_LV:2017}. Our convergence analysis is focused on establishing a precise characterization of the convergence factor as well as natural upper bounds. We provide an exact formula for the convergence factor, which turns out to be the spectral radius of a matrix (the Jacobian of the fixed-point map). Using this exact formula, we derive upper bounds which can be phrased in terms of higher gaps (as we define later in Definition~\ref{def:hogap}) and the action of a linear operator on the outer products of eigenvectors. We also provide an example where the convergence cannot be characterized based on the first gap alone, which illustrates the importance of taking into account the higher gaps in the convergence analysis. This should be viewed in contrast to the analysis in \cite{liu2014convergence,Cai:2017:EIGVEC,Yang:2009:SCF}, which is primarily focused on the first gap. We will use a formulation of the SCF iteration in terms of the density matrix. In our context, a density matrix is defined by \[ P_{k}:=V_{k}V_{k}^H\in M. \] Given $P_k$ we can compute $A(P_k)=A(V_kV_k^H)$ from which we can compute $V_{k+1}$ and in principle construct $P_{k+1}=V_{k+1}V_{k+1}^H$. Hence, the iteration \eqref{eq:scf1} is equivalent to a fixed point iteration in $P_k$. We will refer to this fixed point map as $\Psi$, i.e., \begin{equation}\label{eq:psifp} P_{k+1}=\Psi(P_k). \end{equation} Although our conclusions hold for general problems, we restrict our analysis to the case where the operator $A$ has the form \begin{equation}\label{eq:ap} A(P)=A_0+\LLL(P), \end{equation} where $A_0\in M$ and $\LLL:\CC^{n\times n}\to \CC^{n\times n}$ is a complex linear operator, i.e., $\LLL(zA)=z\LLL(A)$ for all $z\in\CC$. The density matrix formulation in \eqref{eq:psifp} is the starting point of several linear scaling variants of the SCF-algorithm that avoid explicit construction of $V_{k+1}$\cite{BOWLER:MIZAYAKI}. Before we proceed to the next section, we will introduce some necessary notation and definitions. Let $X,\Lambda$ correspond to a complete eigenvalue decomposition of $A$ evaluated in a solution to \eqref{eq:prob}, i.e., \[ A(X_1X_1^H)X=X\Lambda, \] where \begin{equation}\label{eq:lambda} \Lambda=\begin{bmatrix}\Lambda_1 & \\ & \Lambda_2\end{bmatrix} =\diag(\lambda_1,\ldots,\lambda_p,\lambda_{p+1},\ldots,\lambda_n) \end{equation} and \[ X=[X_1\;\;X_2]=[x_1\;\;\ldots\;\;x_n]. \] \begin{definition}[Gap] The smallest distance between the diagonal elements of $\Lambda_1$ and the diagonal elements of $\Lambda_2$ is defined as as the \emph{gap} and denoted $\delta$. \end{definition} Due to the numbering of eigenvalues, the gap is given by \[ \delta=\min_{i\le p,\;j\ge p+1}|\lambda_i-\lambda_j|=\lambda_{p+1}-\lambda_p. \] Throughout this paper, we assume that $\delta \neq 0$. Otherwise the decomposition \eqref{eq:lambda} is not unique and violates the common uniform well-posedness hypothesis of \cite{Cances:2000:SCF}. \begin{definition}[Higher gap]\label{def:hogap} The $j$-th smallest distance between the diagonal elements of $\Lambda_1$ and $\Lambda_2$ is denoted by $\delta_j$. \end{definition} Note that $\delta_1 = \delta$. As an example, the second gap is given by \[ \delta_2=\underset{(i,j)\neq (p,p+1)}{\min_{i\le p,\;j\ge p+1}}|\lambda_i-\lambda_j| =\min(\lambda_{p+2}-\lambda_p,\lambda_{p+1}-\lambda_{p-1}). \] \begin{definition}[Lower triangular vectorization] Let $m = n(n+1)/2$. The operator $\mathit{\vech} \colon M \to \CC^{m}$ is defined as \begin{equation*} \vech(W) = \begin{bmatrix} w_{1,1}& \cdots& w_{n,1}& w_{2,2}& \cdots & w_{n,2}& w_{3,3}& \cdots & w_{n,n} \end{bmatrix}^T. \end{equation*} \end{definition} This is the vectorization operator adapted for hermitian matrices, and returns the vectorizaton of the lower triangular part. Similarly, we define the inverse operator $\vech^{-1} \colon \CC^{m} \to M$ which maps any vector $v \in \CC^m$ to a corresponding $W \in M$. The relation between $\vect$ and $\vech$ is given by \begin{equation}\label{eq:vvech} \vech(W)=T\vect(W), \end{equation} where $T\in\RR^{m\times n^2}$. The matrix $T$ is in general non-unique (as discussed in \cite{McCull:SymMat} and \cite{Henderson:vech}). In this paper, we will specifically use \eqref{eq:vvech} with \begin{equation* T = \diag(I_n,\begin{bmatrix}0& I_{n-1}\end{bmatrix},\begin{bmatrix}0& 0& I_{n-2}\end{bmatrix},\cdots,1). \end{equation*} \section{Convergence characterization} \subsection{Main theory} The following characterization involves the matrix consisting of reciprocal gaps, which we denote $R\in\RR^{n\times n}$, and is given by \begin{equation}\label{eq:Rdef} R_{i,j}= \begin{cases} \frac{1}{\lambda_j(0)-\lambda_i(0)},\;\;& \textrm{ if }i\le p \textrm{ and } j>p\\ \frac{1}{\lambda_i(0)-\lambda_j(0)},\;\;& \textrm{ if }i> p \textrm{ and } j\le p\\ 0&\textrm{ otherwise}. \end{cases} \end{equation} where $\lambda_1(t),\ldots,\lambda_n(t)$ are eigenvalues of a parameter dependent matrix $B(t)$.\footnote{Note that the matrix $R$ has appeared with different names in other papers before. For example, in \cite{Stanton:1981:CLOSEDSHELL}, it is referred to as the "density perturbation". In \cite{liu2015analysis}, it is called "first divided difference matrix".} The matrix $R$ is symmetric with the following structure: \begin{equation* R = \begin{bmatrix}0 &R_p^T\\R_p& 0\end{bmatrix}, \end{equation*} where $R_p\in \mathbb{R}^{(n-p)\times p}$. We need the following perturbation result whose variants exist in quantum mechanical perturbation theory, for example in chapter 15.III of \cite{Messiah1999}. \begin{lemma}[Density matrix derivatives]\label{thm:densitymatder} Consider a matrix-valued function $B$ depending on a complex parameter such that $B(t) = B_0+B_1t$, where $B_0,B_1$ are Hermitian. Let $X,\Lambda$ correspond to a parameter dependent diagonalization for a sufficiently small neighborhood of $t_0=0$, i.e., \begin{equation*} X(t)\Lambda(t)X(t)^H=B(t) \end{equation*} with $X(t)^HX(t)=I$ and $\Lambda(0)=\operatorname{diag}(\lambda_1(0),\ldots,\lambda_n(0))$, where $\lambda_1(0)<\cdots< \lambda_p(0)<\lambda_{p+1}(0)<\cdots< \lambda_n(0)$. Let $X(t)\in\CC^{n\times n}$ be decomposed as $X(t)=[X_1(t),X_2(t)]$, where $X_1(t)\in\CC^{n\times p}$ and $P(t):=X_1(t)X_1(t)^H$. Then, \begin{equation* \vect(P'(0))= -(\overline{X(0)}\otimes X(0))D(X(0)^T\otimes X(0)^H)\vect(B_1) \end{equation*} where \begin{equation}\label{eq:Ddef} D:=\diag(\vect(R)). \end{equation} \end{lemma} \begin{proof} Let us consider the domain $\mathcal{D} = \left(-\infty,\lambda_{mid}\right)\cup \left(\lambda_{mid},\infty\right)$, where $\lambda_{mid} = \frac{\lambda_p(0)+\lambda_{p+1}(0)}{2}$ and define the step function $h:\mathcal{D}\to \mathbb{R}$, which acts as a filter for the $p$ smallest eigenvalues of $B(0) = B_0$: \begin{equation}\label{eq:heavyside} h(x) = \begin{cases}1,\quad x<\lambda_{mid}\\0,\quad x>\lambda_{mid}\end{cases}. \end{equation} We can generalize $h:M\to M$ as a matrix function (in the sense of \cite{HIGHAMFOM}) and rewrite the density matrix function as \begin{equation}\label{eq:densityheavy} P(t) = X_1(t)X_1(t)^H = X(t)\begin{bmatrix}I_p& 0\\ 0& 0\\ \end{bmatrix}X(t)^H = X(t)h(\Lambda(t))X(t)^H = h(B(t)), \end{equation} where we assume that $t$ lies in a sufficiently small neighbourhood of $t_0=0$ such that $\lambda_p(t)<\lambda_{mid}<\lambda_{p+1}(t)$. Then, \begin{eqnarray}\label{eq:dfrechet} P'(0) =& \lim\limits_{\epsilon\to 0} &\frac{h(B(\epsilon))-h(B(0))}{\epsilon} \nonumber\\ =& \lim\limits_{\epsilon\to 0, \norm{E}\to 0} &\frac{X(0)\Big(h(\Lambda(0)+X(0)^HEX(0))-h(\Lambda(0))\Big)X(0)^H}{\epsilon}, \end{eqnarray} where \begin{equation}\label{eq:eb0} E = \epsilon B_1. \end{equation} If we denote by $L(F,G)$ the Fr\'echet derivative of $h$ evaluated at $F$ and applied to $G$, then from \eqref{eq:dfrechet}, we get \begin{align}\label{eq:dfrechet2} P'(0) &=& \lim\limits_{\epsilon\to 0, \norm{E}\to 0} &\frac{X(0)\Big(L\big(\Lambda(0),X(0)^HEX(0)\big)+O(\norm{E}^2)\Big)X(0)^H}{\epsilon}\nonumber\\ &=& \lim\limits_{\epsilon\to 0, \norm{E}\to 0} &\frac{X(0)L\big(\Lambda(0),X(0)^HEX(0)\big)X(0)^H}{\epsilon}\nonumber\\ &=& &X(0)L\big(\Lambda(0),X(0)^HB_1X(0)\big)X(0)^H& \end{align} The last equation is a consequence of \eqref{eq:eb0} and the fact that $O(\norm{E}^2)/\epsilon$ goes to zero as $\epsilon$ goes to zero and $L$ being linear in the second argument. From the Dalecki{\u i}-Kre{\u i}n theorem\cite[Theorem ~3.11]{HIGHAMFOM},\cite{daleckiikrein}, we have \begin{equation}\label{eq:Lfrechet} L(\Lambda(0),X(0)^HB_1X(0)) = U\circ (X(0)^HB_1X(0)), \end{equation} where $U$ denotes the matrix of divided differences \begin{equation* U := \begin{cases}\frac{h(\lambda_i(0))-h(\lambda_j(0))}{\lambda_i(0)-\lambda_j(0)}&i \neq j\\h'(\lambda_i(0))&i=j.\end{cases} \end{equation*} By definition of $h$ in \eqref{eq:heavyside} and from \eqref{eq:Rdef}, we see that $U$ reduces to $-R$. Vectorizing \eqref{eq:dfrechet2} and using \eqref{eq:Lfrechet} leads us to \begin{align} \vect(P'(0)) &= -\vect\Big(X(0)\big(R\circ (X(0)^HB_1X(0))\big)X(0)^H\Big)\\ \nonumber &= -(\overline{X(0)}\otimes X(0))D(X(0)^T\otimes X(0)^H)\vect(B_1), \end{align} which is a result of repeated application of the matrix product vectorization identity $\vect(LMN) = (N^T\otimes L)\vect(M)$. \end{proof} In practice, the computation of the density matrix as in \eqref{eq:psifp} can be done by using \eqref{eq:densityheavy}. Moreover, in SCF iterations, variations of the problem can be solved, where an approximation of the step function is used. A common choice for this approximation is the Fermi-Dirac distribution: $f_{\mu,\beta}(t) = \frac{1}{1+e^{\beta(t-\mu)}}$, where the parameter $\mu$ is usually selected such that $\trace(P(0)) = p$. The function $f_{\mu,\beta}$ tends to the step function in the limit $\beta \to 0$. Note that Lemma \ref{thm:densitymatder} can be generalized for such an approximation of the density matrix $P_f$ as follows: \begin{equation* \vect(P'_f(0)) = (\overline{X(0)}\otimes X(0))D_f(X(0)^T\otimes X(0)^H)\vect(B'(0)), \end{equation*} where $D_f = \diag(\vect(R_f))$ and \begin{equation* R_f = \begin{cases}\frac{f(\lambda_i(0))-f(\lambda_j(0))}{\lambda_i(0)-\lambda_j(0)}&i \neq j\\f'(\lambda_i(0))&i=j\end{cases}. \end{equation*} \begin{thm}[Density matrix local convergence]\label{thm:densitymat} Let $P_*=X_{1*}X_{1*}^H\in\CC^{n\times n}$ be a fixed point of $\Psi$, i.e., $P_*=\Psi(P_*)$. Then, the SCF iteration satisfies \[ \vech(P_{k+1}-P_*)= \vech(\Psi(P_k)-P_*)=J_{P}\vech(P_k-P_*)+O(\|\vech(P_k-P_*)\|^2) \] where \begin{equation} J_P=-T(\overline{X_*}\otimes X_*) D({X_*}^T\otimes {X_*}^H)L' \label{eq:JPdef} \end{equation} and $L'\in\CC^{n^2\times m}$ is defined by \begin{equation}\label{eq:Lprime} L'= \left(\vect(\mathcal{L}(\vech^{-1}(e_1)),\ldots, \vect(\mathcal{L}(\vech^{-1}(e_{m})))\right), \end{equation} $X_* = [X_{1*},X_{2*}]$ and $e_1,e_2,\ldots,e_m$ are the first $m$ columns of the identity matrix $I_n$. \end{thm} \begin{proof} Applying the operator $\vech(\cdot)$ to our iteration, we get \begin{equation* \begin{aligned} \vech(P_{k+1}) &= \vech(\Psi(P_k))\\ &= \vech(\vect^{-1}(\vect(\Psi(\vech^{-1}(\vech(P_k)))))). \end{aligned} \end{equation*} Hence, the fixed point iteration can be re-written as \begin{equation*} \vech(P_{k+1}) = f(\vech(P_{k})), \end{equation*} where $f: \mathbb{R}^{m} \to \mathbb{R}^{m}$, \begin{equation*} f(v) = \vech(\vect^{-1}(\vect(\Psi(\vech^{-1}(v)))))\quad \forall v \in \mathbb{R}^m. \end{equation*} A Taylor expansion around the fixed-point $\vech(P_*)$ gives us, \begin{equation* \vech(P_{k+1}-P_*) = J_P\vech(P_k-P_*) + O(\|\vech(P_k-P_*)\|^2), \end{equation*} where $J_P$ is the Jacobian of $f$ evaluated in $\vech(P_*)$. The $j$-th column of $J_P$ is given by \begin{equation}\label{eq:jach} \begin{aligned} J_P(:,j) &= \lim_{\epsilon \to 0} \frac{f(\vech(P_*)+\epsilon e_j) - f(\vech(P_*))}{\epsilon}\\ &= \lim_{\epsilon \to 0} \frac{\vech(\vect^{-1}(\vect(\Psi(P_*+\epsilon \vech^{-1}(e_j))-\Psi(P_*))))}{\epsilon}.\\ \end{aligned} \end{equation} By using linearity of the vectorization operators and of $\mathcal{L}$, we can now invoke Lemma~\ref{thm:densitymatder} with $B(\alpha) = A+\mathcal{L}(P_*+\alpha \vech^{-1}(e_j))$ to get \begin{multline*} \lim_{\epsilon \to 0} \frac{\vect(\Psi(P_*+\epsilon \vech^{-1}(e_j))-\Psi(P_*))}{\epsilon} =\\ (\overline{X_*}\otimes X_*)D({X_*}^T\otimes {X_*}^H)\vect(\mathcal{L}(\vech^{-1}(e_j))). \end{multline*} Using this and equation \eqref{eq:jach}, we have \begin{equation} \begin{aligned} J_P(:,j) &= \vech(\vect^{-1}((\overline{X_*}\otimes {X_*})D({X_*}^T\otimes {X_*}^H)\vect(\mathcal{L}(\vech^{-1}(e_j)))))\\ &= T(\overline{X_*}\otimes {X_*})D({X_*}^T\otimes {X_*}^H)\vect(\mathcal{L}(\vech^{-1}(e_j))). \label{eq:JPX} \end{aligned} \end{equation} Due to the fact that $\vect(\mathcal{L}(\vech^{-1}(e_j)))$ is the only component in \eqref{eq:JPX} depending on $j$, we obtain \eqref{eq:JPdef} by factorizing the matrix $T(\overline{X_*}\otimes {X_*})D({X_*}^T\otimes {X_*}^H)$. \end{proof} \newpage \section{Convergence factor bounds and their interpretation} \subsection{Spectral-norm bounds} Since \eqref{eq:psifp} is a nonlinear fixed-point map and the Jacobian is evaluated at a fixed point in Theorem~\ref{thm:densitymat}, the convergence factor is \begin{equation* c=\rho(J_P) \end{equation*} where $\rho(J_P)$ denotes the spectral radius of $J_P$. Moreover, a sufficient and almost necessary condition for local convergence is $c < 1$. Due to the fact that the spectral radius is smaller than any operator norm, we have in particular for the spectral norm: \begin{equation* c\leq \norm{J_P}_2 =:c_2. \end{equation*} \subsubsection{Naive bounds} Now note that $\overline{X_*}\otimes X_*$ and $X_*^T\otimes X_*^H$ are orthogonal matrices, and that $D$ defined by \eqref{eq:Ddef} is a diagonal matrix whose largest element is the reciprocal gap, such that \[ \|D\|_2=\max_{i}|d_{i,i}| =\frac{1}{\delta}. \] By using this and the Cauchy-Schwartz inequality we obtain a straight-forward upper bound for $c$, \begin{equation}\label{eq:mnaive} c\leq \norm{T}_2 \norm{\overline{X}\otimes {X}}_2 \norm{D}_2 \norm{({X}^T\otimes {X}^H)}_2 \norm{L'}_2 \leq \frac{\norm{L'}_2}{\delta} := c_{\rm naive}, \end{equation} where we dropped subscript $*$ in the eigenvector matrix $X_*$ for notational convenience. We can conclude from \eqref{eq:mnaive} that a small gap implies a larger value of the upper bound $c_{naive}$, indicating slow convergence. This is consistent with the well-known fact that problems with a small gap are more difficult to solve using the SCF iteration which is concluded in several convergence analysis works, e.g. \cite{Yang:2009:SCF}. Note that the bound \eqref{eq:mnaive} does not depend on the gap alone but also on $\norm{L'}_2$, which can be large and difficult to analyze. The matrix $L'$ depends on the action of the operator $\mathcal{L}$, which leads us to the pursuit of other bounds which may quantify this dependence in a way that is easier to interpret. \subsubsection{Cycled permutation} Different bounds can be derived by using the fact that the spectral radius does not change when we reverse the order of multiplication of matrices, i.e., $\rho(AB)=\rho(BA)$. Therefore, from the definition of $J_P$ and \[ c=\rho(J_P) = \rho(T(\overline{X}\otimes X)D({X}^T\otimes {X}^H)L'), \] we obtain variants based on cyclic permutation \begin{subequations}\label{eq:cyclic1} \begin{eqnarray} c&=&\rho((\overline{X}\otimes X)D({X}^T\otimes {X}^H)L'T)\\ &=&\rho(D({X}^T\otimes {X}^H)L'T(\overline{X}\otimes X)). \end{eqnarray} \end{subequations} Both equations in \eqref{eq:cyclic1} lead to the bound \begin{equation* c\le \|D({X}^T\otimes {X}^H)L'T\|_2=:c_{2,a}. \end{equation*} The cyclic permutation can be continued such that \begin{subequations}\label{eq:cyclic2} \begin{eqnarray} c&=&\rho(({X}^T\otimes {X}^H)L'T(\overline{X}\otimes X)D)\\ &=&\rho(L'T(\overline{X}\otimes X)D({X}^T\otimes {X}^H)). \end{eqnarray} \end{subequations} Equation \eqref{eq:cyclic2} leads to the bound \begin{equation} \label{eq:c2b} c\le \|L'T(\overline{X}\otimes X)D\|_2=:c_{2,b}. \end{equation} In the following we need the symmetrization operator formally defined as \begin{equation}\label{eq:symmet} S(X):=\sum_{j=1}^m\vech(X)_j\vechinv(e_j) \end{equation} or equivalently $S(L+D+R)=L+D+L^T$, where $L+D+R$ is the decomposition into the lower triangular, diagonal and upper triangular matrices. Using \eqref{eq:symmet}, we have the identity \begin{eqnarray}\label{eq:lpident} L'\vech(X) &=& \sum_{j=1}^m\vect(\LLL(\vechinv(e_j)))\vech(X)_j \\ &=&\vect(\mathcal{L}(S(X))), \nonumber \end{eqnarray} due to the definition of $L'$ in \eqref{eq:Lprime} and the linearity of $\LLL$. The columns of the matrix in \eqref{eq:c2b} can be expressed as \begin{eqnarray*} (L'T(\overline{X}\otimes X)D)_{:,j}&=&L'T\vect(x_\ell x_m^H)d_{j,j}\\ &=&L'\vech(x_\ell x_m^H)d_{j,j} \end{eqnarray*} where $j=n(m-1)+\ell$ and where we used \eqref{eq:vvech} in the last step. Hence, from \eqref{eq:lpident} we have \begin{equation} \label{eq:c2b_columns} (L'T(\overline{X}\otimes X)D)_{:,j}=\vect(\LLL(S(x_\ell x_m^H)))d_{j,j}. \end{equation} The columns of the matrix inside the norm in \eqref{eq:c2b} is given by \eqref{eq:c2b_columns}. This formula can be interpreted as follows. We clearly see that the action of the linear operator applied to the outer products of eigenvectors $\LLL(S(x_\ell x_m^H))$ has significance. The weighting with $d_{j,j}$ implies that only pairs of eigenvectors of different occupancy are relevant (in this bound). This quantity is further described in the following section. \subsection{Higher gaps} We begin by decomposing the matrix obtained by cycled permutation in \eqref{eq:cyclic2} as follows, \begin{multline} \label{eq:Dsep} L'T(\overline{X}\otimes X)D({X}^T\otimes {X}^H)=\\ L'T(\overline{X}\otimes X)(D-D_1-\cdots-D_{2q})({X}^T\otimes {X}^H) +L'T(\overline{X}\otimes X)(D_1+\cdots+D_{2q})({X}^T\otimes {X}^H) \end{multline} where $D_j$ are rank one diagonal matrices (and we take $2q$ terms for symmetry reasons). We have $q=p(n-p)$ unique gaps and they occur twice each on the diagonal of $D$. Hence, this decomposition reveals the dependence of the convergence on eigenvalue gaps other than just the smallest gap $\delta$. In the following theorem we quantify this dependence along with the dependence on the outer products of the eigenvectors $\LLL(S(x_\ell x_m^H))$ as introduced in the previous subsection. In the formulation of the theorem, we use the set $\Omega_q\subset [1,n]\times [1,n]$ which contains the indices of $R$ that have entries corresponding to the $q$ smallest gaps. As a result, $\Omega_q$ contains $2q$ elements. \begin{figure}[h] \begin{center} \scalebox{0.6}{\includegraphics{gap3.eps}} \caption{ Schematic illustration of elements of $\Omega_3$ as indices of $R$ for the real-valued problem in subsection \ref{subsec:dlap} with $n=7, p=3, \alpha = 10.0$. } \label{fig:schema} \end{center} \end{figure} In figure~\ref{fig:schema}, we visualize the elements of $\Omega_3$. The set $\Omega_3$ comprises the indices of the reciprocal gap matrix $R$ which correspond to these gaps, that is, $\Omega_3 = \{(4,3),(3,4),(4,2),(2,4),(5,3),(3,5)\}$. \begin{theorem}[Higher gaps]\label{theo:hogaps} The convergence factor of the SCF-iteration is bounded by \begin{equation} \label{eq:hgaps} \rho(J_p)\le \frac{\|L'\|_2}{\delta_{q+1}}+ \sum_{(\ell,m)\in\Omega_q}\frac{1}{|\lambda_\ell-\lambda_m|}\|\LLL(S(x_\ell x_m^H))\|_F := c_{gap,q} \end{equation} for any $q\in [0,p(n-p)]$, where $\delta_{p(n-p)+1}:=\infty$ and $\delta_1 = \delta$. \end{theorem} \begin{proof} For notational convenience, we express $D$ in terms of $R$ (defined in \eqref{eq:Rdef}), i.e., \begin{eqnarray*} D&=&\diag(\vect(R))=\sum_{\ell,m}^n r_{\ell,m}\vect(e_\ell e_m^T)\vect(e_\ell e_m^T)^T \\ &=&\sum_{(\ell,m)\not\in\Omega_q} r_{\ell,m}\vect(e_\ell e_m^T)\vect(e_\ell e_m^T)^T+ \\ &&\sum_{(\ell,m)\in\Omega_q} r_{\ell,m}\vect(e_\ell e_m^T)\vect(e_\ell e_m^T)^T. \end{eqnarray*} The idea of the proof is to take the last sum in this equation as $\sum_{j=1}^{2q} D_j$ with the decomposition in \eqref{eq:Dsep}. By the triangle inequality, we have \begin{equation} \label{eq:gapbound_first} c \leq \|L'T(\overline{X}\otimes X)(D-D_1-\cdots-D_q)\| +\sum_{j=1}^{2q}\|L'T(\overline{X}\otimes X)D_j\|. \end{equation} The first term in \eqref{eq:gapbound_first} is of the form used in the naive bound \eqref{eq:mnaive}, except that the diagonal matrix is modified by setting the contribution corresponding to the $q$ first gaps to zero. We obtain directly the first term in \eqref{eq:hgaps}, \[ \|L'T(\overline{X}\otimes X)(D-D_1-\cdots-D_{2q})\|\le \frac{\|L'\|_2}{\delta_{q+1}}. \] The second term in \eqref{eq:gapbound_first} is \[ \sum_{j=1}^{2q}\|L'T(\overline{X}\otimes X)D_j\|= \sum_{(\ell,m)\in\Omega_q} r_{\ell,m}\|L'T(\overline{X}\otimes X)\vect(e_\ell e_m^T)\vect(e_\ell e_m^T)^T\|. \] This can be simplified with the identity \eqref{eq:c2b_columns} which implies that \begin{subequations}\label{eq:gapbound_second} \begin{eqnarray} \|L'T(\overline{X}\otimes X)\vect(e_\ell e_m^T)\vect(e_\ell e_m^T)^T\|&=&\|\vect(\LLL(S(x_\ell x_m^H)))\vect(e_\ell e_m^T)^T\| \\ &=&\|\vect(\LLL(S(x_\ell x_m^H)))\| \\ &=&\|\LLL(S(x_\ell x_m^H))\|_F. \end{eqnarray} \end{subequations} The last two equalities follow from the fact that the spectral norm of a matrix with one non-zero column, is the two-norm of that column vector, which is the Frobenius norm of the corresponding matrix. The proof is concluded by combining \eqref{eq:gapbound_first} with \eqref{eq:gapbound_second} and noting that $r_{\ell,m}=\frac{1}{|\lambda_\ell-\lambda_m|}$. \end{proof} Theorem \ref{theo:hogaps} should be further interpreted as follows. The parameter $q$ is free and the theorem therefore provides us with a family of bounds parameterized by $q$. For example, $q=0$ gives us $\rho(J_P) \leq \frac{\norm{L'}_2}{\delta_1}$, which is the naive bound from \eqref{eq:mnaive}. For $q = 1$, we have \begin{equation* \rho(J_p) \leq \frac{\norm{L'}_2}{\delta_2}+\frac{\norm{\mathcal{L}(S(x_px_{p+1}^H))}_F+\norm{\mathcal{L}(S(x_{p+1}x_p^H))}_F}{\delta_1}. \end{equation*} By induction, $q=k$ gives us a bound that is a function of the $k+1$ smallest gaps and the norm of the action of $\mathcal{L}$ on the outer products of eigenvector-pairs corresponding to the gap indices of the $k$ smallest gaps. \subsection{Illustrative example}\label{sec:acadexample} In order to illustrate the insight provided by Theorem~\ref{theo:hogaps}, we provide an example showing how in certain situations, using a bound with a higher value of $q$ provides us tighter upper bounds. Consider the following problem parameterized by $\epsilon$. \begin{equation}\label{eq:acadproblem} \left(\begin{bmatrix}0 & \epsilon & 0\\ \epsilon & 1+d & \epsilon\\ 0 & \epsilon & 10\\\end{bmatrix} + \begin{bmatrix}1 & 0 & 0\\ 0 & 1 & 0\\ 0 & 0 & 100\\\end{bmatrix} \circ x_1x_1^H\right)X = X\Lambda. \end{equation} Let $d=0.16$. In the notation in equation \eqref{eq:ap}, we have, \[ A_0 = \begin{bmatrix}0 & \epsilon & 0\\ \epsilon & 1+d & \epsilon\\ 0 & \epsilon & 10\\\end{bmatrix}\quad \LLL(P) = L\circ P = \begin{bmatrix}1 & 0 & 0\\ 0 & 1 & 0\\ 0 & 0 & 100\\\end{bmatrix} \circ P . \] First note that for $\epsilon = 0$, the solution is given by \[ \Lambda(0) = \begin{bmatrix}\lambda_1& 0& 0\\0& \lambda_2& 0\\0& 0& \lambda_3\\\end{bmatrix} =\begin{bmatrix}1 & 0 & 0\\ 0 & 1+d & 0\\ 0 & 0 & 10\\\end{bmatrix},\quad X(0) = \begin{bmatrix}1 & 0 & 0\\ 0 &1 & 0\\ 0 & 0 & 1\\\end{bmatrix}. \ \begin{figure}[h!] \centering \scalebox{1}{\includegraphics{conveps.eps}} \caption{Convergence factor and bounds for the illustrative example} \label{fig:acadbounds} \end{figure \begin{figure}[h!] \centering \scalebox{1}{\includegraphics{action.eps}} \caption{Norm of $\mathcal{L}(x_1x_2^H)$ and $\mathcal{L}(x_2x_3^H)$} \label{fig:actionofl} \end{figure \begin{figure}[h!] \centering \scalebox{1}{\includegraphics{gapseps.eps}} \caption{Variance of $\delta_1$ and $\delta_2$ with $\epsilon$} \label{fig:gaps} \end{figure Varying $\epsilon$, solving the resulting problem instances and plotting the convergence factor and its bounds gives us figure~\ref{fig:acadbounds}. In figure~\ref{fig:actionofl}, we plot the norm of $L'$ along with the norm of action of $\mathcal{L}$ on the outer products of the eigenvectors ($x_1x_2^H$ and $ x_1x_3^H$) with $\epsilon$. In figure~\ref{fig:gaps}, we visualize how the gaps $\delta_1$ and $\delta_2$ vary with $\epsilon$. Note that we have $\norm{L'}_2 = 100$, which is independent of $\epsilon$. As can be seen from figures ~\ref{fig:acadbounds} and ~\ref{fig:gaps}, $c_{naive}$ shows a direct inverse dependence on $\delta_1$, which is expected from \eqref{eq:mnaive} because $\norm{L'}$ is constant. We can see from figures \ref{fig:acadbounds} and \ref{fig:actionofl} that $c_{naive}$ is not a good approximation to the convergence factor $c$, in comparison to our bounds $c_{gap,2}$, $c_2$ and $c_{2,a}$. The increase in $c$ as we increase $\epsilon$ (which means slower convergence for larger values of $\epsilon$) is not captured by $c_{naive}$ or $c_{gap,1}$ as they are essentially constant for small $\epsilon$. However, from figure~\ref{fig:acadbounds}, we see that the increase in $c$ coincides with the increase of the norm of action of $\LLL$ on the outer products of the eigenvectors ($x_1x_2^H$ and $ x_1x_3^H$), as seen in figure~\ref{fig:actionofl}. This behaviour is better captured in the formula for the upper bound $c_{gap,2}$, \begin{align*} c_{gap,2} &= \frac{\norm{\mathcal{L}(S(x_1x_{2}^H))}_F+\norm{\mathcal{L}(S(x_{2}x_1^H))}_F}{\delta_1}\\ &+\frac{\norm{\mathcal{L}(S(x_1x_{3}^H))}_F+\norm{\mathcal{L}(S(x_{3}x_1^H))}_F}{\delta_2}. \end{align*} Although the bounds $c_{gap,2}$ , $c_2$ and $c_{2,a}$ are better approximations of $c$ as compared to $c_{naive}$, there is still a discrepancy in the slopes in figure~\ref{fig:acadbounds}, and the rate of increase of $c$ is faster than that of the bounds. We now provide a more detailed analysis. First note that differentiating \eqref{eq:acadproblem} with respect to $\epsilon$, and setting $\epsilon = 0$, we obtain \begin{equation}\label{eq:xp0} X'(0) = \begin{bmatrix}0& \frac{1}{\lambda_2-\lambda_1}& 0\\ \frac{1}{\lambda_1-\lambda_2}& 0& \frac{1}{\lambda_3-\lambda_2}\\0& \frac{1}{\lambda_2-\lambda_3}& 0\end{bmatrix},\quad \Lambda'(0) = 0,\quad D'(0) = 0. \end{equation} Let $J(\epsilon)$ denote the parameter dependent Jacobian evaluated at the solution, \begin{equation*} J(\epsilon) = -T\left(X(\epsilon)\otimes X(\epsilon)\right)D(\epsilon)\left(X(\epsilon)^H\otimes X(\epsilon)^H\right)L'. \end{equation*} Differentiating w.r.t $\epsilon$, setting $\epsilon = 0$, and using $X(0) = I$, \begin{equation}\label{eq:jprime0} \begin{split} J'(0) = -T\bigg(&\left(X'(0)\otimes I+I\otimes X'(0)\right)D(0)+D'(0)\\ &- D(0)\left(X'(0) \otimes I+I\otimes X'(0)\right)\bigg)L'.\\ \end{split} \end{equation} Using the formulae from \eqref{eq:xp0} and substituting into \eqref{eq:jprime0}, we get, \begin{equation* J'(0) = \frac{1}{{(\lambda_2-\lambda_1)}^2}\begin{bmatrix} -e_2& 0& 0& e_2& 0& 0 \end{bmatrix}. \end{equation*} From the structure of $J'(0)$ and the fact that all eigenvalues of $J'(0)$ are zero, we have, \begin{eqnarray}\label{eq:rhojp0} \rho(J'(0)) &=& 0,\\ \norm{J'(0)}_2 &=& \frac{1}{{(\lambda_2-\lambda_1)}^2}\quad. \nonumber \end{eqnarray} This allows us to carry out a Taylor series analysis for $\rho\left(J(\epsilon)\right)$ and $\norm{J(\epsilon)}_2$ around 0, \begin{equation}\label{eq:rhoj} c = \rho(J(\epsilon)) = \rho(J(0)+\epsilon J'(0)+\mathcal{O}(\epsilon^2)) \approx \epsilon\rho(J'(0))+\mathcal{O}(\epsilon^2) = \mathcal{O}(\epsilon^2). \end{equation} Similarly, \begin{equation}\label{eq:normj} c_2 = \norm{J(\epsilon)}_2 = \epsilon\norm{J'(0)}_2+\mathcal{O}(\epsilon^2) = \frac{\epsilon}{{(\lambda_2-\lambda_1)}^2}+\mathcal{O}(\epsilon^2) = \mathcal{O}(\epsilon). \end{equation} Hence, from \eqref{eq:rhoj} and \eqref{eq:normj}, we expect $c$ to vary at a rate that is an order of magnitude faster than $c_2$ for very small values of $\epsilon$, which is exactly what we observe in figure~\ref{fig:acadbounds}. We clearly see from \eqref{eq:rhojp0},\eqref{eq:rhoj} and \eqref{eq:normj} that this is because $J'(0)$ has zero eigenvalues (and hence zero spectral radius) but non-zero norm. This illustrates how the two-norm based bounds can overestimate $c$. \section{Numerical examples} \subsection{Discrete Laplacian example}\label{subsec:dlap} In this subsection, we apply our theory to a minor variation of the problem type discussed in \cite[Section~5]{liu2014convergence}. In the context of this paper, this translates to \[ A_0 = \begin{bmatrix} 2/h^2& -1/h^2+i/2h& 0& \ldots & 0\\ -1/h^2-i/2h& 2/h^2& -1/h^2+i/2h& \ddots& \vdots \\ 0& -1/h^2-i/2h& 2/h^2& \ddots& 0\\ \vdots& \ddots& \ddots& \ddots& -1+i/2h\\ 0& \ldots& 0& -1/h^2-i/2h& 2/h^2\\ \end{bmatrix}, \] which is the discretized 1D differential operator: $\dfrac{\partial^2}{\partial x^2}+i\dfrac{\partial}{\partial x}$. In a PDE setting, this would correspond to a diffusion term added with a complex convection term discretized with a central difference scheme with grid spacing $h$. We also have \[ \mathcal{L}(P) = \alpha Diag\left(Re(A_0)^{-1}diag\left(P\right)\right). \] Note that $\mathcal{L}(\cdot)$ depends only on the diagonal of $P$. \begin{figure} \subfigure[Convergence factor and bounds]{{\includegraphics[height=5.cm,width=6.2cm]{complexconv.eps}}} \subfigure[Distribution of eigenvalues]{{\includegraphics[height=5.cm,width=6.2cm]{complexeigs1.eps}}} \subfigure[Variance with $n$]{{\includegraphics[height=5.cm,width=6.2cm]{nvsconv.eps}}} \subfigure[Variance with $\alpha$]{{\includegraphics[height=5.cm,width=6.2cm]{alphavsconv.eps}}} \caption{Complex-valued problem for $n=30$\big((a),(b) and (d)\big), $p= 15, \alpha=40.0$\big((a),(b) and (c)\big)} \label{fig:cproblem} \end{figure} Figure~\ref{fig:cproblem}(a) shows us that the predicted convergence rate $c$ agrees perfectly with SCF convergence history. The norm based bounds $c_2$ and $c_{naive}$ are slightly worse than the exact rate $c$. As expected, $c_{naive}$ is the least accurate upper bound, but $c_{gap,2}$ is only slightly better. As seen from figure~\ref{fig:cproblem}(b) the gaps between the eigenvalues are not very well separated, that is, the higher gaps are not much larger than $\delta$. More precisely, $\delta_3$ is not much larger than $\delta_1$ and $\delta_2$. This explains why $c_{gap,2}$ is not a good approximation to the norm based bounds for this problem. Figure~\ref{fig:cproblem}(d) shows a linear increase in the value of the convergence factor and the upper bounds with $\alpha$, which is expected from the linear dependence of $L'$ on $\alpha$ and equation \eqref{eq:cyclic1}. From figure~\ref{fig:cproblem}(c), we also see that convergence becomes faster(that is $c$ decreases) with increase in problem size for a constant value of $\alpha$ and $p$. To make a comparison of our upper bounds with the convergence factor derived from \cite[Theorem~4.2]{liu2014convergence}, we change the problem by setting \[ A_0 = \begin{bmatrix} 2/h^2& -1/h^2& 0& \ldots & 0\\ -1/h^2& 2/h^2& -1/h^2& \ddots& \vdots \\ 0& -1/h^2& 2/h^2& \ddots& 0\\ \vdots& \ddots& \ddots& \ddots& -1/h^2\\ 0& \ldots& 0& -1/h^2& 2/h^2\\ \end{bmatrix} \] since the analysis in that paper is presented for real-valued problems. The operator $\mathcal{L}(\cdot)$ is the same as before. We plot the upper bound $c_{Liu} = \frac{2\alpha\sqrt{n}\norm{A_0^{-1}}_2}{\delta_1}$ along with the other upper bounds for this modified real-valued problem in figure~\ref{fig:realconv}. \begin{figure}[h!] \centering \scalebox{0.7}{\includegraphics{realconv.eps}} \caption{Real-valued problem for $n = 60$, $\alpha = 5.0$, $p = 25$} \label{fig:realconv} \end{figure} Figure~\ref{fig:realconv} suggests to us that the upper bounds discussed in this paper are improvements over the upper bound in \cite{liu2014convergence}. \subsection{Water molecule example}\label{subsec:watermolecule} In this example, we apply the SCF iteration to a problem that originates from the modelling of a water molecule system. The discretization involves a restricted Hartree-Fock approximation with a set of $n=13$ basis functions and $p=5$. For any nonlinear eigenvalue problem that results from a Hartree-Fock approximation, we have \[ A(X_1X_1^H) = H_{core}+2G(R^{-1}X_1X_1^H{R^{-1}}^H) \] where $G(\cdot)$ is a linear operator and $H_{core}$ is a sum of two matrices that correspond to terms for kinetic energy and the nuclear-electron interaction energy. In our context, $\mathcal{L}(P) = 2G(R^{-1}P{R^{-1}}^H)$. Here, $R$ is a lower triangular matrix that results from a cholesky decomposition of the "overlap matrix". The overlap matrix is hermitian and obtained by computing integrals of products of basis functions, as explained in \cite[Section~2.4]{Rudberg}. For the purpose of reproducibility, we provide the coordinates of the nuclei of the Oxygen and Hydrogen atoms in the following table. Note that all data is in atomic units. \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline Atom& Charge($\mathbf{e}$)& x($\mathbf{a_0}$)& y($\mathbf{a_0}$)& z($\mathbf{a_0}$)\\ \hline O& 8.0& 0.0& 0.0& 0.0\\ \hline H& 1.0& -1.809& 0.0& 0.0\\ \hline H& 1.0& 0.453549& 1.751221& 0.0\\ \hline \end{tabular} \end{center} The computation was performed using Ergo\cite{ERGO-JCTC-2011,ERGO-SOFTWAREX}, which is a software package for large-scale SCF calculations. The standard Gaussian basis set 3-21G was used and the starting guess for the density matrix was projected from the calculation using a smaller STO-3G basis set. We plot the SCF convergence history and the exact convergence factor $c$. We have not plotted the other upper bounds that we derived because they overestimate $c$ by a large margin. Instead, based on the theory in Theorem~\ref{theo:hogaps} we investigate the bound that neglects certain terms such that \[ \widetilde{c_{2}} = \norm{T(\overline{X}\otimes X)D_1({X}^T\otimes {X}^H)L'}_2 \] which is the spectral norm of a 2-rank approximation of the Jacobian $J_P$ (taking into account only the entries that contain $\delta_1$ in $D$). As we can see from figure ~\ref{fig:watemolecule}, the observed behaviour of SCF convergence agrees with that predicted by the exact value of the spectral radius, $c$. \begin{figure}[h!] \subfigure[Convergence factor and bounds]{{\includegraphics[height=5.cm,width=6.2cm]{water.eps}}} \subfigure[Distribution of eigenvalues]{{\includegraphics[height=5.cm,width=6.2cm]{water_eigs.eps}}} \caption{Water molecule problem with $n=13, p=5$} \label{fig:watemolecule} \end{figure} \section{Conclusions and outlook} The SCF algorithm is an important algorithm in many fields. We have provided a new convergence characterization for the algorithm using a density matrix based analysis of a fixed point map. The upper bounds derived for the spectral radius of the Jacobian of the fixed point map illustrate how the convergence depends on the different problem parameters and physical properties. In particular, Theorem~\ref{theo:hogaps} provides a mathematical footing for studying how the gaps interact with the outer products of eigenvectors to affect the convergence properties. This is a quantification of Stanton's observation in \cite[Section~IV]{Stanton:1981:CLOSEDSHELL}, where he points out that typically, divergence in SCF calculations is not due to a single very low energy excitation, but to the interaction of several moderately low excitations. The discussion of the illustrative example in section \ref{sec:acadexample} explains how when the Hessian has zero spectral radius but non-zero norm, an upper bound based on the interaction of higher gaps is needed to give a more accurate picture of the convergence behaviour. Finally, the application of our upper bounds to practical problems in subsections ~\ref{subsec:dlap} and ~\ref{subsec:watermolecule} reveal that our bounds are slightly better approximations to the convergence factor than the bounds that exist in previous literature.

2c682c7bded8d79c2e3ee3660b330b12c137412c

\section{Introduction} Advances in deep learning have mobilized the research community to adopt deep reinforcement learning (RL) agents for challenging control problems, typically in complex environments with raw sensory state-spaces. Breakthroughs by~\cite{mnih2015human} show RL-agents can reach above-human performance in Atari 2600 games, and AlphaGo Zero~\cite{silver2017mastering} becomes the world champions on the game of \textit{Go}. Still, training RL agents is non-trivial. Off-policy methods typically require days of training to obtain competitive performance, while on-policy methods could be trapped in local minima. Recent techniques featuring on-policy learning~\cite{mnih2016asynchronous,schulman2017proximal,wu2017scalable} have shown promising results in stabilizing the learning processes, enabling an agent to solve a variety of tasks in much less time. In particular, the state-of-the-art on-policy ACKTR agent by~\cite{wu2017scalable} shows improved sample efficiency with the help of Kronecker-factored (K-Fac) approximate curvature for natural gradient updates, resulting in stable and effective model updates towards a more promising direction. However, tasks with sparse rewards remain challenging to on-policy methods. An agent could take massive amount of exploration before reaching non-zero rewards; and as the agent learns on-policy, the sparseness of reward feedback (receiving all-zero rewards from most actions performed by the agent) could be malicious and render an agent to falsely predict all states in an environment towards a value of zero. As there does not exist a universal criterion for measuring ``task sparseness'', we show an ad-hoc metric in Figure~\ref{fig:acktr1} in an attempt to provide intuition. For example, we observe that the ACKTR agent is unable to receive sufficient non-zero immediate rewards that can provide instructive agent updates in Atari games ``Freeway'' and ``Enduro'', resulting in failures when solving these two games. Moreover, if ACKTR gets drawn to and trapped in unfavorable states (as in games like Boxing and WizardOfWor), it could again take long hours of exploration before the agent can get out of the local minima. Such evidence shows that on-policy agent could indeed suffer from the insufficiencies of guidance provided by the exclusive immediate reward signals from the environment. In this paper, we introduce an effective auxiliary reward signal in tasks with sparse rewards to remedy the deficiencies of learning purely from standard immediate reward feedbacks. As on-policy agents may take many explorations before reaching non-zero immediate rewards, we argue that we can leverage past reward statistics to provide more instructive feedbacks to agents in the same environment. To this end, we propose to characterize the past reward statistics in order to gauge the ``long-term'' performance of an agent (detailed in Section~\ref{sec:vwr}). After performing an action, an agent will receive a long-term reward signal describing its past performance upon this step, as well as the conventional immediate reward from the environment. To effectively characterize the long-term performance of the agent, we take into considerations both the crude amount of rewards and the volatility of rewards received in the past, where highly volatile distributions of long-term rewards are explicitly penalized. This enables complex value functions to be more easily approximated in multi-critics supervision. We find in practice that by explicitly penalizing highly volatile long-term rewards while maximizing the expectation of short-term rewards, the learning process and the overall performance are improved regarding both sample efficiency and final rewards. We further propose a ``hot-wiring'' exploration mechanism that can boost seemingly trapped agent in the earlier stage of learning. By leveraging the characterization of long/short-term reward statistics, our proposed advantage actor multi-critic model (A2MC) shows significantly improved performance on the Atari 2600 games and the MuJoCo tasks as compared to the state-of-the-art on-policy methods. \begin{figure}[t] \begin{center} \includegraphics[width=\linewidth]{Figures/Best.pdf} \vskip -0.3 cm \caption{Performance of A2MC on Atari games trained with 15 million timesteps. Our method has a winning rate of $55.3\%$ among all the Atari games tested, as compared to the ACKTR. Our A2MC learns quickly in some of the hardest games for on-policy methods, such as ``Boxing'', ``Enduro'', ``Freeway'', ``Robotank'' and ``WizardOfWor''. The sparseness of a game is defined as the sparseness of average rewards $\mathbf{x}$ obtained by ACKTR within the first $n=10^6$ timesteps by $\varphi (\mathbf{x})= \left( \sqrt{n} - \frac{\left \| \mathbf{x} \right \|_{1}}{\left \| \mathbf{x} \right \|_{2}}\right) / (\sqrt{n}-1)$. A higher value of sparseness indicates sparser rewards. A relative performance margin (in terms of final reward) larger than $10\%$ is deemed as winning / losing. The shaded region denotes the standard deviation over 2 random seeds.} \label{fig:acktr1} \end{center} \end{figure} \section{Preliminary} Consider the standard reinforcement learning setting where an agent interacts with an environment over a number of discrete time step. At each time step $t$, the agent receives an environment state $s_t$, then executes an action $a_t$ based on policy $\pi_t$. The environment produces reward $r_t$ and next state $s_{t+1}$, according to which the agent gets feedback of its immediate action and will decide its next action $a_{t+1}$. The process <$\mathbf{S}, \mathbf{A}, \mathbf{R}, \mathbf{S}$>, typically considered as a Markov Decision Process, continues until a terminal state $s_{T}$ upon which the environment resets itself and produces a new episode. Under conventional settings, the return is calculated as the discounted summation of rewards $r_{t}$ accumulated from time step $t$ onwards $R_t = \sum_{k=0}^{\infty}\gamma^{k}r_{t+k} $. The goal of the agent is to maximize the expected return from each state $s_{t}$ while following policy $\pi$. Each policy $\pi$ has a corresponding action-value function defined as $ Q^{\pi}(s, a) = \mathbb{E}[R_{t}|s_t=s, a_t=a; \pi] $. Similarly, each state $s \in S$ under policy $\pi$ has a value function defined as: $ V^{\pi}(s) = \mathbb{E}[R_t|s_t=s] $. In \textbf{value-based approaches} (e.g., Q-learning~\cite{mnih2015human}), function approximator $Q(s, a; \theta)$ can be used to approximate the optimal action value function $Q^*(s, a)$. This is conventionally learned by iteratively minimizing the below loss function: \begin{equation} \label{eq:dqn} L(\theta) = \mathbb{E}[(y^{target}_t - Q(s_t,a_t; \theta))^2], \\ \end{equation} where $y^{target}_{t} = r_t + \gamma \max_{a'} Q(s_{t+1}, a'; \theta)$ and $s_{t+1}$ is the next state following state $s_t$. In \textbf{policy-based approaches} (e.g., policy gradient methods), the optimal policy $\pi^{*}(a|s)$ is approximated using the approximator $\pi(a|s; \theta)$. The policy approximator is then learned by gradient ascent on $\nabla_{\theta}\mathbb{E}[R_t] \approx \nabla_{\theta}\log\pi(a_t | s_t; \theta)R_t $. The REINFORCE method~\cite{williams1992simple} further incorporates a baseline $b({s_t})$ to reduce the variance of the gradient estimator: $ \nabla_{\theta}\mathbb{E}[R_t]_{REINFORCE} \approx \nabla_{\theta}\log\pi(a_t | s_t; \theta)(R_t - b({s_t})) $ In \textbf{actor-critic based approaches}, the variance reduction further evolves into the advantage function $ A(s_t, a_t) = Q(s_t, a_t) - V(s_t) $ in~\cite{mnih2016asynchronous}, where the action value $Q^{\pi}(s_t, a_t)$ is approximated by $R_t$ and $b(s_t)$ is replaced by $V^{\pi}(s_t)$, deriving the advantage actor-critic architecture where actor-head $\pi(\cdot|s)$ and the critic-head $V(s)$ share low-level features. The gradient update rule w.r.t. the action-head is $\nabla_{\theta}\log\pi(a_t | s_t; \theta)(R_{t} - V(s_{t}; \theta))$. The gradient update w.r.t. the critic-head is: $\nabla_{\theta} (R_{t} - V(s_t; \theta))^2$, where $R_{t} = r_{t} + \gamma V(s_{t+1})$. \section{Characterization of Past Reward Statistics} \label{sec:vwr} The conventional reward $r_{t}$ received from the environment at time step $t$ after an action $a_{t}$ is performed represents the immediate reward regarding this particular action. This ``immediacy'' could be interpreted as a \textit{short-term} horizon of how the agent is doing, \textit{i.e., } evaluating the agent via judging its actions by immediate rewards. We argue that the deficiencies of learning solely from immediate rewards mainly come from this limitation that the agent is learning from one single type of exclusive short-term feedback. As the goal of providing reward feedback to an agent is to inform the agent of its performance, we seek to find an auxiliary performance metric that can measure whether the agent is performing \textit{consistently} well. Inspired by the formulation of \textit{Sharpe Ratio}~\cite{sharpe1994sharpe} in evaluating the long-term performance of fund performance and trading strategies, an effective characterization of historical reward statistics should take into account at least two factors, namely 1) how high the immediate reward is and 2) how varied past rewards were. \subsection{Variability-Weighted Reward} To this end, we follow insight behind~\cite{dowd2000adjusting} and define a variability-weighted characterization of rewards in the past. This is illustrated in Figure~\ref{fig:vwr}. In particular, we consider a historical sequence of $T$ rewards upon timestep $t$ (looking backward $T-1$ timesteps): $ \vec{\mathbf{r}} = [r_{t-(T-1)}..., r_{t-2}, r_{t-1}, r_{t}] $. In order to evaluate how high and varied the reward sequence is, a few steps of pre-processing $\mathcal{G}$ is applied, denoted as $\vec{\mathcal{R}} = \mathcal{G}(\vec{\mathbf{r}})$. In particular, we first derive the reward change at each timestep by extracting the first-order difference with $d_n = r_n - r_{n-1}$: \begin{equation} \label{eq:diff} \begin{split} \vec{\mathbf{d}} = [d_{t-(T-1)}, d_{t-(T-2)}, \ldots, d_{t}] = [r_{t-(T-1)}, r_{t-(T-2)} - r_{t-(T-1)}, \ldots, r_{t}-r_{t-1}]. \end{split} \end{equation} Then we re-order the sequence by flipping \footnote{By flipping, we further encourage recent stable rewards and penalize the volatility of recent past rewards. A concrete example is given in the Appendix A.} with $f_n = d_{t+1-n}$: \begin{equation} \label{eq:diff} \begin{split} \vec{\mathbf{f}} = [f_1, f_2, \ldots, f_{T}] = [d_{t}, d_{t-1}, \ldots, d_{t-(T-1)}]. \end{split} \end{equation} Next we append $f_{0} = 1$ to the head of sequence $\vec{\mathbf{f}}$ and take the normalized cumulative sum to obtain post-processed reward sequence $\vec{\mathcal{R}}$ as: \begin{equation} \begin{split} \vec{\mathcal{R}} = [\mathcal{R}_0, \mathcal{R}_1, \ldots, \mathcal{R}_{T}] = \frac{1}{T+1}[f_0, f_0+f_1, \ldots, \sum_{i=0}^{T}f_{i}]. \end{split} \end{equation} Under such processing, $\vec{\mathcal{R}}$ is a reward sequence with $\mathcal{R}_T - \mathcal{R}_0 = \frac{1}{T+1}r_t$, and $\mathcal{R}_n - \mathcal{R}_{n-1} = \frac{1}{T+1}(r_{t+1-n} - r_{t-n})$. Therefore, the difference between $\mathcal{R}_T$ and $\mathcal{R}_0$ represents the immediate reward and the whole sequence $\vec{\mathcal{R}}$ reflects the volatility of past rewards. In Figure~\ref{fig:vwr}, three examples of processed sequence are presented in the second row with the corresponding raw rewards shown in the first row. We account for \textit{how high the immediate reward is} by defining the average log total return as: \begin{equation} \begin{split} \mathcal{R}_{H} = 100\times(e^{{\frac{1}{T} \ln \frac{\mathcal{R}_{T}}{\mathcal{R}_0}}}-1). \end{split} \end{equation} To account for \textit{how varied past rewards were}, we first define a smooth \textit{zero-variability reference} as: $ \vec{\mathcal{R}}^{zero} = [\mathcal{R}^{zero}_0, \mathcal{R}^{zero}_1, \ldots, \mathcal{R}^{zero}_T] = \mathcal{R}_{0}[e^{0\times\widetilde{\mathcal{R}}}, e^{1\times\widetilde{\mathcal{R}}}, \ldots, e^{T\widetilde{\mathcal{R}}}] $ with $\widetilde{\mathcal{R}} = \frac{1}{T} \ln \frac{\mathcal{R}_{T}}{\mathcal{R}_0}$, representing a smooth monotonic reference sequence from $\mathcal{R}_0$ to $\mathcal{R}_T$. Then we define the reward differential $\delta_{\mathcal{R}}$ as the differential reward versus its zero-variability reference as $ \delta_{\mathcal{R}}(n) = \frac{\mathcal{R}_n - \mathcal{R}^{zero}_n}{\mathcal{R}^{zero}_n} $, whose statistics are sketched in the third row of Figure~\ref{fig:vwr}. With maximally allowed volatility as $\sigma_{max}$, the variability weights can be defined as: \begin{equation} \omega = 1-[\frac{\sigma(\delta_{\mathcal{R}})}{\sigma_{max}}]^{\tau}, \end{equation} where $\sigma(\cdot)$ is the standard deviation and $\tau$ controls the rate to penalize highly volatile reward distribution. Finally we can derive the variability-weighted past reward indicator \ $r^{vwr}$ for the characterization of past reward statistics: \begin{equation} r^{vwr} = \left\{\begin{matrix} \mathcal{R}_{H}(1-[\frac{\sigma(\delta_{\mathcal{R}})}{\sigma_{max}}]^{\tau}) & \text{if} \ \sigma(\delta_{\mathcal{R}}) < \sigma_{max} \text{, } \mathcal{R}_T > 0 \\ 0 & \text{otherwise} \end{matrix}\right. \end{equation} Example computed values of $r^{vwr}$ for the characterization of different reward statistics are shown in Figure~\ref{fig:vwr}. \\ \subsection{Multi-Critic Architecture} A higher value of $r^{vwr}$ indicates better agent performance as the result of the historical sequence of actions. The same set of optimization procedures for conventional value function (\textit{i.e.,} via maximization of immediate reward signal $r$) update can be applied accordingly. The actual returns computed from both the ``long-term'' and ``short-term'' rewards are discounted by the same factor $\gamma$. In particular, for $N$-step look-ahead approaches, we have: \begin{equation} R_{t}^{\text{short-term}} = \sum_{n=0}^{N-1} \gamma^{n}{r_{t+n}} + \gamma^{N}V(s_{t+N}), \end{equation} \begin{equation} R_{t}^{\text{long-term}} = \sum_{n=0}^{N-1} \gamma^{n}{r_{t+n}^{vwr}} + \gamma^{N}V^{vwr}(s_{t+N}). \end{equation} Similar to the standard state value function $V(s)$, we further define $V^{vwr}(s)$ as the value function w.r.t the variability-weighted reward $r^{vwr}$. These value functions form \textit{multiple} critics judging a given state $s$. The gradients w.r.t. the critics then become: \begin{equation} \begin{split} \nabla_{\theta^{\text{short-term}}} [(R^{\text{short-term}}_{t} - V(s_t; \theta^{\text{short-term}}))^2] + \nabla_{\theta^{\text{long-term}}} [(R^{\text{long-term}}_{t} - V^{vwr}(s_t; \theta^{\text{long-term}}))^2]. \end{split} \end{equation} We show the effectiveness of the proposed characterization of past reward statistics in advantage actor-critic frameworks. The two different value functions can share the same low-level feature representation, enabling a single agent to learn multiple critics as parameterized by $\theta^{j}, j \in \{\text{short-term}, \text{long-term}\}$. \section{Hot-Wire $\epsilon$-Exploration} Being handed a game-stick, a human most likely would try out all the available buttons on it to see which particular button entails whatever actions on the game screen, hence receiving useful feedbacks. Inspired by this, we propose to hot-wire the agent to perform an identical sequence of randomly chosen actions in the N-step look-ahead during the initial stage (randomly pressing down a game-stick button for a while): \begin{equation} a_{t+k}= \left\{\begin{matrix} \text{a random action identical for all k} & \text{with prob } \ \epsilon \\ \pi(a_{t+k} | s_{t+k}) \ \ \text{for} \ k = 0,1,2, ..., N-1 & \text{with prob } \ 1-\epsilon \end{matrix}\right. \end{equation} We show that by enabling the ``hot-wiring'' mechanism\footnote{hot-wire is triggered only when the agent is unable to receive meaningful rewards in an initial learning stage. The legend ``vwr + hotWire'' in Fig. 3 indicates that the mechanism is ``enabled'' but not ``enforced''.}, a seemingly trapped agent can be boosted to learn to quickly solve problems where rewards can only be triggered by particular action sequences, as shown in games like ``Robotank'' and ``WizardOfWor'' in Figure~\ref{fig:good}. The full algorithm, advantage actor multi-critic learning (A2MC), is shown in Algorithm~1 in the Appendix E. \section{Related Work} The family of off-policy methods~\cite{wang2015dueling} may be less prone to failure in tasks with sparse rewards at the cost of large amount of explorations before performing agent updates. To tackle the challenge in tasks with rarely observed rewards, pseudo-rewards maximization is adopted in earlier works~\cite{konidaris2009skill,silver2012compositional}. Auxiliary vision tasks (e.g., learning pixel changes or network features) are adopted in the off-policy \textit{UNREAL} agent~\cite{jaderberg2016reinforcement} in order to facilitate learning better feature representations, particularly for sparse reward environments. Another direction of effort aims to design a better reward function for improving sample efficiency via experience replay. \cite{andrychowicz2017hindsight} enhances off-policy learning by re-producing informative reward in hindsight for sequences of actions that do not lead to success previously. The HRA approach~\cite{van2017hybrid} exploits domain knowledge to define a set of environment-specific rewards based on reward categories. In contrast to heuristically defining vision-related auxiliary tasks, our proposed on-policy A2MC agent learns from the characterization of intrinsic past reward statistics obtainable from any environment; and different from the hybrid architecture tailored specific to Ms. Pacman, our A2MC agent can generalize well to a variety of tasks without the need to engineer a decomposition of problem-specific environment rewards. The \textit{multi-agent} approaches~\cite{lanctot2017unified, lowe2017multi, jin2018real} present another promising direction for learning. They propose to train multiple agents in parallel when solving a task, and to combine multiple action-value functions with a centralized action-value function. The multi-critics supervision in our proposed A2MC model can be seen as a form of joint-task or multi-task learning~\cite{teh2017distral} for both long-term and short-term rewards. Our empirical results based on learning the characterization of long/short-term reward statistics also echo the effectiveness of a recently proposed off-policy reinforcement learning framework~\cite{bellemare2017distributional} that features a distributional variant of Q-learning, wherein the value functions are learned to match the distribution of standard immediate returns. Also, \cite{wang2016sample} shows that applying experience replay to on-policy methods can further enhance learning stability. \cite{schulman2015high} proposes a variant of advantage function that provides both low-variance and low-bias gradient estimates. These works are orthogonal to our approach can potentially be combined with the proposed characterization of past reward statistics to further enhance learning performance. \section{Experiments} We use the same network architecture and natural gradient optimization method as in the ACKTR model~\cite{wu2017scalable}. We set $\sigma_{max} = 1.0$, $\tau = 2.0$ and $T=20$ in the computation of variability-weighted reward. For hot-wiring exploration, we choose $\epsilon = 0.20$ and initial stage to be first $\frac{1}{40}$ of the total training period for all experiments. Other hyperparameters such as learning rate and gradient clipping remain the same as in the ACKTR model~\cite{wu2017scalable}. We first present results of evaluating the proposed A2MC model in two standard benchmarks, the discrete Atari experiments and the continuous MuJoCo domain. Then we show further ablation studies on the robustness of the hyper-parameters involved as well as evaluating the extensibility of the proposed long/short-term reward characterizations to other on-policy methods. \begin{figure}[t] \begin{center} \includegraphics[width=\linewidth]{Figures/Good.pdf} \vskip -0.3cm \caption{Performance of A2MC on Atari games. ``Hot-Wiring'' exploration makes the agent easier to figure out how to play challenging games like ``Robotank'' and ``WizardOfWor'', and for most games, it provides a better initial state for the agent to start off at a game and hence to obtain better final results. The number in figure legend shows the average reward among the last 100 episodes and the percentage shows the performance margin as compared to ACKTR. The shaded region denotes the standard deviation over 2 random seeds.} \vskip -0.3cm \label{fig:good} \end{center} \end{figure} \subsection{ATARI 2600 Games} We follow standard evaluation protocol to evaluate A2MC in a variety of Atari game environments (starting with 30 no-op actions). We train our models for 15 million timesteps for each game environment and score each game based on the average episode rewards obtained among the last 100 episodes as in~\cite{wu2017scalable}. The learning results on 12 Atari games are shown in Figure~\ref{fig:good} where we also included an ablation experiment of A2MC without hot-wiring. We observe that on average A2MC improves upon ACKTR in terms of final performance under the same training budget. Our A2MC is able to consistently improve agent performance based on the proposed characterization of reward statistics, hence the agent is able to get out of local minima in less time (higher sample efficiency) compared to ACKTR. The complete learning results on all games are attached in the Appendix B. We further expand the training budget and continue learning the games until 50 million timesteps as in~\cite{wu2017scalable}. As shown in Table~\ref{tab:humanLevel}, our A2MC model can solve games like ``Boxing'', ``Freeway'' and ``Enduro'' that are challenging for the baseline ACKTR model. For a full picture of model performance in Atari games, A2MC has a human-level performance rate of $74.5\%$ (38 out of 51 games) in the Atari benchmarks, compared to $63.6\%$ reached by ACKTR. Individual game scores for all the Atari games are reported in the Appendix B. \begin{figure}[t] \begin{center} \includegraphics[width=\linewidth]{Figures/MuJoCo1.pdf} \vskip -0.3cm \caption{Performance on the MuJoCo benchmark. A2MC is also competitive on MuJoCo continuous domain when compared to ACKTR. The shaded region denotes the standard deviation over 3 random seeds.} \label{fig:mujoco1} \end{center} \end{figure} \begin{table}[h] \centering \caption{Comparison of average episode rewards at the end of 50 million timesteps in Atari experiments. The reward scores and the first episodes reaching human-level performance~\cite{mnih2015human} are reported as in~\cite{wu2017scalable}. A2MC is able to solve games that are challenging to ACKTR and also retain comparable performance in the rest of games.} \label{tab:humanLevel} \begin{tabular}{lrrrrrrr} \toprule & & & \multicolumn{2}{c}{ACKTR} & & \multicolumn{2}{c}{A2MC} \\ \cmidrule{4-5}\cmidrule{7-8} Domain & Human Level& & Rewards & Episode & & Rewards & Episode \\ \midrule Asteroids & 47388.7 & & 34171.0 & N/A & & \textbf{830232.5} & \textbf{11314} \\ Beamrider & 5775.0 & & 13581.4 & 3279 & & 13564.3 & \textbf{3012} \\ Boxing & 12.1 & & 1.5 & N/A & & \textbf{99.1} & \textbf{158} \\ Breakout & 31.8 & & \textbf{735.7} & 4097 & & 411.4 & \textbf{3664} \\ Double Dunk & -16.4 & & -0.5 & 742 & & \textbf{21.3} & \textbf{544} \\ Enduro & 860.5 & & 0.0 & N/A & & \textbf{3492.2} & \textbf{730} \\ Freeway & 29.6 & & 0.0 & N/A & & \textbf{32.7} & \textbf{1058} \\ Pong & 9.3 & & 20.9 & 904 & & 19.4 & \textbf{804} \\ Q-bert & 13455.0 & & 21500.3 & \textbf{6422} & & \textbf{25229.0} & 7259 \\ Robotank & 11.9 & & 16.5 & - & & \textbf{25.7} & 4158 \\ Seaquest & 20182.0 & & 1776.0 & N/A & & \textbf{1798.6} & N/A \\ Space Invaders & 1652.0 & & \textbf{19723.0} & 14696 & & 11774.0 & \textbf{11064} \\ Wizard of Wor & 4756.5 & & 702 & N/A & & \textbf{7471.0} & \textbf{8119} \\ \bottomrule \end{tabular} \vskip -0.3cm \end{table} \subsection{Continuous Control} For the evaluations on continuous control tasks simulated in MuJoCo environment, we first follow~\cite{wu2017scalable} and tune a different set of hyper-parameters from Atari experiments. Specifically, all MuJoCo experiments are trained with a larger batch size of 2500. The results of eight MuJoCo environments trained for 1 million timesteps are shown in Figure~\ref{fig:mujoco1}. We observe that A2MC also performs well in all MuJoCo continuous control tasks. In particular, A2MC has obtained significant improvement as compared to ACKTR on the tasks of \textit{HalfCheetah}, \textit{Swimmer} and \textit{Walker2d} (see Table~\ref{tab:mujoco}). To test the robustness of A2MC, we perform another set of evaluations on MuJoCo tasks by keeping an identical set of hyper-parameters used in the Atari experiments. Figure~\ref{fig:mujoco2} shows this ablation result. We observe that even under sub-optimal hyper-parameters, our A2MC model can still learn to solve the MuJoCo control tasks in the long run. Moreover, it is less prone to overfitting when compared to ACKTR under such ``stress testing''. Additional hyper-parameter studies are shown in Appendix C. \begin{figure}[t] \begin{center} \includegraphics[width=\linewidth]{Figures/MuJoCo1_stress.pdf} \vskip -0.4cm \caption{``Stress testing'' ablation study on the MuJoCo continuous benchmark using hyperparameters tuned in Atari discrete control. Although this set of hyperparameters is suboptimal for the MuJoCo continuous control tasks, A2MC still obtain reasonable performance in the long run and it is less prone to overfitting.} \label{fig:mujoco2} \end{center} \vskip -0.5cm \end{figure} We also evaluate a multi-critics variant of the proximal policy optimization (PPO) model on the MuJoCo tasks with our proposed long/short-term rewards. In particular, we observe that our proposed variability-weighted reward generalizes well with the vanilla PPO, and our multi-critics PPO variant (MC-PPO) shows more favorable performance, as shown in Table~\ref{tab:mujoco}. Specifically, MC-PPO shows the best performance on \textit{Hopper} and \textit{Walker-2d} among all models under the 1-million timesteps training budget. Both of our multi-critics variants (A2MC and MC-PPO) have won 6 out of the 8 MuJoCo tasks with relative performance margins (percentages in parentheses) larger than $25\%$ (see Table~\ref{tab:mujoco}). \begin{table}[h] \centering \caption{Average episode rewards obtained among the last 10 episodes upon 1 million timesteps of training in MuJoCo experiments. } \label{tab:mujoco} \setlength\tabcolsep{2.0pt} \begin{tabular}{lrrrrrrrccrrrr \toprule GAMES & \multicolumn{3}{l}{ACKTR} && \multicolumn{2}{c}{Our A2MC} &\ \ && \multicolumn{2}{c}{PPO} && \multicolumn{2}{c}{Our MC-PPO} \\ \midrule Ant & 1671.6 &&& & \textbf{2216.1}&\textbf{(32.5\%)} &\ \ && 411.4 &($\pm$ 107.7) && $\mathbf{618.9 }$&$\mathbf{(50.4\%)}$ \\ HalfCheetah & 1676.2 &&& & \textbf{2696.6}&\textbf{(60.8\%)} &\ \ && 1433.7 &($\pm$ 83.9) && $\mathbf{2473.4 }$&$\mathbf{(72.5\%)}$ \\ Hopper & 2259.1 &&& & \textbf{2835.7}&\textbf{(25.5\%)} &\ \ && 2055.8 &($\pm$ 274.6) && $\mathbf{3131.3 }$&$\mathbf{(52.3\%)}$ \\ InvertedDoublePendulum & 6295.4 &&& & \textbf{7872.6}&\textbf{(25.0\%)} &\ \ && 4454.1 &($\pm$ 1098.1) && $\mathbf{7648.7 }$&$\mathbf{(71.7\%)}$ \\ InvertedPendulum & 1000.0 &&& & 957.2 & (-4.2\%) &\ \ && 839.7 &($\pm$ 127.1) && $777.4 $&$(-7.4\%)$ \\ Reacher & -4.2 &&& & -3.9 & (0.4\%) &\ \ && -5.47 &($\pm$ 0.3) && $-10.3 $&$(-8.5\%)$ \\ Swimmer & 43.2 &&& & \textbf{187.4}&\textbf{(333.7\%)} &\ \ && 79.1 &($\pm$ 31.2) && $\mathbf{102.9 }$&$\mathbf{(30.2\%)}$ \\ Walker2d & 1090.8 &&& & \textbf{2405.9}&\textbf{(120.5\%)} &\ \ && 2300.8 &($\pm$ 397.6) && $\mathbf{3718.1 }$&$\mathbf{(61.6\%)}$ \\ \midrule Win | Fair | Lose & \multicolumn{2}{c}{N/A} &&& \multicolumn{2}{c}{ 6 | 2 | 0 } &\ \ && \multicolumn{2}{c}{N/A} & & \multicolumn{2}{c}{ 6 | 2 | 0 } \\ \bottomrule \end{tabular} \vskip -0.3cm \end{table} \section{Effects of Flipping} While introducing the variability-weighted reward, a flipping operation is conducted in the pre-processing of the reward sequence as formulated in Eq. (3). In Figure~\ref{fig:flip1} and \ref{fig:flip2}, we construct 4 reward sequences to show that the flipping operation can further penalize the oscillation in the recent past rewards while encourage recent stable rewards. \textit{(a1, a2, b1, b2)} share the same value of immediate reward at $t=9$ and thus the $\mathcal{R}_H$ of all reward sequences are the same. \begin{figure}[h] \begin{minipage}{.5\linewidth} \centering\includegraphics[width=.87\linewidth]{Figures/Fig_flip_a.pdf} \vskip -0.3cm \caption{Calculation without flipping.} \label{fig:flip1} \end{minipage} \begin{minipage}{.5\linewidth} \centering\includegraphics[width=.9\linewidth]{Figures/Fig_flip_b.pdf} \vskip -0.3cm \caption{Calculation with flipping.} \label{fig:flip2} \end{minipage} \end{figure} Therefore, the variability-weighted reward only depends on the volatility statistics of $\delta_R$, i.e., how varied past rewards were. \textbf{Without flipping.} In Figure~\ref{fig:flip1}, sequence \textit{(a1)} and \textit{(a2)} are mirror symmetrical to the $y$-axis, and the only difference between them is that the recent past rewards ($t=5,6,7,8$) of \textit{(a2)} are more stable than \textit{(a1)}. Intuitively, we want to encourage stable past rewards like \textit{(a2)} while penalizing oscillation in \textit{(a1)}. As presented in the third row of Figure~\ref{fig:flip1}, the $r^{vwr}$ difference of \textit{(a1)} and \textit{(a2)} is less than 1 without flipping in the pre-processing. \textbf{With flipping}. In Figure~\ref{fig:flip2}, \textit{(b1, b2)} exactly have the same reward sequence as \textit{(a1, a2)}, respectively. However, flipping is performed as a step of pre-processing, largely increasing the $r^{vwr}$ gap (from less than 1 to nearly 4) between the two constructed sequences. Comparing \textit{(b1, b2)} with \textit{(a1, a2)} , the post-processed sequences $\vec{\mathcal{R}}$ (shown in green) become centrosymmetric to those without flipping. Specifically, the recent reward drops at $t=6,7,8$ are reflected as high values at the beginning of $\vec{\mathcal{R}}$ as shown in \textit{(b1)}, while oscillations long ago are transformed into high values at the end of $\vec{\mathcal{R}}$ as presented in \textit{(b2)}. When compared to the zero-variability reference (shown in orange), which is designed as an exponential function, the flipping leads to a higher variability for the former sequence while a lower variability for the latter one, enlarging the $r^{vwr}$ gap between those two sequences. \section{Complete results in Atari 2600 Games} We show the learning curves for 15 million timesteps on all Atari games in Figure~\ref{fig:atari-all} and in Table~\ref{tab:atari-full} we show the complete results of training til 50 million timesteps. report the mean episode reward as in~\cite{wu2017scalable}. Entries with $\sim$ indicates approximated value as retrieved from learning figures published by~\cite{wu2017scalable}. Results from other models are taken from~\cite{wu2017scalable} and ~\cite{mnih2015human}. We show that A2MC has reached a human-level performance rate of $74.5\%$ (38 out of 51 games) as compared to $63.6\%$ reached by ACKTR. The relative performance margin of A2MC as compared to ACKTR is also shown. \begin{table}[] \centering \caption{Raw scores across all games, starting with 30 no-op actions. Scores are reported by averaging the last 500 episodes upon 50 million timesteps of training as in~\cite{wu2017scalable}. A relative margin comparing A2MC to ACKTR is shown. Scores from other models are taken from~\cite{wu2017scalable} and~\cite{mnih2015human}.} \label{tab:atari-full} \setlength\tabcolsep{2.8pt} \begin{tabular}{l|rrrrrr|r} GAME & Human & DQN & DDQN & Prior. Duel & ACKTR & Our A2MC & (Margin) \\ \hline Alien & 7127.7 & 1620 & 3747.7 & 3941 & 3197.1 & 2986.3 & -6.6\% \\ Amidar & 1719.5 & 978 & 1793.3 & 2296.8 & 1059.4 & 2040.1 & \textbf{92.6\%} \\ Assault & 742.0 & 4280.4 & 5393.2 & 11477 & 10777.7 & 9892.4 & -8.2\% \\ Asterix & 8503.3 & 4359 & 17356.5 & 375080 & 31583.0 & 32671.0 & 3.4\% \\ Asteroids & 47388.7 & 1364.5 & 734.7 & 1192.7 & 34171.6 & 828931.6 & \textbf{2325.8\%} \\ Atlantis & 29028.1 & 279987 & 106056 & 395762 & 3433182.0 & 2886274.0 & -15.9\% \\ Bankheist & 753.1 & 455 & 1030.6 & 1503.1 & 1289.7 & 1290.6 & 0.1\% \\ Battlezone & 37187.5 & 29900 & 31700 & 35520 & 8910.0 & 10570.0 & \textbf{18.6\%} \\ Beamrider & 16926.5 & 8627.5 & 13772.8 & 30276.5 & 13581.4 & 13715.6 & 1.0\% \\ Berzerk & 2630.4 & 585.6 & 1225.4 & 3409 & 927.2 & 974.0 & 5.0\% \\ Bowling & 160.7 & 50.4 & 68.1 & 46.7 & 24.3 & 31.6 & \textbf{30.0\%} \\ Boxing & 12.1 & 88 & 91.6 & 98.9 & 1.5 & 93.5 & \textbf{6344.8\%} \\ Breakout & 30.5 & 385.5 & 418.5 & 366 & 735.7 & 420.6 & -42.8\% \\ Centipede & 12017.0 & 4657.7 & 5409.4 & 7687.5 & 7125.3 & 12096.5 & \textbf{69.8\%} \\ Choppercommand & 9882.0 & N/A & N/A & N/A & $\sim$8000 & 12149.0 & \textbf{$\sim$42.5\%} \\ Crazyclimber & 35829.4 & 110763 & 117282 & 162224 & 150444.0 & 152439.0 & 1.3\% \\ Demonattack & 1971.0 & 12149.4 & 58044.2 & 72878.6 & 274176.7 & 361807.1 & \textbf{32.0\%} \\ Doubledunk & -16.4 & -6.6 & -5.5 & -12.5 & -0.5 & 20.6 & \textbf{3907.5\%} \\ Enduro & 860.5 & 729 & 1211.8 & 2306.4 & 0.0 & 3550.6 & $\infty$\% \\ Fishingderby & -38.7 & -4.9 & 15.5 & 41.3 & 33.7 & 38.4 & \textbf{13.9\%} \\ Freeway & 29.6 & 30.8 & 33.3 & 33 & 0.0 & 32.7 & $\infty$\% \\ Frostbite & 4335.0 & N/A & N/A & N/A & $\sim$280 & 293.7 & $\sim$5.1\% \\ Gopher & 2412.5 & 8777.4 & 14840.8 & 104368.2 & 47730.8 & 86101.4 & \textbf{80.4\%} \\ Gravitar & 2672.0 & N/A & N/A & N/A & $\sim$300 & 995.0 & -2.9\% \\ Icehockey & 0.9 & -1.9 & -2.7 & -0.4 & -4.2 & -2.1 & \textbf{16.3\%} \\ Jamesbond & 302.8 & 768.5 & 1358 & 812 & 490.0 & 545.0 & \textbf{11.2\%} \\ Kangaroo & 3035.0 & 7259 & 12992 & 1792 & 3150.0 & 11269.0 & \textbf{257.7\%} \\ Krull & 2665.5 & 8422.3 & 7920.5 & 10374.4 & 9686.9 & 10245.4 & 5.8\% \\ Kungfumaster & 22736.3 & 26059 & 29710 & 48375 & 34954.0 & 39773.0 & \textbf{13.8\%} \\ Mspacman & 15693.0 & N/A & N/A & N/A & $\sim$3500 & 5006.1 & $\sim$\textbf{34.5\%} \\ Namethisgame & 4076.0 & N/A & N/A & N/A & $\sim$12500 & 12569.9 & $\sim$0.6\% \\ Phoenix & 7242.6 & 8485.2 & 12252.5 & 70324.3 & 133433.7 & 221288.3 & \textbf{65.8\%} \\ Pitfall & 6463.7 & -286.1 & -29.9 & 0 & -1.1 & -2.5 & -0.3\% \\ Pong & 14.6 & 20.9 & 21 & 20.9 & 20.9 & 19.7 & -5.9\% \\ Privateeye & 69571.0 & N/A & N/A & N/A & $\sim$560 & 507.0 & -9.5\% \\ Qbert & 13455.0 & 13117.3 & 15088.5 & 18760.3 & 23151.5 & 24075.8 & 4.0\% \\ Riverraid & 17118.0 & 7377.6 & 14884.5 & 20607.6 & 17762.8 & 18671.9 & 5.1\% \\ Roadrunner & 7845.0 & 39544 & 44127 & 62151 & 53446.0 & 50071.0 & -6.3\% \\ Robotank & 11.9 & 63.9 & 65.1 & 27.5 & 16.5 & 26.5 & \textbf{60.5\%} \\ Seaquest & 42054.7 & 5860.6 & 16452.7 & 931.6 & 1776.0 & 1805.6 & 1.7\% \\ Solaris & 12326.7 & 3482.8 & 3067.8 & 133.4 & 2368.6 & 2277.2 & -3.9\% \\ Spaceinvaders & 1668.7 & 1692.3 & 2525.5 & 15311.5 & 19723.0 & 13544.2 & -31.3\% \\ Stargunner & 10250.0 & 54282 & 60142 & 125117 & 82920.0 & 89616.0 & 8.1\% \\ Tennis & -8.9 & N/A & N/A & N/A & $\sim$-12 & -4.7 & $\sim$\textbf{20.4\%} \\ Timepilot & 5229.2 & 4870 & 8339 & 7553 & 22286.0 & 21992.0 & -1.3\% \\ Tutankham & 167.6 & 68.1 & 218.4 & 245.9 & 314.3 & 193.7 & -38.4\% \\ Upndown & 11693.2 & 9989.9 & 22972.2 & 33879.1 & 436665.8 & 563659.3 & \textbf{29.1\%} \\ Videopinball & 17667.9 & 196760.4 & 309941.9 & 479197 & 100496.0 & 127452.4 & \textbf{26.8\%} \\ Wizardofwor & 4756.5 & 2704 & 7492 & 12352 & 702.0 & 7864.0 & \textbf{1020.2\%} \\ YarsRevenge & 54576.9 & 18098.9 & 11712.6 & 69618.1 & 125169.0 & 143141.5 & \textbf{14.4\%} \\ Zaxxon & 9173.3 & 5363 & 10163 & 13886 & 17448.0 & 19365.0 & \textbf{11.0\%} \\ \hline \begin{tabular}[c]{@{}l@{}} Human-level \\ (Win / Total)\end{tabular} & N/A & \begin{tabular}[c]{@{}r@{}}21 / 44 \\ (47.7\%)\end{tabular} & \begin{tabular}[c]{@{}r@{}}31 / 44 \\ (70.4 \%)\end{tabular} & \begin{tabular}[c]{@{}r@{}}34 / 44 \\ (77.3 \%)\end{tabular} & \begin{tabular}[c]{@{}r@{}}28 / 44 \\ (63.6 \%)\end{tabular} & \begin{tabular}[c]{@{}r@{}} \textbf{38 / 51} \\ (74.5 \%)\end{tabular} & \end{tabular} \end{table} \section{Hyper-parameter Studies} The proposed variability-weighted reward mechanism considers the volatility of rewards by keeping a $T$-step history of agent's performance. The hyper-parameter $T=20$ is empirically chosen to be the same as the look-ahead parameter $N$ in standard on-policy methods, so as to keep the same period ($T=N=20$) in ``T-step history'' and ``N-step forward''. And $\sigma_{max} = 1$ is chosen as the average of the observed volatility based on statistics in the $T$ history rewards of the ACKTR models. As parameter choices could be vital, we perform an additional ablation study shown below. The result shows that the performance of A2MC is robust across different parameters of choice and is not too sensitive to changes on either of the hyper-params. \begin{center} \vskip -0.2cm \begin{tabular}{lrrrrrrr} \multirow{2}{*}{Games} & \multirow{2}{*}{ACKTR} & \multirow{2}{*}{A2MC w/} & T=20 & T=10 & T=10 & T=40 & T=40 \\ & & & $\sigma_{max}$=1 & $\sigma_{max}$=1 & $\sigma_{max}$=2 & $\sigma_{max}$=1 & $\sigma_{max}$=2 \\ \midrule Boxing & 1.23 & & \textbf{99.19} & 94.76 & 98.51 & 99.18 & 98.07 \\ Jamesbond & 409.50 & & 453.50 & 438.50 & \textbf{470.00} & 442.25 & 457.75 \\ Wizard of Wor & 744.50 & & 5448.00 & \textbf{5601.00} & 5363.50 & 2528.50 & 3287.50 \\ \end{tabular} \vskip -0.7cm \end{center} \section{Extension to Multi-Critic PPO (MC-PPO)} The learning results of the proposed MC-PPO model on the MuJoCo tasks are shown in Figure~\ref{fig:mujoco-mcppo}. MC-PPO shows the best performance on \textit{Hopper} and \textit{Walker-2d} among all models under the 1-million timesteps training budget. Both of our multi-critics variants (A2MC and MC-PPO) have won 6 out of the 8 MuJoCo tasks with relative performance margins (percentages in parentheses) larger than $25\%$. \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{Figures/MuJoCo_mcppo.pdf} \vskip -0.3cm \caption{Performance on the MuJoCo continuous control benchmarks using PPO-based methods. Our proposed long/short-term reward characterization can be extended to the PPO method, i.e., the proposed multi-critic variant of PPO (MC-PPO). The shaded region denotes the standard deviation over 3 random seeds.} \label{fig:mujoco-mcppo} \end{center} \end{figure} \section{Algorithm} The learning algorithm of A2MC is shown in Algorithm 1. \clearpage \begin{algorithm}[h] \label{algo} \caption{Advantage Actor Multi-Critic Learning (A2MC)}\label{euclid} \color{black} \begin{algorithmic}[1] \color{black} \State $\text{Initialize parameters:} \ \theta_{a}, \theta_{v}^{j}, j \in \{\text{short-term}, \text{long-term}\}$ \State $\text{Initialize look-ahead steps:} \ N$, step counter: $T=0$, maximum step: $T_{max}$ \State $\text{Initialize hot-wire probability:} \ \epsilon$ \State $\text{Initialize environment:} \ Env$ \State $\text{Initialize reward history:} \ \vec{\mathbf{r}}$ \Repeat \State $\text{Reset gradients:} \ d\theta \gets 0 \ \text{and} \ d\theta_{v}^{j} \gets 0, j \in \{\text{short-term}, \text{long-term}\}$ \State $\text{Get state:} \ s_{t} \gets Env$ \State $flag = 1$, $a_{rand}$ is uniformly sampled in action space with probability $\epsilon$, otherwise $flag = 0$ \For{$t = 0 : N-1$} \State $\text{Perform\ } a_{t} \text{\ according to policy\ } \pi(a_t | s_t; \theta_{a}) \ \textbf{if not} \ flag \ \textbf{else} \ a_t = a_{rand} $ \State $\text{Received reward\ } r_{t} \text{\ and new state\ } s_{t+1}, \ \text{append\ } r_{t} \text{\ to\ } \vec{\mathbf{r}}$ \State Calculate $r_{t}^{vwr}$ from $\vec{\mathbf{r}}$ based on Eq. (2-7) \State $T \gets T+1$ \EndFor \State $R^{\text{short-term}} = V(s_{N}; \theta_v^{\text{short-term}}) $ \State $R^{\text{long-term}} = V(s_{N}; \theta_v^{\text{long-term}}) $ \For{$i = N-1$ \textbf{to} $0$ \textbf{step} $-1$} \State $R^{\text{short-term}} \gets r_{i} + \gamma R^{\text{short-term}}$ \State $R^{\text{long-term}} \gets r_{i}^{vwr} + \gamma R^{\text{long-term}}$ \State $\text{Advantange gradients wrt\ } \theta_{a}: d\theta_{a} \gets d\theta_{a} + \nabla_{\theta_{a}}\log \pi(a_i|s_i;\theta_a)\sum_j(R^{j}-V(s_i; \theta_v^{j}))$ \For{$j \in \{\text{short-term}, \text{long-term}\}$} \State $\text{Accumulate gradients wrt\ } \theta_{v}^{j}: d\theta_{v}^{j} \gets d\theta_{v}^{j} + \partial (R^{j} - V(s_i; \theta_{v}^{j}))^2 / \partial \theta_{v}^{j}$ \EndFor \EndFor \Until $T\geq{T_{max}}$ \end{algorithmic} \end{algorithm} \begin{figure}[h] \begin{center} \includegraphics[width=\linewidth]{Figures/Atari_All.pdf} \vskip -0.3cm \caption{Performance of A2MC on Atari games. The number in figure legend shows the average reward among the last 100 episodes upon 15 million timesteps and the percentage shows the performance margin as compared to ACKTR. The shaded region denotes the standard deviation over 2 random seeds.} \label{fig:atari-all} \end{center} \end{figure} \section{Conclusion} In this work, we introduce an effective auxiliary reward signal to remedy the deficiencies of learning solely from the standard environment rewards. Our proposed characterization of past reward statistics results in improved learning and higher sample efficiencies for on-policy methods, especially in challenging tasks with sparse rewards. Experiments on both discrete tasks in Atari environment and MuJoCo continuous control tasks validate the effectiveness of utilizing the proposed long/short-term reward statistics for on-policy methods using multi-critic architectures. This suggests that expanding the form of reward feedbacks in reinforcement learning is a promising research direction.

246cee145cafa63e6be6c616661c6db0017244fe

\section{Data Generation} \label{app:data-generation} \begin{algorithm}[tb]\captionsetup{labelfont={sc,bf}} \caption{\textsc{GenerateIHDP}: Semi-synthetic Data Generation Algorithm for Fair Causal Inference} \label{alg:ihdp-main} \begin{algorithmic} \STATE {\bfseries Step 1:} Remove all children from the dataset with non-white mothers who received the original treatment (as in \citet{hill2011bayesian}). \STATE {\bfseries Step 2:} Optional: Remove extra features from $X$. \STATE {\bfseries Step 3:} Normalize data (for each feature of $X$, subtract mean and divide by standard deviation). \STATE {\bfseries Step 4:} Remove some features $Z$ from the data to act as unobserved confounders. \STATE {\bfseries Step 5:} Remove some feature $A$ from the data to act as the sensitive attribute. \STATE {\bfseries Step 6:} Sample factual and counterfactual outcomes $\{y_{T=t,A=a} \ \forall t, a\} = \textsc{GenerateOutcomes(X, Z)}$. \STATE {\bfseries Step 7:} Sample factual and counterfactual treatments $\{t_{A=a} \ \forall a\} = \textsc{GenerateTreatments(Z)}$. \STATE {\bfseries Return} $Z, A, X, \{y_{T=t,A=a} \ \forall t, a\}, \{t_{A=a} \ \forall a\}$ \end{algorithmic} \end{algorithm} \begin{algorithm}[tb]\captionsetup{labelfont={sc,bf}} \caption{\textsc{GenerateOutcomes}: Generate outcomes for each value of the treatment and sensitive attribute (style of \citep{hill2011bayesian}, Resp. B)} \label{alg:ihdp-response} \begin{algorithmic} \STATE {\bfseries Input:} Features $X$, unobserved confounders $Z$ \STATE Let $[X, Z]$ denote the horizontal concatenation of $X$ and $Z$, and let the offset matrix $W$ be the shape of $[X, Z]$ with 0.5 in every position. \STATE Sample $\beta \sim P_{\beta}$, choose $\omega, \beta_A \in \mathds{R}$. \STATE Sample $y_{T=0,A=0} \sim \mathcal{N}(\exp(([X, Z] + W)\beta^T), 1)$ \STATE Sample $y_{T=1,A=0} \sim \mathcal{N}([X, Z]\beta^T - \omega, 1)$ \STATE Sample $y_{T=0,A=1} \sim \mathcal{N}(\exp(([X, Z] + W)\beta^T) + \beta_A, 1)$ \STATE Sample $y_{T=1,A=1} \sim \mathcal{N}([X, Z]\beta^T - \omega + \beta_A), 1)$ \STATE \textbf{Return} \{$y_{T=0,A=0}, y_{T=1,A=0}, y_{T=0,A=1}, y_{T=1,A=1}$\} \end{algorithmic} \end{algorithm} \begin{algorithm}[tb]\captionsetup{labelfont={sc,bf}} \caption{\textsc{GenerateTreatments}: Generate treatments for each value of the sensitive attribute} \label{alg:ihdp-treatment} \begin{algorithmic} \STATE {\bfseries Input:} Unobserved confounders $Z$ \STATE Choose $\alpha_0, \alpha_1 \in [0, 1], \zeta \in \mathds{R}$. \STATE Let $p_{A=0} = Clip(\alpha_0 + \zeta Z, 0, 1), p_{A=1} = Clip(\alpha_1 + \zeta Z, 0, 1)$ \STATE Sample $t_{A=0} \sim Bern(p_{A=0})$. \STATE Sample $t_{A=1} \sim Bern(p_{A=1})$. \STATE \textbf{Return} $\{t_{A=0}, t_{A=1}\}$ \end{algorithmic} \end{algorithm} We detail our dataset generation progess in Algorithm \ref{alg:ihdp-main}. We denote the outcome $Y$ under interventions $do(T=t), do(A=a$) as $y_{T=t,A=a}$. The subroutines in Algorithms \ref{alg:ihdp-response} and \ref{alg:ihdp-treatment} generate all factual and counterfactual outcomes and treatments for each example, one for each possible setting of $A$ and/or $T$. In Algorithm \ref{alg:ihdp-main}, we have several undefined constant variables. We use the following values for those variables: \begin{itemize} \item $\beta \sim P_{\beta}$: \begin{itemize} \item for continuous variables, \\ \ $\beta_i \sim Cat([0,.1,.2,.3,.4],[.5, .125, .125, .125, .125])$ \item for binary variables $\beta_i \sim Cat([0,.1,.2,.3,.4], [.6, .1, .1, .1, .1])$ \item for $Z$, $\beta_i = Cat([.4,.6], [.5,.5])$ \end{itemize} where $Cat(x, p)$ selects values from $x$ according to the array of probabilities $p$. \item $\beta_A = 6$ \item $\omega = -11$ \item $\alpha_0, \alpha_1, \zeta = 0.7, 0.4, 0.1$ \end{itemize} We also use the function $Clip$, which is defined as: \begin{equation} Clip(x, m, M) = \min(\max(x, m), M); x, m, M \in \mathds{R} \end{equation} \section{Identifiability of Causal Effects} \label{app:proofs} Here we show that if we can successfully recover the joint distribution $P(Z, A, X, T, Y)$, we can recover all three treatment effects we are interested in: \begin{enumerate} \item The effect of $T$ on $Y$ ($T \rightarrow Y$): $\mathds{E}(Y | do(T = 1), X, A) - \mathds{E}(Y | do(T = 0), X, A)$ \item The effect of $A$ on $T$ ($A \rightarrow T$): $\mathds{E}(T | do(A = 1), X) - \mathds{E}(T | do(A = 0), X)$ \item The effect of $A$ on $Y$ ($A \rightarrow Y$): $\mathds{E}(Y | do(A = 1), X) - \mathds{E}(Y | do(A = 0), X)$ \end{enumerate} Our proof will closely follow \citet{louizos2017causal}. For each effects, it will suffice to show that we can recover the first term on the right-hand side of each expression. (The argument for the second term is the same). We will show only the proof for the effect of $T$ on $Y$ --- the others are very similar. \\ {\bf Theorem.} Given the causal model in Fig. \ref{subfig:cevae-fair}, if we recover the joint distribution $P(Z, A, X, T, Y)$, then we can recover $\mathds{E}(Y | do(T = 1), X, A)$. {\it Proof.} We have that \begin{equation} \begin{aligned} P(Y | do(T = 1), X, A) &= \int_Z P(Y | do(T = 1), X, A, Z) P(Z | do(T = 1), X, A)\\ \end{aligned} \end{equation} By the $do$-calculus, we can reduce further: \begin{equation} \begin{aligned} &= \int_Z P(Y | T = 1, X, A, Z) P(Z | X, A) \end{aligned} \end{equation} If we know the joint distribution $P(Z, A, X, T, Y)$, we can identify the value of each term in this expression; hence we can identify the value of the whole expression. \hfill \qedsymbol\\ \section{Experimental details} \label{app:experiments} We run each model on 500 distinct data seed/model seed pairs, in order to get robust confidence estimates on the error of each model. We parametrize each function in our causal model with a neural network. Our networks between $X$ and $Z$ have a single hidden layer of 20 hidden units. The size of the learned hidden confounder $Z$ was 10 units. Each of our TARNets consist of a network outputting a shared representation, and two networks making predictions from that representation. Each of these network have 1 hidden layer with 100 hidden units. The size of the shared representation in the TARNets was 20 units. For simplicity, we set $g_X^{\sigma} = 1$ for all experiments (but not $g_Z^{\sigma}$)---this amounts to assuming unit variance for the data $X$, a sensible assumption because they are normalized during pre-processing. We used ELU non-linear activations \cite{clevert2015fast}. We trained our model with ADAM \citep{kingma2014adam} with a learning rate of 0.001, calculating the ELBO on a validation set and stopping training after 10 consecutive epochs without improvement. We sample 10 times from the posterior $q(Z|\cdot)$ at both training and test time for each input example. At training time we compute the average ELBO across the ten samples, while at test time we use the average prediction. \section{Introduction} \label{sec:intro} In this work, we consider the problem of fair decision-making from biased datasets. Much work has been done recently on the problem of fair classification \citep{zafar2015fairness,hardt2016equality,bechavod2017penalizing,agarwal2018reductions}, yielding an abundant supply of definitions, models, and algorithms for the purposes of learning classifiers whose outputs satisfy distributional constraints. Some of the canonical problems for which these algorithms have been proposed are loan assignment \citep{hardt2016equality}, criminal risk assessment \citep{chouldechova2017fair}, and school admissions \citep{friedler2016possibility}. However, none of these problems are fully specified by the classification paradigm. Rather, they are decision-making problems: each problem requires an action (or ``treatment'') to be taken in the world, which in turn yields an outcome. In other words, the central question is how to intervene in an ongoing and evolving process, rather than predict outcomes alone \citep{barabas2017interventions}. Decision-making, i.e. learning to intervene, requires a fundamentally different approach from learning to classify: historical training data are the product of past interventions and thus provide an incomplete view of all possible outcomes. Only actions which were previously chosen yield observable outcomes in the training data, while the implicit counterfactual outcomes (the outcome that would have occurred had another action been taken) are never observed. The incompleteness of this data can have great impact on learning and inference \cite{rubin1976inference}. It has been widely argued that biased data yields unfair machine learning systems \citep{kallus2018residual,hashimoto2018fairness,pmlr-v81-ensign18a}. In this work we examine dataset bias through the lens of causal inference. To understand how past decisions may bias a dataset, we first must understand how sensitive attributes may have affected the generative process which created the dataset, including the (historical) decision makers' actions (treatments) and results (outcomes). Causal inference is well suited to this task: since we are interested in decision-making rather than classification, we should be interested in the causal effects of actions rather than correlations. Causal inference has the added benefit of answering counterfactual queries: What would this outcome have been under another treatment? How would the outcome change if the sensitive attribute were changed, all else being equal? These questions are core to the mission of learning fair systems which aim to inform decision-making \citep{kusner2017counterfactual}. While there is much that causal inference can offer to the field of fair machine learning, it also poses several significant challenges. For example, the presence of \emph{hidden confounders}---unobserved factors that effect both the historical choice of treatment and the outcome---often prohibits the exact inference of causal effects. Additionally, understanding effects at the individual level can be especially complex, particularly if the outcome is non-linear in the data and treatments. These technical difficulties are often amplified by the problem scope of modern machine learning, where large and high-dimensional datasets are commonplace. To address these challenges, we propose a model for fairly estimating individual-level causal effects from biased data, which combines causal modeling \citep{pearl2009causality} with approximate inference in deep latent variable models \citep{kucukelbir2017automatic,louizos2017causal}. Our focus on individual-level causal effects and counterfactuals provides a natural fit for application areas requiring fair policies and treatments for individuals, such as finance, medicine, and law. Specifically, we incorporate the sensitive attribute into our model as a confounding factor, which can possibly influence both the treatment and the outcome. This is a first step towards achieving ``fairness through awareness'' \citep{dwork2012fairness} in the interventional setting. Our model also leverages recent advances in deep latent-variable modeling to model potential hidden confounders as well as complex, non-linear functions between variables, which greatly increases the class of relationships which it can represent. Through experimental analysis, we show that our model can outperform non-causal models, as well as causal models which do not consider the sensitive attribute as a confounder. We further explore the performance of this model, showing that fair-aware causal modeling can lead to more accurate, fairer policies in decision-making systems. \section{Background} \label{sec:causal-background} \subsection{Causal Inference}\label{sec:causal-inference} We employ Structural Causal Models (SCMs), which provide a general theory for modeling causal relationships between variables \citep{pearl2009causality}. An SCM is defined by a directed graph, containing vertices and edges, which respectively represent variables in the world and their pairwise causal relationships. There are two types of vertices: exogenous variables $\mathcal{U}$ and endogenous variables $\mathcal{V}$. Exogenous variables are unspecified by the model; we model them as unexplained noise distributions, and they have no parents. Endogenous variables are the objects we wish to understand; they are descendants of endogenous variables. The value of each endogenous variable is fully determined by its ancestors. Each $V \in \mathcal{V}$ has some function $f_V$ which maps the values of its immediate parents to its own. This function $f_V$ is deterministic; any randomness in an SCM is due to its exogenous variables. In this paper, we are primarily concerned with three endogenous variables in particular: $X$, the observable \emph{features} (or covariates) of some example; $T$, a \textit{treatment} which is applied to an example; and $Y$, the \textit{outcome} of a treatment. Our decision problem is: given an example with particular values for its features, $X=x$, what value should we assign to treatment $T$ in order to produce the best outcome $Y$? This is fundamentally different from a classification problem, since typically we observe the result of only one treatment per example \footnote{ {\color{black} Note that we use the terms \textit{treatment} and \textit{outcome} as general descriptors of a decision made/action taken and its result, respectively. These terms are associated with an alternative theory of causal inference \citep{rubin2005causal} which can also be used to describe the methods we propose, but which we will not discuss in this paper. } } . To answer this decision problem, we need to understand the value $Y$ will take if we \textit{intervene} on $T$ and set it to value $t$. Our first instinct may be to estimate $P(Y | T = t, X = x)$. However, this is unsatisfactory in general. If we are estimating these probabilities from observational data, then the fact that $x$ received treatment $t$ \textit{in the past} may have some correlation with the historical outcome $Y$. This ``confounding'' effect---the fact that $X$ has an effect on both $T$ and $Y$ is depicted in Figure \ref{subfig:pearl}, by the arrows pointing out of $X$ into $T$ and $Y$. For instance, in an observational medical trial, it is possible that young people are more likely to choose a treatment, and also that young people are more likely to recover. A supervised learning model, given this data, may then overestimate the average effectiveness of the treatment on a test population. Broadly, to understand the effect of assigning treatment $t$, supervised learning is not enough; we need to model the functions $\{f_V\}$ of the SCM. Once we have a fully defined SCM, we can use the $do$ operation \cite{pearl2009causality} to simulate the distribution over $Y$ given that we assign some treatment $t$---we denote this as $P(Y | do(T = t), X = x)$. We do the $do$ through graph surgery: we assign the value $t$ to $T$ by removing all arrows going into $T$ from the SCM and setting the corresponding structural equation output to the desired value regardless of its input $f_T(\cdot) = t$. We then set $X = x$ and continue with inference of $Y$ as we normally would. A common assumption in causal modelling is the ``no hidden confounders'' assumption, which states that there are no unobserved variables affecting both the treatment and outcome. We follow \citet{louizos2017causal}, and use variational inference to model confounders that are not directly observed but can be abstracted from proxies. {\color{black} In Sec. \ref{sec:models} we consider the implications of this approach and discuss alternative assumptions. } \subsection{Approximate Inference} Individual and population-level causal effects can be estimated via the \emph{do} operation when the values of all confounding variables are observed \citep{pearl2009causality}, which motivates the common no-hidden-confounders assumption in causal inference. However this assumption is rather strong and precludes classical causal inference in many situations relevant to fair machine learning, e.g., where ill-quantified and hard-to-observe factors such socio-economic status (SES) may significantly confound the observable data. Therefore we follow \citet{louizos2017causal} in modeling unobserved confounders using a high dimensional latent variable $Z$ to be inferred for each observation $(X, T, Y)$. They prove that if the full joint distribution is successfully recovered, individual treatment effects are identifiable, even in the presence of hidden confounders. In other words, causal effects are identifiable insofar as exact inference can be carried out, and the observed covariates are sufficiently informative. Because exact inference of $Z$ is intractable for many interesting models, we approximately infer $Z$ by variational inference, specifying $q(Z|X, T, Y)$ using a parametric family of distributions and learning the parameters that best approximate the true posterior $p(Z|X, T, Y)$ by maximizing the evidence lower bound (ELBO) of the marginal data likelihood \citep{wainwright2008graphical}. In particular, we amortize inference by training a neural network (whose functional form is specified separately from the causal model) to predict the parameters of $q$ given $(X, T, Y)$ \cite{kingma2013auto}. Amortized inference is much faster but less optimal than local inference \cite{kim2018semi}; alternate inference strategies could be explored for applications where the importance of accuracy in individual estimation justifies the additional computational cost. \subsection{TARNets} \label{sec:tarnets} TARNets \citep{shalit2016estimating} are {a \color{black} class of neural network} architectures for estimating outcomes of a binary treatment. {\color{black} The network comprises two separate arms---each predicts the outcomes associated with a separate treatment---that share parameters in the lower layers. The entire network is trained end to end using gradient-based optimization, but with } only one arm (the one with the treatment which was actually given) receiving error signal for any given example. The TARNet prediction of result $R$ and input variables $V$ and potential intervention $I$ is expressed by combining the shared representation function $\Phi$ with the two functions $h_0, h_1$ corresponding to the separate prediction arms. This yields two composed functions, \begin{equation} \begin{aligned} g^{I=0}_R(V, I) &= h_0(\Phi(V))\\ g^{I=1}_R(V, I) &= h_1(\Phi(V)) \end{aligned} \end{equation} with $h_0, h_1, \Phi$ realized as neural networks. \citet{shalit2016estimating} explore a group-wise MMD penalty on the outputs of $\Phi$; we do not use this. \section{Fair Causal Inference} \label{sec:problem-setup} As stated in Sec. \ref{sec:causal-inference}, we are interested in modeling the causal effects of treatments on outcomes. However, when attempting to learn fairly from a biased dataset, this problem takes on an extra dimension. In this context, we become concerned with understanding causal effects in the presence of a \emph{sensitive attribute} (or protected attribute). Examples include race, gender, age, or SES. When learning from a historical data, we may believe that one of these attributes affected the observable treatments and outcomes, resulting in a biased dataset. \citet{lum2016predict} give an example in the domain of predictive policing of how a dataset of drug crimes may become biased with respect to race through unfair policing practices. They note that it is impossible to collect a dataset of all drug crimes in some area; rather, these datasets are really tracking drug \emph{arrests}. Due to a higher level of police presence in heavily Black than heavily White communities, recorded drug arrests will by nature over-represent Black communities. Therefore, a predictive policing algorithm which attempts to fit this data will continue the pattern of over-policing Black communities. \citet{lum2016predict} provide experimental validation of this hypothesis through simulation, contrasting the output of a common predictive policing algorithm with independent, demographic-based estimates of drug use by neighborhood. Their work shows that wrongly specifying a learning problem as one of supervised classification can lead to replicating past biases. In order to account for this in the learning process, we should be aware of the biases which shaped the data --- which may include sensitive attributes that historically affected the treatment and/or outcome. Using the above example for concreteness, we specify the variables at play. The decision-making problem is: should police be sent to neighborhood $X$ at a given time? The variables are: \begin{itemize} \item $A \in \{0, 1\}$: a sensitive attribute. For example the majority race of a neighborhood. \item $T \in \{0, 1\}$: a treatment. For example the presence or absence of police in a certain neighborhood on a particular day. \item $Y \in \mathds{R}$: an outcome. For example the number of arrests recorded in a given neighborhood on a particular day. \item $X \in \mathds{R}^D$: $D$-dimensional observed features. For example statistics about the neighborhood, which may change day-to-day \end{itemize} We will represent sensitive attributes and treatments as binary throughout this paper; we recognize this is not always an optimal modeling choice in practice. Note that the choice of treatment will causally alter the outcome---an arrest cannot occur if there are no police in the area. Furthermore, the sensitive attribute can causally effect the outcome as well; research has shown that policing can disparately effect various races, even controlling for police presence \cite{gelman2007analysis} (the treatment in this case). We note that in various domains, there may be more variables of interest than the ones we list here, and more appropriate causal models than those shown in Fig. \ref{fig:cevae}. However, we believe that the setup we describe is widely applicable and contains the minimal set of variables to be useful for fairness-aware causal analysis. We are interested in calculating causal effects between the above variables. In particular, we seek answers to the following three questions: \paragraph{What is the effect of the treatment on the outcome?} This will help us to understand which $T$ is likely to produce a favorable outcome for a given $X$. Let us denote~$y_{T=t}(x, a) = \mathds{E}[y | do(T = t), X = x, A = a]$~as the expected conditional outcome under $T = t$, that is, the ground truth value taken by $Y$ when the treatment $T$ is assigned the value $t$, and conditioning on the values $x, a$ for the features and sensitive attribute respectively. Then, we can express the individual effect of $T$ on $Y$ ($IE_{T \rightarrow Y}$) as \begin{equation} IE_{T \rightarrow Y}(x, a) = y_{T=1}(x, a) - y_{T=0}(x, a). \end{equation} \paragraph{What is the effect of the sensitive attribute on the treatment?} This allows us to understand how the treatment assignment was biased in the data. Similarly, we can define $t_{A=a}(x) = \mathds{E}[t | do(A = a), X = x]$, which is the expected conditional treatment in the historical data when the value $a$ is assigned to the sensitive attribute. Then, the individual effect of $A$ on $T$ can be expressed as \begin{equation} IE_{A \rightarrow T}(x) = t_{A=1}(x) - t_{A=0}(x). \end{equation} \paragraph{What is the effect of the sensitive attribute on the outcome?} This allows us to understand what bias is introduced into the historically observed outcome. We can also define $y_{A=a}(x) = \mathds{E}[y | do(A = a), X = x, T=t_{A=a}(x)]$ as the expected conditional outcome under $A = a$; the ground truth value of $Y$ conditioned on the features being $x$ if the sensitive attribute were assigned the value $a$, and the treatment $T$ were assigned the ground truth value $t_{A=a}(x)$. Then, we can express the individual effect of $A$ on $Y$ as \begin{equation} IE_{A \rightarrow Y}(x) = y_{A=1}(x) - y_{A=0}(x). \end{equation} {\color{black} \subsection{Intervening on Sensitive Attributes} There has been some disagreement around the notion of intervening on an immutable (or effectively immutable) sensitive attribute. \citet{holland1986statistics} argue that there is ``no causation without manipulation'' --- i.e. an attribute can never be a cause; only an experience undergone can be. Briefly stated, they argue that if the factual and counterfactual versions cannot be ``defined in principle, it is impossible to define the causal effect''. In a counterargument, \citet{marini1988causality} claim that a ``synthesis of intrinsic and extrinsic determination [provides] a more adequate picture of causal relations'' --- meaning that both externally imposed experiences (extrinsic) and internally defined attributes (intrinsic) are valid conceptual components of a theory of causation. We agree with this view --- that the notion of a causal effect of an immutable attribute is valid, and believe that it is particularly useful in a fairness context. Specifically pertaining to race, some argue it is possible to understand the causal effect of an immutable attribute in terms of the effects of more manipulable attributes (proxies). \citet{vanderweele2014causal} argue that, rather than interpreting a causal effect estimate of $A$ as a hypothetical randomized intervention on $A$, one can interpret it as a particular type of intervention on some other set of manipulable variables related to $A$ (under certain graphical and distributional assumptions on those variables). \citet{sen2016race} take a constructivist approach, and consider race to be composed of constituent parts, some of which \textit{can} be theoretically manipulated. They describe several experimental designs which could estimate the effects of immutable attributes. Another issue with intervening on sensitive attributes is that, since many are ``assigned at conception'', all observed covariates $X$ are post-treatment \citep{sen2016race} (as reflected in the design of our SCM in Fig. \ref{subfig:cevae-fair}). In statistical analysis, a frequent approach is to ignore all post-treatment variables to avoid introducing collider biases \citep{gelman2007analysis,king1994designing}. However, in our model, the purpose of the covariates is to deduce the true (unobserved) values of the latent $Z$ for that individual. Therefore, when conditioning on the observed covariates, correlation of $A$ and $Z$ is the objective, rather than an undesired side effect. This is the first step (``Abduction'') of computing counterfactuals (according to \citet{pearl2009causality}); we can think of this as adjusting for bias (of the sensitive attribute) in the $X$-generating process. } \section{Proposed Method} \label{sec:models} In this section we first conceptualize and describe our proposed causal model---depicted in Fig. \ref{fig:proposed}---then discuss the parameterization of the corresponding SCMs and learning procedure. {\color{black} A common causal modelling approach is to define a new SCM for each problem \citet{pearl2009causality}, taking advantage of domain specific knowledge for that particular problem. This stands in contrast to a classic machine learning (ML) approach, which aims to process data and draw conclusions as generally as possible, by automatically discovering patterns of correlation in the data. While the causal modelling approach is capable of detecting effects the ML approach cannot, the ML approach is attractive since it provides modularity, generality and a more automated data processing pipeline. In this work, we aim to interpolate between the two approaches by considering a single, general causal model for observational data. Our model contains what we argue are a minimal set of fairly general causal variables for discovering treatment effects and biases in the data-generation process, allowing us to interface causally with arbitrary data that fits the proposed structure. } Two features of our causal model are noteworthy. First is the explicit consideration of the sensitive attribute---a potential source of dataset bias---as a confounder, which causally affects both the treatment $T$ and the outcome $Y$. This contrasts with approaches from outside the fairness literature (e.g. \citep{louizos2017causal}, Fig. \ref{subfig:cevae}), which in a fairness setting (Fig. \ref{subfig:cevae-sens}) would treat potential sensitive attributes as equivalent to other observed features. Our model accounts for the possibility that a sensitive attribute may have causal influence on the observed features, treatments and outcomes and the historical process which generated them. It makes the sensitive attribute distinct from the other attributes of $X$, which we understand not as confounders but observed proxies. We can think of this as a causal modeling analogue of ``fairness through awareness''. By actively adjusting for causal confounding effects of sensitive attributes, we can build a model which accounts for the interplay between the treatment and outcome for both values of the sensitive attribute. The other noteworthy aspect of our model is the latent variable $Z$. Together, $Z$ and $A$ make up all the confounding variables. {\color{black} We note two important points about these confounders. Firstly, we clarify that the model class we propose (a latent Gaussian and a deep neural network), is not necessarily the definitive model of the confounders of $T$ and $Y$; however, it is a flexible one, with numerous applications in machine learning \citep{rezende2014stochastic}. Secondly, we note that causal inference and machine learning have different conventions around unobserved (i.e. latent) variables --- in causal inference, these variables are generally considered to be nameable objects in the world (e.g. SES, historical predjudice), whereas in machine learning they represent some unspecified (and perhaps abstract) structure in the data. Our $Z$ follows the machine learning convention. } As in \citet{louizos2017causal}, $Z$ represents all the unobserved confounding variables which effect the outcomes or treatments (other than $A$). The features $X$ can be seen as proxies (noisy observations) for the confounders ($Z, A$). Altogether, the endogenous variables in our model are $X$, $A$, $Z$, $T$, and $Y$. We also have exogenous variables $\epsilon_X, \epsilon_A, \epsilon_Z, \epsilon_T, \epsilon_Y$ (not shown), each the immediate parent of (only) their respective endogenous variable. The structural equations are: \begin{align}\label{eq:structural-functions} \nonumber &Z = f_Z(\epsilon_Z) &A = f_A(\epsilon_A) \\ \nonumber &X = f_X(Z, A, \epsilon_X) &T = f_T(Z, A, \epsilon_T) \\ &Y = f_Y(Z, A, T, \epsilon_Y) &\medspace \epsilon_V \sim P_V(\epsilon_V) \ \forall V \in \{Z, A ,X, T, Y\} \end{align} {\color{black} Since $Z$ does not necessarily refer to tangible objects in the world, it is reasonable that $Z \perp A$ in our model. This does not prevent a characteristic such as SES (which may be correlated with $A$) from being a confounder --- rather, $Z$ could represent the component of SES which is not based on $A$. Since both confounders are inputs to all other variables in the SCM, the model can learn to represent variables which \textit{are} based on $A$, (e.g. SES) as a joint distribution of $Z$ and $A$. } With this SCM in hand, we can estimate various interventional outcomes, if we know the values of $f_V \ \forall \ V \in \{Z, A ,X, T, Y\}$. For instance, we might estimate: \begin{equation} \label{eq:scm-estimations} \begin{aligned} \mathds{E} \left[ Y | Z=z, A=a, do(T=1) \right] &= \mathds{E}_{\epsilon_Y \sim P_Y(\epsilon_Y)} [f_Y(z, a, 1, \epsilon_Y)]\\ \mathds{E} \left[ Y | Z=z, do(A=1), do(T=1) \right] &= \mathds{E}_{\epsilon_Y \sim P_Y(\epsilon_Y)} [f_Y(z, 1, 1, \epsilon_Y)]\\ \mathds{E} \left[ Y | Z=z, do(A=1) \right] &= \\ \mathds{E}_{\epsilon_Y \sim P_Y(\epsilon_Y)} \mathds{E}_{\epsilon_T \sim P_T(\epsilon_T)} &[f_Y(z, 1, f_T(z, 1, \epsilon_T), \epsilon_Y)] \end{aligned} \end{equation} which are the expected values over outcomes of interventions on $T$, $T$ and $A$, and just $A$, respectively. However, the problem with the calculations in Eq. \ref{eq:scm-estimations} is that $Z$ is unobserved, so we cannot simply condition on its value. Rather, we observe some proxies $X$. Since the structural equations go the other direction --- $X$ is a function of $Z$, not the other way around --- inferring $Z$ from a given $X$ is a non-trivial matter. In summary, we need to learn two things: a generative model which can approximate the structural functions $f$, and an inference model which can approximate the distribution of $Z$ given $X$. Following the lead of \citet{louizos2017causal}, we use variational inference parametrized by deep neural networks to learn the parameters of both of these models jointly. In variational inference, we aim to learn an approximate distribution over the joint variables $P(Z, A, X, T, Y)$, by maximizing a variational lower bound on the log-probability of the observed data. As demonstrated in \citet{louizos2017causal}, the causal effects in the model become identifiable if we can learn this joint distribution. We extend their proof in Appendix \ref{app:proofs} to show identifiability holds when including the sensitive attribute in the model (as in Fig. \ref{subfig:cevae-fair}). {\color{black} We discuss here the identifiability condition from \citet{louizos2017causal}. Given some treatment $T$ and outcome $Y$, the classic ``no hidden confounders'' assumption asserts that the set of observed variables $O$ blocks all backdoor paths from $T$ to $Y$. \citet{louizos2017causal} weaken this: they assume that there is a set of confounding variables $Z = O_Z \cup U_Z$ such that $Z$ blocks all backdoor paths from $T$ to $Y$, where $O_Z$ are observed and $U_Z$ are unobserved. They claim that if we recover the full joint distribtuion of $p(Z, X, T, Y)$, then we can identify the causal effect $T \rightarrow Y$. However, this is only possible if we have sufficiently informative proxies $X$. While recovering the full joint distribution does not mean we have to measure every confounder, we do have to at least measure some proxy for each confounder. This is a weaker assumption, but not fully general. There may be confounding factors which cannot be inferred from the proxies $X$ --- in this case, our model will be unable to learn the joint distribution, and the causal effect will be unidentifiable. In this case, we are back to square one; our causal estimates may be inaccurate. Determining the exact fairness implications of this remains an open problem --- it would depend on which confounders were missing, and which proxies were already collected. A complicating factor is that testing for unconfoundedness is difficult, and usually requires making further assumptions \citep{tran2016model}. Therefore we might unintentionally make unfair inferences if we are unaware that we cannot infer all confounders. If we think this is the case, one solution is to collect more proxies. This provides an alternative motivation for the idea of increasing fairness by measuring additional variables \citep{chen2018my}. } To learn a generative model of the data which is faithful to the structural model defined in Eq. \ref{eq:structural-functions}, we define distributions $p$ which will approximate various conditional probabilities in our model. We model the joint probability assuming the following factorization: \begin{equation} \label{eq:joint-probability} \begin{aligned} P(Z, A, X, T, Y) = p(Z) p(A) p(X | Z, A) p(T | Z, A) p(Y | Z, A, T) \end{aligned} \end{equation} Each of these $p$ corresponds to an $f$ in Eq. \ref{eq:structural-functions} --- formally, $p(V = v | W = w) = P_{\epsilon_{V}}[f_V(W, \epsilon_V)]$ for an endogenous variable $V$ and subset of endogenous variables $W$, where $\{V\}, W \subset \{Z, A, X, T, Y\}$. For simplicity, we choose computationally tractable probability distributions for each conditional probability in Eq. \ref{eq:joint-probability}: \begin{equation} \label{eq:distributions} \begin{aligned} p(Z) &= \prod_{j=1}^{D_Z} \mathcal{N}(Z_{j} | 0, 1)\\ p(A) &= Bern(A|\pi_A)\\ p(X | Z, A) &= \prod_{j=1}^{D_X} \mathcal{N}(X_{j} | \mu_X(Z, A), \sigma_X^2(Z, A) \\ p(T | Z, A) &= Bern(T | \pi_T(Z,A)) \end{aligned} \end{equation} where $D_Z, D_X$ are the dimensionalities of $Z$ and $X$ respectively, and $\pi_A \in [0, 1]$ is the empirical marginal probability of $A=1$ across the dataset (if this is unknown, we could use a Beta prior over that distribution; in this paper we assume $A$ is observed for every example). For $p(Y | Z, A, T)$, we use either a Bernoulli or a Gaussian distribution, depending on if $Y$ is binary or continuous: \begin{equation} \label{eq:distributions-Y} \begin{aligned} p_{binary}(Y | Z, A, T) &= Bern(Y | \pi_Y(Z,A,T))\\ p_{cont}(Y | Z, A, T) &= \mathcal{N}(Y | \mu_Y(Z,A,T), \sigma^2_Y(Z,A,T)) \end{aligned} \end{equation} To flexibly model the potentially complex and non-linear relationships in the true generative process, we specify several of the distribution parameters from Eqs. \ref{eq:distributions} and \ref{eq:distributions-Y} as the output of a function $g_V$, which is realized by a neural network (or TARNet \citep{shalit2016estimating}) with parameters $\theta_V$. We parametrize the model of $X$ with neural networks $g_X^{\mu}, g_X^{\sigma}$: \begin{equation}\label{eq:reparamatrization-mlp} \begin{aligned} \mu_X(Z,A) &= g^{\mu}_X(Z, A) \\ \sigma_X^2(Z,A) &= \exp{2g^{\sigma}_X(Z, A)} \end{aligned} \end{equation} We use TARNets \cite{shalit2016estimating} (see Sec. \ref{sec:tarnets}) to parameterize the distributions over $T$ and $Y$. In our model, $A$ acts as the ``treatment'' for the TARNet that outputs $T$. Likewise $A$ and $T$ are joint treatments affecting $Y$ --- our $Y$ model can be seen as a \textit{hierarchical TARNet}, with one TARNet for each value of $A$, where each TARNet has an arm for each value of $T$. In all, this yields the following parametrization: \begin{equation}\label{eq:reparamatrization-tarnets} \begin{aligned} p_T(Z,A) &= (1 - A) \sigma(g^{A=0}_T(Z, A)) + A \sigma(g^{A=1}_T(Z, A)); \\ p_Y(Z,A,T) &= (1 - T) (1 - A) \sigma(g^{T=0,A=0}_Y(Z, A, T))\\ &\quad + T (1 - A)\sigma(g^{T=1,A=0}_Y(Z, A, T)) \\ &\quad + (1 - T) A \sigma(g^{T=0,A=1}_Y(Z, A, T)) \\ &\quad + T A \sigma(g^{T=1,A=1}_Y(Z, A, T)); \end{aligned} \end{equation} and the same for $\mu_Y(Z,A,T)$ and $\sigma_Y^2(Z,A,T)$; where $\sigma$ is the sigmoid function $\sigma(x) = \frac{1}{1 + \exp(-x)}$ and $g^{I=0}_R(V, I)$ are defined as in Sec. \ref{sec:tarnets}. We further define an inference model $q$, to determine the values of the latent variables $Z$ given observed $X, A$. This takes the form: \begin{equation} \label{eq:inference-model} \begin{aligned} q(Z | X, A) = \mathcal{N}(\mu_Z(X,A), \sigma_Z^2(X,A)) \end{aligned} \end{equation} where the normal distribution is reparametrized analogously to Eq. \ref{eq:reparamatrization-mlp} with networks $g_Z^{\mu}, g_Z^{\sigma}$. Since $A$ is always observed, we do not need to infer it, even though it is a confounder. We note that this is a different inference network from the one in \citet{louizos2017causal} --- we do not use the given treatments and outcomes in the inference model. We found it to be a simpler solution (no auxiliary networks necessary), and did not see a large change in performance. This is similar to the approach taken in \citet{parbhoo2018causal}. To learn the parameters of this model, we can maximize the expected lower bound on the log probability of the data (the ELBO), which takes the form below, which we note is also a valid ELBO to optimize for lower-bounding the conditional log-probability of the treatments and outcomes given the data. \begin{equation} \label{eq:elbo} \begin{aligned} \mathcal{L} = \sum_{i=1}^n &\mathds{E}_{q(z_i | x_i, a_i)} [\log{p(x_i | z_i, a_i)} + \log{p(t_i | z_i, a_i)} \\ &+ \log{p(y_i | z_i, a_i, t_i)} + \log{p(z_i)} - \log{q(z_i | x_i, a_i)}] \end{aligned} \end{equation} \section{Related Work} \label{sec:related-work} Our work most closely relates to the Causal Effect Variational Autoencoder \citep{louizos2017causal}. Some follow-up work is done by \citet{parbhoo2018causal}, who suggest a purely discriminative approach using the information bottleneck. Our model differs from this work in that they did not include a sensitive attribute in their model, and their model does not contain a ``reconstruction fidelity'' term, in this case $\log{p(x_i | z_i, a_i)}$. Previous papers which learn causal effects using deep learning (with all confounders observed) include \citet{shalit2016estimating} and \citet{johansson2016learning}, who propose TARNets as well as some form of balancing penalty. The intersection of fairness and causality has been explored recently. Counterfactual fairness --- the idea that a fair classifier is one which doesn't change its prediction under the counterfactual value of $X$ when $A$ is flipped --- is a major theme \citep{kusner2017counterfactual}. Criteria for fairness in treatments are proposed in \citet{nabi2018fair}, and fair interventions are further explored in \citet{kusner2018causal}. \citet{zhang2018fairness} present a decomposition which provides a different way of understanding of unfairness in a causal inference model. Other work focuses on the causal relationship between sensitive attributes and proxies in fair classification \citep{kilbertus2017avoiding}. \citet{kallus2018residual} explore the idea of learning from biased data, making the point that a ``fair'' predictor learned on biased data may not be fair under certain forms of distributional shift, while not touching on causal ideas. Some conceptually similar work has looked at the ``selective labels'' problem \citep{lakkaraju2018selective,dearteaga2018learning}, where only a biased selection of the data has labels available. There has also been related work on \textit{feedback loops} in fairness, and the idea that past decisions can affect future ones, in the predictive policing \citep{lum2016predict,pmlr-v81-ensign18a} and recommender systems \citep{hashimoto2018fairness} contexts, for example. \citet{barabas2017interventions} advocate for understanding many problems of fair prediction as ones of intervention instead. Another variational autoencoder-based fairness model is proposed in \citet{louizos2015variational}, but with the goal of fair representation learning, rather than causal modelling. \citet{dwork2012fairness} originated the term ``fairness through awareness'', and argued that the sensitive attribute needed to be given a place of privilege in modelling in order to reduce unfairness of outcomes. \section{Experiments} \label{sec:experiments} In this section we compare various methods for causal effect estimation. The three effects we are interested in are \iffalse \begin{wraptable}{r}{7cm} \caption{A wrapped table going nicely inside the text.}\label{wrap-tab:1} \begin{tabular}{ccc}\\\toprule Shorthand & Description & Individual-level quantity \\\midrule $A \rightarrow T$ &The effect of $A$ on $T$ & $P(T = 1 | do(A = 1), X) - P(T = 1 | do(A = 0), X)$ \\ \midrule 2 &3 & 5\\ \midrule 2 &3 & 5\\ \bottomrule \end{tabular} \end{wraptable} \fi \begin{itemize} \item $A \rightarrow T$, the causal effect of $A$ on $T$: \begin{equation*} \mathds{E}(T = 1 | do(A = 1), X) - \mathds{E}(T = 1 | do(A = 0), X) \end{equation*} \item $A \rightarrow Y$, the causal effect of $A$ on $Y$: \begin{equation*} \mathds{E}(Y = 1 | do(A = 1), X) - \mathds{E}(Y = 1 | do(A = 0), X) \end{equation*} \item $T \rightarrow Y$, the causal effect of $T$ on $Y$: \begin{equation*} \mathds{E}(Y = 1 | do(T = 1), X, A) - \mathds{E}(Y = 1 | do(T = 0), X, A) \end{equation*} \end{itemize} Note that all three effects are individual-level; that is, they are conditioned on some observed $X$ (and possibly $A$), and then averaged across the dataset. \subsection{Data} We evaluate our model using semi-synthetic data. The evaluation of causal models using non-synthetic data is challenging, since a random control trial on the intervention variable is required to validate correctness --- this is doubly true in our case, where we are concerned with two different possible interventions. Additionally, while data from random control trials for treatment variables exists (albeit uncommon), conducting a random control trial for a sensitive attribute is usually impossible. We have adapted the IHDP dataset \citep{multisite1990enhancing,brooks-gunn_liaw_klebanov_1994}---a standard semi-synthetic causal inference benchmark---for use in the setting of causal effect estimation under a sensitive attribute. The IHDP dataset is from a randomized experiment run by the Infant Health and Development Program (in the US), which "targeted low-birth-weight, premature infants, and provided the treatment group with both intensive high-quality child care and home visits from a trained provider" \cite{hill2011bayesian}. Pre-treatment variables were collected from both child (e.g. birth weight, sex) and the mother at time of birth (e.g. age, marital status) and behaviors engaged in during the pregnancy (e.g. smoked cigarettes, drank alcohol), as well as the site of the intervention (where the family resided). We choose our sensitive attribute to be mother's race, binarized as White and non-White. We follow a similar method for generating outcomes to the Response B surface proposed in \citet{hill2011bayesian}. However, our setup differs since we are interested in additionally modelling a sensitive attribute and hidden confounders, so there are three more steps which must be taken. First, we need to generate outcomes $Y$ for each example for $T \in \{0, 1\}$ under the counterfactual sensitive attribute $A$. Second, we need to generate a treatment assignment for each example for the counterfactual value of the sensitive attribute. Finally, we need to remove some data from the observable measurements to act as a hidden confounder, as in \citet{louizos2017causal}. We detail our full data generation method in Appendix \ref{app:data-generation}. We denote the outcome $Y$ under interventions $do(T=t), do(A=a$) as $y_{T=t,A=a}$. The subroutines in Algorithms \ref{alg:ihdp-response} and \ref{alg:ihdp-treatment} generate all factual and counterfactual outcomes and treatments for each example, one for each possible setting of $A$ and/or $T$. Values of the constants that we use for data generation can be found in Appendix \ref{app:data-generation}. We choose our hidden confounding feature $Z$ to be birth weight. In the second (optional) step of data generation, we choose to remove 0, 1, or 2 other features. Especially if we choose features which are highly correlated with the hidden confounder, this has the effect of making the estimation problem more difficult. When removing 0 features, we do nothing. When removing 1 feature, we remove the feature which is most highly correlated with $Z$ (head size). When removing 2 features, we remove two features most highly correlated with $Z$ (head size \& weeks born preterm). \subsection{Experimental Setup} \label{sec:exp-setup} We run four different models for comparison, including the one we propose. Since we are interested in estimating three different causal effects simultaneously ($A \rightarrow T, A \rightarrow Y, T \rightarrow Y$), we cannot compare against most standard causal inference benchmark models for treatment effect estimation. The models we test are the following: \begin{itemize} \item \textbf{Counterfactual MLP (CFMLP)}: a multilayer perception (MLP) which takes the treatment and sensitive attribute as input, concatenated to $X$, and aims to predict outcome. Counterfactual outcomes are calculated by simply flipping the relevant attributes and re-inputting the modified vector to the MLP. A similar auxiliary network learns to predict the treatment from a vector of $X$ concatenated to $A$. \item \textbf{Counterfactual Multiple MLP (CF4MLP)}: a set of four MLPs --- one for each combination of $(A, T) \in \{0, 1\}^2$. Examples are inputted into the appropriate MLP for the factual outcome, and simply inputted into another MLP for the appropriate counterfactual outcome. A similar pair of auxiliary networks predict treatment. \item \textbf{Causal Effect Variational Autoencoder with Sensitive Attribute (CVAE-A, Fig. \ref{subfig:cevae-sens})}: a model similar to \citet{louizos2017causal}, but with the simpler inference model we propose. We incorporate a sensitive attribute $A$ by concatenating $X$ to $A$ as input; counterfactuals along $A$ are taken by flipping $A$ and re-inputting the modified vector. Counterfactuals along $T$ are taken as in \citet{louizos2017causal}. \item \textbf{Fair Causal Effect Variational Autoencoder (FCVAE, Fig. \ref{subfig:cevae-fair})}: our proposed fair-aware causal model, with $A$ concatenated to $Z$ as confounders. We run two versions: one where $A$ is used to help with reconstructing $X$ and inferring $Z$ from $X$ (FCVAE-1), and one where it is not (FCVAE-2). Formally, the inference model and generative model of $X$ in FCVAE-1 are $q(Z | X, A)$ and $p(X | Z, A)$, and in FCVAE-2 are $q(Z | X)$ and $p(X | Z)$ respectively. In both versions, $A$ is a confounder of both the treatment and the outcome. \end{itemize} The CFMLP is purely a classification baseline. It learns a mapping from input to output, estimating the conditional distribution $P(Y | X, A, T)$. The CF4MLP shares this goal, but has a more complex architecture---it learns a disjoint set of parameters for each setting of interventions, allowing it to model completely separate generative processes. However, it is still ultimately concerned with supervised prediction. Furthermore, neither of these models is built to consider the impact of hidden confounders. The CVAE-A is a model for causal inference of outcomes from treatments. Therefore, we should expect it to perform well in estimating $T \rightarrow Y$. It is also created to model these effects under hidden confounders. Therefore, the difference between CVAE-A and the MLPs will tell us the improvement which comes from appropriate causal modelling rather than classification. However, the CVAE-A does not consider the sensitive attribute $A$ as a confounder; rather, it treats it simply as another covariate of $X$. So in comparing the FCVAE to the CVAE-A, we observe the improvement that comes from causally modelling the dataset unfairness stemming from a sensitive attribute. In comparing the FCVAE to the MLPs, we observe the full impact of the FCVAE --- joint causal modelling of treatments, outcomes, sensitive attributes, and hidden confounders. See Appendix \ref{app:experiments} for experimental details. \subsection{Results} \label{sec:results} \subsubsection{Estimating Causal Effects} In this section, we evaluate how well the models from Sec. \ref{sec:exp-setup} can estimate the three causal effects $A \rightarrow T, A \rightarrow Y, T \rightarrow Y$. To avoid confusion with the words \textit{treatment} and \textit{outcome}, in each of these three causal interactions, we will refer to to the causing variable as the \textit{intervention} variable, and the affected variable as the \textit{result} variable. To evaluate how well our model can estimate causal effects, we use PEHE: Precision in Estimation of Heterogeneous Effects \cite{hill2011bayesian}. This is calculated as: $PEHE = \sqrt[]{\mathds{E}[( (r_1 - r_0) - (\hat{r}_1 - \hat{r}_0)) ^2]}$, where $r_i$ is the ground truth value of result from the intervention $i$, and $\hat{r}_i$ is our model's estimate of that quantity. PEHE measures our ability to model both the factual (ground truth) and the counterfactual results. \begin{table}[] \begin{tabular}{llll} \hline Model & A $\rightarrow$ T & T $\rightarrow$ Y & A $\rightarrow$ Y \\ \hline CFMLP & 0.681 $\pm$ 0.00 & 4.51 $\pm$ 0.13 & 3.28 $\pm$ 0.07 \\ CF4MLP & 0.667 $\pm$ 0.00 & 4.58 $\pm$ 0.13 & 3.71 $\pm$ 0.09 \\ CVAE-A & 0.665 $\pm$ 0.00 & \textbf{3.80} $\pm$ 0.10 & 3.04 $\pm$ 0.06 \\ FCVAE-1 & \textbf{0.659} $\pm$ 0.00 & \textbf{3.82} $\pm$ 0.11 & \textbf{2.88}$\pm$ 0.06 \\ FCVAE-2 & \textbf{0.659} $\pm$ 0.00 & \textbf{3.81} $\pm$ 0.11 & \textbf{2.78 }$\pm$ 0.06 \\ \hline \end{tabular} \caption{PEHE for each model on IHDP data (no extra features removed). Mean and standard errors shown, as calculated over 500 random seedings.} \label{table:ihdp-pehe-0} \end{table} \begin{table}[] \begin{tabular}{llll} \hline Model & A $\rightarrow$ T & T $\rightarrow$ Y & A $\rightarrow$ Y \\ \hline CFMLP & 0.675 $\pm$ 0.00 & 4.30 $\pm$ 0.11 & 3.42 $\pm$ 0.08 \\ CF4MLP & \textbf{0.661} $\pm$ 0.00 & 4.37 $\pm$ 0.11 & 3.89 $\pm$ 0.07\\ CVAE-A & 0.672 $\pm$ 0.00 & \textbf{4.05} $\pm$ 0.10 & 3.53 $\pm$ 0.07 \\ FCVAE-1 & 0.663 $\pm$ 0.00 & \textbf{4.00} $\pm$ 0.10 & \textbf{3.39} $\pm$ 0.08 \\ FCVAE-2 & 0.663 $\pm$ 0.00 & \textbf{3.99} $\pm$ 0.10 & \textbf{3.25} $\pm$ 0.07 \\ \hline \end{tabular} \caption{PEHE for each model on IHDP data (1 most informative feature removed). Mean and standard errors shown, as calculated over 500 random seedings.} \label{table:ihdp-pehe-1} \end{table} \begin{table}[] \begin{tabular}{llll} \hline Model & A $\rightarrow$ T & T $\rightarrow$ Y & A $\rightarrow$ Y \\ \hline CFMLP & 0.666 $\pm$ 0.00 & 6.03 $\pm$ 0.21 & 4.30 $\pm$ 0.12 \\ CF4MLP & \textbf{0.659} $\pm$ 0.00 & 5.77 $\pm$ 0.18 & 4.59 $\pm$ 0.10 \\ CVAE-A & 0.672 $\pm$ 0.00 & \textbf{5.46} $\pm$ 0.18 & 4.19 $\pm$ 0.10 \\ FCVAE-1 & \textbf{0.659} $\pm$ 0.00 & \textbf{5.40} $\pm$ 0.18 & \textbf{4.07}$\pm$ 0.11 \\ FCVAE-2 & \textbf{0.659} $\pm$ 0.00 & \textbf{5.39} $\pm$ 0.18 & \textbf{3.95} $\pm$ 0.10 \\ \hline \end{tabular} \caption{PEHE for each model on IHDP data (2 most informative features removed). Mean and standard errors shown, as calculated over 500 random seedings. Lower is better.} \label{table:ihdp-pehe-2} \end{table} In Tables \ref{table:ihdp-pehe-0}-\ref{table:ihdp-pehe-2}, we show the PEHE for each of the models described in Sec. \ref{sec:exp-setup}, for each causal effect of interest. Each table shows results for a version of the dataset with 0-2 of the most informative features removed (as measured by correlation with the hidden confounder). Therefore, the easiest problem is with zero features removed, the hardest is with two. Note that in IHDP, $Y \in \mathds{R}$. Generally, as expected, we observe that the causal models achieve lower PEHE for most estimation problems. Also as expected, we observe that that the PEHE for the more complex estimation problems ($A \rightarrow Y, T \rightarrow Y$) increases as the most useful proxies are removed from the data. We suspect there is less variation in the results for $A \rightarrow T$ since it is a simpler problem: there are no extra confounders (other than $Z$) or mediating factors to consider. We find that our model (the FCVAE) compares favorably to the other models in this experiment. We see that in general, the fair-aware models (FCVAE-1 and FCVAE-2) have lower PEHE than all other models when estimating the causal effects relating to the sensitive attribute ($A \rightarrow Y, A \rightarrow T$). Furthermore, the FCVAE also performs similarly to the CVAE-A at $T \rightarrow Y$ estimation as well, demonstrating a slight improvement (at least in the more difficult 1, 2 features removed cases). One interesting note is that FCVAE-1 (where $A$ is used in reconstruction of $X$ and in inference of $Z$) and FCVAE-2 seem to perform similarly, with FCVAE-2 being slightly better, if anything. This may seem surprising at first, since one might imagine that using $A$ would allow the model to learn better representations of $X$, particularly for the purpose of doing counterfactual inference across $A$. To explore this further, we examine in table \ref{table:ihdp-mizx} the latent representations $Z$ learned by each model in terms of their encoder mutual information between $Z$ and $X$, which is calculated as $KL(q(Z | X) || p(Z))$, the KL-divergence from the encoder posterior to the prior. This quantity is roughly the same for both versions of the FCVAE, implying that the inference network $q(Z|\cdot)$ does not leverage the additional information provided by $A$ in its latent code $Z$. This is in fact sensible because FCVAE has access to $A$ as an observed confounder in modeling the structural equations. We also noticed that CVAE contains about one bit of extra information in its latent code, implying some degree of success in capturing relevant information about $A$ in $Z$. But if CVAE models all confounders during inference, why does it underperform relative to FCVAE estimating the downstream causal effects, especially $A \rightarrow Y$? By making explicit the role of $A$ as confounder, we hypothesize that FCVAE can learn the interventional distributions with respect to $A$ (e.g., $p(Y|T, do(A=a), Z))$ rather than the conditional distributions of CVAE (e.g., $p(Y|T, Z(A)))$; we suspect that the gating mechanism of the TARNet implementation of the structural equations to be important in this regard. \begin{table}[] \begin{tabular}{ll} \hline Model & $KL \left[ q(z|\cdot) || p(z) \right] $ \\ \hline CVAE-A & 4.28 $\pm$ 0.10 \\ FCVAE-1 & 3.50 $\pm$ 0.12 \\ FCVAE-2 & 3.53 $\pm$ 0.12 \\ \hline \end{tabular} \caption{KL divergence from the encoder posterior $q(z|\cdot)$ to prior $p(z)$ after training on IHDP; equivalent to encoder mutual information \cite{alemi2018fixing}. CVAE and FCVAE-1 use $(X, A)$ as input to encoder, while FCVAE-2 uses $X$ only. Mean and standard errors shown, as calculated over 500 random seedings.} \label{table:ihdp-mizx} \end{table} \subsubsection{Learning a Treatment Policy} \label{sec:treatment-policy} The next natural question is: how does estimating these causal effects contribute to a fair decision-making policy? We examine two dimensions of this. We define a \textit{policy} $\pi: X, A \rightarrow T$ as a function which maps inputs (features and sensitive attribute) to treatments. We suppose the goal is to assign treatments $T$ using a policy $\hat T = \pi(X, A)$ that maximizes its expected \textit{value} $V(\pi)$, defined here as the expected outcome $Y$ it achieves over the data, i.e. $V(\pi) = \mathds{E}_{x, a}[Y | do(T=\pi(x, a)), A=a, X=x]$. For example, we could imagine the treatments to be various medications, and the outcome to be some health indicator (e.g. number of months survived post-treatment). We can derive a policy from an outcome prediction model simply by outputting the predicted argmax value over treatments, i.e. $\pi(x, a) = \argmax_{t \in T} \mathds{E}_{\hat{Y}}[\hat{Y} | do(T=t), A=a, X=x]$, where $\hat{Y}$ is the model's prediction of the true outcome $Y$. The optimal policy $\pi^\star(x, a )= \argmax_{t \in T} \mathds{E}_Y[Y | do(T=t), A=a, X=x]$ takes the argmax over ground truth outcomes every time. First, we look at the mean \textit{regret} of the policy $\pi$, which is the difference between its achieved value and the the value of the optimal policy: $R(\pi) = V(\pi^\star) - V(\pi)$. {\color{black} We note that in general, a policy's regret is not easy to compute or bound without assumptions on the outcome distribution in the data. } In Table \ref{table:ihdp-regret}, we display the expected regret values for the learned policies. We observe that the fair-aware model achieves lower regret than the unaware causal model, and much lower regret than the non-causal models, for both the easier and more difficult settings of the IHDP data. \begin{table}[] \begin{tabular}{llll} \hline Model & 0 removed & 1 removed & 2 removed \\ \hline CFMLP & 0.37 $\pm$ 0.02 & 0.42 $\pm$ 0.02 & 0.81 $\pm$ 0.04 \\ CF4MLP & 0.31 $\pm$ 0.02 & 0.43 $\pm$ 0.02 & 0.59 $\pm$ 0.02 \\ CVAE-A & 0.21 $\pm$ 0.01 & 0.38 $\pm$ 0.01 & 0.59 $\pm$ 0.02 \\ FCVAE-1 & \textbf{0.19} $\pm$ 0.01 & \textbf{0.36} $\pm$ 0.01 & \textbf{0.55} $\pm$ 0.02 \\ FCVAE-2 & \textbf{0.19} $\pm$ 0.01 & \textbf{0.35} $\pm$ 0.01 & \textbf{0.55} $\pm$ 0.02\\ \hline \end{tabular} \caption{Regret for each model's policy on IHDP data with 0, 1, or 2 of the most useful covariates removed. Mean and standard errors shown, as calculated over 500 random seedings. Lower regret is better.} \label{table:ihdp-regret} \end{table} Next, we attempt to measure the policy's fairness. Most fairness metrics are designed for evaluating classification, not for intervention. However, \citet{chen2018my} explore an idea which is easily adjusted to the interventional setting: that an algorithm is unfair if it is much less accurate on one subgroup. Here, we adapt this notion to evaluate treatment policy fairness. For any $x$, let us say the policy $\pi$ is \textit{accurate} if it chooses the treatment which in fact yields the best outcome for that individual; i.e. if $\pi(x, a) = \pi^\star(x, a)$. We can define the \textit{accuracy} of the policy $Acc(\pi) = \mathds{E}_{x, a}[\mathds{1}(\pi(x, a) = \pi^\star(x, a))]$, where $\mathds{1}$ is an indicator function. We can define the subgroup accuracy $Acc_{\alpha}$ as accuracy calculated while conditioning (not intervening) on a particular value $\alpha$ of $A$: $Acc_\alpha(\pi) = \mathds{E}_{x | A =\alpha}[\mathds{1}(\pi(x, \alpha) = \pi^\star(x, \alpha))]$. We condition rather than intervene on $A$ here since we are interested in measuring the impact of the policy on real, existing populations, rather than hypothetical ones. Finally, to evaluate the fairness of the policy, we can look at the \textit{accuracy gap}: $| Acc_1(\pi) - Acc_0(\pi) |$. If this is high, the model is more unfair, since the policy has been more successful at modelling one group than the other, and is much more consistently choosing the correct treatment for individuals in that group. In Table \ref{table:ihdp-acc-gap} we display the accuracy gaps for our models and baselines on the IHDP dataset. We observe that the FCVAE achieves a smaller accuracy gap than those which do not consider the effect of the sensitive attribute. This is an encouraging sign that by understanding the confounding influence of sensitive attributes in biasing historical datasets, we can learn treatment policies which are more accurate for all subgroups of the data. \begin{table}[] \begin{tabular}{llll} \hline Model & 0 removed & 1 removed & 2 removed \\ \hline CFMLP & 0.042 $\pm$ 0.002 & 0.033 $\pm$ 0.002 & 0.062 $\pm$ 0.002 \\ CF4MLP & 0.034 $\pm$ 0.002 & 0.038 $\pm$ 0.002 & 0.054 $\pm$ 0.002 \\ CVAE-A & 0.033 $\pm$ 0.001 & \textbf{0.028} $\pm$ 0.001 & 0.051 $\pm$ 0.002 \\ FCVAE-1 & \textbf{0.031} $\pm$ 0.001 & \textbf{0.028} $\pm$ 0.001 & \textbf{0.046} $\pm$ 0.001 \\ FCVAE-2 & \textbf{0.030} $\pm$ 0.001 & \textbf{0.027} $\pm$ 0.001 & \textbf{0.047} $\pm$ 0.001 \\ \hline \end{tabular} \caption{Accuracy gap for each model's policy on IHDP data with 0, 1, or 2 of the most useful covariates removed. Mean and standard errors shown, as calculated over 500 random seedings. Lower gap is more fair.} \label{table:ihdp-acc-gap} \end{table} \section{Discussion} \label{sec:conclusions} In this paper, we proposed a causally-motivated model for learning from potentially biased data. We emphasize the importance of modeling the potential confounders of historical datasets: we model the sensitive attribute as an observed confounder contributing to dataset bias, and leverage deep latent variable models to approximately infer other hidden confounders. In Sec. \ref{sec:treatment-policy}, we demonstrated how to use our model to learn a simple treatment policy from data which assigns treatments more accurately and fairly than several causal and non-causal baselines. Looking forward, the estimation of sensitive attribute causal effects suggests several compelling new research directions, which we non-exhaustively discuss here: \begin{itemize} \item \textbf{Counterfactual Fairness:} Our model learns outcomes for counterfactual values of both $T$ and $A$. This means we could choose to implement a policy where we assess everyone under the same value $a'$, by assigning treatments to all individuals, no matter their original value $a$ of $A$, based on the inferred outcome distribution $P(Y | do(A=a'), X, T)$. Such a policy respects the definition of \textit{counterfactual fairness} proposed by \citet{kusner2017counterfactual}, which requires invariance to counterfactuals in $A$ at the individual level. \item \textbf{Path-Specific Effects:} Our model allows us to decompose $A \rightarrow Y$ into direct and indirect effects through mediation analysis of $T$ \citep{robins1992identifiability}. By estimating this decomposition, we could learn a policy which respects \textit{path-specific} fairness, as proposed by \citet{nabi2018fair}. \item \textbf{Analyzing Historical Bias:} Estimating causal effects between $A$, $T$, and $Y$ permits for the analysis and comparison of bias in historical datasets. For instance, the effect $A \rightarrow T$ is a measure of bias in a historical policy, and the effect $A \rightarrow Y$ is a measure of bias in whatever system historically generated the outcome. This could serve as the basis of a \textit{bias auditing technique} for data scientists. \item \textbf{Data Augmentation:} The absence of data (especially not-at-random) has strong implications for downstream modeling in both fairness \citep{kallus2018residual} and causal inference \citep{rubin1976inference}. Our model outputs counterfactual outcomes for both $A$ and $T$, which could be used for \textit{fair missing data imputation} \citep{van2018flexible,sterne2009multiple}. This could in turn enable the application of simpler methods like supervised learning to interventional problems. \item \textbf{Fair Policies Under Constraints:} In this paper, we consider an approach to fairness where understanding dataset bias is paramount, rather than the more common fairness-accuracy constraint-based tradeoff \citep{hardt2016equality,menon2017cost}. However, in some domains we may be interested in policies which satisfy a \textit{fairness constraint} (e.g., the same distribution of treatments are given to each group). Estimating the underlying causal effects would be useful for constrained policy learning. \item \textbf{Incorporating Prior Knowledge:} Graphical models (both probabilistic and SCM) permit the specification of \textit{prior knowledge} when modeling data, and provide a framework for inference that balances these beliefs with evidence from the data. This is a powerful fairness idea---we may believe a priori that a dataset \textit{should} look a certain way if not for some bias. In the context of a fair machine learning pipeline that considers many datasets, this relates to the AutoML task of learning distributions over datasets that share global parameters \cite{edwards2016towards}. \end{itemize} In automated decision making, the focus on intervention over classification \cite{barabas2017interventions} suggests the more equitable deployment of machine learning when only biased data are available, but also raises significant technical challenges. We believe causal modeling to be an invaluable tool in addressing these challenges, and hope that this paper contributes to the discussion around how best to understand and make predictions from existing datasets without replicating existing biases.

43dc7520a7c3f650f8d382f5856b75347a3384b0

\section{The variable-order time fractional advection-diffusion equation} \section{The modeling equation and time discretization approximation} \label{timeSec} The variable-order time fractional advection-diffusion equation \cite{TAYEBI2017655} can be formulated as \begin{equation} _0^cD_t^{\alpha({\bf x},t)}u({\bf x},t)=\kappa({\bf x},t)\Delta u({\bf x},t)-{\bf v}({\bf x},t)\cdot\nabla u({\bf x},t)+f({\bf x},t),\quad {\bf x}=(x,y) \in\Omega\subset \mathbb{R}^2,\quad t>0, \label{tfade} \end{equation} subject to the following general initial and boundary conditions \begin{equation}\ba{l} u({\bf x},0)=g({\bf x}),\quad {\bf x} \in \Omega,\\ \mathbb{B}u({\bf x},t)=h({\bf x},t), \quad {\bf x} \in \partial\Omega,\quad t>0, \ea \label{initialBoundaryCon} \end{equation} where $\Delta u=\pxx u+\pyy u$, $\nabla u=(\px u, \py u)$, and $\mathbb{B}$ denotes the operator with Dirichlet or Neumann boundary conditions. $\Omega$ is a bounded domain in $\mathbb{R}^2$, $\partial \Omega$ is the boundary of $\Omega$, $\kappa({\bf x},t)>0$ is the diffusion coefficient function, {\bf v}({\bf x},t) is advection velocity, and $f({\bf x},t)$, $g({\bf x})$, $h({\bf x},t)$ are given functions. In Equation (\ref{tfade}), the variable-order time fractional derivative $\fracD u$ with $0<\alpha({\bf x},t)\leq 1$ in the Caputo definition as \begin{equation} \fracD u=\left\{ \ba{l} \frac{1}{\Gamma(1-\alpha({\bf x},t))}\int_0^t\frac{1}{(t-\xi)^{\alpha({\bf x},t))}} \frac{\partial u({\bf x},\xi)}{\partial \xi} d\xi,\quad 0<\alpha({\bf x},t)<1, \\ \frac{\partial u({\bf x},t)}{\partial t},\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \alpha({\bf x},t)=1. \ea \right.\label{CapDef} \end{equation} In this study, we focus on the RBF-DQ method for the variable-order time fractional advection-diffusion equation on complex geometries with different boundary conditions. First, in this section, we give the time discretization approximation. Use the time discretization method proposed in the paper \cite{TAYEBI2017655} for Caputo definition, with the notation $b_j({\bf x},t_{k+1})=(j+1)^{1-\alpha ({\bf x},t_{k+1})}-j^{1-\alpha ({\bf x},t_{k+1})},$ we have \begin{equation}\ba{l} \fracdd u({\bf x},t_{k+1})=\frac{1}{\Gamma(1-\alpha({\bf x},t_{k+1}))} \sum\limits_{j=0}^{k}\int_{t_{j}}^{t_{j+1}} \frac{1}{(t_{k+1}-\xi)^{\alpha({\bf x},t_{k+1})}} \frac{\partial u({\bf x},\xi)}{\partial \xi} d\xi,\\ \qquad \qquad \qquad \qquad=\frac{1}{\Gamma(1-\alpha({\bf x},t_{k+1}))} \sum\limits_{j=0}^{k}\frac{ u({\bf x},t_{j+1}) - u({\bf x},t_{j}) }{\Delta t} \int_{t_{j}}^{t_{j+1}}\frac{1}{(t_{k+1}-\xi)^{\alpha({\bf x},t_{k+1})}} d\xi +R_{k+1},\\ \qquad \qquad \qquad \qquad=\frac{\Delta t^{-\alpha({\bf x},t_{k+1})}}{\Gamma(2-\alpha({\bf x},t_{k+1}))}\sum\limits_{j=0}^{k} b_{k-j}({\bf x}, t_{k+1})(u({\bf x},t_{j+1}) - u({\bf x},t_{j}))+R_{k+1} \\ \qquad \qquad \qquad \qquad=\frac{\Delta t^{-\alpha({\bf x},t_{k+1})}}{\Gamma(2-\alpha({\bf x},t_{k+1}))}\sum\limits_{j=0}^{k} b_{j}({\bf x}, t_{k+1})(u({\bf x},t_{k-j+1}) - u({\bf x},t_{k-j}))+R_{k+1} \ea \label{CapTimeDisc} \end{equation} where $t_k=k\Delta t$ for $k=0,1,\dots,M$ and $M=T/\Delta t$. The truncation error $R_{k+1}$ \cite{LIN20071533,TAYEBI2017655,Zhuang2011} is subjected to \begin{equation} |R_{k+1}|\leq C \Delta t ^{2-\alpha ({\bf x}, t_{k+1})}.\label{TruncErr} \end{equation} Substituting Equation (\ref{CapTimeDisc}) into Equation (\ref{tfade}), and with the $\theta$-weighted scheme ($\theta\in [0,1]$), we can obtain \begin{equation}\ba{l} u({\bf x},t_{k+1})-\theta\mu\xt{k+1}\left[\kappa \xt{k+1} \Delta u\xt{k+1} -{\bf v}\xt{k+1} \cdot \nabla u\xt{k+1}\right] \\ \qquad=u({\bf x},t_{k})+(1-\theta)\mu\xt{k+1}\left[\kappa \xt{k+1} \Delta u\xt{k} -{\bf v}\xt{k+1} \cdot \nabla u\xt{k}\right] \\ \qquad - \sum\limits_{j=1}^{k} b_{j}({\bf x}, t_{k+1})(u({\bf x},t_{k-j+1}) - u({\bf x},t_{k-j}))+\mu\xt{k+1} f\xt{k+1} +\overline{ R}_{k+1} +\overline{M}_{k+1}, \ea \label{thetaWeiSch} \end{equation} where \begin{equation} \mu\xt{k+1}=\Delta t ^{\alpha\xt{k+1}}\Gamma (2-\alpha({\bf x},t_{k+1})), \label{muFunc} \end{equation} and \begin{equation} |\overline{ R}_{k+1} |\leq \overline{C} \Delta t^2,\quad |\overline{ M}_{k+1} |\leq \overline{C}_1 (1-\theta)\Delta t^{1+\alpha\xt{k+1}}. \label{newTruncErr} \end{equation} Using the notations $u^k=u^k({\bf x})$ as the numerical approximation to $u\xt{k}$, $\mu ^k=\mu\xt{k}$, $\kappa^k=\kappa\xt{k},\ {\bf v}^k={\bf v}\xt{k}=(v_1\xt{k},v_2\xt{k}), \ b_j^k=b_j\xt{k}$ and $f^k=f\xt{k}$, then Equation (\ref{tfade}) can be discretized as follows \begin{equation}\ba{l} u^{k+1}-\theta\mu^{k+1}\left[\kappa^{k+1} \Delta u^{k+1} -{\bf v}^{k+1} \cdot \nabla u^{k+1}\right] \\ \qquad=u^k+(1-\theta)\mu^{k+1}\left[\kappa ^{k+1} \Delta u^{k} -{\bf v}^{k+1} \cdot \nabla u^{k}\right] - \sum\limits_{j=1}^{k} b_{j}^{k+1}(u^{k-j+1} - u^{k-j})+\mu^{k+1} f^{k+1} . \ea \label{timeFinalSch} \end{equation} For $\theta=1$, the Equation (\ref{timeFinalSch}) is the same as the scheme proposed by the paper \cite{Zhuang2011} for constant-order time fractional advection-diffusion equation. The equation (\ref{timeFinalSch}) can be rearranged as \begin{equation}\ba{l} u^{k+1}-\theta\mu^{k+1}\left[\kappa^{k+1} \Delta u^{k+1} -{\bf v}^{k+1} \cdot \nabla u^{k+1}\right] \\ \qquad=(1-b_1^{k+1})u^k+(1-\theta)\mu^{k+1}\left[\kappa ^{k+1} \Delta u^{k} -{\bf v}^{k+1} \cdot \nabla u^{k}\right] \\ \qquad\quad +\sum\limits_{j=1}^{k-1} (b_{j}^{k+1}-b_{j+1}^{k+1})u^{k-j}+b_k^{k+1}u^0+\mu^{k+1} f^{k+1} . \ea \label{timeFinalSch1} \end{equation} \begin{comment} \subsection{Stability and convergence} \label{stabConSubsec} Following the thought in \cite{Zhuang2011} for constant-order time fractional advection-diffusion equation with $\theta=1$, we use the following inner product \begin{equation} (v,w)=\int\int_{\Omega}v({\bf x})w(\bfx x)dxdy, \label{innProd} \end{equation} and norm in $L_2$ \begin{equation} \|v\|=[(v,v)]^{\frac 12}=\left[ \int\int_{\Omega}v^2({\bf x})dxdy \right]^{\frac 12} \label{l2norm} \end{equation} \end{comment} \section{RBF-based differential quadrature method} \label{fbfdqSec} \subsection{Basic RBF-based differential quadrature method} The idea of differential quadrature method came from the numerical integral, that any integral over a closed domain can be approximated by a linear weighted sum of all the functional values in the integral domain. In differential quadrature method, as the numerical integral, the derivative value $u^{(m)}({\bf x})$ at the centre ${\bf x}_i$ are approximated by a linear weighted sum of the function values at a set of nodes in a closed domain as \begin{equation} u^{(m)}_x({\bf x_i})\approx \sum\limits_{j=1}^N w_{ij}^{(m)} u({\bf x}_j), \quad u^{(m)}_y({\bf x_i})\approx \sum\limits_{j=1}^N \overline{w}_{ij}^{(m)} u({\bf x}_j), \label{grbfdq} \end{equation} for $i=1,\cdots,N,$ where $w_{ij}^{(m)},\ \overline{w}_{ij}^{(m)} $ are the weighting coefficients for derivatives of order $m$ with respect to $x$ and $y$, respectively. In RBF-DQ method, the weighting coefficients of $w_{ij}^{(m)},$ and $ \overline{w}_{ij}^{(m)} $ are determined by all the base functions as test function in Equation (\ref{grbfdq}) \cite{Wu2002Development,Shu2003Local}. There are many RBFs available. In this study, due to the better performance for the interpolation of 2D scattered data, multiquadrics (MQ) basis function is used as the test function. The function in the region of $\Omega$ can be locally approximated by MQ RBFs as \begin{equation} h(x,y)=\sum\limits_{j=1}^N\lambda_j \sqrt{(x-x_j)^2+(y-y_j)^2+c_j^2} +\lambda_{N+1}, \label{FuncIPbyMQ} \end{equation} with shape parameter $c_j$. To make the problem be well-posed, the equation \begin{equation} \sum\limits_{j=1}^N\lambda_j=0, \label{lweq} \end{equation} is enforced. Substituting Equation (\ref{lweq}) into equation (\ref{FuncIPbyMQ}) gives \begin{equation} h(x,y)=\sum\limits_{j=1,j\neq i}^N\lambda_j g_j(x,y)+\lambda_{N+1}, \label{funcByg} \end{equation} where \begin{equation} g_j(x,y)=\sqrt{(x-x_j)^2+(y-y_j)^2+c_j^2} -\sqrt{(x-x_i)^2+(y-y_i)^2+c_i^2}. \label{gbasis} \end{equation} The number of unknowns in Equation (\ref{funcByg}) is $N$. As the setting in the paper \cite{Shu2003Local}, $\lambda_{N+1}$ can be replaced by $\lambda_i$, and Equation (\ref{funcByg}) can be written as \begin{equation} h(x,y)=\sum\limits_{j=1,j\neq i}^N\lambda_j g_j(x,y)+\lambda_{i}. \label{funcByg2} \end{equation} Where $g_i(x,y)=1$ and $g_j(x,y),\ j=1,\cdots,N$ but $j \neq i$ given by Equation (\ref{gbasis}) are a base vector for the function space of $h(x,y)$. In RBF-DQ method, the weighting coefficients of $w_{ij}^{(m)},$ and $ \overline{w}_{ij}^{(m)} $ are determined by all the base functions as the test function in Equation (\ref{grbfdq}), and we can obtain \begin{subequations} \begin{equation} \sum\limits_{k=1}^{N}w_{ik}^{(m)}=0, \end{equation} \begin{equation} \frac{\partial ^m g_j(x_i,y_i)}{\partial x^m}=\sum\limits_{k=1}^{N}w_{ik}^{(m)}g_j(x_k,y_k),\ j=1,2,\cdots,N,\ \mbox{but}\ j\neq i. \end{equation} \label{coeBase} \end{subequations} The $N$ equations of (\ref{coeBase}) form a linear system with $N$ unknowns for the given $i$, and the weighting coefficients $w_{ik}^{(m)}$ can be solved by a numerical method. In a similar manner, the weighting coefficients $ \overline{w}_{ij}^{(m)} $ of the $y$-derivatives can also be computed by Equation (\ref{coeBase}) with $x$ substituted by $y$. Substitution of Equation (\ref{grbfdq}) into time discretized equation (\ref{timeFinalSch1}) at point ${\bf x_i}=(x_i,y_i)$, with $u^{k+1}_i$ as the approximation solution of $u({\bf x}_i, t_{k+1})$, yields \begin{equation}\ba{l} u^{k+1}_i-\theta\mu^{k+1}_i\left[\kappa^{k+1}_i \sum\limits_{j=1}^N \left(w_{ij}^{(2)}+\overline{w}_{ij}^{(2)}\right) u_j^{k+1} -\sum\limits_{j=1}^N \left(w_{ij}^{(1)}v_{1,i}^{k+1}+\overline{w}_{ij}^{(1)}v_{2,i}^{k+1}\right)u_j^{k+1}\right] \\ \qquad=(1-b_{i,1}^{k+1})u_i^k+(1-\theta)\mu_i^{k+1}\left[\kappa_i^{k+1} \sum\limits_{j=1}^N \left(w_{ij}^{(2)}+\overline{w}_{ij}^{(2)}\right) u_j^{k} -\sum\limits_{j=1}^N \left(w_{ij}^{(1)}v_{1,i}^{k+1}+\overline{w}_{ij}^{(1)}v_{2,i}^{k+1}\right)u_j^{k}\right] \\ \qquad\quad +\sum\limits_{j=1}^{k-1} (b_{i,j}^{k+1}-b_{i,j+1}^{k+1})u_i^{k-j}+b_{i,k}^{k+1}u_i^0+\mu_i^{k+1} f_i^{k+1} . \ea \label{fullSch} \end{equation} In Equation (\ref{fullSch}), we use the following notations: \begin{subequations} \begin{equation*} \mu^{k+1}_i=\mu(\bfx x_i, t_{k+1}), \quad v_{1,i}^{k+1}=v_1(\bfx x_i,t_{k+1}), \quad v_{2,i}^{k+1}=v_2(\bfx x_i,t_{k+1}), \end{equation*} \begin{equation*} \kappa^{k+1}_i=\kappa(\bfx x_i, t_{k+1}),\quad b_{i,j}^{k+1}=b_j({\bf x}_i,t_{k+1}), \quad f_i^{k+1}=f(\bfx x_i, t_{k+1}). \end{equation*} \end{subequations} \subsection{Local RBF-based differential quadrature method} When the number of knots, $N$, is large, the coefficient matrix of (\ref{coeBase}) may be ill-conditioned. This limits its application. Hence, we mainly use the local RBF-DQ method to solve the variable-order time fractional advection-diffusion equation. The key of local RBF-based differential quadrature method is that the $m$-order derivatives of $u(\bfx x)$ at $\bfx x_i$ are approximated by the function values at a set of nodes in the neighborhood of $\bfx x_i$ with $N_i$ nodes (including $\bfx x_i$). That is \begin{equation} u^{(m)}_x({\bf x_i})\approx \sum\limits_{j=1}^{N_i} w_{ij}^{(m)} u({\bf x}_j), \quad u^{(m)}_y({\bf x_i})\approx \sum\limits_{j=1}^{N_i} \overline{w}_{ij}^{(m)} u({\bf x}_j),\quad i=1,2,\cdots,N. \label{lrbfdq} \end{equation} The corresponding coefficients $w_{ij}^{(m)} $ and $\overline{w}_{ij}^{(m)} $ can be determined by Equation (\ref{coeBase}) with $N_i$ local support nodes in the neighbor of $\bfx x_i$. Substitution of Equation (\ref{lrbfdq}) into time discretized equation (\ref{timeFinalSch1}) at point ${\bf x_i}$, by the local RBF-based differential quadrature method, yields \begin{equation}\ba{l} u^{k+1}_i-\theta\mu^{k+1}_i\left[\kappa^{k+1}_i \sum\limits_{j=1}^{N_i} \left(w_{ij}^{(2)}+\overline{w}_{ij}^{(2)}\right) u_j^{k+1} -\sum\limits_{j=1}^{N_i} \left(w_{ij}^{(1)}v_{1,i}^{k+1}+\overline{w}_{ij}^{(1)}v_{2,i}^{k+1}\right)u_j^{k+1}\right] \\ \qquad=(1-b_{i,1}^{k+1})u_i^k+(1-\theta)\mu_i^{k+1}\left[\kappa_i^{k+1} \sum\limits_{j=1}^{N_i} \left(w_{ij}^{(2)}+\overline{w}_{ij}^{(2)}\right) u_j^{k} -\sum\limits_{j=1}^{N_i}\left(w_{ij}^{(1)}v_{1,i}^{k+1}+\overline{w}_{ij}^{(1)}v_{2,i}^{k+1}\right)u_j^{k}\right] \\ \qquad\quad +\sum\limits_{j=1}^{k-1} (b_{i,j}^{k+1}-b_{i,j+1}^{k+1})u_i^{k-j}+b_{i,k}^{k+1}u_i^0+\mu_i^{k+1} f_i^{k+1} , \ea \label{localfullSch} \end{equation} for $i=1,2,\cdots, N$. As shown in the previous subsection, the RBF-DQ approximation of the function contains a shape parameter $c$ that could be knot-dependent and must be determined by the user. In this study, we use the method of normalization of supporting region suggested in the paper of \cite{Shu2003Local} and set $c=5$. \section{Boundary treatment} \label{bndyTreat} The boundary treatment for Dirichlet boundary condition in collocation method is trivial. However, the collocation scheme has difficulties in dealing with Neumann boundary conditions. In this paper, for the variable-order time fractional advection-diffusion equation, we present a new Hermite-type RBF-DQ method for the Neumann boundary condition. \subsection{Hermite RBF interpolation} The approximation of a function $u({\bf x})$ (see the papers \cite{Liu2005Radial, Liu2005An}) can be written in a linear combination of RBFs at all the $N_i$ nodes (including the $N_i^b$ Neumann boundary points) within the local support domain and the normal derivatives along the normal direction ${\bf n}=(n_x,n_y)$ at the Neumann boundary points as \begin{equation} u({\bf x})\approx h(x,y)=\sum\limits_{j=1}^{N_i} \lambda_j \phi_j+ \sum\limits_{l=1}^{N_i^b } \gamma_l \frac{\partial \phi_l^b}{\partial{\bf n}_l}+ \lambda_{N_i+1}, \label{nrbfdq} \end{equation} where $\lambda_j (j=1,\cdots,N_i),\gamma_l (l=1,\cdots,N_i^b)$ are interpolation coefficients, and $\phi_j, \phi_l^b$ are radial basis function, respectively. In this study, as the Section \ref{fbfdqSec}, the MQ basis function is used as the radial basis function, that is \begin{equation} \phi_j=\sqrt{(x-x_j)^2+(y-y_j)^2+c_j^2},\quad \phi_l^b=\sqrt{(x-x_l^b)^2+(y-y_l^b)^2+c_l^2}, \label{mqphi} \end{equation} where ${\bf x} _l^b=(x_l^b,y_l^b)$ is the coordinate for the $l$th normal derivative boundary point. So we can get \begin{equation} \frac{\partial\phi_l^b}{\partial {\bf n}_l}= \frac{\partial\phi_l^b}{\partial x }n_x+ \frac{\partial\phi_l^b}{\partial y}n_y=\frac{(x-x_l^b)n_x+(y-y_l^b)n_y}{\sqrt{(x-x_l^b)^2+(y-y_l^b)^2+c_l^2}}. \label{ndmqphi} \end{equation} With the constrains of (\ref{lweq}) and the notation of (\ref{gbasis}), the equation of (\ref{nrbfdq}) can be reformulated as \begin{equation} u({\bf x})\approx h(x,y)=\sum\limits_{j=1,j\neq i}^{N_i} \lambda_j g_j(x,y)+ \lambda_{i}+ \sum\limits_{l=1}^{N_i^b } \gamma_l \frac{\partial \phi_l^b}{\partial{\bf n}_l}, \label{nrbfdq2} \end{equation} and \begin{equation} \frac{\partial u({\bf x}) }{\partial {\bf n}}\approx \sum\limits_{j=1,j\neq i}^{N_i} \lambda_j \frac{\partial g_j(x,y)}{\partial {\bf n}}+ \sum\limits_{l=1}^{N_i^b } \gamma_l \frac{\partial }{\partial {\bf n}}\left(\frac{\partial \phi_l^b}{\partial{\bf n}_l}\right), \label{nrbfdq3} \end{equation} where $g_i(x,y)=1$, $g_j(x,y),\ j=1,\cdots,N_i$ but $j \neq i$ and $\frac{\partial\phi_l^b}{\partial {\bf n}_l},\ l=1,\cdots,N_i^b$ are a base vector for the function space of $h(x,y)$. Let \begin{equation} \hat{u}=\left[ u_1 \ \cdots\ u_{N_i}\ \frac{\partial u_1^b}{\partial {\bf n}_1} \ \cdots\ \frac{\partial u_{N_i^b}^b}{\partial {\bf n}_{N_i^b}}\right]^T_{1\times(N_i+N_i^b)}, \label{ue} \end{equation} and the coefficients vector \begin{equation} \hat{a}=\left[ \lambda_1 \ \cdots\ \lambda_{N_i}\ \gamma_1 \ \cdots\ \gamma_{N_i^b}\right]^T_{1\times(N_i+N_i^b)}, \label{ue} \end{equation} can be obtained by the interpolations at $N_i$ points and the normal derivatives at $N_i^b$ points on Neumann boundary in Equations (\ref{nrbfdq2}) and (\ref{nrbfdq3}). The process can also be expressed by matrix formulation as follows \begin{equation} \hat{u}=\Psi \hat{a}. \label{uematrix} \end{equation} Thus, the unknown coefficients vector \begin{equation} \hat{a}=\Psi^{-1} \hat{u}. \label{ae} \end{equation} Finally, the approximation form of function can be obtained like \begin{equation} u({\bf x})\approx \phi \hat{a}=\phi\Psi^{-1} \hat{u}=\psi \hat{u}, \label{uappr} \end{equation} where \begin{equation}\phi=\left[g_1(x,y), \cdots,g_{i-1}(x,y),1,g_{i+1}(x,y), \cdots, g_N(x,y), \frac{\partial \phi_1^b}{\partial{\bf n}_1},\cdots,\frac{\partial \phi_{N_i^b}^b}{\partial{\bf n}_{N_i^b}} \right], \label{nbase} \end{equation} and the matrix of shape functions $\psi$ can be expressed as follows \begin{equation} \psi=\left[ \psi_1,\psi_2,\cdots, \psi_{N_i}, \psi_1^H, \psi_2^H,\cdots, \psi_{N_i^b}^H \right]. \label{psiBase} \end{equation} Hence, the function $u({\bf x})$ can be approximately expressed as \begin{equation} u({\bf x})\approx \sum\limits_{k=1}^{N_i} \psi_k u_k + \sum\limits_{j=1}^{N_i^b} \psi_j^H \frac{\partial u_{j}^b}{\partial {\bf n}_j}. \label{uappr2} \end{equation} \subsection{Hermite RBF-DQ method} From the interpolation formulation (\ref{uappr2}), we can directily get \begin{equation} \frac{\partial^m u({\bf x})}{\partial x^m}\approx \sum\limits_{k=1}^{N_i} \frac{\partial^m \psi_k }{\partial x^m} u_k + \sum\limits_{j=1}^{N_i^b}\frac{ \partial^m\psi_j^H}{\partial x^m} \frac{\partial u_{j}^b}{\partial {\bf n}_j}. \label{uapprDS} \end{equation} Then the $m$-th order derivative value $u^{(m)}_x({\bf x})=\frac{\partial^m u({\bf x})}{\partial x^m}$ at the centre ${\bf x}_i$ can be approximated by a linear weighted sum of the function values at $N_i$ points and the normal derivatives at $N_b$ points on Neumann boundary as \begin{equation} u^{(m)}_x({\bf x_i}) \approx \sum\limits_{k=1}^{N_i} w_{ik} ^{(m)}u_k + \sum\limits_{j=1}^{N_i^b} v_{ij}^{(m)} \frac{\partial u_{j}^b}{\partial {\bf n}_j}. \label{uapprDSpx} \end{equation} In order to get the coefficients in Equations (\ref{uapprDSpx}), by the idea of differential quadrature in Section \ref{fbfdqSec}, we use all the base functions in (\ref{nbase}) as the test functions and solve the following system \begin{subequations} \begin{equation} \sum\limits_{k=1}^{N_i}w_{ik}^{(m)}=0, \end{equation} \begin{equation} \frac{\partial^m g_j(x_i,y_i)}{\partial x^m}=\sum\limits_{k=1}^{N_i}w_{ik}^{(m)}g_j(x_k,y_k)+\sum\limits_{l=1}^{N_i^b} v_{il}^{(m)} \frac{\partial g_{j}(x_l^b,y_l^b)}{\partial {\bf n}_l} ,\ j=1,2,\cdots,N_i,\ \mbox{but}\ j\neq i, \end{equation} \begin{equation} \frac{\partial ^m}{\partial x^m}\left( \frac{\partial\phi_p^b}{\partial {\bf n}_p} \right)(x_i,y_i)=\sum\limits_{k=1}^{N_i}w_{ik}^{(m)} \frac{\partial\phi_p^b(x_k,y_k)}{\partial {\bf n}_p} +\sum\limits_{l=1}^{N_i^b} v_{il}^{(m)} \frac{\partial }{\partial {\bf n}_l}\left(\frac{\partial \phi_p^b}{\partial{\bf n}_p}\right)(x_l^b,y_l^b) ,\ p=1,2,\cdots,N_i^b. \end{equation} \label{coeBaseDS} \end{subequations} In a similar manner, the weighting coefficients $ \overline{w}_{ik}^{(m)}, \overline{v}_{ij}^{(m)} $ of the $m$-th order $y$-derivatives can also be computed by Equation (\ref{coeBaseDS}) with $x$ substituted by $y$. Then the Neumann boundary condition at point $(x_i^b,y_i^b)$ can be formulated as \begin{equation} \frac{\partial u(x_i^b,y_i^b)}{\partial {\bf n}}\approx \sum\limits_{k=1}^{N_i} (w_{ik} ^{(1)}n_x+ \overline{w}_{ik}^{(1)}n_y)u_k + \sum\limits_{j=1}^{N_i^b} (v_{ij}^{(1)} n_x+\overline{v}_{ij}^{(1)}n_y)\frac{\partial u_{j}^b}{\partial {\bf n}_j}. \label{HRBFNB} \end{equation} We must mention that the present Hermite RBF point interpolation method is very different from the ones in \cite{Liu2005Radial, Liu2005An}. The method in \cite{Liu2005Radial, Liu2005An} firstly required to obtain the matrix of shape functions $\psi$ in (\ref{uappr}); then the differential of components of $\psi$ must be calculated to obtain Equation (\ref{uapprDS}). Furthermore, for time related problems, their method demand that the calculation of the matrix of shape functions $\psi$ with their differential of components should be solved at every time step. In our method, the coefficients in (\ref{uapprDSpx}) is only related to the support points set; and for time related problems, we just solve the system of (\ref{coeBaseDS}) once. So the coefficients in (\ref{uapprDSpx}) can be stored before the time evolution. In the present Hermite RBF-DQ method, the normal derivatives at the Neumann boundary nodes are added as additional DOFs and Equations (\ref{HRBFNB}) are enforced. For an internal collocation node at ${\bf x}_i$, if its local support nodes do not include Neumann boundary node, the conventional differential quadrature formulas (\ref{lrbfdq}) are used to approximate the derivatives in Equation (\ref{timeFinalSch1}), and the Equation (\ref{localfullSch}) is used to discretize the variable-order time fractional advection-diffusion equation. If its support neighboring nodes include Neumann boundary nodes, the formulation (\ref{uapprDS}) is used. Although, the derivatives in the Neumann condition can also be discretized by the local RBF-DQ method of (\ref{lrbfdq}) as \begin{equation} \frac{\partial u(x_i^b,y_i^b)}{\partial {\bf n}}\approx \sum\limits_{k=1}^{N_i} (w_{ik} ^{(1)}n_x+ \overline{w}_{ik}^{(1)}n_y)u_k, \label{LRBFDQNB} \end{equation} the accuracy is far lower than the present Hermite RBF-DQ method, which will be shown in the next section. \section{Numerical examples} \label{sec:numEx} In this section, numerical experiments are carried out to demonstrate the effectiveness of RBF-DQ method developed in this paper for variable-order time fractional advection-diffusion equation with Dirichlet or Neumann boundary condition. Three different test problems are chosen to show the capability and accuracy of the proposed method. These test problems come from the paper \cite{TAYEBI2017655}, but the boundary conditions with Neumann condition and different solution domains are considered. Although Equation (\ref{localfullSch}) is valid for any value of $\theta \in [0,1]$, we use $\theta=1$ as the famous implicit scheme in all numerical examples. To show the accuracy of the proposed method, the $L^2,\ L^{\infty}$ errors and Root-Mean-Square (RMS) of errors are measured using the following definitions: \begin{equation*} L^2= \sqrt{ \left.{\sum\limits_{i=1}^{N} \left | u({\bf x}_i)-u^h({\bf x}_i) \right |^2} \middle/ { \sum\limits_{i=1}^{N} \left | u({\bf x}_i) \right |^2} \right.}, \end{equation*} \begin{equation*} L^{\infty}=\max \limits_{1 \leq i \leq N} \left | u({\bf x}_i)-u^h({\bf x}_i) \right |, \ \mathrm{ RMS}= \sqrt{ \frac{1}{N} \left(\sum\limits_{i=1}^{N} \left | u({\bf x}_i)-u^h({\bf x}_i) \right |^2\right)}, \end{equation*} where $u^h({\bf x} _i)$ is the numerical solution of $u({\bf x}_i)$. \\ \\ {\bf Example 1.} Consider the following variable-order time fractional advection-diffusion equation \begin{equation} \frac{\partial^{ \alpha({\bf x},t)} u({\bf x},t)}{\partial t^{ \alpha({\bf x},t)}}=\frac{\partial ^2 u({\bf x},t)}{\partial x^2}+ \frac{\partial ^2 u({\bf x},t)}{\partial y^2}- \frac{\partial u({\bf x},t)}{\partial x}- \frac{\partial u({\bf x},t)}{\partial y}+ \frac{2t^{2-\alpha ({\bf x},t)}}{\Gamma(3-\alpha ({\bf x},t))}+2x+2y-4. \label{ex1eq} \end{equation} \begin{comment} Here, the rectangular domain $\Omega=[0,1]\times[0,1]$ and unit circular domain are considered for the simulation. As a test of RBF-DQ method, we firstly solve the Equation (\ref{ex1eq}) with Dirichlet boundary condition. The corresponding initial and boundary conditions for the problem can be computed by the exact solution $u({\bf x},t)=x^2+y^2+t^2.$ We have solved this problem by the developed method for $\Delta t=1/200$. The point distribution for the rectangular and circular domains are shown in Figure \ref{fig:Ex1points}. The total numbers of points are 494 and 1452 for the rectangular and circular domain, respectively. The approximate solutions at time $t=1$ with $\alpha ({\bf x},t) =0.8-0.1\cos(xt)\sin(x)-0.1\cos(yt)\sin(y)$ and the absolute errors on different domains are demonstrated in Figure \ref{fig:Ex1RecDir} and \ref{fig:Ex1CirDir}. From the numerical results, the approximate solutions obtained by the presented local RBF-DQ method are in agreement with the exact solutions for regular and irregular domains. To quantify the accuracy of the presented method, the $L^2,\ L^{\infty}$ and RMS errors at time $t=1$ for two constant orders $\alpha =0.5,\ 0.8$ and $\alpha ({\bf x},t) =0.8-0.1\cos(xt)\sin(x)-0.1\cos(yt)\sin(y)$ are reported in Table \ref{tab:ex1Accu}. It has proven that the local RBF-DQ method has very good accuracy on constant and variable order time fractional advection-diffusion equation. \setlength{\tabcolsep}{10pt} \begin{table}[h]\caption{$L^2/L^{\infty}$ errors and RMS of errors obtained by local RBF-DQ method on different domains with Dirichlet boundary conditions }\label{tab:ex1Accu} \begin{center} \begin{tabular}{c|cccccc} \hline &\multicolumn{3}{c}{Rectangular domain} &\multicolumn{3}{c}{Circular domain } \\ \hline $\alpha$& $L^2$ & $L^{\infty} $ &RMS & $L^2 $ &$L^{\infty} $ & RMS \\ \hline 0.5 &2.1106e-05 & 7.6664e-05 &3.6774e-05 &6.9255e-05 & 1.8004e-04 & 1.0763e-04 \\ 0.8 &1.0315e-05 & 3.59885e-05 &1.7973e-05 &6.3183e-06 & 1.8717e-05 & 9.8190e-06 \\ $\alpha({\bf x},t)$ &8.6187e-06& 3.83095e-05 &1.5017e-05 &3.3338e-05 & 1.2044e-05 & 5.1809e-05 \\ \hline \end{tabular} \end{center} \end{table} \begin{figure}[h] \centering \subfigure[The rectangular domain.]{ \includegraphics[width=3.125in]{pictures/ex1pointRec.eps} } \subfigure[The circular domain.]{ \includegraphics[width=3.125in]{pictures/ex1pointCir.eps} } \caption{Point distribution of the rectangular domain and circular domain for Example 1.} \label{fig:Ex1points} \end{figure} \begin{figure}[h] \centering \subfigure[The approximate solution.]{ \includegraphics[width=3.125in]{pictures/ex1uapproxRec.eps} } \subfigure[The absolute error.]{ \includegraphics[width=3.125in]{pictures/ex1errorRec.eps} } \caption{The space graphs of the approximate solution and absolute error for Example 1 on the rectangular domain.} \label{fig:Ex1RecDir} \end{figure} \begin{figure}[h] \centering \subfigure[The approximate solution.]{ \includegraphics[width=3.125in]{pictures/ex1uapproxCir.eps} } \subfigure[The absolute error.]{ \includegraphics[width=3.125in]{pictures/ex1errorCir.eps} } \caption{The space graphs of the approximate solution and absolute error for Example 1 on the unit circular domain.} \label{fig:Ex1CirDir} \end{figure} \end{comment} \section{Conclusion} \label{con_sec} \bibliographystyle{plain}

9c880732213f771986279f2baefdef3958d69a38

\section{\textsc{Introduction}} Over the last decades, trained networks have been used as a test-bed for understanding how complex computations can be performed by large ensembles of elementary non-linear units \citep{Rosenblatt1958, Tsoi, LeCun2015}. Trained recurrent networks, moreover, can be interpreted as models of cortical circuits that one can easily probe and analyze in an \emph{in-silico} setup \citep{Mante2013, Sussillo2014, Rajan2016, Wang2018, Barak2017}. Because of the long temporal dependencies generated by the intricate connectivity, designing efficient learning algorithms for recurrent networks is, however, a difficult task. Only recently, advances in technology and machine learning algorithms \citep{Atiya, MartensSutskever, Pascanu2013} have made possible to effectively train large recurrent networks by overcoming severe stability issues \citep{Doya, Bengio}. A first strategy to circumvent the problem of dealing with recurrent temporal dependencies was proposed in the work by Jaeger \citep{JaegerHaas} and Maass \citep{Maass2007}. In these frameworks, external feedback loops were implemented to control and reshape the disordered dynamics that spontaneously appear in large random networks. In such feedback architectures, the task-specific component of the recurrent dynamics is totally specified by the feedback input from the readout unit to the random reservoir. As a result, learning can be restricted to the readout weights, which are trained by clamping the feedback input to the target of the task \citep{Jaeger, Lukosevicius}, or by fast weights modification within an online training scheme \citep{Jaeger2002, SussilloAbbott}. From a mathematical perspective, feedback loops can be interpreted as rank-one perturbations of the originally random connectivity. Such a low-rank perturbation can significantly modify the characteristics of the original matrix \citep{Tao2013}, and generates the articulate dynamics which are required by complex tasks like pattern generation and decision making \citep{SussilloAbbott}. In order to obtain arbitrary activity profiles, the entries of the rank-one term are often tightly fine-tuned to the entries of the original random matrix. As learning potentially relies on amplifying the correlations which exist between the readout vector and the finite-size random bulk, developing a formal understanding of the resulting network dynamics has proved to be a difficult theoretical task \citep{SussilloBarak, MassarMassar, RivkindBarak}. In this work, we focus on a simple, yet not fully explored scenario: the feedback network is trained to produce a stationary output signal in response to a constant pattern of external inputs \citep{RivkindBarak}. To characterize the problem from a theoretical perspective, we focus on a restricted class of simple training solutions for which a complete analytical description can be developed. These training solutions correspond to readouts which are strongly aligned with the external and feedback input weights, but are uncorrelated with respect to the random synaptic weights within the reservoir. For such readouts, we derive a mean-field theory which captures the network behaviour in terms of a few representative macroscopic variables \citep{Sompolinsky1988, MastrogiuseppeOstojic2} and allows us to isolate the relevant geometrical components of the readout vector which dominantly contribute to the final network dynamics. In particular, we find that different combinations of these components can be mapped into different output states of the final feedback network. In a second step, we consider the more general training solutions which are found by standard learning techniques, such as least-squares (LS) \citep{Jaeger, Lukosevicius} and recursive least-squares (RLS) algorithms \citep{SussilloAbbott, LajeBuonomano}. We use our mean-field framework to derive approximate analytical descriptions of such solutions, where the full trained readout vector is replaced by its projection on the hyperplane spanned by the external and feedback input vectors. We show that such simplified portraits predict with good precision the output states that are displayed by trained networks, together with their stability: our simplified description completely neglects the correlations existing between the readout vector and the random bulk, but still accurately predicts how the network phase space is reshaped by training. Specifically, we show that learning through the rank-one feedback can induce severe dynamical instabilities even in an extremely simple task. We use our theoretical framework to unveil and compare the strategies that the LS and RLS algorithms adopt to solve the task, and to clarify the reasons of their failures and successes. \section{\textsc{The setup}} We consider a rate-based network model, where every unit $i$ is characterized by a continuous activation variable $x_i$ and its non-linear transform $\phi(x_i)$ \citep{Sompolinsky1988}. In line with many classical studies \citep{Jaeger, SussilloAbbott, LajeBuonomano, RivkindBarak}, we chose $\phi(x)=\tanh(x)$. The case of rate networks characterized by a threshold-linear activation function is discussed in \emph{Appendix C}. To begin with, we consider a random reservoir network driven by an external input consisting of an $N-$dimensional stationary pattern $\mathbf{I}$. Its dynamics read: \begin{equation}\label{eq:dyn_nofeedback} \dot{x_i}(t) = -x_i(t) + g\sum_{i=1}^N \chi_{ij} \phi(x_j(t)) + I_i, \end{equation} where we have rescaled time to set the time constant to unity. The parameter $g$ quantifies the strength of the random connectivity $\bm{\chi}$, whose entries are generated from a centered Gaussian distribution of variance $1/N$ \citep{Sompolinsky1988}. The network output is defined at the level of a linear readout: \begin{equation}\label{eq:readout} z(t) =\sum_{i=1}^N n_i \phi(x_i(t)), \end{equation} where vector $\mathbf{n}$ sets the output readout direction. A task specifies how the network needs to map the external input $\mathbf{I}$ into the output signal $z(t)$. The task is solved when an appropriate readout vector $\mathbf{n}$ has been learned, such that the readout $z$ correctly matches the target. Here we consider a specific fixed point task, in which the network has to hold an arbitrary stationary output signal: $z(t) = A$. The random reservoir in Eq.~\ref{eq:dyn_nofeedback} admits a unique input-driven stable state \citep{Kadmon2015, MastrogiuseppeOstojic}, where single units activity is a combination of the feedforward input and the recurrent dynamics generated by random synapses. For reasons that will become clear in a moment, we refer to this attractor as the \emph{feedback-free} activity state. Once the network activity has relaxed into such a state, the task can be solved by tuning the readout weights $n_i$ until the network output matches the target $A$. This training protocol results into a \emph{feedback-free} implementation of the task, for which the dynamics of the reservoir is independent of the choice of the readout weights. While they are suitable for implementing many simple input-output associations, feedback-free reservoirs have a limited range of applications. As pointed out in a series of seminal studies \citep{Jaeger, Maass2007, SussilloAbbott}, more flexible architectures can be obtained by using the readout signal $z(t)$ as an additional external input, which acts as an external feedback to the random reservoir (Fig.~\ref{fig:general} \textsf{a}). The dynamics of the network obey: \begin{equation}\label{eq:full_dynamics} \dot{x_i}(t) = -x_i(t) + g\sum_{i=1}^N \chi_{ij} \phi(x_j(t)) + m_i z(t) + I_i, \end{equation} where the $N$-dimensional vector $\mathbf{m}$ defines the input direction of the feedback signal. In standard training procedures \citep{Jaeger, SussilloAbbott}, vector $\mathbf{m}$ is considered to be fixed across learning, and is generated with random entries. The feedback loop from the readout to the reservoir introduces a novel, one-dimensional element of recurrent dynamics in the network. Substituting Eq.~\ref{eq:readout} into Eq.~\ref{eq:full_dynamics} allows to verify that the final feedback architecture can be described by an equivalent connectivity matrix, which contains the original random bulk together with a rank-one matrix specified by the feedback vectors: \begin{equation}\label{eq:full_dynamics_rankone} \dot{x_i}(t) = -x_i(t) + \sum_{i=1}^N \left( g\chi_{ij} + m_i n_j \right) \phi(x_j(t)) + I_i. \end{equation} As observed in a variety of studies \citep{Jaeger, Lukosevicius}, such a low-rank matrix perturbation can deeply reshape the dynamical landscape of the reservoir network. The number of activity states and their stability properties, which directly affect training performance in practical applications, strongly depend on the choice of the readout vector $\mathbf{n}$. Building an exact causal relationship between the different readout solutions and the corresponding emerging dynamics is, however, a non-trivial theoretical problem \citep{MassarMassar, RivkindBarak, Landau}. In this work, we characterize the dynamics emerging in feedback networks (Eq.~\ref{eq:full_dynamics}) trained to solve the fixed point task. The input vectors $\mathbf{m}$ and $\mathbf{I}$ determine the fixed geometry of the task by specifying two preferred directions in the $N$-dimensional phase space spanned by the reservoir population. In our framework, vectors $\mathbf{m}$ and $\mathbf{I}$ are generated as random patterns of variance $\sigma_m^2$ and $\sigma_I^2$. Their entries are extracted from joint Gaussian distributions of null mean. The two vectors can thus be constructed as: \begin{equation}\label{eq:mI} \begin{split} &\mathbf{m} = \sigma_m \left( \rho \bm{\xi} + \sqrt{1 - \rho^2} \: \bm{\eta}_m \right) \\ &\mathbf{I} = \sigma_I \left( \rho \bm{\xi} + \sqrt{1 - \rho^2} \: \bm{\eta}_I \right) \\ \end{split} \end{equation} where $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta}_I$ are standard Gaussian vectors. The parameter $\rho$ quantifies the strength of correlations between the input vectors $\mathbf{m}$ and $\mathbf{I}$, defined onto the overlap direction $\bm{\xi}$ (alternatively, the overlap direction $\bm{\xi}$ can be taken to be parallel to the unitary vector $\mathbf{u} = (1, ..., 1)$, resulting in input vectors $\mathbf{m}$ and $\mathbf{I}$ characterized by random and uncorrelated entries of non-zero mean). From this setup, the feedback-free implementations can be retrieved as a special case by setting the feedback inputs to zero through $\sigma_m=0$. \section{\textsc{General classes of solutions}} As a first step, we approach the problem from a theoretical perspective. We analyze the space of all the possible task solutions by focusing on the limit of a large random bulk ($N\rightarrow\infty$). The specific task we consider requires a constant readout signal, which is trivially obtained in stationary network states. Equilibrium solutions, which are obtained by setting $\dot{x}_i(t)=0$ in Eq.~\ref{eq:full_dynamics}, can be written as: \begin{equation}\label{eq:xi} x_i = m_i z + I_i + \delta_i, \end{equation} where the term $\delta_i$ corresponds to self-consistent recurrent input generated by the random part of the connectivity: \begin{equation} \delta_i = g\sum_{j=1}^N \chi_{ij} \phi(x_j(t)). \end{equation} The value of the readout $z$ can thus be re-expressed as: \begin{equation}\label{eq:kappa} \begin{split} z & = \sum_{i=1}^N n_i \phi(m_i z + I_i + \delta_i)\\ & = N \langle n_i \: \phi (m_i z + I_i + \delta_i) \rangle, \end{split} \end{equation} where the notation $\langle . \rangle$ indicates an average over the population of units. For a broad range of $\phi(x)$ functions, the output $z$ is thus determined by the correlations between the readout $\mathbf{n}$ and the vectors defining the direction of the network activity: $\mathbf{m}$, $ \mathbf{I}$ and $\bm{\delta}$. While vectors $\mathbf{m}$ and $\mathbf{I}$ are considered to be fixed, the direction of the residual input $\bm{\delta}$ varies from one realization to the other of the random connectivity $\bm{\chi}$. For a given matrix $\bm{\chi}$, furthermore, the direction of vector $\bm{\delta}$ varies in the different dynamical states admitted by the network. For these reasons, the residual input $\bm{\delta}$ represents the hardest term to evaluate and control. One approach consists in fixing the random connectivity matrix and computing $\bm{\delta}$ in the \emph{open-loop} configuration, obtained by clamping the readout signal to the target: $z(t)=A$ \citep{Jaeger, RivkindBarak}. Such a value for $\bm{\delta}$, however, holds only locally -- that is, in the vicinity of the target fixed point. As a consequence, this approach does not allow us to control the full, global dynamics that the readout $\mathbf{n}$ imposes on the final feedback network. For example, it does not allow to predict whether additional, spurious attractors are created together with the desired output state. A different approach is used in classical mean-field studies of purely random networks \citep{Sompolinsky1988, Molgedey, Cessac94, Schuecker2016, Marti2018}. In that case, the recurrent input $\bm{\delta}$ is approximated by a vector of uncorrelated white noise of self-consistent amplitude. Such an approximation is justified by the disorder in the random connectivity, which effectively decorrelates the input to different units. This procedure washes out any dependence on the specific instantiation of the random matrix $\bm{\chi}$, so that classical mean-field theories cannot be easily used to describe solutions where the readout vector $\mathbf{n}$ is tightly correlated with $\bm{\delta}$. For the simple task we consider here, one can focus on the restricted class of solutions where the readout $\mathbf{n}$ is only correlated with the input vectors $\mathbf{m}$ and $ \mathbf{I}$. We show in the next paragraph that the classical mean-field approaches can be easily used in this case to map any readout vector into a full description of the network output states. As they do not rely on correlations with the random bulk $\bm{\chi}$, readout solutions in this class generate an output which is robust with respect to changes in the random part of the connectivity. From Eq.~\ref{eq:mI}, a minimal and general solution in this class can be written as: \begin{equation}\label{eq:sol} \mathbf{n} = \frac{c}{N} \left(p \bm{\xi} + p_m \bm{\eta}_m + p_I \bm{\eta}_I\right). \end{equation} The geometry of the readout vector is determined by the relative weights of the three coefficients $(p, p_m, p_I)$, which quantify how much the readout direction is distributed on the three axes defined by the non-trained part of the network architecture. When the readout vector is aligned with one of the three axes $\bm{\xi}$, $\bm{\eta}_m$ or $\bm{\eta}_I$, the population average in Eq.~\ref{eq:kappa} takes finite $\mathcal{O}(1)$ values (see below). As a consequence, training solutions in the form of Eq.~\ref{eq:sol} need to scale as the inverse of the network size ($n_i \sim \mathcal{O}(1/N)$) in order to prevent diverging activity and outputs. The weights $(p, p_m, p_I)$ are furthermore normalized through the constant $c$, which fixes the output $z$ to the exact target value specified by the task. In the remaining of this section, we characterize analytically the dynamics emerging in trained networks where the readout vector belongs to the restricted class defined by Eq.~\ref{eq:sol}. To this end, we introduce the key equations of the mean-field description and we analyze in detail three simple geometries $(p, p_m, p_I)$ which satisfy the task through rather different output dynamics. In Sections 4 and 5, we exploit our simplified mean-field framework to analyze the more general readout solutions which are generated through learning by standard training algorithms. \subsection{Mean-field description} For every fixed readout vector in the form of Eq.~\ref{eq:sol}, we follow \citet{MastrogiuseppeOstojic2} to derive a macroscopic, effective description of the network activity. As in standard mean-field theories for random recurrent networks \citep{Sompolinsky1988, Rajan2010, Kadmon2015}, we approximate the equilibrium activation vector $\mathbf{x}$ by the statistical distribution of its elements. In particular, every element $x_i$ can be thought as extracted from a Gaussian distribution of variance $\Delta \equiv \langle [x_i^2]\rangle - \langle[x_i]\rangle^2$, where $\langle.\rangle$ indicates an average over the population and $[.]$ an average across different realizations of the random bulk. From here on, we consider the overlap direction $\bm{\xi}$ to be generated as a standard Gaussian vector. In this case, the mean of the distribution of $\mathbf{x}$ vanishes, as $\langle m_i \rangle = \langle I_i \rangle = [\chi_{ij}] =0$. As shown in \citet{MastrogiuseppeOstojic2}, if the overlap direction $\bm{\xi}$ coincides instead with the unitary vector, the mean-field equations take a slightly different form, which includes finite mean values but result in qualitatively identical results. In order to derive a self-consistent expression for $\Delta$, we first consider the average over the bulk connectivity $\bm{\chi}$. By direct averaging of Eq.~\ref{eq:xi} we get: \begin{equation}\label{eq:mf_i} \begin{split} &\mu_i = [x_i] = m_i [z] + I_i\\ &\Delta_i = [x_i^2] - [x_i]^2= m_i^2 \left([z^2]-[z]^2\right) + [\delta_i^2]. \end{split} \end{equation} As in standard mean-field derivations \citep{Sompolinsky1988}, we have $[x_ix_j]=[x_i][x_j]$ and: \begin{equation} [\delta_i^2] = g^2 \sum_{j=1}^N \sum_{k=1}^N [\chi_{ij}\chi_{ik}] [\phi(x_j)\phi(x_k)] = g^2 \sum_{j=1}^N [\chi_{ij}^2][\phi(x_j)^2] = g^2 \langle [\phi(x_i) ^ 2]\rangle, \end{equation} since $[\chi_{ij}^2]=1/N$. In the thermodynamic limit, furthermore, the variance of $z$ vanishes, as $n_i$ is characterized by a weaker scaling with $N$ (from Eq.~\ref{eq:sol}, $n_i^2 \sim \mathcal{O}(1/N^2)$). The readout signal $z$ is thus self-averaging in large networks. For this reason, we drop the brackets in $[z]$, and we compute the readout signal self-consistently through: \begin{equation}\label{eq:kappa_mf} z = c \:\langle \left\{ p \xi_i + p_m \eta_{mi} + p_I \eta_{Ii} \right\} [\phi(x_i)] \rangle. \end{equation} Note that Eq.~\ref{eq:mf_i} allows to effectively replace the recurrent input $\bm{\delta}$ with white noise of self-consistent amplitude: \begin{equation}\label{eq:xi_approx} x_i = \mu_i + \delta_i \sim \mu_i + \sqrt{\Delta_i} w_i \end{equation} where $\Delta_i = g^2 \langle [\phi(x_i) ^ 2]\rangle$ and $w_i$ is a standard Gaussian variable. By averaging again across units, we find that the population distribution of the equilibria $x_i$ obeys the following statistics: \begin{equation}\label{eq:mf} \begin{split} & \mu = \langle [x_i] \rangle = 0 \\ & \Delta = \langle[x_i^2]\rangle - \langle[x_i]\rangle^2 = g^2 \langle [\phi(x_i) ^ 2]\rangle +\sigma_m^2 z^2 + 2\sigma_{mI} z + \sigma_I^2, \end{split} \end{equation} where $\sigma_{mI} = \langle m_i I_i \rangle = \rho^2 \sigma_m\sigma_I$. In order to obtain a closed form for our equations, we derive a self-consistent expression for the average quantities $z$ and $\langle [\phi(x_i) ^ 2]\rangle$. As in standard mean-field theories, we re-express averages through Gaussian integrals: \begin{equation}\label{eq:gauss_int} \langle [\phi(x_i) ^ 2]\rangle = \int \mathcal{D}w \: \phi^2 (\sqrt{\Delta} w), \end{equation} where we used the short-hand notation $\mathcal{D}w = \diff w e^{-\frac{w^2}{2}}/\sqrt{2\pi} $. By combining Eqs.~\ref{eq:mI} and \ref{eq:kappa_mf}, and by integrating by parts over $\xi_i$, $\eta_{mi}$ and $\eta_{Ii}$, we furthermore get \citep{MastrogiuseppeOstojic2}: \begin{equation} \begin{split}\label{eq:kappa_fin} z = c \: \left\{ p (\sigma_m \rho z + \sigma_I \rho) + p_m \sigma_m \sqrt{1-\rho^2} z + p_I \sigma_I \sqrt{1-\rho^2}\right\} \langle [\phi'(x_i)]\rangle, \end{split} \end{equation} where the average $\langle [\phi'(x_i)]\rangle$ can be evaluated similarly to Eq.~\ref{eq:gauss_int}. For any fixed set of readout weights $(p,p_m,p_I)$, Eqs.~\ref{eq:mf} and \ref{eq:kappa_fin} can be solved together through standard numerical techniques to return the values of $\Delta$ and $z$ which characterize the different output fixed points. To conclude, we observe that the equations above only apply to networks which satisfy the task through stationary output states. In general, stationary outputs do not require stationary reservoir activity. For example, solutions in the form of Eq.~\ref{eq:sol} can be used to solve the task even when activity in the bulk is chaotic, but fluctuations cancel out at the level of the readout \citep{MastrogiuseppeOstojic2}. In that case, the mean-field description above can easily be extended to take temporal fluctuations into account \citep{Sompolinsky1988}. Standard training techniques, however, typically fail to converge when the target of the task is not strong enough to suppress chaotic activity in the bulk \citep{RivkindBarak}. For this reason, here and in the following, we concentrate on fixed point dynamics. \subsection{Determining task solutions} For every fixed set of weights $(p, p_m, p_I)$, first the normalization parameter $c$ needs to be determined. We fix $c$ by requiring that, in the final network output, the readout value exactly matches the target $A$. To this end, we consider the {open-loop} formulation of Eqs.~\ref{eq:mf} and \ref{eq:kappa_fin}, obtained by setting $z=A$, and we solve the system for $c$ and $\Delta^{ol}$. This gives: \begin{equation}\label{eq:c} c = \frac{A}{\left\{ p (\sigma_m \rho A + \sigma_I \rho) + p_m \sigma_m \sqrt{1-\rho^2} A + p_I \sigma_I \sqrt{1-\rho^2}\right\} \langle [\phi'(x_i^{ol})]\rangle} \end{equation} where $\langle [\phi'(x_i^{ol})]\rangle$ is computed as Gaussian integral over a distribution of variance $\Delta^{ol}$ (see Eq.~\ref{eq:gauss_int}). Once the normalizing factor has been derived, the full mean-field equations can be solved. Because of the normalization, the mean-field system of equations always admits a stationary solution which satisfies the task for $N \rightarrow \infty$ (i.e.~its readout obeys $z=A$). This solution, however, is not guaranteed to be locally stable from the point of view of the full feedback dynamics (Fig.~\ref{fig:general} {\bf b}). Apart from the target solution, furthermore, the mean-field equations might admit other solutions corresponding to stable fixed points. In those cases, depending on the initial conditions, the dynamics of the final network might converge to spurious output states where the readout $z$ significantly deviates from the target. For every solution of the mean-field equations, we predict local stability by evaluating the eigenspectrum of the linear stability matrix of the corresponding fixed point. For the class of readouts that we consider, the eigenspectra consist of a dense set of eigenvalues, distributed within a circle in the complex plane, together with a single real outlier (Fig.~\ref{fig:general} {\bf b}) \citep{MastrogiuseppeOstojic2, Tao2013, RivkindBarak}. The radius of the circular set and the position of the outlier eigenvalue can be evaluated within the mean-field framework \citep{MastrogiuseppeOstojic2}; their value is determined by the interplay between the amplitude of the random bulk and the relative geometrical arrangement of vectors $\mathbf{n}$, $\mathbf{m}$ and $\mathbf{I}$. Details of the calculations are provided in \emph{Appendix A}. \begin{figure} \centering \includegraphics{NECO-09-18-3251R2-figure1.pdf} \caption{ Implementing the fixed point task: network model and mean-field analysis in three example geometries. {\bfseries a.} Architecture of the feedback network (Eq.~\ref{eq:full_dynamics_rankone}). {\bfseries b.} Sample from the dynamics of two finite-size networks, $N=1000$. In the left example, the output state which corresponds to the target fixed point is locally stable, and the readout signal $z(t)$ converges to the correct value. In the right example, instead, the target corresponds to a locally unstable state, and the readout converges to an incorrect value. Gray traces indicate the time course of the activation variable $\mathbf{x}$ for five randomly selected units. The readout trajectory is instead displayed in blue. The top-right insets display the mean-field prediction for the stability eigenspectrum of the target fixed point (see \emph{Appendix A}). Local instabilities are due to the outlier eigenvalue crossing the instability boundary (dashed). The parameters which have been used in the left and the right examples are indicated by colored arrows in {\bfseries e}. {\bfseries c-d-e.} Mean-field characterization of the network fixed points for three different readout geometries (see text). Continuous (resp. dashed) lines correspond to locally stable (resp. unstable) mean-field solutions. Shaded areas indicate the values of the target $A$ for which the association is locally unstable. Top row: value of the readout signal. Bottom row: position of the outlier eigenvalue in the stability eigenspectra. Gray lines indicate the value of the radius of the eigenvalues bulk. The instability boundary is indicated on the ordinate axis by the black arrows. The results of simulations are displayed as dots ($N=3000$, average over 8 network realizations). We integrate numerically the dynamics of finite-size networks where the readout vector $\mathbf{n}$ is normalized through Eq.~\ref{eq:c}. In order to reach different solutions, we initialize the network dynamics in two different initial conditions, centered around $\mathbf{n}$ and $-\mathbf{n}$. Note that the position of the outlier eigenvalue cannot be measured when the outlier is absorbed in the circular set of the eigenspectrum. In order to ease the comparison between theory and simulations, we thus chose a small $g$ value. Parameters: $g=0.3$, $\sigma_m=1.2$, $\sigma_I=0.5$. $\rho=0.5$. In {\bfseries e}, we take $p=p_m = 1$, $p_I=0.3$. Here and in the following figures: error bars, when visible, correspond to the standard error. } \label{fig:general} \end{figure} \subsection{Categories of readout solutions} We specifically look at the results obtained for three readout vectors characterized by different geometries. In the first example we analyze (Fig.~\ref{fig:general} {\bf c}), the readout $\mathbf{n}$ overlaps solely with the feedback input $\mathbf{m}$ along $\bm{\eta}_m$. This configuration corresponds to setting $p = p_I = 0$, and can be realized also in absence of the external input vector. In this case, the network admits three stationary states for every value of the target $A$. Among them, only two states are locally stable. Bistability is due to the strong overlap between the feedback vectors $\mathbf{m}$ and $\mathbf{n}$, and emerges through a pitchfork bifurcation induced by the outlier eigenvalue in the stability eigenspectrum of the feedback-free fixed point \citep{Hopfield, MastrogiuseppeOstojic2}. As a consequence, stationary states tuned to small target values are characterized by long relaxation timescales. For most of the target values, the solution corresponding to the target (i.e.~$z=A$) is locally stable. However, the network admits a second stable fixed point, characterized by a different readout value. This additional stable state is reached when the network is initialized within the appropriate basin of attraction. The position of the unstable fixed point (dashed lines in Fig.~\ref{fig:general}) can be used to estimate the size of the basins of attraction of the two stable fixed points, which in this case is approximately equal. Within a small parameter region corresponding to very small target values, the target fixed point is built in correspondence of the intermediate, locally unstable branch of the mean-field solutions. The mean-field theory indicates that the local instability is due to the outlier eigenvalue of the fixed point stability matrix laying above the critical line (Fig.~\ref{fig:general} {\bf b} right). The amplitude of the instability region is controlled by the overlap between the two fixed vectors $\mathbf{m}$ and $\mathbf{I}$, and disappears when $\rho = 0$ \citep{RivkindBarak}. In the second example we consider (Fig.~\ref{fig:general} {\bf d}), the readout $\mathbf{n}$ overlaps solely with the external input weights $\mathbf{I}$ along $\bm{\eta}_I$. We thus set: $p = p_m = 0$. Analogously to a feedback-free network, aligning the readout vector with the external input generates a unique stationary solution, corresponding to the target of the task ($z=A$). Such a solution is always stable, and it is characterized by negative values of the outlier eigenvalue. As a last example, we consider the general case where the readout displays non-vanishing overlaps with both vectors $\mathbf{m}$ and $\mathbf{I}$ (Fig.~\ref{fig:general} {\bf e}). Such a category includes the simplest configuration where the readout $\mathbf{n}$ is aligned with the overlap direction $\bm{\xi}$ ($p_m=p_I=0$), but more general solutions characterized by non-zero projections along $\bm{\eta}_m$ and $\bm{\eta}_I$ are comprised as well. As in the first example we considered (Fig.~\ref{fig:general} {\bf c}), the components of the readout which overlap with the feedback input $\mathbf{m}$ tend to generate three different fixed points. Because of the components along the external input $\mathbf{I}$, however, these solutions are no longer characterized by a strong symmetry around zero. As a result, the unstable intermediate fixed point is always characterized by finite output values. This unstable state corresponds to the task solution within a finite parameter region at small and negative target values. This instability area is delimited by a critical target value $A^*$ where the readout amplitude $c$, together with the stable fixed points, diverges: \begin{equation} A^* = - \frac{ p \sigma_I \rho + p_I \sigma_I \sqrt{1-\rho^2}}{p \sigma_m \rho + p_m\sigma_m\sqrt{1-\rho^2}}. \end{equation} Note that this geometry produces a unique and stable input-output association within a parameter window corresponding to small and positive target values. For large (positive or negative) target values, instead, the projection along $\mathbf{m}$ dominates and a second stable fixed point exists. To conclude, we characterized the output states generated by readout solutions in the class defined by Eq.~\ref{eq:sol}. We found that, for different solutions belonging to that class, the task is satisfied through output states which are characterized by different local and global stability properties. Stability can be predicted via the mean-field theory by examining the projections of the readout $\mathbf{n}$ on the input vectors $\mathbf{m}$ and $\mathbf{I}$. In particular, an optimal solution -- characterized by both local and global stability -- can be obtained by aligning the readout $\mathbf{n}$ along the non-shared direction of the external input $\mathbf{I}$. This solution requires a non-vanishing input pattern and generates a feedback network that is dynamically analogous to the feedback-free implementation. Components along the feedback input $\mathbf{m}$, on the other hand, tend to generate additional stable solutions which can attract the dynamics to different readout values. Combining both readout components along $\mathbf{m}$ and $\mathbf{I}$, finally, generates non-symmetric bistable solutions which result in broad regions of local and global instability. \section{\textsc{Least-Squares training}} In Section 3, we derived a theoretical framework which allows us to directly map the geometry of the readout vector into the final output states of the feedback network. This description applies to networks which obey three closely related assumptions (Eq.~\ref{eq:sol}): (i) the readout is a linear combination of the input axes $\bm{\xi}$, $\bm{\eta}_{{m}}$, $\bm{\eta}_{{I}}$; (ii) no correlations exist between the readout and the random bulk; (iii) the scaling of the solution with the network size $N$ is weak (i.e. $n_i$ is of order $1/N$). Importantly, these assumptions might not hold in finite-size networks obtained through standard training techniques. Trained networks might indeed converge to more complex, non-linear solutions, where strong correlations with the random connectivity significantly influence the output dynamics. In Sections 4 and 5, we investigate how far our approximate theory can be used to describe trained networks. To this end, we construct for every training solution a crude approximation in the form of Eq.~\ref{eq:sol}, for which the mean-field equations can be solved and the global dynamics of the resulting network can be predicted. Since in the mean-field description correlations with the random bulk are neglected, comparing mean-field predictions with training performance allows us to quantify how much training relies on a fine-tuning of the readout to the random connectivity $\bm{\chi}$. In cases where the theory predicts that no stable solutions exist, furthermore, our analysis allows to investigate whether finite-size trained networks converge to more effective solutions that the simplified theoretical description cannot capture. We consider in two specific cases, corresponding to two classical training protocols. The first training procedure we consider is the simple batch update through least-squares (LS) inversion \citep{Jaeger, Lukosevicius, RivkindBarak}. To begin with, a finite-size architecture defined by a random bulk $\bm{\chi}$ and two vectors $\mathbf{m}$ and $\mathbf{I}$ is generated. The activity of the whole population in the open-loop configuration is then simulated, yielding: \begin{equation}\label{eq:xi_star} x_i^{ol} = m_i A + I_i + \delta_i^{ol}. \end{equation} The task imposes a unique constraint on the readout, namely: $\phi(\mathbf{x}^{ol})^T \mathbf{n} = A$. This equation represents a underdetermined linear system for the $N$ entries of vector $\mathbf{n}$. The minimum norm LS solution can be computed through the pseudo-inverse of $\phi(\mathbf{x}^{ol})^T$ \citep{Bretscher}: \begin{equation}\label{eq:LS} \mathbf{n} = \frac{A}{N \langle \phi^2(x_i^{ol})\rangle}\phi(\mathbf{x}^{ol}) \end{equation} and thus scales with the network size as $1/N$. Through the open-loop activation vector $\mathbf{x}^{ol}$, the LS readout inherits three $N$-dimensional components along vectors $\mathbf{m}$, $\mathbf{I}$ and $\bm{\delta}^{ol}$ (Eq.~\ref{eq:xi_star}). The non-linear activation function $\phi(x)$, which is applied to $\mathbf{x}^{ol}$ (Eq.~\ref{eq:LS}), mixes and reshapes these three directions, by generating a continuum of additional components along orthogonal axes. The component of $\mathbf{x}^{ol}$ aligned with the recurrent input $\bm{\delta}^{ol}$, furthermore, introduces finite, $\mathcal{O}(1)$ correlations between the readout and the finite-size instantiation of the random bulk used for training. Note that a fraction of those correlations is used by the LS algorithm to exactly match the readout $z$ to the target $A$ within network instantiations of any size. Mean-field solutions (Eq.~\ref{eq:sol}), instead, are independent of the random bulk and imply small readout errors (of the order of $1/\sqrt{N}$) which disappear only in the thermodynamic limit. \subsection{Mean-field approximation} We construct an approximate description of $\mathbf{n}$ (Eq.~\ref{eq:LS}) which conserves the relative strength of its components along the input vectors $\mathbf{m}$ and $\mathbf{I}$ (Eq.~\ref{eq:sol}), thus totally neglecting the components originating from the recurrent input $\bm{\delta^{ol}}$ and the non-linearity. To this end, we extract the linear projections of the readout vector along the orthogonal axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta}_I$, resulting in a set of weights $(p, p_m, p_I)$ which effectively replaces the LS solution with an approximate readout in the form of Eq.~\ref{eq:sol}. For such an approximate readout, the network output states can be exactly predicted. As we show in the next paragraphs, corrections to the mean-field description due to both neglected terms $\bm{\delta^{ol}}$ and $\phi(x)$ can, in this case, be exactly computed \citep{RivkindBarak}. A minimal approximation is however of interest per-se, as it directly generalizes to more complex readout vectors for which analytical expressions do not exist, so that the exact form of correlations and the shape of non-linear components is not known a priori (see Section 5). The projection of $\mathbf{n}$ on $\bm{\xi}$ can be computed as: \begin{equation}\label{eq:LS_alpha} \begin{split} p &=\frac{A}{N \langle \phi^2(x_i^{ol})\rangle} \langle \xi_i \: \phi(x_i^{ol}) \rangle\\ & =\frac{A}{N \langle \phi^2(x_i^{ol})\rangle} \langle \xi_i \: \phi\left( \sigma_m(\rho \xi_i + \sqrt{1-\rho^2}\eta_{mi}) A +\sigma_I(\rho \xi_i + \sqrt{1-\rho^2}\eta_{Ii}) + \delta_i^{ol}\right)\rangle. \end{split} \end{equation} We re-express the population average as a Gaussian integral and we integrate by parts over $\xi_i$, yielding: \begin{equation} p = \frac{A }{N \langle \phi^2(x_i^{ol})\rangle}\rho(\sigma_mA + \sigma_I ) \langle \phi'(x_i^{ol}) \rangle . \end{equation} Similar calculations return the values of $p_m$ and $p_I$. The final set of weights is given by: \begin{equation}\label{eq:LS_weights} \begin{split} & p = \gamma \rho(\sigma_mA + \sigma_I ) \\ & p_m = \gamma \sigma_m \sqrt{1-\rho^2}A \\ & p_I = \gamma \sigma_I \sqrt{1-\rho^2} , \end{split} \end{equation} where we defined the common multiplicative factor $\gamma = A \langle \phi'(x_i^{ol}) \rangle / N \langle \phi^2(x_i^{ol})\rangle $. As in Fig.~\ref{fig:general} {\bf e}, the LS solution thus includes non-vanishing components along the three axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta}_I$. In contrast to Fig.~\ref{fig:general}, however, the values of $(p, p_m, p_I)$ that have been derived from the LS solution display an explicit dependence on the target value $A$. For fixed vectors $\mathbf{m}$ and $\mathbf{I}$, the geometrical arrangement of the readout vector thus changes according to the target of the task. Starting from Eq.~\ref{eq:LS_weights}, we evaluate the normalization factor (Eq.~\ref{eq:c}) and we solve the system of mean-field equations (Eqs.~\ref{eq:mf}, \ref{eq:kappa_fin}). In Fig.~\ref{fig:LS}, the solutions and their stability properties are illustrated. \begin{figure}[t] \centering \includegraphics{NECO-09-18-3251R2-figure2.pdf} \caption{ The least-squares solution: approximate mean-field description. {\bf a-b-c.} Mean-field characterization for three different configurations of the non-trained input vectors $\mathbf{m}$ and $\mathbf{I}$. In {\bf a}, we take $\rho=0.6$; in {\bf b}, $\rho=0.97$; in {\bf c}, $\rho=1$. Continuous and dashed lines indicate the solutions of the simplified mean-field description, where the readout is approximated by a vector in the form of Eq.~\ref{eq:sol}. Details are as in Fig.~\ref{fig:general}. The light green line indicates the exact value of the outlier eigenvalue which measures the local stability of the target solution \citep{RivkindBarak}; details are provided in \emph{Appendix B}. Dots display the results of simulations from finite-size networks ($N=3000$, averages over 8 network realizations). We integrate numerically the dynamics of finite-size networks where the readout vector $\mathbf{n}$ is given by the LS solution (Eq.~\ref{eq:LS}). In order to reach different solutions, we initialize the network dynamics in two different initial conditions, centered around $\mathbf{n}$ and $-\mathbf{n}$. Parameters as in Fig.~\ref{fig:general}. } \label{fig:LS} \end{figure} We find that the shape of the mean-field solutions critically depends on the geometrical arrangement of the non-trained part of the network architecture. Specifically, the network output is strongly influenced by the value of the overlap between the input vectors $\mathbf{m}$ and $\mathbf{I}$, quantified by the parameter $\rho$. When the input vectors $\mathbf{m}$ and $\mathbf{I}$ are orthogonal or share a weak overlap, the theory admits a continuous stable solution -- on which the target output is built -- together with two symmetric stable branches at large target values (Fig.~\ref{fig:LS} {\bf a}). In this case, the theory predicts that training always generates locally stable dynamics \citep{RivkindBarak}, but spurious fixed point attractors are generated at large $A$. As the overlap $\rho$ is increased, the branch of the solution corresponding to the target merges with one of the two additional fixed points, generating a region of local instability at small and negative target values (Fig.~\ref{fig:LS} {\bf b}). In the limit where the input vectors $\mathbf{m}$ and $\mathbf{I}$ are fully aligned ($\rho=1$, Fig.~\ref{fig:LS} {\bf c}), finally, mean-field solutions are similar to the results in Fig.~\ref{fig:general} {\bf e}. The network is characterized by a region of local instability which is delimited by a critical target value $A^*$ where the dynamics diverge. The normalization factor $c$ is in this case given by: \begin{equation} c = \frac{A}{ \rho \left\{\sigma_I + \sigma_m A \right\}^2 \langle [\phi'(x_i^{ol})]\rangle} , \end{equation} which diverges at $A^* = - \sigma_I/\sigma_m$. In Fig.~\ref{fig:LS}, mean-field predictions are compared with the outcome of simulations performed in finite-size networks which have been trained through LS inversion. Despite the approximate nature of the theory, a good agreement between the two is found. As predicted by the theory, output states of trained networks are strongly affected by the value of the input overlap $\rho$. In particular, the theoretical approximation correctly predicts the existence of an area of local instability for large $\rho$ values. Within this parameter region, training generates an output $z(t)$ which immediately diverges from the desired target. For every value of $\rho$, furthermore, the theory captures the existence of a second, spurious fixed point at large target values. The mean-field prediction for the outlier stability eigenvalue can be further compared with an exact analytical expression derived by following the analysis in \citet{RivkindBarak}. This expression, provided in \emph{Appendix B}, includes the components of the readout $\mathbf{n}$ which are neglected in the simplified mean-field description: correlations with the bulk $\bm{\chi}$ and non-linear geometry. We find that, apart from minor quantitative discrepancies, the exact outlier prediction is well matched by the approximate one. Note that this exact analytical expression only provides a measure of local stability, as it can be evaluated only in correspondence to the target solution. To conclude, we found that dynamics in LS trained networks can be explained by a highly simplified, approximate theoretical description which conserves the projection of the readout vector on the hyperplane spanned by the feedback and external input vectors. The approximate mean-field theory specifically predicts that LS readouts often generate bistable fixed points, while unstable solutions can be encountered for large values of the overlap $\rho$. Note that, as long as the input vectors $\mathbf{m}$ and $\mathbf{I}$ are not completely parallel (i.e.~$\rho<1$), the mean-field framework indicates that an optimal readout solution can be designed by aligning the readout $\mathbf{n}$ with the non-shared component of $\mathbf{I}$ (Fig.~\ref{fig:general} {\bf d}). We conclude that, even for the simple task we consider, LS training does not lead to an optimal solution that prevents dynamical instabilities. Apart from an overall quantitative agreement, mismatches between simulations and mean-field solutions can be observed. Their size depends on the value of the target and on the network parameters. A brief description of the role and impact of the different sources of mismatch is presented in the next two paragraphs. \subsection{Non-linear corrections} \begin{figure} \centering \includegraphics{NECO-09-18-3251R2-figure3.pdf} \caption{ The least-squares solution: mean-field description with non-linear corrections ($\rho=1$). {\bf a.} We compare the solutions which are obtained from the linear (blue) and the non-linear (orange) set of mean-field equations. Note that the two set of solutions almost completely overlap. {\bf b.} We plot a magnified view of the solutions within the shaded yellow areas of {\bf a}. We compare both mean-field solutions with the results of simulations performed in LS-trained networks (dots). Parameters and simulations settings are as in Fig.~\ref{fig:LS}. } \label{fig:nonlinear} \end{figure} Together with its components along the input axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta}_I$, the full LS readout (Eq.~\ref{eq:LS}) includes a continuum of orthogonal directions which can potentially affect the network dynamics and output values. We first focus on the orthogonal components which are generated when the non-linear function $\phi(x)$ is applied to $\mathbf{x}^{ol}$ (Eq.~\ref{eq:LS}). We build a more precise mean-field approximation which conserves the non-linearity of the LS solution (Fig.~\ref{fig:nonlinear}). Details of the analysis are provided in \emph{Appendix D}. By solving the novel set of mean-field equations, we altogether find that non-linear solutions lie very close to the ones we obtained for linearly approximated readouts (Fig.~\ref{fig:nonlinear} left). The simple linear approximation, therefore, appears to capture well the readout geometry that is relevant for the final network dynamics. At a finer scale, we find that non-linear solutions in some cases explain with higher precision the output states observed in trained networks. In particular, we observe that the agreement significantly improves in the parameter regions corresponding to large and positive target values, where the dynamics are strongly non-linear and admit bistable fixed points (Fig.~\ref{fig:nonlinear} right). \subsection{The effect of correlations} Our approximate mean-field description neglects a second set of orthogonal readout directions, generated by the recurrent input $\bm{\delta}^{ol}$ (Eq.~\ref{eq:xi_star}). When this vector is replaced with Gaussian noise, correlations between the LS readout and the random part of the connectivity are effectively washed out. In order to boost the effect of correlations, we increase the strength of random connections $g$ up to the critical point where dynamics become chaotic (Fig.~\ref{fig:correlations_a} {\bf a-b}; note that the rank-one structure and the external input pattern shift the critical coupling from $g=1$ to slightly larger values \citep{Rajan2010, Schuecker2017}). For every value of $g$, we observe that the approximate mean-field description correctly captures performance in LS-trained networks when the two input vectors $\mathbf{m}$ and $\mathbf{I}$ are not completely aligned ($\rho<1$, Fig.~\ref{fig:correlations_a} {\bf a}). On the other hand, we find that the theory fails to fully describe degenerate architectures with parallel input vectors ($\rho=1$, Fig.~\ref{fig:correlations_a} {\bf b}): a significant qualitative mismatch is observed within the instability region corresponding to small and negative target values. In such a region, non-linear corrections have very little effect (Fig.~\ref{fig:nonlinear}), and the mismatch is mostly due to correlations between the readout vector and the random bulk. The mismatch indeed increases with the random strength $g$, as the recurrent input deriving from the random connectivity becomes large with respect to the feedback input (Fig.~\ref{fig:correlations_a} {\bf c}). \begin{figure} \begin{adjustwidth}{-0.3in}{0in} \centering \includegraphics{NECO-09-18-3251R2-figure4.pdf} \end{adjustwidth} \caption{ The least-squares solution: effect of correlations. {\bf a-b.} Comparison between the linear mean-field predictions and the trained networks output states for two increasing values of $g$, corresponding to pink ($g=0.6$) and magenta ($g=1.1$) traces. We consider network architectures characterized by two different values of the input vectors overlap: in {\bf a}, we take $\rho = 0.6$, in {\bf b}, $\rho=1$. Note that, especially in {\bf a}, the results for the two values of $g$ strongly overlap. A significative mismatch between theory and simulations is observed in {\bf b}, for $g=1.1$, within the instability area at small and negative target values. {\bf c.} For $\rho=1$, we fix the value of the target $A$ within the instability window (grey arrows in {\bf b}) and we measure the mismatch between mean-field predictions and trained networks as the random strength $g$ is increased. The mismatch is measured at the level of the readout $z$ in correspondence of the spurious fixed point characterized by positive readout values. Parameters and simulations settings are as in Fig.~\ref{fig:LS}. } \label{fig:correlations_a} \end{figure} \begin{figure}[t] \begin{adjustwidth}{-0.3in}{0in} \centering \includegraphics{NECO-09-18-3251R2-figure5.pdf} \end{adjustwidth} \caption{ The least-squares solution: effect of correlations; detailed analysis of the case $\rho=1$ (Fig.~\ref{fig:correlations_a}). The results for the two different values of $g$ are displayed separately in the top ($g=0.6$) and in the bottom ($g=1.1$) panels. {\bf a.} Magnified view of Fig.~\ref{fig:correlations_a} {\bf b}. Arrows on the abscissa axis indicate the target values used in {\bf c}. {\bf b.} Exact prediction for the outlier eigenvalue of the stability matrix of the target fixed point (in green, see \emph{Appendix B}). The outlier eigenvalue lies above the instability boundary within a parameter region corresponding to exact instability area, indicated in light gray. The instability area predicted by the approximate mean-field theory is instead indicated in dark gray. Gray lines indicate the value of the radius of the eigenvalues bulk. {\bf c.} Probing the basin of attraction of the target solution for two different values of $A$, indicated by arrows in {\bf a}. We simulate the post-training network dynamics starting from initial conditions given by Gaussian vectors of variable amplitude added on top of the open-loop activity $\mathbf{x}^{ol}$. The absolute error is then measured at the level of the readout $z$, and is averaged over 40 trained networks. Parameters are as in Fig.~\ref{fig:LS}. } \label{fig:correlations_b} \end{figure} A detailed analysis of degenerate feedback architectures ($\rho=1$) is provided in Fig.~\ref{fig:correlations_b}. In the top row, we select an intermediate value of $g$. In this case, dynamics in trained networks diverge at the border of the instability window as predicted by the mean-field theory (Fig.~\ref{fig:correlations_b} {\bf a}). In the bottom row, we consider a larger value of $g$, laying right below the instability to chaos. In this case, the trained network output appears to obey a different qualitative pattern: the largest stable solution, which is predicted to diverge at $A^*$, converges instead at the border of the instability window. Our approximate mean-field description predicts furthermore a large instability window, wider than the instability region which is predicted by the exact analytical expression \citep{RivkindBarak} (light shaded regions in Fig.~\ref{fig:correlations_b} {\bf a} and {\bf b}). The target fixed point remains thus locally stable within a parameter region where the approximate theory predicts instability. Within the same window, however, the target solution is characterized by a stability eigenvalue laying very close to the instability boundary and a very narrow basin of attraction (Fig.~\ref{fig:correlations_b} {\bf b} and {\bf c}). Trained networks, indeed, converge to the target fixed point only when the initial conditions of the dynamics are very carefully tuned to the open-loop activation vector $\mathbf{x}^{ol}$. Since the dynamics systematically diverge from the target solution when the initial conditions are perturbed along random directions, Fig.~\ref{fig:correlations_b} {\bf c} indicates that the basin of attraction of such a fixed point is particularly narrow with respect to any axis of the phase space. To conclude, we found that correlations impact the agreement between mean-field theory and trained networks within a restricted region of parameters, corresponding to almost parallel input vectors. In that region, a quantitative and qualitative match cannot be obtained when the ratio $g/A$, which measures the relative strength of the recurrent over the feedback dynamics, is large. The approximate mean-field description, however, successfully captures the presence of an instability area, where the target fixed point is unstable or confined into a very narrow basin of attraction. \section{\textsc{Recursive Least-Squares training}} Our analysis of least-squares solutions revealed the presence of a finite parameter region where the target fixed point is constructed on a locally unstable state. Dynamical feedback instabilities have been observed in a variety of studies where, as in the LS case we examined so far, the readout vector $\mathbf{n}$ is fixed through a unique and batch weights update \citep{Jaeger, Lukosevicius, Reinhart}. A convenient strategy to overcome local instabilities consists of training feedback networks with an online algorithm so that the effect of synaptic changes is immediately propagated and tested on the network dynamics. FORCE (first-order reduced and controlled error) learning \citep{SussilloAbbott} is a popular example which relies on an online and recursive formulation of the least-squares algorithm (RLS) \citep{Liu, Jaeger2002}. This algorithm includes a regularization term, whose strength is controlled by a parameter $r$, which penalizes large amplitude readouts. Details of the training procedure are reported in \emph{Appendix E}. In this section, we train feedback networks with FORCE and apply our theoretical framework to derive an approximate but effective description of the reached solutions. Similarly to the previous section, we can investigate how the algorithm implements the task by replacing the trained readout vector with an approximation in the form of Eq.~\ref{eq:sol}. Specifically, we are interested in understanding whether -- and eventually, how -- RLS solutions overcome the local instabilities that are encountered by the batch version of the LS algorithm. \subsection{Parallel geometry} We start our analysis from a specific case, corresponding to a particularly simple network architecture: the two input vectors $\mathbf{m}$ and $\mathbf{I}$ are taken to be parallel ($\rho=1$). In this case, the mean-field theory for uncorrelated solutions (Eq.~\ref{eq:sol}) predicts that a unique solution exists, for which the readout vector $\mathbf{n}$ is parallel to the overlap axis $\bm{\xi}$, while orthogonal readout directions do not affect the output $z$. As in Figs.~\ref{fig:general} {\bf e} and \ref{fig:LS} {\bf c}, such a readout unavoidably generates regions of local and global instability. Here we ask whether RLS solutions implement the task through a similar strategy, or whether alternative and more effective solutions are found outside of the restricted class of readouts which are described by the mean-field theory (Eq.~\ref{eq:sol}). \begin{figure}[t] \centering \includegraphics{NECO-09-18-3251R2-figure6.pdf} \caption{ Recursive least-squares solutions ($\rho=1$): analysis of the readout vector. We consider a set of $N$ unitary and orthogonal vectors $\{\mathbf{v}^k\}_{k=1,..,N}$ which includes the input direction $\bm{\xi}$, and we decompose the trained readout $\mathbf{n}$ on such an orthonormal basis. We construct an histogram by collecting the strengths of the projections of $\mathbf{n}$ along the different basis vectors. Different histogram colors indicate three different training trials, where the norm of the initial guess for $\mathbf{n}$ has been varied. The initial guess is generated as a Gaussian random vector of std $N^s$. The value of $s$ thus controls the amplitude of the initial guess with respect to the network size. The inset reports a magnified view of the histogram for small ordinate values. The grey arrow indicates the average strength of the component of $\mathbf{n}$ along the input axis $\bm{\xi}$. The amplitude of the projections along the remaining basis vectors, instead, is strongly modulated by the initial amplitude $s$. Parameters: $g=0.3$, $\sigma_m=\sigma_I=1.2$, $\rho=1$, $A=1.6$. } \label{fig:RLS_parallel} \end{figure} To start with, we run the RLS algorithm up to convergence and we then analyze numerically the geometry of the reached solution $\mathbf{n}$ (Fig.~\ref{fig:RLS_parallel}). We find that, regardless of the network size, the value of the parameters and the initial conditions, the readout vector is systematically characterized by a large component aligned with the input direction $\bm{\xi}$ (grey arrow in the inset of Fig.~\ref{fig:RLS_parallel}). The readout vector, however, includes also strong components oriented in orthogonal directions: such additional components are induced by non-linearities, correlations with the random bulk and initial conditions of training. In particular, since training convergence is often very fast, a large fraction of orthogonal components is directly inherited from the geometry of the initial guess for the readout $\mathbf{n}$, which is typically generated at random (Fig.~\ref{fig:RLS_parallel}). As a consequence, the amplitude of orthogonal components is strongly affected by the norm of the random vector which is used as initial guess for $\mathbf{n}$. In Fig.~\ref{fig:RLS_parallel}, the initial readout is generated as a Gaussian vector of standard deviation $N^{s}$, so that the amplitude of orthogonal components is controlled by parameter $s$, which measures the amplitude of the initial guess with respect to the bulk size. As a second step, we train feedback networks to different target values, and we attempt to predict performance through the help of the mean-field framework (Fig.~\ref{fig:RLS_parallel2}). As in the previous section, our approximate theoretical description conserves the component of $\mathbf{n}$ which is parallel to $\bm{\xi}$, and discards any orthogonal direction. In Eq.~\ref{eq:sol}, we thus set $p=1$, $p_m = p_I = 0$. \begin{figure}[t] \begin{adjustwidth}{-0.3in}{0in} \centering \includegraphics{NECO-09-18-3251R2-figure7.pdf} \end{adjustwidth} \caption{ Recursive least-squares solutions ($\rho=1$): comparison between mean-field predictions and trained networks. The region of local instability predicted by the theory is displayed in gray. The regions where the mean-field theory predicts the existence of a second stable fixed point are instead indicated in {\bf d}, {\bf e} and {\bf f} by horizontal red stripes. In {\bf a}, {\bf b} and {\bf c}, we quantify the local stability of the training solution. To this end, the post-training activity is simulated for 50 normalized time units, where the initial condition is set by the activation variable during the last training step $\mathbf{x}^{end}$. The error is finally measured as $|z-A|$. In {\bf d}, {\bf e} and {\bf f}, we test whether the network admits two bistable states. We simulate dynamics starting from $\mathbf{x}^{end}$ and $-\mathbf{x}^{end}$ as initial conditions ($\mathbf{n}$ and $-\mathbf{n}$ give similar results), and we compare the final value of the readout $z$ in the two cases. The trained dynamics are considered to be bistable if these two values have opposite sign. Dots show results averaged over 50 network realizations of size $N=600$. Here and in the following, continuous lines do not indicate analytical results, but are drawn to guide the eye. {\bf a, d.} Training results for different values of the regularization parameter $r$ (details provided in \emph{Appendix E}). {\bf b, e.} Training results for different values of the random strength $g$. {\bf c, f.} Training results for different values of the scaling $s$ of the initial guess for the readout $\mathbf{n}$. Parameters as in Fig.~\ref{fig:RLS_parallel}. } \label{fig:RLS_parallel2} \end{figure} As in Figs.~\ref{fig:general} {\bf e} and \ref{fig:LS} {\bf c}, the mean-field theory predicts an area of local instability at small and negative target values (gray shaded regions in Fig.~\ref{fig:RLS_parallel2}). In order to evaluate the local stability of the fixed point generated by FORCE, we compute the post-training dynamics by using the activation vector from the last training step ($\mathbf{x}^{end}$) as an initial condition, and we finally measure the error at the level of the readout $z$. Remarkably, we find that the RLS algorithm often fails to converge to a locally stable solution when the target $A$ is taken within the instability window predicted by the mean-field theory (Fig.~\ref{fig:RLS_parallel2} {\bf a}, {\bf b} and {\bf c}). The failure of the algorithm is thus correctly explained by our crude theoretical approximation, which replaces the complex readout emerging from training with a unique component along $\bm{\xi}$. Although some variability exists, this result is robust with respect to a broad choice of network and training parameters. In particular, the average post-training error decays but remains positive for both small and large values of the random strength $g$ (Fig.~\ref{fig:RLS_parallel2} {\bf b}). Interestingly, training evolves differently in the two cases (Fig.~\ref{fig:training}). For small values of $g$, the algorithm never converges (Fig.~\ref{fig:training} {\bf a}). The readout error remains large during training, while the amplitude of the weights modification imposed by the algorithm smoothly decays. When the value of $g$ is large and close to the instability to chaos, instead, continuous weight modifications make the readout $z$ stably converge to the target within the majority of the trials (Fig.~\ref{fig:training} {\bf b} bottom), resulting in a small average error (Fig.~\ref{fig:RLS_parallel2} {\bf b}). These strongly correlated training solutions, which allow feedback networks to overcome the dynamical instability at small and negative target values, cannot be captured within our approximate theoretical framework. We conclude that large $g$ values, corresponding to strong random connectivities, help the algorithm to find fine-tuned solutions whenever weak and uncorrelated readout vectors (Eq.~\ref{eq:sol}) cannot generate stable fixed points. In certain trials, however, the algorithm fails to converge (Fig.~\ref{fig:training} {\bf b} top), or generates fine-tuned solutions which diverge from the target after learning. As a consequence, the average post-training error never completely vanishes (Fig.~\ref{fig:RLS_parallel2} {\bf b}). \begin{figure}[t] \centering \includegraphics{NECO-09-18-3251R2-figure8.pdf} \caption{ Recursive least-squares solutions ($\rho=1$): evolution of training for a target value within the instability window ($A=-1.1$). The top rows display the time course of the readout variable $z$, while the bottom rows show the amplitude of the synaptic modifications in $\mathbf{n}$ imposed by the algorithm. The training window corresponds to the shaded region, and the target $A$ is indicated by the dashed black line. In {\bf a}: $g=0.3$, in {\bf b}: $g=1.1$. In {\bf b}, we show a non-converging (top) and a converging (bottom) trial. Parameters as in Fig.~\ref{fig:RLS_parallel}. } \label{fig:training} \end{figure} As in Figs.~\ref{fig:general} {\bf e} and \ref{fig:LS} {\bf c}, the mean-field approximation further predicts the existence of an additional stable fixed point for large -- positive and negative -- target values (red shaded regions in Fig.~\ref{fig:RLS_parallel2} {\bf d}, {\bf e} and {\bf f}). Such a fixed point is characterized by a readout value of opposite sign, and is responsible for large readout errors whenever the dynamics is initialized within its basin of attraction. In order to assess bistability, we integrate the post-training dynamics starting from initial conditions centered around $\mathbf{x}^{end}$ and $-\mathbf{x}^{end}$. Again, a good agreement with the mean-field prediction is found: trained networks converge to a second stable fixed point when the target value is taken within the parameter regions where the approximate mean-field network is bistable. The systematic agreement between theory and simulations (Fig.~\ref{fig:RLS_parallel2}) altogether indicates that the readout component along $\bm{\xi}$ plays a major role in shaping the dynamical landscape of trained networks. Orthogonal readout directions, which are artificially included in the readout vector during training, can have strong amplitude (Fig.~\ref{fig:RLS_parallel}) but contribute little to the network output states. \begin{figure}[t] \centering \includegraphics{NECO-09-18-3251R2-figure9.pdf} \caption{ Recursive least-squares solutions ($\rho=1$): analysis of the synaptic connectivity eigenspectra. {\bf a.} Sample eigenspectra of $\bm{\chi} + \mathbf{mn}^T$. Top (resp. bottom) example: the initial guess for $\mathbf{n}$ is weak- (resp. strong-) amplitude: $s=-\infty$ (resp. $s=-0.5$). The dark blue dot indicates the position of the eigenvalue of the rank-one matrix $\mathbf{mn}^T$. The dark blue star indicates the position of the theoretical mean-field prediction. For a readout in the form of Eq.~\ref{eq:sol}, the two values coincide. {\bf b.} Simple estimate for the position of the outlier eigenvalue, measured as the largest real part across all the eigenvalues of the trained connectivity matrix. {\bf c.} Simple estimate for the radius of the circular set of eigenvalues, measured as the largest imaginary part across all the eigenvalues of the trained connectivity matrix. The theoretical predictions are indicated by the black dashed lines. Dots show results averaged over 200 network realizations. The shaded areas indicate the typical size of finite-size fluctuations, measured as the standard deviation of the collected sample. The target is fixed to $A=0.4$, parameters as in Fig.~\ref{fig:RLS_parallel}. } \label{fig:RLS_eig} \end{figure} As a final step, we look at the eigenvalues of the effective connectivity matrix $\chi + \mathbf{mn}^T$ (Eq.~\ref{eq:full_dynamics_rankone}). The theory predicts that, for a solution in the form of Eq.~\ref{eq:sol}, one real outlier appears in the eigenspectrum. In a large network, its value corresponds to the overlap between the fedback vectors $\mathbf{m}^T\mathbf{n}$ \citep{Tao2013, MastrogiuseppeOstojic2}. We find that when the initial guess for the readout $\mathbf{n}$ is small, the amplitude of those additional components is weak (Fig.~\ref{fig:RLS_parallel}), and the eigenspectrum is close to the theoretical prediction (Fig.~\ref{fig:RLS_eig} {\bf a}, top). Spefically, the eigenspectrum includes a single real outlier whose position is well approximated by the scalar product $\mathbf{m}^T\mathbf{n}$. The outlier eigenvalue gets closer to the predicted value as the training parameter $r$, which penalizes large readout weights, is increased (Fig.~\ref{fig:RLS_eig} {\bf b} and {\bf c}). When the initial guess is characterized by strong scaling, instead, the shape of the eigenspectrum becomes very variable from one trial to the other. Typically, the eigenspectrum includes more than one outlier, the position of which fluctuates strongly independently of the theoretical prediction (Fig.~\ref{fig:RLS_eig} {\bf a}, bottom). Taken together, those results suggest that the connectivity eigenspectrum may bear little information about the overall dynamics of the trained network. The eigenspectrum of the effective connectivity matrix $\chi + \mathbf{mn}^T$, indeed, appears to be very sensitive to the readout components which are spuriously introduced by the training algorithm but contribute little to the final output states. Our mean-field description, which specifically ignores such components, successfully captures the fundamental traits of local and global dynamics in trained networks but fails to predict the shape of the complex eigenspectra displayed by the synaptic connectivity matrices. \subsection{Arbitrary geometry} We complete our analysis by considering the RLS solutions which emerge from training in feedback architectures characterized by arbitrary initial geometries ($\rho<1$). In Section 4, we found that the least-squares solution generates a region of local instability when the overlap between the two input vectors $\mathbf{m}$ and $\mathbf{I}$ is larger than a finite limit value (Fig.~\ref{fig:LS}). On the other hand, our theoretical framework suggests that -- as long as the two input vectors are not completely aligned ($\rho<1$) -- a stable mean-field solution always exists. The optimal readout $\mathbf{n}$ includes a strong component along $\eta_{I}$, i.e.~the direction of the external input $\mathbf{I}$ which is not shared with the feedback input vector $\mathbf{m}$ (Fig.~\ref{fig:general} {\bf d}). In Fig.~\ref{fig:RLS_arbitrary} {\bf a}, we select a value of the target for which the LS solution is unstable within a finite parameter region at large $\rho$ values. We then test the stability of the fixed point generated by FORCE training by continuously varying the value of the overlap parameter $\rho$. In agreement with the mean-field theory, we find that the RLS algorithm systematically converges to a locally stable solution whenever the value of the overlap between the two input vectors is significantly smaller than one. Specifically, the algorithm converges inside and outside the shaded region where the least-square solution is unstable. \begin{figure} \centering \includegraphics{NECO-09-18-3251R2-figure10.pdf} \caption{ Recursive least-squares solutions (arbitrary $\rho$): comparison between mean-field predictions and trained networks. {\bf a.} Local stability of the target fixed point in feedback networks trained through the RLS algorithm. Details are as in Fig.~\ref{fig:RLS_parallel2} {\bf a}. We change the initial geometry of the network architecture by increasing values of the overlap $\rho$. The target value is fixed to $A=-1.1$. The yellow shaded area indicates the parameter region where the LS solution is unstable. {\bf b.} Geometry of the LS solution, characterized in terms of the analytical projections of the readout $\mathbf{n}$ onto the three axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta_I}$ (Eq.~\ref{eq:LS_weights}). The top and the bottom rows display results for two different values of $\rho$, indicated in {\bf a} by the black arrows. {\bf c.} Geometry of the RLS solution, computed numerically by projecting the readout $\mathbf{n}$ onto the three axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta_I}$. {\bf d.} Fraction of bistable trained networks, compared with the mean-field prediction extracted from the readout weights in {\bf c}. Details are as in Fig.~\ref{fig:RLS_parallel2} {\bf d}. Results are displayed for three different values of the regularization parameter $r$. Dots show results averaged over 15 network realizations. Parameters as in Fig.~\ref{fig:RLS_parallel}. } \label{fig:RLS_arbitrary} \end{figure} In order to understand how the instability is overcome, we fix the value of the overlap $\rho$ within the instability window, and we systematically compare the geometry of the LS and the RLS solutions (Fig.~\ref{fig:RLS_arbitrary} {\bf b}, {\bf c} and {\bf d} top). As in Section 4, we extract the dominant readout geometries by projecting the readout vectors onto the three orthogonal axes $\bm{\xi}$, $\bm{\eta}_m$ and $\bm{\eta_I}$ (Eq.~\ref{eq:sol}), resulting into sets of weights $(p, p_m, p_I)$ from which dynamics can be predicted. We find that the LS solution contains strong components along $\bm{\xi}$ and $\bm{\eta}_m$, together with a weak and positive component along the non-shared input direction $\bm{\eta}_I$ (Fig.~\ref{fig:RLS_arbitrary} {\bf b} top). Importantly, because of Eq.~\ref{eq:LS_weights}, the value of $p_I$ is constant with respect to the target value $A$. This solution generates a window of local instability at small target values (Fig.~\ref{fig:LS} {\bf b}) because of the interaction between the strong components along $\bm{\eta}_m$ and $\bm{\xi}$, which tend to produce symmetric and bistable solutions, and the small component along $\bm{\eta}_I$, which weakly suppresses bistability and disrupts such a symmetry. In particular, the instability area appears at negative target when the value of $p_I$ is positive. The RLS algorithm, on the other hand, converges to a readout vector which includes a strong component along $\bm{\eta}_I$ (Fig.~\ref{fig:RLS_arbitrary} {\bf c} top). The value of $p_I$, furthermore, is not constant but varies with the value of the target $A$. In particular, the sign of $p_I$ coincides with the sign of the target. The instability region thus gets pushed to positive target values when $A$ is negative, and vice versa. The approximate mean-field description we derive from the RLS weights (Fig.~\ref{fig:RLS_arbitrary} {\bf c} top) confirms that such a region never emerges. Consistently, training always results in a locally stable fixed point. The mean-field description further predicts the existence of a second, locally stable solution at large and positive target values (Fig.~\ref{fig:RLS_arbitrary} {\bf d} top). In agreement, we find that a large fraction of trained networks displays bistability when the target $A$ is taken within such a window. As a last step, we examine a network architecture characterized by a smaller value of the overlap $\rho$, taken outside of the LS instability window (Fig.~\ref{fig:RLS_arbitrary} {\bf b}, {\bf c} and {\bf d} bottom). We observe that, for smaller values of $\rho$, the bistability region predicted by the mean-field theory at large target values disappears (Fig.~\ref{fig:RLS_arbitrary} {\bf d} bottom). Bistability is suppressed because, in correspondence of the same target values, the readout component along the non-shared input direction $\bm{\eta}_I$ is characterized by large relative weight (Fig.~\ref{fig:RLS_arbitrary} {\bf c} bottom). Consistently with the theory, the fraction of trained networks displaying bistability at large $A$ values significantly decreases (Fig.~\ref{fig:RLS_arbitrary} {\bf d} bottom). To conclude, we tested the RLS algorithm on feedback networks characterized by arbitrary geometry, where our simplified theoretical framework predicts the existence of globally stable solutions. We found that the algorithm is systematically able to find readout solutions which generate the stable dynamics that is suitable to solve the task. Stability is achieved thanks to the flexibility of the online algorithm, which reshapes the geometry of the readout vector until a final optimal configuration is found. In contrast with the LS solution, such an optimal readout includes a strong and structured component along the non-shared input direction $\bm{\eta}_I$, which suppresses bistability and prevents local instabilities. \section{Conclusions} Predicting activity in trained recurrent networks is a challenging theoretical problem, due to the disordered synaptic structure and the large number of units \citep{Doya, Barak2017}. Through the recurrent dynamics, global bifurcations can appear as a result of minimal modifications to the synaptic weights and lead to dramatic changes in the network dynamical landscape. Among all the possible architectures, feedback networks represent a class of particularly simple recurrent networks. In these networks, training is restricted to a set of $N$ synaptic weights out of $N^2$, and the task-specific part of the recurrent dynamics is specified by the readout signal alone. In this work, we examined how feedback networks can be trained to perform a single, stationary input-output association. Such a task is simple enough to be implemented in a purely feedforward setup, where no recurrent feedback exists and the reservoir dynamics is driven solely by the external input pattern. In such a case, which corresponds to setting $\sigma_m=0$ in Eq.~\ref{eq:mI}, the mean-field theory suggests that a task implementation can be constructed by aligning the readout vector $\mathbf{n}$ with the external input $\mathbf{I}$. This simple feedforward solution implements the task through a single stable fixed point, and thus possesses optimal stability properties. Throughout the paper, we investigated how more complex solutions, which involve a rank-one recurrent feedback, can be constructed. To this end, we derived a approximate mean-field framework which captures the geometry of the $N$-dimensional readout vector $\mathbf{n}$ with respect to the input vectors $\mathbf{m}$ and $\mathbf{I}$. In this theoretical setup, the random bulk connections are treated as quenched noise, so that correlations between the readout vector and the specific instatiation of the random matrix $\bm{\chi}$ are neglected. Note that the amplitude of the bulk connections $g$ affect the solutions of our mean-field equations only through minor quantitative differences. As a consequence, the space of solutions that our theory describes corresponds to network models where the global connectivity matrix is in practice rank-one \citep{MastrogiuseppeOstojic2}. We used the mean-field theory to show that implementing such a simple fixed-point task in a feedback network can generate output states characterized by non-trivial stability properties. Specifically, we showed that the task can be solved if the readout $\mathbf{n}$ picks and amplifies the feedback component of the input, which is parallel to the vector $\mathbf{m}$ (Fig.~\ref{fig:general} {\bf c}). Similarly to \citet{Hopfield}, this implementation strongly relies on the non-linearity of the dynamics and produces additional spurious attractors. Alternatively, the task can be solved by aligning the readout vector with the external input $\mathbf{I}$ (Fig.~\ref{fig:general} {\bf d}). Such a configuration induces a hidden feedforward structure in the feedback network, which generates a single and stable output fixed point. In \emph{Appendix C}, we showed that this class of solutions performs optimally also in feedback networks characterized by a threshold-linear activation function. The mean-field framework was further used to approximate the more complex, disordered solutions that are obtained via training with standard algorithmic techniques. Specifically, we used the theory to predict the local and the global stability of the target fixed point, and thus to understand how network parameters and target values impact training performance. Despite implying very strong assumptions, we found that the mean-field approximation correctly describes trained networks in a surprisingly broad range of parameters. The approximation only fails to predict the value of the spurious output states in LS training when the architecture is strongly degenerate ($\rho \sim 1$) and the value of the random strength $g$ is large. This non-trivial result indicates that the dynamical mechanisms underlying the input-output association in LS and RLS trained networks mimics the way the same computation emerge in a simple rank-one network \citep{MastrogiuseppeOstojic2}. As a result of our theoretical analysis, we found that a common training algorithm, which relies on the least-square inversion of the constraint $\phi(\mathbf{x})^T \mathbf{n} = A$, returns under certain conditions suboptimal readout solutions. This happens because such a constraint, exactly like our system of mean-field equations (Eq.~\ref{eq:mf}), possibly admits multiple solutions but is blind with respect to the underlying dynamics which control stability. Furthermore, we found that both LS and RLS approaches systematically fail to converge to a locally stable solution when the geometry of the initial network architecture is degenerate, and the two input vectors $\mathbf{m}$ and $\mathbf{I}$ are parallel. Such a failure is explained by the approximate mean-field theory, which suggests that the unique task solution in the form of Eq.~\ref{eq:sol} is characterized by locally unstable dynamics. The mean-field framework, moreover, suggests that introducing extra directions in the network geometry through the non-shared axes $\bm{\eta}_m$ and $\bm{\eta}_I$ expands the space of possible readout solutions. In Fig.~\ref{fig:RLS_arbitrary}, we showed that an online training algorithm like FORCE can take advantage of such a broader set of solutions by selecting the appropriate readout vector which generates stable output states. Degenerate architectures consisting of strongly overlapping input vectors are not common in practical applications, where vectors $\mathbf{m}$ and $\mathbf{I}$ are typically generated at random. Note, however, that strong overlaps can be accidentally introduced in the network architecture when the two vectors $\mathbf{m}$ and $\mathbf{I}$ are uniform or generated through random distributions of large mean. In such cases, the overlap direction coincides with the unitary vector $\mathbf{u}$. Our analysis indicates a possible important role for the parameter $g$, which scales the amplitude of the random bulk of connectivity. We found that increasing the value of $g$ does not significantly influence the structure of the feedback instabilities, which are mostly controlled by the geometry of the input vectors. Furthermore, large $g$ values have essentially no effect on training performance when the target fixed point can be implemented through uncorrelated, mean-field solutions. On the other hand, large random bulks significantly help the RLS algorithm to converge to strong, correlated solutions when uncorrelated solutions cannot satisfy the task. Correlated solutions are likely to play a major role in feedback networks trained to more complex tasks beyond the single input-output association. Understanding the mechanism through which large $g$ values actively supports training is thus an open but crucial question. As a hint, classical studies indicate that convergence of the RLS algorithm is guaranteed under the condition of \emph{persistent excitation} \citep{Bittanti, Kubin}. This condition requires that activity in the bulk, which serves as a set of basis functions, samples during training a large portion of the $N$-dimensional space spanned by the network population. From this perspective, large random bulks contribute to improve \emph{excitation} in two different directions: they slow down the decay of network activity, thus enlarging the training sample, and they increase the amplitude of the activity components that are orthogonal to the input vectors, thus increasing the sampling variance. Our simplified mean-field description is complementary to the approach proposed in \cite{RivkindBarak}, where a similar fixed point task was studied. The mean-field analysis developed in \cite{RivkindBarak} is exact, and takes into account both the non-linear components of the readout vector and its correlations with the random bulk. Such an approach requires an analytical expression for trained readouts, and returns the local stability of the target fixed point. On the other hand, our analysis is approximate, but extends more easily to any readout vector obtained through algorithmic training. Our approach, furthermore, captures dynamics on a global scale, and allows us to predict the existence of spurious, bistable fixed points. The mean-field framework we adopt is flexible enough to be directly extended to network architectures where the input weights are independently generated from non-Gaussian probability distributions or the activation function $\phi(x)$ has a different shape (see \emph{Appendix C}). Significant additional work will be required, instead, to extend our results to networks of spiking units, for which different training algorithms have been proposed \citep{EliasmithAnderson, Boerlin2013, Thalmeier2016, Kim2018}. The mean-field approach we considered here is directly adapted from \citet{MastrogiuseppeOstojic2}. Although the two studies build on the same theoretical tools, they are motivated by deeply different perspectives. In \citet{MastrogiuseppeOstojic2}, the low-rank part of the network connectivity is designed by hand: the overall understanding of the dynamics which emerges from the mean-field framework is exploited to construct low-rank connectivity structures which stably and efficiently implement a variety of behavioral tasks. Crucially, the tasks are in that case defined as qualitative input-output associations rules which fix the overall network behaviour but not the exact value of the readout $z$. In this work, in contrast, the mean-field framework serves as a tool for understanding how an extremely simple task is implemented through training in standard network architectures. As in supervised learning applications, the task is specified by a fixed quantitative constraint on the readout value $z$. Consequently, the resulting target fixed point is not guaranteed to be a stable state for the network dynamics. In \citet{MastrogiuseppeOstojic2}, moreover, both connectivity vectors $\mathbf{m}$ and $\mathbf{n}$ are considered to be plastic. In the network models that are derived, the two vectors serve two different computational roles: vector $\mathbf{n}$ selectively amplifies the external input patterns, while vector $\mathbf{m}$ defines the output direction to be picked by the readout. In the present framework, instead, the network architecture is constrained as in machine learning applications \citep{Jaeger, Lukosevicius, SussilloBarak}: the input weights $\mathbf{m}$ are generated at random and considered to be fixed, while the readout direction is given by the trained vector $\mathbf{n}$. The mean-field analysis adopted in both studies is exact when two specific assumptions are satisfied: the readout vector $\mathbf{n}$ obeys a weak ($1/N$) scaling with the network size, and its entries are statistically uniform and uncorrelated with the random part of the connectivity $\bm{\chi}$. If these hypotheses hold, the resulting feedback network can only display simple or bistable stationary dynamics \citep{MastrogiuseppeOstojic2}, which was suitable in our case for implementing a simple fixed-point task. For a feedback architecture trained to solve a more articulate task, on the other hand, the entries of $\mathbf{n}$ need to be structured and/or fine-tuned to the random bulk. In that case, a mean-field description with the same characteristics would fail to correctly describe the network dynamics. However, one can hypothesize that -- as long as the task is reasonably simple -- the overall connectivity can still be approximated by the sum of a random noise term and a weak and uncorrelated connectivity matrix, where the rank of the latter is larger than one but still much smaller than the network size $N$. If such a low-dimensional decomposition can be extracted from the trained synaptic matrix, then a similar mean-field approach can be directly applied. Specifically, one can hope to predict dynamics by looking at the relative geometrical arrangement of a restricted number of high-dimensional vectors, which define the external inputs together with the low-rank part of the connectivity. Ongoing work indicates that this approach can be used to successfully extract the relevant dynamical mechanisms from recurrent networks trained on more complex tasks, where the spectrum of possible network implementations is not known a priori. \section*{Acknowledgements} This work was funded by the Programme Emergences of City of Paris, Agence Nationale de la Recherche grants ANR-16-CE37-0016-01 and ANR-17-ERC2-0005-01, and the program ``Investissements d'Avenir'' launched by the French Government and implemented by the ANR, with the references ANR-10-LABX-0087 IEC and ANR-11-IDEX-0001-02 PSL University. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

cefebb06e316e54b34851b35882acc72478f6d67

\subsubsection*{\bibname}} \bibliographystyle{apalike} \usepackage{hyperref} \usepackage{url} \usepackage{booktabs} \usepackage{amsfonts} \usepackage{nicefrac} \usepackage{microtype} \usepackage{amsmath,amssymb} \usepackage{lastpage,fancyhdr,graphicx} \usepackage{rotating} \newcommand{\mathop{\mathrm{Tr}}}{\mathop{\mathrm{Tr}}} \newcommand{\mathop{\mathrm{diag}}}{\mathop{\mathrm{diag}}} \newcommand{\mathbf{e}}{\mathbf{e}} \newcommand{\mathbf{h}}{\mathbf{h}} \newcommand{\mathbf{g}}{\mathbf{g}} \newcommand{\mathbf{s}}{\mathbf{s}} \newcommand{\mathbf{c}}{\mathbf{c}} \newcommand{\mathbf{n}}{\mathbf{n}} \newcommand{\mathbf{d}}{\mathbf{d}} \newcommand{\mathbf{b}}{\mathbf{b}} \newcommand{\mathbf{w}}{\mathbf{w}} \newcommand{\mathbf{x}}{\mathbf{x}} \newcommand{\mathbf{y}}{\mathbf{y}} \newcommand{\mathbf{z}}{\mathbf{z}} \newcommand{\mathbf{A}}{\mathbf{A}} \newcommand{\mathbf{E}}{\mathbf{E}} \newcommand{\mathbf{R}}{\mathbf{R}} \newcommand{\mathbf{W}}{\mathbf{W}} \newcommand{\mathcal{R}}{\mathcal{R}} \newcommand{\mathcal{N}}{\mathcal{N}} \newcommand{\mathcal{D}}{\mathcal{D}} \newcommand{\mathcal{L}}{\mathcal{L}} \newcommand{\mathcal{E}}{\mathcal{E}} \newcommand{\mathrm{KL}}{\mathrm{KL}} \newcommand{\hat{x}}{\hat{x}} \newcommand{\hat{y}}{\hat{y}} \newcommand{\hat{\mathbf{y}}}{\hat{\mathbf{y}}} \newcommand{\hat{z}}{\hat{z}} \DeclareMathOperator*{\argmax}{argmax} \DeclareMathOperator*{\argmin}{argmin} \DeclareGraphicsRule{.ai}{pdf}{.ai}{} \begin{document} \twocolumn[ \aistatstitle{A simple probabilistic deep generative model for learning generalizable disentangled representations from grouped data} \aistatsauthor{ Haruo Hosoya } \aistatsaddress{ ATR International, Kyoto, Japan \\ {\tt [email protected]} } ] \begin{abstract} The disentangling problem is to discover multiple complex factors of variations hidden in data. One recent approach is to take a dataset with grouping structure and separately estimate a factor common within a group (content) and a factor specific to each group member (transformation). Notably, this approach can learn to represent a continuous space of contents, which allows for generalization to data with unseen contents. In this study, we aim at cultivating this approach within probabilistic deep generative models. Motivated by technical complication in existing group-based methods, we propose a simpler probabilistic method, called group-contrastive variational autoencoders. Despite its simplicity, our approach achieves reasonable disentanglement with generalizability for three grouped datasets of 3D object images. In comparison with a previous model, although conventional qualitative evaluation shows little difference, our qualitative evaluation using few-shot classification exhibits superior performances for some datasets. We analyze the content representations from different methods and discuss their transformation-dependency and potential performance impacts. \end{abstract} \section{Introduction} \label{sec:intro} Human can effortlessly disentangle complex factors of variations in sensory inputs. For example, when seeing a set of face images like in Figure~\ref{fig:model}A, we can easily recognize that the explaining factors are the content---what the face is shaped like---and the transformation---how the face is posed or expressing. Furthermore, once we discover the factors, we can generalize this knowledge for faces of unseen identities. How can such ability be achieved computationally? The disentanglement problem has recently been attracting much attention in the context of deep generative models (\cite{Kingma:2013tz,Goodfellow:2014td}). Among others, one promising approach is to use a dataset that comes with grouping structure (\cite{Mathieu:2016tn,Bouchacourt:2018vt,Chen:2017tk}). That is, when data items (like images) containing the same content but transformed differently are grouped together, effective disentangling can be achieved by extracting the content as the factor common within a group and the transformation as the factor differentiating the group members.\footnote{Such assumption is not abnormal since a grouped dataset would easily be derived from class-labeled data (by grouping images of the same class) or from video data (by grouping frames in the same or nearby sequence); e.g., Figure~\ref{fig:model}A.} Notably, this approach allows for generalization over novel contents: the learned representation forms a general continuous ``space'' of contents (e.g., a space of facial shapes), which can accommodate unseen contents. Such generalizability is not commonly attainable, in particular, in approaches that only separate classes from transformations (\cite{Kingma:2014uq,Cheung:2014ug,Siddharth:2017ud,Chen:2016tp}). The purpose of this study is to further pursue the group-based approach within a probabilistic generative model. We here develop a new method called group-contrastive variational autoencoders (GVAE), extending variational autoencoders (VAE; \cite{Kingma:2013tz}). The main motivation is that previous methods of the same goal (\cite{Mathieu:2016tn,Bouchacourt:2018vt}) have used rather complicated techniques for exploiting the grouping structure but necessity of such complexity has been unclear. For example, \cite{Mathieu:2016tn} used adversarial learning in combination with VAE, requiring optimization of multiple objectives; \cite{Bouchacourt:2018vt} used ``evidence accumulation'' technique for encoding contents in a VAE-based model, involving complex multiplication of deep net outputs in the objective. In contrast, we use only a simple ``averaging'' of individual content representations to obtain the common content, which is technically not much more than the basic VAE, yet achieve reasonable disentanglement and generalizability over novel contents. To evaluate our approach, we concentrate here on comparison with Multi-Level VAE (MLVAE; \cite{Bouchacourt:2018vt}), which is the closest method to ours. We introduce a quantitative evaluation method using few-shot classification. The rationale is: since a successful representation would cleanly separate content from transformation, it would achieve high performance in classifying test data of novel classes using a small number of examples. While conventional qualitative evaluation indicates little discernible difference between GVAE and MLVAE, our quantitative evaluation shows significant superiority of GVAE for some datasets. For further understanding, we analyze the latent representations and expose the way of coding uncertainty in MLVAE that leads to transformation-dependent representations, potentially conflicting with the goal of disentangling. \begin{figure}[t] \begin{center} \includegraphics[width=8.5cm]{figures-pdf/model.pdf} \caption{(A) Example of grouped data. Each column groups face images of the same person with different views. (B) The graphical model. Each group member $x_k$ is generated from the member-specific transformation variable $y_k$ and the member-general content variable $z$. } \label{fig:model} \end{center} \end{figure} \section{Related work} \label{sec:related} As already mentioned, several studies have investigated disentangling using grouped data. \cite{Mathieu:2016tn} have proposed a combination of VAE and generative adversarial networks (GAN; \cite{Goodfellow:2014td}). They used adversarial learning as a regularization to encourage generated images to be classified as the same when changing the transformation factor. They verbally reported that omitting adversarial learning abolished disentangling, although we show here that it can be achieved without such technique; we believe that this is because we chose a very low dimension for the transformation variable (Section~\ref{sec:disentangling}). In addition, although they pointed out the potential of this approach for generalization over new contents, they demonstrated it only for one dataset with a rather unclean result possibly due to the small data size. The more recent study by \cite{Bouchacourt:2018vt}, developed concurrently with ours, proposed MLVAE, which estimates the common content using evidence accumulation; we provide detailed comparison in Section~\ref{sec:results}. \cite{Chen:2017tk} have extended GAN for disentangling from grouped data. Although their method is simple yet powerful enough to learn content-transformation separation and generate clean and sharp images. However, we should stress that GAN-based approaches usually do not offer a probabilistic generative model (in the form of a joint distribution of input and latent variables) and therefore cannot enjoy the flexibility and extensibility that otherwise come for free (semi-supervised learning, partial observation, etc.). Grouped data are related to sequential data like video. \cite{Yang:2016wg} and \cite{Denton:2017uf} have investigated disentangling for video data. They crucially assumed that the content variable remains similar in consecutive frames, as inspired by the temporal coherence principle (\cite{Foldiak:1991up}). Although this idea itself is somewhat similar to the group-based approach, the sequence-based methods have used additional mechanisms to exploit ordering among data items, e.g., a recurrent neural network to predict future states from past states (\cite{Yang:2016wg}) or adversarial training to take temporal structure into account (\cite{Denton:2017uf}). Supervised or semi-supervised approaches can also achieve effective disentangling by explicitly supplying the content information and inferring the transformation variable as the remaining factor (\cite{Kingma:2014uq,Cheung:2014ug,Siddharth:2017ud}). However, the prior studies in this approach have typically supplied class labels for supervision (since these are the only labels available in most datasets), represented them in categorical variables, and estimated a generative model conditioned on the class. Since such estimated model would have little useful information for unseen classes, generalization for new contents seems difficult. Some unsupervised approaches require no label or grouping in data, but optimizes the efficiency of the latent representation, either by maximizing mutual information between hidden variables in a GAN-based model (\cite{Chen:2016tp}) or by adjusting a so-called $\beta$-term in a variational lower bound (\cite{Higgins:2016vm}). Either study focused on disentangling capability, but not generalizability over novel contents. \section{Method} \label{sec:method} \subsection{Model} \label{sec:model} We assume a dataset $\mathcal{D}$ consisting of $N$ groups, in which each group has $K$ data members, $(x_1^{(n)},\ldots,x_K^{(n)})$, where $x_k^{(n)}\in \mathcal{R}^D$ is a data member indexed by the group number $n$ and the member number $k$. For example, Figure~\ref{fig:model}A shows a set of 5 groups of 3 data members, where each member is an image. (For brevity, we sometimes elide the superscript $(n)$ below.) We assume independence between groups but not members within a group. We intend that each group in the dataset contains different transformations of the same content. (We do not require alignment in transformation at each group member number.) In other words, the factor common among the group members correspond to the content, while the factor differentiating them correspond to the transformation. To extract such common and differentiating factors from the grouped data, we consider the following probabilistic generative model with two types of hidden variables: the (member-specific) transformation variables $y_1,\ldots,y_K\in\mathcal{R}^L$ and the (member-common) content variable $z\in\mathcal{R}^M$ (Figure~\ref{fig:model}B): \begin{align} p(z) &= \mathcal{N}(0,I) \\ p(y_k) &= \mathcal{N}(0,I) \\ p_\theta(x_k|y_k,z) &= \mathcal{N}( f_\theta(y_k,z), I ) \end{align} for $k=1,\ldots,K$. That is, each hidden variable, $z$ or $y_k$, is generated from the standard Gaussian prior. Then, each observed variable $x_k$ is generated from the decoder $f_\theta$ applied to the corresponding individual transformation $y_k$ and the common content $z$, added with Gaussian noise of unit variance, where the decoder $f_\theta$ is a deep net with weight parameters $\theta$. \subsection{GVAE} \label{sec:gvae} For learning, we extend the VAE approach introduced by \cite{Kingma:2013tz}, which uses encoder models based on deep nets to approximate posterior distributions for hidden variables. First, we estimate each transformation $y_k$ from the corresponding input image $x_k$ as follows: \begin{align} q_{\phi,\xi}(y_k|x_k) &= \mathcal{N}\left( g_\phi(x_k), r_\xi(x_k)\right) \end{align} where we use an encoder deep net $g_\phi$ with weight parameters $\phi$ for estimating the mean and another deep net $r_\xi$ (positively valued) with weight parameters $\xi$ for estimating the variance. For inference of content, we could likewise assume a pair of deep nets to estimate the content $z$ from all images $x_1,\ldots,x_K$, but it cannot exploit symmetry in the members of the same group. Instead, our approach estimates the individual content for each image $x_k$ by a pair of deep nets and thereafter averages all the individual contents to obtain the common content $z$: \begin{align} q_{\psi,\pi}(z|x_1,\ldots,x_K) &= \nonumber \\ \quad \mathcal{N}\left( \frac{1}{K} \sum_{k=1}^K h_\psi(x_k), \frac{1}{K} \sum_{k=1}^K s_\pi(x_k) \right) \label{eq:gvae-inf} \span\omit \end{align} The encoder deep nets $h_\psi$ and $s_\pi$ each estimate the mean and variance of the posterior distribution of an individual content. The common content $z$ for all group members is inferred as the average of the individual contents. Note that, therefore, the variance of $z$ becomes the average of the individual variances. The intention here is that, as we attempt to reconstruct each image $x_k$ by the common content $z$ with the individual transformation $y_k$, all the individual contents $h_\psi(x_1),\ldots,h_\psi(x_K)$ are encouraged to converge to an equal value in the course of learning. Thus, $z$ will eventually become a common factor of all $x_k$, while each $y_k$ will become a differentiating factor. To train the model, we consider the following variational lower bound of the log likelihood: \begin{align} \log p_\theta(\mathbf{x}) &\geq \mathcal{E}_{q_{\theta,\phi,\xi,\psi,\pi}(\mathbf{y},z|\mathbf{x})}\left[\sum_{k=1}^K \log p_\theta(x_k|y_k,z)\right] \nonumber \\ &\quad - \sum_{k=1}^K \mathrm{KL}( q_{\phi,\xi}(y_k | x_k) \parallel p(y_k) ) \nonumber \\ &\quad - \mathrm{KL}( q_{\psi,\pi}(z | \mathbf{x}) \parallel p(z) ) = \mathcal{L}(\mathbf{x}) \label{eq:obj} \end{align} where $\mathbf{x}=(x_1,\ldots,x_K)$ and $\mathbf{y}=(y_1,\ldots,y_K)$. Then, our goal is to maximize the lower bound for the whole dataset $\mathcal{D}$: \begin{align} \mathcal{L}(\mathcal{D}) = \frac{1}{N}\sum_{n=1}^{N}\mathcal{L}(\mathbf{x}^{(n)}) \label{eq:obj-full} \end{align} This can be solved by the standard VAE techniques (stochastic variational Bayes, reparametrization, etc.; \cite{Kingma:2013tz}). \subsection{MLVAE} \label{sec:mlvae} MLVAE differs from GVAE only in the way of inferring the content (equation~\ref{eq:gvae-inf}), where the posterior distribution is estimated as the product of Gaussians for individual members (``evidence accumulation''): \begin{align} q_\psi(z|x_1,\ldots,x_K) &= \frac{1}{Z}\prod_{k=1}^K\mathcal{N}\left( h_\psi(x_k), s_\pi(x_k) \right) \label{eq:mlvae-inf-eviacc} \end{align} where $Z$ is the normalizing constant. Since a product of Gaussians is a Gaussian, the above definition can be rewritten as: \begin{align} \mbox{\eqref{eq:mlvae-inf-eviacc}} = \mathcal{N}\left( \frac{ \sum_k{h_\psi(x_k)}/s_\pi(x_k) }{ \sum_k{1/s_\pi(x_k)} }, \frac{ 1 }{ \sum_k 1/s_\pi(x_k) } \right) \span\omit \label{eq:mlvae-inf} \end{align} Note that the mean of the Gaussian has the form of weighted average where the weights are the precisions $1/s_\pi(x_k)$. Thus, MLVAE may appear to be more expressive than GVAE, which uses simple averaging (equation~\ref{eq:gvae-inf}). However, GVAE is not a special case of MLVAE since the variance has a very different form. Due to the weighted average form, the objective of MLVAE is significantly more complex than GVAE. Indeed, the form introduces multiplication and division between the outputs of deep nets, which makes optimization not easy. (This could be slightly simplified by letting a deep net represent the precision $1/s_\pi(x_k)$ instead of the variance $s_\pi(x_k)$, though). In fact, we often encountered training runs that resulted in two kinds of degenerate models, one with all $h(x)$ yielding $0$ and the other with all $s(x)$ yielding $1$. The first kind is clearly a failure in terms of disentangling, while the second kind makes no use of variance estimation and therefore defeats any potential advantage over GVAE. On the other hand, we did not have such difficulty in the case of GVAE. \section{Experimental set-up} \label{sec:exper} We prepared the following three datasets. % (1) Multi-PIE: multi-viewed natural face images derived from \cite{Gross:2010fj}. Grouping is done by the subject and (cloth/hair) fashion, while varying the view and expression. The training and test sets have disjoint 268 and 69 subjects. % (2) Chairs: multi-viewed synthetic chair images derived from \cite{Schuldt:2004ee}. Grouping is done by the chairs type, while varying the view. The training and test sets have disjoint 650 and 159 types. % (3) KTH: image frames from video clips of human (only pedestrian) motion derived from \cite{Schuldt:2004ee}. Grouping is done by the video clip id, while varying the position of the subject. The training and test sets have disjoint 20 and 5 subjects. % See Appendix~\ref{sec:dataset-add} for details of these datasets. Apart from these, we also tested smaller-sized datasets, but these generally produced rather poor disentangling results. For each dataset, we built a GVAE model (Section~\ref{sec:gvae}) or an MLVAE model (Section~\ref{sec:mlvae}) with convolutional neural nets for the encoders ($g$, $r$, $h$, and $s$) and a deconvolutional neural net as the decoder ($f$). We used the same architecture for all models with the transformation dimension $L=3$ and the content dimension $M=100$; See Appendix~\ref{sec:architecture-add} for details of the architecture. To train each model, we first randomly initialized the weight parameters of the encoders and decoder and then optimized the objective~\eqref{eq:obj-full} (either using the inference distribution~\eqref{eq:gvae-inf} or~\eqref{eq:mlvae-inf}) with respect to the weight parameters using the training set. Training proceeded by mini-batches (size 100), where each group was formed on the fly by randomly choosing 5 images according to the dataset-specific grouping strategy described above ($K=5$). We used Adam optimizer (\cite{Kingma:2014us}) with the recommended optimization parameters. \section{Results} \label{sec:results} \subsection{Disentangling and generalizablity} \label{sec:disentangling} \begin{figure*}[t] \begin{center} \includegraphics[angle=90,height=13.5cm]{figures-pdf/result-multipie.pdf} \caption{Qualitative evaluation of disentangling in models trained with Multi-PIE. (A,B) A swapping matrix from a GVAE model for training images (A) or test images (B). Each image in the matrix was generated from the content representation of a sample image in the top row and the transformation representation of another sample image in the left-most column. Red box: sample image. (C,D) Analogous swapping matrices from an MLVAE model. (E,F) Ditto from a VAE model.} \label{fig:result-multipie} \end{center} \end{figure*} We started with a conventional qualitative evaluation of disentangling called swapping. In this, given two lists of inputs $x_1,\ldots,x_I$ and $x_1',\ldots,x_J'$, we show a matrix of images, each generated from the content representation estimated from an input $x_i$ in the first list and the transformation representation estimated from an input $x_j'$ in the second list: $f(g(x_j'),h(x_i))$. Note that we estimated the latent representation from a single image (with no grouping) and thus the encoders $r$ and $s$ were unused. In Figure~\ref{fig:result-multipie}A, we show a swapping matrix for a GVAE model trained with Multi-PIE dataset. The top row and the left-most column show two lists of sample images in the training set. In the matrix, we can observe reasonably successful disentangling: each generated image reflects the subject of the corresponding input in the first list and the view of the corresponding input in the second list. Figure~\ref{fig:result-multipie}B shows a swapping matrix for sample test images. Generalization of disentangling to the test case can be observed with a quality more or less similar to the training case, which is remarkable, given the fact that none of the subjects here had been seen during the training. \begin{figure*}[t] \begin{center} \includegraphics[width=15cm]{figures-pdf/result-fewshot.pdf} \caption{Quantitative evaluation of disentangling using few-shot classification. (A) The success rates (y-axis) of each model trained with Multi-PIE for different numbers of shots (x-axis). The error bars show the standard deviations across the splits. Three instances were examined for each method (legend). (B,C) Analogous results for Chairs and KTH. (D) The success rate (y-axis) for each view (x-axis) from a GVAE model trained with Multi-PIE. Two down-looking views are omitted. (E) Analogous result for Chairs. Two colors correspond to different vertical angles of view (legend). } \label{fig:fewshot} \end{center} \end{figure*} Figure~\ref{fig:result-multipie}CD shows analogous swapping matricies for an MLVAE model also trained with Multi-PIE. Although the generated images were not the same as the GVAE model, the overall quality was quite similar. However, we will see a quantitative difference in Section~\ref{sec:fewshot}. As another comparison, to examine whether grouping of data items was crucial, we redid the same experiment using a dataset with no group ($K=1$). This case in fact corresponds to a basic VAE model with a single variable (\cite{Kingma:2013tz}) since, without grouping, the content and transformation variables can be integrated without loss of generality. The results clearly indicated failure of disentangling: no clear content-transformation separation could be observed and the generated images were often corrupted (Figure~\ref{fig:result-multipie}EF). We found that it was crucial for successful disentangling to choose a very low dimension for the transformation variable $y$ (we used 3 dimensions here). This is because this variable would otherwise learn to represent all aspects of inputs including the contents and therefore the content variable $z$ would become degenerate. Apart from this, we did not observe much effect from changes of other model parameters such as architectures of the encoders and decoders. We conducted the same qualitative evaluation for models trained with Chairs or KTH datasets, each shown in Appendices~\ref{sec:chairs-add} and~\ref{sec:kth-add}. For both datasets, qualitative observations were similar: disentangling was generally reasonable and generalizable to new contents; the qualities of the results were similar between the training and test cases and between GVAE and MLVAE models. \subsection{Few-shot classification} \label{sec:fewshot} \begin{figure*}[t] \begin{center} \includegraphics[width=15.5cm]{figures-pdf/result-analysis.pdf} \caption{Analysis of latent content representations. (A) The distributions of estimated means $h(x)$ for each content dimension in an MLVAE model trained with Multi-PIE. The content dimensions are sorted according to the standard deviation. (C) Blue: the distributions of estimated precisions $1/s(x)$ for each view for each of top 8 content dimensions in the MLVAE model. Red: the view-wise normalized distances between two images generated from latent variables differing in one dimension. (E) The images generated from the latent representations of three sample images while varying the content dimension \#6, \#3, or \#5 in the MLVAE model. The three sample images were chosen so that each gives a high estimated precision at one of the three content dimensions in the MLVAE model. (B,D,F) Analogous results for the GVAE model. } \label{fig:analysis} \end{center} \end{figure*} Since the qualitative evaluation did not reveal any difference between GVAE and MLVAE, we next conducted a quantitative evaluation by measuring performances of few-shot classification. The rationale is that, since the learned content representation is expected to eliminate information on transformation, this should allow for transformation-invariant recognition of objects. In particular, even for novel test classes, it should ideally be sufficient to hold the content representation of a single image for each class in order to classify the remaining images (one-shot learning). Thus, the better the disentangled representation is, the more accurate the few-shot classification should be. Our evaluation procedure goes as follows. For a dataset, we formed 100 splits of gallery/probe of test images, where each gallery included $S$ random images for each class and the corresponding probe included the remaining images. Here, a class refers to each subject/fashion combination in Multi-PIE (378 test classes), to each chair type in Chairs (159 test classes), and to each video clip in KTH (240 test classes). Then, for a trained model, we classified each probe image as the class of the gallery image that had the maximal cosine similarity with the input probe image in the space of content variable $z$; classification succeeded if the inferred class matched the actual class. We measured success rates for each $S=1,\ldots,10$ for each trained model. % We compared three methods, GVAE, MLVAE, and VAE, where each model had the same encoder/decoder architectures (the VAE models were in fact GVAE models trained with ungrouped data). We examined three separately trained models for each method and each dataset. Figure~\ref{fig:fewshot} summarizes the results. First, notice that performance variability across model instances was not negligible even within the same method, which means that comparing different methods by single instances would be meaningless. In particular, in the cases of Multi-PIE and Chairs, performance varied quite largely in MLVAE, possibly because the complex objective introduced many under-performing local optima. The average performance of GVAE was higher compared to MLVAE in the cases of Multi-PIE and Chairs. However, in either case, the best performing MLVAE model was comparable with GVAE. This might mean that the best solution represented somewhat similar information on the content, but MLVAE had a more difficulty in finding such solution. We also found that the best performing MLVAE model for Multi-PIE (not the one shown in Section~\ref{sec:disentangling}) was actually degenerate where the estimated variances $s(x)$ were always $1$. This indicates that the evidence accumulation technique may not actually be so effective for generalizable disentangling; we return to this point in Section~\ref{sec:coding}. However, these results are dependent on the dataset: MLVAE performed moderately better in the case of KTH. The VAE models generally performed quite badly compared to other methods. This is not surprising since the disentangled representations were qualitatively poor as shown in Figure~\ref{fig:result-multipie}EF. As a side interest, we wondered which view of 3D objects led to more successful recognition. Figure~\ref{fig:fewshot}DE show the view-wise success rates of one-shot classification in GVAE models. We found that, in both Multi-PIE and Chairs, diagonal views always gave better success rates than profile or frontal views. This result is intuitive since we can perceive better the entire shape of a face or chair from diagonal views than other views. Perhaps, this is the reason why photos of actors or furniture items are typically taken in diagonal views. \subsection{Encoding strategy of contents} \label{sec:coding} To further understand the quantitative differences, we conducted some analyses on the latent content representation. First of all, Figure~\ref{fig:analysis} plots the estimated mean values $h(x)$ of each content dimension for test images from an (A) MLVAE or (B) GVAE model trained with Multi-PIE, where the content dimensions are sorted by the standard deviation. In both, only a part of the full dimensions were effective and the rest became degenerate. However, the GVAE model used a much fewer dimensions. This tendency was generally observed in all other models we trained. To see why, it is important to see view-specific structure in the estimated variances in each model. In Figure~\ref{fig:analysis}C, each scatter plot (blue) shows the estimated precisions $1/s(x)$ for each view (left-profile, frontal, right-profile, etc.) for one of top 8 content dimensions in the MLVAE model. The estimated precisions tended to be peaked at a particular view while they went down to very low values elsewhere, and the peak view was different from dimension to dimension. This was starkly different from the GVAE case (Figure~\ref{fig:analysis}D), where the estimated precisions were only slightly higher for frontal views compared to profile views for most dimensions. Recall the weighted-average form in MLVAE (equation~\eqref{eq:mlvae-inf}), where each estimated mean is multiplied with the corresponding estimated precision. Since the estimated precisions were very low in the non-peak views, we would expect that the content dimension tended to have particularly strong meaning around the peak views. To confirm this, we estimated the latent variables for each test image and generated two new images using the same variables but modifying one dimension of the content variable to either a small value ($\mu-3\sigma$ using the mean $\mu$ and s.d. $\sigma$ of $h(x)$ for that dimension) or a large value ($\mu+3\sigma$); we then calculated the normalized Euclidean distance between the two new images. In Figure~\ref{fig:analysis}CD, each curve (red) shows such distances for each view for one content dimension. We can see that the distances consistently followed the magnitudes of precisions in both cases. As a consequence, in MLVAE, each dimension had much more influence on the generated images in high-precision views than low-precision views, whereas the contrast was more moderate in GVAE (Figure~\ref{fig:view-dep}B). Also, note that the peak views covered a broad range in MLVAE, while they were concentrated around the frontal view (Figure~\ref{fig:view-dep}A). \begin{figure}[t] \begin{center} \includegraphics[width=8cm]{figures-pdf/result-view-dep.pdf} \caption{(A) The distribution of peak views (in terms of changes in generated images) for the effective dimensions in each model shown in Figure~\ref{fig:analysis} (legend). (B) The distribution of ratios of maximum and minimum changes in generated images in each model. } \label{fig:view-dep} \end{center} \end{figure} To illustrate, we chose three sample images of left-profile, frontal, and right-profile views that elicited large precisions values for dimensions \#6, \#3, and \#5, respectively, in the MLVAE model (each emphasizing left-profile, frontal, and right-profile views). Figure~\ref{fig:analysis}E shows generated images from the three images in the MLVAE model, while varying those three dimensions. We can observe that changes of images tended to be larger for views with high precisions. For example, varying dimension \#6 led to large changes in the left-profile view, but small changes in the other views. Figure~\ref{fig:analysis}F shows analogously generated images for the GVAE model, but changes in all content dimensions tended to alter the generated images equally in all views. This view-dependent coding in MLVAE may explain the larger number of effective dimensions in MLVAE (Figure~\ref{fig:analysis}AB). It is perhaps because different dimensions were used for representing different views and therefore encoding of all views would require a larger number of dimensions than GVAE, in which each dimension seemed to play a role in all views. The view dependency of the estimated precisions may be related to uncertainty coding of facial features. That is, since a single image provides only partial information for the content, the inference necessarily includes ambiguity. For example, from a frontal face image, we are sure about the eyes, nose, and mouth, but less sure about the ears; from a profile face image, we are sure about the visible side of the face, but much less sure about the invisible side. Thus, we might be able to infer a more accurate content representation if we integrate estimated content information for different views with the uncertainty taken into account. However, the view-dependent representation seems to go somewhat in the opposite direction to the goal of disentangling---to discover view-invariant representation---which might be one explanation for the observed lower performance of MLVAE in few-shot classification. \section{Conclusion} \label{sec:concl} To investigate a probabilistic method for learning generalizable disentangled representations, we proposed group-contrastive VAE, which exploits grouping structure in a dataset to extract the content as the common factor and the transformation as the remaining factor. Our approach achieved, for three datasets, effective disentangling of content and transformation and generalizes for test images with new contents. While qualitative comparison indicated little difference from MLVAE, quantitative comparison using few-shot classification showed superiority for some datasets. Our detailed analysis of the content representation revealed transformation-dependent coding arising from the evidence accumulation technique in MLVAE, which may potentially conflict with the goal of disentangling. \section*{Acknowledgments} The author expresses special thanks to Aapo Hyv\"arinen, Mohammad Emtiyaz Khan, and Motoaki Kawanabe for precious comments and suggestions. This work has been supported by Grants-in-Aid for Scientific Research.

9d21a28034ca86857e286bcd61eaef2a7996a45b

\section{Introduction} A complete description of geophysical and astrophysical flows involves taking into account the nonlinear interaction of fluids on a vast range of spatial and temporal scales. Such an undertaking via direct solution of the partial differential equations (often termed Direct Numerical Simulation or DNS) is currently beyond the capability of numerical models, even utilising state-of-the art high performance computation on massively parallel architectures \cite{bauer2015quiet,lorenz1967nature}. A major problem for such schemes is that, for the rotating and stratified systems that are typical of geophysical and astrophysical flows, the large scales influence and are in turn influenced by the smaller scales (which are typically difficult to model)\cite{vallis:2016}. \end{fmtext} \maketitle For this reason, much effort has been dedicated to deriving alternative approaches to DNS that in some way takes into account the small-scale interactions in the large-scale dynamics (and the corresponding effect of large scales on the small scale turbulence). One such approach is that termed Direct Statistical Simulation (DSS). In this approach the statistics of the geophysical and astrophysical turbulence are solved for (rather than the detailed dynamics) \cite{tobias2011astrophysical}. In principle, numerical solution of Fokker-Planck equations may lead to the determination of the pdf of the statistics of turbulent flows, or large deviation theory may give indications of the probabilities of rare events. One such example of DSS that has received much attention recently is a generalisation of the methods of Kolomogorov and Kraichnan (see e.g. \cite{frisch1995turbulence}) to flows that are inhomogeneous and anisotropic (as is typically the case for geophysical and astrophysical flows --- owing to the presence of mean flows, rotation, stratification and possibly magnetic fields. This (equal-time) cumulant expansion method of DSS is able to capture the large scales of turbulent flows with fewer degrees of freedom, as the low-order statistics are spatially smoother than the corresponding dynamical fields and are described by evolution on a slow manifold \cite{marston2014direct}. In order to yield tractable computational problems, the cumulant expansion should be truncated as soon as possible. Truncated at second order (CE2) the method has been shown adequately to represent the statistics of planetary jets and those of the nonlinear magnetorotational instability \cite{sb2015}. Furthermore at this level of truncation the method can be shown to be formally equivalent to the stochastic structural stability theory (SSST or S3T) of Farrell, Ioannou and collaborators \cite{Farrell:2007fq,Constantinou:2013fh}. This method has been utilised to model a number of physical systems and can be justified for systems near equilibrium for which there is a separation of time-scales \cite{Bouchet:2018er}. At this level of truncation the cumulant expansion is the the statistical representation of the Quasilinear (QL) approximation (see e.g. \cite{marston65conover}). However it can be shown that for systems far from statistical equilibrium that CE2 based on zonal averaging (and hence QL) ceases to be an accurate representation \cite{tobias2013direct}. For such systems there are three possibilities for deriving a system that yields a more accurate representation of the statistics of out-of-equilibrium systems. The first is to include eddy/eddy $\rightarrow$ eddy in the cumulant expansion truncation. This leads to the CE3 (or CE2.5) approximation, which has been shown to improve performance of DSS \cite{marston2014direct}. A second possibility is to extend the averaging procedure from zonal averaging to ensemble averaging (see e.g. \cite{bakas2013emergence,atm2017}), thus enabling the description of coherent structures other than zonal means. Finally one may generalise the quasilinear approximation \cite{mct2016}, as described below, and derive a corresponding statistical theory (GCE2). The Generalised Quasilinear Approximation (GQL approximation) was introduced by Marston et al (2016) and its effectiveness in reproducing the fully nonlinear results has been compared with that of the QL approximation for a number of systems including the driving of barotropic jets, the helical magnetorotational instability (HMRI) and rotating couette flow \cite{mct2016,chmt2016,tm2017}. In all of these systems the driving of the velocity arises either via a small-scale forcing or through the boundaries of the system and is independent of the dynamics of the system. Another class of problems, important for geophysical and astrophysical flows, exists however. In this class of convective systems the driving arises through buoyancy and the effectiveness of the driving input into kinetic energy may depend on the state of the system. In this paper we examine the simplest geophysically realistic system that involves the interaction of rotating convection with mean (zonal) flows --- the Busse annulus model. We shall determine the effectiveness of the QL and GQL approximations for this system which may undergo many types of non-trivial dynamics. In the next section we give a brief derivation of the model and the equations, before showing some representative dynamics in section 3. In Section 4 we consider the QL and GQL approximations before giving our conclusions in the final section. \section{Model and Equations} \begin{figure}[!h] \centering \includegraphics[width=4.5in]{busse_diag.pdf} \caption{Schematic diagram of computational domain for the thermal annulus..} \label{fig_busse_pic} \end{figure} The model we consider here is of rotating incompressible Boussinesq fluid (with viscosity $\nu$ and thermal diffusivity $\kappa$) in an annulus of length $L_x$ with vertical sidewalls and weakly tilted top and bottom boundaries (see figure~\ref{fig_busse_pic} and \cite{busse:1976,bh1993,rj2006}). The system rotates at angular velocity ${\boldsymbol \Omega = \Omega {\bf e}_z}$ and gravity is uniform and in the $y$-direction. The angle of inclination of the top and bottom boundaries are given by $\theta_T$ and $\theta_B$ respectively. The system is non-dimensionalised, scaling lengths with the width of the annulus in the $y$-direction ($d$), time with the viscous timescale $d^2/\nu$ velocites with $\nu/d$ and temperatures with the temperature difference between the inner and outer walls ($\Delta T$). The dynamics of the system is then controlled by evolution equations for the non-dimensional velocity ${\bf u}$ and temperature $T$, which depend on three non-dimensional parameters; the modified Rayleigh number, the Ekman number and the Prandtl number given by \begin{equation} Ra^\prime = \dfrac{\alpha g \Delta T d}{\nu \Omega}, \quad E = \dfrac{\nu}{\Omega d^2}, \quad Pr = \dfrac{\nu}{\kappa}. \end{equation} Here $\alpha$ is the coefficient of thermal expansion at constant pressure. The temperature is decomposed $T = T_{BS} +{\hat \theta(x,y)}$ into a basic state profile $T_{BS}$ satisfying $\nabla^2 T_{BS} = 0$ and a perturbation ${\hat \theta}$. Here we take $T_{BS} = y$ for the basic state. Once stress-free boundary conditions have been imposed on the vertical walls and no-slip boundary conditions are imposed on the top and bottom boundaries then for two-dimensional quasi-geostrophic models the velocity and temperature can be written as \begin{equation} {\bf u} = -\boldsymbol{\nabla} \times (\psi(x,y) {\bf e}_z) + {\bf u}^\prime({\bf r}) \end{equation} with \begin{equation} | {\bf u}^\prime| \ll | \boldsymbol{\nabla} \times (\psi(x,y) {\bf e}_z)| \end{equation} and \begin{equation} {\hat \theta} = {\hat \theta}(x,y), \end{equation} and the evolution equations for $\psi$ and ${\hat \theta}$ are given by (see \cite{rj2006}) \begin{eqnarray} \dfrac{\partial \nabla^2 \psi}{\partial t} + J(\psi, \nabla^2 \psi) - \beta \dfrac{\partial \psi}{\partial x} &=& -\dfrac{Ra}{Pr} \, \dfrac{\partial {\hat \theta}}{\partial x} - C |\beta|^{1/2} \nabla^2 \psi + \nabla^2 \nabla^2 \psi, \label{psieqn}\\ \dfrac{\partial {\hat \theta}}{\partial t} + J(\psi, {\hat \theta}) &=& - \dfrac{\partial {\psi}}{\partial x} + \dfrac{1}{Pr} \nabla^2 {\hat \theta}, \label{psiTeqns} \end{eqnarray} where $J(A,B) = \frac{\partial A}{ \partial x} \frac{\partial B}{ \partial y} - \frac{\partial A}{ \partial y} \frac{\partial B}{ \partial x}$ is the Jacobian and the parameters are given by \begin{equation} Ra = \dfrac{\alpha g \Delta T d^3}{ \nu \kappa}, \quad \beta = \dfrac{(2 (\theta_T - \theta_B)d)}{(L_z E)}, \quad C = \left(\dfrac{2d}{|\theta_T-\theta_B| L_z}\right)^{1/2}\quad Pr = \dfrac{\nu}{\kappa}. \end{equation} Hence $\beta$ measures the degree of vortex stretching engendered by the sloping endwalls and $C$ measures the degree of friction. In general increasing $\beta$ and $C$ is expected to lead to an increase in the number of the jets (see later). It is equations~(\ref{psieqn}-\ref{psiTeqns}) and their QL and GQL counterparts (given in section~\ref{GQL}) that form the basis of this paper. These are integrated numerically using the pseudospectral PDE solving package Dedalus (REF) \subsection{Representative Dynamics.} The system described above has been extensively studied utilising DNS by a number of authors (for example \cite{busse:1976,bh1993,rj2006}). The dynamics of the system is complicated and involves the interaction of thermal convection, rotation and zonal flows. The dynamics is known to depend critically on the rotation rate (as measured by the Ekman number and hence $\beta$) and the degree of supercriticality (as measured by the Rayleigh number). Less is known about the role of the Prandtl number in the system, though in related systems this has been shown to have a significant role in determining the form of the solutions \cite{gc2016}. In this paper we fix $Pr=1$ and and show representative dynamics for varying $\beta$, $C$ and Rayleigh number $Ra$. \begin{center} \begin{table}[!h] \caption{Parameters for DNS runs \label{table} \begin{tabular}{cccccc \hline Run &$\beta$ &$Ra$ &$C$&Solution & Resolution ($ny \times nz$) \\ \hline A &$2.8 \times 10^3$ &$7.6 \times 10^4$ &$0$ &Large-scale jets& $256 \times 64$\\ B &$7.07 \times 10^5$ &$1 \times 10^8$ & $0.316$&Multiple jets& $512 \times 256$ \\ C &$5.00 \times 10^5$ & $8 \times 10^7$& $0$&Bursting Jets& $512 \times 256$\\\hline \end{tabular} \vspace*{-4pt} \end{table \end{center} \begin{figure}[!h] \centering \includegraphics[width=2.5in]{./theta_snap_r78.png} \includegraphics[width=2.5in]{./ux_snap_r78.png} \includegraphics[width=2.5in]{./theta_snap_b75.png} \includegraphics[width=2.5in]{./ux_snap_b75.png} \includegraphics[width=2.5in]{./theta_snap_b55.png} \includegraphics[width=2.5in]{./ux_snap_b55.png} \includegraphics[width=2.5in]{./theta_snap_b55_conv.png} \includegraphics[width=2.5in]{./ux_snap_b55_conv.png} \caption{Snapshots of horizontal velocity $u_x$ and $\theta$ in the saturated regime for (a,b) Case A (c,d) Case B (e,f) CASE C (during the strong shear phase) (d) CASE C (during the weak shear phase).} \label{fig_dyn_snap} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in, height=2in]{./uxRa76000_all.png} \includegraphics[width=2.5in, height=2in]{./ux_7e5_all.png} \includegraphics[width=2.5in, height=2in]{./ux5e5Ra8e8_all.png} \caption{Hovmoller plots of DNS solutions for (a) Case A (b) Case B (c) Case C as defined in Table 1.} \label{fig_dyn} \end{figure} A summary of the dynamics and the parameters for which they are found is given in Table~\ref{table}. Case A is moderately rotating and the Rayleigh number is not high. For this choice of parameters, the convective turbulence interacts with the rotation to produce large-scale jets --- here large-scale indicates that the jets are found on the scale of domain whilst the convective cells take a smaller scale. Snapshots of the temperature perturbuation $\theta$ and the horizontal velocity $u_x$ in the saturated regime shows both the jet structure and the mechanism for the formation of the jet see Figure~\ref{fig_dyn_snap}(a,b). Convective cells, in this case stretching across the domain, interact with the rotation and tilt away from the vertical. This non-trivial systematic tilt leads to the formation of non-zero Reynolds stresses and hence to zonal flows (see \cite{bh1993,jra2003}). In this case the zonal flows are prograde at the top of the domain and retrograde at the bottom. However, owing to the Boussinesq up-down symmetry of the system, solutions with the other parity can be found for different initial conditions. Similar behaviour leading to the generation of mean flows can be found for regular rotating convection in a plane layer in the presence of a non-aligned gravity and rotation vector \cite{ct2016}. Increase of the rotation rate (and corresponding increase in $Ra$) as in Case B, leads to convection cells being driven on much smaller lengthscales (see Figure~\ref{fig_dyn_snap}(c)). In turn the Reynolds stresses are modified so that multiple jets are found. In this case 7 jets are formed. Although these these are now on a scale smaller than the computational domain, they are still on a scale larger than that for the convective cells, as shown in figure~\ref{fig_dyn_snap}(c,d). In both Case A and B the dynamics reaches a statistically steady state where the mean flows are quasi-steady as shown in the Hovmoller plots in Figure~\ref{fig_dyn}(a,b). These demonstrate that the jets driven by the Reynolds stresses are fixed in position. \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./ux_clup.png} \includegraphics[width=2.5in]{./theta_clup.png} \caption{Close up Hovmoller plots for (a) $\langle u_x \rangle$ and (b) $\theta$ for DNS of the bursting solution of Case C showing relaxation oscillations.} \label{fig_clup} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=5.5in]{./Triads.pdf} \caption{Feynman Diagrams showing retained and discarded mode interactions.} \label{interactions} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./uxRa76000_all.png} \includegraphics[width=2.5in]{./uxRa76000_five.png} \includegraphics[width=2.5in]{./uxRa76000_one.png} \includegraphics[width=2.5in]{./uxRa76000_zero.png} \caption{Hovmoller plots for Case A for (a) DNS (b) GQL with $\Lambda=5$ (c) GQL with $\Lambda=1$ and (d) QL.} \label{fig_hov_ls} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./SC_Ra76000_test0p25_all_theta.png} \includegraphics[width=2.5in]{./SC_Ra76000_test0p25_five_theta.png} \includegraphics[width=2.5in]{./SC_Ra76000_test0p25_one_theta.png} \includegraphics[width=2.5in]{./SC_Ra76000_test0p25_zero_theta.png} \caption{Second Cumulants $c_{\theta \theta}$ for Case A for (a) DNS (b) GQL with $\Lambda=5$ (c) GQL with $\Lambda=1$ and (d) QL.} \label{fig_sc_76} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./ux_7e5_all.png} \includegraphics[width=2.5in]{./ux_7e5_five.png} \includegraphics[width=2.5in]{./ux_7e5_one.png} \includegraphics[width=2.5in]{./ux_7e5_zero.png} \caption{Hovmoller plots for Case B for (a) DNS (b) GQL with $\Lambda=5$ (c) GQL with $\Lambda=1$ and (d) QL.} \label{fig_hov_ss} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./ra78_avg.pdf} \includegraphics[width=2.5in]{./Beta75_avg.pdf} \caption{Comparison of mean flows averaged over last third of evolution for (a) Case A (b) Case B.} \label{fig_av} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./SC_BETA7E5_test0p25_all_theta.png} \includegraphics[width=2.5in]{./SC_BETA7E5_test0p25_five_theta.png} \includegraphics[width=2.5in]{./SC_BETA7E5_test0p25_one_theta.png} \includegraphics[width=2.5in]{./SC_BETA7E5_test0p25_zero_theta.png} \caption{Second Cumulants $c_{\theta \theta}$ for Case B for (a) DNS (b) GQL with $\Lambda=5$ (c) GQL with $\Lambda=1$ and (d) QL.} \label{fig_sc_ss} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.75in]{./ke_all.pdf} \includegraphics[width=2.75in]{./ke_all_zoom.pdf} \caption{Timeseries of kinetic energy density for Case C (a) full timeseries (b) zoom of timeseries showing relaxation oscillations.} \label{fig_ke_burst} \end{figure} \begin{figure}[!h] \centering\includegraphics[width=2.5in]{./ux5e5Ra8e8_all.png} \includegraphics[width=2.5in]{./ux5e5Ra8e8_fifteen.png} \includegraphics[width=2.5in]{./ux5e5Ra8e8_ten.png} \includegraphics[width=2.5in]{./ux5e5Ra8e8_five.png} \includegraphics[width=2.5in]{./ux5e5Ra8e8_one.png} \includegraphics[width=2.5in]{./ux5e5Ra8e8_zero.png} \caption{Hovmoller plots for Case C for (a) DNS (b) GQL with $\Lambda=15$ (c) GQL with $\Lambda=10$ (d) GQL with $\Lambda=5$ (e) GQL with $\Lambda=1$ and (f) QL.} \label{fig_hov_burst} \end{figure} Figures~\ref{fig_dyn}(c) show the dynamics of the large-scale jets for Case C via a long time-scale Hovmoller plot. Here the jets although large-scale exhibit non-trivial temporal variation. Up until $t \approx 1$ the large-scale flow has three quasi-steady jets; however this behaviour is transient and the solution eventually takes the form of two jets. Interestingly at this point the system exhibits bursting between states in the form of a relaxation oscillator. A typical cycle of the oscillation proceeds as follows (as shown in Figure~\ref{fig_clup}). Convection interacts with the rotation to drive a mean shear (as before). However in this case the shear is strong enough to act as a barrier to transport and hence suppress the convection (which is itself the source of the shear). The shear therefore slowly decays and the transport barrier disappears. The cycle is then free to proceed again. This predator-prey type dynamics between the mean flows and the turbulence has been seen before in a number of systems involving zonal flows; these include models of convection, the interaction of zonal flows with turbulence in tokamak plasmas, dynamo models \cite{t1997,kgd2015,bs2006} and the magnetorotational instability \cite{OML2011}. Figure~\ref{fig_clup} show snapshots of $u_x$ and $\theta$ in both the weak and strong convection phases. The dynamics of the annulus system is therefore rich and varied. Complicated interactions between mean flows, turbulence and driving lead to non-trivial dynamics and mean flows and so this system represents a formidable challenge for approximations such as QL and GQL. It is the efficacy of these approximations that we shall investigate in the next section. \subsection{Evaluation of the QL and GQL approximations} \label{GQL} In order to construct the QL and GQL systems we perform an idempotent decomposition of the streamfunction and temperature into large-scale and small-scale modes, i.e.\ we set, say, $\psi(x, y)= \psi_l + \psi_h$, where \begin{equation} \psi_l(x, y) =\sum_{k=-\Lambda}^{\Lambda}{\psi}_{k}(y) \,e^{i {2 \pi k x}/{L_x}}, \quad \quad \psi_h = \psi - \psi_l, \end{equation} where $\psi_l$ and $\psi_h$ are the `low' and `high' wavenumber modes respectively. Similarly ${\hat \theta}(x,y) = {\hat \theta}_l+{\hat \theta}_h$. Having made this decomposition we derive equations for the evolution of the low and high modes, making use of the Generalised Quasilinear Approximation. That is we set retain the nonlinear interaction between low modes and low modes and that between high modes and high modes to yield low modes in the low mode equation and also retain the interaction between low modes and high modes to yield high modes in the high mode evolution equation. All other nonlinear interactions are discarded (as shown in Figure~\ref{interactions}. As this ensures that triad interactions are removed in pairs, this process is an example of constrained triad decimation in pairs \cite{kraichnan1985} and is therefore guaranteed to conserve quadratic invariants in the dissipationless system. Crucially we perform this decimation for both the inertial nonlinearity in the Navier-Stokes equation {\it and} the advective nonlinearity in the temperature equation. Hence we set \begin{eqnarray} \dfrac{\partial \nabla^2 {\psi}_l}{\partial t} + J({\psi}_l, \nabla^2 {\psi}_l) + J({\psi}_h,\nabla^2 {\psi}_h)- \beta \dfrac{\partial {\psi}_l}{\partial x} &=& -\dfrac{Ra}{Pr} \, \dfrac{\partial {\hat \theta}_l}{\partial x} - C |\beta|^{1/2} \nabla^2 {\psi}_l + \nabla^2 \nabla^2 {\psi}_l, \nonumber\\ \dfrac{\partial \nabla^2 {\psi}_h}{\partial t} + J({\psi}_l, \nabla^2 {\psi}_h) +J({\psi}_h, \nabla^2 {\psi}_l) - \beta \dfrac{\partial {\psi}_h}{\partial x} &=& -\dfrac{Ra}{Pr} \, \dfrac{\partial {\hat \theta}_h}{\partial x} - C |\beta|^{1/2} \nabla^2 \psi_h + \nabla^2 \nabla^2 \psi_h,\nonumber\\ \dfrac{\partial {\hat \theta}_l}{\partial t} + J(\psi_l, {\hat \theta}_l) + J(\psi_h, {\hat \theta}_h)&=& - \dfrac{\partial {\psi_l}}{\partial x} + \dfrac{1}{Pr} \nabla^2 {\hat \theta}_l,\nonumber\\ \dfrac{\partial {\hat \theta}_h}{\partial t} + J(\psi_l, {\hat \theta}_h) + J(\psi_h, {\hat \theta}_l) &=& - \dfrac{\partial {\psi}_h}{\partial x} + \dfrac{1}{Pr} \nabla^2 {\hat \theta}_h. \label{psiTeqns_gql} \end{eqnarray} The QL system can therefore be recovered by setting $\Lambda=0$ (so that only the $k_x=0$ mode is counted as a low mode), whilst DNS is reproduced by setting the cut-off $\Lambda$ to be the highest wavenumber of the spectral truncation of DNS. Intermediate values of $\Lambda$ yields the GQL approximation. We start by considering how well QL and GQL perform for the steady large-scale jets of Case A. Figure~\ref{fig_hov_ls} shows a comparison of the Hovmoller plots for $\langle u_x \rangle$ as a function of $y$ and $t$ for the cases of DNS, $\Lambda=5$, $\Lambda=1$ and QL ($\Lambda=0$). These indicate that GQL performs well even at $\Lambda=5$ and $\Lambda=1$, with both of these settling down into a two-jet solution (note that the sign of the jet is not important owing to the Boussinesq symmetry). However QL performs not so well, selecting a weaker three-jet solution. This is confirmed in the time averages of $\langle u_x \rangle$ (denoted $\langle u_x \rangle_t$) shown in Figure~\ref{fig_av}(a). Here the time average is taken over the last third of the evolution. Clearly GQL (even at a cutoff $\Lambda=1$) is performing well in representing the first cumulant (or zonal mean). Accurately describing the second cumulants represents more of a test for the QL and GQL approximation. Figure~\ref{fig_sc_76} shows, the second cumulant for temperature $c_{\theta \theta}$ defined as \begin{equation} c_{\theta \theta}(\xi,y_1,y_2) = \int \theta'(x_1,y_1) \theta'(x_1+\xi,y_2) dx_1, \end{equation} where $\xi = x_2 - x_1$. The cumulant is shown, averaged over the last third of the temporal evolution, for a representative choice of $y_1 = 0.25$ as a function of $\xi$ and $y_2=y$, for DNS, QL and GQL (at $\Lambda=1$, $5$). For DNS the second cumulant, as expected peaks around the reference point $(\xi,y)=(0,0.25)$ and with a lengthscale of variation given by the width of a typical convective structure and a height of half the domain. The structure of the second cumulant also displays a characteristic tilt, which leads to the non-trivial Reynolds stress and is also shaped by the shear --- note that the sense of the tilt is reversed if the shear is reversed Far from the reference point the correlations die away and hence the second cumulant tends to zero. GQL (at both $\Lambda=1$, $5$) does a good job of reproducing this structure, but QL overstates the importance of long-range correlations, as it overrepresents the importance of thermal Rossby waves in the system. This type of behaviour has been seen before for QL systems \cite{mct2016,chmt2016,tm2017}. Figure~\ref{fig_hov_ss} compares the evolution of the mean flows for Case B for the various approximations with that of DNS. All the approximations are able to capture the driving of multiple jets by the convection, however the QL approximation does not reproduce the correct number of jets (having two more). The GQL approximation performs better again, with the correct number of jets of approximately the correct amplitude (as shown in Figure~\ref{fig_av}(b)). The second cumulants in Figure~\ref{fig_sc_ss} show the localised nature of the interactions owing to the reduction in lengthscale of the convection. GQL again provides a good approximation of the correlations. As the number of modes in the GQL approximation is decreased the approximation becomes unsurprisingly worse, with the amplitude of the local correlation being underestimated and the correlations becoming delocalised, again because of the overestimation of the stability of thermal Rossby waves. FInally in this section we consider the bursting case C. For this case, in addition to QL and GQL at $\Lambda=1$ and $\Lambda=5$ we also consider GQL at $\Lambda=10$ and $\Lambda=15$. Figure~\ref{fig_ke_burst}(a) shows the timeseries of the kinetic energy for DNS, GQL (at the various truncations) and QL, with a zoom of the saturated regime shown in Figure~\ref{fig_ke_burst}(b). It is clear that, for this complicated solution, more modes are required in GQL in order for the bursting solution to be accurately represented. If QL or GQL at $\Lambda=1$ or $\Lambda=5$ is used then the energy is significantly lower than the ``true'' solution. The reason for this is shown in the Hovmoller plot of Figure~\ref{fig_hov_burst}. QL and low truncation GQL gets both the amplitude and the number of jets incorrect. Interestingly, although QL exhibits bursting behaviour, low truncation GQL does not (at least for these initial conditions). However by the time $\Lambda=10$, GQL does manage to reproduce faithfully the amplitude, number and bursting of jets. For such a complicated solution, it is not clear what is the correct procedure averaging the second cumulants, i.e. whether to average over the whole solution or over bursts and troughs separately so we do not include such a plot in the description of the results. \section{Conclusions} We have constructed GQL models for three distinct flow regimes in a Busse annulus model. For each, GQL accurately reproduces key features from the DNS results that QL calculations do not. In particular, for the steady cases (case A and B), GQL with a single extra retained mode ($\Lambda = 1$) correctly predicts the number and amplitude of jets and compares favorably with the DNS for the second cumulant. With $\Lambda = 5$, the GQL second cumulant matches very well to the DNS. For Case C, the significantly more complex bursting dynamics can also be reproduced using GQL, though $\Lambda \ge 10$ in order to do so. The $\Lambda = 10$ case reproduces the amplitude, jet structure, and time-dependence of the relaxation oscillations. These flows demonstrate that GQL significantly outperforms QL in three dynamically different, far-from-equilibrium flows. In each case, a small number of retained low modes are sufficient to approximate the flow. We have also demonstrated that the rich dynamics offered by Busse annulus flows are particularly challenging to standard QL techniques. Given that closures at second order (e.g. CE2) represent an efficient trade-off between dimensionality and performance, we take the present results to bolster previous results \cite{mct2016,chmt2016,tm2017} that point toward the development of GCE2 techniques as a particularly promising path toward efficient, accurate numerical models of complex geophysical and astrophysical flows. \enlargethispage{20pt} \ethics{This work only involves data obtained from computer simulations.} \dataccess{The computer code that solved the GQL equations is available as supplementary material.} \aucontribute{All authors contributed equally to the design, implementation and analysis of the numerical experiments. All authors gave final approval for publication.} \competing{We have no competing interests.} \funding{Insert funding text here.} \ack{Insert acknowledgment text here.}

b713f5f882366e493ffb51e1ffb37db7dbdd8f6c

\section*{\refname}} \begin{abstract} Buried-channel semiconductor heterostructures are an archetype material platform to fabricate gated semiconductor quantum devices. Sharp confinement potential is obtained by positioning the channel near the surface, however nearby surface states degrade the electrical properties of the starting material. In this paper we demonstrate a two-dimensional hole gas of high mobility ($5\times 10^{5}$ cm$^2$/Vs) in a very shallow strained germanium channel, which is located only 22 nm below the surface. This high mobility leads to mean free paths $\approx6 \mu m$, setting new benchmarks for holes in shallow FET devices. Carriers are confined in an undoped Ge/SiGe heterostructure with reduced background contamination, sharp interfaces, and high uniformity. The top-gate of a dopant-less field effect transistor controls the carrier density in the channel. The high mobility, along with a percolation density of $1.2\times 10^{11}\text{ cm}^{-2}$, light effective mass (0.09 m$_e$), and high g-factor (up to $7$) highlight the potential of undoped Ge/SiGe as a low-disorder material platform for hybrid quantum technologies. \end{abstract} \pacs{} \maketitle \setlength{\textwidth}{183mm} \section*{Introduction} Germanium (Ge) has the highest hole mobility of common semiconductors and is integrated onto Si substrates within a foundry-qualified process\cite{Pillarisetty2011academic}. These properties make high-speed Ge transistors appealing for extending chip performance in classical computers beyond the limits imposed by miniaturization. Ge is also emerging as a promising material for quantum technology as it contains crucial parameters for semiconducting, superconducting, and topological quantum electronic devices. The high mobility of holes and their low effective mass promote the confinement of spins in low-disorder Ge quantum dots by uniform potential landscapes\cite{Hendrickx2018gate}. Holes in Ge have large and tunable $g$-factors, with inherent strong spin-orbit interaction. These properties facilitate fast all-electrical qubit control\cite{watzinger_ge_2018}, qubit coupling at a distance via superconductors\cite{kloeffel2013circuit}, and are key ingredients for the emergence of Majorana zero modes for topological quantum computing. Like Si, Ge can be isotopically purified into a nuclear spin-free material to achieve long spin lifetimes\cite{itoh_isotope_2014}. In contrast, virtually every metal on Ge, including superconductors with high critical fields, show a Fermi level pinned close to the valence band\cite{dimoulas_fermi-level_2006}. This facilitates the injection of holes and thus the formation of Ohmic superconductor/semiconductor contacts, a key building block in hybrid quantum devices. These enticing prospects have motivated the theoretical framework for Ge-based spin-qubits\cite{terrazos2018light} and Majorana fermions\cite{maier2014majorana}. Experimental milestones in self-assembled Ge nanostructures include gate-tunable superconductivity in Ge/Si nanowires\cite{xiang2006ge} and the demonstration of electrically driven spin qubits\cite{watzinger_ge_2018} and single-shot readout of single spins\cite{vukuvsic2018single} in Ge hut wires. \begin{figure}[H] \includegraphics[width=\linewidth]{Fig1_raster}% \caption{Schematics of SiGe heterostructures band-edge profiles as a function of the Ge concentration \textit{x} and \textit{y} in the active material and in the relaxed substrate, respectively. Star, circle, and triangle refer to Ge/SiGe, Si/SiGe and Ge/Si heterostructures, respectively. Adapted from Ref.\cite{virgilio2006type}} \label{fig:play} \end{figure} \begin{figure*} \includegraphics[width=\linewidth]{Fig2} \caption{Structural characterisation of a Ge/SiGe heterostructure. (a) Layer schematics. (b) Ge, Si, and O signals from transmission electron microscopy with energy dispersive X-Ray analysis (STEM/EDX) of the Ge quantum well and nearby Si$_{0.2}$Ge$_{0.8}$; (c) STEM and (d-f) TEM showing the SiGe virtual substrate, quantum well, quantum well/barrier interface, barrier/surface interface. (g) X-ray diffraction reciprocal space map of (-2-24) reflection.} \label{fig:str} \end{figure*} Recently, Ge/SiGe heterostructures have emerged as a planar technology that can bring together low disorder, potential for fast qubit driving, and avenue for scaling due to the compatibility with large scale manufacturing. In Ge/SiGe, the band-edge profiles between compressively strained Ge and relaxed Ge-rich Si$_{1-y}$Ge$_y$ (Fig. \ref{fig:play}; star, $y\approx0.8$) produce a type I band alignment\cite{virgilio2006type}. This is different from Si/SiGe heterostructures (Fig. \ref{fig:play}; circle, $y\approx0.3$) and Ge/Si core-shell nanowires (Fig. \ref{fig:play}; triangle), where a type II band alignment promotes confinement of either electrons or holes, respectively. Charge carriers can populate the quantum well either by doping the heterostructure or via top gating. Holes confined in modulation doped Ge/SiGe have shown exceptionally high mobility of 1.5 million cm$^2$/Vs, strong spin orbit coupling\cite{failla_terahertz_2016}, and fractional quantum hall physics\cite{shi2015spinless}. \begin{figure*} \includegraphics[width=\linewidth]{Fig3} \caption{Ge/SiGe heterostructure field effect transistor. (a) STEM/EDX (top) and TEM (bottom) in the ohmic contact region. (b) Device cross-section and optical image (upper and lower panels, respectively). (c) STEM/EDX (top) and TEM (bottom) in the channel region. (d) Bandstructure simulation with heavy holes (HH), light holes (LH) and charge distribution in the first subband HH0.} \label{fig:dev} \end{figure*} In analogy to electron spin qubits in Si/SiGe\cite{maune2012coherent} it is preferable to completely eliminate dopant atoms, which are a major source of scattering, charge noise, and instability for the low-temperature operation of quantum devices\cite{borselli2011pauli}. Gate controlled quantum dots\cite{Hendrickx2018gate}, ballistic one dimensional channels\cite{mizokuchi2018ballistic}, and ballistic phase coherent superconductivity\cite{hendrickx2018ballistic} were demonstrated recently by using undoped Ge/SiGe. The added complexity in developing reliable gate-stacks compared to modulation-doped structures has limited the investigation of quantum transport properties in undoped Ge/SiGe so far\cite{laroche_magneto-transport_2016,su_effects_2017-1,hardy2018single}. In this paper we demonstrate state of the art, very shallow, undoped Ge/SiGe heterostructures and devise a process for fabricating dopant-less heterostructure field effect transistors (H-FETs) without compromising on material quality. A comprehensive investigation of key electrical properties such as carriers mobility, critical density for conduction, effective mass, and g-factor, establishes Ge/SiGe as a promising platform for future hybrid quantum technologies. \vspace{-0.25\baselineskip} \section*{Results} \noindent\textbf{Ge/SiGe heterostructures.} Figure \ref{fig:str} shows a schematic of the investigated undoped Ge/SiGe heterostructures along with the results of structural characterization to elucidate the crystallographic, morphological, and chemical properties of the stack. The Ge/SiGe heterostructure is grown on a 100 mm Si (001) substrate in a high-throughput reduced-pressure chemical vapour deposition reactor. The layer sequence comprises a Si$_{0.2}$Ge$_{0.8}$ virtual substrate obtained by reverse grading\cite{shah_reverse_2008,capellini2010strain}, a 16-nm-thick Ge quantum well, a 22-nm-thick Si$_{0.2}$Ge$_{0.8}$ barrier, and a thin sacrificial Si cap (see Methods). The purpose of the Si cap is to provide a well-known starting surface for subsequent high-$\kappa$ metal gate stack deposition\cite{vincent2011si} and to possibly achieve a superior dielectric interface than what SiGe could offer. Both C and O concentration within the quantum well are below the secondary ion mass spectroscopy detection limit of ($4\times 10^{16}\text{ cm}^{-3}$) and ($8\times 10^{16}\text{ cm}^{-3}$), respectively, pointing to a very low impurity background level (see Supplementary Informations). Figure \ref{fig:str}(c) highlights the crystalline quality of the Si$_{0.2}$Ge$_{0.8}$ virtual substrate. Defects and dislocations are confined to the lower layers, at the Si/Ge interface and in the graded Si$_{1-y}$Ge$_y$. As the Si (Ge) concentration in the SiGe alloy is increased (decreased), relaxation of the upper layers is promoted. By performing preferential etching (see Supplementary Informations) we estimate an upper bound for the threading dislocation density of $(3.0\pm0.5)\times10^7$ for the Si$_{0.2}$Ge$_{0.8}$. The Si and Ge concentrations profiles across the virtual substrate (Fig. \ref{fig:str}(c); blue and red curves respectively) confirm the achievement of linear reverse-graded SiGe with targeted alloy composition. In-plane and out-of plane lattice parameters are obtained from the X-Ray diffraction reciprocal space map (XRD-RSM) in Fig. \ref{fig:str}(g). The Ge and Si$_{0.2}$Ge$_{0.8}$ buffer layers are over-relaxed compared to the Si substrate with a residual tensile strain of $\epsilon_{\parallel}=0.2\%$ and 0.26\%, respectively. This is typical in SiGe virtual substrates due to the difference in thermal contraction of the materials after cooling from the high growth temperature\cite{shah_reverse_2008,capellini2012high}. The peak corresponding to the Ge quantum well is vertically aligned to the peak of the Si$_{0.2}$Ge$_{0.8}$ buffer layer, indicating a pseudomorphic growth of the quantum well and resulting in an in-plane compressive strain of $\epsilon_{\parallel}=-0.63\%$. Figure \ref{fig:str}(b) shows the individual and combined signals of Si, Ge, and O signals from the strained Ge quantum well embedded between Si$_{0.2}$Ge$_{0.8}$. The Ge QW appears as a uniform layer of constant thickness and with sharp interfaces to the adjacent Si$_{0.2}$Ge$_{0.8}$. The increasing O and Si signals at the top of the heterostructure indicate that the Si cap has readily oxidized upon exposure to air. The absence of extended defects in the high resolution TEM images in Fig. \ref{fig:str}(d)-(f) indicates high crystalline quality in the quantum well and adjacent Si$_{0.2}$Ge$_{0.8}$. The high degree of control achieved in the deposition process results in the Si and Ge composition profiles in Fig. \ref{fig:str}(d)-(f), with constant Ge composition within each layer of the Si$_{0.2}$Ge$_{0.8}$/Ge/Si$_{0.2}$Ge$_{0.8}$ structure. \begin{figure*}% \includegraphics[width=\linewidth]{Fig4}% \caption{Transport measurements at $T=1.7$ K as a function of magnetic field $B$ and carrier density $p_{2D}$. (a) Conductivity $\sigma_{xx}$ (blue line) and Hall density $p_{2D}$ (black line) as a function of gate voltage $V_g$ at 1.7 K. Dashed line is a linear fit of the gate-dependent density at low $V_g$. (b) Density dependent mobility $\mu$ (solid circles) and power law fit (red curve) (c) Density dependent $\sigma_{xx}$ (solid circles) and fit to percolation theory (red line). (d) Magnetoresistivity $\rho_{xx}$ (black line) and filling factor (solid circles) at saturation density as a function of inverse magnetic field $1/B$. Red line is the fit to the filling factor progression. (e) Density obtained by the analysis of the Shubnikov de Haas oscillations $p_{SdH}$ plotted against the corresponding Hall density $p_{2D}$. Open and solid circles are measured at $1.7$ K and $50$ mK, respectively. (f) Temperature dependence of the SdH oscillations $\Delta\rho$ in the range $T=$ 1.7-10 K, after background subtractions. (g) $\Delta\rho$ (solid circles) as a function of $T$, normalized at $\Delta\rho_0=\Delta\rho(T_0=1.7\text{K})$. The red line is the fit used to extract $m^{*}$. (h) Dingle plot at $T_0=1.7$ K (see Methods). The red line is the linear fit used to extract $\tau_q$.} \label{fig:cla} \end{figure*} \noindent\textbf{Heterostructure field effect transistors.} Hall-bar shaped heterostructure field effect transistors (H-FETs) were fabricated to investigate the magnetotransport properties of the undoped Ge/SiGe. An external electric field is applied to the insulated top-gate thereby populating the Ge quantum well and creating a two-dimensional hole gas (2DHG). Compared to undoped Si/SiGe H-FETs\cite{mi2015magnetotransport}, we adopted a low-thermal budget, implantation-free process to obtain direct contact between diffused Pt metallic ohmics and the induced 2DHG (see Methods). This approach is possible due to the very low platinum germanosilicide hole Schottky barrier height\cite{kittl2008silicides,alptekin2009platinum}. The optical image of the final device is displayed in Fig. \ref{fig:dev}(b), together with the schematics of the transistor cross section at the gate/ohmic contact overlap region. The upper panels in Fig. \ref{fig:dev}(a) and (c) show STEM/EDX analysis in the ohmic contact region and under the top-gate, respectively. In the ohmic region (Fig. \ref{fig:dev}(a), top panel) Pt diffuses inside the SiGe barrier and surpasses the quantum well. A PtSiGe alloy is formed, with a Ge concentration less than the value of 0.8 in the as-grown material, due to the Pt dilution within. The formation of a PtSiGe alloy is also supported by the presence of crystalline grains, as visible by TEM (Fig. \ref{fig:dev}(a), bottom panel). Since Pt diffusion is achieved at significantly lower temperature than the quantum well growth - 300 $^\circ$C vs 500 $^\circ$C, respectively - the crystalline quality of the heterostructure under the gate-stack is preserved. The critical Ge/SiGe interfaces after device processing (Fig. \ref{fig:dev}(c), top panel) are as sharp as in the as grown material (Fig. \ref{fig:str}(c)). Furthermore, the high-resolution TEM image in the bottom panel of Fig. \ref{fig:dev}(c) highlights the atomically flat semiconductor/oxide interface. Figure \ref{fig:dev}(d) shows the band-structure in the H-FET at a carrier density of $1\times 11^{10}\text{ cm}^{-2}$ by solving the Schroedinger-Poisson equation as a function of the applied electric field at low temperatures. States in the quantum well with heavy hole (HH) symmetry are favored compared to light holes (LH) states, with the HH and LH band-edges split in energy by 40 meV. The wave-function of the fundamental HH state (HH0) is well confined in the Ge quantum well, with an energy splitting between the HH0 and LH0 states of 47 meV, and between HH0 and the first excited HH state (HH1) of 15 meV. These obtained energy splittings in Ge/SiGe are more than one order of magnitude larger than the valley splitting in the conduction band of Si/SiGe or Si/SiO$_2$ systems\cite{zwanenburg2013silicon}, supporting the possibility of obtaining well defined qubits in this material platform. \noindent\textbf{Mobility, critical density, and effective mass.} Magnetotransport characterisation of the Ge/SiGe H-FETs was performed at low temperatures to elucidate the quantum transport properties of the 2DHG. The device is operated in accumulation mode, in which carriers populate the quantum well by applying a negative DC voltage bias ($V_g$) to the top gate. Upon applying a fixed AC voltage bias to source and drain contacts ($V_{sd}$), standard four-probe lock-in techniques allow to measure the longitudinal and transverse components of the resistivity tensor ($\rho_{xx}$, $\rho_{xy}$, respectively), from which longitudinal ($\sigma_{xx}$) and transverse ($\sigma_{xy}$) conductivity are extracted. The active carrier density $p_{2D}$ is measured by the Hall effect and, consequently, the carrier mobility $\mu$ (see Methods). \begin{figure*}% \includegraphics[width=120mm]{Fig5_raster}% \caption{Transport measurements at $T=50$ mK as a function of magnetic field $B$ and carrier density $p_{2D}$. (a) Quantized Hall conductance and (b) Shubnikov de Haas oscillations. (c) The local Fermi energy $E$ for different Landau levels. (d) Effective g-factor as a funcion of filling factor (left panel) and density (right panel).} \label{fig:Fig6} \end{figure*} Figure \ref{fig:cla}(a) shows the conductivity and the carrier density as a function of gate voltage (blue and black curves, respectively). At zero applied $V_g$ there are no carriers induced in the quantum well, the undoped heterostructure is insulating, and no conduction is measured. Above a threshold bias ($V_g$ = -3.8 V), current starts flowing in the channel and $\sigma_{xx}$ increases monotonically until saturation. After turn-on, $p_{2D}$ increases linearly as $V_g$ sweeps more negative. This is consistent with a simple parallel-plate capacitor model in which the Si$_{0.2}$Ge$_{0.8}$ and Al$_2$O$_3$ layers are the dielectric layers between the Ge quantum well and the metallic top-gate. The extracted effective gate capacitance obtained by fitting the linear portion of the $p_{2D}$ $vs.$ $V_g$ curve, however, is reduced to $\approx60\%$ of the expected value. At larger $V_g$ the $p_{2D}$ $vs.$ $V_g$ curve deviates from linearity and eventually $p_{2D}$ saturates at a value of $6\times 10^{11}\text{ cm}^{-2}$. Figure \ref{fig:cla}(b) shows the density-dependent mobility. The mobility increases strongly with density over the entire range of accessible densities, without signs of saturation. By fitting the data to a power law dependence $\mu = p_{2D}$$^\alpha$ we find a large exponent $\alpha = 2.1$. Including local field corrections\cite{dolgopolov2003remote,gold2010mobility}, exponents ($\alpha \geq$ 1.5) indicate that the mobility is still limited by scattering from remote impurities at the dielectric/semiconductor interface, as seen previously in Si/SiGe heterostructures\cite{lu2011upper,laroche_magneto-transport_2016}. At saturation density $p_{2D}=6\times 10^{11}\text{ cm}^{-2}$ we measure a maximum mobility of $5\times 10^{5}\text{ cm}^2$/Vs, corresponding to a mean free path of $\approx6 \mu m$, setting new benchmarks for holes in shallow FET devices. The critical density for establishing metallic conduction in the channel is extracted by fitting the density-dependent conductivity (Fig. \ref{fig:cla}(c)) in the low density regime to percolation theory\cite{tracy2009observation,kim_lyon_2017} $\sigma \propto A(p-p_p)^p$ cm$^{-2}$.. By fixing the exponent $p=1.31$, as expected in a 2D system, we estimate a percolation threshold density $p_p = 1.15\times10^{11}$ cm$^{-2}$ at 1.7 K, which sets an upper bound for $n_p$ at the temperature at which qubits typically operate ($T\leq$ 100 mK). Note that this value is in agreement with the qualitative estimate $p_c = 1.19\times10^{11}$ cm$^{-2}$ obtained by extrapolating to zero the linear region of mobility $\mu$ vs. log($p_{2D}$) curve above the critical density \cite{mccamey2005donor}. The obtained percolation threshold density is similar to the values reported in high quality Si MOSFET\cite{tracy2009observation,kim2017annealing}, and $\approx 2 \times$ higher than the values reported in undoped Si/SiGe\cite{mi2015magnetotransport}, possibly because the dielectric interface in our samples is much closer to channel (22 nm compared to 50 nm in Ref\cite{mi2015magnetotransport}). Figure \ref{fig:cla}(d) shows the Shubnikov de Haas (SdH) oscillations in the magnetoresistivity at saturation density. The oscillations minima reach zero value at high fields and are periodic against the inverse of magnetic field $1/B$. From the linear filling factor progression (Fig. \ref{fig:cla}(d), red circles) we extract a density $p_{SdH}=6.1\times 10^{11}\text{ cm}^{-2}$, in agreement with the Hall data. The agreement $p_{SdH}=p_{2D}$ is verified throughout the range of investigated density and temperature (Fig. \ref{fig:cla}(e)). Figure \ref{fig:cla}(f) shows the temperature dependent magnetoresistance oscillation amplitude $\Delta\rho$ at a density $p_{2D}=5.4\times 10^{11}\text{ cm}^{-2}$ as a function of the inverse of the magnetic field $1/B$, after the subtraction of a polynomial background. The hole effective mass $m^*$ is extracted by fitting the damping of the SdH oscillation amplitude with increasing temperature at $B = 1.4~$T (Fig. \ref{fig:cla}(g), see Methods). The obtained value $m^* = (0.090 \pm 0.002)m_e$, where $m_e$ is the free electron mass, is comparable to previous reports in Ge/SiGe.\cite{laroche_magneto-transport_2016,hardy2018single}. The quantum lifetime $\tau_q$ at 1.7 K is extracted by fitting the SdH oscillation amplitudes envelope\cite{coleridge1989low}. From the Dingle plot in Fig. \ref{fig:cla}(h), we estimate $\tau_q = 0.74~$ps, corresponding to a large Dingle ratio $\tau_t/\tau_q = 27$, where $\tau_t$ is the transport lifetime. The obtained large Dingle ratio confirms that the mobility is limited by scattering from charges trapped at the dielectric/heterostructure interface. \noindent\textbf{Landau fan diagram and effective g-factor.} In Fig. \ref{fig:Fig6}(a) and (b) we show color plots of both $\rho_{xx}$ as well as the transversal Hall signal $\sigma_{xy}$ at $T=50$ mK, as a function of out-of-plane magnetic field $B$ and carrier density $p_{2D}$, as obtained from the low-field Hall data. In this Landau fan diagram, both the quantum Hall effect (Fig. \ref{fig:Fig6}(a)) and SdH-oscillations (Fig. \ref{fig:Fig6}(b)) fan out linearly towards higher field and density. Observation of a Landau fan diagram of such quality is a sign of the very low disorder in the 2D channel. As in the measurements at 1.7 K (Fig. \ref{fig:cla}(f)), we do not observe beatings in the SdH oscillations, which prevent us to directly measure the strength of the spin-orbit interaction. However, we estimate an upper bound for spin-splitting in the 2DHG of 1.5 meV from the peak width of 1.5 T observed in the Fourier transform of $\rho_{xx}$ against 1/B. We estimate the out-of-plane g-factor by analyzing the difference in slope of the spin-up and spin-down level of the same Landau level\cite{huang2013direct}. We convert the carrier density to an energy (Fig. \ref{fig:Fig6}(c)) by assuming a free electron dispersion $E_F=\pi\hbar^2p_{2D}/m^*$, with $\hbar$ Planck's constant and $m^*=0.09m_e$ the effective mass in the system. The difference between the spin split levels is directly given by the Zeeman splitting energy $E_Z=g\mu_BB$. We plot this derived $g$-factor in the left panel of Fig. \ref{fig:Fig6}(d) and observe that it decays from $g=7$ to $g=4$ as a function of the filling factor. Alternatively we compare the field strength at which SdH oscillations appear, $B_{Tr}$, to the field at which the spin-splitting occurs $B_{SO}$, such that the $g$-factor can be estimated by $g=\frac{2m_e}{m^*}\frac{1}{1+B_{SO}/B_{Tr}}$\cite{mi2015magnetotransport}. The analysis is shown in Fig. \ref{fig:Fig6}(d) (right panel), with the range of the $g$-factor being in good agreement with the fan diagram analysis and other reports in Ge\cite{drichko2018effective}. The observed enhancement of the $g$-factor with decreasing densities has been previously observed and attributed to an increase in the strength of the effective Coulomb interaction\cite{sadofyev2002large,lu2017effective} or to the non-parabolicity of the valence band\cite{drichko2018effective}. \section*{Discussion} The observed $p_{2D}-V_g$ and $\mu-p_{2D}$ dependences are in line with previous studies on shallow undoped Si and Ge/SiGe heterostructures \cite{laroche2015scattering, su_effects_2017-1}. At small electric fields, carrier tunneling can occur from the shallow Ge quantum well to defect states in the band-gap of the dielectric/SiGe interface. Whilst tunneling reduces the gate capacitance, the passivated surface impurities by tunneled carriers lead also to an enhanced mobility\cite{huang2014screening}. At higher electric fields, the Fermi level aligns with the valence band edge at the dielectric/SiGe interface. Population of the surface quantum well prevents, by screening, further carrier accumulation in the buried channel, which reaches saturation. Nevertheless, only the buried quantum well contributes to transport, since the surface quantum well carrier concentration is likely below the mobility edge threshold, which is typically high for a channel at the Al$_2$O$_3$/Si interface. This interpretation is supported by the matching densities $p_{SdH}$ and $p_{SdH}$ (Fig. \ref{fig:cla}(d)), with no beating observed in the SdH oscillations: only one high mobility subband contributes to the measured transport. In conclusion, by measuring key transport metrics at low temperatures we have shown that shallow and undoped Ge/SiGe heterostructures are a promising low-disorder platform for Ge quantum devices. The reported half-million mobility sets new benchmarks for Si and Ge shallow-channel H-FETs\cite{huang2014screening,laroche2015scattering,su_effects_2017-1}, while even higher mobilities may be obtained by by further improving the semiconductor/dielectric interface. Possible avenues in these directions include the removal of the native silicon oxide layer prior to high-\textit{$\kappa$} dielectric deposition and/or post-metallization thermal anneals. A better quality semiconductor/dielectric interface should also improve the critical density, which is a critical metric for quantum devices. The obtained effective mass of 0.09m$_{e}$ is much lighter than the 0.19m$_{e}$ value for electron in silicon and close to the value of 0.067m$_{e}$ for electrons in GaAs. A light effective mass is beneficial for spin qubits since it corresponds to larger energy level spacing in quantum dots and relaxes lithographic fabrication requirements due to a larger extent of the wavefunction. Further improvements towards achieving the even lighter theoretical value of 0.05m$_{e}$\cite{terrazos2018light} will include lowering the growth temperature of the quantum well to reduce Ge diffusion from the quantum well in the barrier, thereby improving the abruptness of the quantum well/barrier interface and preventing possible spilling of the wave-function in the SiGe barrier. \section*{Methods} \small \textbf{Heterostructure growth.} The Ge/SiGe heterostructure is grown in a high-throughput reduced-pressure chemical vapour deposition (RP-CVD) reactor (ASMI Epsilon 2000) in one deposition cycle using germane (GeH$_4$) and dichlorosilane (SiH$_2$Cl$_2$) as precursor gases. Starting with a 100 mm Si (001) substrate, a 1.4-$\mu$m-thick layer of Ge is grown using a dual-step process. An initial low-temperature (400 $^\circ$C) growth of a Ge seed layer is followed by a higher temperature (625 $^\circ$C) overgrowth of a thick relaxed Ge buffer layer. Cycle anneals at 800 $^\circ$C are performed to promote full relaxation of the Ge. The subsequent 900-nm-thick reverse-graded Si$_{1-y}$Ge$_y$ layer\cite{shah_reverse_2008} is grown at 800$^\circ$ C with $y$ changing from 1 to 0.8. The SiGe virtual substrate is completed by a Si$_{0.2}$Ge$_{0.8}$ buffer layer of uniform composition, which is initially grown at 800 $^\circ$C. For the final 160 nm, the growth temperature is lowered to match the growth temperature of the subsequent layers (500 $^\circ$C). In this way growth interruption for temperature equilibration is avoided at the critical quantum well/virtual substrate interface. The heterostructure is completed with the deposition of a 16 nm-thick Ge quantum well, a 22 nm-thick Si$_{0.2}$Ge$_{0.8}$ barrier, and a thin ($<$2 nm) sacrificial Si cap. \newline \\ \small \textbf{Structural analysis.} X-ray diffraction measurements were performed with a 9 kW SmartLab diffractometer from Rigaku equipped with a Ge(400x2) crystal collimator and a Ge(220x2) crystal analyzer using CuK1 radiation. The asymmetric (-2-24) reflection was used for reciprocal space mapping to determine in-plane and out-of-plane lattice parameters. Transmission electron microscope investigation was carried out using a FEI Tecnai Osiris. For energy dispersive X-ray spectroscopy (EDX), the TEM was operated in scanning TEM (STEM) mode. The beam energy for all TEM measurements was 200 keV. \newline \\ \small \textbf{Device fabrication.} The process for undoped Ge/SiGe H-FETs comprises the deposition of metallic ohmics, a high-$\kappa$ dielectric, and a metallic top-gate. Ohmic pads are deposited on top of a mesa structure by e-beam evaporation of 60 nm of Pt. An HF dip is performed prior Pt deposition to etch the native oxide at the surface and ensure that the Pt film is in direct contact with the underlying Si$_{0.2}$Ge$_{0.8}$. The subsequent atomic layer deposition of 30 nm of Al$_2$O$_3$ at 300 $^\circ$C has the twofold purpose of electrically isolating the transistors top-gate from the channel as well as providing the thermal budget needed to drive the Pt ohmics in the Si$_{0.2}$Ge$_{0.8}$. Finally, the top-gate is realized by depositing 10/150 nm of Ti/Au. \newline \\ \small \textbf{Electrical measurements.} Magnetotransport data in the temperature range of 1.7 to 10K has been obtained in a $^4$He variable temperature insert refrigerator equipped with a $9$ T magnet. Magnetotransport data at 50 mK were obtained in a $^3$He dilution refrigerator equipped with a $9$ T magnet. A bias in the 0.1-1 mV range, frequency 7.7 Hz, is applied to the source and drain contacts. The source drain current $I_{SD}$, the longitudinal voltage $V_{xx}$, and the transverse Hall voltage $V_{xy}$ are measured; the longitudinal resistivity $\rho_{xx}$ and transverse Hall resistivity $\rho_{xx}$ are calculated as $\rho_{xx}=V_{xx}/I_{SD}\times W/L$ and $\rho_{xy}=V_{xy}/I_{SD}$, respectively (aspect ratio $L/W=5$). Longitudinal ($\sigma_{xx}$) and transverse ($\sigma_{xy}$) conductivity are calculated from $\rho_{xx}$ and $\rho_{xy}$ by tension inversion. The electrically active Hall carrier density $p_{2D}$ is obtained from the linear dependence of the Hall resistivity with perpendicular magnetic field ($\rho_{xy}=B/p_{2D}e$) at low magnetic field values ($B \leq 0.5$ Tesla). The carrier mobility $\mu$ is obtained from the relationship $1/\rho_{xx}=p_{2D}e\mu$. The effective mass is fitted from the damping of the SdH oscillations by using the expression\cite{de1993effective}: \begin{equation} \dfrac{\Delta\rho_{xx}}{\Delta\rho_{0}} = \dfrac{\Delta\rho / \rho_{0}~(T)}{\Delta\rho / \rho_{0}~(T_{0})} = \dfrac{A_{T}}{A_{T_{0}}} = \dfrac{T~sinh(\alpha T_{0})}{T_{0}~sinh(\alpha T)}~, \end{equation} where $\alpha = \dfrac{2\pi k_{B} m^*}{\hbar eB}$, $k_{B}$ is Boltzmann constant, $\hbar$ the Planck constant, $\rho_{0}$ is the zero field resistivity, and $T_{0} = 1.7K$ is the coldest temperature at which the oscillations were measured. \newline \section*{Additional information} {\small \textbf{Competing interests:} The authors declare no competing interests.}

c58e16f0d9bfda1c67eac93977a71bf272b4ca42

\section{Introduction} Let $\scalA$ be a Dedekind domain with fraction field $K$. Let ${P}$ be a set of maximal ideals of $\scalA$ such that for each ${\mathfrak p}\in{P}$, the residue field $k({\mathfrak p})=\scalA/{\mathfrak p}$ has finite cardinality $N({\mathfrak p})$. We will be most interested in the case where $K$ is a number field or local field, $\scalA$ is the full ring of integers $\maxO_K$ of $K$, and ${P}$ is the set $M_K$ of all maximal ideals of $\maxO_K$. Let $\ringb$ be a (commutative) $\scalA$-algebra. Then for each ${\mathfrak p}\in{P}$ the algebra $\ringb/{\mathfrak p} \ringb= \ringb \otimes_\scalA k({\mathfrak p})$ over $k({\mathfrak p})$ has a natural $k({\mathfrak p})$-algebra endomorphism $F_{\mathfrak p}\colon x\mapsto x^{N({\mathfrak p})}$, which is called the Frobenius endomorphism. By a Frobenius lift on $\ringb$ at ${\mathfrak p}$ we mean an $\scalA$-algebra endomorphism $\psi_{\mathfrak p}\:\ringb\to\ringb$ such that $\psi_{\mathfrak p}\otimes k({\mathfrak p})=F_{\mathfrak p}$. We define a \emph{$\Lambda_{\scalA,{P}}$-structure} on $\ringb$ to be a set map ${P}\to{\mathrm{End}}_{\scalA{\mbox{-}\mathrm{alg}}}({E})$, denoted ${\mathfrak p}\mapsto\psi_{\mathfrak p}$, such that \begin{enumerate} \item $\psi_{\mathfrak p}$ is a Frobenius lift at ${\mathfrak p}$ for each ${\mathfrak p}\in {P}$. \item $\psi_{\mathfrak p} \circ\psi_{\mathfrak q} =\psi_{\mathfrak q} \circ\psi_{\mathfrak p}$ for all ${\mathfrak p},{\mathfrak q}\in{P}$. \end{enumerate} By a \emph{$\Lambda_{\scalA,{P}}$-ring} we mean an $\scalA$-algebra with $\Lambda_{\scalA,{P}}$-structure. (In fact, this definition of $\Lambda_{\scalA,{P}}$-structure is well-behaved only when $\ringb$ is torsion free as an $\scalA$-module, but since all the $\Lambda_{\scalA,{P}}$-rings in this paper have that property, we will use the simple definition given above. For the general one, see \cite{Borger:BGWV-I}.) For example, if $\scalA$ is the ring of integers of a number field $K$ and $\ringb$ is the ring of integers of a subfield $L$ of the strict Hilbert class field of $K$, then $\ringb$ has a unique $\Lambda_{\scalA,{P}}$-structure: $\psi_{\mathfrak p}$ is the Artin symbol of ${\mathfrak p}$ in the field extension $K\subseteq L$. Observe that if ${\mathfrak p}$ satisfies $\ringb\otimes_A k({\mathfrak p})=0$, then the lifting condition (1) is vacuous. In particular, if $\ringb$ is an algebra over $K$, then any commuting collection of $K$-automorphisms of $\ringb$ indexed by the maximal ideals of ${P}$ is a $\Lambda_{\scalA,{P}}$-structure on $\ringb$. At a different extreme, if ${P}$ consists of one maximal ideal, for example if $\scalA$ is a local ring, then the commutation condition (2) is vacuous. When $\scalA={\mathbb Z}$ and ${P}=M_{\mathbb Q}$, Wilkerson and Joyal have shown that a $\Lambda_{\scalA,{P}}$-structure on a ring without ${\mathbb Z}$-torsion is the same as a $\lambda$-ring structure in the sense of algebraic K-theory \cite{Wilkerson}\cite{Joyal:Lambda}. For instance, for any abelian group $M$ we have a natural $\Lambda_{{\mathbb Z},{P}}$-structure on the group ring ${\mathbb Z}[M]$ given by $\psi_{(p)}(m)=m^p$ for $m\in M$ and prime $p$. In an earlier paper \cite{Borger-deSmit:Integral-models}, we showed that a $\Lambda_{{\mathbb Z},{P}}$-ring that is reduced and finite flat over ${\mathbb Z}$ is a sub-$\Lambda_{{\mathbb Z},{P}}$-ring of ${\mathbb Z}[C]^n$ for some finite cyclic group $C$ and integer $n\geq 0$. The proof uses the explicit description of ray class fields over ${\mathbb Q}$ as cyclotomic fields. Over a general number field, class field theory is less explicit, and the generalizations we present in the present paper are consequently less explicit. However, we can still give a very similar criterion for a finite \'etale $K$-algebra ${E}$ with $\Lambda_{\scalA,{P}}$-structure to admit an integral $\Lambda_{\scalA,{P}}$-model, by which we mean a sub-$\Lambda_{\scalA,{P}}$-ring $\ringb\subseteq {E}$ which is finite flat as an $\scalA$-module such that the induced map $K\otimes_{\scalA}\ringb \to {E}$ is a bijection. See theorem \ref{global-thm-intro} below. Let $\mathrm{Id}_{{P}}$ denote the set of non-zero ideals of $\scalA$ divisible only by the primes in ${P}$. It as a monoid under ideal multiplication, the free commutative monoid on the set ${P}$. Let ${K^{\mathrm{sep}}}$ be a separable closure of $K$, and let ${G_K}$ denote the Galois group of ${K^{\mathrm{sep}}}$ over $K$. It is a profinite group. By a ${G_K}$-set $X$ we mean a finite discrete set with a continuous ${G_K}$-action. By Grothendieck's formulation of Galois theory, a finite \'etale $K$-algebra ${E}$ is determined by the ${G_K}$-set $S$ consisting of all $K$-algebra homomorphisms ${E}\to {K^{\mathrm{sep}}}$. Giving a $\Lambda_{\scalA,{P}}$-structure on ${E}$ then translates to giving a monoid map $\mathrm{Id}_{{P}}\to{\mathrm{Map}}_{G_K}(S,S)$. By giving $\mathrm{Id}_{{P}}$ the discrete topology, we see that the category of $\Lambda_{\scalA,{P}}$-rings whose underlying $\scalA$-algebra is a finite \'etale $K$-algebra is anti-equivalent to the category of finite discrete sets with a continuous action of the monoid ${G_K}\times \mathrm{Id}_{{P}}$. Let us first consider the case where $\scalA$ is a complete discrete valuation ring and ${P}$ consists of the single maximal ideal ${\mathfrak p}$. Then $\mathrm{Id}_{{P}}$ is isomorphic as a monoid to the monoid of non-negative integers under addition. Let ${I_K}\subseteq{G_K}$ be the inertia subgroup. Then ${I_K}$ is normal in ${G_K}$ and ${G_K}/{I_K}$ is the absolute Galois group of $k({\mathfrak p})$, which contains the Frobenius element $F\in {G_K}/{I_K}$ given by $x\mapsto x^{N({\mathfrak p})}$. Thus, $F$ acts on any ${G_K}$-set on which ${I_K}$ acts trivially. \begin{thm} \label{local-thm} Suppose that $\scalA$ is a complete discrete valuation ring and that ${P}$ consists of the single maximal ideal ${\mathfrak p}$. Let ${E}$ be a finite \'etale $K$-algebra with a $\Lambda_{\scalA,{P}}$-structure, and let $S$ be the set of $K$-algebra maps from ${E}$ to ${K^{\mathrm{sep}}}$. Then $K$ has an integral $\Lambda_{\scalA,{P}}$-model if and only if the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ satisfies the following two conditions: \begin{enumerate} \item the group ${I_K}$ acts trivially on $S_{{\mathrm{unr}}}=\bigcap_{n\geq 0}{\mathfrak p}^n S$; \item the elements ${\mathfrak p}\in \mathrm{Id}_{{P}}$ and $F\in {G_K}/{I_K}$ act in the same way on $S_{{\mathrm{unr}}}$. \end{enumerate} \end{thm} \smallskip\noindent See section~\ref{sec:local} for the proof. Next, consider the global case, where $\scalA$ is the ring of integers in a number field. In order to express our result, let us first recall the definition of the ray class\ monoid. A cycle of $K$ is a formal product ${\mathfrak f}=\prod_{\mathfrak p}\gp^{n_{\mathfrak p}}$, where the product ranges over all primes of $K$, both finite and infinite, all $n_{\mathfrak p}$ are non-negative integers, only finitely many of which are non-zero, and we have $n_{\mathfrak p}\in\{0,1\}$ for real primes ${\mathfrak p}$, and $n_{\mathfrak p}=0$ for complex primes ${\mathfrak p}$. The finite part of ${\mathfrak f}$ is ${\mathfrak f}_{{\mathrm{fin}}}=\prod_{{\mathfrak p}<\infty}{\mathfrak p}^{n_{\mathfrak p}}$, which can be viewed as an ideal of $\scalA$. We write ${\mathrm{ord}}_{\mathfrak p}({\mathfrak f})=n_{\mathfrak p}$. For a cycle ${\mathfrak f}$, we say that two integral ideals ${\mathfrak a},{\mathfrak b}$ are \emph{${\mathfrak f}$-equivalent}, and write ${\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}$, if the following two conditions are satisfied: \begin{enumerate} \item ${\mathfrak a}$ and ${\mathfrak b}$ have the same greatest common divisor ${\mathfrak d}$ with ${\mathfrak f}_{{\mathrm{fin}}}$ \item ${\mathfrak a}{\mathfrak d}^{-1}$ and ${\mathfrak b}{\mathfrak d}^{-1}$ represent the same class in the ray class group ${\mathrm{Cl}}({\mathfrak f}{\mathfrak d}^{-1})$ of conductor ${\mathfrak f}{\mathfrak d}^{-1}$. \end{enumerate} This is an equivalence relation, and we write ${\mathrm{DR}}_{{P}}({\mathfrak f})$ for the quotient $\mathrm{Id}_{{P}}/{\sim_{\mathfrak f}}$. Because $\sim_{{\mathfrak f}}$ is preserved by multiplication of ideals, ${\mathrm{DR}}_{{P}}({\mathfrak f})$ inherits a unique monoid structure from $\mathrm{Id}_{P}$. We call it the \emph{ray class\ monoid (or Deligne--Ribet monoid)} of conductor ${\mathfrak f}$ supported at ${P}$. It was introduced in Deligne--Ribet \cite{Deligne-Ribet} in the case where ${P}=M_K$ and every real place divides ${\mathfrak f}$. For alternative definitions of ${\mathfrak f}$-equivalence and ${\mathrm{DR}}_{{P}}({\mathfrak f})$, see section~\ref{sec:DR}. Let us say that ${P}$ is \emph{Chebotarev dense} if any element of any ray class group ${\mathrm{Cl}}({\mathfrak f})$ can be represented by an ideal supported at ${P}$, or equivalently by infinitely many such ideals. For example, by Chebotarev's theorem, any set ${P}$ consisting of all but finitely many maximal ideals is Chebotarev dense. Whenever ${P}$ is Chebotarev dense, any element of ${\mathrm{Cl}}({\mathfrak f})$ can be written as the class of an ideal supported at ${P}$, and hence gives a well-defined element of ${\mathrm{DR}}_{P}({\mathfrak f})$. This defines a map ${\mathrm{Cl}}({\mathfrak f})\to {\mathrm{DR}}_{{P}}({\mathfrak f})$, which is in fact injective. Composing with the Artin symbol defines a map \begin{equation} \label{GalDRmap} {G_K}{\,\longrightarrow\,} {\mathrm{Cl}}({\mathfrak f}){\,\longrightarrow\,} {\mathrm{DR}}_{{P}}({\mathfrak f}), \end{equation} and hence a surjective map \begin{equation} \label{DR-recmap} {G_K}\times \mathrm{Id}_{{P}} {\,\longrightarrow\,} {\mathrm{DR}}_{{P}}({\mathfrak f}) \end{equation} whose restriction to the first factor is the map (\ref{GalDRmap}) and whose restriction to the second factor is the canonical quotient map $\mathrm{Id}_{{P}}\to{\mathrm{DR}}_{{P}}({\mathfrak f})$. \begin{thm} \label{global-thm-intro} Suppose that $K$ is a number field and that ${P}$ is Chebotarev dense. Let ${E}$ be a finite \'etale $K$-algebra with a $\Lambda_{\scalA,{P}}$-structure, and let $S$ be the set of $K$-algebra maps from ${E}$ to ${K^{\mathrm{sep}}}$. Then ${E}$ has an integral $\Lambda_{\scalA,{P}}$-model if and only if there is a cycle ${\mathfrak f}$ of $K$ such that the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ factors (necessarily uniquely) through the map ${G_K}\times \mathrm{Id}_{{P}} {\,\longrightarrow\,} {\mathrm{DR}}_{{P}}({\mathfrak f})$ above. \end{thm} It follows that the category of such $\Lambda_{\scalA,{P}}$-rings is anti-equivalent to the category of finite discrete sets with a continuous action by the profinite monoid $\lim_{\mathfrak f} {\mathrm{DR}}_{{P}}({\mathfrak f})$, where the inverse limit is taken over all cycles ${\mathfrak f}$ with respect to the canonical surjective maps ${\mathrm{DR}}({\mathfrak f})\to {\mathrm{DR}}({\mathfrak f}')$ when ${\mathfrak f}'\mid {\mathfrak f}$. When $K={\mathbb Q}$, $\scalA={\mathbb Z}$, and ${P}=M_{\mathbb Q}$, this limit is the monoid $\hat{{\mathbb Z}}^{\circ}$ of all profinite integers under multiplication. In this case, the theorem above reduces to the first theorem of our earlier paper \cite{Borger-deSmit:Integral-models}. It was Lenstra who suggested that the ray class\ monoid could play this role over general number fields. When ${E}$ admits an integral $\Lambda_{\scalA,{P}}$-model, there must be a maximal one. (See section~\ref{sec:max-model}.) In the example above, ${\mathbb Z}[x]/(x^n-1)$ is the maximal integral model of ${\mathbb Q}[x]/(x^n-1)$. This is the second theorem in~\cite{Borger-deSmit:Integral-models}. In general, let us write $\rcl{\scalA,{P}}{{\mathfrak f}}$ for the maximal integral $\Lambda_{\scalA,{P}}$-model associated to the free ${\mathrm{DR}}_{{P}}({\mathfrak f})$-set on one generator, namely ${\mathrm{DR}}_{{P}}({\mathfrak f})$. We call $\rcl{\scalA,{P}}{{\mathfrak f}}$ the \emph{ray class algebra} of conductor ${\mathfrak f}$---just as $K({\mathfrak f})$, the ray class field of conductor ${\mathfrak f}$, is the extension of $K$ corresponding to the free ${\mathrm{Cl}}({\mathfrak f})$-set on one generator. The ray class algebra is an order in a product of ray class fields: $$ K\otimes_{\scalA}\rcl{\scalA,{P}}{{\mathfrak f}} = \prod_{{\mathfrak d}\mid{\mathfrak f},{\mathfrak d}\in\mathrm{Id}_{P}}K({\mathfrak f}{\mathfrak d}^{-1}). $$ It is typically smaller than the maximal order in the non-$\Lambda$ sense. For example, $$ \rcl{{\mathbb Z},M_{\mathbb Q}}{n\infty}={\mathbb Z}[x]/(x^n-1)\subsetneq \prod_{d\mid n}{\mathbb Z}[\zeta_d]. $$ The theorems in~\cite{Borger-deSmit:Integral-models} for $K={\mathbb Q}$, however, give us something more than the abstract existence theorem above. They give explicit presentations of the ray class algebras $\rcl{{\mathbb Z},M_{\mathbb Q}}{n\infty}$, namely ${\mathbb Z}[x]/(x^n-1)$. More importantly, the presentations are all as quotients of a single finitely generated $\Lambda$-ring---in this case ${\mathbb Z}[x^{\pm 1}]$, or ${\mathbb Z}[x]$, where each $\psi_p$ is defined by $\psi_p(x)=x^p$. One can view this as a $\Lambda$-refinement of the Kronecker--Weber theorem, telling us that the function algebra $\sheafO(\mu_n)$ of the $n$-torsion subscheme $\mu_n\subset{\mathbb G}_{\mathrm{m}}=\Spec{\mathbb Z}[x^{\pm 1}]$ is isomorphic as a $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-ring to the ray class algebra $\rcl{{\mathbb Z},M_{\mathbb Q}}{n\infty}$; this is instead of the statement that its set of $\bar{{\mathbb Q}}$-points, $\mu_n(\bar{{\mathbb Q}})$, generates the ray class field ${\mathbb Q}(n\infty)$. This refinement gives us Frobenius lifts at all primes, even those dividing the conductor, by treating integral structures with more care. But the ray class algebras will have zero divisors and be non-normal over the primes dividing the conductor, which some might consider a drawback. Then again, they are normal in a $\Lambda$-ring theoretic sense, by definition. It is natural to ask whether something like this holds for number fields $K$ larger than ${\mathbb Q}$. Do the ray class algebras $\rcl{\scalA,{P}}{{\mathfrak f}}$ have a common origin in the algebraic geometry of $\Lambda_{\scalA,{P}}$-rings? If so, this would give a systematic way of generating ray class algebras and hence ray class fields. Or more modestly, is it at least true that the known explicit class field theories admit a $\Lambda$-refinement as above? There is also a converse question: when does a $\Lambda_{\scalA,{P}}$-structure on an $\scalA$-scheme $X$ give rise to a family of abelian extensions, as the $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-structure on ${\mathbb G}_{\mathrm{m}}$ does? The converse is the easier direction, and we will consider it first. Let $X$ be a (flat and separated) $A$-scheme with a $\Lambda_{\scalA,{P}}$-structure, by which we mean a commuting family of Frobenius lifts $\psi_{\mathfrak p}\:X\to X$, for ${\mathfrak p}\in{P}$. If we are going to follow the model of $\mu_n\subseteq{\mathbb G}_{\mathrm{m}}$ and produce abelian extensions of our number field $K$ by finding finite flat sub-$\Lambda_{\scalA,{P}}$-schemes $Z\subseteq X$ and applying the theorem above, then by this theorem, there must exist a cycle ${\mathfrak f}$ such that the Frobenius operators $\psi_{\mathfrak a}$ on $Z$ are ${\mathfrak f}$-periodic in ${\mathfrak a}$, meaning that they depend only on the class of ${\mathfrak a}$ in ${\mathrm{DR}}_{P}({\mathfrak f})$. So it is natural to consider the largest such subscheme, the locus $\per{X}{{\mathfrak f}}\subseteq X$ where the operations $\psi_{\mathfrak a}$ depend only the class of ${\mathfrak a}$ in ${\mathrm{DR}}_{P}({\mathfrak f})$. We call $\per{X}{{\mathfrak f}}$ the ${\mathfrak f}$-periodic locus. It is defined by a large equalizer diagram; so it does indeed exist and is a closed subscheme of $X$. For example, in the cyclotomic setting with the $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-scheme $X={\mathbb G}_{\mathrm{m}}$ as above and ${\mathfrak f}=n\infty$, with $n\geq 1$, the ${\mathfrak f}$-periodicity condition is $\psi_{(m+n)}=\psi_{(m)}$ for all $m\geq 1$. In other words, it is that the operators $\psi_{(m)}$ are periodic in $m$ with period dividing $n$. It follows that the ${\mathfrak f}$-periodic locus is just the $n$-torsion locus $\mu_n$. In general, while the ${\mathfrak f}$-periodic locus is similar in spirit to the ${\mathfrak f}$-torsion locus when $X$ is a group, they can be different: for example if $X$ is still ${\mathbb G}_{\mathrm{m}}$ but ${\mathfrak f}$ is trivial at $\infty$, so ${\mathfrak f}=(n)$, then the periodic locus must also be invariant under the involution $x\mapsto x^{\pm 1}$ and is hence just $\mu_2$ if $n$ is even, or $\mu_1$ if $n$ is odd. But the definition of $\per{X}{{\mathfrak f}}$ does not even require $X$ to be a group. Thus the group scheme structure in the traditional frameworks for explicit class field theory is replaced by a $\Lambda$-structure in ours. This will allow us some flexibility that is not available when working with group schemes. For example, we can divide out a CM elliptic curve, say, by its automorphism group. Although the group structure is lost, the $\Lambda$-structure is retained and hence we can still speak of the periodic locus on the quotient. Note that whereas in the group-scheme setting, the abelian nature of the Galois theory comes from the torsion locus being of rank $1$ over some commutative ring of complex multiplications and then from the commutativity of the general linear group $\mathrm{GL}_1$, in our setting, it comes from the assumption that the Frobenius lifts $\psi_{\mathfrak p}$ commute with each other and then from Chebotarev's theorem. We can now state the answer to the converse question: \begin{thm} \label{thm:intro-periodic} Let $X$ be a $\Lambda_{\scalA,{P}}$-scheme of finite type over $\scalA$, as above, and assume ${P}$ is Chebotarev dense. Then possibly after inverting some primes, $\per{X}{{\mathfrak f}}$ is an affine $\Lambda_{\scalA,{P}}$-scheme which is reduced and finite flat over $A$. \end{thm} The idea of the proof is that for primes ${\mathfrak p}\nmid{\mathfrak f}$, the endomorphism $\psi_{\mathfrak p}$ is an automorphism of finite order because the class $[{\mathfrak p}]\in{\mathrm{DR}}_{P}({\mathfrak f})$ is invertible and hence has finite order; therefore the Frobenius endomorphism of the fiber of $X$ over ${\mathfrak p}$ is an automorphism of finite order, and so the fiber is finite and geometrically reduced. Then apply the semicontinuity theorems of scheme theory. For details, see theorem~\ref{thm:periodic-locus3}. It then follows from theorem~\ref{global-thm-intro} that $\sheafO(\per{X}{{\mathfrak f}}_K)$, the function algebra of the generic fiber of $X({\mathfrak f})$, is a finite product of abelian extensions of $K$ of conductor dividing ${\mathfrak f}$. Thus any $\Lambda_{\scalA,{P}}$-scheme $X$ of finite type (still flat and separated) gives rise to a uniform geometric way of constructing abelian extensions indexed by cycles ${\mathfrak f}$. It will not always produce arbitrarily large abelian extensions---for example, the Chebyshev line below will only produce the ray class fields over ${\mathbb Q}$ with trivial conductor at infinity, namely ${\mathbb Q}(\zeta_n+\zeta_n^{-1})$. We can now state some precise versions of our original question of whether the ray class algebras have a common origin in $\Lambda$-algebraic geometry. It is cleaner to restrict to cycles ${\mathfrak f}$ of a fixed type away from ${P}$, and in particular of a fixed type at infinity. So fix a cycle ${\mathfrak r}$ supported away from ${P}$, and let $Z({P},{\mathfrak r})$ denote the set of cycles ${\mathfrak f}$ which agree with ${\mathfrak r}$ away from the primes in ${P}$. We will refer to the triple $(\scalA,{P},{\mathfrak r})$ as the \emph{context}. \begin{enumerate} \item[(Q1)] Does there exist a $\Lambda_{\scalA,{P}}$-scheme $X$ of finite type such that for all ${\mathfrak f}\inZ({P},{\mathfrak r})$, the direct factors of the \'etale $K$-algebra $\sheafO(\per{X}{{\mathfrak f}}_K)$ generate the ray class field $K({\mathfrak f})$? \end{enumerate} The Kronecker--Weber theorem states that the answer is positive in the cyclotomic context $({\mathbb Z},M_{\mathbb Q},\infty)$, with $X$ being ${\mathbb G}_{\mathrm{m}}$ with the usual $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-structure defined above. We will show it is true in two other classical contexts of explicit class field theory, namely the real-cyclotomic context $({\mathbb Z},M_{\mathbb Q},(1))$ and the context $(\maxO_K,M_K,(1))$ where $K$ is an imaginary quadratic field. In fact, we will show a stronger $\Lambda$-refinement holds. The strongest question one might ask is the following: \begin{enumerate} \item[(Q2)] Does there exist a $\Lambda_{\scalA,{P}}$-scheme $X$ of finite type such that for all ${\mathfrak f}\inZ({P},{\mathfrak r})$, the function algebra $\sheafO(\per{X}{{\mathfrak f}})$ of the periodic locus is isomorphic to the ray class algebra $\rcl{\scalA,{P}}{{\mathfrak f}}$? \end{enumerate} As mentioned above, the answer is positive in the cyclotomic context with $X={\mathbb G}_{\mathrm{m}}$. But it appears to be slightly too much to ask for in general. For instance, consider the real-cyclotomic context $({\mathbb Z},M_{\mathbb Q},(1))$, and let $X={\mathbb G}_{\mathrm{m}}/(x\mapsto x^{-1})\cong{\mathbb A}^1_{{\mathbb Z}}$. Then the $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-structure on ${\mathbb G}_{\mathrm{m}}$ descends to one on $X$. The operations $\psi_{(p)}$ are given by Chebyshev polynomials $\psi_{(p)}(y)\in{\mathbb Z}[y]$ determined by $\psi_{(p)}(x+x^{-1})=x^p+x^{-p}$. Then for any $n\geq 1$, the periodic locus $X(n)$ does indeed generate the real ray class fields, $$ {\mathbb Q}(X(n)(\bar{{\mathbb Q}}))={\mathbb Q}(\zeta_n+\zeta_n^{-1}), $$ but $\sheafO(X(n))$ does not generally agree with the full ray class algebra $\rcl{{\mathbb Z},{P}}{n}$. It is however only of index $2$ in it, once $n$ is even, and this error even disappears in the limit as $n$ grows. So in the real-cyclotomic context, the answer to (Q2) is as about as close to being positive without being so (at least for the given $X$!). This discrepancy is no doubt due to the nontrivial isotropy groups of the points $\pm 1\in{\mathbb G}_{\mathrm{m}}$ under the involution $x\mapsto x^{-1}$, and it may very well disappear in a proper stack-theoretic treatment. But if we stay in the world of schemes, we need to control it. So given a finite reduced $\Lambda_{\scalA,{P}}$-ring $B$, let $\tilde{B}$ denote the maximal integral $\Lambda_{\scalA,{P}}$-model in $K\otimes_A B$. It is the $\Lambda$-analogue of the integral closure of $B$. \begin{enumerate} \item[(Q3)] Does there exist a $\Lambda_{\scalA,{P}}$-scheme $X$ of finite type such that for all ${\mathfrak f}\inZ({P},{\mathfrak r})$, \begin{enumerate} \item[(i)] the function algebra $\sheafO(\per{X}{{\mathfrak f}})$ is of finite index in $\sheafO(\per{X}{{\mathfrak f}})^{\sim}$, \item[(ii)] the morphism of pro-rings $$ \Big(\sheafO(\per{X}{{\mathfrak f}})\Big)_{{\mathfrak f}\inZ({P},{\mathfrak r})} {\,\longrightarrow\,} \Big(\sheafO(\per{X}{{\mathfrak f}})^{\sim}\Big)_{{\mathfrak f}\inZ({P},{\mathfrak r})} $$ is an isomorphism, and \item[(iii)] $\sheafO(\per{X}{{\mathfrak f}})^{\sim}$ is isomorphic to $\rcl{\scalA}{{\mathfrak f}}$ as a $\Lambda_{\scalA,{P}}$-ring? \end{enumerate} \end{enumerate} \begin{thm}\label{thm:main} The answer to (Q3) is positive in the following contexts $(\scalA,{P},{\mathfrak r})$ with the $\Lambda_{\scalA,{P}}$-schemes $X$: \begin{enumerate} \item $({\mathbb Z},M_{\mathbb Q},\infty)$ and $X=\mathbb{G}_{\mathrm{m}/{\mathbb Z}}$ with the $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-structure above \item $({\mathbb Z},M_{\mathbb Q},(1))$ and $X={\mathbb A}^1_{\mathbb Z}$ with the Chebyshev $\Lambda_{{\mathbb Z},M_{\mathbb Q}}$-structure \item $(\maxO_K,M_K,(1))$, where $K$ is imaginary quadratic with Hilbert class field $H$, and $X$ is ${\mathbb P}^1_{\maxO_H}$, viewed as a scheme over $\maxO_K$, with the Latt\`es $\Lambda_{\maxO_K,M_K}$-structure defined in (\ref{subsec:lattes}). \end{enumerate} \end{thm} The $\Lambda$-scheme ${\mathbb P}^1_{\maxO_H}$ in the imaginary quadratic context plays the role of the target of the Weber functions $E\to E/{\mathrm{Aut}}(E)\cong{\mathbb P}^1_{\mathbb C}$ in the traditional treatments of explicit class field theory of imaginary quadratic fields. There are, however, a number of subtleties in constructing this $\Lambda$-structure. For instance, CM elliptic curves are only defined over $H$ and not $K$, there can be more than one of them, but it might be that none of them has good reduction everywhere. These problems were completely clarified in Gurney's thesis~\cite{Gurney:thesis}. He also gave an account of class field theory for imaginary quadratic fields from the point of view of the $\Lambda$-structure on ${\mathbb P}^1_{\maxO_H}$, but he stopped at considering the field extensions generated by the preimages $\psi_{\mathfrak f}^{-1}(\infty)$. In either approach, the package of elliptic curves with complex multiplication and their Weber functions is replaced by the single $\Lambda_{\maxO_K,M_K}$-scheme ${\mathbb P}^1_{\maxO_H}$. We emphasize that ${\mathbb P}^1_{\maxO_H}$ is not so interesting from a cohomological or motivic point of view---there are no interesting Galois representations in the cohomology. All the richness is in the $\Lambda$-structure. It would be interesting to know if something similar holds for other number fields, starting with CM fields. One virtue of the questions above is that they allow for negative answers, which would also be interesting. In the final section, we will raise some further questions in this direction. \vspace{2mm} It is a pleasure to thank Lance Gurney and Hendrik Lenstra for several helpful discussions. \part{Finite $\Lambda$-rings and class field theory} \section{Maximal $\Lambda$-orders} \label{sec:max-model} \subsection{} \emph{Maximal $\Lambda$-orders.} Let $\scalA'$ be a sub-$\scalA$-algebra of $K$, and let $E$ be a finite $K$-algebra with $\Lambda_{\scalA,{P}}$-structure. A sub-$\scalA'$-algebra $B\subseteq E$ with a $\Lambda_{\scalA,{P}}$-structure is said to be a \emph{$\Lambda_{\scalA,{P}}$-order (over $\scalA'$)} if it is finite over $\scalA'$. We say it is \emph{maximal (in $E$)} if it contains all others. Maximal $\Lambda_{\scalA,{P}}$-orders always exist, by an elementary argument (\cite{Borger-deSmit:Integral-models}, prop.\ 1.1), since maximal orders in the ordinary sense exist and since $\scalA'$ is noetherian. We will say a finite flat $\scalA'$-algebra $B$ with a $\Lambda_{\scalA,{P}}$-structure is \emph{normal} (or \emph{$\Lambda_{\scalA,{P}}$-normal over $\scalA'$}) if it is maximal in $K\otimes_{\scalA'} B$, in the sense above. We will need the following basic facts: \begin{proposition}\label{pro:maximality-facts} Let $B$ be a finite flat $\scalA'$-algebra with $\Lambda_{\scalA,{P}}$-structure. \begin{enumerate} \item $B$ is $\Lambda_{\scalA,{P}}$-normal over $\scalA'$ if $\maxO_{K_{\mathfrak p}}\otimes_{\scalA'}B$ is $\Lambda_{\maxO_{K_{\mathfrak p}},{\mathfrak p}}$-normal over $\maxO_{K_{\mathfrak p}}$ for all maximal ideals ${\mathfrak p}$ of $\scalA'$. \item If $B$ is $\Lambda_{\scalA,{P}}$-normal over $\scalA'$, and $G$ is a group acting on $B$ by $\Lambda_{\scalA,{P}}$-automorphisms, then the invariant subring $B^G$ is $\Lambda_{\scalA,{P}}$-normal over $\scalA'$. \item If $B$ is a product $B_1\times B_2$ of $\Lambda_{\scalA,{P}}$-rings, then $B$ is $\Lambda_{\scalA,{P}}$-normal over $\scalA'$ if and only if $B_1$ and $B_2$ are. \end{enumerate} \end{proposition} \begin{proof} (1): Let $C\subseteq K\otimes_{\scalA'}B$ be a $\Lambda_{\scalA,{P}}$-order over $\scalA'$ containing $B$. Then for any maximal ideal ${\mathfrak p}\subset \scalA'$, the base change $\maxO_{K_{\mathfrak p}}\otimes_{\scalA'}C$ is a $\Lambda_{\maxO_{K_{\mathfrak p}},{\mathfrak p}}$-order over $\maxO_{K_{\mathfrak p}}$. Therefore it agrees with $\maxO_{K_{\mathfrak p}}\otimes_{\scalA'}B$. Since this holds for all maximal ideals ${\mathfrak p}\subset\scalA'$, the ring $C$ agrees with $B$. Therefore $B$ is maximal. (2): Let $C\subseteq K\otimes_{\scalA'}B^G$ be a $\Lambda_{\scalA,{P}}$-order over $\scalA'$ containing $B^G$. Then $C$ is contained in $K\otimes_{\scalA'}B$ and hence, by the maximality of $B$, is contained in $B$. But since $C$ is contained in $K\otimes_{\scalA'}B^G$, it is also $G$-invariant. Therefore we have $C\subseteq B^G$. (3): Let $C\subseteq K\otimes_{\scalA'}(B_1\times B_2)$ be a $\Lambda_{\scalA,{P}}$-order over $\scalA'$ containing $B_1\times B_2$. Put $C_i=C\otimes_{B_1\times B_2} B_i$, for $i=1,2$. Then each $C_i$ is a $\Lambda_{\scalA,{P}}$-order over $\scalA'$ in $K\otimes_{\scalA'}B_i$. By the maximality of $B_i$, we have $C_i=B_i$. Therefore we have $$ C = C_1\times C_2 = B_1\times B_2. $$ Thus $B_1\times B_2$ is maximal. \end{proof} \section{The local case} \label{sec:local} In this section, $\scalA$ will be a complete discrete valuation ring with maximal ideal ${\mathfrak p}$, and ${P}$ will be the singleton set $\{{\mathfrak p}\}$. So $\mathrm{Id}_{{P}}$ is the multiplicative monoid of all nonzero ideals of $\scalA$. Write $k=k({\mathfrak p})$ and $\Lambda_{\scalA,{\mathfrak p}}=\Lambda_{\scalA,{P}}$. \begin{proposition} \label{unramified} Let $\ringb$ be a finite \'etale $\scalA$-algebra. Then $\ringb$ has a unique $\Lambda_{\scalA,{\mathfrak p}}$-structure, and the induced action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S={\mathrm{Hom}}_{K{\mbox{-}\mathrm{alg}}}(\ringb\otimes_\scalA K, {K^{\mathrm{sep}}})$ has the property that the inertia group ${I_K}$ acts trivially and that the element ${\mathfrak p}\in\mathrm{Id}_{{P}}$ and the Frobenius element $F\in{G_K}/{I_K}$ act on $S$ in the same way. \end{proposition} \begin{proof} Because $\ringb$ is \'etale, $k\otimes_\scalA \ringb$ is a product of finite fields. Since $\ringb$ is complete in its ${\mathfrak p}$-adic topology, idempotents of $\ringb/{\mathfrak p} \ringb$ lift to $\ringb$, so that $\ringb$ is a finite product of rings of integers in finite unramified extensions of $K$. Therefore the inertia group ${I_K}\subseteq {G_K}$ acts trivially on $S={\mathrm{Hom}}_{\scalA{\mbox{-}\mathrm{alg}}}(\ringb,{K^{\mathrm{sep}}})$. Every finite unramified field extension ${L}$ of $K$ is Galois with an abelian Galois group, and its rings of integers has a unique Frobenius lift. It follows that when $\ringb$ is unramified over $\scalA$, it has a unique $\Lambda_{\scalA,{\mathfrak p}}$-structure. \end{proof} \subsection{} \label{localstructure} \emph{Structure of finite \'etale $K$-algebras with $\Lambda_{\scalA,{\mathfrak p}}$-structure.} Let ${E}$ be a finite \'etale $K$-algebra with a $\Lambda_{\scalA,{\mathfrak p}}$-structure, and write $$ S={\mathrm{Hom}}_{K{\mbox{-}\mathrm{alg}}}({E},{K^{\mathrm{sep}}}). $$ Put $S_0=\bigcap_{n\geq 0}{\mathfrak p}^n S$ and for $i=1,2, \ldots$, let $$ S_i = \{s\in S\colon\; s\not\in S_{i-1} \text{ and } {\mathfrak p} s\in S_{i-1}\}. $$ Then each $S_i$ is a sub-${G_K}$-set of $S$. Since $S$ is finite, there exists an $n\geq 0$ such that $S_{n+1}=\varnothing$. Then we have the decomposition $$ S= S_0{\,\scriptstyle\amalg\,} S_1 {\,\scriptstyle\amalg\,} \cdots {\,\scriptstyle\amalg\,} S_n. $$ Let ${E}_i={\mathrm{Map}}_{{G_K}}(S_i, {K^{\mathrm{sep}}})$ be the corresponding finite \'etale $K$-algebra for each $i$. Then the maps $S_n\to \cdots \to S_1\to S_0\righttoleftarrow$ given by multiplication by ${\mathfrak p}$ give rise to a diagram of $K$-algebras \[ f_0 \lefttorightarrow {E}_0 {\buildrel f_1 \over\longrightarrow} {E}_1 {\buildrel f_2 \over\longrightarrow} \cdots {\buildrel f_{n} \over\longrightarrow} {E}_n \] Since $S=S_0{\,\scriptstyle\amalg\,} S_1 {\,\scriptstyle\amalg\,} \cdots{\,\scriptstyle\amalg\,} S_n$ is a decomposition of $G$ as a ${G_K}$-set, we have a corresponding product decomposition of the finite \'etale $K$-algebras ${E}={E}_0\times {E}_1 \times \cdots\times {E}_n$. In terms of this decomposition, $\psi_p$ is given by \begin{equation} \label{psicomp} \psi_{\mathfrak p}(e_0,e_1, \ldots, e_n)=(f_0(e_0),f_1(e_0),\ldots, f_{n}(e_{n-1})). \end{equation} Since $S_0$ is closed under multiplication by ${\mathfrak p}$, the quotient ring ${E}_0$ of ${E}$ is a quotient $\Lambda_{\scalA,{\mathfrak p}}$-ring of ${E}$, with Frobenius lift $f_0$. We will now construct a splitting of this quotient map ${E}\to {E}_0$ in the category of $\Lambda_{\scalA,{\mathfrak p}}$-rings. Note that we have ${\mathfrak p}^kS=S_0$ for sufficiently large $k$; so we have ${\mathfrak p} S_0=S_0$ and hence ${\mathfrak p}$ act as a bijection on $S_0$. Thus, $f_0$ is an automorphism of ${E}_0$. For $s\in S_i$ we have ${\mathfrak p}^is\in S_0$. Again since ${\mathfrak p}$ acts bijectively on $S_0$, we can define a map $S{\,\longrightarrow\,} S_0$ by sending $s\in S_i$ to the unique element $s'$ of $S_0$ such that ${\mathfrak p}^i s' = {\mathfrak p}^i s$. This map commutes with the ${G_K}\times \mathrm{Id}_{{P}}$-action, and it is a retraction of the inclusion $S_0\to S$. This induces our desired splitting ${E}_0\to E$. In other words, ${E}_0$ is not only a quotient $\Lambda_{\scalA,{\mathfrak p}}$-ring of ${E}$, but also a sub-$\Lambda_{\scalA,{\mathfrak p}}$-ring: \begin{align*} j\colon\; {E}_0 {\,\longrightarrow\,} & {E} \\ e_0 \longmapsto & (e_0,f_1f_0^{-1}e_0, f_2f_1f_0^{-2} e_0, \ldots, f_{n-1} \cdots f_1 f_0^{-n+1} e_0). \end{align*} \subsection{} \emph{Proof of theorem \ref{local-thm}.} Suppose that ${E}$ has an integral $\Lambda_{\scalA,{\mathfrak p}}$-model $\ringb$, that is, \begin{enumerate} \item[(i)] $\ringb$ is finitely generated as an $\scalA$-module and has rank $\dim_{K}({E})$, \item[(ii)] $\psi_{\mathfrak p}(\ringb)\subseteq \ringb$, \item[(iii)] $\psi_{\mathfrak p}\otimes_\scalA k$ is the Frobenius map $x\mapsto x^{N({\mathfrak p})}$ on $\ringb\otimes_A k$. \end{enumerate} Let $\ringb_0$ denote the image of $\ringb$ under the quotient map ${E}\to{E}_0$ (in the notation of \ref{localstructure}). Then $\ringb_0$ is an integral $\Lambda_{\scalA,{\mathfrak p}}$-model of ${E}_0$. Since $f_0$ is an automorphism of ${E}_0$, the ring $\ringb_0$ and its subring $f_0(\ringb_0)$ have the same discriminant. Thus, $f_0(\ringb_0)=\ringb_0$ and hence $f_0$ is an automorphism of $\ringb_0$. This implies that the map $x\mapsto x^{N({\mathfrak p})}$ on $\ringb_0\otimes_\scalA k$ is an automorphism, and so $\ringb_0$ is \'etale over $\scalA$. Conditions (1) and (2) of theorem \ref{local-thm} now follow by proposition \ref{unramified}. For the converse, suppose that conditions (1) and (2) hold. We will produce an integral $\Lambda_{\scalA,{\mathfrak p}}$-model of ${E}={E}_0\times \cdots\times {E}_n$. Let $R_i$ be the integral closure of $\scalA$ in ${E}_i$. Since ${I_K}$ acts trivially on $S_{\mathrm{unr}}=S_0$, the $A$-algebra $R_0$ is finite \'etale and hence has a unique $\Lambda_{\scalA,{\mathfrak p}}$-structure by proposition \ref{unramified}. Our integral model $\ringb\subseteq E$ will be of the form \begin{equation} \label{eq:lambda-split} \ringb = j(R_0) \oplus (0\times {\mathfrak a}_1\times \cdots\times{\mathfrak a}_n), \end{equation} where each ${\mathfrak a}_i$ is an ideal in $R_i$ and $j$ is the map defined in~\ref{localstructure}. Observe that any $\ringb$ of this form is a subring of $E$. For it to be a sub-$\Lambda_{\scalA,{\mathfrak p}}$-ring, we need to have $\psi_{\mathfrak p}(a)\equiv a^{N({\mathfrak p})}\bmod {\mathfrak p} \ringb$ for all $a\in \ringb$. Since both sides of this congruence are additive in $a$, it is enough to consider elements $a$ in each of the summands in~(\ref{eq:lambda-split}). It holds for the summand $j(R_0)$ because $j$ is a $\Lambda_{\scalA,{\mathfrak p}}$-morphism. So by (\ref{psicomp}), a sufficient condition for $\ringb$ to be a sub-$\Lambda_{\scalA,{\mathfrak p}}$-ring is $f_i({\mathfrak a}_{i-1})\subseteq {\mathfrak p}{\mathfrak a}_{i}$ and ${\mathfrak a}_i^{N({\mathfrak p})}\subseteq{\mathfrak p}{\mathfrak a}_i$ for $i=1,\dots,n$, where we take ${\mathfrak a}_0=0$. This holds if, for instance, ${\mathfrak a}_i={\mathfrak p}^{n-i+1}R_i$. So for this choice, $\ringb$ is an integral $\Lambda_{\scalA,{\mathfrak p}}$-model of ${E}$. \qed \subsection{} \emph{Remark.} \label{rmk-group-alg} Note that the integral model produced in the proof above is not always the maximal one. For instance, if $C_n$ denotes a cyclic group of order $n$, then on the group algebra ${\mathbb Q}_2[C_4]$ with its usual $\Lambda_{{\mathbb Z}_2,2}$-structure, the proof produces a strict subring of ${\mathbb Z}_2[C_4]$ and hence cannot be not maximal. (In fact, ${\mathbb Z}_2[C_4]$ is the maximal integral model, as is shown in~\cite{Borger-deSmit:Integral-models}). It can also happen that, for group algebras, the integral model supplied by the proof is strictly larger than the integral group algebra. An example is ${\mathbb Q}_2[C_2\times C_2]$. \subsection{} \emph{Remark.} It is possible to express theorem \ref{local-thm} in a more Galois-theoretic way, similar to the statement of theorem \ref{global-thm-intro}. We can define an inverse system of finite quotients $G_{N,n}$ of the monoid ${G_K}\times\mathrm{Id}_{{P}}$ with the property that ${E}$ has an integral $\Lambda_{\scalA,{\mathfrak p}}$-model if and only if the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ factors through some $G_{N,n}$. The quotients $G_{N,n}$ are defined as follows. Let $N$ be an open normal subgroup of ${G_K}$, and let $n$ be an integer $\geq 0$. Define a relation on ${G_K}\times \mathrm{Id}_{{P}}$ by $(g,{\mathfrak p}^a)\sim (h,{\mathfrak p}^b)$ if either or both of the following conditions hold: \begin{enumerate} \item $a=b$ and $g\equiv h \bmod N$ \item $a,b\geq n$ and $g F^a\equiv h F^{b} \bmod N{I_K}$. \end{enumerate} This is easily seen to be an equivalence relation which is stable under the monoid operation. We then define $G_{N,n}$ to be the quotient monoid. Observe that we have a decomposition of ${G_K}$-sets: \begin{equation} \label{eq:G_N,n} G_{N,n} = \underbrace{{G_K}/N {\;\scriptstyle\amalg\;} \cdots {\;\scriptstyle\amalg\;} {G_K}/N}_{n\text{ times}} {\;\scriptstyle\amalg\;} {G_K}/N{I_K}. \end{equation} For $N'\subseteq N$ and $n'\geq n$, we have evident transition maps $G_{N',n'} \to G_{N,n}$. If we consider the inverse limit $$ \hat{G}=\lim_{N,n} G_{N,n}, $$ then the re-expression of theorem \ref{local-thm} is that ${E}$ has an integral $\Lambda_{\scalA,{\mathfrak p}}$-model if and only if the action of ${G_K}\times\mathrm{Id}_{{P}}$ on $S$ factors (necessarily uniquely) through a continuous action of $\hat{G}$. One might call $\hat{G}$ the $\Lambda_{\scalA,{\mathfrak p}}$-algebraic fundamental monoid of $\Spec \maxO_K$ with ramification allowed along $\Spec k$. \subsection{} \emph{Remark.} In the global case considered in the rest of this paper, we will see that only abelian field extensions arise from integral $\Lambda$-models. But in the local case, we can get nonabelian extensions. In fact, we can get arbitrary extensions. Indeed, any given extension $L$ of $K$ is a direct factor of the $K$-algebra $L_{\mathrm{unr}}\times L$, which has a $\Lambda_{\scalA,{\mathfrak p}}$-structure admitting an integral model. For instance, one can take $\psi_{\mathfrak p}(e_0,e_1)=(F(e_0),e_0)$. \section{The ray class\ monoid} \label{sec:DR} Fix the following notation: \begin{align*} K &= \text{a finite extension of ${\mathbb Q}$} \\ \maxO_K &= \text{the ring of integers in $K$} \\ M_K &= \text{the set of maximal ideals of $\maxO_K$} \\ \scalA &= \text{a Dedekind domain whose fraction field is $K$} \\ {P} &= \text{a set of maximal ideals of $\scalA$} \\ \mathrm{Id}_{{P}} &= \text{the monoid of nonzero ideals of $\scalA$ supported at ${P}$} \\ {\mathfrak f} &= \text{a cycle (or modulus) on $K$} \\ {\mathfrak f}_{{P}} &= \text{the part of ${\mathfrak f}$ supported at ${P}$} \\ {\mathfrak f}^{{P}} &= \text{the part of ${\mathfrak f}$ supported away from ${P}$} \\ {\mathfrak f}_{{\mathrm{fin}}} &= \text{the part of ${\mathfrak f}$ supported at the finite primes} \\ {\mathfrak f}_{\infty} &= \text{the part of ${\mathfrak f}$ supported at the real places} \\ \mathrm{Id}_{{P}}({\mathfrak f}) &= \text{the submonoid of $\mathrm{Id}_{{P}}$ of ideals prime to ${\mathfrak f}_{{\mathrm{fin}}}$} \\ {\mathrm{Cl}}({\mathfrak f}) &= \text{the ray class group of $K$ of conductor ${\mathfrak f}$} \\ {\mathrm{Cl}}_{{P}}({\mathfrak f}) &= \text{the image of the canonical map $\mathrm{Id}_{{P}}({\mathfrak f})\to {\mathrm{Cl}}({\mathfrak f})$} \\ R^\circ &= R \text{ viewed as a monoid under multiplication, for any ring }R \end{align*} Observe that, up to canonical isomorphism, the constructions above depend only on the places of $K$ corresponding to ${P}$, and so they depend on $\scalA$ only in that these places must come from maximal ideals of $\scalA$. Therefore we can take $\scalA=\maxO_K$ without changing anything above. \subsection{} \emph{Structure of the ray class\ monoid.} There is a bijection \begin{equation} \coprod_{{\mathfrak d}\in\mathrm{Id}_{{P}},{\mathfrak d}\mid{\mathfrak f}} {\mathrm{Cl}}_{{P}}({\mathfrak f}{\mathfrak d}^{-1}) \,{\buildrel \sim\over\longrightarrow}\, {\mathrm{DR}}_{{P}}({\mathfrak f}), \end{equation} sending an ideal class $[{\mathfrak a}]\in {\mathrm{Cl}}_{{P}}({\mathfrak f}{\mathfrak d}^{-1})$ in the summand of index ${\mathfrak d}$ to the class $[{\mathfrak d}{\mathfrak a}]\in{\mathrm{DR}}_{{P}}({\mathfrak f})$. Thus we have \begin{equation} \label{DR-decomp} \coprod_{{\mathfrak d}\in\mathrm{Id}_{{P}},{\mathfrak d}\mid{\mathfrak f}} [{\mathfrak d}]{\mathrm{Cl}}_{{P}}({\mathfrak f}{\mathfrak d}^{-1}) = {\mathrm{DR}}_{{P}}({\mathfrak f}). \end{equation} The multiplication law on ${\mathrm{DR}}_{{P}}({\mathfrak f})$ is given in terms of the left-hand side by the formula \begin{equation} \label{eq:DR-mult} [{\mathfrak d}][{\mathfrak a}]\cdot [{\mathfrak d}'][{\mathfrak a}'] = [{\mathfrak d}''][{\mathfrak a}''], \end{equation} where ${\mathfrak d}''=\gcd({\mathfrak d}\gd',{\mathfrak f})$ and ${\mathfrak a}''$ satisfies ${\mathfrak d}''{\mathfrak a}''={\mathfrak d}{\mathfrak a}{\mathfrak d}'{\mathfrak a}'$. It follows, for example, that the submonoid ${\mathrm{DR}}_{{P}}({\mathfrak f})^*\subseteq {\mathrm{DR}}_{{P}}({\mathfrak f})$ of invertible elements agrees with the part of the ray class group supported at ${P}$: $$ {\mathrm{DR}}_{{P}}({\mathfrak f})^*={\mathrm{Cl}}_{{P}}({\mathfrak f}). $$ When ${P}$ is Chebotarev dense, we have ${\mathrm{Cl}}_{{P}}({\mathfrak f})={\mathrm{Cl}}({\mathfrak f})$, and so the invertible part of ${\mathrm{DR}}_{{P}}({\mathfrak f})$ is independent of ${P}$. Observe however that this is not the case for the noninvertible part. For example, if $K={\mathbb Q}$ and ${\mathfrak f}=(n)\infty$, then we have \begin{equation} \label{eq:DR-over-Q} {\mathrm{DR}}_{{P}}({\mathfrak f}) = ({\mathbb Z}/n_{{P}}{\mathbb Z})^{\circ} \times ({\mathbb Z}/n^{{P}}{\mathbb Z})^*, \end{equation} where $n_{{P}}$ is the factor of $n$ supported at ${P}$ and $n^{P}$ is that supported away from ${P}$. The invertible part is $({\mathbb Z}/n_{{P}}{\mathbb Z})^* \times ({\mathbb Z}/n^{{P}}{\mathbb Z})^* = ({\mathbb Z}/n{\mathbb Z})^*$, which does not depend on ${P}$, but the non-invertible part does. \subsection{} \emph{Change of conductor.} \label{subsec:DR-change-of-f} Suppose ${\mathfrak f}\mid{\mathfrak f}'$. Then we have the implication $$ {\mathfrak b}\sim_{{\mathfrak f}'}{\mathfrak c} \Rightarrow {\mathfrak b}\sim_{{\mathfrak f}}{\mathfrak c}. $$ which induces a monoid map \begin{equation} {\mathrm{DR}}_{P}({\mathfrak f}')\to {\mathrm{DR}}_{{P}}({\mathfrak f}), \end{equation} which we call the canonical map. There is also a map in the other direction. Write ${\mathfrak f}'={\mathfrak f}{\mathfrak a}$. Then by the equivalence $$ {\mathfrak b}\sim_{\mathfrak f}{\mathfrak c} \Leftrightarrow {\mathfrak a}{\mathfrak b}\sim_{{\mathfrak a}{\mathfrak f}}{\mathfrak a}{\mathfrak c}, $$ there is a well-defined injective map \begin{equation} {\mathrm{DR}}_{P}({\mathfrak f}) {\,\longrightarrow\,} {\mathrm{DR}}_{P}({\mathfrak f}'), \quad [{\mathfrak b}]\mapsto [{\mathfrak a}{\mathfrak b}]. \end{equation} It is not a monoid map, but it is an equivariant map of ${\mathrm{DR}}_{{P}}({\mathfrak f}')$-sets. The composition with the canonical projection ${\mathrm{DR}}_{P}({\mathfrak f}')\to {\mathrm{DR}}_{{P}}({\mathfrak f})$ on either the left or the right, is given by multiplication by the class of ${\mathfrak a}$. \vspace{5mm} We conclude this section with two more descriptions of ${\mathfrak f}$-equivalence and the ray class\ monoid, although they only have small parts in this paper. The first is the one that appears in Deligne--Ribet~\cite{Deligne-Ribet}, and a form of the second was pointed out to us by Bora Yalkinoglu. \begin{proposition} \label{pro:DR-original-def} Two ideals ${\mathfrak a},{\mathfrak b}$ of $\maxO_K$ are ${\mathfrak f}$-equivalent if and only if ${\mathfrak a}=x{\mathfrak b}$ for some element $x\in 1+{\mathfrak f}_{{\mathrm{fin}}}{\mathfrak b}^{-1}$ which is positive at all real places dividing ${\mathfrak f}$. \end{proposition} \begin{proof} Let ${\mathfrak d}=\gcd({\mathfrak f},{\mathfrak b})$. By definition, we have ${\mathfrak a}\sim_{\mathfrak f}{\mathfrak b}$ if and only if there exists an element $x\in K^*$ satisfying the following: \begin{enumerate} \item ${\mathfrak a}{\mathfrak d}^{-1}=x{\mathfrak b}{\mathfrak d}^{-1}$ \item $x\equiv 1\bmod {\mathfrak p}^{n_{\mathfrak p}}$, where $n_{\mathfrak p}={\mathrm{ord}}_{\mathfrak p}({\mathfrak f}{\mathfrak d}^{-1})$, whenever $n_{\mathfrak p}\geq 1$ \item $x$ is positive at all real places dividing ${\mathfrak f}{\mathfrak d}^{-1}$. \end{enumerate} Observe that (1) is equivalent to ${\mathfrak a}=x{\mathfrak b}$ and that (3) is equivalent to the positivity condition in the statement of the proposition. It is therefore enough to show that, under (1), condition (2) is equivalent to $x\in 1+{\mathfrak f}_{{\mathrm{fin}}}{\mathfrak b}^{-1}$, or equivalently to the condition that that for all ${\mathfrak p}$, we have $x\equiv 1\bmod {\mathfrak p}^{m_{\mathfrak p}}$, where $m_{\mathfrak p}={\mathrm{ord}}_{\mathfrak p}({\mathfrak f}{\mathfrak b}^{-1})$. So fix a prime ${\mathfrak p}$. In the case ${\mathrm{ord}}_{\mathfrak p}({\mathfrak f})\geq {\mathrm{ord}}_{\mathfrak p}({\mathfrak b})$, we have $n_{\mathfrak p}=m_{\mathfrak p}$ and so this condition is indeed equivalent to (2). Now consider the remaining case ${\mathrm{ord}}_{\mathfrak p}({\mathfrak f})<{\mathrm{ord}}_{\mathfrak p}({\mathfrak b})$. Then we have $n_{\mathfrak p}=0$ and $m_{\mathfrak p}<0$. Because $n_{\mathfrak p}=0$, condition (2) is vacuous. Therefore it is enough to show that $x\equiv 1\bmod {\mathfrak p}^{m_{\mathfrak p}}$ necessarily holds. Since $\gcd({\mathfrak f},{\mathfrak b})=\gcd({\mathfrak f},{\mathfrak a})$, we have ${\mathrm{ord}}_{\mathfrak p}({\mathfrak f})<{\mathrm{ord}}_{\mathfrak p}({\mathfrak a})$ and hence $${\mathrm{ord}}_{\mathfrak p}(x-1)\geq \min\{{\mathrm{ord}}_{\mathfrak p}(x),0\}= \min\{{\mathrm{ord}}_{\mathfrak p}({\mathfrak a}{\mathfrak b}^{-1}),0\} \geq \min\{m_{\mathfrak p},0\} = m_{\mathfrak p}.$$ \end{proof} \begin{proposition} \label{pro:Bora-DR-definition} Assume that ${P}$ is Chebotarev dense. Then there is an isomorphism \begin{equation} \label{map:bora} (\maxO_K/{\mathfrak f}_{{P}})^{\circ}\oplus_{(\maxO_K/{\mathfrak f}_{{P}})^*} {\mathrm{Cl}}({\mathfrak f})\,{\buildrel \sim\over\longrightarrow}\,{\mathrm{DR}}_{{P}}({\mathfrak f}), \end{equation} which is given by the canonical inclusion ${\mathrm{Cl}}({\mathfrak f})\to{\mathrm{DR}}_{P}({\mathfrak f})$ on the second factor and which on the first factor sends the residue class of any element $x\in\maxO_K$ with $(x)\in\mathrm{Id}_{P}$ to the class $[(x)]\in{\mathrm{DR}}_{P}({\mathfrak f})$. \end{proposition} The notation $A\oplus_G B$ above refers to the push out in the category of commutative monoids, which when $G$ is a group is the quotient of $A\oplus B=A\times B$ by the action of $G$ given by $g\cdot(a,b)=(ga,g^{-1}b)$. \begin{proof} First observe that this morphism is well defined. Indeed, any element $x\in (\maxO_K/{\mathfrak f}_{P})^*$ is the image of an element $\tilde{x}\in\maxO_K$ relatively prime to ${\mathfrak f}_{P}$, and its class $[(\tilde{x})]$ in ${\mathrm{DR}}_{P}({\mathfrak f})$ is indeed the image of its class in the ray class group ${\mathrm{Cl}}({\mathfrak f})$. The fact that this morphism is an isomorphism is a consequence of the following equalities, which will be justified below, and where ${\mathfrak d}$ runs over $\mathrm{Id}_{P}$ with ${\mathfrak d}\mid{\mathfrak f}$: \begin{equation} \label{eq:sss} {\mathrm{DR}}_{P}({\mathfrak f}) = \coprod_{{\mathfrak d}} [{\mathfrak d}]{\mathrm{Cl}}_{P}({\mathfrak f}{\mathfrak d}^{-1}) = \coprod_{{\mathfrak d}} [{\mathfrak d}]{\mathrm{Cl}}({\mathfrak f}{\mathfrak d}^{-1}) \end{equation} Let us simplify the right-hand side further and show \begin{equation} \label{eq:ttt} [{\mathfrak d}]{\mathrm{Cl}}({\mathfrak f}{\mathfrak d}^{-1}) = {\mathrm{Cl}}({\mathfrak f})/(1+{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1}/{\mathfrak f}_{\mathrm{fin}}). \end{equation} We use the ad\`elic description of ray class groups: ${\mathrm{Cl}}({\mathfrak m})={\mathbb A}_K^*/K^*{U}_{\mathfrak m}$. Then we have the exact sequence $$ {U}_{{\mathfrak f}{\mathfrak d}^{-1}}/{U}_{\mathfrak f} {\,\longrightarrow\,} {\mathrm{Cl}}({\mathfrak f}) {\,\longrightarrow\,} {\mathrm{Cl}}({\mathfrak f}{\mathfrak d}^{-1}) {\,\longrightarrow\,} 1. $$ Since the archimedean parts of ${\mathfrak f}{\mathfrak d}^{-1}$ and ${\mathfrak f}$ agree, we have $$ {U}_{{\mathfrak f}{\mathfrak d}^{-1}}/{U}_{\mathfrak f} = {U}_{{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1}}/{U}_{{\mathfrak f}_{\mathrm{fin}}} = 1+{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1}/{\mathfrak f}_{\mathrm{fin}}. $$ Equation (\ref{eq:ttt}) follows. Combining this with (\ref{eq:sss}), we have \begin{align*} {\mathrm{DR}}_{P}({\mathfrak f}) &\labeleq{} \coprod_{{\mathfrak d}} [{\mathfrak d}]{\mathrm{Cl}}({\mathfrak f})/(1+{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1}/{\mathfrak f}_{\mathrm{fin}}) \\ &\labeleq{} \coprod_{{\mathfrak d}} [{\mathfrak d}](\maxO_K/{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1})^*\oplus_{(\maxO_K/{\mathfrak f}_{\mathrm{fin}})^*}{\mathrm{Cl}}({\mathfrak f}) \\ &\labeleq{} \Big(\coprod_{{\mathfrak d}} [{\mathfrak d}](\maxO_K/{\mathfrak f}_{\mathrm{fin}}{\mathfrak d}^{-1})^*\Big)\oplus_{(\maxO_K/{\mathfrak f}_{\mathrm{fin}})^*}{\mathrm{Cl}}({\mathfrak f}) \\ &\labeleq{} \big((\maxO_K/{\mathfrak f}_{{P}})^{\circ}\times (\maxO_K/{\mathfrak f}^{{P}})^*\big) \oplus_{(\maxO_K/{\mathfrak f}_{\mathrm{fin}})^*}{\mathrm{Cl}}({\mathfrak f}) \\ &\labeleq{} \big((\maxO_K/{\mathfrak f}_{{P}})^{\circ}\times (\maxO_K/{\mathfrak f}^{{P}})^*\big) \oplus_{(\maxO_K/{\mathfrak f}_{P})^*\times (\maxO_K/{\mathfrak f}^{P})^*}{\mathrm{Cl}}({\mathfrak f}) \\ &\labeleq{} (\maxO_K/{\mathfrak f}_{{P}})^{\circ} \oplus_{(\maxO_K/{\mathfrak f}_{P})^*}{\mathrm{Cl}}({\mathfrak f}). \end{align*} \end{proof} \subsection{} \emph{Examples.} If $K$ has class number $1$ and ${P}$ is still Chebotarev dense, then we have \begin{equation} \label{eq:DR-PID2} {\mathrm{DR}}_{{P}}({\mathfrak f}) = (\maxO_K/{\mathfrak f}_{{P}})^{\circ}\oplus_{\maxO_{K,{\mathfrak f}_\infty}^*} (\maxO_K/{\mathfrak f}_{\mathrm{fin}}^{{P}})^*, \end{equation} where $\maxO_{K,{\mathfrak f}_\infty}^*$ is the subgroup of $\maxO_K^*$ consisting of units which are positive at all places dividing ${\mathfrak f}_\infty$. At a different extreme, if $K$ is arbitrary but ${P}=M_K$, then in the limit we have \begin{equation} \label{eq:Bora-DR} \lim_{\mathfrak f} {\mathrm{DR}}_{P}({\mathfrak f}) = \hat{\maxO}_K^\circ \oplus_{\hat{\maxO}_K^*}G_K^{\mathrm{ab}}. \end{equation} If $K$ has class number $1$ and ${P}=M_K$, then we have $$ {\mathrm{DR}}_{{P}}({\mathfrak f}) = (\maxO_K/{\mathfrak f}_{\mathrm{fin}})^{\circ}/\maxO_{K,{\mathfrak f}_\infty}^*. $$ \section{Global arguments} The purpose of this section is to prove theorem \ref{global-thm-intro} from the introduction. It will follow immediately from proposition \ref{abelian} and theorem \ref{global-thm2} below. We continue with the notation of the previous section. Also fix the following notation: \begin{align*} {E} &= \text{a finite \'etale $K$-algebra with a $\Lambda_{\scalA,{P}}$-structure} \\ S &= {\mathrm{Hom}}_K({E},{K^{\mathrm{sep}}}), \text{ with its continuous action of } {G_K}\times \mathrm{Id}_{{P}}. \end{align*} Define ${\mathfrak r}\in \mathrm{Id}_{{P}}$ by setting \begin{equation} \label{equ:r-def} {\mathrm{ord}}_{\mathfrak p}({\mathfrak r})=\inf\{i\ge 0\colon\; {\mathfrak p}^{i+1}S={\mathfrak p}^i S\} \end{equation} for each prime ${\mathfrak p}\in{P}$. This is well defined because ${\mathfrak p} S=S$ whenever ${\mathfrak p}$ is unramified in $\ringb$, by proposition \ref{unramified}. \begin{lemma} \label{gcd} Assume that ${E}$ has an integral $\Lambda_{\scalA,{P}}$-model $\ringb$. Let ${\mathfrak d}$ be an ideal in $\mathrm{Id}_{{P}}$, and put ${\mathfrak b}=\gcd({\mathfrak d},{\mathfrak r})$. Then ${\mathfrak d} S$ equals ${\mathfrak b} S$, and this $G_K$-set is unramified at all primes dividing ${\mathfrak d}{\mathfrak b}^{-1}$. \end{lemma} \begin{proof} Observe that for any prime ${\mathfrak p}\mid{\mathfrak d}{\mathfrak b}^{-1}$, we have ${\mathfrak p}{\mathfrak b} S={\mathfrak b} S$. Indeed, we have ${\mathrm{ord}}_{{\mathfrak p}}({\mathfrak b})={\mathrm{ord}}_{{\mathfrak p}}({\mathfrak r})$ and hence ${\mathfrak p}^{1+{\mathrm{ord}}_{{\mathfrak p}}({\mathfrak b})}S={\mathfrak p}^{{\mathrm{ord}}_{{\mathfrak p}}({\mathfrak b})}S$, by the definition of ${\mathfrak r}$. Then ${\mathfrak p}{\mathfrak b} S={\mathfrak b} S$ follows. This implies by induction that ${\mathfrak d} S={\mathfrak b} S$. It also implies that for each prime ${\mathfrak p}\mid{\mathfrak d}{\mathfrak b}^{-1}$, the action of ${\mathfrak p}$ on ${\mathfrak b} S={\mathfrak d} S$ is bijective. Therefore by theorem~\ref{local-thm}(1), this $G_K$-set is unramified at all primes ${\mathfrak p}\mid{\mathfrak d}{\mathfrak b}^{-1}$. \end{proof} \begin{proposition} \label{abelian} If ${E}$ has an integral $\Lambda_{\scalA,{P}}$-model and ${P}$ is Chebotarev dense, then the action of ${G_K}$ on $S$ factors through the abelianization of ${G_K}$. \end{proposition} \begin{proof} Let $\ringb$ be an integral $\Lambda_{\scalA,{P}}$-model of ${E}$. For each prime ${\mathfrak p}\in{P}$ we consider the completion $\scalA_{\mathfrak p}$, and its fraction field $K_{\mathfrak p}$. Then we obtain an ${\Lambda_{\gpO_\gp,\gp}}$-structure on the finite \'etale $K_{\mathfrak p}$-algebra ${E}_{\mathfrak p}= {E}\otimes_K K_{\mathfrak p}$, and then $\ringb\otimes_\scalA \scalA_{\mathfrak p}$ is an integral ${\Lambda_{\gpO_\gp,\gp}}$-model of ${E}_{\mathfrak p}$. Fixing an embedding ${K^{\mathrm{sep}}}\to {K_{\mathfrak p}^{\mathrm{sep}}}$ for each ${\mathfrak p}$ we can view ${G_\gp}$ as a subgroup of ${G_K}$. The finite \'etale $K_{\mathfrak p}$-algebra ${E}_{\mathfrak p}$ then corresponds to the ${G_\gp}$-set that one gets by restricting the action of ${G_K}$ on $S$ to ${G_\gp}$. Now let ${\bar{G}}$ be the image of the action map ${G_K}\to{\mathrm{Map}}(S,S)$. Because ${P}$ is Chebotarev dense, for each $g\in{\bar{G}}$ there is a prime ${\mathfrak p}={\mathfrak p}_g$ in ${P}$ such that \begin{enumerate} \item $\ringb$ is unramified at ${\mathfrak p}$, \item the image of $F_{\mathfrak p}\in {G_\gp}/{I_\gp}$ under the induced map ${G_\gp}/{I_\gp}\to{\bar{G}}$ is $g$. \end{enumerate} By proposition \ref{unramified}, the action of $g$ on $S$ is the same as the action of ${\mathfrak p}_g$ on $S$. But by the definition of $\Lambda_{\scalA,{P}}$-structure, the ${\mathfrak p}_g$ commute with each other. Therefore ${\bar{G}}$ is abelian. \end{proof} \subsection{} \emph{Conductors.} By class field theory, any continuous action of the abelianization of ${G_K}$ on a finite discrete set $T$ factors, by the Artin map, through the ray class group ${\mathrm{Cl}}(c(T))$ for a minimal cycle $c(T)$ on $K$, which we call the conductor of $T$. \begin{lemma}\label{lem:equiv} Assume that ${E}$ has an integral $\Lambda_{\scalA,{P}}$-model $\ringb$ and that the action of ${G_K}$ on $S$ factors through its abelianization. Let ${\mathfrak f}$ be a cycle on $K$, and let ${\mathfrak r}$ be as in (\ref{equ:r-def}). Then following are equivalent: \begin{enumerate} \item the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ factors through an action of ${\mathrm{DR}}_{{P}}({\mathfrak f})$, \item ${\mathfrak r}$ divides ${\mathfrak f}$, and for each ideal ${\mathfrak d}\mid{\mathfrak f}$ we have $c({\mathfrak d} S) \mid {\mathfrak f}{\mathfrak d}^{-1}$, \item ${\mathfrak r}$ divides ${\mathfrak f}$, and for each ideal ${\mathfrak d}\mid{\mathfrak r}$ we have $c({\mathfrak d} S) \mid {\mathfrak f}{\mathfrak d}^{-1}$, \item the least common multiple ${\mathrm{lcm}}_{{\mathfrak d}\mid {\mathfrak r}} \big({\mathfrak d}\, c({\mathfrak d} S)\big)$ divides ${\mathfrak f}$. \end{enumerate} \end{lemma} \begin{proof} (1)$\Rightarrow$(2): To show ${\mathfrak r}\mid{\mathfrak f}$, we will show that for all ${\mathfrak p}\in{P}$, we have ${\mathfrak p}^{n}S \subseteq {\mathfrak p}^{n+1} S$, where $n={\mathrm{ord}}_{\mathfrak p}({\mathfrak f})$. Using the decomposition (\ref{DR-decomp}), we see $[{\mathfrak p}]^{n+1}{\mathrm{DR}}_{P}({\mathfrak f})=[{\mathfrak p}]^n{\mathrm{DR}}_{P}({\mathfrak f})$ and hence $[{\mathfrak p}]^n=[{\mathfrak p}]^{n+1}x$ for some $x\in{\mathrm{DR}}_{P}({\mathfrak f})$. Then because the action of $\mathrm{Id}_P$ is assumed to factor through ${\mathrm{DR}}_{{P}}({\mathfrak f})$, we have ${\mathfrak p}^n S = {\mathfrak p}^{n+1} xS \subseteq {\mathfrak p}^{n+1} S$. Second, the condition $c({\mathfrak d} S) \mid {\mathfrak f}{\mathfrak d}^{-1}$ of (2) is equivalent to the condition that the action of ${G_K}$ on ${\mathfrak d} S$ factors through the Artin map ${G_K}\to{\mathrm{Cl}}({\mathfrak f}{\mathfrak d}^{-1})$. But this holds by the assumption (1) and the decomposition~(\ref{DR-decomp}). (2)$\Rightarrow$(1): Consider an element $(\sigma,{\mathfrak a})\in{G_K}\times\mathrm{Id}_{P}$. We will show its action on $S$ depends only on its class in ${\mathrm{DR}}_{P}({\mathfrak f})$. First observe that by the assumption $c({\mathfrak d} S) \mid {\mathfrak f}{\mathfrak d}^{-1}$, taken when ${\mathfrak d}=(1)$, we have $c(S) \mid {\mathfrak f}$. This implies that the action of ${G_K}$ factors through the Artin map ${G_K}\to{\mathrm{Cl}}({\mathfrak f})$, and so it is enough to show that the action of ${\mathfrak a}$ depends only on its class $[{\mathfrak a}]\in{\mathrm{DR}}_{P}({\mathfrak f})$. Now put ${\mathfrak d}=\gcd({\mathfrak a},{\mathfrak f})$ and ${\mathfrak a}'={\mathfrak a}{\mathfrak d}^{-1}$. Then we have ${\mathfrak a}'\in\mathrm{Id}_{P}({\mathfrak f})$. Therefore, because of (\ref{DR-decomp}), it is enough to show that the action of ${\mathfrak a}'\in\mathrm{Id}_{P}({\mathfrak f})$ on ${\mathfrak d} S$ depends only on its class $[{\mathfrak a}']\in{\mathrm{Cl}}_{P}({\mathfrak f}{\mathfrak d}^{-1})$. In particular, it is enough to show \begin{equation} \label{eq23498} {\mathfrak a}'{\mathfrak d} s = F_{{\mathfrak a}'}{\mathfrak d} s, \end{equation} for all $s\in S$, where $F_{{\mathfrak a}'}\in{G_K}$ is an element mapping to $[{\mathfrak a}']$ under the Artin map ${G_K}\to{\mathrm{Cl}}({\mathfrak f})$. To do this, it is enough to consider the case where ${\mathfrak a}'$ is a prime ${\mathfrak p}\in\mathrm{Id}_{P}({\mathfrak f})$. Now by our assumption ${\mathfrak r}\mid{\mathfrak f}$, we have ${\mathfrak p}\nmid{\mathfrak r}$ and hence ${\mathfrak p}$ acts bijectively on $S$. Therefore by theorem~\ref{local-thm}, we have $$ {\mathfrak p} s = F_{\mathfrak p} s, $$ for all $s\in S$. This implies (\ref{eq23498}) and hence (1). (2) $\Leftrightarrow$ (3): The implication (2) $\Rightarrow$ (3) is clear. So consider the other direction. Given an ideal ${\mathfrak d}\mid{\mathfrak f}$, let ${\mathfrak b}$ denote $\gcd({\mathfrak d},{\mathfrak r})$. Then by lemma \ref{gcd}, we have equivalences $$ c({\mathfrak d} S) \mid {\mathfrak f}{\mathfrak d}^{-1} {\;\Leftrightarrow\;} c({\mathfrak b} S)\mid{\mathfrak f}{\mathfrak d}^{-1} {\;\Leftrightarrow\;} c({\mathfrak b} S)\mid{\mathfrak f}{\mathfrak d}^{-1}{\mathfrak d}{\mathfrak b}^{-1} {\;\Leftrightarrow\;} c({\mathfrak b} S)\mid{\mathfrak f}{\mathfrak b}^{-1}. $$ Therefore since ${\mathfrak b}$ is a divisor of ${\mathfrak r}$, the condition $c({\mathfrak d} S)\mid {\mathfrak f}{\mathfrak d}^{-1}$ holds for all ${\mathfrak d}\mid{\mathfrak f}$ if it holds for all ${\mathfrak d}\mid{\mathfrak r}$. (3) $\Leftrightarrow$ (4): Clear. \end{proof} \begin{thm} \label{global-thm2} Let ${\mathfrak f}$ be a cycle on $K$. Then the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ factors (necessarily uniquely) through the map ${G_K}\times \mathrm{Id}_{{P}} {\,\longrightarrow\,} {\mathrm{DR}}_{{P}}({\mathfrak f})$ of (\ref{DR-recmap}) if and only if the following hold: \begin{enumerate} \item the action of ${G_K}$ on $S$ factors through its abelianization, \item ${\mathrm{lcm}}_{{\mathfrak d}\mid {\mathfrak r}} \big({\mathfrak d}\, c({\mathfrak d} S)\big)$ divides ${\mathfrak f}$, \item ${E}$ has an integral $\Lambda_{\scalA,{P}}$-model. \end{enumerate} If ${P}$ is Chebotarev dense, then condition (1) can be removed. \end{thm} \begin{proof} If (1)--(3) hold, then by lemma~\ref{lem:equiv}, the action of ${G_K}\times \mathrm{Id}_{{P}}$ on $S$ factors through ${\mathrm{DR}}_{{P}}({\mathfrak f})$; and if ${P}$ is Chebotarev dense, then (1) can be removed because, by proposition~\ref{abelian}, it follows from (3). Let us now consider the converse direction. Suppose that the ${G_K}\times \mathrm{Id}_{{P}}$-action on $S$ factors through ${\mathrm{DR}}_{{P}}({\mathfrak f})$. Then (1) holds because ${\mathrm{DR}}_{P}({\mathfrak f})$ is commutative. Further, (3) implies (2) by lemma~\ref{lem:equiv}. Therefore it is enough to show (3). For each ${\mathfrak p}\in {P}$, let $\ringb_{\mathfrak p}$ denote the maximal sub-$\Lambda_{\scalA_{\mathfrak p},{\mathfrak p}}$-ring of ${E}\otimes_K K_{\mathfrak p}$ which is finite over $\scalA_{\mathfrak p}$. As mentioned in~\ref{rmk-group-alg}, it always exists. We will show that $\ringb_{\mathfrak p}\otimes_{\scalA_{\mathfrak p}} K_{\mathfrak p}$ agrees with $E_{\mathfrak p}$. To do this, it is enough to show that ${E}\otimes_K K_{\mathfrak p}$ has an integral $\Lambda_{\scalA_{\mathfrak p},{\mathfrak p}}$-model. So write ${\mathfrak f}={\mathfrak p}^n{\mathfrak f}'$ with $n={\mathrm{ord}}_{\mathfrak p}({\mathfrak f})$. Then $[{\mathfrak p}^k] \in [{\mathfrak p}^n]{\mathrm{Cl}}({\mathfrak f}')\subseteq {\mathrm{DR}}({\mathfrak f})$ for all $k\ge n$. This implies, by say~(\ref{eq:DR-mult}), that the action of ${\mathfrak p}$ on $\bigcap_i{\mathfrak p}^i S={\mathfrak p}^nS$ is given by the Artin symbol of $[{\mathfrak p}]\in {\mathrm{Cl}}({\mathfrak f}')$, which by our local result, theorem \ref{local-thm}, guarantees existence of an integral $\Lambda_{\scalA_{\mathfrak p},{\mathfrak p}}$-model. Now let $R$ denote the integral closure of $\scalA$ in ${E}$, and let $\ringb$ denote the set of elements $a\in R$ such that the image of $a$ in ${E}\otimes_K K_{\mathfrak p}$ lies in $\ringb_{\mathfrak p}$ for all ${\mathfrak p}\in{P}$. We will show that $\ringb$ is what we seek, an integral $\Lambda_{\scalA,{P}}$-model for $E$. For all ${\mathfrak p}\nmid {\mathfrak f}$, we are in the unramified case, and so $\ringb_{\mathfrak p}$ is $R\otimes_\scalA{{\mathcal O}_{\mathfrak p}}$, by proposition \ref{unramified}. It follows that $\ringb$ is of finite index in $R$. Therefore, since $R$ is finite and flat over $\scalA$ and we have ${E}=R\otimes_\scalA K$, the same hold for $\ringb$. Further, $\ringb$ is closed under all $\psi_{\mathfrak q}$ with ${\mathfrak q}\in{P}$. Indeed it is enough to show $\psi_{\mathfrak q}(\ringb_{\mathfrak p})\subseteq \ringb_{\mathfrak p}$ for all ${\mathfrak p},{\mathfrak q}\in{P}$; and this holds because is a $\psi_{\mathfrak q}(\ringb_{\mathfrak p})$ sub-$\Lambda_{\scalA_{\mathfrak p},{\mathfrak p}}$-ring of $E_{\mathfrak p}$ which is finite over $\scalA_{\mathfrak p}$, and so it is contained in the maximal one $\ringb_{\mathfrak p}$. Finally, for each ${\mathfrak p}\in {P}$, the induced endomorphism $\psi_{\mathfrak p}$ of $\ringb$ is a Frobenius lift, because $\ringb\otimes_\scalA {{\mathcal O}_{\mathfrak p}}$ is a $\Lambda_{\scalA_{\mathfrak p},{\mathfrak p}}$-ring. Therefore $\ringb$ is an integral $\Lambda_{\scalA,{P}}$-model for $E$. This establishes (3) and, hence, (1) and (2) as explained above. \end{proof} \begin{corollary} \label{cor:lambda-Galois-category-equiv} If ${P}$ is Chebotarev dense, there is a contravariant equivalence between the category of finite \'etale $\Lambda_{\scalA,{P}}$-rings over $K$ which admit an integral model and the category of finite discrete sets with a continuous action of the profinite monoid $\lim_{\mathfrak f} {\mathrm{DR}}_{{P}}({\mathfrak f})$. \end{corollary} \part{Periodic loci and explicit class field theory} \section{$\Lambda$-schemes} Below $X$ will denote a flat $\scalA$-scheme. \subsection{} \emph{$\Lambda$-schemes.} Let ${\mathfrak p}$ be a maximal ideal of $\scalA$. As in the affine case, the fiber $X\times_{\Spec\scalA}\Spec k({\mathfrak p})$ has a natural $k({\mathfrak p})$-scheme endomorphism $F_{{\mathfrak p}}$ which is the identity map on the underlying topological space and such that for each affine open subscheme $\Spec \ringb$, the induced the endomorphism of $\ringb$ is the affine Frobenius map $x\mapsto x^{N({\mathfrak p})}$. Let ${\mathrm{End}}_{\scalA}(X)$ denote the monoid of $\scalA$-scheme endomorphisms of $X$. An endomorphism $\psi\in{\mathrm{End}}_{\scalA}(X)$ is said to be a \emph{Frobenius lift} at ${\mathfrak p}$ if the induced endomorphism on the fiber $X\times_{\Spec \scalA}\Spec k({\mathfrak p})$ agrees with $F_{{\mathfrak p}}$. A \emph{$\Lambda_{\scalA,{P}}$-structure} on a flat $\scalA$-scheme $X$ is defined to be a set map ${P}\to{\mathrm{End}}_{\scalA}(X)$, denoted ${\mathfrak p}\mapsto \psi_{{\mathfrak p}}$ such that $\psi_{{\mathfrak p}}$ is a Frobenius lift at ${\mathfrak p}$ for each ${\mathfrak p}\in{P}$ and such that $\psi_{{\mathfrak p}}\circ\psi_{{\mathfrak q}}=\psi_{{\mathfrak q}}\circ\psi_{{\mathfrak p}}$ for all ${\mathfrak p},{\mathfrak q}\in{P}$. We will call an $\scalA$-scheme with $\Lambda_{\scalA,{P}}$-structure a \emph{$\Lambda_{\scalA,{P}}$-scheme}. (When $X$ is not flat over $\scalA$, this definition still makes sense; but as in the affine case, it is not well behaved. In general, one should define it to be an action of the Witt vector monad $W_{\scalA,{P}}^*$ as in the introduction to~\cite{Borger:BGWV-II}. We will only consider $\Lambda$-structures on flat schemes in this paper; so the simplified definition above is good enough here.) For any ideal ${\mathfrak a}\in\mathrm{Id}_{{P}}$ with prime factorization ${\mathfrak a}={\mathfrak p}_1\cdots{\mathfrak p}_n$, let $\psi_{{\mathfrak a}}$ denote the composition $\psi_{{\mathfrak p}_1}\circ\cdots\circ \psi_{{\mathfrak p}_n}$. It is independent of the order of the factors because the operators $\psi_{{\mathfrak p}}$ commute with each other. A morphism $X\to Y$ of $\Lambda_{\scalA,{P}}$-schemes is a morphism $f\colon X\to Y$ of $\scalA$-schemes such that $f\circ\psi_{\mathfrak p}=\psi_{\mathfrak p}\circ f$, for all ${\mathfrak p}\in{P}$. In this way, $\Lambda_{\scalA,{P}}$-schemes form a category. \subsection{}\emph{Examples.} \label{subsec:toric-lambda-schemes} The multiplicative group $X={\mathbb G}_{\mathrm{m}}$ over $\scalA$ has a $\Lambda_{\scalA,{P}}$-structure given by $\psi_{\mathfrak p}(x)=x^{N({\mathfrak p})}$. This extends uniquely to $\Lambda_{\scalA,{P}}$-structures on ${\mathbb A}^1$ and ${\mathbb P}^1$. More generally, projective $n$-space ${\mathbb P}^n$ has a $\Lambda_{\scalA,{P}}$-structure where $\psi_{\mathfrak p}$ raises the homogeneous coordinates to the $N({\mathfrak p})$ power. Any product of $\Lambda_{\scalA,{P}}$-schemes is again a $\Lambda_{\scalA,{P}}$-scheme, where the $\psi$-operators act componentwise. In a similar way, coproducts of $\Lambda_{\scalA,{P}}$-schemes are $\Lambda_{\scalA,{P}}$-schemes. \section{Periodic $\Lambda$-schemes} Let ${\mathfrak f}$ be a cycle on $K$. \subsection{} \emph{Periodic $\Lambda$-schemes.} We will say that a $\Lambda_{\scalA,{P}}$-scheme $X$ is \emph{${\mathfrak f}$-periodic} if for all ${\mathfrak f}$-equivalent ideals ${\mathfrak a},{\mathfrak b}\in\mathrm{Id}_{{P}}$, the two maps $\psi_{{\mathfrak a}},\psi_{{\mathfrak b}}\colon X \to X$ are equal---in other words, the monoid map $\mathrm{Id}_{P}\to{\mathrm{End}}_A(X)$ factors through ${\mathrm{DR}}_{P}({\mathfrak f})$. \subsection{} \emph{Examples.} Suppose that $\scalA={\mathbb Z}$, ${P}=M_{\mathbb Q}$, and ${\mathfrak f}=(n)\infty$ with $n$ positive. Then ${\mathfrak f}$-periodicity means that $\psi_{(a)}=\psi_{(a+n)}$ for all integers $a\geq 1$. In other words, the sequence of Frobenius lifts $\psi_{(a)}$ is periodic in $a\geq 1$ with period dividing $n$, which is the reason for the name. Representation rings of finite groups, with their usual $\lambda$-ring structure in algebraic K-theory, are periodic. In fact, periodicity was first introduced in this context by Davydov~\cite{Davydov:periodic-lambda-rings}. On the other hand, when $\scalA$ is general but ${\mathfrak f}$ is $(1)$, then ${\mathfrak f}$-periodicity means that $\psi_{{\mathfrak a}}$ depends only on the class of ${\mathfrak a}$ in the class group ${\mathrm{Cl}}(1)$. If ${\mathfrak f}$ is instead the product of all real places, then it means that $\psi_{{\mathfrak a}}$ depends only on the class of ${\mathfrak a}$ in the narrow class group ${\mathrm{Cl}}({\mathfrak f})$. In particular, if either of these class groups is trivial, then every $\psi_{\mathfrak a}$ is the identity map, and so any (flat) $\scalA$-scheme has at most one $\Lambda_{\scalA,{P}}$-structure with that type of periodicity. If we are in the intersection of the two cases above, a $(1)$-periodic $\Lambda_{{\mathbb Z},{P}}$-ring is just a $\lambda$-ring in which all the Adams operations are the identity. Elliott~\cite{Elliott:binomial-rings} has proved that this is equivalent to being a binomial ring, a notion which dates back to Berthelot's expos\'e in SGA6~\cite{SGA6}, p.\ 323. \begin{proposition} \label{pro:periodic-structure2} Let $X$ be a separated flat ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-scheme of finite type. \begin{enumerate} \item If ${P}$ is infinite, then $X$ is affine, reduced, and quasi-finite over $\scalA$, with \'etale generic fiber $X_K$. \item If ${P}$ is Chebotarev dense, then the generic fiber's function algebra $\sheafO(X_K)$ is a product of subextensions of the ray class field $K({\mathfrak f})$. \end{enumerate} \end{proposition} \begin{proof} (1): Let us first show that $X$ is reduced and quasi-finite over $\scalA$. By periodicity, for every ${\mathfrak p}\in{P}$ satisfying ${\mathfrak p}\nmid{\mathfrak f}$, the Frobenius lift $\psi_{{\mathfrak p}}$ on $X$ is an automorphism of finite order. Therefore the Frobenius map $F_{\mathfrak p}$ on the fiber $k({\mathfrak p})\otimes X$ is an automorphism of finite order, and so the fiber is both geometrically reduced and finite over $k({\mathfrak p})$. Since ${P}$ is infinite and since the set of prime ideals of $\scalA$ with geometrically reduced fibers forms a constructible subset of $\Spec\scalA$, by EGA IV (9.7.7) \cite{EGA-no.28}, the generic fiber $X\times_{\Spec \scalA}\Spec K$ must be geometrically reduced. Similarly, since infinitely many fibers are finite, the generic fiber is also finite, by EGA IV (9.2.6.2)~\cite{EGA-no.28}. We now use the flatness of $X$ over $\scalA$ to pass from the generic fiber to all of $X$. It is clear that flatness implies $X$ is reduced. For quasi-finiteness, apply EGA IV (14.2.4) \cite{EGA-no.28}. Therefore $X$ is reduced and is quasi-finite over $\scalA$. Affineness then follows from Zariski's Main Theorem, as we now explain. Let $\ringd$ denote the integral closure of $\scalA$ in $\sheafO(X)$. Let $X'$ denote $\Spec \ringd$, called the normalization of $\scalA$ in $X$ in the Stacks Project \cite[\href{http://stacks.math.columbia.edu/tag/035H}{Tag 035H}]{stacks-project}. Then $\sheafO(X)$ is reduced and is flat over $\scalA$, and hence so is $\ringd$. It is also finite over $\scalA$, by \cite[\href{http://stacks.math.columbia.edu/tag/03GR}{Tag 03GR}]{stacks-project}. Since $X$ is quasi-finite, the canonical map $X\to X'$ is an open immersion by Zariski's Main Theorem \cite[\href{http://stacks.math.columbia.edu/tag/03GW}{Tag 03GW}]{stacks-project}. Finally, since $X'$ is finite flat over $\scalA$, which is a Dedekind domain, its Krull dimension is $1$ and hence its open subscheme $X$ must be affine \cite[\href{http://stacks.math.columbia.edu/tag/09N9}{Tag 09N9}]{stacks-project}. Thus we can write $X=\Spec \ringb$, where $\ringb$ is an ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-ring which is reduced and is flat, quasi-finite, and of finite type over $\scalA$. It follows that $K\otimes_{\scalA}\ringb$ is \'etale over $K$. (2): It follows from statement (1) that there is a nonzero element $t\in \scalA$ such that $\ringb[1/t]$ is a finite product $\prod_i \ringd_i$, where each $\ringd_i$ is the integral closure of $\scalA[1/t]$ in a finite extension $L_i$ of $K$: Indeed, since $\ringb$ is of finite type over $\scalA$ and since $\ringb\otimes_{\scalA}K$ is finite over $K$, there is an element $t\in\scalA$ such that $\ringb[1/t]$ is finite flat over $\scalA[1/t]$. Since $\ringb$ is reduced, so is $\ringb[1/t]$, and hence the discriminant ideal of $\ringb[1/t]$ over $\scalA[1/t]$ is nonzero. Then by scaling $t$ so that it lies in the discriminant ideal, and is nonzero, we may assume that $\ringb[1/t]$ is finite \'etale over $\scalA[1/t]$. It then follows that $\ringb[1/t]$ is the integral closure of $\scalA[1/t]$ in $K\otimes_{\scalA}\ringb$ and is hence of the required form. Now let ${P}_t$ denote the set of primes ${\mathfrak p}$ in ${P}$ that do not divide $t$. Then we can consider ${P}_t$ as a set of primes of $\scalA[1/t]$; and since ${P}$ is Chebotarev dense, so is ${P}_t$. Applying proposition \ref{abelian} to the $\Lambda_{\scalA[1/t],{P}_t}$-ring $\ringb[1/t]$, we see each field $K\otimes_\scalA \ringd_i$ is an abelian extension of $K$. Since the Frobenius elements act on each $\ringd_i$ with period ${\mathfrak f}$, the conductor of $K\otimes_\scalA \ringd_i$ divides ${\mathfrak f}$. \end{proof} \subsection{} \emph{Ray class algebras.} \label{subsec:ray-class-lambda-rings} These are $\Lambda$-ring analogues of the ray class fields of $K$. We can view ${\mathrm{DR}}_{{P}}({\mathfrak f})$ as a pointed ${\mathrm{DR}}_{{P}}({\mathfrak f})$-set: the distinguished point is the identity element, and the action is translation. By theorem~\ref{global-thm2}, the corresponding $\Lambda_{\scalA,{P}}$-ring over $K$ has an integral model. Define $\rcl{\scalA,{P}}{{\mathfrak f}}$, the \emph{ray class algebra of conductor ${\mathfrak f}$}, to be the maximal $\Lambda_{\scalA,{P}}$-order in this $K$-algebra. Thus we have $$ K\otimes_{\scalA} \rcl{\scalA,{P}}{{\mathfrak f}} = \prod_{\substack{{\mathfrak d}\in\mathrm{Id}_{{P}}\\{\mathfrak d}\mid{\mathfrak f}_{{\mathrm{fin}}}}} K({\mathfrak f}{\mathfrak d}^{-1}), $$ where $K({\mathfrak f}{\mathfrak d}^{-1})$ is the ray class field of $K$ with conductor ${\mathfrak f}{\mathfrak d}^{-1}$. Under this identification, the map $\beta\colon \rcl{\scalA,{P}}{{\mathfrak f}}\to{K^{\mathrm{sep}}}$ coming from the distinguished point of ${\mathrm{DR}}_{{P}}({\mathfrak f})$ is the projection to the component with ${\mathfrak d}=(1)$. In our previous paper \cite{Borger-deSmit:Integral-models}, we considered the case where $\scalA={\mathbb Z}$ and ${P}$ is all maximal ideals. There we showed that the ray class algebra of conductor $(n)\infty$ is ${\mathbb Z}[x]/(x^n-1)$, or more naturally, the group ring on the cyclic group $\mu_n({K^{\mathrm{sep}}})$ of $n$-th roots of unity in ${K^{\mathrm{sep}}}$. Observe that the ray class algebra is not usually a domain, and in particular the map $\beta\:\rcl{\scalA,{P}}{{\mathfrak f}}\to K({\mathfrak f})$ to the ray class field is not usually injective. Also, unlike the ray class field, it depends not only on ${\mathfrak f}$ but also on $\scalA$ and ${P}$. On the other hand, $K\otimes_{\scalA} \rcl{\scalA,{P}}{{\mathfrak f}}$ is independent of $\scalA$. If ${P}$ is Chebotarev dense, the ray class algebra $\rcl{\scalA,{P}}{{\mathfrak f}}$ satisfies the following maximality property: if $\ringd$ is a reduced finite flat ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-ring equipped with a map $\alpha\colon \ringd\to{K^{\mathrm{sep}}}$ of $\scalA$-algebras, then there is a unique map $\varphi\colon \ringd\to \rcl{\scalA,{P}}{{\mathfrak f}}$ of $\Lambda_{\scalA,{P}}$-rings making the following diagram commute: $$ \xymatrix{ & {K^{\mathrm{sep}}} \\ \ringd \ar^{\alpha}[ur] \ar@{-->}^{\varphi}[rr] & & \rcl{\scalA,{P}}{{\mathfrak f}}. \ar_{\beta}[ul] } $$ This is simply because, under the anti-equivalence with ${\mathrm{DR}}_{{P}}({\mathfrak f})$-sets, $\rcl{\scalA,{P}}{{\mathfrak f}}$ corresponds to ${\mathrm{DR}}_{{P}}({\mathfrak f})$, which is the free ${\mathrm{DR}}_{{P}}({\mathfrak f})$-set on one generator. \subsection{} \emph{Change of ${\mathfrak f}$.} Suppose ${\mathfrak f}'={\mathfrak f}{\mathfrak a}$. Then the maps ${\mathrm{DR}}_{P}({\mathfrak f}')\to {\mathrm{DR}}_{{P}}({\mathfrak f})$ and ${\mathrm{DR}}_{P}({\mathfrak f}) {\,\longrightarrow\,} {\mathrm{DR}}_{P}({\mathfrak f}')$ of~\ref{subsec:DR-change-of-f} induce an inclusion \begin{equation} u\:\rcl{\scalA,{P}}{{\mathfrak f}} {\,\longrightarrow\,} \rcl{\scalA,{P}}{{\mathfrak f}'} \end{equation} and a surjection \begin{equation} v\:\rcl{\scalA,{P}}{{\mathfrak f}'} {\,\longrightarrow\,} \rcl{\scalA,{P}}{{\mathfrak f}} \end{equation} of $\Lambda_{\scalA,{P}}$-rings, and the compositions $u\circ v$ and $u\circ v$ agree with the two $\psi_{\mathfrak a}$ endomorphisms. In the case where $K={\mathbb Q}$, ${P}$ is all maximal ideals, ${\mathfrak f}=(n)\infty$, and ${\mathfrak f}'=(n')\infty$, these maps can be identified with the maps on group rings corresponding to the inclusion $\mu_{n}({K^{\mathrm{sep}}})\subseteq \mu_{n'}({K^{\mathrm{sep}}})$, in the case of $u$, and the $n'/n$-th power map $\mu_{n'}({K^{\mathrm{sep}}})\to \mu_{n}({K^{\mathrm{sep}}})$, in the case of $v$. \section{Periodic loci and abelian extensions} Let ${\mathfrak f}$ be a cycle on $K$. Let $X$ be a separated (flat) $\Lambda_{\scalA,{P}}$-scheme. \subsection{} \emph{Periodic locus $\per{X}{{\mathfrak f}}$.} \label{subsec:periodic-loci} Define the ${\mathfrak f}$-periodic locus $\per{X}{{\mathfrak f}}$ of $X$ to be the scheme-theoretic intersection $$ \per{X}{{\mathfrak f}} = \bigcap_{\substack{{\mathfrak a},{\mathfrak b}\in\mathrm{Id}_{{P}}\\{\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}}} X(\psi_{{\mathfrak a}}=\psi_{{\mathfrak b}}), $$ where $X(\psi_{{\mathfrak a}}=\psi_{{\mathfrak b}})$ denotes the equalizer of the two maps $\psi_{{\mathfrak a}},\psi_{{\mathfrak b}}\colon X\to X$. Since $X$ is separated, $\per{X}{{\mathfrak f}}$ is a closed subscheme. The functor $\per{X}{{\mathfrak f}}$ represents is $$ \per{X}{{\mathfrak f}}(C) = \{x\in X(C): \psi_{{\mathfrak a}}(x)=\psi_{{\mathfrak b}}(x) \text{ for all } {\mathfrak a},{\mathfrak b}\in\mathrm{Id}_{{P}} \text{ with } {\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}\}. $$ We emphasize that $\per{X}{{\mathfrak f}}$ depends on ${P}$ as well as ${\mathfrak f}$, although the notation does not reflect this. Observe that $X({\mathfrak f})$ is functorial in $X$. It also behaves well under change of ${\mathfrak f}$: if ${\mathfrak f}'$ is another cycle and ${\mathfrak f}\mid{\mathfrak f}'$, then we have $$ \per{X}{{\mathfrak f}} \subseteq \per{X}{{\mathfrak f}'}. $$ Finally, let $\perfl{X}{{\mathfrak f}}$ denote the maximal $\scalA$-flat subscheme of $\per{X}{{\mathfrak f}}$. It is the closed subscheme of $\per{X}{{\mathfrak f}}$ defined by the ideal sheaf of $\scalA$-torsion elements. Although the actual periodic locus $\per{X}{{\mathfrak f}}$ is more fundamental than $\perfl{X}{{\mathfrak f}}$, for the purposes of this paper it is enough to consider $\perfl{X}{{\mathfrak f}}$, and doing so will allow us to avoid the subtleties of $\Lambda$-rings with torsion. \begin{proposition} \label{pro:preimage-of-periodic-locus} For any ideal ${\mathfrak a}\in\mathrm{Id}_{{P}}$, we have the subscheme inclusion \begin{equation} \label{eq:382} \per{X}{{\mathfrak a}{\mathfrak f}}\subseteq \psi_{{\mathfrak a}}^{-1}(\per{X}{{\mathfrak f}}). \end{equation} \end{proposition} \begin{proof} Let $x$ be a point of $\per{X}{{\mathfrak a}{\mathfrak f}}$ with coordinates in some ring $R$. Let ${\mathfrak b}$ and ${\mathfrak c}$ be ${\mathfrak f}$-equivalent ideals in $\mathrm{Id}_{{P}}$. Then we have ${\mathfrak a}{\mathfrak b}\sim_{{\mathfrak a}{\mathfrak f}}{\mathfrak a}{\mathfrak c}$ and hence $$ \psi_{{\mathfrak b}}\big(\psi_{{\mathfrak a}}(x)\big)=\psi_{{\mathfrak a}{\mathfrak b}}(x)=\psi_{{\mathfrak a}{\mathfrak c}}(x)=\psi_{{\mathfrak c}}\big(\psi_{{\mathfrak a}}(x)\big). $$ It follows that $\psi_{{\mathfrak a}}(x)\in \per{X}{{\mathfrak f}}$, and this implies~(\ref{eq:382}). \end{proof} \subsection{} \emph{Torsion locus $X[{\mathfrak f}]$.} Write ${\mathfrak f}={\mathfrak f}_{{\mathrm{fin}}}{\mathfrak f}_{\infty}$. Then we define the \emph{${\mathfrak f}$-torsion} locus by $$ X[{\mathfrak f}] = \psi_{{\mathfrak f}_{{\mathrm{fin}}}}^{-1}(\per{X}{{\mathfrak f}_{\infty}}). $$ It follows from proposition~\ref{pro:preimage-of-periodic-locus} that we have an inclusion of subschemes \begin{equation} \label{eq:periodic-in-torsion} \per{X}{{\mathfrak f}} \subseteq X[{\mathfrak f}]. \end{equation} For example, if $\scalA={\mathbb Z}$, ${P}=M_{\mathbb Q}$, and ${\mathfrak f}=(n)\infty$, then this inclusion is an equality because both sides are $\mu_n$. If however ${\mathfrak f}=(n)$, then the ${\mathfrak f}$-torsion locus is again $\mu_n$, but the ${\mathfrak f}$-periodic locus ${\mathbb G}_{\mathrm{m}}({\mathfrak f})$ is $\mu_m$, where $m={\gcd(2,n)}$. So the containment~(\ref{eq:periodic-in-torsion}) can be far from an equality. \vspace{3mm} Summing up the results above with the previous section, we have the following: \begin{thm} \label{thm:periodic-locus3} If $X$ is of finite type over $\scalA$ and ${P}$ is Chebotarev dense, then we have the following: \begin{enumerate} \item The flat ${\mathfrak f}$-periodic locus $\perfl{X}{{\mathfrak f}}$ is a closed ${\mathfrak f}$-periodic sub-$\Lambda_{\scalA,{P}}$-scheme of $X$, and it is the maximal flat closed subscheme with these properties. \item We have $\perfl{X}{{\mathfrak f}}=\Spec \ringb$, where $\ringb$ is a finitely generated $\scalA$-algebra, and $K\otimes_\scalA \ringb$ is a finite product of abelian extensions of $K$ of conductor dividing ${\mathfrak f}$. \item If $X$ is proper, then $\ringb$ is a finite $\scalA$-algebra. \end{enumerate} \end{thm} \begin{proof} (1) By functoriality, the subscheme $\per{X}{{\mathfrak f}}$ is stable under the operators $\psi_{{\mathfrak p}}$, for all ${\mathfrak p}\in{P}$. Again by functoriality, $\perfl{X}{{\mathfrak f}}$ is also stable under them. Since the $\psi_{{\mathfrak p}}$ are Frobenius lifts on $X$, so are the endomorphisms they induce on the closed subscheme $\perfl{X}{{\mathfrak f}}$. Since $\perfl{X}{{\mathfrak f}}$ is flat, this defines a $\Lambda_{\scalA,{P}}$-structure on $\perfl{X}{{\mathfrak f}}$. It is obviously ${\mathfrak f}$-periodic. Maximality is also clear. (2) This follows from proposition \ref{pro:periodic-structure2}. (3) When $X$ is proper, so is $\perfl{X}{{\mathfrak f}}$, since it is a closed subscheme of $X$. It therefore must be finite over $\scalA$ because it is flat and generically finite. \end{proof} \subsection{} \emph{Conjecture.} \label{conj:points-exist} Let ${\mathfrak r}$ denote the product of all the real places. If $X$ is proper and nonempty, we conjecture that $\per{X}{{\mathfrak r}}$ is nonempty. When $K={\mathbb Q}$, this was proved in~\cite{Borger:LRFOE}. It is possible, however, to give an easier argument that avoids the deep theorems in \'etale cohomology and $p$-adic Hodge theory used there. One would expect this argument to go through for general $K$. \subsection{} \emph{Computing the periodic locus.} Is there an algorithm to find equations describing $X({\mathfrak f})$, given equations for $X$ and formulas for the Frobenius lifts $\psi_{\mathfrak p}$? Without such an algorithm, our approach to generating abelian extensions would not be so explicit. But it also might be an indication that, for theoretical purposes, there really is more freedom in this approach. \section{Interlude on periodic Witt vectors} In this section, we define periodic Witt vectors and show how they recover the ray class algebras of the previous section. It will not be used elsewhere in the paper. \subsection{} \emph{${P}$-typical Witt vector rings $W_{\scalA,{P}}({R})$.} Let us review the generalized Witt vector rings as defined in the first section of \cite{Borger:BGWV-I}. Let ${R}$ be a flat $\scalA$-algebra. Then the monoid $\mathrm{Id}_{{P}}$ acts on the product $\scalA$-algebra ${R}^{\mathrm{Id}_{{P}}}$ (the so-called ghost ring) by translation in the exponent. Explicitly, if $x_{{\mathfrak a}}$ denotes the ${\mathfrak a}$-th component of a vector $x\in {R}^{\mathrm{Id}_{{P}}}$, then $\psi_{{\mathfrak b}}\colon {R}^{\mathrm{Id}_{{P}}}\to {R}^{\mathrm{Id}_{{P}}}$ is defined by the formula $$ \big(\psi_{{\mathfrak b}}(x)\big)_{{\mathfrak a}} = x_{{\mathfrak a}{\mathfrak b}}. $$ Now consider the set of sub-$\scalA$-algebras $\ringd\subseteq {R}^{\mathrm{Id}_{{P}}}$ such that $\ringd$ is taken to itself by the action of $\mathrm{Id}_{{P}}$ and such that for each prime ${\mathfrak p}\in{P}$, the induced endomorphism $\psi_{{\mathfrak p}}\colon \ringd\to \ringd$ is a Frobenius lift at ${\mathfrak p}$. An elementary argument shows that this collection of subrings has a maximal element $W_{\scalA,{P}}({R})$. It is called the ring of ${P}$-typical Witt vectors with entries in ${R}$. It recovers the usual $p$-typical Witt vector functor (restricted to torsion-free rings) when $\scalA$ is ${\mathbb Z}$ and ${P}$ consists of the single maximal ideal $p{\mathbb Z}$; it recovers the big Witt vector functor when instead ${P}$ consists of all maximal ideals of ${\mathbb Z}$. This construction is functorial in ${R}$, and one can show that the functor $W_{\scalA,{P}}$ is representable by a flat $\scalA$-algebra $\Lambda_{\scalA,{P}}$. (Incidentally, this shows that $W_{\scalA,{P}}$ extends to a functor on all $\scalA$-algebras, namely the one represented by $\Lambda_{\scalA,{P}}$. Thus we can extend the theory of $\Lambda$-structures and Witt vectors to $\scalA$-algebras with torsion, but we will not need this generality here.) The endomorphisms $\psi_{{\mathfrak a}}$ of the functor $W_{\scalA,{P}}$ induce endomorphisms of $\Lambda_{\scalA,{P}}$. They are in fact Frobenius lifts, and hence $\Lambda_{\scalA,{P}}$ is a $\Lambda_{\scalA,{P}}$-ring. \subsection{} \emph{Universal property of Witt vector rings.} \label{subsec:witt-universal-property} The Witt vector functor is the right adjoint of the forgetful functor from $\Lambda_{\scalA,{P}}$-rings to $\scalA$-algebras. Let us spell out the universal property for future reference. Let ${R}$ be a flat $\scalA$-algebra, let $\ringd$ be a flat $\Lambda_{\scalA,{P}}$-ring, and let $\varphi\colon\ringd\to {R}$ be an $\scalA$-algebra map. Then $\varphi$ lifts to a unique $\Lambda_{\scalA,{P}}$-ring map $\tilde{\varphi}$ to the Witt vector ring: $$ \xymatrix{ \ringd\ar_{\varphi}[dr]\ar@{-->}_-{\exists!}^-{\tilde{\varphi}}[rr] & & W_{\scalA,{P}}({R}) \ar^{x\mapsto x_{(1)}}[dl] \\ & {R}. } $$ Indeed, it lifts to a unique $\mathrm{Id}_{P}$-equivariant morphism $\ringd\to {R}^{\mathrm{Id}_{P}}$, given by $a\mapsto x$, where $x_{\mathfrak a}=\varphi(\psi_{\mathfrak a}(a))$. It remains to check that the image $S$ of this map lies in $W_{\scalA,{P}}({R})$. But $S$ is torsion free, as a subalgebra of ${R}^{\mathrm{Id}_{P}}$; it has an action of $\mathrm{Id}_{P}$, as the image of an equivariant map of rings with an $\mathrm{Id}_{P}$-action; and it satisfies the Frobenius lift condition because it is an $\mathrm{Id}_{P}$-equivariant quotient of $\ringd$, which satisfies the Frobenius lift condition. Therefore $S$ is contained in $W_{\scalA,{P}}({R})$ by the maximality property of $W_{\scalA,{P}}({R})$. \subsection{} \emph{${\mathfrak f}$-periodic Witt vector rings $\peru{W}{{\mathfrak f}}({R})$.} Let ${\mathfrak f}$ be a cycle on $K$. We define the set of ${\mathfrak f}$-periodic Witt vectors with entries in an $\scalA$-algebra ${R}$ as follows \begin{equation} \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) := \{x\in W_{\scalA,{P}}({R}): \psi_{{\mathfrak a}}(x)=\psi_{{\mathfrak b}}(x) \text{ whenever } {\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}\}. \end{equation} In other words, if we view the functor $W_{\scalA,{P}}$ as a scheme, then $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({R})$ is the set of ${R}$-valued points on its ${\mathfrak f}$-periodic locus. When ${R}$ is flat, the periodic Witt vectors can be described simply in terms of their ghost components: \begin{equation} \label{def:periodic-witt} \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) = \{x\in W_{\scalA,{P}}({R}): x_{{\mathfrak a}}=x_{{\mathfrak b}} \text{ whenever } {\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}\}. \end{equation} Indeed, this follows from the implication ${\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b} \Rightarrow {\mathfrak a}{\mathfrak c}\sim_{{\mathfrak f}}{\mathfrak b}{\mathfrak c}$. In other words, we have \begin{equation} \label{eq:periodic-witt-intersection} \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) = W_{\scalA,{P}}({R}) \cap {R}^{{\mathrm{DR}}_{{P}}({\mathfrak f})}, \end{equation} as subrings of the ghost ring ${R}^{\mathrm{Id}_{{P}}}$. \subsection{} \emph{Example.} Suppose $\scalA={\mathbb Z}$, ${P}=M_{\mathbb Q}$, and ${\mathfrak f}=(n)\infty$ where $n\geq 1$. Then an ${\mathfrak f}$-periodic Witt vector with entries in a torsion-free ring ${R}$ is just a big Witt vector whose ghost components are periodic with period dividing $n$. This is the reason for the name. For example, if $\zeta_n$ is an $n$-th root of unity, then the Teichm\"uller element $$ [\zeta_n] := \langle \zeta_n,\zeta_n^2,\zeta_n^3,\dots \rangle $$ is $n\infty$-periodic. (This is even true when ${R}$ is not torsion free, by functoriality and because the universal ring with an $n$-th root of unity is ${\mathbb Z}[x]/(x^n-1)$, which is torsion free.) \begin{proposition} \label{pro:w-per-is-sub-lambda-ring} $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({R})$ is an ${\mathfrak f}$-periodic sub-$\Lambda_{\scalA,{P}}$-ring of $W_{\scalA,{P}}({R})$, for any flat $\scalA$-algebra ${R}$. \end{proposition} \begin{proof} It is clearly a sub-$\scalA$-algebra of $W_{\scalA,{P}}({R})$. It is also preserved by all $\psi_{{\mathfrak p}}$ operators (${\mathfrak p}\in{P}$) because of the implication ${\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b} \Rightarrow {\mathfrak p}{\mathfrak a}\sim_{{\mathfrak f}} {\mathfrak p}{\mathfrak b}$. The family of operators $\psi_{\mathfrak a}$ is also ${\mathfrak f}$-periodic on $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({R})$ by definition. So all that remains is to check the Frobenius lift property. For ${\mathfrak p}\in{P}$ and any $x\in \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R})$, we have $$ \psi_{{\mathfrak p}}(x)-x^{N({\mathfrak p})}\in {\mathfrak p} W_{\scalA,{P}}({R}) \cap \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}). $$ Thus it is enough to show that the containment \begin{equation} \label{eq:per-Frob-lift} {\mathfrak p}\peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) \subseteq {\mathfrak p} W_{\scalA,{P}}({R}) \cap \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) \end{equation} is an equality. This follows from the diagram $$ \xymatrix{ 0 \ar[r] & \peru{W}{{\mathfrak f}}({R}) \ar[r] & W({R}) \ar^-{\gamma}[r] & \prod_{{\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}} {R} \\ 0 \ar[r] & {\mathfrak p}\otimes_{\scalA}\peru{W}{{\mathfrak f}}({R}) \ar[r]\ar@{>->}[u] & {\mathfrak p}\otimes_{\scalA}W({R}) \ar^-{{\mathfrak p}\otimes\gamma}[r]\ar@{>->}[u] & {\mathfrak p}\otimes_{\scalA}\prod_{{\mathfrak a}\sim_{{\mathfrak f}}{\mathfrak b}} {R}\ar@{>->}[u] \\ } $$ of exact sequences, where $\gamma(x)=(\dots,x_{{\mathfrak a}}-x_{{\mathfrak b}},\dots)$. The vertical arrows are the evident multiplication maps $a\otimes b\mapsto ab$; they are injective because all modules on the top row are flat. This plus the exactness of the bottom row implies that (\ref{eq:per-Frob-lift}) is an equality. \end{proof} \subsection{} \emph{Remark: Plethysic algebra.} The formal concepts above can be expressed in the language of plethystic algebra \cite{Borger-Wieland:PA}. Let $\Lambda_{\scalA,{P}}$ be the $A$-algebra representing the functor $W_{\scalA,{P}}$, and let $\peru{\Lambda}{{\mathfrak f}}_{\scalA,{P}}$ be the one representing the functor $\peru{W}{{\mathfrak f}}_{\scalA,{P}}$. Then the inclusion of functors $\peru{W}{{\mathfrak f}}_{\scalA,{P}} \subseteq W_{\scalA,{P}}$ induces a surjection $\Lambda_{\scalA,{P}}\to \peru{\Lambda}{{\mathfrak f}}_{\scalA,{P}}$. It is not hard to show that this has the structure of a morphism of $\scalA$-plethories and an action of $\peru{\Lambda}{{\mathfrak f}}_{\scalA,{P}}$ is the same as an ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-structure. It follows that the forgetful functor from ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-rings to all $\Lambda_{\scalA,{P}}$-rings has both a left and a right adjoint. The right adjoint outputs the ${\mathfrak f}$-periodic elements of the given $\Lambda_{\scalA,{P}}$-ring. This provides one approach to $\Lambda$-structures and Witt vectors that works smoothly in the presence of torsion. \subsection{} \emph{Universal property of periodic Witt vector rings.} \label{subsec:periodic-witt-univ-prop} Let ${R}$ be a flat $\scalA$-algebra, let $\ringd$ be a flat ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-ring, and let $\varphi\colon \ringd\to {R}$ be an $\scalA$-algebra map. Then $\varphi$ lifts to a unique $\Lambda_{\scalA,{P}}$-ring map to the ${\mathfrak f}$-periodic Witt vector ring: $$ \xymatrix{ \ringd\ar_{\varphi}[dr]\ar@{-->}_-{\exists!}^-{\tilde{\varphi}}[rr] & & \peru{W}{{\mathfrak f}}_{\scalA,{P}}({R}) \ar^{x\mapsto x_{(1)}}[dl] \\ & {R}. } $$ Indeed, by the universal property of Witt vectors (\ref{subsec:witt-universal-property}), it lifts to a unique $\Lambda_{\scalA,{P}}$-map $\ringd\to W_{\scalA,{P}}({R})$. But since $\ringd$ is ${\mathfrak f}$-periodic, the image is contained in $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({R})$. \begin{proposition} Let ${K^{\mathrm{sep}}}$ be a separable closure of $K$, and let ${\scalA}^{\mathrm{int}}$ denote the integral closure of $\scalA$ in ${K^{\mathrm{sep}}}$. Let $\rcl{\scalA,{P}}{{\mathfrak f}}$ denote the ray class $\Lambda_{\scalA,{P}}$-ring of conductor ${\mathfrak f}$, as defined in \ref{subsec:ray-class-lambda-rings}. Then we have an isomorphism \begin{equation} \tilde{\beta}\:\rcl{\scalA,{P}}{{\mathfrak f}} \,{\buildrel \sim\over\longrightarrow}\, \peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}}) \end{equation} of $\Lambda_{\scalA,{P}}$-rings, where the map $\tilde{\beta}$ is the lift, in the sense of \ref{subsec:periodic-witt-univ-prop}, of the projection $\beta\:\rcl{\scalA,{P}}{{\mathfrak f}}\to{\scalA}^{\mathrm{int}}$ defined in \ref{subsec:ray-class-lambda-rings}. \end{proposition} \begin{proof} The $\Lambda_{\scalA,{P}}$-rings $\rcl{\scalA,{P}}{{\mathfrak f}}$ and $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}})$ are characterized by the same universal property, except that the one for $\rcl{\scalA,{P}}{{\mathfrak f}}$ is restricted to algebras $\ringd$ that are reduced and finite flat over $\scalA$. So it is enough to show that $\peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}})$ is itself reduced and finite flat over $\scalA$. Since it is a subalgebra of $\prod_{{\mathrm{DR}}_{{P}}({\mathfrak f})}{\scalA}^{\mathrm{int}}$ by definition, it is reduced and flat. So it is enough to show that it is finite over $\scalA$. First, observe that we have \begin{equation} \label{eq:24} \colim_L \peru{W}{{\mathfrak f}}_{\scalA,{P}}(\tO{L}) = \peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}}), \end{equation} where $L$ runs over the finite extensions of $K$ contained in ${K^{\mathrm{sep}}}$. Indeed, for any $x\in\peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}})$, let $L$ denote the extension $K(\dots,x_{\mathfrak a},\dots)$, where the $x_{\mathfrak a}$ are the (ghost) components of $x$. It follows that $x\in W_{\scalA,{P}}(L)\cap W_{\scalA,{P}}({\scalA}^{\mathrm{int}})$. Because $W_{\scalA,{P}}$ is representable, we also have $$ W_{\scalA,{P}}(L)\cap W_{\scalA,{P}}({\scalA}^{\mathrm{int}}) = W_{\scalA,{P}}(L\cap{\scalA}^{\mathrm{int}}), $$ and hence $x\in W_{\scalA,{P}}(L\cap{\scalA}^{\mathrm{int}})$. Because $x$ is also ${\mathfrak f}$-periodic, we have $$ x\in\peru{W}{{\mathfrak f}}_{\scalA,{P}}(\tO{L}), $$ as desired. It remains to show that $L$ is a finite extension of $K$. This holds because $x$ is ${\mathfrak f}$-periodic and ${\mathrm{DR}}_{{P}}({\mathfrak f})$ is finite, and so $x$ has only finitely many distinct ghost components. This proves~(\ref{eq:24}). Therefore it is enough to prove that $\peru{W}{{\mathfrak f}}_{\scalA,{P}}(\tO{L})$ has bounded rank as $L$ runs over all finite extensions of $K$. By proposition \ref{pro:w-per-is-sub-lambda-ring}, $\peru{W}{{\mathfrak f}}_{\scalA,{P}}(\tO{L})$ is a flat reduced ${\mathfrak f}$-periodic $\Lambda_{\scalA,{P}}$-ring. Being a subring of $\prod_{{\mathrm{DR}}_{{P}}({\mathfrak f})}\tO{L}$, it is also finitely generated as an $\scalA$-module. Therefore by theorem \ref{global-thm2}, it is contained in the product ring $\prod_{{\mathfrak d}} \tO{K({\mathfrak f}{\mathfrak d}^{-1})}$, where ${\mathfrak d}$ runs over $\mathrm{Id}_{{P}}$ with ${\mathfrak d}\mid{\mathfrak f}_{{\mathrm{fin}}}$. Therefore its rank as $L$ varies is bounded. \end{proof} \subsection{} \emph{Example.} \label{subsec:cyclotomic-ray-class-algebra} If $\scalA={\mathbb Z}$, $P=M_{\mathbb Q}$, and ${\mathfrak f}=(n)\infty$ with $n\geq 1$, then for any primitive $n$-th root of unity $\zeta_n$, there is an isomorphism $$ {\mathbb Z}[x]/(x^n-1) \,{\buildrel \sim\over\longrightarrow}\, \peru{W}{n\infty}_{{\mathbb Z},{P}}(\maxO_{\bar{{\mathbb Q}}}) $$ given by $x\mapsto [\zeta_n]$. Given the above, this is the second theorem of our first paper~\cite{Borger-deSmit:Integral-models}. On the other hand, if ${\mathfrak f}=(1)$ but $K$ is arbitrary and $P=M_K$, then we have $$ \peru{W}{1}_{\maxO_K,{P}}(\maxO_{\bar{K}}) = \maxO_H $$ where $H$ is the Hilbert class field of $K$. More generally, if ${\mathfrak f}$ is supported at infinity, then $\peru{W}{{\mathfrak f}}_{\maxO_K,{P}}(\maxO_{\bar{K}})$ is the ring of integers in the corresponding narrow Hilbert class field. \begin{corollary} \label{cor:periodic-W-class-fields} The periodic Witt vector ring is generically a product of ray class fields: \begin{equation} K\otimes_{\scalA} \peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}}) = \prod_{\substack{{\mathfrak d}\in\mathrm{Id}_{{P}}\\{\mathfrak d}\mid{\mathfrak f}_{{\mathrm{fin}}}}}K({\mathfrak f}{\mathfrak d}^{-1}). \end{equation} The ray class field $K({\mathfrak f})$ is the image of the projection \begin{equation} K\otimes_{\scalA} \peru{W}{{\mathfrak f}}_{\scalA,{P}}({\scalA}^{\mathrm{int}}) {\,\longrightarrow\,} {K^{\mathrm{sep}}}, \quad x\mapsto x_{(1)}. \end{equation} \end{corollary} \subsection{} \emph{Remark: Periodic Witt vectors and explicit class field theory.} \label{subsec:periodic-W-and-explicit-CFT} It follows from corollary~\ref{cor:periodic-W-class-fields} that any ray class field $K({\mathfrak f})$ is generated by the first coordinate $x_{(1)}$ of the ${\mathfrak f}$-periodic ${\scalA}^{\mathrm{int}}$-points $x$ on the affine scheme $W_{\scalA,{P}}=\Spec(\Lambda_{\scalA,{P}})$. We emphasize however that $W_{\scalA,{P}}$ is not of finite type, and so this does not give an explicit method of producing a polynomial whose roots generate $K({\mathfrak f})$. In fact, the periodic locus $\peru{W}{{\mathfrak f}}_{\scalA,{P}}$ is itself not even of finite type. For instance, if $K={\mathbb Q}$, $P=M_{\mathbb Q}$, and ${\mathfrak f}=\infty$, then $\Lambda^{({\mathfrak f})}_{\scalA,{P}}$ is isomorphic to the binomial ring, the subring of ${\mathbb Q}[x]$ generated by the binomial coefficients $\binom{x}{n}$, and this is not finitely generated as a ring. Therefore to find a point of $W^{({\mathfrak f})}_{\scalA,{P}}$, one has to solve infinitely many simultaneous polynomial equations with coefficients in $\scalA$. To be sure, $K\otimes_\scalA \Lambda^{({\mathfrak f})}_{\scalA,{P}}$ is finitely generated as a $K$-algebra, because on $K$-algebras it represents the periodic ghost functor ${R}\mapsto {R}^{{\mathrm{DR}}_{P}({\mathfrak f})}$ and ${\mathrm{DR}}_{P}({\mathfrak f})$ is finite. Therefore a periodic Witt vector is determined by finitely many components, namely its ghost components. However to give a criterion for a periodic ghost vector to be a Witt vector, we need infinitely many congruences between polynomials in the ghost components. If we add variables to express these congruences as equations, we will need infinitely many new variables. \section{$K={\mathbb Q}$: the toric line and the Chebyshev line} In this section, we consider the case where $\scalA$ is ${\mathbb Z}$ and ${P}$ is the set of all maximal ideals of ${\mathbb Z}$. For any integer $n\geq 1$, let us write $\psi_n=\psi_{(n)}$ and $\Lambda=\Lambda_{{\mathbb Z},{P}}$. \subsection{} \emph{The toric line and cyclotomic extensions.} \label{subsec:toric-line} Define the toric $\Lambda$-structure on ${\mathbb G}_{\mathrm{m}}=\Spec{\mathbb Z}[x^{\pm 1}]$ to be the one given by $\psi_{p}(x)=x^p$, for all primes numbers $p$. (We use this name because it is a particular case of the natural $\Lambda$-structure on any toric variety.) Observe that it extends uniquely to the affine and projective lines. For each cycle of the form $(n)\infty$, with $n\geq 1$, the periodic locus $\per{{\mathbb G}_{\mathrm{m}}}{n\infty}$ is simply $\mu_n=\Spec{\mathbb Z}[x]/(x^n-1)$. In other words, the containment (\ref{eq:periodic-in-torsion}) is an equality. Thus we have $$ {\mathbb Q}\otimes_{{\mathbb Z}}\per{{\mathbb G}_{\mathrm{m}}}{n\infty}=\coprod_{d|n} \Spec {\mathbb Q}(\zeta_d). $$ So in this case, the $n\infty$-periodic locus of the $\Lambda$-scheme ${\mathbb G}_{\mathrm{m}}$ does in fact generate the ray class field of conductor $n\infty$. \subsection{} \emph{The Chebyshev line and real-cyclotomic extensions.} \label{subsec:chebyshev} The toric $\Lambda$-ring ${\mathbb Z}[x^{\pm 1}]$ above has an automorphism $\sigma$ defined by $\sigma(x)=x^{-1}$. The fixed subring is easily seen to be a sub-$\Lambda$-ring and freely generated as a ring by $y=x+x^{-1}$. The elements $\psi_{p}(y)\in{\mathbb Z}[y]$ are given by Chebyshev polynomials $$ \psi_n(y)=2+\prod_{i=0}^{n-1} (y-\zeta_n^i-\zeta_n^{-i}), $$ where $\zeta_n$ is a primitive $n$-th root of unity. For example, we have $$ \psi_{2}(y) = y^2-2, \quad\quad \psi_{3}(y)=y^3-3y, \quad\quad \psi_{5}(y)=y^5-5y^3+5y, \quad\quad\dots. $$ This gives the affine line $Y=\Spec {\mathbb Z}[y]$ a $\Lambda$-scheme structure. We call it the Chebyshev $\Lambda$-structure. It also extends uniquely to the projective line. (Incidentally, ${\mathbb Z}[y]$ is isomorphic as a $\Lambda$-ring to the Grothendieck group of the Lie algebra $\mathfrak{sl}_2$. By Clauwens's theorem~\cite{Clauwens:Line}, this and the toric $\Lambda$-structure are the only two $\Lambda$-structures on the affine line, up to isomorphism.) Now consider a cycle ${\mathfrak f}$ with trivial infinite part. Write ${\mathfrak f}=(n)$, where $n\geq 1$, and write $Y(n)=Y({\mathfrak f})$ for the periodic locus. We have $$ Y[1] = Y(1) = \Spec {\mathbb Z}[y]/(y-2), $$ and hence the $n$-torsion locus is $$ Y[n] = \psi_n^{-1} (\Spec {\mathbb Z}[y]/(y-2)) $$ or more simply, $Y[n]=\psi_n^{-1}(2)$. Observe that $Y[n]$ is not reduced when $n\geq 3$: for instance if $n$ is odd, we have $$ \psi_n(y)-2 = \prod_{i=0}^{n-1} (y-\zeta_n^i-\zeta_n^{-i}) = (y-2)\prod_{i=1}^{\frac{n-1}{2}} (y-\zeta_n^i-\zeta_n^{-i})^2. $$ However the periodic locus is reduced. In fact, we have $$ {\mathbb Q}\otimes_{\mathbb Z} \sheafO(Y(n)) = \prod_{d|n} {\mathbb Q}(\zeta_d+\zeta_d^{-1}), $$ by the following more precise integral result: \begin{proposition}\label{pro:chebyshev-periodic-locus2} \begin{enumerate} \item The periodic locus $\per{Y}{n}$ is flat and reduced, and the inclusion $$ i:\per{Y}{n}\to Y[n]_{\mathrm{red}} $$ is an isomorphism. \item The map $\sheafO(\per{Y}{n})\to {\mathbb Z}[x]/(x^n-1)$ is injective. If $n$ is odd, its image is the subring of invariants under the involution $\sigma:x\mapsto x^{-1}$. If $n$ is even, its image is the span of $\{1,x+x^{-1},\dots,x^{n/2-1}+x^{1-n/2},2x^{n/2}\}$. \end{enumerate} \end{proposition} \begin{proof} (1): Consider the differences of Chebyshev polynomials $$ P_{a,b}(y) := \psi_a(y)-\psi_b(y) \in {\mathbb Z}[y]. $$ Then the periodic locus $\per{Y}{n}$ is $\Spec{\mathbb Z}[y]/I$, where $$ I= \big(P_{a,b}(y) \,\mid\, a\equiv \pm b\bmod n,\ a,b\geq 1\big). $$ Let $Q(y)$ denote the product of the monic irreducible factors of $(\psi_n(y)-2)$, each taken only with multiplicity $1$. If $n$ is odd, we have $$ Q(y) = \prod_{i=0}^{\frac{n-1}{2}}(y-\zeta_n^i-\zeta_n^{-i}). $$ and if $n$ is even, we have $$ Q(y) = \prod_{i=0}^{\frac{n}{2}}(y-\zeta_n^i-\zeta_n^{-i}). $$ Then we have $$ I\subseteq (Q(y)) $$ since each $P_{a,b}(y)$ vanishes at $y=\zeta_n^{i}+\zeta_n^{-i}$, for each $i$. Therefore to prove (1), it is enough to show this containment is an equality. Since ${\mathbb Z}[x^{\pm 1}]$ is a faithfully flat extension of ${\mathbb Z}[y]$, it is enough to show this after base change to ${\mathbb Z}[x^{\pm 1}]$. In other words, it is enough to show $$ Q(x+x^{-1}) \in \big(P_{a,b}(x+x^{-1}) \,\mid\, a\equiv\pm b\mod n\big). $$ We have $$ P_{a,b}(x+x^{-1}) = x^a+x^{-a}-x^b-x^{-b} = x^{-a}(x^{a+b}-1)(x^{a-b}-1). $$ Therefore we have \begin{align*} \big(P_{n+1,1}(x+x^{-1})\big) &=\big((x^n-1)(x^{n+2}-1)\big), \\ \big(P_{n+2,2}(x+x^{-1})\big) &=\big((x^n-1)(x^{n+4}-1)\big) \end{align*} and hence $$ \big(P_{n+1,1}(x+x^{-1}),P_{n+2,2}(x+x^{-1})\big) = \big(x^n-1\big)\big((x^{2}-1),(x^{n}-1)\big). $$ If $n$ is odd, then we have $$ \big(x^n-1\big)\big((x^2-1),(x^n-1)\big) = \big((x^n-1)(x-1)\big) = \big(Q(x+x^{-1})\big). $$ Similarly, if $n$ is even $$ \big(x^n-1\big)\big((x^2-1),(x^n-1)\big) = \big((x^n-1)(x^2-1)\big) = \big(Q(x+x^{-1})\big). $$ Thus in either case, we have \begin{align*} Q(x+x^{-1}) &\in \big(P_{n+1,1}(x+x^{-1}),P_{n+2,2}(x+x^{-1})\big) \\ &\subseteq \big(P_{a,b}(y) \,\mid\, a\equiv \pm b\bmod n,\ a,b\geq 1\big). \end{align*} (2): Suppose a polynomial $f(y)\in{\mathbb Z}[y]$ maps to zero in ${\mathbb Z}[x]/(x^n-1)$. Then we have $f(\zeta_n^i+\zeta_n^{-i})=0$ for all $i$. Therefore we have $(y-\zeta_n^i-\zeta_n^{-i})\mid f(y)$. Thus if $n$ is odd, we have $$ \prod_{i=0}^{\frac{n-1}{2}}(y-\zeta_n^i-\zeta_n^{-1})\mid f(y) $$ and if $n$ is even, $$ \prod_{i=0}^{\frac{n}{2}}(y-\zeta_n^i-\zeta_n^{-1})\mid f(y). $$ In either case, we have $Q(y)\mid f(y)$. Therefore the map ${\mathbb Z}[y]/(Q(y))\to {\mathbb Z}[x]/(x^n-1)$ is injective. Let us now consider surjectivity. When $n$ is odd, the set $$ \{1,x+x^{-1},\dots,x^{\frac{n-1}{2}}+x^{-\frac{n-1}{2}}\} $$ is a ${\mathbb Z}$-basis for the subring $\big({\mathbb Z}[x]/(x^n-1)\big)^\sigma$ of $\sigma$-invariants, and this is contained in the image of ${\mathbb Z}[y]$. When $n$ is even, $\{1,x+x^{-1},\dots,x^{n/2-1}+x^{1-n/2},x^{n/2}\}$ is a ${\mathbb Z}$-basis. All these elements but $x^{n/2}$ lie in the image of ${\mathbb Z}[y]$. However $2x^{n/2}$ does lie in the image. \end{proof} \begin{corollary} \label{cor:chebyshev-periodic-ray-class-algebra} The map $$ \sheafO(\per{Y}{n}) {\,\longrightarrow\,} \rcl{{\mathbb Z},{P}}{n} $$ to the ray class algebra given by the point $\zeta_n+\zeta_n^{-1}\in\per{Y}{n}(\maxO_{\bar{{\mathbb Q}}})$ is injective. If $n$ is odd, it is surjective; if $n$ is even, its cokernel is a group of order $2$ generated by the class of $[-1]=\langle \dots,(-1)^{d},\dots \rangle\in \prod_{d\mid n}{\mathbb Q}(\zeta_{n/d}+\zeta_{n/d}^{-1})$. \end{corollary} \begin{proof} By the map ${\mathbb Z}[x]/(x^n-1)\to \maxO_{\bar{{\mathbb Q}}}$ given by $x\mapsto \zeta_n$ results in a diagram $$ \xymatrix{ {\mathbb Z}[x]/(x^n-1)\ar[r] & \rcl{{\mathbb Z},{P}}{n\infty} \\ \big({\mathbb Z}[x]/(x^n-1)\big)^{\sigma}\ar[r]\ar@{>->}[u] & \rcl{{\mathbb Z},{P}}{n}. \ar@{>->}[u] } $$ By the second theorem in our first paper~\cite{Borger-deSmit:Integral-models}, the top map is an isomorphism. Since taking maximal $\Lambda$-orders commutes with taking group invariants, by part (2) of proposition~\ref{pro:maximality-facts}, the bottom arrow is also an isomorphism. Observe that when $n$ is even, the image of $x^{n/2}$ is $[-1]$. Now invoke proposition~\ref{pro:chebyshev-periodic-locus2}. \end{proof} \section{$K$ imaginary quadratic: CM elliptic curves and the Latt\`es scheme} Let $K$ be an imaginary quadratic field. For convenience, let us fix an embedding $K\subset{\mathbb C}$. In this section, we show how explicit class field theory over $K$, due to Kronecker and his followers, can be set naturally in the framework of this paper. This builds on Gurney's thesis~\cite{Gurney:thesis}. The arguments are similar in spirit to those in the real-cyclotomic context in the previous section. Let us write $$ \Lambda=\Lambda_{\maxO_K,M_K} \quad\text{and}\quad \Lambda_{\mathfrak p}=\Lambda_{\maxO_{K_{\mathfrak p}},{\mathfrak p}}, $$ where ${\mathfrak p}\in{P}$ is any given prime. \subsection{} \emph{The moduli space of CM elliptic curves.} Let $R$ be an $\maxO_K$-algebra. Then a \emph{CM elliptic curve} over $R$ is an elliptic curve $E$ over $R$ together with a ring map $\maxO_K\to{\mathrm{End}}_R(E)$ such that the induced action of $\maxO_K$ on the tangent space $T_0(E)$ agrees with the $\maxO_K$-algebra structure map $\maxO_K\to R={\mathrm{End}}_R(T_0(E))$. Let $\mathcal{M}_{\mathrm{CM}}(R)$ denote the category whose objects are the CM elliptic curves over $R$ and whose morphisms are the $\maxO_K$-equivariant isomorphisms of elliptic curves. Since all morphisms are isomorphisms, $\mathcal{M}_{\mathrm{CM}}(R)$ is a groupoid by definition. As the $\maxO_K$-algebra $R$ varies, the usual base-change maps make $\mathcal{M}_{\mathrm{CM}}$ a fibered category over the category of affine $\maxO_K$-schemes. Further, this fibered category satisfies effective descent for the fppf topology because all ingredients in its definition can be expressed in terms which are fppf-local. In other words, $\mathcal{M}_{\mathrm{CM}}$ is a stack. (See~\cite{Laumon-MB:Champs-algebriques} the theory of stacks.) Let $\mathcal{E}\to\mathcal{M}_{\mathrm{CM}}$ denote the universal object. Then $\mathcal{E}(R)$ is the groupoid of pairs $(E,x)$, where $E$ is a CM elliptic curve over $R$, and $x$ is a point of $E(R)$. Both $\mathcal{E}$ and $\mathcal{M}_{\mathrm{CM}}$ have certain endomorphisms $\psi_{\mathfrak a}$, for any integral ideal ${\mathfrak a}\subseteq \maxO_K$, defined as follows. On $\mathcal{M}_{\mathrm{CM}}$, the map $\psi_{\mathfrak a}$ is defined by $$ \psi_{\mathfrak a}: E\mapsto {\mathfrak a}^{-1}\otimes E, $$ where the elliptic curve ${\mathfrak a}^{-1}\otimes E$ is the Serre tensor product~\cite{Serre:complex-mult}, defined as a functor by $$ \big({\mathfrak a}^{-1}\otimes E\big)(C):={\mathfrak a}^{-1}\otimes_{\maxO_K}\big(E(C)\big), $$ for any algebra $C$ over the base $R$ of $E$. (This is again an elliptic curve: since ${\mathfrak a}^{-1}$ is a direct summand of $\maxO_K^2$, the functor ${\mathfrak a}^{-1}\otimes E$ is a direct summand of the abelian variety $E^2$.) In particular, if $R={\mathbb C}$ and the period lattice of $E$ is $L$, so that $E({\mathbb C})={\mathbb C}/L$, then ${\mathfrak a}^{-1}\otimes E$ is the elliptic curve with period lattice ${\mathfrak a}^{-1}\otimes_{\maxO_K}L$, up to canonical isomorphism. Using the isogeny $E\to {\mathfrak a}^{-1}\otimes E$ defined by $x\mapsto 1\otimes x$, we can then define the endomorphism $\psi_{\mathfrak a}:\mathcal{E}\to\mathcal{E}$ by $$ \psi_{\mathfrak a}: (E,x) \mapsto ({\mathfrak a}^{-1}\otimes E,1\otimes x). $$ Observe that it lies over the endomorphism $\psi_{\mathfrak a}$ of $\mathcal{M}_{\mathrm{CM}}$. The $\psi_{\mathfrak a}$ operators should be thought of as providing a $\Lambda$-structure on $\mathcal{M}_{\mathrm{CM}}$ and $\mathcal{E}$, but $\mathcal{M}_{\mathrm{CM}}$ and $\mathcal{E}$ are stacks and we will not define what $\Lambda$-structures on stacks are here. Let us just say that the $\psi_{\mathfrak a}$ operators, whether on $\mathcal{M}_{\mathrm{CM}}$ or $\mathcal{E}$, commute up to a coherent family of canonical isomorphisms coming from the usual associators $(U\otimes V)\otimes W\to U\otimes (V\otimes W)$ on the category of $\maxO_K$-modules. Similarly, if ${\mathfrak a}$ is a prime ${\mathfrak p}$, then the reduction of $\psi_{\mathfrak p}$ modulo ${\mathfrak p}$ agrees with the $N({\mathfrak p})$-power Frobenius map $F_{\mathfrak p}$, in the sense that for any elliptic curve $E$ over an ${\mathbb F}_p$-algebra, there is unique isomorphism $\psi_{\mathfrak p}(E)\to F_{\mathfrak p}(E)$ compatible with the canonical maps from $E$. \subsection{} \emph{Latt\`es scheme.} \label{subsec:lattes} We follow Gurney's thesis~\cite{Gurney:thesis}. Let $\mathcal{L}$ denote the coarse space underlying $\mathcal{E}$. It is the functor defined, for any $\maxO_K$-algebra $C$, by $$ \mathcal{L}(C) = \{\text{local-isomorphism classes of triples $(C',E,x)$}\}, $$ where $C'$ is an fppf cover of $C$, and $E$ is a CM elliptic curve over $C'$, and $x\in E(C')$ and where two pairs $(C'_1,E_1,x_1)$ and $(C'_2,E_2,x_2)$ are in the same local-isomorphism class if $C'_1$ and $C'_2$ have a common fppf cover $C''$ such that when pulled back to $C''$, there is an isomorphism between $E_1$ and $E_2$ identifying $x_1$ and $x_2$. Observe that if $C={\mathbb C}$ (or any algebraically closed field), we have $$ \mathcal{L}({\mathbb C}) = \{\text{isomorphism classes of pairs $(E,x)$}\}, $$ where $E$ is a CM elliptic curve over ${\mathbb C}$ and $x\in E({\mathbb C})$. Thus $\mathcal{L}({\mathbb C})$ is the union of $h_K$ copies of ${\mathbb P}^1({\mathbb C})$, where $h_K$ is the class number of $K$. The functor $\mathcal{L}$ is an $\maxO_K$-scheme (see 4.3.10 of~\cite{Gurney:thesis}), which we call the \emph{Latt\`es scheme}, and unless we say otherwise we will view it as an $\maxO_K$-scheme. Note however that the structure map $\mathcal{L}\to\Spec\maxO_K$ factors naturally through the coarse space underlying $\mathcal{M}_{\mathrm{CM}}$, which Gurney shows (2.6.10 of~\cite{Gurney:thesis}) is $\Spec \maxO_H$, where $H$ is the Hilbert class field of $K$ in ${\mathbb C}$: $$ \xymatrix{ \mathcal{L} \ar@{-->}[rr]\ar[dr] & & \Spec \maxO_H \ar[dl] \\ & \Spec \maxO_K } $$ So one can also view $\mathcal{L}$ as an $\maxO_H$-scheme, and sometimes we will. For example, given any CM elliptic curve $E$ over $R$, we have an identification \begin{equation} \label{eq:lattes-quotient} \mathcal{L} \times_{\Spec(\maxO_H)}\Spec(R) = E/\maxO_K^*. \end{equation} (This can in fact serve as an alternative definition of $\mathcal{L}$.) The universal map $\mathcal{E}\to\mathcal{L}$ is then nothing more than the classical Weber function expressed in our language. Gurney (4.3.16 of~\cite{Gurney:thesis}) proves there exists an isomorphism $\mathcal{L}\to \mathbb{P}^1_{\maxO_H}$ of $\maxO_H$-schemes. (The fact that $\mathcal{L}$ is a $\mathbb{P}^1$-bundle over $\maxO_H$ is expected, but the fact that it is the trivial bundle requires more work.) It follows that as an $\maxO_K$-scheme, $\mathcal{L}$ has geometrically disconnected fibers if $h_K>1$. We also emphasize that there appears to be no canonical isomorphism $\mathcal{L}\to \mathbb{P}^1_{\maxO_H}$. There is a canonical $\maxO_H$-point $\infty\in\mathcal{L}(\maxO_H)$, which corresponds to pairs of the form $(E,0)$. The isomorphism $\mathcal{L}\to\mathbb{P}^1_{\maxO_H}$ would then typically be chosen to send $\infty$ to $\infty$, but we cannot make any further restrictions, it seems. So the isomorphism $\mathcal{L}\to\mathbb{P}^1_{\maxO_H}$ is canonically defined only up to the action of the stabilizer group of $\infty$, the semi-direct product $\maxO_H\rtimes\maxO_H^*$. The endomorphisms $\psi_{\mathfrak a}$ of $\mathcal{E}$ induce endomorphisms $\psi_{\mathfrak a}:\mathcal{L}\to\mathcal{L}$ (morphisms of $\maxO_K$-schemes). Then $\psi_{\mathfrak p}$ for ${\mathfrak p}$ prime reduces to the $N({\mathfrak p})$-power Frobenius map $F_{\mathfrak p}$ modulo ${\mathfrak p}$, and hence the $\psi_{\mathfrak a}$ define a $\Lambda $-structure on $\mathcal{L}$. (See 4.3.11 of~\cite{Gurney:thesis}.) Further the map $\mathcal{L}\to\Spec\maxO_H$ above is a morphism of $\Lambda$-schemes, where $\maxO_H$ is given its unique $\Lambda$-ring structure. When expressed in terms of $\mathbb{P}^1_{\maxO_H}$, the endomorphisms $\psi_{\mathfrak a}$ are often call the Latt\`es functions, as for example in Milnor's book~\cite{Milnor:dynamics-book} (page 72), which explains the name. \begin{proposition} \label{pro:lattes-torsor} \begin{enumerate} \item $\mathcal{L}[{\mathfrak f}]({\mathbb C})$ is the set of isomorphism classes of pairs $(E,x)$, where $E$ is a complex CM elliptic curve and $x\in E[{\mathfrak f}]({\mathbb C})$. \item The action of $\mathrm{Id}_{{P}}$ on $\mathcal{L}[{\mathfrak f}]({\mathbb C})$, where ${\mathfrak a}$ acts as $\psi_{\mathfrak a}$, factors through an action (necessarily unique) of ${\mathrm{DR}}_{{P}}({\mathfrak f})$ on $\mathcal{L}[{\mathfrak f}]({\mathbb C})$. The resulting ${\mathrm{DR}}_{{P}}({\mathfrak f})$-set is a torsor generated by any class of the form $(E_0,x_0)$, where $E_0$ is a CM elliptic curve and $x_0$ is a generator of $E_0[{\mathfrak f}]({\mathbb C})$ as an $\maxO_K$-module. \end{enumerate} \end{proposition} \begin{proof} (1): Any point of $y\in\mathcal{L}({\mathbb C})$ is the isomorphism class of a pair $(E,x)$, where $E$ is a complex CM elliptic curve and $x\in E({\mathbb C})$. Under $\psi_{\mathfrak f}$, the pair $(E,x)$ maps to $({\mathfrak f}^{-1}\otimes E, 1\otimes x)$. So if $y$ lies in the ${\mathfrak f}$-torsion locus $\mathcal{L}[{\mathfrak f}]({\mathbb C})$, the object $({\mathfrak f}^{-1}\otimes E, 1\otimes x)$ must be isomorphic to one of the form $(E',0)$. Therefore the point $1\otimes x\in ({\mathfrak f}^{-1}\otimes E)({\mathbb C})$ is $0$, and hence $x$ is an ${\mathfrak f}$-torsion point of $E$. (2): Let us first show that the orbit of $(E_0,x_0)$ under $\mathrm{Id}_{{P}}$ is all of $\mathcal{L}[{\mathfrak f}]({\mathbb C})$. Let $(E,x)$ be an element of $\mathcal{L}[{\mathfrak f}]({\mathbb C})$. Then there is an integral ideal ${\mathfrak a}\subseteq\maxO_K$ such that there exists an isomorphism of elliptic curves $f\:{\mathfrak a}^{-1}\otimes E_0 \to E$. We can choose ${\mathfrak a}$ such that it is coprime to ${\mathfrak f}$. It follows that the image $1\otimes x_0 \in {\mathfrak a}^{-1}\otimes E_0({\mathbb C})$ of $x_0$ is a generator of $({\mathfrak a}^{-1}\otimes E_0)[{\mathfrak f}]({\mathbb C})$, and hence that $f(1\otimes x_0)$ is a generator of $E[{\mathfrak f}]({\mathbb C})$. In other words, there exists an element $b\in \maxO_K$ such that $bf(1\otimes x_0) = x$. Then we have \begin{align*} \psi_{b{\mathfrak a}}(E_0,x_0) &= (b^{-1}{\mathfrak a}^{-1}\otimes E_0, 1\otimes x_0) \cong ({\mathfrak a}^{-1}\otimes E_0, b\otimes x_0)\\ &\cong (E, f(b\otimes x_0)) = (E,b f(1\otimes x_0)) = (E,x). \end{align*} Therefore $(E_0,x_0)$ generates $\mathcal{L}[{\mathfrak f}]({\mathbb C})$ as a ${\mathrm{DR}}_{{P}}({\mathfrak f})$-set. It remains to show $$ {\mathfrak a}\sim_{\mathfrak f}{\mathfrak b} \Longleftrightarrow \psi_{\mathfrak a}((E_0,x_0))=\psi_{\mathfrak b}((E_0,x_0)), $$ for any two ideals ${\mathfrak a},{\mathfrak b}\in\mathrm{Id}_{{P}}$. First consider the direction $\Rightarrow$. The equality $\psi_{\mathfrak a}((E_0,x_0))=\psi_{\mathfrak b}((E_0,x_0))$ is equivalent to the existence of an isomorphism $({\mathfrak a}^{-1}\otimes E_0, 1\otimes x_0)\cong({\mathfrak b}^{-1}\otimes E_0, 1\otimes x_0)$. Since ${\mathfrak a}\sim_{\mathfrak f}{\mathfrak b}$, by~\ref{pro:DR-original-def}, there exists an element $t\in K$ such that ${\mathfrak a}={\mathfrak b} t$ and $t-1\in {\mathfrak f}{\mathfrak b}^{-1}$. So it is enough to show that the isomorphism \begin{equation} \label{eq:583} {\mathfrak a}^{-1}\otimes E_0 \longlabelmap{t\otimes 1} {\mathfrak b}^{-1}\otimes E_0 \end{equation} given by multiplication by $t\otimes 1$, sends $1\otimes x_0$ on the left to $1\otimes x_0$ on the right. In other words, it is enough to show that the two elements $1\otimes x_0, t\otimes x_0\in {\mathfrak b}^{-1}\otimes E_0({\mathbb C})$ agree. We give an argument using period lattices. Write $E_0({\mathbb C})={\mathbb C}/{\mathfrak c}$, where ${\mathfrak c}$ is a fractional ideal of $K$, and let $y\in{\mathbb C}$ be a coset representative of $x_0\in{\mathbb C}/{\mathfrak c}$. Then the morphism (\ref{eq:583}) above is identified with $$ {\mathbb C}/{\mathfrak a}^{-1}{\mathfrak c} \longlabelmap{t} {\mathbb C}/{\mathfrak b}^{-1}{\mathfrak c}. $$ So what we want to show is equivalent to the congruence $$ ty \equiv y \mod {\mathfrak b}^{-1}{\mathfrak c}. $$ But since $x_0$ is an ${\mathfrak f}$-torsion element of ${\mathbb C}/{\mathfrak c}$, we have $y\in{\mathfrak f}^{-1}{\mathfrak c}$, and hence we have $$ (t-1)y \in {\mathfrak f}{\mathfrak b}^{-1}\cdot {\mathfrak f}^{-1}{\mathfrak c} = {\mathfrak b}^{-1}{\mathfrak c}, $$ as desired. Now consider the direction $\Leftarrow$. So assume $\psi_{\mathfrak a}(E_0,x_0)\cong \psi_{\mathfrak b}(E_0,x_0)$. This means there is an isomorphism $$ {\mathbb C}/{\mathfrak a}^{-1}{\mathfrak c} \longlabelmap{t} {\mathbb C}/{\mathfrak b}^{-1}{\mathfrak c}, $$ given by multiplication by some element $t\in K^*$ with $(t)={\mathfrak a}{\mathfrak b}^{-1}$, such that $tx_0=x_0$. Therefore the diagram $$ \xymatrix{ {\mathfrak f}^{-1}{\mathfrak c}/{\mathfrak c} \ar^{{\mathrm{id}}}[r]\ar[d] & {\mathfrak f}^{-1}{\mathfrak c}/{\mathfrak c} \ar[d] \\ {\mathbb C}/{\mathfrak a}^{-1}{\mathfrak c} \ar^{t}[r] & {\mathbb C}/{\mathfrak b}^{-1}{\mathfrak c} } $$ commutes, since the two ways around the diagram agree on $x_0$, which is a generator of ${\mathfrak f}^{-1}{\mathfrak c}/{\mathfrak c}$. Therefore the difference map $t-1\:{\mathfrak f}^{-1}{\mathfrak c}/{\mathfrak c}\to {\mathbb C}/{\mathfrak b}^{-1}{\mathfrak c}$ is zero. This implies $(t-1){\mathfrak f}^{-1}{\mathfrak c}\subseteq {\mathfrak b}^{-1}{\mathfrak c}$ and hence $t-1\in {\mathfrak f}{\mathfrak b}^{-1}$. It follows from~\ref{pro:DR-original-def} that ${\mathfrak a}\sim_{\mathfrak f} {\mathfrak b}$. \end{proof} \begin{proposition} \label{pro:lattes-periodic-torsion} Let ${\mathfrak f}$ be an integral ideal in $\maxO_K$. Then we have equalities $$ \perfl{\mathcal{L}}{{\mathfrak f}} = \perred{\mathcal{L}}{{\mathfrak f}} =\mathcal{L}[{\mathfrak f}]_{\mathrm{red}} $$ of closed subschemes of $\mathcal{L}$. \end{proposition} \begin{proof} We have the following diagram of containments of closed subschemes of $\mathcal{L}$: $$ \xymatrix{ \mathcal{L}({\mathfrak f})_{\mathrm{red}} \ar@{>->}^{\alpha}[r]\ar@{>->}^{\beta}[d] & \mathcal{L}({\mathfrak f})_{\mathrm{fl}} \ar@{>->}[r]\ar@{>->}[d] & \mathcal{L}({\mathfrak f}) \ar@{>->}[d]\\ \mathcal{L}[{\mathfrak f}]_{\mathrm{red}} \ar@{>->}[r] & \mathcal{L}[{\mathfrak f}]_{\mathrm{fl}} \ar@{>->}[r] & \mathcal{L}[{\mathfrak f}] } $$ (As an aside, we note that $\mathcal{L}[{\mathfrak f}]$ is finite flat of degree $N({\mathfrak f})$ over $\maxO_H$, by~4.3.11 of~\cite{Gurney:thesis}, and hence $\mathcal{L}[{\mathfrak f}]_{\mathrm{fl}}=\mathcal{L}[{\mathfrak f}]$.) Observe that $\alpha$ is an isomorphism, by~(\ref{pro:periodic-structure2}), and so it is enough to show that $\beta$ is an isomorphism. It is enough to show this after base change to ${\mathbb C}$, since $\beta$ is a closed immersion and $\mathcal{L}[{\mathfrak f}]_{\mathrm{red}}$ is flat over $\maxO_K$. Further, since the schemes in question are of finite type, it is enough to show that $\beta$ induces a surjection on complex points, or in other words that the inclusion $$ \per{\mathcal{L}}{{\mathfrak f}}({\mathbb C}) \hookrightarrow \mathcal{L}[{\mathfrak f}]({\mathbb C}) $$ is surjective. But this follows from~\ref{pro:lattes-torsor}, which says that $\mathcal{L}[{\mathfrak f}]({\mathbb C})$ is ${\mathfrak f}$-periodic. \end{proof} \begin{corollary}\label{cor:weber-factorization} The map $\mathcal{E}[{\mathfrak f}]\to\mathcal{L}$ factors through the closed subscheme $\mathcal{L}({\mathfrak f})_\mathrm{fl}$ of $\mathcal{L}$. \end{corollary} \begin{proof} By functoriality, it factors through $\mathcal{L}[{\mathfrak f}]$. Because $\mathcal{E}[{\mathfrak f}]$ is reduced, it factors further through $\mathcal{L}[{\mathfrak f}]_\mathrm{red}$. Therefore, by~\ref{pro:lattes-periodic-torsion}, it factors through $\mathcal{L}({\mathfrak f})_\mathrm{fl}$. \end{proof} \begin{proposition}\label{pro:ell-finite-level} Let $R$ be a flat $\maxO_K$-algebra over which there is a CM elliptic curve $E$. Fix an ideal ${\mathfrak f}\subseteq\maxO_K$, and write $G=\maxO_K^*$. Let $\mathcal{L}_R({\mathfrak f})_\mathrm{fl}$ denote the base change of $\mathcal{L}({\mathfrak f})_\mathrm{fl}$ from $\maxO_H$ to $R$. Then we have the following: \begin{enumerate} \item The map $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})\to \sheafO(E[{\mathfrak f}])^G$ induced by corollary~\ref{cor:weber-factorization} is injective. \item Its cokernel is a finitely generated $R/nR$-module, where $n$ is the order of $\maxO_K^*$. \item For any sufficiently divisible integer $m$, there exists a (unique) morphism making the following diagram commute: $$ \xymatrix{ \sheafO(\mathcal{L}_R(m{\mathfrak f})_\mathrm{fl})\ar@{>->}[r] \ar[d] & \sheafO(E[m{\mathfrak f}])^G \ar[d]\ar@{-->}[dl] \\ \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})\ar@{>->}[r] & \sheafO(E[{\mathfrak f}])^G . } $$ \end{enumerate} \end{proposition} \begin{proof} Write $\mathcal{L}_R=E/G\cong\mathbb{P}^1_R$. Then $\mathcal{L}_R({\mathfrak f})_\mathrm{fl}$ is a closed subscheme of $\mathcal{L}_R$, and by~\ref{cor:weber-factorization}, we have the following containment of closed subschemes of $E$: $$ E[{\mathfrak f}] \subseteq \mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R} E. $$ Let $I_{\mathfrak f}(R)$ denote the deal sheaf on $\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R} E$ corresponding to $E[{\mathfrak f}]$; so we have the short exact sequence \begin{equation} \label{seq:hiho} 0 {\,\longrightarrow\,} I_{\mathfrak f}(R) {\,\longrightarrow\,} \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E) {\,\longrightarrow\,} \sheafO(E[{\mathfrak f}]) {\,\longrightarrow\,} 0. \end{equation} Its long exact sequence of group cohomology begins \begin{equation} \label{diag:g-mod-les} 0 {\,\longrightarrow\,} I_{\mathfrak f}(R)^G {\,\longrightarrow\,} \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E)^G {\,\longrightarrow\,} \sheafO(E[{\mathfrak f}])^G {\,\longrightarrow\,} H^1(G,I_{\mathfrak f}(R)). \end{equation} We would like to simplify this sequence using the fact that the map \begin{equation} \label{map:gme} \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})\to \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E)^G \end{equation} is an isomorphism, which we will now show. We have the following diagram of $G$-equivariant quasi-coherent sheaves on $\mathcal{L}_R$ (dropping the usual direct-image notation for simplicity): $$ \xymatrix{ 0 \ar[r] & \sheafO_{\mathcal{L}_R} \ar[r]\ar@{->>}[d] & \sheafO_E \ar[r] \ar@{->>}[d] & M \ar[r]\ar@{->>}[d] & 0 \\ 0 \ar[r] & \sheafO_{\mathcal{L}_R({\mathfrak f})_\mathrm{fl}} \ar[r] & \sheafO_{\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E} \ar[r] & \bar{M} \ar[r] & 0, } $$ where the rows are exact, $M$ and $\bar{M}$ being defined to be the cokernels as shown. We know $\sheafO_{\mathcal{L}_R}=\sheafO_E^G$, and so we have $({\mathbb Q}\otimes_{\mathbb Z} M)^G=0$ and hence $({\mathbb Q}\otimes_{\mathbb Z}\bar{M})^G=0$. Therefore the map (\ref{map:gme}) becomes an isomorphism after tensoring with ${\mathbb Q}$. To show it is an isomorphism, it is therefore enough to show it is surjective. Also observe that both sides are torsion free. So let $b$ be a $G$-invariant element of $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E)$. Then we have $b=a/n$, for some $a\in\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$ and some integer $n\geq 1$. Therefore $a$ becomes a multiple of $n$ in $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E)$. But because $\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R}E$ is faithfully flat over $\mathcal{L}_R({\mathfrak f})_\mathrm{fl}$, it must be a multiple of $n$ already in $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$. Therefore $b$ is in $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$, and so the map (\ref{map:gme}) is surjective, and hence an isomorphism. Thus we can rewrite the long exact sequence (\ref{diag:g-mod-les}) as \begin{equation} \label{diag:g-mod-les2} 0 {\,\longrightarrow\,} I_{\mathfrak f}(R)^G {\,\longrightarrow\,} \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}) {\,\longrightarrow\,} \sheafO(E[{\mathfrak f}])^G {\,\longrightarrow\,} H^1(G,I_{\mathfrak f}(R)). \end{equation} Using this sequence, we will prove parts (1)--(3). (1): We will show $I_{\mathfrak f}(R)^G=0$. First, observe that $I_{\mathfrak f}(R)^G$ is a flat-local construction in $R$: if $R'/R$ is flat, we have $I_{\mathfrak f}(R')=R'\otimes_R I_{\mathfrak f}(R)$ and hence $$ I_{\mathfrak f}(R')^G = (R'\otimes_R I_{\mathfrak f}(R))^G = R'\otimes_R I_{\mathfrak f}(R)^G $$ since taking invariants under the action of a finite group commutes with flat base change. Second, since $I_{\mathfrak f}(R)$ is a submodule of $\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$, which by construction is flat over $O_K$, its invariant subgroup $I_{\mathfrak f}(R)^G$ is also flat over $\maxO_K$ and hence maps injectively to $$ {\mathbb C}\otimes_{\maxO_K}I_{\mathfrak f}(R)^G=({\mathbb C} \otimes_{\maxO_K} R)\otimes_R I_{\mathfrak f}(R)^G=I_{\mathfrak f}({\mathbb C}\otimes_{\maxO_K}R)^G. $$ Therefore to show $I_{\mathfrak f}(R)^G=0$, it is enough to do it in the case where $R$ is a ${\mathbb C}$-algebra; and because any ${\mathbb C}$-algebra is flat over ${\mathbb C}$, we can apply the flat-local property again and conclude that it is enough to assume $R={\mathbb C}$, which we will now do. Summing up, we have the diagram $$ \xymatrix{ E[{\mathfrak f}] \ar@{>->}[r] & \mathcal{L}_{\mathbb C}({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_{\mathbb C}}E \ar@{>->}[r]\ar[d] & E \ar[d] \\ & \mathcal{L}_{\mathbb C}({\mathfrak f})_\mathrm{fl} \ar@{>->}[r] & \mathcal{L}_{\mathbb C} } $$ where $E\to \mathcal{L}_{\mathbb C}$ is a $G$-Galois cover of complex curves, the horizontal maps are closed immersions, and $E[{\mathfrak f}]$ and $\mathcal{L}_{\mathbb C}({\mathfrak f})_\mathrm{fl}$ are reduced. Write $I_{\mathfrak f}({\mathbb C})=\bigoplus_Z I_Z$, where $Z$ runs over the $G$-orbits of $E[{\mathfrak f}]({\mathbb C})$ and the ideal $I_Z$ of $\sheafO(\mathcal{L}_{\mathbb C}({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_{\mathbb C}}E)$ is the part of $I_{\mathfrak f}({\mathbb C})$ supported at $Z$. It is then enough to show $I_Z^G=0$ for each $Z$. So consider the filtration of $I_Z$ by its powers: $$ I_Z \supseteq I_Z^2 \supseteq \cdots \supseteq I_Z^e=\{0\}, $$ where $e$ is the ramification index of $E$ over $\mathcal{L}$ at $Z$, or equivalently the order of the stabilizer subgroup $H\subseteq G$ of a point $x\in Z$. Then it is enough to show $(I_Z^n/I_Z^{n+1})^G=0$ for $n=1,\dots,e-1$. This holds vacuously if $e=1$. So assume $e>1$. For $n\leq e-1$, the $G$-representation $I_Z^n/I_Z^{n+1}$ is the induced representation $\mathrm{Ind}_{H}^G(U^{\otimes n})$ where and $U=I_x/I_x^2$ is the cotangent space of $E$ at $x$. Therefore we have $$ (I_Z^n/I_Z^{n+1})^G = (U^{\otimes n})^{H}. $$ Observe that as a representation of $H$, the cotangent space $U$ is isomorphic to the restriction to $H$ of the representation of $G=\maxO_K^*$ on ${\mathbb C}$ given by usual multiplication. Therefore $U^{\otimes n}$ is the one-dimensional representation on which a generator $\zeta\in H$ acts as multiplication by $\zeta^n$. But since $n<e$, and since $e$ is the order of $H$, we have $\zeta^n\neq 1$. Therefore $\zeta$ acts nontrivially on $U^{\otimes n}$ and hence we have $$ (I_Z^n/I_Z^{n+1})^G = (U^{\otimes n})^H = 0, $$ as desired. (2): By general properties of group cohomology, $H^1(G,I_{\mathfrak f}(R))$ is an $R$-module of exponent dividing $n$, and hence an $R/nR$-module. The cokernel in question is a sub-$R$-module of $H^1(G,I_{\mathfrak f}(R))$ and is therefore also an $R/nR$-module. It remains to show it is finitely generated. Since it is a quotient of $\sheafO(E[{\mathfrak f}])^G$, it is enough to show that $\sheafO(E[{\mathfrak f}])^G$ is finitely generated, and hence enough to show this locally on $R$. But locally $E$ descends to some finitely generated $\maxO_K$-algebra over which $R$ is flat. Since the formation of $\sheafO(E[{\mathfrak f}])^G$ commutes with flat base change, it is enough to show finite generation in the case where $R$ is a finitely generated $\maxO_K$-algebra. Here it holds because $E[{\mathfrak f}]$ is finite flat and $R$ is noetherian. (3): Write $I_{\mathfrak f}=I_{\mathfrak f}(R)$. Then by part (1), for any $m\geq 1$,we have a morphism of exact sequences $$ \xymatrix{ 0 \ar[r] & \sheafO(\mathcal{L}_R(m{\mathfrak f})_\mathrm{fl}) \ar[r] \ar[d] & \sheafO(E[m{\mathfrak f}])^G \ar[r] \ar[d] & H^1(G,I_{m{\mathfrak f}}) \ar^\alpha[d] \\ 0 \ar[r] & \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}) \ar[r] & \sheafO(E[{\mathfrak f}])^G \ar[r] & H^1(G,I_{{\mathfrak f}}). } $$ To prove (3), it is enough to show the map $\alpha$ above is zero for sufficiently divisible $m$. Since $I_{\mathfrak f}$ is torsion free, it is enough by lemma~\ref{lem:H-vanishing} below to show that the map $I_{m{\mathfrak f}}\to I_{\mathfrak f}$ is zero modulo $n$ for sufficiently divisible $m$. Further, by part (2), it is enough to show this instead modulo each prime dividing $n$---that is, it is enough to show that for each prime $p\mid n$, there exists an integer $m$ such that the map $I_{m{\mathfrak f}}\to I_{\mathfrak f}$ is zero modulo $p$. So fix a prime $p\mid n$. Let $E_p$ denote the reduction of $E$ modulo $p$, which is an elliptic curve over $R/pR$. Then for any $j\geq 0$, we have a morphism of short exact sequences $$ \xymatrix{ 0 \ar[r] & I_{p^j{\mathfrak f}}/pI_{p^j{\mathfrak f}} \ar[r] \ar[d] & \sheafO(\mathcal{L}_R(p^j{\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R} E_p) \ar[r]\ar[d] & \sheafO(E_p[p^f{\mathfrak f}]) \ar[r]\ar[d] & 0 \\ 0 \ar[r] & I_{{\mathfrak f}}/pI_{{\mathfrak f}} \ar[r] & \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}\times_{\mathcal{L}_R} E_p) \ar[r] & \sheafO(E_p[{\mathfrak f}]) \ar[r] & 0. } $$ (Left-exactness is because $E[{\mathfrak f}]$ and $E[p^j{\mathfrak f}]$ are flat.) Now whether the fibers of $E_p$ are supersingular or ordinary, the closed subschemes $E_p[p^j{\mathfrak f}]$ contain arbitrary nilpotent thickenings of $E_p[{\mathfrak f}]$, for large enough $j$. Since $\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}}\times_{\mathcal{L}}E_p$ is a nilpotent thickening of $E_p[{\mathfrak f}]$, we can take $j$ such that $E_p[p^j{\mathfrak f}]$ contains $\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}}\times_{\mathcal{L}}E_p$. Thus the map $$ I_{p^j{\mathfrak f}}/pI_{p^j{\mathfrak f}} {\,\longrightarrow\,} I_{{\mathfrak f}}/pI_{{\mathfrak f}} $$ is zero, and hence so is the map $I_{p^j{\mathfrak f}}\to I_{\mathfrak f}/pI_{\mathfrak f}$, as desired. \end{proof} \begin{lemma}\label{lem:H-vanishing} Let $G$ be a finite group, and let $n$ denote its order. Let $M\to N$ be a morphism of $G$-modules which vanishes modulo $n$, and assume $N$ is $n$-torsion free. Then the induced map $H^i(G,M)\to H^i(G,N)$ is zero, for all $i\geq 1$. \end{lemma} \begin{proof} Since our given map $M\to N$ vanishes modulo $n$ and $N$ is $n$-torsion free, it factors as $$ \xymatrix{ M \ar@{-->}[d] \ar[dr] \\ N \ar^{n}[r] & N. } $$ Hence the induced map on cohomology factors similarly $$ \xymatrix{ H^i(G,M) \ar@{-->}[d] \ar[dr] \\ H^i(G,N) \ar^{n}[r] & H^i(G,N) } $$ However $H^i(G,N)$ is a $n$-torsion group for $i\geq 1$, because $G$ has exponent dividing $n$. Therefore the bottom map is zero and hence so is the diagonal map. \end{proof} \begin{corollary}\label{cor:lattes-elliptic-comparison} With the notation of proposition~\ref{pro:ell-finite-level}, the map $$ \Big(\sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})\Big)_{\mathfrak f} {\,\longrightarrow\,} \Big(\sheafO(E[{\mathfrak f}])^G\Big)_{\mathfrak f} $$ of pro-rings is an isomorphism. \end{corollary} \begin{proposition}\label{pro:local-cm-maximality} Let ${\mathfrak p}$ be a prime of $\maxO_K$, let $R$ be a finite unramified extension of $\maxO_{K_{\mathfrak p}}$, and let $E$ be a CM elliptic curve over $R$. Then for any integer $r\geq 0$, $\sheafO(E[{\mathfrak p}^r])$ is $\Lambda_{{\mathfrak p}}$-normal over $\maxO_{K_{\mathfrak p}}$. \end{proposition} \begin{proof} We follow (3.1) of~\cite{Borger-deSmit:Integral-models}, which is the analogous result for ${\mathbb G}_{\mathrm{m}}$. For $r=0$, it is clear. So we may assume that $r\geq 1$ and, by induction, that $\sheafO(E[{\mathfrak p}^{r-1}])$ is $\Lambda_{{\mathfrak p}}$-normal over $\maxO_{K_{\mathfrak p}}$. The ${\mathfrak p}^r$-torsion $E[{\mathfrak p}^r]$ can be obtained by gluing the ${\mathfrak p}^{r-1}$-torsion $E[{\mathfrak p}^{r-1}]$ and the locus of exact order ${\mathfrak p}^r$, as follows. Let $\ringd_r$ denote $\sheafO(E[{\mathfrak p}^r])$. Let $\pi$ be a generator of ${\mathfrak p}\maxO_{K_{\mathfrak p}}$, and consider the Lubin--Tate polynomial $h(z)\in R[z]$ for $E$ with respect to the uniformizer $\pi$ and some fixed local coordinate. So we have $h(z)=zg(z)$, where $g(z)$ is a monic Eisenstein polynomial with $g(0)=\pi$. Then $\ringd_r$ is identified with $R[z]/(h^{\circ r}(z))$. Because $h^{\circ r}(z)=h^{\circ r-1}(z)\cdot g(h^{\circ r-1}(z))$, we have the following diagram of quotients of $R[z]$: $$ \xymatrix{ R[z]/(h^{\circ r}(z)) \ar[d]\ar[r] & R[z]/(h^{\circ r-1}(z)) \ar[d] \\ R[z]/(g(h^{\circ r-1}(z))) \ar[r] & R[z]/(h^{\circ r-1}(z),g(h^{\circ r-1}(z))). } $$ Observe that $g(h^{\circ r-1}(z))$ is a monic Eisenstein polynomial. For degree reasons, it does not divide $h^{\circ r-1}(z)$, and so the least common multiple of these two polynomials is their product. Therefore we have $$ (h^{\circ r}(z)) = (h^{\circ r-1}(z))\cap (g(h^{\circ r-1}(z))), $$ and hence the diagram above is a pull-back diagram. Further, if we put $\ringb_r=R[z]/(g(h^{\circ r-1}(z)))$, then $\ringb_r$ is the ring of integers in a totally ramified extension of $K_{\mathfrak p}$. Let $\pi_r\in \ringb_r$ denote the corresponding uniformizer, namely the coset of $z$. Thus we have a pull-back diagram $$ \xymatrix{ \ringd_r \ar@{->>}[d]\ar@{->>}[r] & \ringd_{r-1} \ar@{->>}[d] \\ \ringb_r \ar@{->>}[r] & \bar{\ringb}_r, } $$ where $\bar{\ringb}_r=\ringb_r/(h^{\circ r-1}(\pi_r))$. Further observe that the element $h^{\circ r-1}(\pi_r)\in \ringb_r$ is a root of $g(z)$, which is an Eisenstein polynomial of degree $q-1$, where we write $q=N({\mathfrak p})$. Therefore we have $v_{\mathfrak p}(h^{\circ r-1}(\pi_r))=1/(q-1)$, where $v_{\mathfrak p}$ is the valuation normalized such that $v_{\mathfrak p}(\pi)=1$. Now suppose $C$ is a sub-$\Lambda_{\mathfrak p}$-ring of $K_{\mathfrak p}\otimes_{\maxO_{K_{\mathfrak p}}} \ringd_r$ which is finite over $\maxO_{K_{\mathfrak p}}$. The maximality statement we wish to prove is that $C$ is contained in $\ringd_r=\ringb_r\times_{\bar{\ringb}_r} \ringd_{r-1}$. By induction, $\ringd_{r-1}$ is $\Lambda_{\mathfrak p}$-normal over $\maxO_{K_{\mathfrak p}}$, and hence the image of $C$ in $K_{\mathfrak p}\otimes_{\maxO_{K_{\mathfrak p}}}\ringd_{r-1}$ is contained in $\ringd_{r-1}$. Similarly, since $\ringb_r$ is a maximal order in the usual sense, the image of $C$ in $K_{\mathfrak p}\otimes_{\maxO_{K_{\mathfrak p}}}\ringb_r$ is contained in $\ringb_r$. Putting the two together, we have the containment $C\subseteq \ringb_r\times \ringd_{r-1}$. To show $C \subseteq \ringb_r\times_{\bar{\ringb}_r} \ringd_{r-1}$, let us suppose that this does not hold. Then there is an element $(b,\overline{f(z)})\in C\subseteq \ringb_r\times \ringd_{r-1}$ such that $b$ and $\overline{f(z)}$ do not become equal in $\bar{\ringb}_r$. In other words, we have $$ v_{\mathfrak p}(b-f(\pi_r))< 1/(q-1). $$ Further we can choose $(b,\overline{f(z)})\in C$ such that $v_{\mathfrak p}(b-f(\pi_r))$ is as small as possible. Write $a=v_{\mathfrak p}(b-f(\pi_r))$. Since $C$ is a sub-$\Lambda_{\mathfrak p}$-ring of $K_{\mathfrak p}\otimes_{\maxO_{K_{\mathfrak p}}}\ringd_r$, we know $$ (b,\overline{f(z)})^q-\psi_{\mathfrak p}((b,\overline{f(z)})) \in \pi C $$ The left side can be simplified using the fact $\psi_{\mathfrak p}(z)=h(z)$ on $\ringd_r$: $$ (b,\overline{f(z)})^q-\psi_{\mathfrak p}((b,\overline{f(z)})) = (b^q-f^*(h(\pi_r)),\overline{f(z)}^q-\overline{f^*(h(z))}) $$ where $f^*(z)$ the polynomial obtained by applying the Frobenius map to each coefficient of $f(z)$. Therefore by the minimality of $a$, we have \begin{align*} a &\leq v_{\mathfrak p}\Big(\frac{b^q-f^*(h(\pi_r))}{\pi} - \frac{f(z)^q-f^*(h(z))}{\pi}\Big|_{z=\pi_r}\Big) = v_{\mathfrak p}\Big(\frac{b^q-f(\pi_r)^q}{\pi}\Big) \\ &= v_{\mathfrak p}\Big(\frac{(b-f(\pi_r))^q}{\pi} + (b-f(\pi_r))\cdot c(b,f(\pi_r))\Big) \end{align*} where $c(X,Y)$ denotes the polynomial $$ \frac{(X^q-Y^q) - (X-Y)^q}{\pi(X-Y)} \in \maxO_{K_{\mathfrak p}}[X,Y]. $$ But again by the minimality of $a$, we know $a\leq v_{\mathfrak p}(b-f(\pi_r))$. Therefore we have $$ a\leq v_{\mathfrak p}\Big(\frac{(b-f(\pi_r))^q}{\pi}\Big) = qa-1 $$ and hence $a\geq 1/(q-1)$. This contradicts the assumption $a<1/(q-1)$. \end{proof} \begin{lemma}\label{lem:global-cm-maximality} Let $\scalA=\maxO_K[1/t]$, for some $t\in\maxO_K$, and let $R$ be a $\Lambda$-ring, finite \'etale over $A$, over which there exists a CM elliptic curve $E$. Then for any ideal ${\mathfrak f}\subseteq\maxO_K$, the $\Lambda$-ring $\sheafO(E[{\mathfrak f}])$ is $\Lambda$-normal over $\scalA$. \end{lemma} \begin{proof} Let $S$ denote the maximal $\Lambda$-order over $A$ in $K\otimes_{\maxO_K}\sheafO(E[{\mathfrak f}])$. Then we have an inclusion $\sheafO(E[{\mathfrak f}])\subseteq S$ of finite flat $\scalA$-algebras. Therefore it is an equality if $\maxO_{K_{\mathfrak p}}\otimes_\scalA \sheafO(E[{\mathfrak f}])=\maxO_{K_{\mathfrak p}}\otimes_\scalA S$ for all primes ${\mathfrak p}\nmid t$. Thus it is enough to show that $\maxO_{K_{\mathfrak p}}\otimes_\scalA \sheafO(E[{\mathfrak f}])$ is $\Lambda_{\mathfrak p}$-normal over $\maxO_{K_{\mathfrak p}}$ for all ${\mathfrak p}\nmid t$. Write ${\mathfrak f}={\mathfrak p}^r{\mathfrak g}$, where ${\mathfrak p}\nmid{\mathfrak g}$. Then putting $R_{\mathfrak p}=\maxO_{K_{\mathfrak p}}\otimes_\scalA R$, we have \begin{align*} \maxO_{K_{\mathfrak p}}\otimes_\scalA \sheafO(E[{\mathfrak f}]) &= \maxO_{K_{\mathfrak p}}\otimes_\scalA (\sheafO(E[{\mathfrak p}^r])\otimes_R \sheafO(E[{\mathfrak g}])) \\ &= (\maxO_{K_{\mathfrak p}}\otimes_\scalA \sheafO(E[{\mathfrak p}^r]))\otimes_{R_{\mathfrak p}} (\maxO_{K_{\mathfrak p}}\otimes_\scalA \sheafO(E[{\mathfrak g}])) \\ &= \sheafO(E_{R_{\mathfrak p}}[{\mathfrak p}^r]) \otimes_{R_{\mathfrak p}} \sheafO(E_{R_{\mathfrak p}}[{\mathfrak g}]) \\ &= \prod_{R'} \sheafO(E_{R'}[{\mathfrak p}^r]), \end{align*} where $R'$ runs over all irreducible direct factors of the finite \'etale $\maxO_{K_{\mathfrak p}}$-algebra $\sheafO(E_{R_{\mathfrak p}}[{\mathfrak g}])$. Therefore it is enough to show that each $\sheafO(E_{R'}[{\mathfrak p}^r])$ is $\Lambda_{\mathfrak p}$-normal over $\maxO_{K_{\mathfrak p}}$. But this follows from~\ref{pro:local-cm-maximality}. \end{proof} \begin{proposition}\label{pro:cm-maximality} Let $\scalA=\maxO_K[1/t]$, for some $t\in\maxO_K$. Put $R=\maxO_H[1/t]$ and assume there exists a CM elliptic curve $E$ over $R$. Write $G=\maxO_K^*$. Then the maximal $\Lambda$-order in $K\otimes_{\scalA}\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$ over $\scalA$ is $\sheafO(E[{\mathfrak f}])^G$. \end{proposition} \begin{proof} By lemma~\ref{lem:global-cm-maximality} and part (2) of proposition~\ref{pro:maximality-facts}, the $G$-invariant subring $\sheafO(E[{\mathfrak f}])^{G}$ is $\Lambda$-normal over $\scalA$. Since $K\otimes_{\scalA}\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})= K\otimes_{A}\sheafO(E[{\mathfrak f}])^G$, by part (2) of proposition~\ref{pro:ell-finite-level}, we can conclude that $\sheafO(E[{\mathfrak f}])^{G}$ is the maximal $\Lambda$-order over $\scalA$ in $K\otimes_{\scalA}\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})$. \end{proof} \begin{thm}\label{thm:cm-pro-ring} Let $\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}$ denote the maximal $\Lambda$-order over $\maxO_K$ in $K\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})$. Then \begin{enumerate} \item $\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}$ is isomorphic as a $\Lambda$-ring to the ray class algebra $\rcl{\maxO_K,{P}}{{\mathfrak f}}$. \item The map $$ \big(\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})\big)_{\mathfrak f} \longlabelmap{} \big(\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}\big)_{\mathfrak f} $$ is an isomorphism of pro-rings. \item If a prime number divides the index $[\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}:\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})]$ then it divides the order of the unit group $\maxO_K^*$. \end{enumerate} \end{thm} \begin{proof} (1): Since $\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}$ is the maximal $\Lambda$-order over $\maxO_K$ in $K\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})$, and since $\rcl{\maxO_K,{P}}{{\mathfrak f}}$ is that in $K\otimes_{\maxO_K}\rcl{\maxO_K,{P}}{{\mathfrak f}}$, it is enough to show there is an isomorphism $$ \mathcal{L}({\mathfrak f})(\bar{{\mathbb Q}})\cong {\mathrm{DR}}_{{P}}({\mathfrak f}) $$ of $\mathrm{Id}_{{P}}$-sets, but this is follows immediately from part (2) of proposition~\ref{pro:lattes-torsor}. (2): Since the rings $$ \scalA=\maxO_K[{\mathfrak q}^{-1}] $$ for any two primes ${\mathfrak q}\subset \maxO_K$ form a cover of $\maxO_K$, it is enough to prove the map of pro-rings is an isomorphism after base change from $\maxO_K$ to any two such $\scalA$. By Shimura's theorem (\cite{Gurney:thesis}, prop.\ 4.2.2), there are at least two primes ${\mathfrak q}\subset\maxO_K$ such that there exists a CM elliptic curve over $R=\maxO_H\otimes_{\maxO_K}A$. (In fact, he proves there are infinitely many.) So fixing $\scalA$ and a CM elliptic curve $E$ over $\maxO_H\otimes_{\maxO_K}A$, it is enough to show that \begin{equation} \label{eq:sdfkl} \Big(\scalA\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})\Big)_{\mathfrak f} \longlabelmap{} \Big(\scalA\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}\Big)_{\mathfrak f} \end{equation} is an isomorphism of pro-rings. But we also have $$ \scalA\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_\mathrm{fl}) = R\otimes_{\maxO_H} \sheafO(\mathcal{L}({\mathfrak f})_\mathrm{fl}) = \sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}}) $$ and hence by part (1) of proposition~\ref{pro:maximality-facts} $$ \sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})^{\sim} = \scalA\otimes_{\maxO_K}\sheafO(\mathcal{L}({\mathfrak f})_{\mathrm{fl}})^{\sim}. $$ Therefore (\ref{eq:sdfkl}) can be rewritten as \begin{equation} \label{eq:sdfkl2} \Big(\sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})\Big)_{\mathfrak f} \longlabelmap{} \Big(\sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})^{\sim}\Big)_{\mathfrak f}. \end{equation} We know further that the map $$ \sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})^{\sim} {\,\longrightarrow\,} \sheafO(E[{\mathfrak f}])^G $$ is an isomorphism by~\ref{pro:cm-maximality}. So the map (\ref{eq:sdfkl2}) can in turn be rewritten as \begin{equation} \label{eq:sdfkl3} \Big(\sheafO(\mathcal{L}_R({\mathfrak f})_{\mathrm{fl}})\Big)_{\mathfrak f} \longlabelmap{} \Big(\sheafO(E[{\mathfrak f}])^G\Big)_{\mathfrak f}, \end{equation} and this is an isomorphism of pro-rings by~\ref{cor:lattes-elliptic-comparison}. (3): It is enough to show these properties locally on $\maxO_K$. So as above, it is enough to show them after base change to $A=\maxO_K[{\mathfrak q}^{-1}]$, where ${\mathfrak q}$ is a prime of $\maxO_K$ such that there is a CM elliptic curve $E$ over $R=\maxO_H\otimes_{\maxO_K}A$. Then the map $A\otimes_{\maxO_K}\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}) \to A\otimes_{\maxO_K}\sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl})^{\sim}$ is identified with $$ \sheafO(\mathcal{L}_R({\mathfrak f})_\mathrm{fl}) {\,\longrightarrow\,} \sheafO(E[{\mathfrak f}])^G. $$ By part (2) of proposition~\ref{pro:ell-finite-level}, the image of this map is of finite index divisible only by the primes dividing the order of $\maxO_K^*$. \end{proof} \section{Further questions} Is it possible to use $\Lambda$-schemes of finite type to generate other large abelian extensions, beyond the Kroneckerian explicit class field theories? We will formulate a range of such questions in this section, some of which it is reasonable to hope have a positive answer and some which are more ambitious. \subsection{} \emph{$\Lambda$-geometric field extensions.} So far, we have mostly been interested in what in the introduction we called \emph{$\Lambda$-refinements} of explicit class field theories---that is, in generating ray class algebras instead of abelian field extensions. But hungry for positive answers, we will give weaker, field-theoretic formulations here. Let $K$ be a number field, and write $$ \Lambda=\Lambda_{\maxO_K,M_K}\quad \text{and}\quad \rcl{}{{\mathfrak f}}=\rcl{\maxO_K,M_K}{{\mathfrak f}}. $$ For any separated $\Lambda$-scheme $X$ of finite type over $\maxO_K$, consider the extension of $K$ obtained by adjoining the coordinates of the ${\mathfrak f}$-periodic points, for all cycles ${\mathfrak f}$: $$ K(X) := \bigcup_{\mathfrak f} K(X({\mathfrak f})(\bar{K})). $$ For instance, we have the following: \begin{enumerate} \item If $K={\mathbb Q}$ and $X$ is ${\mathbb G}_{\mathrm{m}}$ with the toric $\Lambda$-structure, then Then $K(X)$ is the maximal abelian extension $\bigcup_n {\mathbb Q}(\zeta_n)$. \item If $K={\mathbb Q}$ and $X={\mathbb A}^1$ with the Chebyshev $\Lambda$-structure, then $K(X)$ is the maximal totally real abelian extension $\bigcup_n {\mathbb Q}(\zeta_n+\zeta_n^{-1})$. \item $K$ is imaginary quadratic, $X$ is ${\mathbb P}^1_{\maxO_H}$ with the Latt\`es $\Lambda$-structure. Then $K(X)$ is the maximal abelian extension of $K$. \item If $K$ is general and $X=\Spec \rcl{}{{\mathfrak f}}$, then $K(X)$ is the ray class field $K({\mathfrak f})$. \item $K(X_1{\,\scriptstyle\amalg\,} X_2)$ is the compositum $K(X_1)K(X_2)$ \end{enumerate} Let us say that an abelian extension $L/K$ is \emph{$\Lambda$-geometric} if there exists an $X$ as above such that $L\subseteq K(X)$. First observe that any finite extension of a $\Lambda$-geometric extension is $\Lambda$-geometric, by (4) and (5) above. Therefore there is no maximal $\Lambda$-geometric extension unless, as in the examples above, the maximal abelian extension itself is $\Lambda$-geometric. \begin{enumerate} \item[(Q4)] Are there number fields other than ${\mathbb Q}$ and imaginary quadratic fields for which the maximal abelian extension is $\Lambda$-geometric? \end{enumerate} It is natural to consider Shimura's generalization of Kronecker's theory to CM fields and abelian varieties. We expect that it can be realized in our framework, or at least that some version of it. But note that, assuming $[K:{\mathbb Q}]>2$, the maximal abelian extension generated by Shimura's method is an infinite subextension of the maximal abelian extension itself---the relative Galois group is an infinite group of exponent $2$. (See~\cite{Shimura:class-fields}\cite{Ovseevich:cm-type}\cite{Wei:cm-motives}.) \begin{enumerate} \item[(Q5)] Let $K$ be a CM field of degree greater than $2$. Is Shimura's extension $\Lambda$-geometric? If so, is there an infinite extension of it which is $\Lambda$-geometric? \end{enumerate} \subsection{} \emph{Production of $\Lambda$-schemes from ray class algebras.} It appears difficult to find $\Lambda$-schemes of finite type which generate large infinite abelian extensions. Every example we know ultimately comes from varieties with complex multiplication. Here we will consider an alternative---the possibility of manufacturing $\Lambda$-schemes of finite type by interpolating the ray class schemes $\Spec(\rcl{}{{\mathfrak f}})$, in the way that ${\mathbb G}_{\mathrm{m}}$ can be viewed as interpolating the $\mu_n$ as $n$ varies. This raises some questions which have the flavor of algebraic number theory more than the geometric questions above, and hence have a special appeal. Let ${\mathfrak r}$ be a product of real places of $K$, and let $X$ a reduced $\Lambda$-scheme of finite type over $\maxO_K$. Assume further that the union $$ \bigcup_{{\mathfrak f}\inZ({P},{\mathfrak r})}X({\mathfrak f}) $$ of the closed subschemes $X({\mathfrak f})$ is Zariski dense in $X$. If it is not, replace $X$ with the closure. Then all the information needed to construct $X$ is in principle available inside the function algebra of $$ \colim_{{\mathfrak f}\inZ({P},{\mathfrak r})} \per{X}{{\mathfrak f}}. $$ To be sure, this ind-scheme and $X$ are quite far apart, much as an abelian variety and its $p$-divisible groups are. But pressing on, if $X$ satisfies the property in (Q2) in the introduction, then this colimit would be isomorphic to $$ \colim_{{\mathfrak f}\inZ({P},{\mathfrak r})}\big(\Spec\rcl{}{{\mathfrak f}}\big), $$ and so all the information need to construct $X$ is in principle available in projective limit $$ \perv{{\mathfrak r}}{} =\lim_{{\mathfrak f}\inZ({P},{\mathfrak r})}\rcl{}{{\mathfrak f}}, $$ which can be viewed as a construction purely in the world of algebraic number theory in that it depends only on $K$ and ${\mathfrak r}$ and not on any variety $X$. For example in the cyclotomic context, where $K={\mathbb Q}$ and ${\mathfrak r}=\infty$, the injective map $$ {\mathbb Z}[x^{\pm 1}] {\,\longrightarrow\,} \lim_n {\mathbb Z}[x]/(x^n-1) = \perv{\infty}{} $$ realizes the function algebra of ${\mathbb G}_{\mathrm{m}}$ as a dense finitely generated sub-$\Lambda$-ring of $\perv{\infty}{}$. We can ask whether similar subrings exist for general $K$: \begin{enumerate} \item[(Q6)] Does $\perv{{\mathfrak r}}{}$ have a dense sub-$\maxO_K$-algebra which is finitely generated as an $\maxO_K$-algebra? Does it have a dense sub-$\Lambda$-ring which is finitely generated as an $\maxO_K$-algebra? \end{enumerate} It might be possible to cook up such a subring purely algebraically, instead of going through geometry. \subsection{}\emph{The cotangent space.} One first step in finding such a subring might be to guess its dimension by looking at the cotangent space of the ray class algebras at a point modulo ${\mathfrak p}$. For example, the cotangent space of $\mu_{p^n}$ modulo $p$ at the origin is $1$-dimensional, at least if $n\geq 1$. For any ideal ${\mathfrak a}\in\mathrm{Id}_{M_K}$ (and ${\mathfrak r}$ still a product of real places), let $I_{\mathfrak a}$ denote the kernel of the morphism \begin{equation} \label{eq:cotan} \rcl{}{{\mathfrak r}{\mathfrak a}}\longlabelmap{\psi_{\mathfrak a}} \rcl{}{{\mathfrak r}}=\maxO_{K({\mathfrak r})}. \end{equation} So we have an exact sequence $$ 0{\,\longrightarrow\,} I_{\mathfrak a} {\,\longrightarrow\,} \rcl{}{{\mathfrak r}{\mathfrak a}} \longlabelmap{\psi_{\mathfrak a}} \maxO_{K({\mathfrak r})} {\,\longrightarrow\,} 0. $$ Therefore $I_{\mathfrak a}/I_{\mathfrak a}^2$ is naturally an $\maxO_{K({\mathfrak r})}$-module. It is finitely generated since $\rcl{}{{\mathfrak r}{\mathfrak a}}$ is noetherian, being finite over ${\mathbb Z}$. \begin{enumerate} \item[(Q7)] Given a maximal ideal ${\mathfrak q}\subset \maxO_{K({\mathfrak r})}$ with residue field $k$, is the dimension $$ \dim_k(k\otimes_{\maxO_{K({\mathfrak r})}}I_{\mathfrak a}/I_{\mathfrak a}^2) $$ constant for large ${\mathfrak a}$? \end{enumerate} If so, can it be expressed in terms of the classical algebraic number theoretic invariants of $K$? One might hope it is the number of places of $K$ at infinity.

523ad89d7f3ff4d7cec4c3f0de81922c55d76b00

\section{Introduction} Tensor decompositions aim to find latent factors in tensor-valued data (i.e., the generalization of multi-dimensional arrays), thereby casting large-scale and intractable tensor problems into a multilinear tensor latent space of low-dimensionality (very few degrees of freedom designated by the rank). The latent factors within tensor decomposition can be considered as the latent features of data, which makes them an ideal set of bases to predict missing entries when the acquired data is incomplete. The specific forms and operations among latent factors determine the type of tensor decomposition. The most classical and successful tensor decomposition models are the Tucker decomposition (TKD) and the CANDECOMP/PARAFAC (CP) decomposition \cite{kolda2009tensor}. More recently, the matrix product state/tensor-train (MPS/TT) decomposition has become very attractive, owing to its super-compression and computational efficiency properties \cite{oseledets2011tensor}. Currently, a generalization of TT decomposition, termed the tensor ring (TR) decomposition, has been studied across scientific disciplines \cite{zhao2016tensor,zhao2018learning}. These tensor decomposition models have found application in various fields such as machine learning \cite{wang2018wide,novikov2015tensorizing,anandkumar2014tensor,kanagawa2016gaussian}, signal processing \cite{cong2015tensor}, image/video completion \cite{liu2013tensor,zhao2016bayesian}, compressed sensing \cite{gandy2011tensor}, to name but a few. Tensor completion is one of the most important applications of tensor decompositions, with the goal to recover an incomplete tensor from partially observed entries. The theoretical lynchpin in tensor completion problems is the tensor low-rank assumption, and the methods can mainly be categorized into two types: (\textrm{i}) tensor-decomposition-based approach and (\textrm{ii}) rank-minimization-based approach. Tensor decomposition based methods find latent factors of tensor using the incomplete tensor, and then the latent factors are used to predict the missing entries. Many completion algorithms have been proposed based on alternating least squares (ALS) method \cite{grasedyck2015variants,wang2017efficient}, gradient-based method \cite{yuan2017completion,acar2011scalable}, to mention but a few. Though ALS and gradient-based algorithms are free from burdensome hyper-parameter tuning, the performance of these algorithms is rather sensitive to model selection, i.e., rank selection of the tensor decomposition. Moreover, since the optimal rank is generally data-dependent, it is very challenging to specify the optimal rank beforehand. This is especially the case for Tucker, TT, and TR decompositions, for which the rank is defined as a vector; it is therefore impossible to find the optimal ranks by cross-validation due to the immense possibilities. Rank minimization based methods employ convex surrogates to minimize the tensor rank. One of the most commonly-used surrogates is the nuclear norm (a.k.a. Schatten norm, or trace norm), which is defined as the sum of singular values of a matrix and it is the most popular convex surrogate for rank regularization. Based on different definitions of tensor rank, various nuclear norm regularized algorithms have been proposed \cite{liu2013tensor,imaizumi2017tensor,liu2014generalized,liu2015trace}. Rank minimization based methods do not need to specify the rank of the employed tensor decompositions beforehand, and the rank of the recovered tensor will be automatically learned from the limited observations. However, these algorithms face multiple large-scale singular value decomposition (SVD) operations on the 2D unfoldings of the tensor when employing the nuclear norm and numerous hyper-parameter tuning, which in turn leads to high computational cost and low efficiency. To address the problems of high sensitivity to rank selection and low computational efficiency which are inherent in traditional tensor completion methods, in this paper, we propose a new algorithm named tensor ring low-rank factors (TRLRF) which effectively alleviates the burden of rank selection and reduces the computational cost. By virtue of employing both nuclear norm regularization and tensor decomposition, our model provides performance stability and high computational efficiency. The proposed TRLRF is efficiently solved by the ADMM algorithm and it simultaneously achieves both the underlying tensor decomposition and completion based on TR decomposition. Our main contributions in this paper are: \begin{itemize} \item A theoretical relationship between the multilinear tensor rank and the rank of TR factors is established, which allows the low-rank constraint to be performed implicitly on TR latent space. This has led to fast SVD calculation on small size factors. \item The nuclear norm is further imposed to regularize the TR-ranks, which enables our algorithm to always obtain a stable solution, even if the TR-rank is inappropriately given. This highlights rank-robustness of the proposed TRLRF algorithm. \item An efficient algorithm based on ADMM is developed to optimize the proposed model, so as to obtain the TR-factors and the recovered tensor simultaneously. \end{itemize} \section{Preliminaries and Related Works} \subsection{Notations} The notations in \cite{kolda2009tensor} are adopted in this paper. A scalar is denoted by a standard lowercase letter or a uppercase letter, e.g., $x, X \in\mathbb{R}$, and a vector is denoted by a boldface lowercase letter, e.g., $\vect{x}\in\mathbb{R}^{I}$. A matrix is denoted by a boldface capital letter, e.g., $\mat{X}\in\mathbb{R}^{I\times J}$. A tensor of order $N\geq 3$ is denoted by calligraphic letters, e.g., $\tensor{X}\in\mathbb{R}^{I_1\times I_2\times\cdots \times I_N}$. The set $\{ \tensor{X}^{(n)}\}_{n=1}^N:=\{ \tensor{X}^{(1)},\tensor{X}^{(2)},\ldots,\tensor{X}^{(N)}\} $ denotes a tensor sequence, with $\tensor{X}^{(n)}$ being the $n$-th tensor of the sequence. Where appropriate, a tensor sequence can also be written as $[\tensor{X}]$. The representations of matrix sequences and vector sequences are designated in the same way. An element of a tensor $\tensor{X} \in\mathbb{R}^{I_1\times I_2\times\cdots \times I_N}$ of index $(i_{1},i_{2},\ldots,i_{N})$ is denoted by $\tensor{X}(i_{1},i_{2},\ldots, i_{N})$ or $x_{i_{1}i_{2}\ldots i_{N}}$. The inner product of two tensors $\tensor{X}$, $\tensor{Y}$ with the same size $\mathbb{R}^{I_1\times I_2\times\cdots \times I_N}$ is defined as $\langle \tensor{X},\tensor{Y} \rangle=\sum_{i_1}\sum_{i_2}\cdots\sum_{i_N}x_{i_1 i_2\ldots i_N}y_{i_1 i_2\ldots i_N}$. Furthermore, the Frobenius norm of $\tensor{X}$ is defined by $\left \| \tensor{X} \right \|_F=\sqrt{\langle \tensor{X},\tensor{X} \rangle}$. We employ two types of tensor unfolding (matricization) operations in this paper. The standard mode-$n$ unfolding \cite{kolda2009tensor} of tensor $\tensor{X} \in\mathbb{R}^{I_1\times I_2\times\cdots \times I_N}$ is denoted by $\mat{X}_{(n)}\in\mathbb{R}^{I_n \times {I_1 \cdots I_{n-1} I_{n+1} \cdots I_N}}$. Another mode-$n$ unfolding of tensor $\tensor{X}$ which is often used in TR operations \cite{zhao2016tensor} is denoted by $\mat{X}_{<n>}\in\mathbb{R}^{I_n \times {I_{n+1} \cdots I_{N} I_{1} \cdots I_{n-1}}}$. Furthermore, the inverse operation of unfolding is matrix folding (tensorization), which transforms matrices to higher-order tensors. In this paper, we only define the folding operation for the first type of mode-$n$ unfolding as $\text{fold}_n(\cdot)$, i.e., for a tensor $\tensor{X}$, we have $\text{fold}_n(\mat{X}_{(n)})=\tensor{X}$. \subsection{Tensor ring decomposition} The tensor ring (TR) decomposition is a more general decomposition model than the tensor-train (TT) decomposition. It represents a tensor of higher-order by circular multilinear products over a sequence of low-order latent core tensors, i.e., TR factors. For $n=1,\ldots,N$, the TR factors are denoted by $\tensor{G}^{(n)} \in\mathbb{R}^{R_{n} \times I_{n} \times R_{n+1}}$ and each consists of two rank-modes (i.e mode-$1$ and mode-$3$) and one dimension-mode (i.e., mode-$2$). The syntax $\{R_1, R_2, \ldots, R_N\}$ denotes the TR-rank which controls the model complexity of TR decomposition. The TR decomposition applies trace operations and all of the TR factors are set to be 3-order; thus the TR decomposition relaxes the rank constraint on the first and last core of TT to $R_1=R_{N+1}$. Moreover, TR decomposition linearly scales to the order of the tensor, and in this way it overcomes the `curse of dimensionality'. In this case, TR can be considered as a linear combination of TTs and hence offers a powerful and generalized representation ability. The element-wise relation of TR decomposition and the generated tensor is given by: \begin{equation} \label{tr_relation1} \tensor{X}(i_1,i_2,\ldots,i_N)=\text{Trace}\left \{ \prod_{n=1}^N \mat{G}^{(n)}_{i_n} \right \}, \end{equation} where $\text{Trace}\{ \cdot \}$ is the matrix trace operation, $ \mat{G}^{(n)}_{i_n} \in\mathbb{R}^{R_n\times R_{n+1}}$ is the $i_n$-th mode-$2$ slice matrix of $\tensor{G}^{(n)}$, which can also be denoted by $\tensor{G}^{(n)}(:,i_n,:)$ according to the Matlab notation. \subsection{Tensor completion} \subsubsection{Completion by TR decomposition} Tensor decomposition based algorithms do not directly employ the rank constraint to the object tensor. Instead, they try to find the low-rank representation (i.e., tensor decompositions) of the incomplete data from the observed entries. The obtained latent factors of the tensor decomposition are used to predict the missing entries. For model formulation, the tensor completion problem is set as a weighted least squares (WLS) model. Based on different tensor decompositions, various tensor completion algorithms have been proposed, e.g., weighted CP \cite{acar2011scalable}, weighted Tucker \cite{filipovic2015tucker}, TRWOPT \cite{yuan2018higher} and TRALS \cite{wang2017efficient}. To the best of our knowledge, there are two proposed TR-based tensor completion algorithms: the TRALS and TRWOPT. They apply the same optimization model which is formulated as: \begin{equation} \min\limits_{[\tensor{G}]} \Vert P_{\Omega}(\tensor{T}-\Psi([\tensor{G}])) \Vert_F^2, \label{tdc} \end{equation} where the optimization objective is the TR factors, $[\tensor{G}]$, $P_{\Omega}(\tensor{T})$ denotes all the observed entries w.r.t. the set of indices of observed entries represented by $\Omega$, and $\Psi([\tensor{G}])$ denotes the approximated tensor generated by $[\tensor{G}]$. Every element of $\Psi([\tensor{G}])$ is calculated by equation \eqref{tr_relation1}. The two algorithms are both based on the model in (\ref{tdc}). However, TRALS applies alternative least squares (ALS) method and TRWOPT uses a gradient-based algorithm to solve the model, respectively. They perform well for both low-order and high-order tensors due to the high representation ability and flexibility of TR decomposition. However, these algorithms are shown to suffer from high sensitiveness to rank selection, which would lead to high computational cost. \subsubsection{Completion by nuclear norm regularization} The model of rank minimization-based tensor completion can be formulated as: \begin{equation} \label{bmc} \min \limits_{\tensor{X}} \ \ \text{Rank}(\tensor{X})+\frac{\lambda}{2}\Vert P_{\Omega}(\tensor{T}-\tensor{X})\Vert_F^2, \end{equation} where $\tensor{X}$ is the recovered low-rank tensor, and $\text{Rank}(\cdot)$ is a rank regularizer. The model can therefore find the low-rank structure of the data and approximate the recovered tensor. Because determining the tensor rank is an NP-hard problem \cite{hillar2013most,kolda2009tensor}, work in \cite{liu2013tensor} and \cite{signoretto2014learning} extends the concept of low-rank matrix completion and defines tensor rank as a sum of the rank of mode-$n$ unfolding of the object tensor. Moreover, the convex surrogate named nuclear norm is applied to the tensor low-rank model and it simultaneously regularizes all the mode-$n$ unfoldings of the object tensor. In this way, the model in \eqref{bmc} can be reformulated as: \begin{equation} \label{overlapped} \min_{\tensor{X}}\sum_{n=1}^N\Vert \mat{X}_{(n)} \Vert_*+\frac{\lambda}{2}\Vert P_{\Omega}(\tensor{T}-\tensor{X})\Vert_F^2, \end{equation} where $\Vert \cdot \Vert_*$ denotes the nuclear norm regularization in the form of a sum of the singular values of the matrix. Usually, the model is solved by ADMM algorithms and it is shown to have fast convergence and good performance when data size is small. However, when dealing with large-scale data, the multiple SVD operations in the optimization step will be intractable due to high computational cost. \section{Tensor Ring Low-rank Factors} To solve the issues traditional tensor completion methods have, we impose low-rankness on each of the TR factors and so that our basic tensor completion model is formulated as follow: \begin{equation} \label{ori_mod} \begin{aligned} \min \limits_{[\tensor{G}],\tensor{X}} \ \ &\sum_{n=1}^N \Vert \tensor{G}^{(n)}\Vert_*+ \frac{\lambda}{2}\Vert \tensor{X}-\Psi([\tensor{G}])\Vert_F^2,\\& s.t.\ P_\Omega(\tensor{X})=P_\Omega(\tensor{T}). \end{aligned} \end{equation} To solve (\ref{ori_mod}), we first need to deduce the relation of the tensor rank and the corresponding core tensor rank, which can be explained by the following theorem. \begin{theorem} Given an $N$-th order tensor $\tensor{X}\in\mathbb{R}^{I_1\times I_2\times\cdots \times I_N}$ which can be represented by equation (\ref{tr_relation1}), then the following inequality holds for all $n=1,\ldots,N$: \begin{equation} \text{Rank}(\mat{G}_{(2)}^{(n)}) \geq \text{Rank}(\mat{X}_{(n)}). \end{equation} \end{theorem} \begin{proof} For the $n$-th core tensor $\tensor{G}^{(n)}$, according to the work in \cite{zhao2016tensor}, we have: \begin{equation} \mat{X}_{<n>}=\mat{G}_{(2)}^{(n)}(\mat{G}_{<2>}^{(\neq n)})^T, \end{equation} where $\tensor{G}^{(\neq n)}\in\mathbb{R}^{R_{n+1}\times \prod_{i=1, i\neq n}^N I_i \times R_n}$ is a subchain tensor generated by merging all but the $n$-th core tensor. Hence, the relation of the rank satisfies: \begin{equation} \begin{aligned} \text{Rank}(\mat{X}_{<n>})& \leq \text{min}\{\text{Rank}(\mat{G}^{(n)}_{(2)}) ,\text{Rank}(\mat{G}_{<n>}^{(\neq n)})\}\\& \leq \text{Rank}(\mat{G}^{(n)}_{(2)}). \end{aligned} \end{equation} The proof is completed by \begin{equation} \begin{aligned} \text{Rank}(\mat{X}_{<n>})=\text{Rank}(\mat{X}_{(n)}) \leq \text{Rank}(\mat{G}^{(n)}_{(2)}). \end{aligned} \end{equation} \end{proof} This theorem proves the relation between the tensor rank and the rank of the TR factors. The rank of mode-$n$ unfolding of the tensor $\tensor{X}$ is upper bounded by the rank of the dimension-mode unfolding of the corresponding core tensor $\tensor{G}^{(n)}$, which allows us to impose a low-rank constraint on $\tensor{G}^{(n)}$. By the new surrogate, our model (\ref{ori_mod}) is reformulated by: \begin{equation} \label{mod} \begin{aligned} \min \limits_{[\tensor{G}],\tensor{X}} \ &\sum_{n=1}^N \Vert \mat{G}^{(n)}_{(2)} \Vert_*+ \frac{\lambda}{2}\Vert \tensor{X}-\Psi([\tensor{G}])\Vert_F^2\\& s.t.\ P_\Omega(\tensor{X})=P_\Omega(\tensor{T}). \end{aligned} \end{equation} The above model imposes nuclear norm regularization on the dimension-mode unfoldings of the TR factors, which can largely decrease the computational complexity compared to the algorithms which are based on model (\ref{overlapped}). Moreover, we consider to give low-rank constraints on the two rank-modes of the TR factors, i.e., the unfoldings of the TR factors along mode-$1$ and mode-$3$, which can be expressed by $\sum_{n=1}^N\Vert \mat{G}^{(n)}_{(1)}\Vert_*$+$\sum_{n=1}^N\Vert \mat{G}^{(n)}_{(3)}\Vert_*$. When the model is optimized, nuclear norms of the rank-mode unfoldings and the fitting error of the approximated tensor are minimized simultaneously, resulting in the initial TR-rank becoming the upper bound of the real TR-rank of the tensor, thus equipping our model with robustness to rank selection. The tensor ring low-rank factors (TRLRF) model can be finally expressed as: \begin{equation} \label{mod_o} \begin{aligned} \min \limits_{[\tensor{G}],\tensor{X}} \ &\sum_{n=1}^N\sum_{i=1}^3 \Vert \mat{G}^{(n)}_{(i)} \Vert_*+ \frac{\lambda}{2}\Vert \tensor{X}-\Psi([\tensor{G}])\Vert_F^2\\& s.t.\ P_\Omega(\tensor{X})=P_\Omega(\tensor{T}). \end{aligned} \end{equation} Our TRLRF model has two distinctive advantages. Firstly, the low-rank assumption is placed on tensor factors instead of on the original tensor, this greatly reduces the computational complexity of the SVD operation. Secondly, low-rankness of tensor factors can enhance the robustness to rank selection, which can alleviate the burden of searching for optimal TR-rank and reduce the computational cost in the implementation. \subsection{Solving scheme} To solve the model in (\ref{mod_o}), we apply the alternating direction method of multipliers (ADMM) which is efficient and widely used \cite{boyd2011distributed}. Moreover, because the variables of TRLRF model are inter-dependent, we impose auxiliary variables to simplify the optimization. Thus, the TRLRF model can be rewritten as \begin{equation} \label{L_mod} \begin{aligned} &\min \limits_{[\tensor{M}], [\tensor{G}],\tensor{X}} \ \sum_{n=1}^N\sum_{i=1}^3 \Vert \mat{M}^{(n,i)}_{(i)} \Vert_*+ \frac{\lambda}{2}\Vert \tensor{X}- \Psi([\tensor{G}])\Vert_F^2 ,\\& \quad s.t. \ \mat{M}^{(n,i)}_{(i)}=\mat{G}^{(n)}_{(i)}, n=1,\ldots,N,\ i=1,2,3, \\&\quad\quad\;\; P_\Omega(\tensor{X})=P_\Omega(\tensor{T}), \end{aligned} \end{equation} where $[\tensor{M}]:=\{\tensor{M}^{(n,i)}\}_{n=1,i=1}^{N,3}$ are the auxiliary variables of $[\tensor{G}]$. By merging the equal constraints of the auxiliary variables into the Lagrangian equation, the augmented Lagrangian function of TRLRF model becomes \begin{equation} \label{o_Lfunc} \begin{aligned} &L \left( [\tensor{G}],\tensor{X},[\tensor{M}], [\tensor{Y}]\right)\\&=\sum_{n=1}^N\sum_{i=1}^3 \big(\Vert \mat{M}^{(n,i)}_{(i)} \Vert_*+<\tensor{Y}^{(n,i)}, \tensor{M}^{(n,i)}-\tensor{G}^{(n)}>\\ &+\frac{\mu}{2}\Vert \tensor{M}^{(n,i)}-\tensor{G}^{(n)}\Vert_F^2 \big) +\frac{\lambda}{2}\Vert \tensor{X}- \Psi([\tensor{G}])\Vert_F^2 ,\\ &\quad s.t.\ P_\Omega(\tensor{X})=P_\Omega(\tensor{T}), \end{aligned} \end{equation} where $[\tensor{Y}]:=\{\tensor{Y}^{(n,i)}\}_{n=1,i=1}^{N,3}$ are the Lagrangian multipliers, and $\mu>0$ is a penalty parameter. For $n=1,\ldots,N$, $i=1,2,3$, $\tensor{G}^{(n)}$, $\tensor{M}^{(n,i)}$ and $ \tensor{Y}^{(n,i)}$ are each independent, so we can update them by the updating scheme below. \subsubsection{Update of $\tensor{G}^{(n)}$.} By using \eqref{o_Lfunc}, the augmented Lagrangian function w.r.t. ${\tensor{G}^{(n)}}$ can be simplified as \begin{equation} \label{LG_o} \begin{aligned} L(\tensor{G}^{(n)})&=\sum_{i=1}^3\frac{\mu}{2}\Big\Vert \tensor{M}^{(n,i)}-\tensor{G}^{(n)}+\frac{1}{\mu}\tensor{Y}^{(n,i)} \Big\Vert^2_F\\&+\frac{\lambda}{2}\big\Vert \tensor{X}-\Psi([\tensor{G}]) \big\Vert^2_F+C_{\tensor{G}}, \end{aligned} \end{equation} where the constant $C_{\tensor{G}}$ consists of other parts of the Lagrangian function which is irrelevant to updating $\tensor{G}^{(n)}$. This is a least squares problem, so for $n=1, \ldots,N$, $\tensor{G}^{(n)}$ can be updated by \begin{equation} \label{LG_u} \begin{aligned} &\tensor{G}^{(n)}=\text{fold}_{2}\Big(\big(\sum_{i=1}^{3}(\mu \mat{M}_{(2)}^{(n,i)} +\mat{Y}_{(2)}^{(n,i)})\\&+\lambda \mat{X}_{<n>}\mat{G}^{(\neq n)}_{<2>} \big)\big(\lambda\mat{G}^{(\neq n),T}_{<2>}\mat{G}^{(\neq n)}_{<2>}+3\mu \mat{I}\big)^{-1}\Big), \end{aligned} \end{equation} where $\mat{I}\in\mathbb{R}^{R_n^2\times R_n^2}$ denotes the identity matrix. \subsubsection{Update of $\tensor{M}^{(n,i)}$.} For $i=1,2,3$, the augmented Lagrangian functions w.r.t. $[\tensor{M}]$ is expressed as \begin{equation} \label{LM_o} \begin{aligned} L(\tensor{M}^{(n,i)})&=\frac{\mu}{2}\big\Vert \tensor{M}^{(n,i)}-\tensor{G}^{(n)}+\frac{1}{\mu}\tensor{Y}^{(n,i)} \big\Vert_F^2\\&+ \big\Vert \mat{M}^{(n,i)}_{(i)}\big \Vert_*+C_{\tensor{M}}. \end{aligned} \end{equation} The above formulation has a closed-form \cite{cai2010singular}, which is given by \begin{equation} \label{LM_u} \begin{aligned} \tensor{M}^{(n,i)}=\text{fold}_i\Big(D_{\frac{1}{\mu}}\big({\mat{G}^{(n)}_{(i)}}-\frac{1}{\mu}\mat{Y}^{(n,i)}_{(i)}\big)\Big), \end{aligned} \end{equation} where $D_{\beta}(\cdot)$ is the singular value thresholding (SVT) operation, e.g., if $\mat{U}\mat{S}\mat{V}^T$ is the singular value decomposition of matrix $\mat{A}$, then $D_\beta(\mat{A})=\mat{U}max\{\mat{S}-\beta \mat{I},0\}\mat{V}^T$. \subsubsection{Update of $\tensor{X}$.} The augmented Lagrangian functions w.r.t. $\tensor{X}$ is given by \begin{equation} \label{LX_o} \begin{aligned} L&(\tensor{X})=\frac{\lambda}{2}\big\Vert \tensor{X}-\Psi([\tensor{G}]) \big\Vert^2_F+C_{\tensor{X}}, \\&s.t. \ P_\Omega(\tensor{X})=P_\Omega(\tensor{T}), \end{aligned} \end{equation} which is equivalent to the tensor decomposition based model in (\ref{tdc}). The expression for $\tensor{X}$ is updated by inputing the observed values in the corresponding entries, and by approximating the missing entries by updated TR factors $[\tensor{G}]$ for every iteration, i.e., \begin{equation} \label{LX_u} \begin{aligned} \tensor{X}=P_{\Omega}(\tensor{T})+P_{\bar{\Omega}}(\Psi([\tensor{G}])), \end{aligned} \end{equation} where $\bar{\Omega}$ is the set of indices of missing entries which is a complement to $\Omega$. \subsubsection{Update of $\tensor{Y}^{(n,i)}$.} For $n=1,\ldots,N$ and $i=1,2,3$, the Lagrangian multiplier $ \tensor{Y}^{(n,i)}$ is updated as \begin{equation} \label{LY_u} \begin{aligned} \tensor{Y}^{(n,i)}=\tensor{Y}^{(n,i)}+\mu\big(\tensor{M}^{(n,i)}-\tensor{G}^{(n)}\big). \end{aligned} \end{equation} In addition, the penalty term of the Lagrangian functions $L$ is restricted by $\mu$ which is also updated for every iteration by $\mu=max\{\rho\mu,\mu_{max}\}$, where $1<\rho<1.5$ is a tuning hyper parameter. The ADMM based solving scheme is updated iteratively based on the above equations. Moreover, we consider to set two optimization stopping conditions: (\textrm{i}) maximum number of iterations $k_{max}$ and (\textrm{ii}) the difference between two iterations (i.e., $\Vert \tensor{X}-\tensor{X}_{last} \Vert_F / \Vert \tensor{X}\Vert_F$) which is thresholded by the tolerance $tol$. The implementation process and hyper-parameter selection of TRLRF is summarized in Algorithm 1. It should be noted that our TRLRF model is non-convex, so the convergence to the global minimum cannot be theoretically guaranteed. However, the convergence of our algorithm can be verified empirically (see experiment details in Figure \ref{conver}). Moreover, the extensive experimental results in the next section also illustrate the stability and effectiveness of TRLRF. \begin{figure}[htb] \begin{center} \includegraphics[width=1\linewidth]{convergence_new.pdf} \end{center} \caption{Illustration of convergence property for TRLRF under different hyper-parameter choices. A synthetic tensor with TR structure (size $7\times8\times 7\times8$ with TR-rank \{4,4,4,4\}, missing rate 0.5) is tested. The experiment records the change of the objective function values along the number of iterations. Each independent experiment is conducted 100 times and the average results are shown in the graphs. Panels (a) and (b) represent the convergence curve when TR-rank and $\lambda$ are changed respectively.} \label{conver} \end{figure} \begin{table}[h] \centering \begin{center} \begin{tabular}{l} \hline \textbf{Algorithm 1.} Tensor ring low-rank factors (TRLRF) \\ \hline 1: \textbf{Input}: $P_{\Omega}(\tensor{T})$, initial TR-rank $\{R_n\}_{n=1}^N$ . \\ 2: \textbf{Initialization}: For $n=1,\ldots,N$, $i=1,2,3$, \\\quad random sample $\tensor{G}^{(n)}$ by distribution $N\sim(0,1)$, \\\quad$\tensor{Y}^{(n,i)}=0$, $\tensor{M}^{(n,i)}=0$, $\lambda=5$, $\mu^0=1$, $\mu_{max}$ \\\quad$=10^2, \rho=1.01$, $tol=10^{-6}$, $k=0$, $k_{max}=300$. \\ 3:\; \textbf{For} $k=1$ to $k_{max}$ \textbf{do}\\ 4:\;\;\;\; $\tensor{X}_{last}=\tensor{X}$. \\ 5:\;\;\;\; Update $\{\tensor{G}^{(n)}\}_{n=1}^N$ by (\ref{LG_u}). \\ 6:\;\;\;\; Update $\{\tensor{M}^{(n,i)}\}_{n=1,i=1}^{N,3}$ by (\ref{LM_u}). \\ 7:\;\;\;\; Update $\tensor{X}$ by (\ref{LX_u}). \\ 8:\;\;\;\; Update $\{ \tensor{Y}^{(n,i)}\}_{n=1,i=1}^{N,3}$ by (\ref{LY_u}). \\ 9:\;\;\;\; $\mu=max(\rho\mu,\mu_{max})$ \\ 6:\;\;\;\; \textbf{If} $\Vert \tensor{X}-\tensor{X}_{last} \Vert_F / \Vert \tensor{X}\Vert_F<tol$, \textbf{break} \\ 7: \;\textbf{End for} \\ 8: \;\textbf{Output}: completed tensor $\tensor{X}$ and TR factors $[\tensor{G}]$. \\ \hline \end{tabular} \end{center} \end{table} \subsection{Computational complexity} We analyze the computational complexity of our TRLRF algorithm as follows. For a tensor $\tensor{X}\in\mathbb{R}^{I_1\times I_2\times \cdots\times I_N}$, the TR-rank is set as $R_1=R_2=\cdots=R_N=R$, then the computational complexity of updating $[\tensor{M}]$ represents mainly the cost of SVD operation, which is $\tensor{O}(\sum_{n=1}^N2I_nR^3+I_n^2R^2)$. The computational complexities incurred calculating $\mat{G}_{<2>}^{(\neq n)}$ and updating $[\tensor{G}]$ are $\tensor{O}(NR^3\prod_{i=1, i\neq n}^N I_i)$ and $\tensor{O}(NR^2\prod_{i=1}^N I_i+NR^6)$, respectively. If we assume $I_1=I_2=\cdots=I_N=I$, then overall complexity of our proposed algorithm can be written as $\tensor{O}(NR^2I^N+NR^6)$. Compared to HaLRTC and TRALS which are the representative of the nuclear-norm-based and the tensor decomposition based algorithms, the computational complexity of HaLRTC is $\tensor{O}(NI^{N+1})$. Since TRALS is based on ALS method and TR decomposition, its computational complexity is $\tensor{O}(PNR^4I^N+NR^6)$, where $P$ denotes the observation rate. We can see that the computational complexity of our TRLRF is similar to that of the two related algorithms. However, the desirable characteristic of rank selection robustness of our algorithm can help relieve the workload for model selection in practice, and thus the computational cost can be reduced. Moreover, though the computational complexity of TRLRF is of high power in $R$, due to the high representation ability and flexibility of TR decomposition, the TR-rank is always set as a small value. In addition, from experiments, we find out that our algorithm is capable of working efficiently for high-order tensors so that we can tensorize the data to a higher-order tensor and choose a small TR-rank to reduce the computational complexity. \section{Experimental Results} \subsection{Synthetic data} We first conducted experiments to testify the rank robustness of our algorithm by comparing TRALS, TRWOPT, and our TRLRF. To verify the performance of the three algorithms, we tested two tensors of size $20\times20\times20\times20$ and $7 \times8\times7\times8\times7\times8$. The tensors were generated by TR factors of TR-ranks $\{6,6,6,6\}$ and $\{4,4,4,4,4,4\}$ respectively. The values of the TR factors were drawn from i.i.d. Gaussian distribution $N \sim (0,0.5)$. The observed entries of the tensors were randomly removed by a missing rate of $0.5$, where the missing rate is calculated by $1-M/\text{num}(\tensor{T}_{real})$ and $M$ is the number of sampled entries (i.e., observed entries). We recorded the completion performance of the three algorithms by selecting different TR-ranks. The evaluation index was RSE which is defined by $\text{RSE}=\Vert \tensor{T}_{real}-\tensor{X} \Vert_F / \Vert \tensor{T}_{real}\Vert_F$, where $\tensor{T}_{real}$ is the known tensor with full observations and $\tensor{X}$ is the recovered tensor calculated by each tensor completion algorithm. The hyper-parameters of our TRLRF were set according to Algorithm 1. All the hyper-parameters of TRALS and TRWOPT are set according to the recommended settings in the corresponding papers to get the best results. \begin{figure}[htb] \begin{center} \includegraphics[width=0.9\linewidth]{rank_sensitive_new.pdf} \end{center} \caption{Completion performance of three TR-based algorithms in the synthetic data experiment. The RSE values of different selected TR-ranks are recorded. The missing rate of the two target tensors is $0.5$ and the real TR-ranks are 6 and 4 respectively.} \label{rs_fig} \end{figure} Figure \ref{rs_fig} shows the final RSE results which represent the average values of 100 independent experiments for each case. From the figure, we can see that all the three algorithms had their lowest RSE values when the real TR-ranks of the tensors were chosen and the best performance was obtained from our TRLRF. Moreover, when the TR-rank increased, the performance of TRLRF remained stable while the performance of the other two compared algorithms fell drastically. This indicates that imposing low-rankness assumption on the TR factors can bring robustness to rank selection, which largely alleviates the model selection problem in the experiments. \subsection{Benchmark images inpainting} \begin{figure}[htb] \begin{center} \includegraphics[width=0.7\linewidth]{benchmarks.pdf} \end{center} \caption{The eight benchmark images. The first image is named ``Lena" and is used in the next two experiments.} \label{bm_fig} \end{figure} \begin{figure}[htb] \begin{center} \includegraphics[width=0.95\linewidth]{img_rank_robust_new.pdf} \end{center} \caption{Visual completion results of the TRLRF (proposed), TRALS, and TRWOPT on image ``Lena'' with different TR-ranks, when the missing rate is $0.8$. The selected TR-ranks are 4, 6, 8, 10, 12 respectively, from the first column to the last column. The RSE results are noted under each picture.} \label{img_robust} \end{figure} \begin{figure}[htb] \begin{center} \includegraphics[width=1\linewidth]{img_compare_new.pdf} \end{center} \caption{Average completion performance of the eight considered algorithms, under different data missing rates.} \label{img_compare} \end{figure} In this section, we tested our TRLRF against the state-of-the-art algorithms on eight benchmark images which are shown in Figure \ref{bm_fig}. The size of each RGB image was $256\times\ 256 \times3$ which can be considered as a three-order tensor. For the first experiment, we continued to verify the TR-rank robustness of TRLRF on the image named ``Lena". Figure \ref{img_robust} shows the completion results of TRLRF, TRALS, and TRWOPT when different TR-ranks for each algorithm are selected. The missing rate of the image was set as $0.8$, which is the case that the TR decompositions are prone to overfitting. From the figure, we can see that our TRLRF gives better results than the other two TR-based algorithms in each case and the highest performance was obtained when the TR rank was set as 12. When TR-rank increases, the completion performance of TRALS and TRLRF decreases due to redundant model complexity and overfitting of the algorithms, while our TRLRF shows better results even the selected TR-rank is larger than the desired TR-rank. In the next experiment, we compared our TRLRF to the two TR-based algorithm, TRALS and TRWOPT, and the other state-of-the-art algorithms, i.e., TenALS \cite{jain2014provable}, FBCP \cite{zhao2015bayesian}, HaLRTC \cite{liu2013tensor}, TMac \cite{xu2013parallel} and t-SVD \cite{zhang2014novel}. We tested these algorithms on all the eight benchmark images and for different missing rates: $0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9$ and $0.95$. The relative square error (RSE) and peak signal-to-noise ratio (PSNR) were adopted for the evaluation of the completion performance. For RGB image data, PSNR is defined as $\text{PSNR}=10\log_{10}(255^2/\text{MSE})$ where MSE is calculated by $\text{MSE}=\Vert \tensor{T}_{real}-\tensor{X} \Vert_F^2/\text{num}(\tensor{T}_{real})$, and num($\cdot$) denotes the number of element of the fully observed tensor. For the three TR-based algorithms, we assumed the TR-ranks were equal for every core tensor (i.e., $R_1=R_2=\ldots=R_N$). The best completion results for each algorithm were obtained by selecting best TR-ranks for the TR-based algorithms by a cross-validation method. Actually, finding the best TR-rank to obtain the best completion results is very tedious. However, this is much easier for our proposed algorithm because the performance of TRLRF is fairly stable even though the TR-rank is selected from a wide large. For the other five compared algorithms, we tuned the hyper-parameters according to the suggestions of each paper to obtain the best completion results. Finally, we show the average performance of the eight images for each algorithm under different missing rates by line graphs. Figure \ref{img_compare} shows the RSE and PSNR results of each algorithm. The smaller RSE value and the larger PSNR value indicate the better performance. Our TRLRF performed the best among all the considered algorithms in most cases. When the missing rate increased, the completion results of all the algorithms decreased, especially when the missing rate was near $0.9$. The performance of most algorithm fell drastically when the missing rate was $0.95$. However, the performance of TRLRF, HaLRTC, and FBCP remained stable and the best performance was obtained from our TRLRF. \begin{table*}[h] \centering \caption{HSI completion results (RSE) under three different tensor orders with different rank selections} \label{hsi} \footnotesize \begin{tabular}{c|c|c|c|c|c|c|c|c} \hline & TRLRF & TRALS & TRWOPT & TMac & TenALS & t-SVD & FBCP &HaLRTC \\ \hline $\text{3-order}, \text{high-rank} (R_n=12)$&\textbf{0.06548} &0.07049 &0.06695 &0.1662&0.3448&0.4223 &0.2363&0.1254\\ $\text{3-order}, \text{low-rank} (R_n=8)$ &\textbf{0.1166} &0.1245 &0.1249 &0.2963&0.3312&- &-&-\\ $\text{5-order}, \text{high-rank} (R_n=22)$ &\textbf{0.1035}&0.1392&0.1200 &0.8064&-&0.9504 &0.3833&0.3944\\ $\text{5-order}, \text{low-rank} (R_n=18)$ &\textbf{0.1062}&0.1122&0.1072 &0.7411&-&- &-&-\\ $\text{8-order}, \text{high-rank} (R_n=24)$ &\textbf{0.1190}&0.1319 &0.1637&0.9487&-&0.9443 &0.4021 &0.9099\\ $\text{8-order}, \text{low-rank} (R_n=20)$ &\textbf{0.1421}&0.1581 &0.1767&0.9488&-&0.9450 &0.4135 &0.9097\\ \hline \end{tabular} \end{table*} \begin{figure*}[h] \begin{center} \includegraphics[width=0.95\linewidth]{hsi_img_new.pdf} \end{center} \caption{Completion results under the $0.9$ missing rate HSI data. The channels 80, 34, 9 are picked to show the visual results. The rank selection of TRLRF, TRALS, TRWOPT, TMac and TenALS are given under the corresponding images.} \label{hsi_img} \end{figure*} \subsection{Hyperspectral image} A hyperspectral image (HSI) of size $200\times\ 200\times80$ which records an area of the urban landscape was tested in this section\footnote{\url{http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes}}. In order to test the performance of TRLRF on higher-order tensors, the HSI data was reshaped to higher-order tensors, which is an easy way to find more low-rank features of the data. We compared our TRLRF to the other seven tensor completion algorithms in 3-order tensor ($200\times\ 200\times80$), 5-order tensor ($10\times 20\times 10 \times 20 \times 80$) and 8-order tensor ($8\times 5\times 5 \times 8 \times 5\times 5\times 8\times10$) cases. The higher-order tensors were generated from original HSI data by directly reshaping it to the specified size and order. The experiment aims to verify the completion performance of the eight algorithms under different model selection, whereby the experiment variables are the tensor order and tensor rank. The missing rates of all the cases are set as $0.9$. All the tuning parameters of every algorithm were set according to the statement in the previous experiments. Besides, for the experiments which need to set rank manually, we chose two different tensor ranks: high-rank and low-rank for algorithms. It should be noted that the CP-rank of TenALS and the Tucker-rank of TMac were set to the same values as TR-rank. The completion performance of RSE and visual results are listed in Table \ref{hsi} and shown in Figure \ref{hsi_img}. The results of FBCP, HaLRTC and t-SVD were not affected by tensor rank, so the cases of the same order with different rank are left blank in Table \ref{hsi}. The TenALS could not deal with tensor more than three-order, so the high-order tensor cases for TenALS are also left blank. As shown in Table \ref{hsi}, our TRLRF gives the best recovery performance for the HSI image. In the 3-order cases, the best performance was obtained when the TR-rank was 12, however, when the rank was set to 8, the performance of TRLRF, TRALS, TRWOPT, TMac, and TenALS failed because of the underfitting of the selected models. For 5-order cases, when the rank increased from 18 to 22, the performance of TRLRF kept steady while the performance of TRALS, TRWOPT, and TMac decreased. This is because the high-rank makes the models overfit while our TRLRF performs without any issues, owing to its inherent TR-rank robustness. In the 8-order tensor cases, similar properties can be obtained and our TRLRF also performed the best. \section{Conclusion} We have proposed an efficient and high-performance tensor completion algorithm based on TR decomposition, which employed low-rank constraints on the TR latent space. The model has been efficiently solved by the ADMM algorithm and it has been shown to effectively deal with model selection which is a common problem in most traditional tensor completion methods, thus providing much lower computational cost. The extensive experiments on both synthetic and real-world data have demonstrated that our algorithm outperforms the state-of-the-art algorithms. Furthermore, the proposed method is general enough to be extended to various other tensor decompositions in order to develop more efficient and robust algorithms. \bibliographystyle{aaai}

c82380c6afaf5a63734a2c4d8cee614c16f96dcc

\section{Introduction} For any positive integer $m$ with $m$ odd or $m$ divisible by 4, let $\zeta_m$ be a primitive $m$-th root of unity, $\QQ(\zeta_m)$ the corresponding cyclotomic field and $\QQ(\zeta_m)^+ = \QQ(\zeta_m + \zeta_m^{-1})$ its maximal totally real subfield. The units in $\QQ(\zeta_m)^+$ together with $\zeta_m$ are a subgroup of finite index in the group of units of $\QQ(\zeta_m)$ (this index is 1 when $m$ is a prime power and 2 otherwise, \cite[Corollary 4.13]{W}). Under the various $\phi(m)/2$ real embeddings of $\QQ(\zeta_m)^+$, each unit $\epsilon$ of $\QQ(\zeta_m)^+$ has a sign, and the collection of these signs is called the {\it signature} of $\epsilon$. The collection of all such unit signatures is an elementary abelian 2-group, whose rank is called the {\it unit signature rank} of $\QQ(\zeta_m)^+$ (or, by abuse, of $\QQ(\zeta_m)$). The signature rank measures how many different possible signs arise from the units and determines the difference between the class number and the strict class number. The purpose of this paper is to prove that the unit signature rank of $\QQ(\zeta_m)^+$ tends to infinity with $m$. We do this by demonstrating an explicit lower bound for the signature rank of the subgroup of {\it cyclotomic units} of $\QQ(\zeta_m)^+$. We then show that the difference between the signature rank of the units in the maximal real subfield of the cyclotomic field of $m p_1^{n_1} \dots p_s^{n_s}$-th roots of unity (with all $p_i$ odd) and its maximum possible value is bounded independent of $n_1, \dots , n_s$ (and is constant if all the $n_i$ are sufficiently large). This in particular provides infinitely many families of cyclotomic fields whose unit signature ranks are `nearly maximal'. Finally, we show this difference can be arbitrarily large, conditional on the existence of infinitely many cyclic cubic fields with totally positive fundamental units. \section{Signatures} For any totally real field $F$ of degree $n$ over $\QQ$ let $F_\RR=F \otimes_{\QQ} \RR \simeq \prod_{\text{$v$ real}} \RR$ where the product is taken over the real embeddings $v$ of $F$. Define the {\it archimedean signature space} $V_{\infty,F}$ of $F$ to be \begin{equation*} V_{\infty,F} = F_\RR^* / F_\RR^{*2} \simeq \prod_{\text{$v$ real}} \{\pm 1\} \simeq \FF_2^{n}, \end{equation*} where by identifying $\{\pm 1\}$ with the finite field $\FF_2$ of two elements we view the multiplicative group $V_{\infty,F}$ as a vector space over $\FF_2$, written additively. For any $\alpha \in F^*$ and $v:F \hookrightarrow \RR$ a real place of $F$, let $\alpha_v=v(\alpha)$ and define the sign of $\alpha_v$ as usual by $\sign(\alpha_v)=\alpha_v/\lvert \alpha_v \rvert \in \{\pm 1\}$. Write $\sgn(\alpha_v) \in \FF_2$ for $\sign(\alpha_v)$ when viewed in the additive group $\FF_2$, i.e., $\sgn(\alpha_v) = 0$ if $\alpha_v > 0$ and $\sgn(\alpha_v) = 1$ if $\alpha_v < 0$. The {\it (archimedean) signature map} of $F$ is the homomorphism \begin{equation*} \begin{aligned} \sgn_{\infty,F}: F^* &\to V_{\infty,F} \\ \alpha &\mapsto (\sgn(\alpha_v))_{v}. \end{aligned} \end{equation*} In the case when $F/\QQ$ is Galois and one real embedding of $F$ is fixed, we can index the real embeddings of $F$ by the elements $\sigma$ in $\Gal(F/\QQ)$, and \begin{equation*} \sgn_{\infty,F} (\alpha) = ( \sgn( \sigma( \alpha) ))_{\sigma \in \Gal(F/\QQ)} , \end{equation*} where $\sgn$ is the sign (viewed as an element of $\FF_2$) in the fixed real embedding. The element $\sgn_{\infty,F} (\alpha)$ is called the {\it signature} of $\alpha$. The collection of all the signatures $\sgn_{\infty,F} (\epsilon)$ where $\epsilon$ varies over the units of $F$ is called the {\it unit signature group} of $F$; the rank of this subspace of $V_{\infty,F}$ is called the {\it (unit) signature rank} of $F$ and, as previously mentioned, is a measure of how many different possible sign configurations arise from the units of $F$. Define the {\it (unit signature rank) ``deficiency'' of $F$}, denoted $\delta(F)$, to be the corank of the unit signature group of $F$ in $V_{\infty,F}$, i.e., $[F:\QQ]$ minus the signature rank of the units of $F$. The deficiency of $F$ is just the nonnegative difference between the unit signature rank of $F$ and its maximum possible value---the deficiency is 0 if and only if there are units of every possible signature type. The deficiency is also the rank of the group of totally positive units of $F$ modulo squares. \begin{remark}\label{rem:deficiencies} For any finite extension $L/F$ of totally real fields we have $\delta (F) \le \delta(L)$, a result due to Edgar, Mollin and Peterson (\cite[Theorem 2.1]{EMP}). We briefly recall the reason: the intersection of the Hilbert class field $H_L$ of $L$ with the strict (or narrow) Hilbert class field $H_F^{\textup{st}}$ of $F$ is easily seen to be the Hilbert class field $H_F$ of $F$ since $L$ is totally real. The composite field $H_L H_F^{\textup{st}}$ is a subfield of the strict Hilbert class field, $H_L^{\textup{st}}$, of $L$ and has degree over $H_L$ equal to $[H_F^{\textup{st}} : H_F]$ because $ H_F = H_L \cap H_F^{\textup{st}} $. Since $[H_F^{\textup{st}} : H_F] = 2^{\delta (F)}$ and $[H_L^{\textup{st}} : H_L] = 2^{\delta (L)}$ (see \cite[\S 2]{D-V} for details), the result follows. $$ \beginpicture \setcoordinatesystem units <1 pt, 1 pt> \linethickness= 0.3pt \put {$F$} at 0 0 \put {$L$} at 0 50 \put {$H_F$} at 50 25 \put {$H_L$} at 50 75 \put { ($ = H_L \cap H_F^{\textup{st}} $) } at 100 25 \put {$H_F^{\textup{st}}$} at 100 50 \put {$H_L H_F^{\textup{st}}$} at 100 100 \put {$H_L^{\textup{st}}$} at 155 125 \putrule from 0 10 to 0 40 \putrule from 50 35 to 50 65 \putrule from 100 60 to 100 90 \setlinear \plot 7.5 5.0 40.0 20.0 / \setlinear \plot 7.5 55.0 40.0 70.0 / \setlinear \plot 57.5 30.0 90.0 45.0 / \setlinear \plot 57.5 80.0 85.0 92.7 / \setlinear \plot 117 107.5 145.0 120.4 / \endpicture $$ \end{remark} \section{Circular Units and Signatures in Cyclotomic Fields} Suppose now that $F = \QQ(\zeta_m )^+$ is the maximal (totally) real subfield of the cyclotomic field $\QQ(\zeta_m)$ of $m$-th roots of unity (with $m$ odd or divisible by 4), which is of degree $\phi(m)/2$ over $\QQ$. We can fix an embedding of $\QQ(\zeta_m)$ into $\CC$ by mapping $\zeta_m$ to $e^{2 \pi i /m}$, which also fixes an embedding of $\QQ(\zeta_m)^+$ into $\RR$. The Galois group $\Gal (\QQ(\zeta_m)/\QQ)$ consists of the automorphisms $\sigma_a$ that map $\zeta_m$ to $\zeta_m^a$ for integers $a$, $1 \le a < m$ relatively prime to $m$. We identify $\Gal (\QQ(\zeta_m)^+/\QQ)$ with $\Gal (\QQ(\zeta_m)/\QQ) / \{ 1, \sigma_{-1} \}$ and take as representatives the elements $ \overline{\sigma_a} \in \Gal (\QQ(\zeta_m)^+/\QQ)$ with $1 \le a < m/2$ relatively prime to $m$. To get information on the unit signature rank of $\QQ(\zeta_m)^+$ we consider the signatures of the subgroup of {\it circular} (or {\it cyclotomic}) {\it units}, denoted $C_{\QQ(\zeta_m )}$, which when $m$ is a prime power has a set of generators given by $-1$ and the $\phi(m)/2 -1 $ independent elements \begin{equation} \label{eq:circunits} \zeta_m^{(1-a)/2} \dfrac{1 - \zeta_m^a}{1 - \zeta_m} \end{equation} for $1 < a < m/2$ and $a$ relatively prime to $m$ \cite[Lemma 8.1]{W} (for $m$ not a prime power the definition of $C_{\QQ(\zeta_m )}$ is more complicated, see \cite[Chapter 8]{W}). The group $C_{\QQ(\zeta_m )}$ is contained in $\QQ(\zeta_m)^+$ and when $m$ is a prime power is isomorphic (as an additive abelian group) to $\ZZ / 2 \ZZ \times \ZZ^{\phi(m)/2 - 1}$, with $-1$ and the elements in \eqref{eq:circunits} serving as independent generators over $\ZZ$. If we choose a particular ordering of the elements of $\Gal(\QQ(\zeta_m)^+ / \QQ)$ and an ordering of a set of generators for the cyclotomic units, then their corresponding signatures in $\FF_2^{\phi(m)/2}$ define the rows of a $\phi(m)/2 \times \phi(m)/2$ matrix over $\FF_2$, whose rank is the rank of the subgroup $ \sgn_{\infty,\QQ(\zeta_m)^+} (C_{\QQ(\zeta_m )})$ of signatures of the cyclotomic units. We first consider the case where $m$ is an odd prime power. \subsection*{The case ${\bf m = p^n}$ for an odd prime ${\bf p}$} When $m = p^n$ for an odd prime $p$ every primitive $m$-th root of unity is the square of another primitive $m$-th root of unity, so the circular units in \eqref{eq:circunits} can be written in the form \begin{equation} \label{eq:circunitsoddp} \dfrac{\zeta_m^a - \zeta_m^{-a}}{\zeta_m - \zeta_m^{-1}} \end{equation} for $1 < a < m/2$ and $a$ relatively prime to $m$. If we order the elements in $ \Gal (\QQ(\zeta_m)^+/\QQ) $ by $ \overline{\sigma_b}$ with $1 \le b < m/2$ relatively prime to $m$, then the signature of the element in \eqref{eq:circunitsoddp} is given by the signs of the elements \begin{equation*} \overline{\sigma_b} \left ( \dfrac{\zeta_m^a - \zeta_m^{-a}}{\zeta_m - \zeta_m^{-1}} \right ) = \dfrac{\zeta_m^{ab} - \zeta_m^{-ab}}{\zeta_m^b - \zeta_m^{-b}} , \quad 1 \le b < m/2, \text{$(b,m) = 1$} . \end{equation*} Under the embedding $\zeta_m = e^{2 \pi i /m}$ we have \begin{equation} \label{eq:circunitsab} \dfrac{\zeta_m^{ab} - \zeta_m^{-ab}}{\zeta_m^b - \zeta_m^{-b}} = \dfrac{ \sin (2 \pi a b /m) } { \sin (2 \pi b /m) }. \end{equation} Since $1 \le b < m/2$, the denominator $\sin (2 \pi b /m) $ is positive. Hence $\overline{\sigma_b} ( ( \zeta_m^a - \zeta_m^{-a} )/( \zeta_m - \zeta_m^{-1} ))$ is positive if and only if $ \sin (2 \pi a b /m) $ is positive, which happens precisely when the least positive residue of $a b$ modulo $m$ is in $(0, m/2)$. It follows that the rank of the subspace $ \sgn_{\infty,{\QQ(\zeta_m )}^+} (C_{\QQ(\zeta_m )})$ of $V_{\infty,\QQ(\zeta_m)^+}$ is equal to the rank of a $\phi(m)/2 \times \phi(m)/2$ matrix $C = (c_{a,b})$ over $\FF_2$ (referred to as the {\it circular unit signature matrix}) whose rows are indexed by the elements $a$ relatively prime to $m$ with $1 \le a < m/2 $ and whose columns are indexed by the elements $b$ relatively prime to $m$ with $1 \le b < m/2 $, as follows. The first row of $C$ is $(1,1, \dots, 1)$, corresponding to the signs $(-1,-1, \dots, -1)$ of the element $-1$, viewed in the additive group $\FF_2$, so \begin{equation*} c_{1,b} = 1 , \end{equation*} for $1 \le b < m/2 $, $b$ relatively prime to $m$. For $2 \le a < m/2 $, $1 \le b < m/2$, $a$ and $b$ relatively prime to $m$, we have \begin{equation*} c_{a,b} = \begin{cases} 0 & \text{if $ a b \ (\text{mod } m) \in (0, m/2)$ } \\ 1 & \text{if $ a b \ (\text{mod } m) \in (m/2, m)$ } \\ \end{cases} , \end{equation*} where $a b \ (\text{mod } m)$ is taken to be the least positive residue of $a b$ modulo $m$. For computational purposes, it is useful to note the row indexed by $a$ is the $i$-th row of the matrix where $i = a - \lfloor (a-1)/p \rfloor$ and the $i$-th row is indexed by $a = i + \lfloor (i-1)/ (p-1) \rfloor$ (and similarly for the numbering of the columns). If we add the first row of $C$ to the remaining rows (which does not affect the rank of the matrix), we obtain the {\it modified circular unit signature matrix} $M = (c_{a,b}')$ with \begin{equation*} c_{a,b}' = \begin{cases} 1 & \text{if $ a b \ (\text{mod } m) \in (0, m/2)$ , and } \\ 0 & \text{if $ a b \ (\text{mod } m) \in (m/2, m)$ ,} \\ \end{cases} \end{equation*} for $1 \le a < m/2 $ and $1 \le b < m/2$, $a$ and $b$ relatively prime to $m$ and as before $a b \ (\text{mod } m)$ is taken to be the least positive residue of $a b$ modulo $m$ (this matrix appears in \cite{Da} in the case when $m$ is an odd prime). The entry $c_{2^d,b}'$ of the matrix $M$ in the column indexed by $b$ ($b = 1, \dots, m/2$, $b$ prime to $m$) and the row indexed by $2^d$ ($1 \le 2^d < m/2$ ) is 1 if the least positive residue of $2^d b$ modulo $m$ lies in $(0, m/2)$ and is 0 if the least positive residue lies in $(m/2,m)$. Writing $ 2^d b = A m + r $ with an integer $A$ and least positive remainder $r$ with $0 \le r < m$, i.e., $ 2^{d+1} b / m = 2A + (2 r /m)$, it follows that $ r \in (0,m/2)$ implies $\lfloor 2^{d+1} b / m \rfloor = 2A$ and $ r \in (m/2,m)$ implies $\lfloor 2^{d+1} b / m \rfloor = 2A +1$. As a consequence, the entry of $M$ in the row indexed by $2^d$ and column indexed by $b$ is 0 if $ \lfloor 2^{d+1} b / m \rfloor $ is odd and is 1 if $ \lfloor 2^{d+1} b / m \rfloor $ is even. \begin{nonumlemma} \label{lem:indeppcols} Suppose $m = p^n$ where $p$ is an odd prime and $n \ge 1$. Let $k \ge 1$ be any integer with $ m > 2^{k+2} $. If $b_0(k) = \lfloor \dfrac{(2^k - 2)m}{2^{k+1}} \rfloor + 1$ and $b_1(k) = \lfloor \dfrac{(2^k - 2)m}{2^{k+1}} \rfloor + 2$, then $ \lfloor 2^{d+1} b_0(k) / m \rfloor $ and $ \lfloor 2^{d+1} b_1(k) / m \rfloor $ are both odd for $ d = 1,2, \dots, k-1$ and both even for $d = k$. \end{nonumlemma} \begin{proof} Write \begin{equation*} \dfrac{(2^k - 2)m}{2^{k+1}} = \lfloor \dfrac{(2^k - 2)m}{2^{k+1}} \rfloor + \theta \end{equation*} where $0 \le \theta < 1$. Then \begin{equation*} b_0(k) = \dfrac{2^k - 2}{2^{k+1}} m + (1 -\theta) \quad \text{and} \quad b_1(k) = \dfrac{2^k - 2}{2^{k+1}} m + (2 -\theta), \end{equation*} so \begin{equation*} \dfrac{ 2^{d+1} b_0(k)}{m} = 2^d - \dfrac{2}{2^{k-d}} + \dfrac{2^{d+1}}{m} (1 -\theta) \quad \text{and} \quad \dfrac{ 2^{d+1} b_1(k)}{m} = 2^d - \dfrac{2}{2^{k-d}} + \dfrac{2^{d+1}}{m} (2 -\theta). \end{equation*} Since $0 < 1 - \theta \le 1$, $1 < 2 - \theta \le 2$, and $m > 2^{k+2}$, when $d = k$ this gives \begin{equation*} \lfloor 2^{k+1} b_0(k) / m \rfloor = \lfloor 2^{k+1} b_1(k) / m \rfloor = 2^k - 2, \end{equation*} so $ \lfloor 2^{k+1} b_0(k) / m \rfloor $ and $ \lfloor 2^{k+1} b_1(k) / m \rfloor $ are both even. For $ 1 \le d < k$, we have \begin{equation*} 2^d - 1 < 2^d - \dfrac{2}{2^{k-d}} + \dfrac{2^{d+1}}{m} (1 -\theta) < 2^d - \dfrac{2}{2^{k-d}} + \dfrac{2^{d+1}}{m} (2 -\theta) < 2^d , \end{equation*} so that \begin{equation*} \lfloor 2^{d+1} b_0(k) / m \rfloor = \lfloor 2^{d+1} b_1(k) / m \rfloor = 2^d - 1, \end{equation*} and $ \lfloor 2^{d+1} b_0(k) / m \rfloor $ and $ \lfloor 2^{d+1} b_1(k) / m \rfloor $ are both odd, completing the proof of the lemma. \end{proof} We can use the lemma to give the following lower bound for the number of independent signatures for the circular units in this case. \begin{proposition} \label{prop:pcase} Suppose $p$ is an odd prime and $n \ge 1$. Then the rank of the group of signatures of the circular units in $\QQ(\zeta_{p^n})^+$ is at least $\lfloor \log_2 (p^n) \rfloor - 2$. \end{proposition} \begin{proof} Since $b_0(k)$ and $b_1(k)$ in the lemma differ by 1 and $m = p^n$, at least one is relatively prime to $m$. It then follows from the lemma that for each $k \ge 1$ with $2^{k+2} < m$ there is a $B(k)$, relatively prime to $m$ and satisfying $1 \le B(k) < m/2$, such that $ \lfloor 2^{d+1} B(k) / m \rfloor $ is odd for $ d = 1,2, \dots, k-1$ and even for $d = k$. By the remarks before the lemma, it follows that for each $k \ge 1$ with $2^{k+2} < m$, the entries of the matrix $M$ in the column indexed by $B(k)$ and belonging to the rows indexed by $2,4,8,\dots, 2^k$ are $0,0,\dots,0,1$, respectively. In particular, this shows that the row of $M$ indexed by $2^k$ is not in the span of the rows indexed by $2,4,\dots,2^{k-1}$. Applying this successively for $k = 1,2, \dots, \lfloor \log_2 m \rfloor - 2 $ shows that all these rows are linearly independent, implying that the rank of $M$ is at least $\lfloor \log_2 m \rfloor - 2$, which proves the proposition. \end{proof} \subsection*{The case ${\bf m = 2^n}$, ${\bf n \ge 2}$} In this case let $\zeta_{2^{n+1}}$ be a $2^{n+1}$-st root of unity with $\zeta_{2^{n}} = {\zeta_{2^{n+1}}^2}$. Fix an embedding into $\CC$ mapping $\zeta_{2^{n+1}}$ to $e^{2 \pi i /2^{n+1}}$ and order the elements of $ \Gal ({\QQ(\zeta_m )}^+/\QQ) $ as $ \overline{\sigma_b}$ with $b = 3,5, \dots, 2^{n-1} - 1$, $b$ odd. Then the conjugates of the elements in \eqref{eq:circunits} are given by \begin{equation} \label{eq:2powercircunits} \dfrac{ \zeta_{2^{n+1}}^{ab} - \zeta_{2^{n+1}}^{-ab} } { \zeta_{2^{n+1}}^{b} - \zeta_{2^{n+1}}^{-b} } = \dfrac{\sin ( \pi a b /2^{n} ) } {\sin ( \pi b /2^{n} ) } \end{equation} for $1 < a < 2^{n-1}$, $1 \le b < 2^{n-1}$, with $a$ and $b$ both odd. The sign of the unit in \eqref{eq:2powercircunits} is $+1$ if the least positive residue of $a b$ modulo $2^{n+1}$ lies in $(0,2^n)$ and is $-1$ if the least positive residue lies in $(2^n, 2^{n+1})$. In this case the circular unit signature matrix has first row consisting of all 1's and the entry in the row indexed by $a$ and column indexed by $b$ is given by \begin{equation} \label{eq:sigmatrix2one} c_{1,b} = 1 \end{equation} for $b$ odd, $1 < b < 2^{n-1}$, and \begin{equation} \label{eq:sigmatrix2two} c_{a, b} = \begin{cases} 0 & \text{if $ a b \ (\text{mod } 2^{n+1}) \in (0,2^n)$ , and } \\ 1 & \text{if $ a b \ (\text{mod } 2^{n+1}) \in (2^n, 2^{n+1})$ } \\ \end{cases} , \end{equation} for $a$ and $b$ odd, $1 < a < 2^{n-1} $ and $1 \le b < 2^{n-2} $, where $a b \ (\text{mod } 2^{n+1})$ is taken to be the least positive residue of $a b$ modulo $2^{n+1}$. An argument similar to that for odd prime powers (here for the rows indexed by $2^d - 1$ and column indexed by $2^n - 2^{n-k+1} + 1$) shows the rank of the circular unit signature matrix is at least $n - 2 = \lfloor \log_2 m \rfloor - 2$, but for this case Weber in 1899 proved the following definitive result. \begin{proposition} \label{prop:2case} (Weber, \cite[B, p.~821]{We}) Suppose $n \ge 2$. Then the rank of the group of signatures of the circular units in the maximal totally real subfield of the cyclotomic field of $2^n$-th roots of unity is maximal, i.e., equal to $2^{n-2}$. \end{proposition} This result was generalized by Hasse in \cite{Ha}, whose simpler and more conceptual proof used the fact that the circular unit $( \zeta_{2^{n+1}}^{5} - \zeta_{2^{n+1}}^{-5} ) / ( \zeta_{2^{n+1}} - \zeta_{2^{n+1}}^{-1} ) $ in ${\QQ(\zeta_{2^n} )}^+$ has norm $-1$ and showed the signatures of its Galois conjugates generate a group of maximal signature rank (see \cite[pp.~28-29]{Ha}). A nice proof of this, using the fact that over $\FF_2$ the only irreducible representation of a 2-group is the trivial representation, can be found in \cite{Ga}. Another nice proof of Weber's result can be found in \cite{D}. \begin{remark} Unlike the situation for the 2-power cyclotomic fields, not every possible signature type occurs for the circular units in general cyclotomic fields, even for $m = p$ an odd prime (for example, in the case $p = 29 $ the circular unit signature group has rank 11, not the maximal possible rank of 14 \cite[Appendix I, p.~ 70]{Da}). \end{remark} \section{Signatures in composites of extensions} Propositions \ref{prop:pcase} and \ref{prop:2case} already imply that the signature rank of the units of ${\QQ(\zeta_m )}^+$ tends to infinity with $m$ (since ${\QQ(\zeta_{p^n} )}^+ \subset {\QQ(\zeta_{m} )}^+$ if $p^n$ divides $m$ and the largest prime power divisor of $m$ tends to infinity as $m$ tends to infinity), but this can be made more precise using the following result, which may be of independent interest. Suppose $F/\QQ$ and $F'/\QQ$ are two totally real Galois extensions of $\QQ$ with $F \cap F' = \QQ$. Then the composite field $F F'$ has Galois group $\Gal(FF'/\QQ) = \Gal(F/\QQ) \times \Gal(F'/\QQ)$, and independent signatures in $F$ and $F'$ produce essentially independent signatures in $F F'$: \begin{proposition} \label{prop:compositeranks} With $F$ and $F'$ as above, suppose $\alpha_1, \dots, \alpha_r$ are nonzero elements of $F$ whose signatures $\sgn_{\infty,F}(\alpha_i)$, $i = 1,\dots,r$ are linearly independent in the $\FF_2$-vector space $V_{\infty,F}$. Suppose similarly that $\beta_1, \dots, \beta_s$ are nonzero elements of $F'$ whose signatures $\sgn_{\infty,F'}(\beta_j)$, $j = 1, \dots, s$, are linearly independent in the $\FF_2$-vector space $V_{\infty,F'}$. Then the dimension of the space generated by the signatures of $\alpha_1, \dots, \alpha_r,\beta_1, \dots, \beta_s$ in the $\FF_2$-vector space $V_{\infty,FF'}$ is $r + s$ unless the subgroups generated by $\alpha_1, \dots, \alpha_r$ and by $\beta_1, \dots, \beta_s$ both contain totally negative elements, in which case the dimension is $r + s - 1$. \end{proposition} \begin{proof} If the signatures in $V_{\infty,FF'}$ of $\alpha_1, \dots, \alpha_r, \beta_1, \dots, \beta_s$ are linearly dependent, then there is an element ${\alpha_1}^{a_1} \dots {\alpha_r}^{a_r} {\beta_1}^{b_1} \dots {\beta_s}^{b_s}$ in $FF'$ with $a_1,\dots,a_r,b_1, \dots, b_s \in \{0,1\}$, not all 0, that is totally positive. Assume without loss that at least one of $a_1,\dots,a_r$ is not 0, and let $\alpha = {\alpha_1}^{a_1} \dots {\alpha_r}^{a_r}$ and $\beta = {\beta_1}^{b_1} \dots {\beta_s}^{b_s}$. Then $\sigma \tau (\alpha \beta) = \sigma (\alpha) \tau (\beta)$ is positive for every $\sigma \in \Gal(F/\QQ)$ and every $\tau \in \Gal(F'/\QQ)$. Since the signatures $\sgn_{\infty,F}(\alpha_i)$, $i = 1,\dots,r$ are linearly independent and some $a_i$ is nonzero, there exists a $\sigma_0$ such that $\sigma_0 (\alpha)$ is negative. This implies $\tau (\beta)$ is negative for every $\tau$, i.e., $\beta$ is totally negative. Then, since there is a $\tau_0$ with $\tau_0 (\beta)$ negative, it follows that also $\sigma (\alpha)$ is negative for every $\sigma$, i.e., $\alpha$ is totally negative. Hence there is at most one nontrivial relation among the signatures of $\alpha_1, \dots, \alpha_r, \beta_1, \dots, \beta_s$ in $V_{\infty,FF'}$---this nontrivial relation occurs if and only if $(1,1,\dots,1) \in V_{\infty,F}$ is in the $\FF_2$-space spanned by $\sgn_{\infty,F}(\alpha_i)$, $i = 1,\dots,r$ and $(1,1,\dots,1) \in V_{\infty,F'}$ is in the $\FF_2$-space spanned by $\sgn_{\infty,F'}(\beta_j)$, $j = 1,\dots,s$, completing the proof. \end{proof} \section{The signature rank of the units in $\QQ(\zeta_m)$ for general $m$} If we combine Propositions \ref{prop:pcase} and \ref{prop:2case} on the signature ranks in the prime power case with Proposition \ref{prop:compositeranks} we obtain the following result. \begin{theorem} \label{thm:1} Suppose the positive integer $m$ is odd or is divisible by 4. Then the rank of the group of signatures of the group of circular units in $\QQ(\zeta_m)^+$ is at least $\log_2 (m) - 4 \omega (m) +1 $, where $\omega(m)$ is the number of distinct prime factors of $m$. In particular, the signature rank of the units in $\QQ(\zeta_m)^+$ tends to infinity with $m$. \end{theorem} \begin{proof} Write $m = p_1^{a_1} \dots p_k^{a_k}$. Then $\QQ(\zeta_m)^+$ contains the composite of the totally real fields $\QQ( \zeta_{{p_i}^{a_i}} )^+$, $i = 1, \dots, k$. By Propositions \ref{prop:pcase} and \ref{prop:2case}, the signature rank of the circular units in these latter fields is at least $\lfloor \log_2 ({p_i}^{a_i}) \rfloor - 2$ (and much better when $p = 2$). Applying the previous proposition repeatedly shows that the signature rank of the group generated by the circular units is at least \begin{equation*} \sum_{i=1}^k \big ( \lfloor \log_2 ({p_i}^{a_i}) \rfloor - 2 \big ) - (k-1). \end{equation*} Since $\lfloor \log_2 ({p_i}^{a_i}) \rfloor > \log_2 ({p_i}^{a_i}) - 1$, the lower bound in the theorem follows. The final statement in the theorem follows from standard bounds on the growth of $\omega(m)$ (see, for example, \cite[Section 22.10]{H-W}). \end{proof} \begin{corollary} With the exception of $m = 12$, no maximal real subfield of any cyclotomic field of $m$-th roots of unity has a fundamental system of units that are all totally positive. \end{corollary} \begin{proof} The signature rank of the circular units is at least 2 for $m = 2^3, 3^2, 5, 7, 11, 13$ by direct computation and for all $m = p$ for primes $p \ge 17$ by Proposition \ref{prop:pcase}. It follows that the signature rank of the circular units is at least 2 for all $m$ divisible by $2^3, 3^2$ or any odd prime $p \ge 5$. The only remaining possible values for $m$ are $m = 3, 4, 12$. The first two have no units of infinite order, and the third has maximal real subfield $\QQ(\sqrt{3})$ with totally positive fundamental unit $2 + \sqrt 3$. \end{proof} \section{Signatures in Cyclotomic Towers over Cyclotomic Fields} \label{sec:towers} Computations suggest that the signature rank of the units in the real subfield of the cyclotomic field of $m$-th roots of unity is in fact always close to the maximal possible rank of $\phi(m)/2$ (equivalently, the unit signature rank deficiency for these fields should be close to 0), i.e., nearly all possible signature types arise for units. This is in keeping with the heuristics in \cite{D-V} suggesting that `most' totally real fields have nearly maximal unit signature rank (although these abelian extensions are hardly `typical'). In this section we prove that for infinitely many different families of cyclotomic fields the units do indeed have nearly maximal signature rank. We do this by showing the unit signature rank deficiency is bounded in (finitely many composites of) prime power cyclotomic towers over cyclotomic fields. \begin{theorem} \label{thm:2} Suppose $p_1, \dots , p_s$ ($s \ge 1$) are distinct odd primes and suppose $m$ is any positive integer that is either odd or divisible by 4 and that is relatively prime to $p_1, \dots , p_s$. Let $\delta(m; n_1, \dots , n_s) = \delta( \QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})^+ ) \ge 0$ denote the unit signature rank deficiency of the maximal real subfield of the cyclotomic field of $m p_1^{n_1} \dots p_s^{n_s}$-th roots of unity defined in \S 2, i.e., the nonnegative difference between the signature rank of the units of $\QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})^+ $ and its maximum possible value $\phi( m p_1^{n_1} \dots p_s^{n_s})/2$. Then \begin{enumerate} \item[(a)] $\delta(m; n_1, \dots , n_s) \le \delta(m; n'_1, \dots , n'_s)$ if $n_i \le n'_i$ for all $1 \le i \le s$, \item[(b)] $\delta(m; n_1, \dots , n_s)$ is bounded independent of $n_1, \dots , n_s$, and \item[(c)] $\delta(m; n_1, \dots , n_s)$ is constant (depending on $m$) if $n_1, \dots , n_s$ are all sufficiently large. \end{enumerate} \end{theorem} \begin{proof} The maximal real subfield of the cyclotomic field of $m p_1^{n_1} \dots p_s^{n_s}$-th roots of unity is a subfield of the maximal real subfield of the cyclotomic field of $m p_1^{n'_1} \dots p_s^{n'_s}$-th roots of unity if $n_i \le n'_i$, $i = 1, \dots ,s$, so (a) follows immediately from Remark \ref{rem:deficiencies}. Suppose that $n_i \ge 1$ for $1 \le i \le s$ and let $L$ denote the cyclotomic field $\QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})$, with maximal real subfield $L^+ = \QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})^+ $. Then the strict class number of $L^+ $ divides twice the class number of $L$, as follows. The quadratic extension $L/L^+$ is ramified at a finite prime (if $m = s = 1$) or is ramified only at infinity (otherwise). Hence, if $H_{L^+}^{\textup{st}}$ is the strict Hilbert class field of $L^+$, then the degree over $L$ of the composite $L H_{L^+}^{\textup{st}}$ is either the strict class class number of $L^+$ (if $m = s = 1$) or half that (otherwise). Since $L H_{L^+}^{\textup{st}}$ is an abelian extension of $L$ that is unramified at finite primes, it is contained in the Hilbert class field of the complex field $L$, so $ [ L H_{L^+}^{\textup{st}} : L]$ divides the class number of $L$, which gives the desired divisibility. Next observe that the cyclotomic fields $\QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})$ with $n_i \ge 1$ for $1 \le i \le s$ are the subfields of the composite of the cyclotomic $\ZZ_{p_i}$-extensions of $\QQ(\zeta_{m p_1 \dots p_s})$ for $1 \le i \le s$. Hence the 2-primary part of the class number of $\QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})$ is bounded for all $s$-tuples $(n_1, \dots , n_s)$ and is constant if $n_1, \dots , n_s$ are all sufficiently large by a theorem of Friedman (\cite{F}) extending a result of Washington. By the previous observation, the strict class number of $L^+ = \QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})^+ $ is therefore bounded for all $s$-tuples $(n_1, \dots , n_s)$ (and is in fact constant if $n_1, \dots , n_s$ are all sufficiently large). Finally, since the strict class number of $L^+$ is the product of the usual class number of $L^+$ with $2^{\delta(L^+)}$, we obtain (b). Then by (a), we obtain (c). \end{proof} \begin{corollary} \label{cor:cor1} With notation as in Theorem \ref{thm:2}, the unit signature rank deficiencies for all totally real abelian fields $F$ whose conductor is a product of a divisor of $m$ with an integer whose prime divisors are among the set $\{ p_1, \dots , p_s \}$, are uniformly bounded. \end{corollary} \begin{proof} Any totally real abelian field $F$ having conductor $d p_1^{n_1} \dots p_s^{n_s}$ where $d$ is a divisor of $m$ and with $n_i \ge 0$ for $1 \le i \le s$ is contained in $\QQ(\zeta_{m p_1^{n_1} \dots p_s^{n_s}})^+ $. Hence $\delta(F) \le \delta(m; n_1,\dots,n_s)$ by Remark \ref{rem:deficiencies}, and the result follows immediately from Theorem \ref{thm:2}. \end{proof} \begin{remark} Corollary \ref{cor:cor1} shows that among all the abelian fields whose conductors are supported in a fixed finite set of primes, almost all have nearly maximal unit signature rank (in the precise sense that the signature rank deficiencies are uniformly bounded by a constant depending only on the set of primes chosen). \end{remark} We highlight some particular special cases: \begin{corollary} \label{cor:basicZp} Let $k$ be a finite totally real abelian extension of $\QQ$. For $p$ an odd prime, let $k^{p,\infty}$ denote the cyclotomic $\ZZ_p$-extension of $k$. If $k_n$ is the subfield of $k^{p,\infty}$ of degree $p^n$ over $k$, then the signature rank of the units of $k_n$ differs from $[k_n:\QQ]$ by a constant amount for $k_n$ sufficiently far up the tower. \end{corollary} Applying Corollary \ref{cor:basicZp} to $k = \QQ(\zeta_p)^+$ for $p$ odd, together with Weber's result in Proposition \ref{prop:2case}, gives the following. \begin{corollary} \label{cor:primepower} For any prime $p$, the difference between the signature rank of the units of $\QQ(\zeta_{p^n})^+$ and $\phi(p^n)/2$ is constant for $n$ sufficiently large (the constant depending on $p$). \end{corollary} Corollary \ref{cor:primepower} gives another proof of the results in \S 3 that the signature ranks of the units in the fields $\QQ(\zeta_{p^n})^+$ tend to infinity as $n$ tends to infinity. Corollary \ref{cor:primepower} is far superior, asymptotically, to the results in \S 3 for the cyclotomic fields of odd prime power conductor since it shows the signature rank is `nearly' maximal. The result yields relatively little information for any specific $\QQ(\zeta_{p^N})^+$, however, since the unit signature rank deficiency for this field could conceivably be close to $\phi(p^N)/2$ (although, as mentioned, this is not expected to happen). The only explicit lower bounds for the unit signature rank for $\QQ(\zeta_{p^n})$ for odd $p$ and, more significantly, for general $\QQ(\zeta_{m})$ (for example if $m$ is the product of distinct primes, for which the results in this section have little to say), are those in \S 3. \begin{remark} We have done some computations of the signature ranks for the subgroup of circular units in towers of prime power cyclotomic fields. While the computations are somewhat modest (since $\phi(p^n)$ grows rapidly with $n$), these computations have exhibited the following behavior: if the signature rank of the circular units in $\QQ(\zeta_p)^+$ is $\frac12 \phi(p) - \delta$ ($\delta \ge 0$), then the signature rank of the circular units in the fields $\QQ(\zeta_{p^n})^+$ is $\frac12 \phi(p^n) - \delta$, i.e., the circular unit signature rank deficiency is constant and equal to its value in the first layer. Whether this behavior persists in general is an extremely interesting question. We also note that the deficiency of the circular units is at least the deficiency for the full group of units, but may be strictly larger: for the field $\QQ(\zeta_{163})^+$ the circular unit deficiency is 2, while the deficiency for the full group of units is 0 (see \cite{D}). \end{remark} \section{Unit Signature Rank Deficiencies in Cyclotomic Fields}\label{sec:unbounded} In this section we show that the signature rank deficiency in the maximal real subfields of cyclotomic fields can be arbitrarily large, under the assumption that there exist infinitely many cyclic cubic extensions having a system of totally positive fundamental units. Suppose $k$ is a cyclic cubic extension of $\QQ$ with totally positive fundamental units $\epsilon_1, \epsilon_2$. If $E_k$ is the unit group of $k$, then $E_k = \{ \pm 1 \} \times \langle \epsilon_1, \epsilon_2 \rangle$ and the subgroup $\langle \epsilon_1, \epsilon_2 \rangle$ consists of the totally positive units in $k$. If the Galois group $G$ of $k$ is generated by $\sigma$, then $\langle \epsilon_1, \epsilon_2 \rangle$ is a module for the quotient $\ZZ[G]/(\sigma^2 + \sigma + 1)$ of the group ring $\ZZ[G]$ of $G$ since $\epsilon_1$ and $\epsilon_2$ both have norm $+1$. This quotient of the group ring is isomorphic to the ring of integers in $\QQ(\sqrt{-3})$, which is a principal ideal domain, and it follows that $\langle \epsilon_1, \epsilon_2 \rangle \iso \ZZ[G]/(\sigma^2 + \sigma + 1)$ as $G$-modules (and not just as abelian groups). Modulo squares, $\langle \epsilon_1, \epsilon_2 \rangle$ is therefore isomorphic to $\FF_2[G]/(\sigma^2 + \sigma + 1)$ as a module over the group ring $\FF_2 [G]$, hence affords the unique irreducible 2-dimensional representation of $\FF_2 [G]$. In particular, $G$ acts irreducibly and with no nontrivial fixed points on $\langle \epsilon_1, \epsilon_2 \rangle$ modulo squares. With these preliminaries, we consider the composite of cyclic cubic fields having a system of totally positive fundamental units: \begin{proposition} \label{prop:cubiccomposites} Suppose $k_1, \dots k_n$ are linearly disjoint cyclic cubic fields, each with a totally positive system of fundamental units, i.e., with unit signature rank deficiency $\delta(k_i)$ equal to 2, $i =1,\dots, n$. Then the unit signature rank deficiency $\delta(k_1 \dots k_n)$ for the composite field $k_1 \dots k_n$ is at least $2 n$. \end{proposition} \begin{proof} We need to prove there are at least $2n$ totally positive units in $k_1 \dots k_n$ that are multiplicatively independent modulo squares (in $k_1 \dots k_n$). We proceed by induction on $n$, the case $n = 1$ being trivial. Suppose by induction that the composite $k_1 \dots k_s$ contains $2 s$ totally positive units $\epsilon'_1 , \dots , \epsilon'_{2s}$ that are multiplicatively independent modulo squares in $k_1 \dots k_s$. By assumption, the field $k_{s+1}$ contains two totally positive units $\epsilon_1, \epsilon_2$ that are multiplicatively independent modulo squares in $k_{s+1}$. Suppose $\epsilon \in \langle \epsilon_1, \epsilon_2 \rangle$ and $\epsilon' \in \langle \epsilon'_1 , \dots , \epsilon'_{2s} \rangle$ with $\epsilon \epsilon' = \alpha^2$ for some $\alpha$ in the composite field $F = k_1 \dots k_s k_{s+1}$. Let $\sigma \in \Gal (F/\QQ) $ be a lift of a generator for the cyclic group $\Gal (k_{s+1}/\QQ)$ that is the identity on $k_1 \dots k_s$. Then $\sigma(\epsilon) \epsilon' = \sigma(\alpha)^2$, so $\sigma(\epsilon)/\epsilon = (\sigma(\alpha)/\alpha)^2$ is a square in $F$, hence $k_{s+1} (\sqrt{ \sigma(\epsilon)/\epsilon })$ is a subfield of $F$. Since $F$ has degree $3^s$ over $k_{s+1}$, $k_{s+1} (\sqrt{ \sigma(\epsilon)/\epsilon })$ cannot be a quadratic extension, so $\sigma(\epsilon)/\epsilon$ is in fact a square in $k_{s+1}$. Since $\sigma$ acting on $\langle \epsilon_1, \epsilon_2 \rangle$ modulo squares in $k_{s+1}$ has no nontrivial fixed point, it follows that $\epsilon$ is the square of a unit in $k_{s+1}$. Then $\epsilon' = \alpha^2 /\epsilon$ would be a square in $F$, a cubic extension of $k_1 \dots k_s$, and, as before, this implies that $\epsilon'$ would be a square in $k_1 \dots k_s$. This shows that the totally positive units $\epsilon_1, \epsilon_2, \epsilon'_1 , \dots , \epsilon'_{2s}$ are multiplicatively independent modulo squares in $F$, completing the proof by induction. \end{proof} For a cyclic cubic field, either there is a totally positive system of fundamental units or the units have all possible signatures. Heuristics, supported by computations, in \cite{B-V-V} suggest that, when counted by discriminant, there is a positive proportion of cyclic cubic fields of either type. Roughly 3\% of cyclic cubic fields (see \cite{B-V-V} for the precise value) are predicted to have unit signature rank deficiency 2, so in particular there should exist infinitely many such cubic fields that are linearly disjoint. \begin{theorem} \label{thm:3} Suppose, as expected, that there exist infinitely many cyclic cubic fields having a totally positive system of fundamental units. Then the difference between $\phi(m)/2$ and the unit signature rank of $\QQ(\zeta_m)^+$ can be arbitrarily large. \end{theorem} \begin{proof} By Proposition \ref{prop:cubiccomposites} and Remark \ref{rem:deficiencies}, to obtain a unit signature rank deficiency at least $2n$ it suffices to take the cyclotomic field whose conductor is the product of the distinct primes (which are congruent to 1 mod 3) dividing the conductors of $n$ linearly disjoint cyclic cubic fields having totally positive fundamental units. \end{proof} \begin{remark} The same sort of arguments could be applied to composites of other abelian fields of, for example, odd prime degree. As in the case for cubic fields, it is expected that there are infinitely many such cyclic extensions of $\QQ$ with nonzero deficiencies (see \cite{B-V-V} for specific predictions). This suggests that the unit signature rank deficiency can increase without bound as one moves `horizontally' among cyclotomic fields, that is, over fields $\QQ(\zeta_m)^+$ where $m$ is the product of an increasing number of distinct primes, as opposed to the results of Section \ref{sec:towers} which show the deficiency is bounded as one moves `vertically' among cyclotomic fields. \end{remark} \section{Remarks on 2-adic Unit Signature Rank Deficiencies} \label{sec:unbounded} The results above have implications for analogous deficiencies for the `2-adic signatures' of units in the sense of \cite[Section 4]{D-V}, which we now very briefly outline. We use the notation of \cite{D-V}. If $F$ is a totally real field with $[F:\QQ] = n$ then there is a structure theorem for the image of the 2-Selmer group, $\Sel_2(F)$, under the 2-Selmer signature map $\phi$ (\cite[Theorem A.13]{D-V}). The space $\phi(\Sel_2(F))$ is $n$-dimensional over $\FF_2$ and is an orthogonal direct sum $U \perp S \perp U'$, where $U$ is the subspace of elements whose 2-adic signature is trivial, $U'$ the subspace of elements whose archimedean signature is trivial, and $S$ is a diagonal subspace. Since $F$ is totally real, the dimension of $U'$ is the same as the dimension of $U$ (\cite[Theorem A.13(a)]{D-V}) and so $S$ has dimension $n - 2 \dim(U)$. Suppose now that the unit signature rank deficiency of $F$ is $\delta(F)$. Then the set of signatures of units is a subspace of dimension $n - \delta(F)$. It follows that the dimension of $U$ is at most $\delta(F)$ (so $S$ has dimension at least $n - 2 \delta(F)$) and that the dimension of the image $\phi(E_F)$ of the units $E_F$ of $F$ is at least $n - \delta(F)$. Then $\dim ( \phi(E_F) + S )$ is at most $\dim \phi(\Sel_2(F)) = n$ and \begin{align*} \dim ( \phi(E_F) \cap S ) & = \dim \phi(E_F) + \dim S - \dim ( \phi(E_F) + S ) \\ & \ge (n - \delta(F)) + (n - 2 \delta(F)) - n = n - 3 \delta(F). \end{align*} Since $S$ is a diagonal subspace, it follows that the dimension of the subspace of 2-adic signatures of the units of $F$ is at least $n - 3 \delta(F)$. Hence the 2-adic signature deficiency of the units of $F$ is at most $3 \delta(F)$. As a consequence, Theorem \ref{thm:2}(b) and Corollaries \ref{cor:cor1}, \ref{cor:basicZp}, and \ref{cor:primepower} remain true if the (archimedean) signature rank of the units is replaced by the 2-adic signature rank of the units.

62e9aa79786d1d5bd85391793fd5b0a04e8e50d5

\section{Introduction} \label{sec:1} We consider estimating the parameter $n$ of the binomial distribution from $k$ independent observations when the success probability $p$ is unknown. This situation is relevant in many applications, for example in estimating the population size of a species \cite{Raftery} or the total number of defective appliances \cite{Draper-Guttman}. Another recent application is quantitative nanoscopy, see \cite{Schneider}. There, the total number of fluorescent markers (fluorophores) attached to so-called DNA-origami is estimated from a time series of microscopic images. The number of active fluorophores counted in each image is modeled as binomial observation, where the probability $p$ that a fluorophore is active in the respective image is very small (often below $5\%$). This setting, where the success probability $p$ is small (and $n$ potentially large), is very challenging. The difficulties that arise can be understood by considering the following property of the binomial distribution: if $n$ converges to infinity, $p$ converges to zero, and the product $np$ converges to $\lambda > 0$, then a Bin$(n,p)$ random variable converges in distribution to a Poisson variable with parameter $\lambda$. Thus, the binomial distribution converges to a distribution with a single parameter. This suggests that it gets harder to derive information about the two parameters separately when $n$ is large and $p$ small. In this context, it is instructive to look at the sample maximum $M_k$ as an estimator for $n$, which was suggested by Fisher in 1941 \cite{Fisher}. Although it turns out to be impractical, see \cite{DasGupta}, the sample maximum is consistent and converges in probability for fixed parameters $(n, p)$ exponentially fast to the true $n$, as $k\rightarrow\infty$. This can be seen from \begin{equation}\label{prob-max} \mathbf{P}\left( M_k =n \right)= 1-(1-p^n)^k, \end{equation} which implies, by Bernoulli inequality and since $1-x\leq e^{-x}$, that \begin{equation*} 1-e^{-kp^n} \leq \mathbf{P}\left( M_k =n \right)\leq kp^n. \end{equation*} In an asymptotic setting where $n\rightarrow\infty$ and $p\rightarrow0$ such that $kp^n\rightarrow0$, the probability in \eqref{prob-max} no longer converges to one. Thus, the sample maximum is a consistent estimator for $n$ only as long as $kp^n\rightarrow\infty$. The condition $e^n=O(k)$ is necessary for this to hold.\\ Estimating $n$ in this difficult regime becomes more manageable by including prior knowledge about $p$. We therefore consider random $N$ and $P$, and variables $X_1,\dots, X_k$ that are independently $\operatorname{Bin}(n,p)$ distributed given that $N=n$ and $P=p$. Various Bayesian estimators have been suggested over the last 50 years, see \cite{Draper-Guttman,Raftery,Berger,Chilko,Hamedani}. In all of this work, a product prior for $(N,P)$ is used, and the prior $\Pi_P$ on $P$ is chosen as beta distribution Beta$(a,b)$ for some $a,b > 0$. Since this is the conjugate prior, it is a natural choice. In contrast, there is quite some discussion about the most suitable prior $\Pi_N$ for $N$, see for example \cite{Kahn,Link,Villa,Berger}. Therefore, the asymptotic results in \cite{Schneider} are described flexible in terms of $\Pi_N$, and they only require a condition that ensures that enough weight is put on large values of $n$ (see equation \eqref{tail-bound} in Section \ref{sec:2}). In \cite{Schneider}, we also introduce a new class of Bayesian point estimators for $n$, which we call scale estimators. We choose $\Pi_P\sim\operatorname{Beta}(a,b)$ and set $\Pi_N(m)\propto m^{-\gamma}$ for a positive value $\gamma$. If $\gamma>1$, the prior $\Pi_N$ is a proper probability distribution, but it is sufficient to ensure $\gamma +a>1$ in order to obtain a well-defined posterior distribution. The scale estimator is then defined as the minimizer of the Bayes risk with respect to the relative quadratic loss, $ l(x,y)=( x/y-1)^2. $ Following \cite{Raftery}, it is given by \begin{equation}\label{E:ScaleBayes} \hat{n}:=\frac{\mathbb{E} \left[\frac{1}{N}|\mathbf{X}^k\right]}{\mathbb{E} \left[\frac{1}{N^2}|\mathbf{X}^k\right]}=\frac{\sum_{m=M_k}^\infty \frac{1}{m}L_{a,b}(m)\Pi_N(m)}{\sum_{m=M_k}^\infty \frac{1}{m^2}L_{a,b}(m)\Pi_N(m)}, \end{equation} where $\mathbf{X}^k=(X_1,\dots,X_k)$ denotes the sample, $M_k$ is the sample maximum, and $L_{a,b}$ is the beta-binomial likelihood, see \cite{Carroll-Lombard}. We refer to \cite{Schneider} for a detailed discussion and numerical study of this estimator.\\ The present article is structured as follows. In Section \ref{sec:2}, the main theorem (proven in \cite{Schneider}) is presented, which shows uniform posterior contraction in the introduced Bayes setting for suitable asymptotics of $n$ and $p$. The theorem states that $n^{6+\epsilon}=O(k)$ for $\epsilon>0$ is already sufficient for consistency of the Bayes estimator, improving significantly over the sample maximum. In Section \ref{sec:3}, we then conduct a simulation study to closer investigate the restrictions for the parameters $n$ and $p$ needed to ensure consistency. Our findings indicate that estimation of $n$ is still consistent if $n^5=O(k)$, but that it becomes inconsistent for $n^3=O(k)$. It is hard to pin down the exact transition from consistency to inconsistency when $n^{\alpha}=O(k)$, but our results suggest that it happens close to $\alpha = 4$. We discuss our results and provide several remarks in Section \ref{sec:4}. \section{Posterior Contraction for $n$} \label{sec:2} To study posterior contraction in the binomial model we consider the Bayesian setting described in Section \ref{sec:1}. For fixed parameters $n$ and $p$ that are independent of the number of observations $k$, posterior consistency follows from Doob's theorem, see, e.g., \cite{vanderVaart-Asymptotic}. We extend this result to the class of parameters \begin{equation} \label{M} \mathcal{M}_{\lambda}:=\left\{ (n_k,p_k)_k: 1/\lambda \leq n_kp_k\leq \lambda,\ n_k\leq \lambda\sqrt[6]{k/\log(k)} \right\} \end{equation} for fixed $\lambda > 1$. Since we want to handle a variety of suitable prior distributions for $N$, we only require that $\Pi_N$ is a proper probability distribution on $\mathbb{N}$ that fulfills the condition \begin{equation}\label{tail-bound} \Pi_N(m) \geq \beta e^{-\alpha m^2} \quad \forall\, m\in\mathbb{N} \end{equation} for some positive constants $\alpha$ and $\beta$. \begin{theorem}[see \cite{Schneider}]\label{main_result} Conditionally on $N=n_k$ and $P=p_k$, let $X_1,\dots,X_k\overset{i.i.d.}{\sim} \operatorname{Bin}(n_k,p_k)$. For any prior distribution $\Pi_{(N,P)} = \Pi_N \Pi_P$ on $(N,P)$ with $\Pi_P=\operatorname{Beta}(a,b)$ for $a,b>0$, and where $\Pi_N$ satisfies \eqref{tail-bound}, we have uniform posterior contraction over the set $\mathcal{M}_{\lambda}$ of sequences $(n_k, p_k)_k$ defined in \eqref{M} for any $\lambda>1$, i.e., \begin{equation*}\label{E:Result_sup} \sup_{(n_k, p_k)_k \in \mathcal{M}_{\lambda}} \mathbb{E}_{n_k,p_k}\big[ \Pi\big( N \neq n_k\ |\, \mathbf{X}^k \big) \big] \rightarrow 0,\ \text{as}~ k\rightarrow\infty. \end{equation*} \end{theorem} This result directly implies consistency of the scale estimator (\ref{E:ScaleBayes}) for parameter sequences in $\mathcal{M}_{\lambda}$. The flexible restrictions on the prior distribution allow to apply the result to the estimators derived in \cite{Chilko} and \cite{Hamedani} as well. Furthermore, it is possible to extend the statement of Theorem \ref{main_result} to improper priors on $N$, as done in Theorem 2 in \cite{Schneider}, in order to cover the estimators in \cite{Draper-Guttman} and \cite{Berger}. \section{Simulation Study} \label{sec:3} The theorem presented in the previous section states that the asymptotic behavior $n_k \sim O\big(\sqrt[6]{k/\log(k)}\big)$ leads to posterior contraction of $N$ for suitable priors, as long as $n_k p_k$ stays in a compact interval bounded away from zero. In this section we try to answer the question by how much the constraints on $\mathcal{M}_\lambda$ in Theorem \ref{main_result} can be relaxed. We address this problem by studying the relation between posterior contraction and the order $\alpha > 0$ when $n_k \sim O\big(\sqrt[\alpha]{k}\big)$. More precisely, we are interested in the smallest $\alpha = \alpha^*$ such that the result \begin{equation}\label{eq:contraction} \mathbb{E}_{n_k, p_k}\left[ \Pi\left( N\neq n_k\,|\, \mathbf{X}^k \right) \right] \rightarrow 0,\ ~\text{as}~ \ k\rightarrow\infty, \end{equation} remains valid. Tackling this problem analytically turns out to be extremely challenging, see the proof of Theorem \ref{main_result} in \cite{Schneider}. \\ In our simulations, we consider sequences $(n_k, p_k)_k$ defined by $n_k = w\sqrt[\alpha]{k}$ and $p_k = \mu / n_k$ for parameters $w, \mu > 0$. The values of $w$ and $\mu$ should, ideally, not matter for the asymptotics and thus for the pursuit of $\alpha^*$. Suitable choices of $w$ and $\mu$ for given $\alpha$ are still necessary for practical reasons to ensure that the asymptotic behavior becomes visible for the values of $k$ covered by the simulations. For any selection $(\alpha, w, \mu)$, we calculate the posterior probability of the true parameter $n_k$ and the MSE of different estimators for values of $k$ up to $10^{11}$. In order to achieve these extremely large observation numbers, we take care to minimize the number of operations when expressing the beta-binomial likelihood $L_{a,b}$ in our implementation. Since $L_{a,b}$ does not depend on the order of the observations but only on the frequencies of each distinct outcome $x_i$, the runtime depends on $n_k$ (the number of different values that $x_i$ can take) instead of $k$ itself. \begin{figure} \centering {\scriptsize\input{figure1.tex}} \caption{Asymptotic behavior of the scale estimator and posterior contraction. \textbf{(a)} shows log-log plots of the MSE of several scale estimators in different asymptotic scenarios $(\alpha, w, \mu)$. The value $\mu$ was set to $25$ in each simulation, and the parameters for the scale estimators were picked as all possible combinations of $\gamma\in\{0.5, 1\}$, $a\in\{1, 5\}$, and $b\in\{1, 5\}$. \textbf{(b)} shows the empirical mean of the posterior probabilities $\Pi(N = n_k^0 \,|\, \mathbf{X}^k)$ for the same four settings depicted in (a). \textbf{(c)} shows the MSE of the scale estimator with parameters $\gamma = a = b = 1$ for constant $\alpha = 6$ and varying values of $w$ and $\mu$.} \label{fig:asymptotics} \end{figure} Figures \ref{fig:asymptotics}a--b show the (empirical) mean posterior probability in \eqref{eq:contraction} and the (empirical) mean square error (MSE) between $\hat{n}$ and $n$ for different scale estimators $\hat{n}$ in several scenarios $(\alpha, w, \mu)$. The number of samples was set to 200. It is clearly visible that the choice $\alpha = 6$ leads to posterior consistency (which is in good agreement with Theorem \ref{main_result}), since the posterior probability approaches 1 while the MSE converges to 0. However, the simulations indicate that this also holds true for $\alpha = 5$. For $\alpha = 4$, it becomes questionable whether posterior contraction will eventually happen. The choice $\alpha = 3$, in contrast, leads to a clear increase of the MSE with increasing $k$, and posterior contraction evidently fails. \begin{figure} \centering {\scriptsize\input{figure2.tex}} \caption{Relation between $\alpha$ and $\beta$. For a given order $\alpha$, the corresponding value of $\beta$ was determined by conducting simulations like in Figure \ref{fig:asymptotics}a and fitting the slope for $k$ between $10^7$ and $10^9$. The graph shows that the zero point $\alpha^*$ of the conjectured function $\beta(\alpha)$ has to be in the vicinity of $4$.} \label{fig:slopes} \end{figure} An interesting observation is the power law behavior $\sim k^{-\beta}$ of the MSE, which is revealed by linear segments in the respective log-log plots. Figure \ref{fig:asymptotics}a shows that the slope $\beta$ is independent of the chosen estimator, and \ref{fig:asymptotics}c suggests that it might also be independent of $w$ and $\mu$. We can therefore consider $\beta$ as a function $\beta(\alpha)$ of $\alpha$ alone. A numerical approximation of $\alpha^*$ is then given by the value of $\alpha$ where $\beta$ changes sign, i.e., \begin{equation*} \beta(\alpha^*) = 0. \end{equation*} Since $\beta(\alpha)$ is strictly monotone, as a higher number $k$ of observations will lead to better estimates, such an $\alpha^*$ is uniquely defined. Figure \ref{fig:slopes} displays an approximation of the graph of $\beta(\alpha)$ for values between $\alpha = 2$ and $\alpha = 8$. The respective slopes are estimated by linear least squares regressions for $k$ between $10^7$ and $10^{9}$. Even though our numerical results do not allow us to establish the precise functional relation between $\alpha$ and $\beta$, it becomes clear that $\alpha^*$ indeed has to be close to $4$.\\ \begin{figure} \centering {\scriptsize\input{figure3.tex}} \vspace{0.1cm} \caption{Comparison of alternative asymptotic settings. \textbf{(a)} shows the MSE for three different asymptotic scenarios. In the first plot, $n_k$ and $p_k$ behave like in Figure \ref{fig:asymptotics} with $w = 16$ and $\mu = 25$. In the second plot, $p_k$ is fixed to the value $0.05$, while $n_k$ still increases with $k$ ($w = 16$). The third plot addresses the scenario where both $n_k$ and $p_k$ are held fixed. \textbf{(b)} shows the scenario of growing $n_k$ (with $\alpha = 6$ and $w = 16$) and different fixed values $p_k$. The graph shows that the slope in the linear segment does not depend on $p_k$. \textbf{(c)} shows the relation between $\beta$ and $\alpha$ for the scenario with fixed $p_k$ and growing $n_k$. The values of the slopes $\beta$ are determined as described in Figure \ref{fig:slopes}, with adapted ranges for $k$.} \label{fig:additional} \end{figure} For comparison, we additionally conducted simulations that target other asymptotic regimes. First, we keep $p_k$ constant and let $n_k$ again increase with the sample size, $n_k = w\sqrt[\alpha]{k}$. In this scenario, a properly rescaled binomial random variable converges to a standard normal distribution. Our simulations confirm that estimation of $n$ is easier in this case: the MSE in Figure \ref{fig:additional}a decreases faster when $\alpha=6$ and $p_k=0.05$ is fixed compared to $\alpha=6$ and $p_k\rightarrow0$. Since the rate of convergence $\beta$ in this alternative setting seems to be independent of the specific choice of $p_k = \mathrm{const}$, see Figure \ref{fig:additional}b, we can again look at the smallest order $\alpha$ that still exhibits consistency. Indeed, Figure \ref{fig:additional}c reveals that the estimation of $n$ remains consistent over a larger range of values for $\alpha$ in this setting, approximately as long as $\alpha>2$ (compared to $\alpha > 4$ in the original setting). The last asymptotic regime we consider is the classical one for parameter estimation, where $n_k = n$ and $p_k = p$ both stay constant as $k$ grows to infinity. Figure \ref{fig:additional}a covers this regime in the last plot. It affirms that estimating $n$ is easiest in this setting, and we obtain the expected rate $\sim k^{-1}$ for the convergence of the MSE towards zero. \section{Discussion} \label{sec:4} Theorem \ref{main_result} (see \cite{Schneider}) shows posterior contraction under diverging parameters $n_k$ and $p_k$ as long as $(n_k,p_k)\in \mathcal{M}_{\lambda}$, which implies $n_k=O(\sqrt[6]{k/\log(k)})$. The aim of our simulation study in Section \ref{sec:3} was to explore the minimal rate $\sqrt[\alpha]{k}$ for $n_k$ such that posterior consistency remains valid. The difference in the permissible rates turns out to be rather small, since our investigation suggests that $\alpha=5$ still allows for consistent estimation, whereas $\alpha=3$ clearly leads to inconsistency. Figure \ref{fig:slopes} shows that the true boundary $\alpha^*$ is likely close to 4, indicating that Theorem \ref{main_result} cannot be improved fundamentally. Several aspects of our simulations and findings deserve further commentary. First, Figure \ref{fig:asymptotics}c reveals that the slope $\beta$ is not strongly affected by the parameters $w$ and $\mu$ in the settings that we tested. However, our numerical approach is not suitable to verify questions like this with a high degree of confidence. For example, our numerics become instable for values $k > 10^{11}$. Secondly, we additionally conducted simulations for other estimators than the scale estimator \eqref{E:ScaleBayes} that are not shown in the article. For example, we tested various versions of the Bayesian estimator given in \cite{Draper-Guttman}. While their performance for $k \le 10^3$ varies quite much -- similar to the different estimators shown in Figure \ref{fig:asymptotics}a -- their asymptotic performance is exactly the same as for the scale estimator. Notably, the maximum likelihood estimator also exhibits the very same asymptotic behavior, even though it performs poorly in the regime of smaller $k$. The sample maximum, in contrast, shows a completely different behavior: the MSE diverges even for $n_k \sim \log(k)$. This illustrates the sharpness of the assumptions for Lemma 10 in \cite{Schneider}, which states that the sample maximum is consistent if $n_k\log(n_k)<c\log(k)$ for $c<1$. Finally, we consistently observed a phase transition in all simulations when the MSE drops below a value of about $0.1$, where it changes its behavior and begins to decreases faster than $\sim k^\beta$. Indeed, it seems to decay exponentially from that point on. We conjecture that this happens due to the discreteness of $n$, which means that the MSE cannot measure small deviations $|\hat{n} - n| < 1$ from the real $n$ without dropping to zero. Rather, if the posterior contracts so much that we estimate $n$ correctly most of the time, the MSE essentially captures the probability that $\hat{n}$ lies outside of the interval $(n-1,n+1)$, and such probabilities usually decay exponentially fast. For applications, the rate of the MSE before the exponential decay is often much more interesting. One instructive example in this context is the sample maximum in the setting of fixed $n$ and $p$, for which we know from Section \ref{sec:1} that it converges exponentially fast. However, as argued above, this only takes place when the MSE is already very small, and simulations suggest that the rate of convergence is much slower if the MSE is larger than $0.1$. For instance, if $p = 0.2$ and $n = 25$, we find $\beta \approx -0.13$. Thus, even though the true asymptotic behavior of the sample maximum is exponential, the practically meaningful rate of convergence is considerably worse than the rate $k^{-1}$ of the Bayesian estimators. \section*{Acknowledgements} Support of the DFG RTG 2088 (B4) and DFG CRC 755 (A6) is gratefully acknowledged. \input{YOUR_references} \end{document}

fe2764550928e28e34f6a0ae7359aa7c8af8e451

\section*{ABSTRACT} The complex interconnections between heterogeneous critical infrastructure sectors make the system of systems (SoS) vulnerable to natural or human-made disasters and lead to cascading failures both within and across sectors. Hence, the robustness and resilience of the interdependent critical infrastructures (ICIs) against extreme events are essential for delivering reliable and efficient services to our society. To this end, we first establish a holistic probabilistic network model to model the interdependencies between infrastructure components. To capture the underlying failure and recovery dynamics of ICIs, we further propose a Markov decision processes (MDP) model in which the repair policy determines a long-term performance of the ICIs. To address the challenges that arise from the curse of dimensionality of the MDP, we reformulate the problem as an approximate linear program and then simplify it using factored graphs. We further obtain the distributed optimal control for ICIs under mild assumptions. Finally, we use a case study of the interdependent power and subway systems to corroborate the results and show that the optimal resilience resource planning and allocation can reduce the failure probability and mitigate the impact of failures caused by natural or artificial disasters. \section{INTRODUCTION} Presidential Policy Directive 21 (PPD-21) identifies 16 critical infrastructure sectors including energy, communication and transportation systems as so vital that their malfunctions and incapacitation can lead to an enormous economic loss and public safety threat \cite{obama2013presidential}. Driven by the recent advances in information and communication technologies (ICTs) and the Internet of Things (IoTs), these sectors become highly interconnected, enabling faster information exchange and a higher level of situational awareness for real-time operations. For example, \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{zimmerman2016promoting} and \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{zimmerman2018network} investigate the interdependency between food, energy and water. \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{nwu020} shows that on the one hand, supervisory control and data acquisition (SCADA) systems control the power generation, transmission, and distribution according to the monitoring information from the communication sector. On the other hand, the energy sector provides the power to guarantee the normal operation of the SCADA systems. Protection of interdependent critical infrastructures (ICIs) against natural or human-made disasters is essential for providing support to a reliable and sustainable economy. The increasing interdependencies, however, make systems vulnerable to these catastrophes because a failure of one system will propagate to others and cause cascading failures without appropriate, rapid responses. Lessons of the September 11 attack, hurricanes Sandy and Irma have highlighted that protection and prevention against such disruptions are not always possible. Hence a paradigm shift to emphasize the preparedness and response is indispensable to enhance the resilience of ICIs as shown in Fig. \ref{Resilience}. \begin{figure*}[] \begin{multicols}{2} \includegraphics[height=0.5\linewidth, width=\linewidth]{figure/resilience}\par\caption{The systematic resilience consists of three phases. First is the robustness against disruptions to prevent failures. Second is the impact mitigation to guarantee the essential services under failures. Third is the prompt response and recovery. A paradigm shift from protections to the mitigation and recovery is indispensable for catastrophes.} \label{Resilience} \includegraphics[height=0.5\linewidth, width=\linewidth]{figure/multiple2single}\par\caption{Multi-layer networks representing interdependent infrastructures such as power grid, subway and communication networks. They can be viewed as a large-scale aggregated network. The directed link between two nodes indicates either physical, cyber or logical influences from the source to the sink.} \label{multilayer} \end{multicols} \end{figure*} Many works have investigated the modeling, simulation, and design of interdependent critical infrastructures \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{ouyang2014review,zimmerman2017conceptual}. Network flows \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{lee2007restoration} and the interacting dynamic model \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{rosato2008modelling} are used to assess and manage the risks. Built upon the modeling of interdependent infrastructures, \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{yuan2015resilient} focuses on designing a resilient control for power grid. \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{6603998} has proposed an L1Simplex architecture to address both physical and software failures in the cyber-physical systems. If an intelligent attacker rather than a random accident causes the network component failure, \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{zhu2015game,huang2017large} apply dynamic game frameworks to incorporate the strategic participants who adapt to the defense policy. Despite these pioneering works, challenges still remain to build a holistic model that captures not only distinct characteristics of individual infrastructures but also complex dependencies among sectors due to geographical, logical, cyber-physical, or human connections \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{rinaldi2001identifying}. Without a holistic framework, we cannot integrate the data or models of individual sectors to make a globally optimal planning and defense decisions. For example, system operators should not only make plans for normal operations of their own system but also be concerned about the possible failures of other dependent sectors and prepare for the potential influences. The framework should also be general enough to include various sectors and straightforward enough to enable real-time plannings for large-scale SoS, which motivates the MDP model in Section \ref{model} to abstract the complex networks, unify the multi-layer interdependences, and design resilient control mechanisms. Moreover, the \textbf{probabilistic} and \textbf{persistent} transition of system status requires system operators to prepare for the uncertainties in a dynamic perspective and avoid the following two extremes. First is the blind optimism that satisfies with the current normal operation and makes no sufficient contingency plan for probabilistic failures. Second is the pursuit of absolute security that would result in a huge inspection and maintenance cost. Thus, it is crucial to make \textbf{proactive} and \textbf{cost-effective} controls based on the real-time status of SoS to reduce the long-term risk economically and globally. In this paper, we first establish a holistic network model to capture the relationships between nodes within an individual infrastructure and across the infrastructures. We use a probabilistic model to assess the impact of the couplings between different components in a system, in which the failure probability of one component depends not only on its state but also on state of its connected ones. One fundamental characteristic of the infrastructures is their dynamics in real-time operations \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{zimmerman2018network}. For example, the load on a bus in a power grid changes due to the time-varying demands, and the bus would encounter a load shedding if the demand exceeds the supply. The concept of resilience involves the phase of recovery of infrastructure after disastrous events. This process is naturally a dynamic one \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{ChenSigmetrics2017}. Hence, building on the probabilistic failure model, we propose a Markov decision processes (MDP) model to take into account the underlying dynamic processes of ICIs and develop resilient mechanisms for heterogeneous interdependent infrastructures. The MDP model has naturally a large state-space due to the large-scale feature of infrastructure networks and exponential increases of the system states with respect to the number of components \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{huang2017large,hayel2015resilient}. The curse of dimensionality of the MDP makes the control design computationally intractable even in a moderate-size system with 15 nodes. To this end, we reduce the dimension by first using basis functions to reformulate the problem as an approximate linear program (ALP) and then leveraging the sparsity of the interconnections to simplify the ALP with factored graphs. The approximated problem yields a low computational complexity. Under mild assumptions on the dependence structures and the separability of reward and control cost functions at each node, the resilient control policies can be designed in a distributed fashion. We can show that the system performance under the distributed control scheme is equivalent to the one achieved under the centralized one. We corroborate the established MDP model and analytical results with case studies of an interdependent power and subway network. \textbf{Organization of the Paper}: The rest of the paper is organized as follows. Section \ref{model} establishes the MDP model of the interdependent infrastructures. Problem analysis is given in Section \ref{analysis}. The structural results including the optimal and distributed resilient control policies are obtained in Section \ref{newsetting}. The case study of a power and transit network is presented in Section \ref{case_study}, and Section \ref{conclusion} concludes the paper. \section{SYSTEM MODEL}\label{model} In this section, we introduce the network model to represent the complex interdependent SoS and the MDP model to produce resilient plannings of ICIs. The resilient plannings are both \textbf{proactive}, i.e., maintain high-risk nodes in advance of failures to reduce the future risks, and \textbf{cost-effective}, i.e., delay the repair of low-influence nodes due to the restrained or costly resources. \subsection{Network Model} ICIs can be modeled by a multi-layer network as shown in Fig. \ref{multilayer}, where nodes represent heterogeneous components in multiple infrastructures such as buses in power grid, subway stations in transportation, and base stations in the communication network. The links connecting nodes in the network represent the dependencies either homogeneous, e.g., the failure of one subway station can disrupt the service of the next station, or heterogeneous, e.g., the power loss of the grid leads to the shutdown of subway stations. Moreover, the inter-dependency implies the ping-pong effect. For example, the shutdown of subway stations can, in turn, slow down the power recovery because maintainers may take extra time to commute. The vicious spiral can lead to the cascading failure. Without loss of generality, we use an aggregated network $\mathcal{G}=(\mathcal{N},\mathcal{E})$ to capture the heterogeneous components in the multi-layer network, where $\mathcal{N}$ denotes a set of nodes and $\mathcal{E}$ is a set of links. \subsection{MDP Model} Each infrastructure node $i \in \mathcal{N}$ in the network $\mathcal{G}$ has a state $X_i$ representing its working status. State $X_i$ admits values from a binary set $\{0,1\}$. Specifically, $X_i=1$ indicates that node $i$ works normally and $X_i=0$ means otherwise. In a network of $|\mathcal{N}|=n$ nodes, the system state can be denoted by $\mathbf{X}=[X_i]_{i \in \mathcal{N}} \in \mathcal{X}:=\{0,1\}^n$. Due to the interdependencies between different nodes, the failure probability of one component depends not only on its own state but also on its connected ones. Similarly, each infrastructure node $i\in\mathcal{N}$ can be controlled by taking action $A_i \in \{0,1\}$. Specifically, $A_i=1$ represents repair (if node is faulty) or maintenance (if normal) of node $i$, and $A_i=0$ indicates that no action is applied. The system action is a vector $\mathbf{A}=[A_i]_{i\in\mathcal{N}} \in \mathcal{A}:=\{0,1\}^n$ including all the actions of each node. In order to design resilient mechanisms for the real-time operation of ICIs, we use an MDP model to capture the dynamics of the infrastructure states as well as the underlying failure and recovery processes of each node. Our objective is to determine a stationary policy $\pi: \mathcal{X}\mapsto \mathcal{A}$ that yields the optimal control of nodes to achieve the largest long-term benefit of ICIs. To achieve this goal, we further define the reward function $R: \mathcal{X}\times \mathcal{A} \mapsto \mathcal{R}$ and the objective function $V_\pi(\mathbf{x}): \mathcal{X}\mapsto \mathbb{R}$. The reward function measures the performance of the ICIs at one step by taking the number of working nodes into account. The objective function accumulates the obtained reward at each step over an infinite horizon: $V_{\pi}(\mathbf{x})=\mathbb{E}_{\pi}[\sum_{t=0}^{\infty}\gamma^{t}R\left(\mathbf{X}^{t},\pi(\mathbf{X}^{t})\right)|\mathbf{X}^{0}=\mathbf{x\mathrm{]}}$, where $\gamma<1$ is a positive discounted factor, $\mathbf{x}$ is an initial state, and $\mathbf{X}^t$ is a random variable denoting the system state at stage $t$. The state transition probability follows the Markov property, i.e., the system state at the next time step only depends on the current system state and the adopted action. For example, $P(\mathbf{x}'|\mathbf{x},\mathbf{a})=\mathbb{P}\{\mathbf{X}^{t+1}=\mathbf{x}'| \mathbf{X}^t=\mathbf{x},\mathbf{A}^t=\mathbf{a} \}, \ \forall t$, denotes the transition probability from state $\mathbf{x} \in \mathcal{X}$ to $\mathbf{x}' \in \mathcal{X}$ under control $\mathbf{a} \in \mathcal{A}$. Note that transition probability $P$ captures the interdependencies between connected infrastructure nodes. Further, the value function $V(\mathbf{x})=\max_\pi V_\pi(\mathbf{x}):=V_{\pi^*}(\mathbf{x})$ is the maximum achievable economic benefit of the ICIs starting from state $\mathbf{x}$ under the optimal policy $\pi^*$. A larger value of $V(\mathbf{x})$, $\forall \mathbf{x}$, indicates a better performance of ICIs, and the network is resilient to failures by using the control policy $\pi^*$. \section{PROBLEM ANALYSIS}\label{analysis} In this section, we introduce a sequence of solution techniques of the MDP problem and tackle the curse of dimension by exploiting the sparse nature of the large-scale infrastructure networks. Approximation with the factor graph and the variable elimination technique leads to equivalent distributed policies. \subsection{Linear Programming} One way to characterize the optimal policy of the MDP is via dynamic programming with the Bellman operator $(TV)(\mathbf{x}):= \max_\pi R(\mathbf{x},\pi(\mathbf{x}))+\gamma\sum_{\mathbf{x}'\in\mathcal{X}}P(\mathbf{x}'|\mathbf{x},\pi(\mathbf{x}))V(\mathbf{x}'),\ \forall \mathbf{x}$. Then, the value function $V(\mathbf{x})$ is the fixed point of $T$, i.e., $V(\mathbf{x})=(TV)(\mathbf{x}), \forall \mathbf{x}$. Linear programming (LP) provides a convenient and efficient approach to solve the Bellman equation with value functions as LP variables, i.e., \begin{equation} \begin{split} \mathrm{(LP):}\ \min_{V(\mathbf{x})}:\quad & \sum_{\mathbf{x}\in\mathcal{X}}\alpha(\mathbf{x})V(\mathbf{x})\\ \mathrm{s.t.} \quad & V(\mathbf{x})\ge R(\mathbf{x},\mathbf{a})+\gamma\sum_{\mathbf{x}' \in \mathcal{X}}P(\mathbf{x}'|\mathbf{x},\mathbf{a})V(\mathbf{x}),\\ &\forall\mathbf{x}\in\mathcal{X},\ \forall \mathbf{a}\in \mathcal{A}. \end{split} \label{eq:exact LP} \end{equation} LP has the benefit to be easily extended to include both endogenous and exogenous constraints. For example, if we add an endogenous action constraint $\sum a_i=1$ which limits to single repair due to the limited resources at each stage, the LP formulation remains the same with only a reduced dimension of action space $\mathcal{A}'$ as in \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{hln}. Similarly, we can directly add exogenous action constraints such as $\sum c_i(a_i)<C$ which gives a bound on the action cost due to limited budges at each step. The disadvantage of \eqref{eq:exact LP} is that the curse of dimension appears in both state space $|\mathcal{X}|=2^n:=N$ and action space $|\mathcal{A}|=2^n:=N$ when solving the LP directly. Therefore, we use ALP and FMDP method to reduce the computation complexity \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{guestrin2003efficient}. First, we use the linear value function \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{de2002linear} to approximate the value functions $V(\mathbf{x})=\sum_{i=1}^{k}w_{i}h_{i}(\mathbf{x})$, where $h_i$ and $w_i$ is the $i^{th}$ basis function and its weight, respectively. Hence the number of LP variables is reduced from $N=2^n$ to $k$ and we obtain the ALP with new variables $\{ w_1,w_2,...,w_k\}$. \begin{equation} \begin{split} \mathrm{(ALP):}\ \min_{\mathbf{w}}\ & \sum_{\mathbf{x} \in \mathcal{X}}\alpha(\mathbf{x})\sum_{i \in \mathcal{N}}w_{i}h_{i}(\mathbf{x})\\ \mathrm{s.t.}\ \sum_{i \in \mathcal{N}}w_{i}&h_{i}(\mathbf{x})\ge R(\mathbf{x},\mathbf{a})+\gamma\sum_{\mathbf{x}'\in \mathcal{X}}P(\mathbf{x}'|\mathbf{x},\mathbf{a})\sum_{i \in \mathcal{N}}w_{i}h_{i}(\mathbf{x}'),\\ \forall\mathbf{x}\in&\mathcal{X},\ \forall \mathbf{a}\in \mathcal{A}. \end{split} \end{equation} Our second step is to reduce the dimension of the transition probability matrix $P(\mathbf{x}'|\mathbf{x},\mathbf{a}) \in \mathcal{R}^{N\times N \times N}$. The system transition probability $P(\mathbf{x'}|\mathbf{x},\mathbf{a})$ can be factored into a multiplication of local transition probabilities $\prod_{i=1}^n P(x_i|\mathbf{{x}}_{\bar{\Omega}_i})$ where ${\bar{\Omega}_i}$ is the node set that can affect the local transition probability of node $i$. With the assumption that the network $\mathcal{G}$ is sparse, we have $|\bar{\Omega}_i|<n$. Notice that the co-domain of each basis function $h_i(\mathbf{x})$ is determined by some subset of $\mathcal{X}$. The effective domain of $g_{i}(\mathbf{x},\mathbf{a}):=\sum_{\mathbf{x}'}P(\mathbf{x}'|\mathbf{x},\mathbf{a})h_{i}(\mathbf{x}')$ is also restricted to a set much smaller than $\mathcal{X}$ and a special case is discussed under assumption \ref{actionassumption}. Define $\bar{g}_i(\mathbf{x},\mathbf{a}):=\gamma g_{i}(\mathbf{x},\mathbf{a})-h_{i}(\mathbf{x})$, the constraints $\sum_{i}w_{i} h_{i}(\mathbf{x})\ge R(\mathbf{x},\mathbf{a})+\gamma \sum_{i} w_{i}\ g_{i}(\mathbf{x},\mathbf{a}), \forall \mathbf{x},\mathbf{a}$ is rewritten in a more compact form $ 0\geq \max_{\mathbf{x}, \mathbf{a}} \ R(\mathbf{x},\mathbf{a})+\sum_{i} w_{i} \ \bar{g}_{i}(\mathbf{x},\mathbf{a}), \forall \mathbf{a} \in \mathcal{A}.$ The reward function $R(\mathbf{x},\mathbf{a})$ can also be factored into a sum of functions with the domain restrained to some subsets of $\mathcal{X}$. Then, we can use the variable elimination method to eliminate each element of $\mathbf{x}$ and $\mathbf{a}$ step by step \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{hln}. \subsection{Approximate Optimal Policy} The optimal centralized policy under the approximate form of LP is obtained by maximizing the objective function given the initial state $\mathbf{x}$, i.e., $ \mathbf{a}^* \in \mathrm{arg}\max_{\mathbf{a} \in \mathcal{A}} [R(\mathbf{x},\mathbf{a})+\gamma\sum_{\mathbf{x}' \in \mathcal{X}}P(\mathbf{x}'|\mathbf{x},\mathbf{a})V(\mathbf{x}')]. $ Similarly, an approximation is adopted to reduce the computation to $k+1$ multiplications and $k$ summations for any given pair of $(\mathbf{x},\mathbf{a})$, i.e., \begin{equation} \begin{split} \mathbf{a}^*& \in \mathrm{arg}\max_{\mathbf{a} \in \mathcal{A}}\left[R(\mathbf{x},\mathbf{a})+\gamma\sum_{i=1}^k w_i g_i(\mathbf{x},\mathbf{a})\right]. \end{split} \label{eq:greedy} \end{equation} To solve \eqref{eq:greedy}, for every state $\mathbf{x}$, we need to search for action $\mathbf{a}$, which means that we have to compute (\ref{eq:greedy}) for $|\mathcal{X}|\times |\mathcal{A}|=2^{2n}$ times. However, by exploiting the structure of the problem, we can make assumptions in Section \ref{newsetting} to achieve a distributed computation of the optimal policy. \subsection{Main Structural Results} \label{newsetting} To study the interdependent infrastructure network consisting of $n$ nodes, we choose the basis function $h_i(\mathbf{x})$ to be the indicator function of node $i$'s state, i.e., $h_i(x_i,\mathbf{x}_{-i})=x_i, \forall \mathbf{x}_{-i} \in \{0,1\}^{n-1}$. Hence the number of basis functions is of the same size of the nodes, i.e., $k=n$. Weight parameter $w_i$ shows the importance of node $i$ in the measure of economic benefits. The advantage of choosing such basis functions is that the value function $V(\mathbf{x})$ at state $\mathbf{x}=[x_i]_{i \in \mathcal{N}}$ is approximated by the summation of the weight $w_i$ of all working node $x_i=1$ at state $\mathbf{x}$. With a larger number of critical nodes functioning, the system achieves a higher reward in the long run. This choice of basis functions is particularly suitable for sparse networks because they primarily capture local effects. We can add new basis functions such as $h_i(x_i,x_j,\mathbf{x}_{\mathcal{N}\backslash\{i,j\}})= x_i \cdot x_j, \forall j \in \Omega_i$, which is necessary when non-local effects play a major role. We make the following reasonable assumptions to achieve a distributed repair policy. \begin{assumption} The local transition probability of node $i$ is not affected by the control value of other nodes, i.e., $P(x_i|\mathbf{{x}}_{\bar{\Omega}_i},\mathbf{a})=P(x_i|\mathbf{{x}}_{\bar{\Omega}_i},a_i)$. \label{actionassumption} \end{assumption} Under the above assumption, we obtain the following equations: \begin{eqnarray} \nonumber g_{i}(\mathbf{x},\mathbf{a}) &=&\sum_{\mathbf{x}' \in \mathcal{X}}P(\mathbf{x}'|\mathbf{x},\mathbf{a})h_{i}(x_{i}') =\sum_{i \in \mathcal{N}} \prod_{j \in \mathcal{N}} P(x_{j}'|\mathbf{{x}}_{\bar{\Omega}_{j}},\mathbf{a})h_{i}(x_{i}')\\ \nonumber &=&\sum_{x_{i}'} P(x_{i}'|x_{i},\mathbf{x}_{\Omega_{i}},\mathbf{a})h_{i}(x_{i}') =P(X_{i}^{'}=1|\mathbf{{x}}_{\bar{\Omega}_{i}},a_i). \end{eqnarray} \begin{assumption} We can separate and factor the reward function $R(\mathbf{x},\mathbf{a})$ into the net reward $\sum_{i=1}^n r_i(\mathbf{{x}}_{\bar{\Omega}_i})$ and the action cost $-\sum_{i=1}^n c_i(a_i)$, where $r_i(\mathbf{{x}}_{\bar{\Omega}_i})$ and $c_i(a_i)$ are the local reward and the cost function of node $i$, respectively. \label{separatereward} \end{assumption} \begin{theorem} \label{distributed policy} A distributed policy $a_i \in \mathrm{arg}\max_{a_i} \gamma w_i g_i(\mathbf{{x}}_{\bar{\Omega}_{i}},a_i)-c_i(a_i)$ achieves the same reward obtained by the centralized optimal policy using \eqref{eq:greedy} based on assumption \ref{actionassumption} and \ref{separatereward}. \end{theorem} Theorem. \ref{distributed policy} holds as we can represent the RHS of Eq. \eqref{eq:greedy} in a factored form, i.e., $$ \max_{a_1,...,a_n}[R(\mathbf{x},\mathbf{a})+\gamma\sum_i w_i g_i(\mathbf{x},\mathbf{a})]=\sum_{i=1}^k r_i(\mathbf{{x}}_{\bar{\Omega}_i})+ \sum_{i=1}^n \max_{a_i} \gamma w_i g_i(\mathbf{{x}}_{\bar{\Omega}_{i}},a_i)-c_i(a_i).$$ The distributed policy only takes into account the states of node $j \in \bar{\Omega}_i$ and reduces the computation complexity from $O(|\mathcal{X}|\times |\mathcal{A}|)$ to $n \cdot 2^{\Omega_i}$. In many applications, the policy designer can only control a limited number of nodes. Thm. \ref{distributed policy} still holds as the sparse control condition can be viewed as a special case. Our next step is to find a pattern to quickly identify each node's optimal action, i.e, repair or not. Thm. \ref{pattern} achieves the above objective under an arbitrarily given set of controllable nodes without explicitly computing $a_i \in \mathrm{arg}\max_{a_i} \gamma w_i g_i(\mathbf{{x}}_{\bar{\Omega}_{i}},a_i)-c_i(a_i)$. With assumption \ref{nature}, we can find the repair policy thresholds as summarized in Thm. \ref{pattern}, which means that we switch to the opposite action if the condition is not satisfied. Let $\bar{\Omega}_i=\{ i,\Omega_i\}$ and $\mathbf{{x}}_{\bar{\Omega}_i}=[x_i,\mathbf{x}_{\Omega_i}]$. Besides, we define a shorthand notation $P_ {\mathbf{x}_{\Omega_i}=j}:=P(X_i^{'}=1 | \mathbf{X}_{\Omega_i}=j,A_i=0)$. \begin{assumption} We further assume that a repair is always effective $P(X_i^{'}=1 | \mathbf{x},a_i=1)=1, \forall i, \forall \mathbf{x}$ and the node stays faulty without a repair $P(X_i^{'}=1 | x_i, \mathbf{x}_{\Omega_i},a_i=0)=x_i \cdot P_ {\mathbf{x}_{\Omega_i}}, \forall i, \forall \mathbf{x}$. In addition, the cascading failure means that $P_{\mathbf{{x}}_{\Omega_i}=\mathbf{1}}>P_{\mathbf{{x}}_{\Omega_i} \neq \mathbf{1}}$, where $\mathbf{1}$ represents a vector with all $n$ elements as $1$. \label{nature} \end{assumption} \begin{theorem} With assumptions \ref{actionassumption}, \ref{separatereward} and \ref{nature}, we only repair node $i$ when it is faulty and its action cost is relatively low, i.e., when $c_i$ satisfies $c_i(1)-c_i(0)<w_i$. If $0<c_i(1)-c_i(0)<w_i(1-P_{\mathbf{{x}}_{\bar{\Omega}_i}=\mathbf{1}})$, we repair node $i$ even when nodes in ${\bar{\Omega}_i}$ are all working normally. We choose to repair node $i$, i.e., $A_i=1$, when $i$ is working $X_i=1$ yet other nodes in $\Omega_i$ have states $\mathbf{x}_{\Omega_i} \neq \mathbf{1}$ under the condition $w_i(1-P_{\mathbf{{x}}_{\bar{\Omega}_i}=\mathbf{1}})<c_i(1)-c_i(0)<w_i(1-P_{\mathbf{{x}}_{\Omega_i} \neq \mathbf{1}})$. \label{pattern} \end{theorem} To show Thm. \ref{pattern}, we first look at the case when only one given node is controllable. Since action $a_i$ is binary, we can obtain the optimal policy by comparing which action leads to a larger value. We generalize the one-node case to the case of multiple nodes because the Thm. \ref{distributed policy} guarantees that each node's policy can be decided independently of others' actions. \section{CASE STUDY} \label{case_study} In this section, we use a case study motivated by Hurricane Sandy, which struck New Jersey on Oct. 29, 2012, resulting in an estimated \$50 billion economic loss and at least 147 direct deaths \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{blakeii}. It wreaked havoc on critical infrastructures such as power grid, transportation and communication systems. In this case study, we focus on the impact of the power and subway failures in the lower Manhattan and downtown Brooklyn as shown in Fig. \ref{bidirection}\footnote{Background image source: Google map}. \subsection{Power and Subway} The most direct impact of Sandy has been on the power grid. Much of Manhattan south of 39th Street has suffered a massive power cut because of the floods and the high wind. A \href{http://www.huffingtonpost.com/2012/10/30/coned-explosion-hurricane-sandy-video_n_2044097.html}{blast of one Con Edison substation} on 14th Street at 9 p.m. on Oct. 29 has intensified the power outage. Besides, ConEd took parts of its grid offline due to the rising water. Load shedding arose when the demand strained the capacity during the storm. The massive damages made ConEd unable to restore most of the power outages in lower Manhattan until 2 November. The whole power system has taken weeks to recover completely, and nearly 8.5 million people have suffered from power outages. As shown in Fig. \ref{ConMTA}\footnote{Source: https://www.coned.com/en/about-us/media-center/multimedia-library-or-gallery.}, the normal operation of subway system relies heavily on a stable power supply. \begin{figure} \includegraphics[width=0.6 \linewidth]{figure/ConMTA} \centering \caption{After the high-voltage transmission of 69/138kV, area substations transform the power voltage to 13/27kV, convert it to direct power current and feed it for the 600-volt third rail of the track to propel the train. Besides, transformers in the distribution grid further lower the voltage to 120/208V to power signals, lights, and track switches in subway stations.} \label{ConMTA} \end{figure} Thus, the subway system has been significantly affected as well. The MTA shuts down the entire subway system from 7 p.m. on Oct. 28 to move trains away from lowland and vulnerable area. Floodwaters brought by the storm surge of Sandy began to enter the subway tunnels and stations after 9 p.m. on 29 October. All seven East River subway tunnels connecting Manhattan, Brooklyn, and Queens have suffered from floods a day later. Besides, Hurricane Sandy wiped out tracks on the A train in the Rockaways and severely damaged new South Ferry station. Since after flooding, water can short-circuit electrical devices including signals, switches, and the electrified third rail, the subway has to remain closed before dewatering. The pump system of the normal drainage does not function during Sandy because of the power cut. Thus, it takes a long time to pump out water with limited backup power generators. After seawater has been pumped out, MTA needs extra effort to clean the salt deposits and debris, inspect, test and repair the electrical components to guarantee the safety of train operations. Unfortunately, not only the resilience of electric power has a significant impact on the recovery of the subway system, the repairing process of electric power systems is strongly influenced by the subway as well because the subway shutdown creates difficulty for ConEd crews to reach faulty components. \begin{figure} \includegraphics[width=0.6\linewidth]{figure/bidirection3} \centering \caption{Network representation of the suggested interdependent microgirds and subway system at the lower Manhattan and downtown Brooklyn area. Red and blue links represent power and subway lines, respectively. Power nodes are indexed from 1 to 10 while the coupled subway stations are indexed from 11 to 20. The size of node $i$ is proportional to its reward $r_i(\mathbf{x}_{{\bar{\Omega}_i}})$, indicated by the people flow density data from \href{http://web.mta.info/nyct/facts/ridership/ridership_sub_annual.htm}{MTA}. Power and subway nodes of the same place inside one yellow circle share the same local reward. } \label{bidirection} \end{figure} \subsection{Network Construction} Lessons learned from Sandy has indicated that the lack of sufficient backup power decreases the recovery efficiency due to the failure of conventional pumping systems. Hence, backup generators (e.g., diesel generators) or local energy storage devices should be prepared in case of power failures. For instance, Verizon's Garden City central office powered by seven fuel cells has saved it from the loss of power \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{kwasinski2013lessons}. Thus, Microgrids can be adopted to supply power to local users when the centralized power distribution network breaks down as what has happened during Sandy. In this case study, we construct a 20-node interdependent network illustrated in Fig. \ref{bidirection}, where each microgrid node $1$-$10$ represents a local backup generator located at the corresponding subway station $11$-$20$. Transmission lines in red connect the microgrid so that loads of each node can change due to its working status. The blue lines construct three simplified subway lines. Let $X_i=1$ be the state of normal operation while $X_i=0$ indicates the failure of node $i$. Backups can fail and we set up the failure model of the microgrids based on our previous work \def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{hln}. The failure of microgrids at subway stations can lead to the meltdown of the public transit system. For example, if subway station $17$ is out of service due to the power outage of node $7$, the entire line $4/5$ has to suspend as trains cannot get through. On the other hand, the subway system affects the microgrids through reward $r_i(\mathbf{x}_{{\bar{\Omega}_i}})$ of node $i$. We adopt the proposed framework to illustrate how we achieve the resilience of the microgrids and the subway lines under our optimal policy. For each node $i$, $A_i=1$ means to repair or maintain the node based on the state of the node. On the other hand, $A_i=0$ means that no action is taken. The coupling of the power and subway system exists as an illustration of the interdependent infrastructures. The failure of microgrids at subway stations can lead to the meltdown of the public transit system. The subway system affects the microgrids through reward $r_i(\mathbf{x}_{{\bar{\Omega}_i}})$ to node $i$. \subsection{Computation Result} \begin{table}[t] \centering \caption{Results of ALP with all nodes controllable.} \label{Computation Results} \begin{tabular}{|c|c|c|c|c|} \hline \multicolumn{3}{|c|}{Node Index and Name} & Local Reward $r_i(\mathbf{x}_{{\bar{\Omega}_i}})$ & Weight $w_i$\\ \hline 1 & Jay St. & 11 & 15.3215 & 18.5749 \\ \hline 2 & Court St. & 12 & 7.1197 & 8.7800 \\ \hline 3 & Clark St. & 13 & 2.4738 & 3.8601 \\ \hline 4 & High St. & 14 & 2.9452 & 3.9617 \\ \hline 5 & Broadway & 15 & 15.7205 & 17.5535 \\ \hline 6 & Chamber St. & 16 & 14.1986 & 14.2076 \\ \hline 7 & WTC & 17 & 20.4828 & 25.2162 \\ \hline 8 & Fulton St. & 18 & 1.7065 & 2.6374 \\ \hline 9 & Wall St. & 19 & 9.3785 & 11.2510 \\ \hline 10 & Bowling Gr & 20 & 10.6528 & 12.8872 \\ \hline \end{tabular} \end{table} Table \ref{Computation Results} shows that the weight of each node obtained by ALP follows the same ordering as the local reward except for node 1 and 5. This result arises from the fact that the topology of the network has less influence on the weight than the local reward. The centralized and the distributed search methods yield the same policy as indicated by Thm. \ref{distributed policy} and the optimal policy satisfies the pattern stated in Thm. \ref{pattern}. The optimal policy is not to just myopically repair defective nodes but also maintain some healthy nodes to achieve a long-term system-wide performance. For instance, we observe that the policy of the most critical node (WTC) is not repaired only when node $6$ and $7$ are both working as shown in Fig. \ref{recoverymap}. We choose to maintain node $7$ when $X_6=0, X_7=1$ to prevent node $7$'s outage in later stages. Other neighboring nodes $6$ and $8$ are not influential on the policy of node $7$ because they have relatively small weights as shown in Table \ref{Computation Results}. As for the computation complexity, it takes the centralized search method about $13843.7$ seconds to solve the case after weights $w$ are computed while our distributed search process requires only $2.1$ seconds. \subsection{Simulation Result} We choose $\mathbf{X}=\mathbf{1}$ as the initial state and show the system resilience under our optimal policy in Fig. \ref{systemresilience}. Averaged under 50 iterations, the data of the $y$-axis shows the average number of working nodes and also measures the resilience of the system. Without controls, although the prior failure probability is set to a small number $0.01$, cascading failure still occurs, and the survival number of nodes quickly decreases to zero. On the other hand, our optimal policy repairs the proper nodes so that the system quickly recovers and stays at the healthy state $\mathbf{X}=\mathbf{1}$. The variance of the subway system is larger than that of the microgrid because one subway station out of service makes the entire line break down. \begin{figure*}[h] \begin{multicols}{3} \includegraphics[width=\linewidth]{figure/new_actionstate}\par\caption{The blue circles at the top two subfigures show the state transition of node $7$ and $6$ respectively. The subplot at the bottom shows the corresponding action of node $7$.}\label{recoverymap} \includegraphics[width=\linewidth]{figure/givencontrol}\par\caption{Variations of the average number of working nodes when only node $7$ (WTC) and node $8$ (Fulton St.) are controllable. The red dots indicate an average number of the working node with error bars shown in blue.}\label{givencontrol} \includegraphics[width=0.9\linewidth]{figure/increasevalue}\par\caption{System reward grows with the increase of the number of controllable nodes. The error bar shows the upper and the lower bounds adopting different combinations of controllable nodes.} \label{increasevalue} \end{multicols} \end{figure*} \begin{figure} \subfigure{ \includegraphics[width=0.49\columnwidth]{figure/new_power} } \subfigure{ \includegraphics[width=0.49\columnwidth]{figure/new_sub} } \caption{The resilience of the microgrid and subway system with and without the optimal policy. The dot indicates the mean value while the error bar shows the variance of the sample trajectory. Our optimal policy largely reduces the variance and thus makes the system more stable.} \label{systemresilience} \end{figure} We compare three other policies. Policy $1$ is a straightforward strategy, i.e., to repair every faulty node and leave alone working nodes. Policy $2$ represents a random strategy that chooses to restore each outage node with probability $4/5$ and repair each non-outage node with probability $1/5$. Taking no action is applied as policy $3$. Our simulation results show that policy $1,2,3$ achieve the value of $979.81$,$947.17$ and $796.87$ with initial state $\mathbf{X}=\mathbf{1}$, while our optimal policy achieves a higher value of $985.44$. The optimal value obtained by ALP is $989.2963$, which is close to the `actual value' obtained from the simulation. Note that it is computationally expensive to directly solve the exact LP. We can approximate the exact value by simulating over a long horizon and averaging over a sufficient large number of samples. Finally, we study the case when only part of nodes is controllable. In Fig. \ref{givencontrol}, we maintain the resilience of some nodes with a sparse manageable set. For example, optimal control of nodes $6$ and $8$ keeps on average four nodes healthily working. More controllable nodes lead to a larger value function as shown in Fig. \ref{increasevalue} and the growth is approximately linear with the size of controllable nodes. \subsection{Large-scale Demo } We illustrate the cascading failure of a two-layer network\footnote{https://drive.google.com/drive/u/0/folders/0B6-Q8-SnvO6lYmlxX2FuN3ZuV1U} with 100 nodes representing mixed components. The failure probability of each node is proportional to the number of its connected faulty nodes and is illustrated via the color bar. A lighter color corresponds to a higher failure probability, and the white color means the node has failed. If a node is repaired or maintained, then it will not fail in a few stages afterward and is shown in black. We introduce both edge-based and node-based resilience metrics, i.e., connectivity and the reward of the working nodes, respectively. They are normalized to 100 with their maximum values. As we can see from the video demo, the system starts with a two-node failure and propagates to the neighboring nodes quickly without effective controls. On the contrary, the dynamic recovery demo shows how the network can recover quickly and wholly from the catastrophic event where almost all nodes fail. Also, the resilient planning manages to maintain a stable, healthy state of the entire network in the long run after the recovery. We assume that the limited resources can only support one node recovery at each stage to obtain a clear visualization. Our optimal control policy achieves a tradeoff between a high-level resilience and the cost it takes. \section{CONCLUSION} \label{conclusion} In this paper, we have formulated a factored MDP model for large-scale ICIs and significantly reduced the complexity of computing optimal policies. A distributed optimal control policy is designed to enhance the resilience and security of ICIs. The proposed framework has been applied to a case study motivated by Hurricane Sandy. We have shown that our policy manages to harden the security by reducing the failure probability and achieve resilience by acting on proper nodes. Moreover, the cost achieved by the found policies is the lowest. The framework developed in this work would be useful to mitigate large-scale infrastructure networks through optimal dynamic resilience planning. Even with a small subset of hardened nodes through planning, we have observed a sizable loss reduction for extreme events. We have used microgrids and subway systems in lower Manhattan and downtown Brooklyn as a case study to illustrate the significant resilience improvement in event of Hurricane Sandy. Our framework can be applied to develop resilience policies for other interdependent infrastructures including water distribution systems, food systems, data centers, and communication infrastructures. The future work would aim at understanding implementation and resource constraints on the policies and developing a large-scale solver that would provide usable infrastructure solutions. \section*{ACKNOWLEDGMENTS} This research is partially supported by a DHS grant through Critical Infrastructure Resilience Institute (CIRI) and grants SES-1541164, and ECCS-1550000 from National Science Foundation (NSF). \bibliographystyle{wsc} \section*{\abstractname}} {\par} \newdimen\labelwidthi \newdimen\labelwidthii \settowidth{\labelwidthi}{M} \settowidth{\labelwidthii}{(d)} \leftmargini=1.2cm \leftmarginii=4ex \leftmarginii=4ex \def\@listi{\leftmargin\leftmargini \rightmargin0pt \parsep 0\p@ \topsep 10\p@ \itemsep0\p@} \let\@listI\@listi \@listi \def\@listii{\leftmargin\leftmarginii \rightmargin0pt \labelsep 1ex \parsep 0\p@ \topsep 0\p@ \itemsep0\p@} \let\@listII\@listii \@listii \def\@listiii{\leftmargin\leftmarginiii \rightmargin0pt \parsep 0\p@ \topsep 0\p@ \itemsep0\p@} \let\@listIII\@listiii \@listiii \labelsep=.25in \setlength \labelwidth{\leftmargini} \addtolength\labelwidth{-\labelsep} \renewcommand\labelenumi{\rlap{\rm\theenumi.}} \renewcommand\labelitemi{\rlap{\textbullet}} \def\@seccntformat#1{\hbox to .25in{\csname the#1\endcsname\hss}\relax} \def\@sect#1#2#3#4#5#6[#7]#8{% \ifnum #2>\c@secnumdepth \let\@svsec\@empty \else \refstepcounter{#1}% \protected@edef\@svsec{\@seccntformat{#1}\relax}% \fi \@tempskipa #5\relax \ifdim \@tempskipa>\z@ \begingroup #6{% \@hangfrom{\hskip #3\relax\@svsec}% \interlinepenalty \@M #8\@@par}% \endgroup \csname #1mark\endcsname{#7}% \addcontentsline{toc}{#1}{% \ifnum #2>\c@secnumdepth \else \protect\numberline{\csname the#1\endcsname}% \fi #7}% \else \def\@svsechd{% #6{\hskip #3\relax \@svsec #8}% \csname #1mark\endcsname{#7}% \addcontentsline{toc}{#1}{% \ifnum #2>\c@secnumdepth \else \protect\numberline{\csname the#1\endcsname}% \fi #7}}% \fi \@xsect{#5}} \deftable{table} \long\def\@makecaption#1#2{% \ifx\@captypetable \vskip3pt \else \vskip\abovecaptionskip\fi \sbox\@tempboxa{#1: #2}% \ifdim \wd\@tempboxa >\hsize #1: #2\par \else \global \@minipagefalse \hb@xt@\hsize{\hfil\box\@tempboxa\hfil}% \fi \vskip\belowcaptionskip} \renewcommand\section{\@startsection {section}{1}{\z@}% {-12pt}% {6pt}% {\hyphenpenalty10000\normalfont\normalsize\bfseries}} \renewcommand\subsection{\@startsection{subsection}{2}{\z@}% {-12pt}% {6pt}% {\normalfont\normalsize\hyphenpenalty10000\bfseries}} \renewcommand\subsubsection{\@startsection{subsubsection}{3}{\z@}% {-12pt} {6pt} {\normalfont\normalsize\hyphenpenalty10000\bfseries}} \let\savesubsub\subsubsection \def\subsubsection#1{\savesubsub{\ \ #1}} \parskip=0pt plus .01pt \let\saveparagraph\paragraph \def\paragraph#1{\vskip1sp {\bf #1}\hskip1em\relax} \raggedbottom \newenvironment{hangref}{\begin{list}{}{\setlength{\itemsep}{0pt} \setlength{\parsep}{0pt}\setlength{\rightmargin}{0pt} \setlength{\leftmargin}{+\parindent} \setlength{\itemindent}{-\parindent}}}{\end{list}} \newif\iftitle \def\@oddhead{\iftitle\global\titlefalse \vtop to0pt{\hbox to.9\textwidth{\titlepageheadfont Proceedings of the \currentYear{} Winter Simulation Conference\hfill}% \vskip2pt \hbox to .9\textwidth{\titlepageheadfont \currentEditors , eds.\hfill}% \vss} \else \hbox to\textwidth{\titlepageheadfont\hfill\thetitle\hfill}\fi} \def\@evenhead{\iftitle\global\titlefalse\fi% \hbox to \textwidth{\hss\titlepageheadfont \theauthors\hss}} \let\@oddfoot\relax \let\@evenfoot\@oddfoot \def\ttitle#1{\gdef\thetitle{#1}} \def\vskip12pt{\vskip12pt} \let\vs\vskip12pt \def\bd{\vskip2pt\noindent} \spaceskip=3.5pt plus 2pt minus2pt \parindent=.25in \hfuzz=1pt \widowpenalty=10000 \clubpenalty=10000 \def\verbatim{\spaceskip=0pt \@verbatim \frenchspacing\@vobeyspaces \@xverbatim} \newcommand\smscriptsize{\@setfontsize\scriptsize\@vipt\@viipt} \def\def\@captype{figure}{\def\@captype{figure}} \let\endnofloatfigure\relax \def\def\@captype{table}{\def\@captype{table}} \let\endnofloattable\relax \newcount\itemcount \def\spenumerate{\bgroup\leftskip=.25in \global\itemcount=0 \def\item{\global\advance\itemcount by 1 \vskip1sp \noindent\hskip-.25in\hbox to .25in{\the\itemcount.\hss}}} \def\vskip12pt\egroup{\vskip12pt\egroup} \newif\ifnoindent \def\@begintheorem#1#2{\vskip-12pt\vskip1sp \trivlist \item[\ifnoindent\global\noindentfalse\else \hskip.25in\fi\hskip \labelsep{\bfseries #1\ #2}]\itshape} \def\@opargbegintheorem#1#2#3{\vskip-12pt\vskip1sp \trivlist \item[\ifnoindent\global\noindentfalse\else\hskip.25in\fi% \hskip \labelsep{\bfseries #1\ #2\ (#3)}]\itshape} \def\@endtheorem{\vskip1sp} \newlength{\bibhang} \setlength{\bibhang}{2em} \newdimen\bibindent \bibindent=.25in \@ifundefined{refname}% {\@ifundefined{chapter}% {\newcommand{\refname}{References}}% {\newcommand{\refname}{Bibliography}}% }% {}% \def\thebibliography#1{\section*{\refname\@mkboth {\uppercase{\refname}}{\uppercase{\refname}}}\list {[\arabic{enumi}]}{\settowidth\labelwidth{[#1]} \rightmargin=0pt \leftmargin=0pt \leftmargin\labelwidth \advance\leftmargin\labelsep \advance\leftmargin\bibindent \advance\leftmargin-24pt \itemindent -\bibindent \listparindent \itemindent \parsep \z@ \usecounter{enumi}} \def{} \sloppy \sfcode`\.=1000\relax} \endinput \makeatletter \def\@lbibitem[#1]#2{\item[ \if@filesw {\let\protect\noexpand\immediate\write\@auxout{ \string\bibcite {#2}{#1}}}\fi\ignorespaces} \def\@cite#1#2{{#1\if@tempswa , #2\fi}} \makeatother \def\Box{\vbox to 6pt{\hrule \hbox{\vrule height 4.8pt \hskip 4.8pt \vrule } \hrule}} \end{filecontents*} \begin{filecontents*}{wscbib.tex} \makeatletter \let\@internalcite\cite \def\cite{\def\@citeseppen{-1000}% \def\@cite##1##2{(##1\if@tempswa , ##2\fi)}% \def\citeauthoryear##1##2##3{##1 ##3}\@internalcite} \def\citeNP{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa , ##2\fi}% \def\citeauthoryear##1##2##3{##1 ##3}\@internalcite} \def\citeN{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa, ##2)\else{}\fi}% \def\citeauthoryear##1##2##3{##1 (##3)}\@citedata} \def\citeA{\def\@citeseppen{-1000}% \def\@cite##1##2{(##1\if@tempswa , ##2\fi)}% \def\citeauthoryear##1##2##3{##1}\@internalcite} \def\citeANP{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa , ##2\fi}% \def\citeauthoryear##1##2##3{##1}\@internalcite} \def\def\citeauthoryear##1##2##3{##2, ##3}\@internalcite{\def\@citeseppen{-1000}% \def\@cite##1##2{(##1\if@tempswa , ##2\fi)}% \def\citeauthoryear##1##2##3{##2 ##3}\@internalcite} \def\shortciteNP{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa , ##2\fi}% \def\citeauthoryear##1##2##3{##2 ##3}\@internalcite} \def\shortciteN{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa, ##2\else{}\fi}% \def\citeauthoryear##1##2##3{##2 (##3)}\@citedata} \def\def\citeauthoryear##1##2##3{##2}\@internalcite{\def\@citeseppen{-1000}% \def\@cite##1##2{(##1\if@tempswa , ##2\fi)}% \def\citeauthoryear##1##2##3{##2}\@internalcite} \def\shortciteANP{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa , ##2\fi}% \def\citeauthoryear##1##2##3{##2}\@internalcite} \def\def\citeauthoryear##1##2##3{##3}\@internalcite{\def\@citeseppen{-1000}% \def\@cite##1##2{(##1\if@tempswa , ##2\fi)}% \def\citeauthoryear##1##2##3{##3}\@citedata} \def\citeyearNP{\def\@citeseppen{-1000}% \def\@cite##1##2{##1\if@tempswa , ##2\fi}% \def\citeauthoryear##1##2##3{##3}\@citedata} \def\@citedata{% \@ifnextchar [{\@tempswatrue\@citedatax}% {\@tempswafalse\@citedatax[]}% } \def\@citedatax[#1]#2{% \if@filesw\immediate\write\@auxout{\string\citation{#2}}\fi% \def\@citea{}\@cite{\@for\@citeb:=#2\do% {\@citea\def\@citea{, }\@ifundefine {b@\@citeb}{{\bf ?}% \@warning{Citation `\@citeb' on page \thepage \space undefined}}% {\csname b@\@citeb\endcsname}}}{#1}}% \def\@citex[#1]#2{% \if@filesw\immediate\write\@auxout{\string\citation{#2}}\fi% \def\@citea{}\@cite{\@for\@citeb:=#2\do% {\@citea\def\@citea{, }\@ifundefine {b@\@citeb}{{\bf ?}% \@warning{Citation `\@citeb' on page \thepage \space undefined}}% {\csname b@\@citeb\endcsname}}}{#1}}% \def\@biblabel#1{} \makeatother \newdimen\bibindent \bibindent=0.0em \def\thebibliography#1{\section*{\refname}\list {}{\settowidth\labelwidth{[#1]} \leftmargin\parindent \itemindent -\parindent \listparindent \itemindent \itemsep 0pt \parsep 0pt} \def{} \sloppy \sfcode`\.=1000\relax} \end{filecontents*} \section*{\refname\@mkboth \def\thebibliography#1{\section*{\refname}\list {}{\settowidth\labelwidth{[#1]} \leftmargin\parindent \itemindent -\parindent \listparindent \itemindent \itemsep 0pt \parsep 0pt} \def{} \sloppy \sfcode`\.=1000\relax}

0a8bbaffd2dace0eaa3e3ce43b56e7623b900724

\section{Introduction}\label{sec:intro} Modern astrophysical and cosmological observations have shown that only a small fraction of the universe is composed of ordinary matter. Recent measurements predict that $\sim26\%$ of mass-energy in the universe is dark matter.\cite{PlanckResult} The makeup of the dark matter remains unknown, though several candidates have been proposed. Among them is the axion, a light hypothetical pseudoscalar particle.\cite{PhysRevLett.38.1440,PhysRevD.16.1791,PhysRevLett.40.223,PhysRevLett.40.279,PRESKILL1983127} Detecting a dark matter particle is inherently challenging because the dark matter has extremely weak interactions with ordinary matter. Axion detection schemes are largely based on an axion--photon coupling interaction known as Primakoff conversion.\cite{PhysRevLett.51.1415} In the presence of a strong magnetic field, the axion scatters off of a virtual photon and converts to a real photon that can be detected. The frequency of the resulting photon corresponds directly to the axion mass as \begin{equation} \nu_a = \frac{m_ac^2}{h}\,, \label{eq:nua} \end{equation} where $m_a$ is the axion mass, $c$ is the speed of light, and $h$ is the Planck constant. Both $m_a$ and the coupling strength of the axion--photon conversion $g_{a\gamma\gamma}$ are unknown, but the parameter space is bounded by astrophysical constraints and observations.\cite{Graham} Axion searches aim to scan the possible parameter space for a signal and eliminate regions where no signal is detected. The axion--photon coupling is expected to be very weak and therefore the power of an axion photon signal would be very small (on the order of $10^{-24}$ W).\cite{ALKENANY201711} Existing searches for galactic halo axions use detectors known as haloscopes. The Haloscope At Yale Sensitive To Axion CDM (HAYSTAC) is a haloscope searching specifically for Cold Dark Matter (CDM) axions, which are assumed to have a uniform `cold' virial velocity. In these devices, a tunable electromagnetic cavity is placed in the bore of a solenoid. The goal is to match the frequency of a resonant cavity mode of choice to $\nu_a$ in order to enhance the signal from the converted photon. An antenna is used to extract the signal, which is then amplified and analyzed. A photon resulting from inverse Primakoff conversion would appear as a power excess. A block diagram of this detection technique is shown in Figure \ref{fig:blockdiagram}. Haloscopes operate at cryogenic temperatures and use electronics operating at or near the Standard Quantum Limit for noise in order to be sensitive to $P\sim10^{-24}$ W. \begin{figure} \includegraphics[width=\columnwidth]{blockdiagramnew-eps-converted-to.pdf} \caption{Block diagram of a haloscope detector. A resonant cavity is placed inside the bore of a superconducting solenoid. An axion in the cavity converts to a photon via Primakoff effect. This signal is extracted and amplified. The spectrum in the bottom right corner shows the expected shape of an axion signal, which appears as a power excess above the background. The shape of the axion results from its virial velocity in the galactic halo. \label{fig:blockdiagram}} \end{figure} The haloscope signal power scales as \begin{equation} P \propto B_0^2VQ_LC_{nm\ell}g^2_{a\gamma\gamma}\,, \label{eq:power} \end{equation} where $B_0$ is the magnetic field supplied by the solenoid, $V$ is the cavity volume, and $Q_L$ is the loaded cavity quality factor. $C_{nm\ell}$ is the `form factor' of the resonant cavity mode, defined as \begin{equation} C_{nm\ell} \equiv \frac{\left(\int{d^3\bm{x}\bm{\hat{z}}\cdot\bm{E}_{nm\ell}\left(\bm{x}\right)}\right)^2}{V\int{d^3\bm{x}\varepsilon\left(\bm{x}\right)\left|\bm{E}_{nm\ell}\right|^2}}\,. \label{eq:formfactor} \end{equation} Here, $\bm{E}_{nm\ell}$ is the electric field of the mode, $V$ is the cavity volume, and $\varepsilon$ is the relative permittivity. Since the applied magnetic field is in the $z$ direction, the form factor quantifies the overlap between the external applied magnetic field and the electric field of the mode. Current experiments use cylindrical microwave cavities with cylindrical metal tuning rods. These cavities are simple from a resonator design perspective, but use in axion haloscopes poses unique constraints and challenges. The speed of axion searches is limited by the haloscope signal power, which in turn depends heavily on cavity performance. In designing the first axion haloscope cavities, Hagmann \textit{et al.}\cite{Hagmann} recognized the importance of studying and limiting non-ideal cavity behavior. Their analysis resulted in the development of the haloscope cavities that are in use today. Now, modern computational tools are capable of modeling these cavities in much greater detail. Coupling these simulations with measurement techniques used in other cavity applications, it is possible to study the imperfections of existing cavities and develop a predictive capability for new designs. This paper outlines the use of standard microwave measurement techniques and simulation software to fully characterize the cavity for the HAYSTAC detector. This is the first use of the bead perturbation technique to study an axion haloscope cavity and the first attempt to match these measurements with 3D electromagnetic simulations of the cavity with imperfections. \section{HAYSTAC}\label{sec:HAYSTAC} HAYSTAC is a collaboration of Yale University, the University of Colorado Boulder, and the University of California Berkeley. Housed at Yale, the detector uses a 9 T superconducting magnet and is cooled to 127 mK using a dilution refrigerator. Josephson Parametric Amplifiers (JPAs) are used to amplify the microwave signal with quantum-limited added noise. With the cavity described in Section \ref{subsec:cavity}, HAYSTAC has excluded axion masses of 23.15--24.0 $\mu\text{eV}$ within the axion dark matter model band.\cite{PhysRevLett.118.061302,PhysRevD.97.092001} \subsection{Cavity design}\label{subsec:cavity} The current HAYSTAC cavity consists of a right cylinder barrel with two endcaps and a large (radius $r = 1$ in) cylindrical tuning rod. The barrel has an inner radius $R = 4$ in and length $L = 10$ in. The barrel and endcaps are constructed from stainless steel and plated with copper for good electrical and thermal conductivity at cryogenic temperatures. Measurements for this study were performed at room temperature using an aluminum twin of the main cavity barrel, as the main cavity was in use for data collection. In both the main cavity and the twin cavity used for this study, the tuning rod is hollow with a copper-plated stainless steel body. The tuning rod has an off-axis axle composed of a steel spindle inside the body of the rod and an alumina tube on each end (outer diameter $d = 0.25$ in). The rod is slightly shorter than 10 in, leaving small ($\sim 0.01$ in) gaps between the rod and the endcaps. While these gaps are necessary in order for the rod to rotate, they have a substantial impact on the cavity performance. This is discussed in more detail in Section \ref{subsec:future}. Turrets are attached to the endcaps to hold the alumina portions of the tuning rod axle. A ball bearing is the only point of contact between the axle and the turret, allowing for frictionless rotation. During HAYSTAC data collection, the cavity mode of interest is the lowest order transverse magnetic mode, the TM$_{010}$. This mode has an electric field purely along the cavity axis, giving it the highest possible form factor of any fundamental cavity mode. To tune the cavity, the rod is rotated from being concentric with the cavity body ($0^{\circ}$, the maximum frequency position) to being nearly tangent with the wall ($180^{\circ}$, the minimum frequency position). Photographs of the cavity and tuning rod are shown in Figure \ref{fig:cavityphotos}; Figure \ref{fig:rodangles} shows a schematic of the tuning extrema. This 180$^{\circ}$ rotation tunes the TM$_{010}$ mode between 3.4--5.8 GHz. \begin{figure} \includegraphics[width=\columnwidth]{cavityphotos-eps-converted-to.pdf} \caption{(a) Photograph of the assembled HAYSTAC cavity before being inserted into the magnet. (b) The HAYSTAC cavity with the top endcap removed, showing the tuning rod. Both the cavity and the rod are stainless steel plated with copper. A portion of the top axle is also visible. \label{fig:cavityphotos}} \end{figure} \begin{figure} \includegraphics[width=\columnwidth]{figrodangles-eps-converted-to.pdf} \caption{Top-down schematic of the cavity (light gray) and rod (dark gray) showing the angle definitions for different rod positions. \label{fig:rodangles}} \end{figure} \subsection{Cavity challenges and limitations}\label{subsec:cavchallenges} Right cylindrical cavities are well-suited for axion haloscope experiments at microwave frequencies. However, there are challenges in using even simple cavity designs. Initial scans have been performed in narrow regions of the cavity tuning range, but as the community moves to cover more parameter space it will be necessary to tune through regions where performance is limited. The cavity performance impacts the signal power through three parameters: the volume, the quality factor, and the form factor. It is therefore vital to understand the cavity performance in problematic regions and make improvements wherever possible. \subsubsection{Mode crossings}\label{subsubsec:probsmodecross} Microwave resonators have many fundamental eigenmodes, but few of these modes have large enough $C_{nm\ell}$ values to be useful in current haloscopes because their $E_z$ components are small or zero. When the tuning rod is rotated, all of the TM (transverse magnetic) eigenmodes tune. The transverse electric (TE) and transverse electromagnetic (TEM) modes present in the cavity are unaffected by this tuning motion. Thus as the frequency of the TM$_{010}$ mode is tuned, the mode crosses stationary TE and TEM modes (generically referred to as `intruder modes'). In an ideal cavity there would be no interaction between the modes, but in reality they mix and form hybrid modes. This degrades the $E_z$ field strength and lowers $C_{010}$. Data taken in these regions is effectively unusable for excluding axion parameter space. To date, HAYSTAC and other haloscopes have circumvented this problem by focusing on frequency ranges that are free from mode crossings and by removing any compromised portions of the data from new exclusion limits. As HAYSTAC moves to perform scans over broader frequency ranges and at higher frequencies, this solution is no longer tenable. Given the density of modes in the HAYSTAC cavity, there are few regions that are entirely free from mode crossings. A broad sweep will therefore be broken up by several of these unusable regions. The density of modes is also higher at higher frequencies, making it difficult to find substantial crossing-free regions. Understanding the nature of mode crossings and characterizing the range over which mixing occurs may allow for useful data to be obtained over part or all of the crossing. A major goal of this study was to characterize the mode crossings in the HAYSTAC cavity such that portions of these regions may be used despite the mode hybridization. \subsubsection{Alignment and mode localization}\label{subsubsec:probslocalization} In the ideal case, the TM$_{010}$ electric field strength is constant along the length of the cavity. The entire field strength is in the $z$ component of the field except at boundaries, which gives the TM$_{010}$ its high form factor. Deviations from perfect alignment disrupt this ideal behavior and cause field strength localization. Notably, tilting the tuning rod with respect to the cavity barrel causes an asymmetry in the axial direction, which forces the field to localize towards one end of the cavity. The impact of this mode localization on the cavity parameters $Q_L$ and $C_{010}$ has been largely unstudied, and power calculations have generally relied on idealized $C_{010}$ values. This work details efforts to characterize and quantify the effects of small tilts on these parameters through measurement and simulation, allowing for improved accuracy in power calculations. This study also aims to demonstrate the predictive capability of simulations, which can be used to inform future cavity designs. \section{Measurement and simulation techniques}\label{sec:techniques} \subsection{Bead perturbation}\label{subsec:bead} In order to study mode localization, it is necessary to determine the relative field strength at different locations inside the cavity. To perform a bead perturbation measurement, a small metal or dielectric object is inserted into a resonator to locally perturb the electric field and cause a shift in the resonant frequency of an eigenmode. By performing this measurement at different locations, a `map' of the field strength can be constructed. This technique is commonly used in accelerator physics to study the field profile of a cavity or set of cavities along the axis of acceleration.\cite{BP1,BP2,BP3} For axion haloscope cavities, a series of bead perturbation measurements can be taken along $z$ at one radial and azimuthal position (referred to in this work as a `bead-pull' measurement). Since the TM$_{010}$ mode has a constant $E\left(z\right)$ for any given $r$ and $\theta$, the result should be a profile with a single, constant value for the perturbed frequency at each axial position. Any deviations from this flat profile indicate there is mode localization causing the field to be stronger at some location. The details of how this measurement technique was used to characterize the HAYSTAC cavity are discussed in Section \ref{subsec:setup}. \subsection{Electromagnetic simulations}\label{subsec:fem} Electromagnetic cavities can be simulated in 3D using finite element modeling. The finite element method (FEM) is used across many disciplines to solve systems of equations in complex geometries. In FEM problems, a geometry is broken up into discrete elements over which the system of equations is easier to solve. Together, the elements form a mesh. Numerical solutions are found for each element and are used to construct a solution for the entire geometry. This can be applied to full 3D systems with many constituent materials and components of different sizes. The use of 3D electromagnetic (EM) simulations allows for the design of cavities and other devices which cannot be described analytically. By solving Maxwell's Equations, 3D EM simulations can predict the eigenmodes of electromagnetic resonators. Material information can be included to determine expected $Q$ values and losses. When excitation sources such as antennas or waveguide ports are included, 3D EM simulations can determine the expected transmission and reflection behavior of a structure. All of this information is used in the design of microwave cavities and can be compared to real values measured in a structure. In this study, simulations were performed using a commercial electromagnetic simulation software, Computer Simulation Technology Microwave Studio (CST MWS).\cite{CST} \section{Methodology and results}\label{sec:methodology} \subsection{Measurement setup}\label{subsec:setup} The bead-pull measurement setup has three main components: the beadline and a stepper motor system, a vector network analyzer (VNA), and a LabVIEW script that controls the VNA and the motor system. The beadline consists of a cylindrical sapphire bead of length 0.196 in and diameter 0.169 in that is glued onto a Kevlar\cite{Kevlar} string. The string passes through a hole in the endcaps, situated at $r=1.25$ in (with respect to the center of the cavity). When properly aligned, the bead can traverse the cavity at a constant radial and azimuthal position. Motion of the beadline is controlled by a system of pulleys and a stepper motor. The pulleys are adjusted such that the beadline is parallel to the z-axis of the cavity. The Kevlar line is attached to an Applied Motion Products STM 23S-2EE stepper motor. The motor rotates in predetermined steps which lead to changes in the vertical position of the sapphire bead. A second stepper motor is attached to the tuning rod axle, providing fine control of the rod's rotation for tuning. The cavity and motor are attached to a frame separate from the beadline system. Vibration isolation is included to prevent the motor from inducing vibrations in the cavity. Two antennas made from semi-rigid coax are inserted through the bottom endcap of the cavity. A Keysight E5071C Network Analyzer provides information about the transmission and reflection measurements between the two antenna ports throughout the frequency range of the cavity. Transmission measurements $\left(S_{21}\right)$ are used to determine the frequencies and $Q$ values of the cavity modes, which appear as peaks on a spectrum of power versus frequency. The stepper motors and the VNA are controlled by a LabVIEW program. For each measurement, the user sets the desired number of bead steps and the frequency of the mode of interest. The rest of the process is automated: the bead is moved and the software records the central frequency of the peak in $S_{21}$. After all steps are performed, the software plots the central frequency versus bead position. This `profile' of the mode is used to determine if there is mode localization or hybridization. \subsection{Tilt study}\label{subsec:tiltstudy} Misalignments of the rod with respect to the cavity cause mode localization that can generally be observed through bead-pull measurements. This mode localization manifests as a shift in the resonant frequency of the TM$_{010}$ mode from one end of the cavity to the other. A primary goal of this study was to quantify this shift and relate it to a physical amount of misalignment while studying its impact on cavity performance. The frequency shift was defined as \begin{equation} \Delta f \equiv f\left(z=\text{cavity bottom}\right)-f\left(z=\text{cavity top}\right)\,. \label{eq:delta} \end{equation} Non-zero $\Delta f$ indicates there is some mode localization. Figure \ref{fig:deltaf} shows two example bead-pull profiles with different $\Delta f$ values. \begin{figure} \includegraphics[width=\columnwidth]{Deltaf_comp-eps-converted-to.pdf} \caption{Two profiles of the TM$_{010}$ mode produced using bead-pull measurements. The frequency of the mode shows a dependence on the position of the bead in the cavity. The abrupt changes in frequency on the two ends correspond to the points where the bead enters and exits the cavity. An approximately flat bead-pull profile (a) is nearly symmetric about the cavity center. Profiles for the misaligned case (b) exhibit a non-negligible $\Delta f$. In this case the $\Delta f = -90\ \textrm{kHz}$. The curved shape of the profile arises from mode localization in the gaps between the endcaps and tuning rod. This is discussed in further detail in Section \ref{subsec:comparison}. \label{fig:deltaf}} \end{figure} The rod misalignment was produced by laterally shifting the rod's axle at one end of the cavity. Two orthogonal micrometers, as shown in Figure \ref{fig:microm}, were mounted on the turret on the bottom endcap of the cavity around the alumina axle. By adjusting the micrometers, the rod can be tilted in a controlled and measurable manner. Rod tilt studies were conducted at three approximate rod positions $\varphi$: 0$^\circ$, 90$^\circ$, and 180$^\circ$, which correspond to the positions illustrated in Figure \ref{fig:rodangles}. These positions were chosen because they covered the extremal frequency cases and a central frequency. Once the rod was adjusted to a position of interest, a tight collar was placed on the top alumina axle to hinder rotations of the rod within the cavity. This reduced the degrees of freedom of the system to the two orthogonal tilts of the rod. The micrometers were set such that the starting values corresponded to $\Delta f = 0$, taken to be an aligned configuration. \begin{figure} \includegraphics[width=\columnwidth]{microm-eps-converted-to.pdf} \caption{Micrometer system used to control tilt of rod. The tight collar is not shown in this photo.\label{fig:microm}} \end{figure} The tilt studies were carried out by adjusting the micrometers in the two orthogonal directions and taking a bead-pull measurement at each position. A square grid of micrometer positions corresponding to rod misalignments was created. In the 0$^\circ$ and the 180$^\circ$ cases, the measurements were taken based off of a $9\times9$ square grid of 0.8 mil steps in the two orthogonal directions, i.e. the maximum micrometer displacements were $\pm4\times0.8\ \textrm{mils}=\pm3.2\ \textrm{mils}$. In the 90$^\circ$ case, a decrease in the mode's sensitivity to misalignments demanded an increase in the grid's dimensions in order to see an appreciable impact. This resulting grid was $7\times 7$ with a 1.2 mil step size between points. These measurements kept the angular misalignment between the rod and the cavity below 1 millirad. The behavior of the cavity at the three rod positions resulted in a plane-like surface in $\Delta f$, as shown in Figure \ref{fig:tiltplanes}a. The orientation of the set of points along which $\Delta f \approx 0$ showed a dependence based on the angle of rotation of the rod. \begin{figure} \includegraphics[width=\columnwidth]{tiltplanes-eps-converted-to.pdf} \caption{Planes created by studying $\Delta f$ as a function of rod misalignment from (a) the measurements and (b) the simulations. The misalignment is depicted in terms of the step size of the micrometer. One step in each direction corresponds to an adjustment of the microemeter's position by 0.8 mils. These planes correspond to a rod rotation of $\varphi=3.3^{\circ}$. \label{fig:tiltplanes}} \end{figure} The same misalignments were input into a CST MWS model of the cavity. Bead-pull profiles were constructed by simulating the measurement process: a sapphire bead was stepped through the cavity and the frequency of the TM$_{010}$ mode was determined from $S_{21}$. The simulated bead-pulls were designed such that they matched the conditions of the measured bead-pulls. Stepping of the bead was performed using a parametric sweep based on the bead's position with mesh adaptation between steps. These simulated bead-pulls were performed at each tilt position in the measurement grids. Figure \ref{fig:tiltplanes}b shows the resulting plane in $\Delta f$. Additional eigenmode simulations were performed with various tilts to study the impact of tilt and mode localization on the form factor $C_{010}$. It has previously been assumed that any mode localization would cause a reduction in $C_{010}$. However, simulations showed that small amounts of axial mode localization introduced by the tilts had negligible impacts on $C_{010}$ because the overall $E_z$ is conserved. Under ideal conditions, the form factor is largest when the field configuration is most symmetric at $\varphi = 0^{\circ}$, decreases through $\varphi = 90^{\circ}$, and increases near $\varphi = 180^{\circ}$. Simulations at each of these three positions showed that, for a given tilt, the impact is largest when the form factor is highest. Tilts in the range of 4--5 mils resulted in decreases in the form factor of 2.4\%, 0.5\%, and 1.5\% for $\varphi = 0^{\circ},~90^{\circ},~\text{and}~180^{\circ}$, respectively. For smaller tilts in the range of 1--4 mils, the shift in form factor was generally less than 0.5\%. \subsection{Mode crossings study}\label{subsec:modestudy} The primary goal of the mode crossings study was to determine the width of mode crossings between the TM$_{010}$ mode and other TE and TEM modes in the cavity. The width of the mode crossing was defined to be the frequency range across which the two modes were sufficiently hybridized such that the effects of this hybridization were visible on a bead-pull profile. Effects of hybridization were observed on bead-pull profiles when the profile of a TE or TEM mode was superimposed on the profile of the TM$_{010}$ mode, as shown in Figure \ref{fig:modecrossing}. The regions in which the modes mix are used to determine the frequency ranges over which the TM$_{010}$ mode can no longer be used due to its hybridization. \begin{figure*} \centering \includegraphics[width=\textwidth]{modecrossings-eps-converted-to.pdf} \caption{Examples of bead-pull profiles near the edge and the middle of a mode crossing. (a) Shows the very beginning of a mode crossing. This is identified by the introduction of small oscillations in the TM$_{010}$ mode profile (left). (b) As the modes approach in frequency, they hybridize and their profiles superimpose. At this point, neither mode can be used for data acquisition. The bead-pull profile as well as 3D EM simulations illustrate that the mode with the oscillatory profile is the TE$_{060}$ mode. \label{fig:modecrossing}} \end{figure*} This study was conducted by performing a series of bead-pull measurements with small rotations of the tuning rod. The rod position $\varphi$ was controlled by a stepper motor so very small steps could be made. Bead-pull measurements were taken every 5 MHz in frequency throughout the entire tuning range in order to look for mode crossings. When a mode crossing was encountered, the rod tilt was adjusted to give $\Delta f = 0$ on the TM$_{010}$ mode. The rod was then rotated in steps corresponding to 2.5 MHz in frequency to bring the frequency of the TM$_{010}$ mode closer to that of the intruder mode. This was repeated until the TM$_{010}$ mode emerged at a frequency higher than that of the intruder mode and mixing was no longer observed. During each step in rotation, bead-pull measurements were taken on both peaks in $S_{21}$, producing two profiles. The alignment of the rod was not adjusted throughout the mode crossing as a control. The frequency range of mode crossings was sufficiently small such that the change in $\Delta f$ was negligible across the range of rotation. These studies showed that approximately 15\% of the available frequency range of the TM$_{010}$ mode contains significant mode mixing that must be considered during axion search data collection. Several observed mode crossings proved to be fairly innocuous since no mixing was observed unless the mode with the wider and more dominant peak (in all such cases, the TM$_{010}$ mode) completely overlapped with the intruder mode. Stronger and wider mode crossings exhibited one of two behaviors. In some cases, the frequency of one hybridized mode remained constant while the other mode tuned across the frequency range of interest. The stationary hybridized mode corresponded to the peak which was initially the intruder mode. This type of crossing is illustrated in Figure \ref{fig:avoidedcrossings}a. In the remaining cases, the mode that initially was at the lower frequency tuned to the original frequency of the intruder mode, while the intruder mode tuned away and eventually became the TM$_{010}$ mode. In these cases, the two peaks never truly crossed, making them so-called `avoided crossings,' as shown in Figure \ref{fig:avoidedcrossings}b. \begin{figure} \includegraphics[width=\columnwidth]{avoidedcrossing-eps-converted-to.pdf} \caption{Two types of mode crossings between the TM modes and the intruder modes (IM) in which (a) the intruder mode's frequency remained constant throughout the crossing while the TM mode tuned through it and (b) an `avoided crossing' takes place. \label{fig:avoidedcrossings}} \end{figure} Mode crossing simulations were performed in CST MWS. To model a mode crossing, bead-pulls were simulated as described in Section \ref{subsec:tiltstudy} at multiple rod angles. Profiles were constructed from both the TM$_{010}$ mode and the intruder mode to be compared directly with the measured results. Simulated mode crossings agreed with measured results for the width and behavior of the mode crossing. Additional mode crossing simulations were performed using the properties of copper at cryogenic temperatures. While the $Q$ value increases by roughly a factor of 4 when the cavity is cooled, there is negligible change in the width of mode crossings. For example, the mode crossing shown in Figure \ref{fig:modecrossing} exhibits mixing over 15 MHz, corresponding to $0.4^{\circ}$ in rod rotation. Simulations were performed in steps of $0.05^{\circ}$, corresponding to roughly 1 MHz in frequency. On this order, the widths of crossings were the same at room temperature and cryogenic temperatures. Simulating mode crossings provides additional information which cannot be determined via measurements, including whether the intruder mode is a TE or TEM mode. Bead-pull measurements show variations in the electric field strength along the axis, but do not give information on the orientation of the field which is necessary to distinguish the mode type. In simulations, the full 3D electric and magnetic fields can be calculated for any given frequency thus allowing each intruder mode to be characterized as TE or TEM. \section{Discussion}\label{sec:discussion} \subsection{Comparison of measured and simulated results}\label{subsec:comparison} Across this study, measured and simulated results show good agreement. In all cases, the results are qualitatively the same. Small quantitative differences are found in many cases, which is expected given the limited complexity of the simulated model. One method of illustrating the general agreement between measurements and simulations is to compare the frequencies of intruder modes in a given frequency range. In particular, in the 4--5 GHz frequency range, all measured intruder modes were observed in the simulations. As shown in Table \ref{table:intrudermodes}, their corresponding frequencies showed general agreement with measured values. \begin{table} \begin{tabular}{c c} \toprule \textbf{Simulated Frequencies (GHz)} & \textbf{Measured Frequencies (GHz)} \\ \hline 4.135 & 4.136 \\ \hline \begin{tabular}[c]{@{}c@{}}4.3338\\ 4.335\\ 4.3999\\ 4.4006\end{tabular} & Four modes in 4.327-4.365 region \\ \hline \begin{tabular}[c]{@{}c@{}}4.7264\\ 4.7435\end{tabular} & Two modes in 4.710-4.730 region \\ \hline 4.8257 & 4.820 \\ \hline \begin{tabular}[c]{@{}c@{}}4.8945\\ 4.902\end{tabular} & Two modes in 4.885-4.905 region \\ \botrule \end{tabular} \caption{Tabulated frequencies of the intruder modes in 4--5 GHz range as observed in 3D electromagnetic simulations and in measurements. This corresponds to a range in tuning rod angle $\varphi$ of $49^{\circ}$. \label{table:intrudermodes} } \end{table} Figure \ref{fig:tiltplanes} visually demonstrates that the measured and simulated tilt studies showed the same overall behavior of $\Delta f$ versus tilt. This agreement is found for all three angles where tilt studies were performed. The form factor simulations showed while that $\Delta f$ is a clear measure of a physical misalignment, does not have a perfect correspondence to a certain shift in $C_{nm\ell}$. The impact on $C_{nm\ell}$ has more dependence on whether the rod is tilted towards or away from the center of the cavity. This indicates that while mode localization occurs with tilts, it does not necessarily degrade the form factor. Instead, tilts have the most impact when they cause mode localization in an otherwise highly symmetric field. Thus, $\Delta f$ is best used to match a measured frequency shift to a physical tilt. This physical tilt can then be included in simulations to calculate the form factor. The simulated $\Delta f$ values agree quantitatively with measured values, though there are small differences. There are multiple possible sources for the discrepancies. In the measurement case, the accuracy of the setup is limited. Most notably, very small rotations of the tuning rod are still possible even though the rod is `fixed.' As such, the simulated rod angle may be slightly different from the measurement case. The accuracy of the micrometers is also limited, as they are set manually. There are further known non-idealities which were not included in simulations in order to limit the complexity of the problem. The cavity barrel is not perfectly cylindrical and has small variations both radially and axially. These variations are on the order of 0.005 in or smaller ($\leq 0.15\%$ of the cavity diameter), but it is possible that they impact mode localization. Similarly, the tuning rod has small variations in its radius on the same order. The simulated cavity barrel and tuning rod are perfect cylinders and thus do not capture these potential impacts. Such variations could potentially be included in future studies by creating a full 3D model of the fabricated cavity. An important potential source of error is the size of the gaps between the endcaps and the tuning rod. The length of the rod and the cavity are both known to a high degree of accuracy, but there is no mechanism to ensure the gaps are symmetric on the top and the bottom. When the cavity is assembled, the collar is attached to the top rod axle fixing the axial position of the tuning rod. This position is set such that the rod is not touching either endcap, but it cannot be set with the accuracy necessary to enforce a 0.01 in gap on either side. Thus, it is possible for the gaps to be asymmetric and the value of that asymmetry is unknown. Simulations have shown there is strong mode localization in the gaps which gives rise to the parabolic shape of the profile shown in Figure \ref{fig:deltaf}. The impact varies with rod angle, giving profiles with different concavities. Asymmetry in the gaps can lead to shifts in this bump that may cause errors in calculating $\Delta f$. The simulated model assumed symmetrical gaps and thus did not include this source of error. To achieve better quantitative agreement in $\Delta f$, this gap asymmetry could be studied in more detail. All observed intruder modes were found in the simulations at the expected frequencies. Small differences in resonant frequency between simulations and measurements are expected due to the non-idealities of the cavity that have been discussed. Furthermore, the simulations at cryogenic conditions showed that despite an increase in $Q$, the width of the mode crossings showed no observable temperature dependence. Overall, the simulated room-temperature mode crossings agreed well with measured values, demonstrating that simulations can be used to characterize the severity of mode crossings even before a cavity is fabricated. \subsection{Implications for HAYSTAC and future cavity designs}\label{subsec:future} The purpose of this study was not only to characterize the HAYSTAC cavity using microwave measurement techniques, but also to demonstrate the capability of 3D electromagnetic simulations to replicate the measurements. This opens up the opportunity to obtain more accurate information on cavity performance for the existing HAYSTAC cavity as well as potential future designs. While the $Q$ value of the TM$_{010}$ mode can be measured \textit{in situ} during data collection, the form factor $C_{010}$ can only be calculated from simulated values of the $E$ field. Previous calculations used an idealized cavity. This study demonstrates that commercial 3D EM software is capable of capturing the effects of small non-idealities such as rod misalignments. Future $C_{010}$ calculations can include misalignments and other non-idealities to give more accurate results. After this study was concluded, $C_{010}$ was calculated for the known misalignment in the main HAYSTAC cavity, which was recently refurbished with a new axle. It was observed that the effects on the form factor were negligible, meaning the misalignment and resulting mode localization are at acceptable levels. The form factor for the perfectly aligned case ($\Delta f =0$) at approximately $90^\circ$ was 0.4654 whereas for a tilt of $\Delta f = 87.5$ kHz the form factor reduced to 0.4643, corresponding to a 0.24\% decrease. Similar calculations can be performed for any tuning position used in future data collection. This study also revealed new insight on mode crossings. Mode crossings that were found to be small or benign in measurements showed the same behavior in simulations. In fact, simulations showed that in these cases the intruder mode in question was always a TEM mode. Previously, all mode crossings were viewed as unusable regions. For the existing cavity, these results demonstrate that some regions impacted by a mode crossing have minimal mixing, meaning portions of the mode crossing may still be useful for data collection. While mode crossings will always exist in any conventional cavity, future designs can be optimized to have large regions that are only interrupted by these small TEM mode crossings. \section{Conclusions}\label{sec:conclusions} This study has used microwave measurement techniques to characterize the HAYSTAC detector's resonant cavity in detail. The results will be used to inform how the cavity will be used in future axion search runs. Simulations of each measurement show that 3D electromagnetic models have the ability to capture the effects of non-idealities found in the cavity. There is great potential to continue to expand the model to capture the full complexity of the cavity. Further, future HAYSTAC cavities can be designed with more knowledge of the impact of potential fabrication errors and misalignments on cavity performance. By predicting the impacts of these flaws, designs can be optimized to maximize cavity performance even in non-ideal cases. \begin{acknowledgments} This work was supported by the National Science Foundation, under grants PHY-1362305 and PHY-1607417, and by the Heising-Simons Foundation under grants 2014-182 and 2016-044. We gratefully acknowledge the contributions of Al Kenany for his work on the vibration isolation system in the bead perturbation setup. \end{acknowledgments}

abc15a6be33df7dd5364191fa8e036703401f213

\section{Introduction} 2013 FY27 was discovered on UT March 17, 2013 as part of an ongoing deep and wide survey for extreme Trans-Neptunian objects (TNOs) by Sheppard and Trujillo (2016). The orbit of 2013 FY27, with a semi-major axis near 59 au, eccentricity of 0.39 and inclination of 33 degs, makes it a typical scattered disk object. With a perihelion that comes inside of 36 au, 2013 FY27 can come relatively close to Neptune and have significant gravitational interactions with the planet. The high eccentricity of 2013 FY27's orbit means it experiences large surface temperature variations of some 16 to 22 Kelvin between aphelion and perihelion. Currently 2013 FY27 is the ninth intrinsically brightest TNO and one of the most distant at around 80 au, which is near its aphelion. 2013 FY27 is the intrinsically brightest known TNO that has not had its thermal emission measured (Figure~\ref{fig:KBOalbedo}). The other intrinsically brightest 15 or so TNOs have been observed for their thermal emission by either Spitzer, Herschel or ALMA (Stansberry et al. 2008; Lellouch et al. 2017; Gerdes et al. 2017; Kovalenko et al. 2017; Santos-Sanz et al. 2017). From thermal observations, one can calculate how much sunlight an object absorbs. Optical observations allows one to calculate how much sunlight an object scatters or reflects. These two measurements can then be used to solved for the two unknowns of the size and albedo of an object, using the radiometric method (Lebofsky et al. 1989; Harris 1998; Fernandez et al. 2013). Understanding the size and albedo of 2013 FY27 is important in order to put this intrinsically bright object into context of the other dwarf planets beyond Neptune. The largest TNOs may have formed or evolved in a much different manner than the more moderate and smaller objects. The intrinsic brightness of 2013 FY27 means it is near the interesting transition regime between the largest few TNOs, which have high albedos and high densities, and the more moderate and smaller TNOs, which have moderate and lower albedos and densities (Brown 2013). There also seems to be some correlation between the color of a TNO and its albedo and dynamical classification (Lacerda et al. 2014). Most of the largest TNOs have known satellites, which one can use to find the bulk densities of the objects. The densities of the largest TNOs seem to be much higher ($\sim 2-3$ $g/cm^{3}$) than for the smaller TNOs ($< 1$ $g/cm^{3}$) (Brown 2013; Grundy et al. 2015; Barr \& Schwamb 2016). This suggests the larger TNOs are made of more rock and less ice and have less porosity than the smaller TNOs. The intermediately sized large TNOs, like 2013 FY27, are thus key to understanding where, how and why this transition in albedo and density occurs. Through thermal and optical observations, we place 2013 FY27 into the context of the largest TNOs. \section{Magellan Optical Observations} We observed 2013 FY27 using the Magellan 6.5 meter telescope at Las Campanas, in Chile on UT March 8, 9 and 10 and May 3, 2016. The IMACS camera was used, which has a pixel scale of 0.20 arcsec per pixel and field of view of about 0.16 square deg. The geometry of the observations are shown in Table 1. All images were bias subtracted and flat fielded with nightly dithered sky twilight flats. Images were obtained guiding at sidereal rates in photometric conditions using the Sloan g, r or i-band filters. Most of the Magellan images were obtained in the r-band to look for short-term variability of 2013 FY27 over minutes, hours and days. To determine the optical color of 2013 FY27, we observed 2013 FY27 multiple times on UT March 10, 2016 in the g-band and i-band Sloan filters as well as the r-band filter. The filters were rotated between the r, g and i-bands to prevent any possible short-term variations from effecting the color measurements and were obtained twice, separated by a few hours. We also convert the Sloan g,r,i colors to the Johnson-Morgan-Cousins BVRI filter system for easier comparison to some past works using the transformation equations from Smith et al. (2002): $B=g'+0.47(g'-r') +0.17$; $V=g'-0.55(g'-r')-0.03$; $V-R=0.59(g'-r')+0.11$; $R-I=1.00(r'-i')+0.21$. Sheppard (2010) showed these transformations from g,r,i colors to BVRI colors are good to within a hundredth of a magnitude for most TNOs. All the optical observations were calibrated using the Sloan standard star fields DLS1359-11 and PG1633+099. Seeing was between 0.6 and 1.5 arcseconds, with longer exposure times used when the seeing was worse. The photometry as well as the details of each individual image from the observations are shown in Table 2. All photometry was obtained similar to that described in Sheppard (2007) and used apertures to fit the variable seeing, ranging from 2.4 in the best seeing to 4 arcseconds in the worse seeing. The background counts for each image were subtracted off the photometry of the object using an aperture annulus that was larger than that used for the object's photometry. \subsection{r-band Optical Light Curve} We observed 2013 FY27 in the r-band at Magellan over minutes, hours and days in early March 2016 to look for any short-term variations caused by rotation of the object. The observational results are shown in Table 2. No clear variability was seen, with variability being less than $0.06\pm 0.02$ over the 3 days of observations from March 8 to March 10, 2016 (Figure~\ref{fig:multify27}). This suggests 2013 FY27 either has a very low amplitude rotational variability, or that its rotation is longer than a few days time. As 2013 FY27 is a large TNO, it should be mostly spherical in shape due to the strength of its own gravity. Any short-term variations would likely be from albedo differences on its surface, which we see no strong evidence for. There is also the less likely possibility that we are observing 2013 FY27 pole-on, and thus would see no rotational short-term variability. \subsection{Sloan Optical Colors} Using the Sloan $g$, $r$ and $i$-band observations from Magellan described above, we find the colors of 2013 FY27 are $g-r=0.76\pm0.02$ and $r-i=0.31\pm0.03$ giving $g-i=1.07\pm0.04$ mags (Table 3). Since the majority of TNOs have nearly linear color slopes at visible wavelengths, the basic color of a TNO can be reported as its spectral gradient (see Doressoundiram et al. 2008 and Sheppard 2010). The spectral gradient is the amount of reddening per 100 nanometers and can be found through \begin{equation} S(\lambda_{2} > \lambda_{1}) = (F_{2,V} - F_{1,V}) /(\lambda_{2} - \lambda_{1}) \end{equation} where $\lambda_{1}$ and $\lambda_{2}$ are the midpoint wavelengths for the filters used and $F_{1,V}$ and $F_{2,V}$ are the flux in the filters normalized to the V-band. We use the $g$ and $i$ filters to determine $S$ for 2013 FY27 as these filters have well separated central wavelengths. For 2013 FY27, we find the optical color spectral slope is $S=17\pm 2$. This color is moderately red when compared to other TNO colors taken from Hainaut et al. (2012) and the updated Minor Bodies in the Outer Solar System (MBOSS) data base (Figure~\ref{fig:KBOcolorsDwarf}). It has been known since the first Kuiper Belt objects were discovered that they exhibit a very wide range of colors with possible color groupings (Luu \& Jewitt 1996; Barucci et al. 2005). Ultra-red material is generally considered to have a large red spectral gradient of $S > 25$ (Jewitt 2002; Sheppard 2010,2012). In Sheppard (2012) there is a noticeable gap in objects from the Cold Classical belt with colors just below $S \sim 20$, so we define $20 < S < 25$ as very red objects. Objects with spectral gradients between about 8 and 20 are considered moderately red while objects with color below this are considered neutral to blue in color. It is apparent in Figure~\ref{fig:KBOcolorsDwarf} that the ten largest TNOs only show extremes in colors, being either very neutral or ultra-red. It is not until objects smaller than about 800 km in diameter start to show very red and moderately red colors. This is interesting as the extreme colors are usually associated with organics and freshly exposed ices such as Methane, water ice and methanol (Brown et al. 2012; Dalle Ore et al. 2015; Fraser et al. 2015). This suggests the largest TNOs are continuing or have more recently modified their surfaces compared to the TNOs smaller than 800 km. Moderately red colors might be more expected on objects that have old surfaces as micrometeorite bombardment and irradiation would be expected to dull any exposed ices over time (Grundy 2009). The surfaces of the largest TNOs might be fresher because they can more easily hold on to volatile ices and could further have differentiated more completely and even have cryovolcanism occurring in recent times. 2013 FY27 is one of the largest known moderately red objects. \subsection{Optical Phase Curve} The optical apparent magnitude of a TNO depends on its radius ($r$), distance from the Sun ($R$), distance from Earth ($\Delta$), albedo ($p$) and phase angle ($\alpha$). The apparent optical magnitude can be calculated as, \begin{equation} m_{filter}=m_{\odot}-2.5\mbox{log}\left[p*r^{2}\phi (\alpha )/(2.25\times 10^{16}R^{2}\Delta^{2})\right] \label{eq:appmag} \end{equation} where $m_{\odot}$ is the apparent magnitude of the Sun in the filter being used and the linear phase function $\phi (\alpha)$ can be represented as \begin{equation} \phi(\alpha) = 10^{-0.4\beta \alpha} \label{eq:phangle} \end{equation} where $\alpha$ is the phase angle in degrees and $\beta$ is the linear phase coefficient in magnitudes per degree. At opposition, $\alpha=0$ deg, and thus $\phi (0)=1$. We further observed 2013 FY27 at Magellan on May 3, 2016 over several hours to determine its phase coefficient (Figure~\ref{fig:multify27may}). Since no significant short-term light curve was found in March, any long-term variations in brightness are likely attributed to the different phase angles observed for 2013 FY27. The March 2016 observations were obtained at a phase angle near 0.18 degrees while the May observations had the object further from opposition, near 0.62 degrees (Table 1). Again, 2013 FY27 showed no significant variations over the nearly three hours of observations on May 3, 2016. The average r-band magnitude in March was $21.92\pm 0.02$ mags and in May $22.04\pm 0.02$ mags. 2013 FY27 should have been about 0.01 mags fainter in May because it was slightly further from the Earth at that time. Thus 2013 FY27 was about 0.11 mag fainter in May when accounting for differences in its distance from the Sun and Earth. We attribute the 0.11 mag difference from 2013 FY27 being 0.44 degrees further away from true opposition in May than the March observations. From the above March and May r-band observations, we find a phase coefficient of $\beta = 0.25\pm 0.03$ mags per deg for 2013 FY27. We show in Figure~\ref{fig:KBOphaseDwarf} the phase coefficients for the largest TNOs using data from Buie et al. (1997); Sheppard and Jewitt (2002),(2003); Rabinowitz et al. (2007); Sheppard (2007); Benecchi \& Sheppard (2013). We confirm the finding in Sheppard (2007) that the largest few TNOs, which are all neutral in color with high albedos, show the lowest phase coefficients (Figure~\ref{fig:KBOphase2Dwarf}). The lower phase coefficient is likely from back scattering off a high albedo surface. 2013 FY27 has a higher than typical TNO phase coefficient, which are mostly in the 0.10 to 0.20 mag per degree range. This higher phase coefficient could signify different grain properties on 2013 FY27's surface compared to the typical TNO or might be a sign of a very long and significant rotational light curve that would not be seen in a 3 day or less period of time. For main belt asteroids, it appears the higher the phase coefficient, the lower the surface albedo (Belskaya and Shevchenko 2000). But main belt asteroids likely have different compositions and usually have lower phase coefficients than the TNOs. In addition, most asteroid phase curves are based over a much larger range of phase angles, in which the TNOs are not able to be viewed as they typically stay below about 2 degrees as seen from Earth. The opposition surge, when an object gets much brighter at very low phase angles from back-scattering effects, starts around 0.1 to 0.2 degrees (Belskaya et al. 2008). Thus the 2013 FY27 observations are just outside of the start of the opposition surge and likely not strongly affected by the opposition surge. \subsection{Optical Absolute Magnitude} Using the linear phase coefficient found above, we can calculate the reduced magnitude as \begin{equation} m_{R}(1,1,0) = m_{R} - 5\mbox{log}(R\Delta). \end{equation} The reduced magnitude of a solar system object is the magnitude it would have if it were observed simultaneously at a geocentric and heliocentric distance of 1 au with a phase angle of 0 deg. Thus the reduced magnitude gives you the brightness of an object independent of observing geometry. The reduced magnitude brightness is for the most part only dependent on the size and albedo of an object. Using the above equation, we find the reduced magnitude of 2013 FY27 at $\alpha = 0.18$ deg as $2.90\pm 0.02$ mags in the r-band during the March 2016 observations. For the May 2016 observations, we find the reduced magnitude is $3.01\pm 0.02$ mags at $\alpha = 0.62$ deg because the object is further from opposition and thus fainter from showing a less illuminated face towards Earth. To calculate the reduced magnitude of 2013 FY27 when at zero degrees phase angle, we use the linear phase coefficient found above of $\beta = 0.25$ mags per deg. So 2013 FY27 should be $0.25$ mags$/$deg $\times$ $0.18$ deg $= 0.05$ mags brighter when at 0 deg phase angle compared to 0.18 deg phase angle assuming a linear phase function. Thus we find $m_{r}(1,1,0)=2.85\pm 0.02$ mags. From the transformation equations in Smith et al. (2002), we find the Johnson-Kron-Cousins R-band reduced magnitude is $m_{R}(1,1,0)=2.59\pm 0.02$ mags. Using the color found for 2013 FY27 of $V-R=0.56\pm 0.03$ mags, we find $m_{V}(1,1,0)=3.15\pm 0.03$ mags, which we take as the absolute magnitude $H$ of 2013 FY27. This is slightly fainter than the current value used at the minor planet center for 2013 FY27, which is $H=3.0$ mags. This slight difference in the absolute magnitudes is to be expected as Sheppard (2007) found the Minor Planet Center is routinely off by several tenths of magnitudes from well measured calculated values of the absolute magnitude, likely because the Minor Planet Center uses all photometry from multiple sources and filters, some of which have low signal to noise and thus large uncertainties. Though the absolute magnitude $H$ generally uses a curved phase function as defined in Bowell et al. (1989), Sheppard (2007) found the reduced magnitude for TNOs using a linear phase function with $\beta$ is similar to within a few hundredths of the absolute magnitude phase function used in Bowell et al. (1989). An important optical reference is to determine the optical brightness of 2013 FY27 at the time of the ALMA observations in late December 2017 and early January 2018. The phase angle and distance of 2013 FY27 at the time of the ALMA observations is very similar to the May 3, 2016 observations from Magellan. Thus we should be able to use the May 3, 2016 photometry as the base for the optical photometry during the ALMA observations. At this time $m_{r}=22.04\pm0.02$, $m_{g}=22.80\pm0.03$ and $m_{i}=21.73\pm0.03$ mags, giving from color transformations using Smith et al. (2002) $m_{R}=21.79\pm0.02$, $m_{V}=22.35\pm0.03$ and $m_{I}=21.27\pm0.03$ mags. \section{ALMA Observations} The Atacama Large Millimeter Array (ALMA) observations were taken on UT 29 and 30 December 2017 and 04 January 2018 in ALMA Cycle 5 under Project 2017.1.01662.S (Table 1). At that time, the array was configured with 46 antennas and a maximal distance between antennas (maximum baseline) of 2500~m. We used Band 3 in the standard continuum setup centered at 97.5 GHz, with a total encompassed bandwidth of 8 GHz. For such a cold object, Band 3 provides the best sensitivity in the continuum among all the ALMA bands making it the best to use as a first thermal detection attempt. The ALMA observations consisted of three sets of 66 minutes observations, one each on 29 and 30 December 2017 and another on 04 January 2018. The combined observations totaled of 3.3 hours of ALMA time, with 2.2 hours of that integrating on the source. The rest of the observation time was spent on quasars used as bandpass, flux and gain calibrators. Interferometric measurements consist in visibilities, which are complex numbers corresponding to signal cross-correlations between each pair of antennas. For each of the observations, visibilities were calibrated using the ALMA calibration pipeline in the Common Astronomy Software Applications package (CASA, McMullin et al. 2007), to correct for the spectral response and temporal gain variations of the instrument. The absolute flux scale was assessed using the well monitored quasar J1127-1857 as a reference. The three sets of observations were then combined, spectrally averaged and stacked into one visibility set. To obtain a continuum image, inverse Fourier transform and deconvolution were applied to the combined visibilities set. Since this experiment is a point-source detection which does not depend on spatial resolution, natural weighting was used to maximize sensitivity (at the expense of beam size), giving the final image an rms of 5.5 microJy. The synthesized beam was $0.50 \times 0.42$ arcseconds in the natural-weighted image. 2013 FY27 was found very near the center of the image as expected, and was 25 microJy bright (Figure~\ref{fig:ALMAimage}). This gives the ALMA detection of 2013 FY27 a Signal-to-Noise of about 4.5. To assess the quality of the flux calibration, we compared the results given by the calibration pipeline using J1127-1857 as a reference to the expected flux value for the phase calibrator J1058+0133, which is a relatively bright and well monitored quasar. The good match between the retrieved and expected flux for J1058+0133 allow us to determine that the absolute flux calibration is accurate at the 3 to 4 percent level. \section{Size and Albedo of 2013 FY27} To first order, the thermal flux from an object is proportional to $D^2(1-p_Vq)$, where $D$ is the effective diameter, $p_V$ is the geometric albedo in V-band, and $q$ is the phase integral. The reflected sunlight in the V-band from the same object is proportional to $D^2p_V$. As shown in Brown \& Butler (2017), $q$ for TNOs can be approximated by $0.336p_V + 0.479$. Thus we have two equations that can be solved for the two unknowns of diameter and albedo ($D$ and $p_V$) (see e.g. Fernandez et al. 2013). The most common method to determine the size and albedo of an object with single-epoch thermal photometry, as here with 2013 FY27, is to use the Near-Earth Asteroid Thermal Model (NEATM) of Harris (1998). Some basic assumptions used in NEATM are that (i) the object is spherical; (ii) the phase darkening of the thermal emission is entirely dependent on how much of the lit-up hemisphere of the object is facing Earth, and (iii) the basic dayside surface temperature falls off from the subsolar point as $\cos^{1/4}\theta$, where $\theta$ is the local zenith angle, and there is no thermal emission from the nightside. These NEATM assumptions are reasonable for 2013 FY27. 2013 FY27 is probably close to spherical as it is likely large and there is little if any apparent shape-induced rotational modulation of our optical photometry (section 2.1). Our observations were obtained at very small phase angles and thus phase darkening uncertainties should be minimal. Regarding the temperature map, the NEATM is appropriate for a ``slow rotator," i.e. an object whose thermal inertia is sufficiently low, and rotation period sufficiently long, that the object has no thermal memory. In NEATM, deviation from a zero-thermal memory situation is encapsulated in the beaming parameter, $\eta$, which will account for the cooler temperatures expected for an object with some thermal memory. A value of $\eta=1$ corresponds to the zero thermal memory situation, and a value of $\eta>1$ applies to cases with thermal memory, with the higher the value, the more memory (such as the case of a quick rotator or a large thermal inertia). For our analysis of 2013 FY27, since we have measurements at only one thermal wavelength, we cannot independently determine what the beaming parameter should be. However we can use other TNO measurements to examine what would be appropriate to assume. Lellouch et al. (2013, 2017), in a study of the thermal emission from several TNOs as observed with ALMA, Herschel, and Spitzer, found that the thermal inertias of these bodies are quite low, at least an order of magnitude below that of the Moon. It is likely safe to assume that 2013 FY27 is similar, and so should be close to acting like a slow rotator in the NEATM model for reasonable values of its rotation period. Lellouch et al. (2017) advise using a value of $\eta=1.175$ for TNOs observed with ALMA, so we adopt that value here. The other assumptions that go into NEATM are the emissivity and the visible-wavelength phase law. For the former, we assume a value of $0.70$, as suggested by Lellouch et al. (2017). For the latter, we apply our results from section 2.3, where $\beta = 0.25$ mag/deg. The results of our application of the NEATM to the 2013 FY27 photometry is shown in Figure~\ref{fig:YanAlbedoSize}. We used a flux density from thermal emission of $25.0\pm5.5$ $\mu$Jy at a wavelength of 3.075 mm, and a V-band magnitude of $22.35\pm0.03$ to fit $D$ and $p_V$. The thermal emission error bar incorporates the $\sim$4\% uncertainty in the absolute calibration. Since there are zero degrees of freedom (using two measurements to fit two parameters), we cannot formerly calculate a reduced-$\chi^2$; instead we use $\chi^2$ itself. The contour plot in Figure~\ref{fig:YanAlbedoSize} shows contours of $\chi^2=1$, 2, and 3, and we report an error bar on the best-fitting parameters to correspond to the $\chi^2=1$ level. We find that the best fit for 2013 FY27 is an effective diameter of $D = 765_{-85}^{+80}$ km and $p_V = 0.17_{-0.030}^{+0.045}$. The two error bars are correlated, as the elongated contours show. The errors on $D$ and $p_V$ we report here are from the uncertainty in the photometry including the absolute calibration of the standard sources in both the visible and thermal. Incorporating the potential uncertainty in the assumed parameters such as the beaming parameter and effective slow rotation period is tricky, but could add some five percent uncertainty to the diameter calculation and 10 percent uncertainty to the albedo calculation. As noted in section 5 below, a satellite some 3 magnitudes fainter than 2013 FY27 was found in HST data. Assuming the satellite has a similar albedo as the primary, the satellite is about 190 km in diameter, making the primary slightly smaller than calculated above at $D = 740_{-90}^{+85}$ km. \subsection{Discussion of Dwarf Planet Diameters, Albedos and Colors} An albedo of about 17\% is on the high end for a moderately red object but consistent with the other moderately red and moderately sized dwarf planet TNOs (Figure~\ref{fig:KBOalbedo}). 2013 FY27 is one of the largest, if not the largest, moderately red TNO. All TNOs larger than 2013 FY27, of which there are about ten, have either ultra-red colors or neutral colors. The TNOs with similar sizes and smaller than 2013 FY27 (diameters $< 800$ km) start to show an abundance of moderately red or very red colors, unlike the largest ten objects, which do not show any of these middle surface colors and only the extreme surface colors. This suggests something is or has physically changed the surfaces of the largest ten TNOs with diameters above 800 km compared to the TNOs with diameters smaller than 800 km. This could be because TNOs above about 800 km have enough self gravity to retain certain ices such as methane more readily on their surfaces than the smaller TNOs. Neutral colored surfaces are generally associated with fresh water ice while ultra-red surfaces are associated with organics and possibly other ices such as methanol and methane. The color differences may be because TNOs above 800 km retained enough internal heat to remain active longer or even to this day, to resurface their surfaces from possible cryovolcanism type events. TNOs above 800 km might also be more fully differentiated than TNOs below 800 km, making their two surfaces types different in composition. As TNOs below 800 km still show neutral and ultra-red surfaces as well as very red and moderately red surfaces, the moderately red surfaces might be a signature of a very old surface that has been bombarded by high energy photons, particles and micrometeorites over long periods of time, while the extreme colors are more of a sign of fresher surfaces from recent collisions or activity. Further analysis is required to determine why the largest TNOs do not show moderately red surfaces, but Figure~\ref{fig:KBOalbedo} strongly suggests there is some kind of surface change around 800 km in diameter for TNOs. \section{Satellite Discovered around 2013 FY27} We observed 2013 FY27 with the Hubble Space Telescope (HST) on UT January 15, 2018 to look for possible satellites (HST Program GO-15248). Four 545 second images were taken with HST between 01:24 and 02:30 hours UT in the F350LP wide band filter (central wavelength of 5859 Angstroms) using the WFC3/UVIS instrument with a pixel scale of $0.04$ arcsec per pixel (MacKenty et al. 2014). An obvious point source was detected about $0.17$ arcseconds at a position angle of 135 degs from the primary (Figure~\ref{fig:2013FY27moon}). The satellite was seen in all four images and moved along with the motion of the primary, which was about -1.47 and 0.34 arcseconds an hour in Right Ascension and Declination, respectively. This detection was reported as a satellite of 2013 FY27 to the International Astronomical Union (see CBET 4537: Sheppard 2018). No satellite motion relative to the primary was detected between the first and last (fourth) image from HST, confirming its association with the primary and likely ruling out an extremely fast orbital period of a few days or less around the primary. The four HST images were aligned with respect to the primary and coadded to search for additional fainter satellites, but nothing obvious was detected to about 27.5 mags within a tenth to tens of arcseconds of the primary. At the time of the HST observations, 2013 FY27 was 79.48 au from the Earth. Thus a 0.17 arcsecond separation means the satellite was at least 9800 km from the primary. The newly discovered satellite was $3.0\pm0.2$ mags fainter than the primary in the optical. Thus the diameter of the satellite, if assuming the same albedo as the primary, would be about 3.9 times smaller than the primary or about 190 km in diameter. Additional HST observations of 2013 FY27 and its satellite have been obtained in May and July of 2018 under HST Program 15460. A full analysis of the new HST data will allow the determination of the satellites orbit and with the albedo reported in this work, a bulk density of the system can be determined giving insight into the composition and structure. The full detailed analysis using all of the HST observations for 2013 FY27 will be part of a future paper on the satellite and its orbit around 2013 FY27. We note that in Figure~\ref{fig:ALMAimage}, there appears to be an extension of the ALMA signal to the southwest of the primary. This is unlikely to be the satellite of 2013 FY27 as we believe the satellite is nearly edge-on, and thus should only show up to the southeast and northwest of the primary. Obtaining the full orbit of the satellite will allow us to predict where it should have been during the ALMA observations and further analysis of the ALMA data and 2013 FY27's satellite is left for the next paper on the full orbit of the 2013 FY27 system. \subsection{Discussion of Dwarf Planet Satellites} All of the largest known TNOs, those with diameters well over 1000 km, irrespective of their dynamical class, have known satellites (Pluto, Eris, Makemake, 2007 OR10, Haumea, and Quaoar) and now 11 of the top 15 largest TNOs have known satellites (Figure~\ref{fig:KBOsatDwarf}). All of the satellites of these largest objects are significantly smaller than the primary and have relatively close orbits that are indicative of collisional formation (Brown et al. 2006; Noll et al. 2008; Parker et al. 2016; Kiss et al. 2017; Brown \& Butler 2018). This is remarkable that satellites seem to be the norm and not the exception for the largest objects. This appears to suggest the collisional formation of Earth's moon and the satellites of Mars are normal outcomes of the planet formation process (Mercury and Venus likely have no satellites simply because their closeness to the massive Sun makes tidal interactions too strong for most satellite orbits around these planets to be stable over the age of the solar system). Though we don't know the full orbit of the satellite for 2013 FY27, its small size relative to the primary and relatively close distance suggest the satellite was likely created by a direct impact, similar to the scenarios envisioned for the other known satellites of the dwarf planets (Canup 2011; McKinnon et al. 2017). This formation scenario is quite different than the equal-sized, distant binaries found mostly in the main Kuiper Belt, which likely did not form through direct collisions onto the primary (Schlichting and Re'em 2008; Nesvorny et al. 2010; Parker et al. 2011; Sheppard et al. 2012). Barr and Schwamb (2016) suggest that there are two possible collisional formation scenarios that could create the close-in satellites found around all the largest TNOs. High energy direct collisions could remove the ice from the primary, leaving a dense primary and small ice rich satellite. A glancing indirect collision would leave the primary volatile rich and thus of low density and could create larger satellites of similar composition to the primary. 2013 FY27 can be a further test to this theory through determining its density from knowing the full orbit of the satellite around the primary as well as looking at the surfaces of the primary and secondary for similarities or differences. We note in Figure~\ref{fig:KBOsatDwarf} there appears to be a break in the ratio of the satellite size to the primary size for TNOs larger than about 900 km. TNOs larger than about 900 km have relatively small satellites compared to the primary (though Pluto somewhat breaks this trend). Many of the smaller TNOs appear to have satellites that are relatively large compared to the primary, with some approaching equal-sized binaries starting around 400 km in diameter. \section{Summary} 2013 FY27 is intrinsically the brightest TNO not yet observed for its basic physical properties. 2013 FY27's absolute magnitude of near 3 mags means it likely has a size that is near the transition region between the largest few TNOs that show high albedos and high densities and the smaller TNOs that show low to moderate albedos and densities. We observed 2013 FY27 over optical and thermal wavelengths at several telescopes to determine its physical characteristics for the first time and compare it to the other dwarf planets. 1) The geometric albedo of 2013 FY27 was found to be $p_V = 0.17_{-0.030}^{+0.045}$. The effective diameter of the 2013 FY27 system is $D =765_{-85}^{+80}$ km. Assuming then newly discovered satellite around 2013 FY27 has a similar albedo as the primary, the satellite is about 190 km in diameter, making the primary slightly smaller than the effective diameter of the system at $D = 740_{-90}^{+85}$ km. 2) The color of 2013 FY27 was found to be moderately red with a spectral gradient of $S=17\pm 2$. This makes 2013 FY27 one of the largest known moderately red TNOs. All TNOs larger than about 800 km in size, of which there are about ten known, only show neutral or ultra-red surface colors. This could be because the largest several TNOs have different or fresher surfaces than TNOs smaller than 800 km from possible cryovolcanism, differentiation and/or abundant exposed surface ices. The TNOs only start to show moderately red colors for objects less than about 800 km in diameter, and they appear to be very abundant below this threshold. As there is also neutral and ultra-red TNOs below 800 km, it suggests the moderately red color might be an indication of an old surface while the more neutral and ultra-red colors in the smaller TNOs could be fresh surfaces exposed from recent impacts or other processes. 3) A satellite of 2013 FY27 was found in Hubble Space Telescope observations of 2013 FY27. It is some $3.0\pm0.2$ mags fainter and was $0.17$ arcsec away from the primary at discovery. The satellite diameter to primary diameter ratio is a little larger than most of the ratios found for the largest TNOs. There appears to be a difference in the satellite to primary size ratios starting around the TNOs over about 900 km. Less than 900 km the satellite to size ratio increases until near 400 km in primary size one starts to approach equal-sized binaries. For TNOs larger than 900 km, the satellite to primary size ratio is generally smaller, with Pluto being an exception. 4) 2013 FY27 was monitored over minutes, hours and days with no obvious short-term variability detected in the r-band. The most likely reason for no measureable short-term light curve is that 2013 FY27 is near spherical in shape with no significant albedo variations on its surface to allow for a measurable rotational period. 2013 FY27 was further monitored in the r-band over months to find its linear phase curve of $\beta = 0.25\pm 0.03$ mags/deg. This is a slightly higher than normal TNO phase curves, but reasonable for a moderate albedo surface, unlike the lower phase curves found for very high albedo TNO surfaces. The reduced magnitude of 2013 FY27 was found to be $m_{r}(1,1,0)=2.85\pm 0.02$ and $m_{V}(1,1,0)=3.15\pm 0.03$ mags. \section*{Acknowledgments} We thank D. Ragozzine for comments on the manuscript and sharing the initial analysis of the 2013 FY27 satellite orbit. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2017.1.01662.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This work is based in part on NASA/ESA Hubble Space Telescope Cycle 25 Program 15248 observations. This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile.

f072c205967dfeae8fb4f3f68667878932f203ef

\section{Introduction} The main purpose of our paper is to develop new methods to obtain precise estimates of Lyapunov exponents and the variance for the CLT related to the products of random matrices. Let $\{Y_i\}_{i\ge1}$ be a sequence of i.i.d. random matrices distributed according to a probability measure $\mu$. Further, let $S_n=Y_n Y_{n-1}\cdots Y_2 Y_1$. Assuming that $\mathbb E\left[\log^+\|Y_1\|\right]<\infty$, the \emph{top Lyapunov exponent} $\lambda$ associated with $\mu$ is given by \begin{equation}\label{lyapdefn} \lambda:=\lim_{n\to\infty}\frac1n\mathbb E\big[\log\|S_n\|\big] \end{equation} with $\lambda\in\mathbb R\cup\{-\infty\}$. The top Lyapunov exponent gives the rate of exponential growth of the matrix norm of $S_n$ as $n\to\infty$. Since all finite-dimensional norms are equivalent, $\lambda$ is independent of the choice of norm $\|\cdot\|$. Although $\lambda$ depends on $\mu$, we usually omit this dependence from our notation. While one can also define a spectrum of Lyapunov exponents, in this paper we will only be concerned with the top Lyapunov exponent $\lambda$ and we refer to it as simply the Lyapunov exponent. Occasionally, when we are considering $\lambda$ over a family of distributions parametrized by some variable, we will write $\lambda$ as a function of that variable. Furstenberg and Kesten (1960) and Le Page (1982) found analogues of the Law of Large Numbers and Central Limit Theorem, respectively, for the norm of these partial products. Despite these results having been established for some time, in most cases it is still impossible to compute the Lyapunov exponent explicitly from the distribution of the matrices. Moreover, computing the variance in the CLT has received scant attention in the literature. We point out that because of the difficulty in computing Lyapunov exponents, most authors need to develop new techniques for specific matrix models rather than work in a general framework. In this paper, we investigate the behavior of the Lyapunov exponent as the common distribution of the sequence of random matrices varies with a parameter. While there are works in the literature where explicit expressions have been obtained for some matrix models under certain conditions \cite{bougerol,chassaingetal,newman,limarahibe,mtwInvariant,Newman1986}, besides a few special examples, it is not possible to find a general explicit formula for the Lyapunov exponent. There is, however, an extensive literature on approximating the Lyapunov exponent for models where it cannot be calculated explicitly (see \cite{Protasov-Jungers-2013,pollicott}). For instance, in \cite{pollicott}, $\lambda$ is expressed in terms of associated complex functions and a more general algorithm to numerically approximate $\lambda$ is given. The method is efficient and converges very fast. The method also applies to a large class of matrix models. There is also a significant interest in computing Lyapunov exponents in physics, with some recent work found in \cite{Akemann-Burda-Kieburg2014,Akemann-Kieburg-Wei-2013,Forrester-2013,Forrester-2015,Kargin-2014,Kieburg-Kosters-2019}. The analytic properties of the Lyapunov exponent as a function of the transition probabilities are studied in \cite{Peres-1991,Peres-1992,Ruelle-1979}. Lyapunov exponents are also useful in mathematical biology in the study of population dynamics. A random Fibonacci sequence $g_0,g_1,g_2\dots$ is defined by $g_0=g_1=1$ along with the recursive relation ${g_{n+1} = g_n \pm g_{n-1}}$ (linear case) or ${g_{n+1} = |g_n \pm g_{n-1}|}$ (non-linear case) for all $n\in\mathbb{N}$, where the sign $\pm$ is chosen by tossing a fair or biased coin (positive sign has probability $p$). In \cite{Viswanath2000}, Viswanath studied the exponential growth of $|g_n|$ as $n\to\infty$ in the linear case with $p=\frac{1}{2}$ by connecting it to a product of random matrices and then employing a new computational method to calculate the Lyapunov exponent to any degree of accuracy. The method involves using Stern-Brocot sequences, Furstenberg's Theorem (see Theorem \ref{furstenberg}) and the invariant measure to compute $\lambda$. We also point to the work of \cite{Janvresse-Rittaud-Rue2010,Janvresse-Rittaud-Rue2009,Janvresse-Rittaud-Rue2008} where the authors generalized the results of Viswanath by letting $0<p\leq 1$ and treating $\lambda$ as a function of $p$ which bears some similarity to the model we study in Section \ref{bernp}. They also considered the non-linear case. The model that is most relevant to our results is given in \cite{Goswami2004}, where the authors give an explicit formula for the cumulative distribution function of a random variable $X_p$ on $(0,\infty)$ characterized by the distributional identity $$ X_p \sim \frac{1}{X_p}+ \epsilon_p,$$ where $\epsilon_p$ is a $\text{Bernoulli}\left(p\right)$ random variable independent of $X_p$. Let CDF denote the cumulative distribution function for a random variable. The CDF of $X_p$ is given in terms of a continued fraction expansion. We will later see that the distribution of $X_p$ is the invariant distribution for the product of random matrices studied in Section \ref{bernp}. We summarize the main results of the paper as follows. Consider the random matrices $$ Y_i = \left(\begin{array}{cc} \epsilon_i & 1\\ 1 & 0 \end{array}\right), $$ where $\epsilon_i$ are i.i.d. random variables. \begin{enumerate} \item {\bf Lyapunov exponent when $\epsilon\sim \text{Bernoulli}\left(p\right)$} (See Theorem \ref{UBprop}): The Lyapunov exponent $\lambda(p)$ can be estimated by \[ \frac{p\log 3}{4-p}\leq\lambda(p)\leq\frac{p\log 3}{2}. \] \item {\bf Lyapunov exponent when $\epsilon\sim \text{Bernoulli}\left(\frac{1}{2}\right)$}(See Theorem \ref{thm:bounds}): The Lyapunov exponent $\lambda$ can be estimated by $$ p_n\leq\lambda\leq q_n, $$ where $$ p_n =\frac{\log c_{n}}{\left(n+7\right)2^{n}} \,\,\, \mbox{and}\,\,\, q_n = \frac{\log c_{n}}{\left(n+4\right)2^{n}}, $$ and $c_n$ is given by Definitions \ref{def:recursion} and \ref{def:c_n}. Moreover, \[ \lim_{n\to \infty}p_{n} = \lim_{n\to\infty} q_{n} = \lambda. \] The method we develop differs from that of the papers listed above and requires the study of an interesting multi-level recursion satisfied by $c_n$. \item {\bf Exact Lyapunov exponent involving Cauchy random variable} (See Proposition \ref{CauchyProp}): When $$Y_i= \left(\begin{array}{cc} \xi \epsilon & -1\\ 1 & 0 \end{array}\right), \ \epsilon\sim \text{Cauchy} \left(0,1\right), \ \xi\in\mathbb{\mathbb{R}}, \ \xi\neq 0, $$ then the Lyapunov exponent $\lambda(\xi)$ is given by \[ \lambda(\xi)=\log\left(\frac{|\xi|+\sqrt{\xi^{2}+4}}{2}\right). \] \item {\bf Variance Simulation} (See Figures \ref{fig:BernPVars}, \ref{fig:Cauchy20Vars} and \ref{fig:CauchyZoomVars}) \end{enumerate} The paper is organized as follows. In Section \ref{preliminaries} we give the preliminaries needed for the paper. In Section \ref{bernp}, we provide exact upper and lower bounds on the Lyapunov exponent associated with the product of random matrices where one entry is $\text{Bernoulli}\left(p\right)$ with $0<p<1$. In particular, in Section \ref{bernhalf} we study the $p=\frac{1}{2}$ case and provide a sequence of progressively better bounds. We prove that these bounds converge to the Lyapunov exponent which gives a new characterization for the Lyapunov exponent. Not surprisingly, these bounds are related to Fibonacci sequences as in the work of \cite{Goswami2004,Janvresse-Rittaud-Rue2010,Janvresse-Rittaud-Rue2009,Janvresse-Rittaud-Rue2008,Viswanath2000}. In Section \ref{xicauchy}, we give an example of a well-known model where we can calculate the Lyapunov exponent explicitly. In this model, one entry in the random matrix has the Cauchy distribution. In Section \ref{variance}, we examine the less studied variance associated with a multiplicative Central Limit Theorem for products of random matrices. The multiplicative CLT holds under some reasonable assumptions, see \cite{bougerol}. It states that for $\mathbf{x}\in\mathbb{R}^d\setminus\{\mathbf{0}\}$, \[ \frac{1}{\sqrt{n}}\left(\log\left\Vert S_{n}\mathbf{x}\right\Vert -n\lambda\right)~\mbox{ and }~\frac{1}{\sqrt{n}}\left(\log\left\Vert S_{n}\right\Vert -n\lambda\right) \] converge weakly to a Gaussian random variable with mean $0$ and variance $\sigma^2>0$ as $n\to\infty$. In the special case where the distribution of $\|Y_1\mathbf{x}\|/\|\mathbf{x}\|$ doesn't depend on $\mathbf{x}\in\mathbb{R}^d\setminus\{\mathbf{0}\}$, Cohen and Newman \cite{newman} gave the explicit formulas \begin{equation}\label{eq:newman} \lambda=\mathbb{E}\left[\log\left(\frac{\|Y_1\mathbf{x}\|}{\|\mathbf{x}\|}\right)\right]~\text{ and }~\sigma^2=\mathbb{E}\left[\left(\log\left(\frac{\|Y_1\mathbf{x}\|}{\|\mathbf{x}\|}\right)-\lambda\right)^2\right] \end{equation} that hold whenever the expectations are finite. As far as the authors know, this is the only case where an explicit formula for the variance is given. Compared to the calculation of the Lyapunov exponent, there have been relatively few attempts to explicitly compute or numerically approximate the variance. We address this deficiency in the context of the parameter models that we consider by first describing an easy to implement Monte Carlo simulation scheme and then using it to approximate the variance for some of the models we considered earlier in the paper. \section{Preliminaries}\label{preliminaries} In what follows, we introduce notational conventions and terminology and recall well-known results regarding the Lyapunov exponent. Let $\mathbb{P}^{1}\left(\mathbb{R}\right)$ denote the one-dimensional projective space. Recall that we can regard $\mathbb{P}^{1}\left(\mathbb{R}\right)$ as the space of all one dimensional subspaces of $\mathbb{R}^{2}$. To describe $\mathbb{P}^{1}\left(\mathbb{R}\right)$, let us first define the following equivalence relation $\sim$ on $\mathbb{R}^{2}\backslash\left\{ \mathbf{0}\right\} $. We say that the vectors $\mathbf{x},\mathbf{x'}\in\mathbb{R}^{2}\backslash\left\{ \mathbf{0}\right\} $ are equivalent, denoted by $\mathbf{x}\sim\mathbf{x'}$, if there exists a nonzero real number $c$ such that $\mathbf{x}=c\mathbf{x'}$. We define $\bar{\mathbf{x}}$ to be the equivalence class of a vector $\mathbf{x}\in\mathbb{R}^{2}\backslash\left\{ \mathbf{0}\right\}$. Now we can define $\mathbb{P}^{1}\left(\mathbb{R}\right)$ as the set of all such equivalence classes $\bar{\mathbf{x}}$. We can also define a bijective map $\phi:\mathbb{P}^{1}\left(\mathbb{R}\right)\to\mathbb{R}\cup\left\{ \infty\right\} $ by \[ \phi\left(\bar{\mathbf{x}}\right)=\begin{cases} \frac{x_{1}}{x_{2}} & \mbox{if }x_{2}\neq0\\ \infty & \mbox{if }x_{2}=0 \end{cases} \] where $\mathbf{x}=\left(\begin{array}{c} x_{1}\\ x_{2} \end{array}\right)\in\mathbb{R}^{2}\backslash\left\{ \mathbf{0}\right\} $ is in the equivalence class $\bar{\mathbf{x}}$. Hence with a slight abuse of notation we can identify $\mathbb{P}^1(\mathbb{R})$ with $\mathbb{R}\cup\left\{ \infty\right\}$. Consider the following group action of $\mathrm{GL}(2,\mathbb R)$ on $\mathbb{P}^{1}\left(\mathbb{R}\right)$. For $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}\in\mathrm{GL}(2,\mathbb R)$ and $x\in\mathbb{P}^{1}\left(\mathbb{R}\right)$, we define \[A\cdot x=\frac{ax+b}{cx+d}.\] Let $\mu$ and $\nu$ be probability measures on $\mathrm{GL}(2,\mathbb R)$ and $\mathbb{P}^{1}\left(\mathbb{R}\right)$, respectively. We say that $\nu$ is \emph{$\mu$-invariant} if it satisfies \begin{equation}\label{eq:invariance} \int_{\mathbb{P}^{1}\left(\mathbb{R}\right)}f(x)\,\mathrm d\nu(x)=\int_{\mathbb{P}^{1}\left(\mathbb{R}\right)}\int_{\mathrm{GL}(2,\mathbb R)}f(A\cdot x)\,\mathrm d\mu(A)\,\mathrm d\nu(x) \end{equation} for all bounded measurable functions $f:\mathbb{P}^{1}\left(\mathbb{R}\right)\to \mathbb{R}$. Furthermore, we say that a set $G \subset \mathrm{GL}(2,\mathbb R)$ is \emph{strongly irreducible} if there is no finite family $V_1,\ldots,V_k$ of proper $1$-dimensional vector subspaces of $\mathbb{R}^2$ such that $A(V_1\cup\cdots\cup V_k)=V_1\cup\cdots\cup V_k$ for all $A\in G$. For a real valued function $f$, define $f^+=\max \left\lbrace f,0 \right\rbrace$. The following result by Furstenberg and Kesten in \cite{furstenbergkesten} gives an important analogue to the Law of Large Numbers. \begin{thm}[Furstenberg-Kesten]\label{furstenbergkesten}\ \\ Let $\{Y_i\}_{i\ge1}$ be a sequence of i.i.d. $\mathrm{GL}(d,\mathbb R)$-valued random matrices and $S_n= Y_n Y_{n-1} \cdots Y_2 Y_1$. If \ $\mathbb E\left[\log^+\|Y_1\|\right]<\infty$ and $\lambda$ is the Lyapunov exponent defined in \eqref{lyapdefn}, then almost surely we have \[\lambda=\lim_{n\to\infty}\frac1n\log\|S_n\|.\] \end{thm} For the rest of this paper, we will suppose that $\mu$ is a probability measure on the group $\mathrm{GL}(2,\mathbb R)$ and that the matrices $\{Y_i\}_{i\ge1}$ are distributed according to $\mu$. However, Theorems \ref{Furstenberg-Kifer}, \ref{furstenberg} and \ref{lepageclt} all have statements valid for matrices in $\mathrm{GL}(d,\mathbb R)$ as well. In \cite{Furstenberg-Kifer1983}, Furstenberg and Kifer give an expression for $\lambda$ in terms of $\mu$ and the $\mu$-invariant probability measures $\nu$ on $\mathbb{P}^{1}\left(\mathbb{R}\right)$. The following result is given in \cite[Theorem 2.2]{Furstenberg-Kifer1983}. \begin{thm}[Furstenberg-Kifer]\label{Furstenberg-Kifer}\ \\ Let $\mu$ be a probability measure on the group $\mathrm{GL}(2,\mathbb R)$ and $\{Y_i\}_{i\ge1}$ be a sequence of i.i.d. random matrices distributed according to $\mu$. If $\mathbb E \big[ \log^+\|Y_1\|+\log^+\|Y_1^{-1}\| \big]<\infty$, then the Lyapunov exponent is given by \[\lambda= \sup_{\nu} \int_{\mathbb{P}^{1}\left(\mathbb{R}\right)}\int_{\mathrm{GL}(2,\mathbb R)}\log\frac{\|A\mathbf{x}\|}{\|\mathbf{x}\|}\,\mathrm d\mu(A)\,\mathrm d\nu(\bar{\mathbf{x}}),\] where the supremum is taken over all probability measures $\nu $ on $\mathbb{P}^{1}\left(\mathbb{R}\right)$ that are $\mu$-invariant. \end{thm} If $\nu$ is the unique $\mu$-invariant probability measure on $\mathbb{P}^{1}\left(\mathbb{R}\right)$, then Theorem \ref{Furstenberg-Kifer} implies that the Lyapunov exponent can be written as \[\lambda= \int_{\mathbb{P}^{1}\left(\mathbb{R}\right)}\int_{\mathrm{GL}(2,\mathbb R)}\log\frac{\|A \mathbf x \|}{\|\mathbf x \|}\,\mathrm d\mu(A)\,\mathrm d\nu(\bar{\mathbf x}).\] Sufficient conditions for the existence of such a unique $\nu$ were given by Furstenberg and can be found in \cite[Theorem II.4.1]{bougerol}. \begin{thm}[Furstenberg]\label{furstenberg}\ \\ Let $\mu$ be a probability measure on the group $\mathrm{GL}(2,\mathbb R)$ and $\{Y_i\}_{i\ge1}$ be a sequence of i.i.d. random matrices distributed according to $\mu$. Additionally, let $G_\mu$ be the smallest closed subgroup containing the support of $\mu$. Suppose the following hold: \begin{enumerate}[(i)] \item $\mathbb E \big[\log^+\|Y_1\| \big]<\infty$, \item For $M$ in $G_\mu$, $|\det M|=1$, \item $G_\mu$ is not compact, \item $G_\mu$ is strongly irreducible. \end{enumerate} Then there exists a unique $\mu$-invariant probability measure $\nu$ on $\mathbb{P}^{1}\left(\mathbb{R}\right)$ and $\lambda>0$. Moreover, $\nu$ is atomless. Consequently, \[\lambda=\int_{-\infty}^\infty\int_{\mathrm{GL}(2,\mathbb R)}\log\frac{\|A\mathbf x\|}{\|\mathbf x \|}\,\mathrm d\mu(A)\,\mathrm d\nu(\bar{\mathbf x}).\] \end{thm} Let $A=\begin{pmatrix}a&b\\c&d\end{pmatrix}$ be a $\mathrm{GL}(2,\mathbb R)$-valued random matrix. In this paper, we only study matrices $A$ with entry $a$ random and all other entries constant. Let us suppose that the distribution of $a$ is chosen such that the hypotheses of Theorem \ref{furstenberg} hold. Then by a simple computation \cite[pp. 3421]{mtwInvariant} we have that \[ \lambda= \int_{-\infty}^\infty\log|cx+d|\,\mathrm d\nu(x), \] where $\nu$ is the unique $\mu$-invariant probability measure on $\mathbb{P}^{1}\left(\mathbb{R}\right)$. Hence, if $X$ is a random variable distributed according to $\nu$, then \begin{equation}\label{expectation} \lambda=\mathbb E\big[\log|cX+d|\big]. \end{equation} Moreover, if $A$ and $X$ are independent, we can also conclude that $A\cdot X$ has the same distribution as $X$, which we write as $A\cdot X\sim X$. This follows from the definition of $\mu$-invariance. Thus, a random variable $X$ with law given by the unique $\mu$-invariant distribution on $\mathbb{P}^{1}\left(\mathbb{R}\right)$ must satisfy \begin{equation}\label{Invariant-Formula} X\sim\frac{aX+b}{cX+d}, \end{equation} where $a$ and $X$ are independent. Likewise, the law of any $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable $X$ which satisfies \eqref{Invariant-Formula} is $\mu$-invariant hence it must be $\nu$. We make use of this distributional identity for the $\mu$-invariant distribution in later sections. The following result by Le Page can be found in \cite[Theorem V.5.4]{bougerol} and gives a less-studied analogue to the Central Limit Theorem. \begin{thm}[Le Page]\label{lepageclt}\ \\ Define $\ell(M)=\max\{\log^+\|M\|,\log^+\|M^{-1}\|\}$ for $M\in\mathrm{GL}(2,\mathbb R)$. Let $\mu$ be a probability measure on the group $\mathrm{GL}(2,\mathbb R)$ and $\{Y_i\}_{i\ge1}$ be a sequence of i.i.d. random matrices distributed according to $\mu$. Moreover, let $G_\mu$ be the smallest closed subgroup containing the support of $\mu$. Suppose the following hold: \begin{enumerate}[(i)] \item $\mathbb{E} \left[\exp \left(t~\ell (Y_1) \right) \right]<\infty$ for some $t>0$, \item $G_\mu$ is strongly irreducible, \item $\{|\det M|^{-1/2}M:M\in G_\mu\}$ is not contained in a compact subgroup of $\mathrm{GL}(2,\mathbb R).$ \end{enumerate} Then there exists $\sigma>0$ such that for any $\mathbf{x}\in\mathbb R^2\setminus\{\mathbf{0}\}$, \[\frac1{\sqrt n}\left(\log\|S_n\mathbf{x}\|-n\lambda\right)\textrm{ and }\frac1{\sqrt n}\left(\log\|S_n\|-n\lambda\right)\] converge weakly as $n\to\infty$ to a Gaussian random variable with mean $0$ and variance $\sigma^2$. \end{thm} We remark that the relatively recent paper \cite{GL_CLT} has relaxed the exponential moment condition $(i)$ to a second moment condition which cannot be improved. In Section \ref{variance}, we use Monte Carlo simulations to approximate the value of $\sigma^2$ for two matrix models that satisfy the hypotheses of Theorem \ref{lepageclt}. \section{$\text{Bernoulli} \left(p\right)$ Parameter Model} \label{bernp} In this section we consider a random matrix model where the random entry follows a $\text{Bernoulli}\left(p\right)$ distribution and the parameter of interest is $p$. Recall that a random variable $\epsilon\sim \text{Bernoulli}\left(p\right)$ if ${\mathbb{P}\left(\epsilon=1\right)=p}$ and ${\mathbb{P}\left(\epsilon=0\right)=1-p}$. Let $\mu_p$ be the probability measure on $\mathrm{GL}(2,\mathbb R)$ given by \begin{equation}\label{BernoullypMatrix1} \left(\begin{array}{cc} \epsilon_p & 1\\ 1 & 0 \end{array}\right), \ \epsilon_p\sim\text{Bernoulli}\left(p\right),\ 0<p<1. \end{equation} It is straightforward to verify that $\mu_p$ satisfies hypotheses $(i)$-$(iv)$ of Theorem \ref{furstenberg}. We verify them here for completeness. For $(i)$, we see that $\mathbb{E}\left[\log^{+}\left\Vert Y_{1}\right\Vert \right]<\infty$ since $\epsilon_{p}$ has finite support. For $(ii)$, consider the subgroup $G$ generated by the possible realizations of \eqref{BernoullypMatrix1}. Since the determinant of each realization has absolute value $1$, so to does every matrix in $G$. Clearly, the closure of $G$, call it $\bar{G}$, is a closed subgroup that contains the support of $\mu_p$. Hence $G_{\mu_p}\subset\bar{G}$. Moreover, since the absolute value of the determinant is continuous, every matrix in $\bar{G}$ also has determinant with absolute value $1$. It follows that the same holds for $G_{\mu_p}$ as required. For $(iii)$, we first let $F_{0},F_{1},F_{2},F_{3},\dots$ be the usual Fibonacci sequence $0,1,1,2,3,5,\dots$ Then a simple calculation shows that for each positive integer $n$, we have \[ \left(\begin{array}{cc} 1 & 1\\ 1 & 0 \end{array}\right)^{n}=\left(\begin{array}{cc} F_{n+1} & F_{n}\\ F_{n} & F_{n-1} \end{array}\right). \] Since the powers of the matrix \eqref{BernoullypMatrix1} with $\epsilon_{p}=1$ must be in $G_{\mu_p}$ and the norm of the powers grow arbitrarily large with large $n$, it follows that $G_{\mu_p}$ is unbounded and hence not compact. Lastly, hypothesis $(iv)$ can be checked by way of an equivalent condition given in \cite[Proposition II.4.3]{bougerol}. This condition is met as long as for any $\bar{\mathbf{x}}\in\mathbb{P}^{1}\left(\mathbb{R}\right)$, the set $S_{\bar{\mathbf{x}}}=\left\{ M\cdot\bar{\mathbf{x}}: M\in G_{\mu_p}\right\} $ has more than two elements. To see that this holds, suppose at least one of $x_1,x_2\in\mathbb{R}$ is nonzero and consider $\mathbf{x}=\left(\begin{array}{c} x_1\\ x_2 \end{array}\right)$. Drawing the matrix $M$ from \eqref{BernoullypMatrix1} with $\epsilon_{p}=1$, we have \begin{align*} M\cdot\bar{\mathbf{x}} & =\left(\begin{array}{cc} 1 & 1\\ 1 & 0 \end{array}\right)\cdot\bar{\mathbf{x}}=\overline{\left(\begin{array}{c} x_1+x_2\\ x_1 \end{array}\right)}=1+\frac{x_2}{x_1}\in S_{\bar{\mathbf{x}}},\\ M^2\cdot\bar{\mathbf{x}} &= \left(\begin{array}{cc} 2 & 1\\ 1 & 1 \end{array}\right)\cdot\bar{\mathbf{x}}= \overline{\left(\begin{array}{c} 2x_1+x_2\\ x_1+x_2 \end{array}\right)}=1+\frac{x_1}{x_1+x_2}\in S_{\bar{\mathbf{x}}},\\ M^3\cdot\bar{\mathbf{x}}&= \left(\begin{array}{cc} 3 & 2\\ 2 & 1 \end{array}\right)\cdot\bar{\mathbf{x}} = \overline{\left(\begin{array}{c} 3x_1+2x_2\\ 2x_1+x_2 \end{array}\right)}=1+\frac{x_1+x_2}{2x_1+x_2}\in S_{\bar{\mathbf{x}}}. \end{align*} Since for any $\mathbf{x}$, each of these elements in $S_{\bar{\mathbf{x}}}$ is distinct, it follows that hypothesis $(iv)$ holds. \begin{comment} when $\mathbf{x}=\left(\begin{array}{c} x\\ 1 \end{array}\right)$ and $x\neq0$. For the second case, we take $x=0$ in $\mathbf{x}=\left(\begin{array}{c} x\\ 1 \end{array}\right)$. Letting $M=\left(\begin{array}{cc} 1 & 1\\ 1 & 0 \end{array}\right)$ and similarly computing $M^i\cdot\bar{\mathbf{x}}$ with $i=1,2,3$ also shows that $S_{\bar{\mathbf{x}}}$ has at least $3$ elements. Finally, we consider the case when $\mathbf{x}=\left(\begin{array}{c} 1\\ 0 \end{array}\right)$ where computing $M^i\cdot\bar{\mathbf{x}}$ with $i=1,2,3$ again shows that $S_{\bar{\mathbf{x}}}$ has at least three elements. This verifies that the condition equivalent to $(iv)$ holds for all $\bar{\mathbf{x}}\in\mathbb{P}^{1}\left(\mathbb{R}\right)$. \end{comment} Since $\mu_p$ satisfies hypotheses $(i)$-$(iv)$ of Theorem \ref{furstenberg}, we know there exists a unique $\mu_p$-invariant distribution $\nu_p$ that satisfies \eqref{eq:invariance} and that $\nu_p$ is atomless. Then by \eqref{Invariant-Formula}, any random variable $X_p$ with law $\nu_p$ must satisfy the distributional identity \begin{equation}\label{Bernp2} X_p\sim\frac{1}{X_p}+\epsilon_p, \end{equation} where $\epsilon_p\sim \text{Bernoulli}\left(p\right)$ and is independent of $X_p$. Likewise, the law of any $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable $X_p$ which satisfies \eqref{Bernp2} is $\mu_p$-invariant hence it must be $\nu_p$. Using \eqref{Bernp2} and the fact that $\nu_p$ is atomless, it is not hard to see that $X_p\in (0,\infty)$ almost surely. See Goswami \cite{Goswami2004} for this fact and other facts about $X$, including an expression for its cumulative distribution function in terms of a continued fraction expansion. In Figures \ref{fig:histogram} and \ref{fig:CDF} we show the empirical distribution of $100\,000$ independent draws from $\nu_{1/2}$ and remark that the fractal nature of this probability measure is clearly apparent. \begin{figure}[H] \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[scale=0.5]{histogram_short.pdf} \caption{Histogram} \label{fig:histogram} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[scale=0.5]{cdf_short.pdf} \caption{CDF} \label{fig:CDF} \end{subfigure} \caption{} \end{figure} Let $\lambda(p)$ be the Lyapunov exponent related to $\mu_p$. Using \eqref{expectation} and the fact that $X_p$ is non-negative, we can write the Lyapunov exponent associated with $\mu$ as \begin{equation}\label{eq:lambda_p} \lambda(p)=\mathbb{E}\left[\log X_p\right]. \end{equation} \subsection{The general $0<p<1$ case}\label{bernp1} In this subsection we study $\lambda(p)$ for general $0<p<1$ and obtain two sided bounds depending on the parameter $p$. First we prove some identities for $\mathbb{E}\left[\log X_{p}\right]$. We begin by establishing an identity for $\mathbb{E}\left[\log X_p\right]$ which will be later generalized for the $p=\frac{1}{2}$ case and used in proving a limiting result. \begin{lem}\label{prop:p_identity If $X_p$ is a $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable satisfying \eqref{Bernp2}, then \[ 0<\mathbb{E}\left[\log X_p\right]<\infty \] and \[ \mathbb{E}\left[\log X_{p}\right]=\frac{p}{3}\mathbb{E}\left[\log\left(2X_{p}+1\right)\right]. \] \end{lem} \begin{proof} Let $X_p$ be a random variable satisfying \eqref{Bernp2}. Then $X_p$ has law $\nu_p$ given by Theorem \ref{furstenberg} applied to random matrices of the form \eqref{BernoullypMatrix1}. Consequently, we have that $0<\lambda(p)<\infty$ and it follows from \eqref{eq:lambda_p} that $\mathbb{E}\left[\log X_p\right]$ is positive and finite. Using \eqref{Bernp2}, we start by writing \begin{align} \mathbb{E}\left[\log X_p\right] & =\mathbb{E}\left[\log\left(\frac{1}{X_p}+\epsilon\right)\right]\nonumber \\ & =(1-p)\mathbb{E}\left[\log\left(\frac{1}{X_p}\right)\right]+p\mathbb{E}\left[\log\left(\frac{1}{X_p}+1\right)\right]\nonumber \\ & =-(1-p)\mathbb{E}\left[\log X_p\right]+p\mathbb{E}\left[\log\left(\frac{1+X_p}{X_p}\right)\right]\nonumber \\ & =-\mathbb{E}\left[\log X_p\right]+p\mathbb{E}\left[\log\left(1+X_p\right)\right].\label{eq:7} \end{align} Adding $\mathbb{E}\left[\log X_p\right]$ to both sides of \eqref{eq:7} and dividing by 2 results in \begin{equation} \mathbb{E}\left[\log X_p\right]=\frac{p}{2}\mathbb{E}\left[\log\left(1+X_p\right)\right].\label{eq:8} \end{equation} Continuing in a similar fashion with \eqref{eq:8}, we obtain \begin{align} \mathbb{E}\left[\log X_p\right] & =\frac{p}{2}\mathbb{E}\left[\log\left(1+\frac{1}{X_p}+\epsilon\right)\right]\nonumber \\ & =\frac{p(1-p)}{2}\mathbb{E}\left[\log\left(1+\frac{1}{X_p}\right)\right]+\frac{p^2}{2}\mathbb{E}\left[\log\left(2+\frac{1}{X_p}\right)\right]\nonumber \\ & =\frac{p(1-p)}{2}\mathbb{E}\left[\log\left(\frac{X_p+1}{X_p}\right)\right]+\frac{p^2}{2}\mathbb{E}\left[\log\left(\frac{2X_p+1}{X_p}\right)\right]\nonumber \\ & =\frac{p(1-p)}{2}\mathbb{E}\left[\log\left(X_p+1\right)\right]+\frac{p^2}{2}\mathbb{E}\left[\log\left(2X_p+1\right)\right]-\frac{p}{2}\mathbb{E}\left[\log X_p\right]\nonumber \\ & =\left(1-\frac{3p}{2}\right)\mathbb{E}\left[\log X\right]+\frac{p^2}{2}\mathbb{E}\left[\log\left(2X+1\right)\right],\label{eq:9} \end{align} where we use \eqref{eq:8} in the last equality. Subtracting $\left(1-\frac{3p}{2}\right)\mathbb{E}\left[\log X\right]$ from both sides of \eqref{eq:9} leads to \begin{equation*} \mathbb{E}\left[\log X\right]=\frac{p}{3}\mathbb{E}\left[\log\left(2X+1\right)\right], \end{equation*} completing the proof. \end{proof} \begin{lem}\label{Expectation-Identity1 If $X_p$ is a $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable satisfying \eqref{Bernp2}, then \begin{equation} \mathbb{E}\left[\log\left(X_{p}\right)\cdot\mathbf{1}_{\left(X_{p}<1\right)}\right]=\left(p-1\right)\mathbb{E}\left[\log\left(X_{p}\right)\cdot\mathbf{1}_{\left(X_{p}>1\right)}\right],\label{eq:c} \end{equation} \begin{equation} \mathbb{E}\left[\log \left(X_{p}\right)\cdot\mathbf{1}_{\left(X_{p}>1\right)}\right]=\frac{1}{p}\mathbb{E}\left[\log X_{p}\right],\label{eq:a} \end{equation} and \begin{equation} \mathbb{E}\left[\log\left(X_{p}\right)\cdot\mathbf{1}_{\left(X_{p}<1\right)}\right]=\frac{p-1}{p}\mathbb{E}\left[\log X_{p}\right].\label{eq:b} \end{equation} \end{lem} \begin{proof} Recalling that the distribution of $X_p$ has non-negative support, observe that \begin{eqnarray*} \mathbb{E}\left[\log\left(X_{p}\right)\cdot\mathbf{1}_{\left(X_{p}<1\right)}\right] & = & p~\mathbb{E}\left[\log\left(\frac{1}{X_{p}}+1\right)\cdot \mathbf{1}_{\left(\frac{1}{X_{p}}+1<1\right)}\right]+\left(1-p\right)\mathbb{E}\left[\log\left(\frac{1}{X_{p}}\right)\cdot \mathbf{1}_{\left(\frac{1}{X_{p}}<1\right)}\right]\\ & = & 0+\left(1-p\right)\mathbb{E}\left[\log\left(\frac{1}{X_{p}}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\\ & = & \left(p-1\right)\mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]. \end{eqnarray*} This proves \eqref{eq:c} which, along with the fact that the distribution of $X_p$ is atomless, allows us to write \begin{eqnarray*} \mathbb{E}\left[\log X_{p}\right] & = & \mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]+\mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}<1\right)}\right]\\ & = & \mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]+\left(p-1\right)\mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\\ & = & p~\mathbb{E}\left[\log\left(X_{p}\right)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right] \end{eqnarray*} which proves \eqref{eq:a}. Combining these two identities now leads to \eqref{eq:b}. \end{proof} Next we use these results to establish bounds on the Lyapunov exponent which are dependent on $p$. \begin{thm}\label{UBprop Let $\mu_p$ be the probability measure on $\mathrm{GL}(2,\mathbb R)$ given by \eqref{BernoullypMatrix1}. Then the Lyapunov exponent $\lambda(p)$ associated with $\mu_p$ can be estimated by \[ \frac{p\log 3}{4-p}\leq\lambda(p)\leq\frac{p\log 3}{2}. \] \end{thm} \begin{proof} Beginning with the upper estimate, first note that $\log(2x+1)\leq \log(3x)$ for $x\geq 1$. Now using Lemma \ref{prop:p_identity} and \eqref{eq:a}, we can write \begin{eqnarray} \mathbb{E}\left[\log X_{p}\right] & = & \frac{p}{3}\mathbb{E}\left[\log(2X_{p}+1)\right]\nonumber \\ & = & \frac{p}{3}\mathbb{E}\left[\log(2X_{p}+1)\cdot \mathbf{1}_{\left(X_{p}<1\right)}+\log(2X_{p}+1)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\nonumber \\ & \leq & \frac{p}{3}\Big(\log 3~\mathbb{P}\left(X_{p}<1\right)+\mathbb{E}\left[\log(3X_{p})\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\Big)\nonumber \\ & = & \frac{p}{3}\Big(\log3+\mathbb{E}\left[\log(X_{p})\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\Big)\nonumber \\ & = & \frac{p}{3}\log3+\frac{1}{3}\mathbb{E}\left[\log X_{p}\right].\label{eq:g} \end{eqnarray} Subtracting $\frac{1}{3}\mathbb{E}\left[\log X_{p}\right]$ from both sides of \eqref{eq:g} and recalling \eqref{eq:lambda_p} leads to the desired result. For the lower estimate, we proceed similarly, noting that $\log(2x+1)\geq\log(3x)$ for $0<x\leq 1$ and using \eqref{eq:b} instead of \eqref{eq:a} to write \begin{eqnarray} \mathbb{E}\left[\log X_{p}\right] & = & \frac{p}{3}\mathbb{E}\left[\log(2X_{p}+1)\cdot \mathbf{1}_{\left(X_{p}<1\right)}+\log(2X_{p}+1)\cdot \mathbf{1}_{\left(X_{p}>1\right)}\right]\nonumber \\ & \geq & \frac{p}{3}\Big(\mathbb{E}\left[\log(3X_{p})\cdot \mathbf{1}_{\left(X_{p}<1\right)}\right]+\log3~\mathbb{P}\left(X_{p}>1\right)\Big)\nonumber \\ & = & \frac{p}{3}\Big(\log3+\mathbb{E}\left[\log(X_{p})\cdot \mathbf{1}_{\left(X_{p}<1\right)}\right]\Big)\nonumber \\ & = & \frac{p}{3}\log3+\frac{p-1}{3}\mathbb{E}\left[\log X_{p}\right].\label{eq:h} \end{eqnarray} Now the lower bound follows from a simple rearrangement of \eqref{eq:h}. \end{proof} \subsection{Approximating $\lambda(p)$ by simulation}\label{sec:approx}\hfill\\ Let $\{Y_i\}_{i\ge1}$ be an i.i.d. sequence drawn from $\mu_p$, and for some $\mathbf{x}\in\mathbb{R}^2$ with $\|\mathbf{x}\|=1$, construct $\{U_i\}_{i\geq 0}$ recursively by $U_0=\mathbf{x}$ and $U_i=Y_i\frac{U_{i-1}}{\|U_{i-1}\|}$. Now, with $S_n=Y_n Y_{n-1}\cdots Y_2 Y_1$ and $S_0=Y_0=I$, we have \begin{align} \frac{1}{n}\log\|S_n\mathbf{x}\|&=\frac{1}{n}\sum_{i=1}^n\log\frac{\|S_i\mathbf{x}\|}{\|S_{i-1}\mathbf{x}\|}\nonumber \\ &=\frac{1}{n}\sum_{i=1}^n\log\left\|Y_i\frac{Y_{i-1}\dots Y_1\mathbf{x}}{\|Y_{i-1}\dots Y_1\mathbf{x}\|}\right\|\nonumber \\ &=\frac{1}{n}\sum_{i=1}^n\log\left\|U_i\right\|.\label{eq:lln_approx} \end{align} Hence it follows from Theorem \ref{furstenbergkesten} that we can approximate $\lambda$ by the right-hand side of \eqref{eq:lln_approx} with $n$ large. Since the $\log\left\|U_i\right\|$ terms aren't growing with $i$, this avoids numerical overflow issues and makes for a robust Monte Carlo scheme. In Figure \ref{fig:BernPLambdaBounds}, we plot simulations for $\lambda(p)$ in black and the upper and lower bounds from Theorem \ref{UBprop} in blue. We discretize $[0,1]$ into sub-intervals of length $0.01$ and use $n=1\,000\,000$ in the Monte Carlo scheme described above. \begin{figure}[H] \centering \includegraphics[scale=0.5]{imgBernPLambdaBounds.pdf} \caption{$n=1\,000\,000$} \label{fig:BernPLambdaBounds} \end{figure} \subsection{The $p=\frac{1}{2}$ case}\label{bernhalf} In this section we study $\lambda :=\lambda\left(\frac{1}{2}\right)$ in more detail. To set notation, recall that a random variable $\epsilon\sim \text{Bernoulli}\left(\frac{1}{2}\right)$ if $\mathbb{P}\left(\epsilon=1\right)=\mathbb{P}\left(\epsilon=0\right)=\frac{1}{2}$. The probability measure $\mu$ on $\mathrm{GL}(2,\mathbb R)$ that we consider is given by \begin{equation} \left(\begin{array}{cc} \epsilon & 1\\ 1 & 0 \end{array}\right),\ \epsilon\sim \text{Bernoulli}\left(\frac{1}{2}\right).\,\,\,\,\,\label{Ber(1/2) - Intro} \end{equation} We know by the general $p$ case that there exists a unique $\mu$-invariant distribution $\nu$ that satisfies \eqref{eq:invariance} and that $\nu$ is atomless. Then by \eqref{Invariant-Formula}, any random variable $X$ with law $\nu$ must satisfy the distributional identity \begin{equation}\label{Invariant1} X\sim\frac{1}{X}+\epsilon, \end{equation} where $\epsilon\sim \text{Bernoulli}\left(\frac{1}{2}\right)$ and is independent of $X$. Using \eqref{expectation} and the fact that $X$ is non-negative, we can write the Lyapunov exponent associated with $\mu$ as \begin{equation}\label{eq:lambda_log} \lambda=\mathbb{E}\left[\log X\right]. \end{equation} Unlike in the general case, we will be able to obtain a sequence of upper and lower bounds that converge to $\lambda$. Recall that by Lemma \ref{prop:p_identity} for $p=\frac{1}{2}$ we showed that \[ 0<\mathbb{E}\left[\log X\right]<\infty \] and \begin{equation}\label{FirstIdentity} \mathbb{E}\left[\log X\right]=\frac{1}{6}\mathbb{E}\left[\log\left(2X+1\right)\right]. \end{equation} We will prove a string of identities akin to equation \eqref{FirstIdentity} in a similar fashion. Here we list a few examples. \begin{equation}\label{Expectation of Poly} \begin{aligned} \mathbb{E}\left[\log X\right] & =\frac{1}{6}\mathbb{E}\left[\log\left(2X+1\right)\right] \\ & =\frac{1}{14}\mathbb{E}\left[\log\left(3X+2\right)(X+2)\right] \\ & =\frac{1}{32}\mathbb{E}\left[\log\left(5X+3\right)\left(3X+1\right)\left(2X+3\right)\left(2X+1\right)\right] \\ & =\frac{1}{72}\mathbb{E}\left[\log\left(8X+5\right)\left(4X+3\right)\left(5X+2\right)\left(3X+2\right)\left(3X+5\right)\left(X+3\right)\left(3X+2\right)\left(X+2\right)\right]\\ & \ \ \vdots \end{aligned} \end{equation} The string of identities above is obtained by iteratively exploiting the distributional equivalence of $X$ and $\frac{1}{X}+\epsilon$, the independence of $X$ and $\epsilon$, and elementary logarithmic identities. We will later see that an interesting pattern emerges. At the first step of the iteration, we are looking at the expected value of the $\log$ of one affine function of $X$ that is obtained by taking the inner product of the vector $(2,1)$ and the vector $(X,1)$. As we move to the second step of the iteration, we encounter the expectation of the $\log$ of the product of two affine functions of $X$. The first one is obtained by taking the inner product of $(3,2)$ and $(X,1)$, while the second is obtained by taking the inner product of $(1,2)$ and $(X,1)$. At the third step, we encounter the expected value of the $\log$ of the product of four $\left(=2^{3-1}\right)$ affine functions of $X$; these are obtained by respectively taking the inner product of $(X,1)$ with the vectors $(5,3)$, $(3,1)$, $(2,3)$, and $(2,1)$. In what follows, we represent the vectors generating the aforesaid affine functions of $X$ via inner products with $(X,1)$, which we call ``coefficient pairs'', in an array where the row number corresponding to the $n^{\text{th}}$ step of the iteration is $n-1$. The first four rows of the array are shown below. We use the symbol $\mapsto$ to map the collection of coefficient pairs to the real number representing the product of the sum of entries in each coefficient pair in the row; we make extensive use of these quantities later on. \begin{equation}\label{an,bn array} \begin{aligned} n=0 & \hspace{6mm}\left(2,1\right) \mapsto 3 \\ n=1 & \hspace{6mm}\left(3,2\right)\left(1,2\right) \mapsto 5\cdot3 = 15\\ n=2 & \hspace{6mm}\left(5,3\right)\left(3,1\right)\left(2,3\right)\left(2,1\right) \mapsto 8\cdot4\cdot5\cdot3=480 \\ n=3 & \hspace{6mm}\left(8,5\right)\left(4,3\right)\left(5,2\right)\left(3,2\right)\left(3,5\right)\left(1,3\right)\left(3,2\right)\left(1,2\right) \mapsto 13\cdot7\cdot7\cdot5\cdot8\cdot4\cdot5\cdot3=1528800 \\ \vdots\hspace{3.2mm} & \hspace{75.5mm}\vdots \end{aligned} \end{equation} For the $k^\text{th}$ coefficient pair in row $n$, let $a_{n}^{k}$ denote the first element and $b_{n}^{k}$ the second. To illustrate this notational convention, consider the example $\frac{1}{14}\mathbb{E}\left[\log\left(3X+2\right)\left(X+2\right)\right]$ from \eqref{Expectation of Poly}. This is in row $n=1$, so we would refer to the 3 in $\left(3X+2\right)$ as $a_{1}^{1}$ and the 2 as $b_{1}^{1}$. Similarly, the coefficient of $X$ in $\left(X+2\right)$ would be labeled $a_{1}^{2}$ and the $2$ would be labeled $b_{1}^{2}$. In terms of $a^k_n$ and $b^k_n$, the expression is $\frac{1}{14}\mathbb{E}\left[\log\left(a_{1}^{1}X+b_{1}^{1}\right)\left(a_{1}^{2}X+b_{1}^{2}\right)\right]$. Now we can define the multi-level recursion that describes the array given in \eqref{an,bn array}. \begin{defn}\label{def:recursion} Set $a_{0}^{1}=2$ and $b_{0}^{1}=1$. For any $n\in \mathbb{Z}_{\geq 0}$, define \[ \begin{array}{ccl} \left(a_{n+1}^{k},b_{n+1}^{k}\right):=\left(a_{n}^{k}+b_{n}^{k},a_{n}^{k}\right),\hfill \hfill & \mbox{ for }k=1,\dots,2^{n},\hfill\\ \left(a_{n+1}^{k},b_{n+1}^{k}\right):=\left(b_{n}^{k-2^{n}},a_{n}^{k-2^{n}}\right),\hfill \hfill& \mbox{ for }k=2^{n}+1,\dots,2^{n+1}.\hfill\\ \\ \end{array} \] \end{defn} We observe several conspicuous patterns in \eqref{an,bn array} which are implicit in Definition \ref{def:recursion}. For instance, row $n$ is made up of $2^{n}$ pairs and the second half of row $n$ is simply row $n-1$ where the elements within the coefficient pairs have been switched. One property that will prove useful is the fact that the first coefficient pair in each row dominates the other pairs occurring in that row in the sense that \begin{equation}\label{eq:dominate} a_n^1\geq a_n^k~\text{ and }~b_n^1\geq b_n^k~\text{ for all }~1\leq k\leq 2^n. \end{equation} This follows from the recursion in Definition \ref{def:recursion} and induction on $n$. To exhibit a less obvious pattern, we first recall that a ``Fibonacci-like sequence'' of numbers $f_0,f_1,f_2\dots$ is a sequence determined by the initial values $f_0,f_1$ such that \[ f_{n+1}=f_{n}+f_{n-1} \] for all $n\in \mathbb{N}$. When $f_0=0,f_1=1$, we recover the standard Fibonacci sequence. Fibonacci-like sequences can be expressed by an explicit formula. Let $f_{n}(f_0,f_1)$ represent the $n$th term in the sequence given initial values $f_0,f_1$. If \[ \phi_{1}=\frac{1+\sqrt{5}}{2}\text{ and }\phi_{2}=\frac{1-\sqrt{5}}{2}, \] then \begin{equation} f_{n}(f_{0},f_{1})=\frac{f_{1}-f_{0}\phi_{2}}{\sqrt{5}}\left(\phi_{1}\right)^{n}+\frac{f_{0}\phi_{1}-f_{1}}{\sqrt{5}}\left(\phi_{2}\right)^{n}.\label{fibonacci} \end{equation} Now note that given $n\in\mathbb{N}$ and $k\in\left\{ 1,\dots,2^{n-1}\right\}$, we have \[ a_{n+1}^{k}=a_{n}^{k}+b_{n}^{k}=a_{n}^{k}+a_{n-1}^{k} \] and \[ b_{n+1}^{k}=a_{n}^{k}=a_{n-1}^k+b_{n-1}^{k}=b_{n}^k+b_{n-1}^{k}. \] Thus, for each $k$, the sequences $\{a_{n}^{k}\}$ and $\{b_{n}^{k}\}$ will be Fibonacci-like sequences in $n$ for $n$ large enough. We use these observations to help establish bounds on the Lyapunov exponent. In order to find suitable estimates, we first need to establish some preliminary results. These involve proving the string of identities given in \eqref{Expectation of Poly}. We also need to prove some elementary inequalities involving the logarithm of the polynomials given inside the expectations in \eqref{Expectation of Poly}. First, we extend the identities given in \eqref{Expectation of Poly} to all $n$. \begin{lem}\label{Prop:3.2 If $X$ is a $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable satisfying \eqref{Invariant1}, then \begin{equation} \mathbb{E}\left[\log X\right]=\frac{1}{(n+6)2^{n}}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}X+b_{n}^{k}\right)\right)\right]\label{eq:9-1} \end{equation} for all $n\in \mathbb{Z}_{\geq 0}$. \end{lem} \begin{proof} We begin with $n=0$. By Lemma \ref{prop:p_identity} with $p=\frac{1}{2}$ we have, \begin{align*} \mathbb{E}\left[\log X\right] & =\frac{1}{6}\mathbb{E}\left[\log\left(2X+1\right)\right]\\ & =\frac{1}{(0+6)2^{0}}\mathbb{E}\left[\log\left(a_{0}^{1}X+b_{0}^{1}\right)\right]. \end{align*} Now suppose \eqref{eq:9-1} holds for $n$. We shall show that \eqref{eq:9-1} holds for $n+1$. Note that \begin{align} \mathbb{E}\left[\log X\right] & =\frac{1}{(n+6)2^{n}}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}X+b_{n}^{k}\right)\right)\right]\nonumber \\ & =\frac{1}{(n+6)2^{n}}\left(\frac{1}{2}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}\left(\frac{1}{X}+1\right)+b_{n}^{k}\right)\right)\right]+\frac{1}{2}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}\left(\frac{1}{X}\right)+b_{n}^{k}\right)\right)\right]\right)\nonumber \\ & =\frac{1}{(n+6)2^{n+1}}\left(\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(\frac{a_{n}^{k}}{X}+a_{n}^{k}+b_{n}^{k}\right)\right)\right]+\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(\frac{a_{n}^{k}}{X}+b_{n}^{k}\right)\right)\right]\right)\nonumber \\ & =\frac{1}{(n+6)2^{n+1}}\left(\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(\frac{a_{n}^{k}+\left(a_{n}^{k}+b_{n}^{k}\right)X}{X}\right)\right)\right]+\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(\frac{a_{n}^{k}+b_{n}^{k}X}{X}\right)\right)\right]\right)\nonumber \\ & =\frac{1}{(n+6)2^{n+1}}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\Big(a_{n}^{k}+\left(a_{n}^{k}+b_{n}^{k}\right)X\Big)\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}X\right)\right)\right]-\frac{\mathbb{E}\left[\log X\right]}{(n+6)}.\label{eq:10} \end{align} Moving the last term on the right-hand side of \eqref{eq:10} to the left leads to \begin{align*} \mathbb{E}\left[\log X\right] & =\frac{1}{\big((n+1)+6\big)2^{n+1}}\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n+1}}\left(a_{n+1}^{k}X+b_{n+1}^{k}\right)\right)\right]. \end{align*} Here we have combined and simplified the products appearing in \eqref{eq:10} by using the recursion from Definition \ref{def:recursion}. The result now follows by induction. \end{proof} We now prove the elementary inequalities needed to estimate \eqref{eq:9-1}. \begin{lem Let $n\in \mathbb{Z}_{\geq 0}$. For $x\geq 1$, \begin{equation}\label{lem1} \log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}x+b_{n}^{k}\right)\right)\leq\log\left(x^{2^{n}}\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}\right)\right). \end{equation} Conversely, when $0<x\leq 1$, \begin{equation}\label{lem2} \log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}x+b_{n}^{k}\right)\right)\geq\log\left(x^{2^{n}}\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}\right)\right). \end{equation} \end{lem} \begin{proof} Note that when $x\geq 1$, we have $a^k_n x+b^k_n\leq x (a^k_n+b^k_n)$. Taking products and the $\log$ of both sides gives us the desired result. The proof of the $0<x\leq 1$ case follows similarly. \end{proof} Using \eqref{lem1} and \eqref{lem2}, we can prove that the Lyapunov exponent is bounded by terms dependent only on $n$. First, we define the following quantities that appear as the rightmost entries of \eqref{an,bn array}. \begin{defn}\label{def:c_n} For each $n\in \mathbb{Z}_{\geq 0}$, let $c_{n}$ be the product of the sums of coefficient pairs in row $n$ of \eqref{an,bn array}. That is, \[ c_{n}=\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}\right). \] \end{defn} \noindent For example, $c_0,\dots, c_3$ are displayed in \eqref{an,bn array}. We remark that the recursion from Definition \ref{def:recursion} implies \begin{equation}\label{eq:cn_rec} c_n=c_{n-1}\prod_{k=1}^{2^{n-1}}\left(a_{n}^{k}+b_{n}^{k}\right)=\prod_{k=1}^{2^{n}}a_{n+1}^{k}. \end{equation} Now we can state our main result of this section. \begin{thm}\label{thm:bounds Let $\mu$ be the probability measure on $\mathrm{GL}(2,\mathbb R)$ given by \eqref{Ber(1/2) - Intro}. Then for each $n\in \mathbb{Z}_{\geq 0}$, the Lyapunov exponent $\lambda$ associated with $\mu$ can be estimated by \begin{equation}\label{Bounds} p_n\leq\lambda\leq q_n, \end{equation} where \begin{equation}\label{ConvergingSequence} p_n =\frac{\log c_{n}}{\left(n+7\right)2^{n}} \,\,\, \mbox{and}\,\,\, q_n = \frac{\log c_{n}}{\left(n+4\right)2^{n}}. \end{equation} Moreover, \[ \lim_{n\to \infty}p_{n} = \lim_{n\to\infty} q_{n} = \lambda. \] \end{thm} \begin{proof} Fix $n\in \mathbb{N}\cup \left\lbrace 0 \right\rbrace$ and let $X$ be a $\mathbb{P}^{1}\left(\mathbb{R}\right)$-valued random variable satisfying \eqref{Invariant1}. Since the distribution of $X$ is atomless, we can use Lemma \ref{Prop:3.2} and \eqref{lem1} to write \begin{align} \mathbb{E}\left[\log X\right] & =\frac{1}{(n+6)2^{n}}\left(\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}X+b_{n}^{k}\right)\right)\cdot\mathbf{1}_{\left(X<1\right)}\right]+\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}X+b_{n}^{k}\right)\right)\cdot\mathbf{1}_{\left(X>1\right)}\right]\right)\nonumber \\ & \leq\frac{1}{(n+6)2^{n}}\left(\mathbb{E}\left[\log\left(\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}\right)\right)\cdot\mathbf{1}_{\left(X<1\right)}\right]+\mathbb{E}\left[\log\left(X^{2^{n}}\prod_{k=1}^{2^{n}}\left(a_{n}^{k}+b_{n}^{k}\right)\right)\cdot\mathbf{1}_{\left(X>1\right)}\right]\right).\nonumber \end{align} Moreover, using \eqref{eq:cn_rec} and \eqref{eq:a} from Lemma \ref{Expectation-Identity1}, it follows that \begin{align} \mathbb{E}\left[\log X\right] & \leq\frac{1}{(n+6)2^{n}}\Big(\log(c_n)\cdot\mathbb{P}(X<1)+2^n\mathbb{E}\left[\log\left(X\right)\cdot\mathbf{1}_{\left(X>1\right)}\right]+\log(c_n)\cdot\mathbb{P}(X>1)\Big)\nonumber \\ & =\frac{\log c_{n}}{(n+6)2^{n}}+\frac{2\mathbb{E}\left[\log X\right]}{n+6}.\label{eq:12} \end{align} Subtracting the last term on the right-hand side of \eqref{eq:12} from both sides while recalling \eqref{eq:lambda_log} leads to \[ \lambda=\mathbb{E}\left[\log X\right]\leq\frac{\log c_{n} }{\left(n+4\right)2^{n}}. \] For the lower bound, we can repeat this same procedure using \eqref{lem2} and \eqref{eq:b} instead of \eqref{lem1} and \eqref{eq:a} to arrive at \[ \mathbb{E}\left[\log X\right]\geq \frac{\log c_{n}}{(n+6)2^{n}}-\frac{\mathbb{E}\left[\log X\right]}{n+6}. \] Similarly, this implies \[ \frac{\log c_{n}}{(n+7)2^{n}}\leq\lambda. \] We now show that these bounds converge to the Lyapunov exponent as $n\to\infty$. We first point out the crude estimate $c_{n}\leq \left(F_{n+4}\right)^{2^{n}}$ where $\{F_n\}:= \{f_{n}(0,1) \}$ is the usual Fibonacci sequence. This follows from \eqref{eq:cn_rec}, \eqref{eq:dominate}, and the fact that $a_n^1=F_{n+3}$ for all $n\geq 0$. Also note that the well-known asymptotic \[ F_n\sim\frac{(\phi_1)^n}{\sqrt{5}}~\text{ as }~n\to\infty \] implies \[ \lim_{n\to\infty}\frac{\log\left(\left(F_{n+4}\right)^{2^{n}}\right)}{(n+4)2^{n}}=\log\left(\phi_1\right). \] Hence we have \begin{align} \limsup_{n\to\infty}\left|q_{n}-p_{n}\right| & =\limsup_{n\to\infty}\frac{3\log c_{n}}{\left(n+7\right)\left(n+4\right)2^{n}}\nonumber \\ & \leq \lim_{n\to\infty}\frac{3\log\left(\left(F_{n+4}\right)^{2^{n}}\right)}{(n+7)\left(n+4\right)2^{n}}\nonumber \\ & =0.\nonumber \end{align} Now the result follows from \eqref{Bounds}. \end{proof} We end this section with the following two remarks. \begin{remark} There doesn't seem to be an obvious recursion among the $c_{n}$ values. In order to compute $c_{n}$ using its definition, we must consider $2^{n}$ coefficient pairs. We are able to compute ${p_{25}\approx 0.204266}$ and ${q_{25}\approx 0.225397}$ but going beyond $n=25$ exceeds our computing power. After implementing a simple numerical scheme to compute $\mathbb{E}\left[\log X\right]$ using the CDF of $X$ from Theorem 5.2 of \cite{Goswami2004} along with \eqref{eq:b}, we expect that $\lambda\approx 0.2165$. \end{remark} \begin{remark} The bounds in Lemma \ref{UBprop} from the general $p$ case are analogous to $p_0$ and $q_0$ from \eqref{ConvergingSequence} of the $\text{Bernoulli}\left(\frac{1}{2}\right)$ model. While we can attempt to improve these bounds by mimicking the proof of Theorem \ref{thm:bounds}, unlike in that case, there doesn't appear to be a nice expression for the corresponding bounds $p_n$ and $q_n$ as $n$ gets larger. \end{remark} \section{$\mathbf{\xi} \cdot \text{Cauchy}$ Parameter Model}\label{xicauchy} The parameter model studied in this section is based on the standard Cauchy distribution (that is, Cauchy with location $x_0 = 0$ and scale $\gamma = 1$). Recall that the probability density function of a $\text{Cauchy} \left(x_0,\gamma\right)$ random variable with location $x_0\in\mathbb{R}$ and scale $\gamma>0$ is \begin{equation}\label{eq:Cauchy_dens} f(x)=\frac{1}{\pi \gamma \left(1+\left(\frac{x-x_0}{\gamma}\right)^2\right)},\ -\infty < x < \infty. \end{equation} Let $\mu_\xi$ be the probability measure on $\mathrm{GL}(2,\mathbb R)$ given by \begin{equation}\label{CauchyMatrix1} \left(\begin{array}{cc} \xi \epsilon & -1\\ 1 & 0 \end{array}\right), \ \epsilon\sim \text{Cauchy} \left(0,1\right), \ \xi\in\mathbb{\mathbb{R}}, \ \xi\neq 0. \end{equation} The fact that $\mu_\xi$ satisfies the hypotheses of Theorem \ref{furstenberg} can be seen through a similar analysis as done in the beginning of Section \ref{bernp} with some slight differences which we now point out. To verify hypothesis $(i)$, we can use the Frobenius matrix norm to arrive at $\mathbb E[\log^+\|Y_1\|]=\frac{1}{2}\int\log(2+\xi^{2}x^{2})f(x)dx$ where $f(x)$ is the density for $\mbox{Cauchy}\left(0,1\right)$. By elementary computations, this integral is seen to be finite for all $\xi$. Hypothesis $(ii)$ can be verified in the same manner as for the $\text{Bernoulli}\left(p\right)$ model. Hypothesis $(iii)$ follows from the unbounded support of $\epsilon$. For hypothesis $(iv)$, we can again use the equivalent condition given in \cite[Proposition II.4.3]{bougerol}. More specifically, draw $M$ from \eqref{CauchyMatrix1} with $\epsilon=\frac{1}{\xi}$ and proceed as in the beginning of Section \ref{bernp}. Hence we know there exists a unique $\mu_\xi$-invariant distribution $\nu_\xi$ such that a random variable $X_\xi$ has law $\nu_\xi$ if and only if it satisfies the distributional identity \begin{equation}\label{Cauchy-distributional} X_\xi\sim-\frac{1}{X_\xi}+\xi\epsilon, \end{equation} where $\epsilon\sim \text{Cauchy} \left(0,1\right)$ and is independent of $X_\xi$. The goal of this section is to find the explicit value of the Lyapunov exponent $\lambda (\xi)$ related to $\mu_\xi$. Following the method from \cite[pp. 35]{bougerol}, we have an explicit formula for the Lyapunov exponent in terms of the parameter $\xi$. This will allow us to to study the variance in the Central Limit Theorem related to the products of random matrices of the form \eqref{CauchyMatrix1} as formulated in Theorem \ref{lepageclt}. Since the Lyapunov exponent used in our Monte Carlo simulation scheme will be exact, we can obtain a better approximation for the variance compared to the other models we study. \begin{prop}\label{CauchyProp} Let $\mu_\xi$ be the probability measure on $\mathrm{GL}(2,\mathbb R)$ given by \eqref{CauchyMatrix1}. Then the Lyapunov exponent $\lambda(\xi)$ associated with $\mu_\xi$ is given by \[ \lambda(\xi)=\log\left(\frac{|\xi|+\sqrt{\xi^{2}+4}}{2}\right). \] \end{prop} \begin{proof} According to \eqref{expectation}, we have $\lambda (\xi)=\mathbb{E}\big[\log |X_\xi|\big]$, where $X_\xi$ is a random variable satisfying \eqref{Cauchy-distributional}. To find the law of such an $X_\xi$, we first guess that it is $\text{Cauchy} \left(0,\gamma\right)$ for some $\gamma>0$ and then verify that it satisfies \eqref{Cauchy-distributional} for a particular $\gamma$. Assuming that $X_\xi\sim\text{Cauchy} \left(0,\gamma\right)$, the well-known transformation properties of the Cauchy distribution imply that the right-hand side of \eqref{Cauchy-distributional} is also Cauchy distributed, namely $$-\frac{1}{X_\xi}+\xi\epsilon \sim \text{Cauchy}\left(0,\frac{1}{\gamma}+\left|\xi\right|\right).$$ Hence \eqref{Cauchy-distributional} holds if and only if \[ \gamma=\frac{1}{\gamma}+\left|\xi\right| \] which has as its unique positive solution \[ \gamma=\frac{|\xi|+\sqrt{\xi^2+4}}{2}. \] Now we can use \eqref{eq:Cauchy_dens} to write \begin{align*} \lambda(\xi)=\int_{-\infty}^{\infty}\log|x|\frac{1}{\pi \gamma\left(1+\left(\frac{x}{\gamma}\right)^2\right)}\,\mathrm d x&=\log(\gamma) \\ &=\log\left(\frac{|\xi|+\sqrt{\xi^{2}+4}}{2}\right). \end{align*} The proof is complete because of the uniqueness of the distribution $v_\xi$ such that \eqref{Cauchy-distributional} is satisfied. \end{proof} Figure \ref{fig:Cauchy20Lambda} shows the graph of $\lambda(\xi)$ for $\xi\in[-20,20]$; in Figure \ref{fig:CauchyZoomLambda}, we plot $\lambda(\xi)$ for $\xi\in[-1,1].$ \begin{figure}[H] \centering \begin{subfigure}{.5\textwidth} \centering \includegraphics[scale=0.5]{imgCauchy20Lambda.pdf} \caption{} \label{fig:Cauchy20Lambda} \end{subfigure}% \begin{subfigure}{.5\textwidth} \centering \includegraphics[scale=0.5]{imgCauchyZoomLambda.pdf} \caption{} \label{fig:CauchyZoomLambda} \end{subfigure} \caption{$\lambda(\xi)$ vs. $\xi$} \label{fig:test} \end{figure} \section{Variance Simulation}\label{variance} It is straightforward to verify that the hypotheses of Theorem \ref{lepageclt} are satisfied for the models we studied in Sections \ref{bernp} and \ref{xicauchy}. In fact, much of the reasoning done in the beginning of Sections \ref{bernp} and \ref{xicauchy} to verify the conditions of Theorem \ref{furstenberg} can be used to verify those of Theorem \ref{lepageclt}. For example, in the $\text{Bernoulli} \left(p\right)$ model, hypothesis $(i)$ follows from the finite support of $\mu_p$. For the Cauchy model, we can again use the Frobenius matrix norm to see that $\mathbb{E}\left[\exp\left(t\ell\left(Y_{1}\right)\right)\right]=\int\left(2+\xi^{2}x^{2}\right)^{t/2}f(x)dx$ where $f(x)$ is the density for $\mbox{Cauchy}\left(0,1\right)$. By elementary computations, this integral is seen to be finite when $t<1$ and hence hypothesis $(i)$ is also satisfied for this model. Moreover, hypothesis $(ii)$ has already be verified for both models and hypothesis $(iii)$ follows from conditions $(ii)$ and $(iii)$ of Theorem \ref{furstenberg} which have already been verified. \begin{comment} For the Cauchy model, condition $(i)$ of Theorem \ref{lepageclt} can be verified by noticing that when using the Frobenius matrix norm, one arrives at $\mathbb{E}\left[\exp\left(t\ell\left(Y_{1}\right)\right)\right]=\int\left(2+\xi^{2}x^{2}\right)^{t/2}f(x)dx$ where $f(x)$ is the density for $\mbox{Cauchy}\left(0,1\right)$. By elementary computations, this integral is seen to be finite when $t<1$ and hence condition $(i)$ is satisfied. For $(ii)$, it is checked by way of an equivalent condition given in {[}3, Proposition II.4.3{]}. For $(iii)$, this follows by noticing that if $M=\left(\begin{array}{cc} \xi\epsilon & -1\\ 1 & 0 \end{array}\right)$ then $\left\Vert M\right\Vert =\left(2+\xi^{2}\epsilon^{2}\right)^{1/2}$ gets arbitrarily large for large $\epsilon$. Hence $\left\{ \left|\det M\right|^{-1/2}M:M\in G_{\mu_{\xi}}\right\} $ is not contained in a compact subgroup of $\mbox{GL}\left(2,\mathbb{R}\right)$ \end{comment} Thus for $0<p<1$ and $\xi\neq 0$, we know there exists $\sigma(p),\sigma(\xi)>0$ such that for any $\mathbf{x}\in\mathbb R^2\setminus\{\mathbf{0}\}$, \[ \frac1{\sqrt n}\Big(\log\|S_n\mathbf{x}\|-n\lambda(p)\Big)\textrm{ and }\frac1{\sqrt n}\Big(\log\|S_n\mathbf{x}\|-n\lambda(\xi)\Big) \] converge weakly as $n\to\infty$ to Gaussian random variables with mean $0$ and variance $\sigma^2(p)$ and $\sigma^2(\xi)$. Here the $S_n$ are products of matrices distributed according to the probability measures $\mu_p$ and $\mu_\xi$ given in Sections \ref{bernp} and \ref{xicauchy}, respectively. Motivated by these considerations and following the idea of Section \ref{sec:approx}, we can approximate $\sigma^2(p)$ and $\sigma^2(\xi)$ by computing the variance of \[ L_p:=\frac{1}{\sqrt{n}}\left(\sum_{i=1}^{n}\log\left\|U_i\right\|-n\lambda(p)\right)\text{ and }~ L_\xi:=\frac{1}{\sqrt{n}}\left(\sum_{i=1}^{n}\log\left\|U_i\right\|-n\lambda(\xi)\right) \] with $n$ large. Here, as in Section \ref{sec:approx}, the sequence $\{U_i\}_{i\geq 0}$ is constructed recursively by $U_0=\mathbf{x}$ and $U_i=Y_i\frac{U_{i-1}}{\|U_{i-1}\|}$ for some $\mathbf{x}\in\mathbb{R}^2$ with $\|\mathbf{x}\|=1$ and $\{Y_i\}_{i\ge1}$ an i.i.d. sequence drawn from $\mu_p$ or $\mu_\xi$ as appropriate. While we have an exact expression for $\lambda(\xi)$, we must settle for the approximation of $\lambda(p)$ obtained by simulation in Section \ref{sec:approx}. In what follows, we summarize the simulation procedure for $\sigma^2(p)$. The procedure for $\sigma^2(\xi)$ is practically identical. \begin{enumerate} \item Choose an interval $[a,b]$ as the range of $p$. Divide this interval into sub-intervals of length $k$ where $k$ divides $b-a$. Let $p$ be of the form $a+jk$ for $j=0,1,\dots,\frac{b-a}{k}$. \item Choose a unit vector $\mathbf{x}\in\mathbb{R}^2$. \item Simulate $L_p$ for each $p$ from Step 1 and store the result as a data vector of length $\frac{b-a}{k}+1$. \item Repeat Step 3 an $m$ number of times to obtain an $m\times\frac{b-a}{k}+1$ matrix, where the $j^{\text{th}}$ column contains all of the $L_p$ simulations corresponding to $p=a+(j-1)k$. \item Estimate $\mathrm{Var}\left(L_{a+(j-1)k}\right)$ by the sample variance of the $j^{\text{th}}$ column of the matrix. \end{enumerate} Note that in all of our simulations, we set $\mathbf{x}=\left(\frac{\sqrt{2}}{2},\frac{\sqrt{2}}{2}\right)$ in Step 2. We first approximate the variance for the $\text{Bernoulli} \left(p\right)$ model considered in Section \ref{bernp}. Trivially, we have that ${\sigma^2(0)=\sigma^2(1)=0}$. For $0<p<1$, we simulate $\mathrm{Var}\left(L_p\right)$ with $k=0.01$, $n=1000$, and $m=1\,000\,000$. We plot the resulting points in Figure \ref{fig:BernPVars} and remark that the graph exhibits distinct asymmetry with the maximum variance occurring around $p=0.56$. \begin{figure}[H] \centering \includegraphics[scale=0.6]{imgBernPVars.pdf} \caption{$k=0.01$, $n=1000$, $m=1\,000\,000$} \label{fig:BernPVars} \end{figure} For the Cauchy parameter model from Section \ref{xicauchy}, it is clear that $\sigma^2(0)=0$. For $\xi\neq 0$, we simulate $\mathrm{Var}\left(L_\xi\right)$ over both a large and small range of $\xi$. Figure \ref{fig:Cauchy20Vars} illustrates the results for $\xi\in[-20,20]$ with $k=0.25$. This is the same interval used to produce Figure \ref{fig:Cauchy20Lambda}. \begin{figure}[H] \centering \includegraphics[scale=0.5]{imgCauchy20Vars.pdf} \caption{$k=0.25$, $n=1000$, $m=5\,000\,000$} \label{fig:Cauchy20Vars} \end{figure} In Figure \ref{fig:CauchyZoomVars}, we plot $\mathrm{Var}\left(L_\xi\right)$ for $\xi\in[-1,1]$ with $k=0.01$ to give a much finer resolution of the graph around the origin. \begin{figure}[H] \centering \includegraphics[scale=0.5]{imgCauchyZoomVars.pdf} \caption{$k=0.01$, $n=1000$, $m=1\,000\,000$} \label{fig:CauchyZoomVars} \end{figure} \begin{acknowledgement} The authors are grateful for many helpful and motivating conversations with M. Gordina, L. Rogers and A. Teplyaev. We would also like to thank an anonymous referee, whose comments and suggestions greatly improved the exposition of the paper. \end{acknowledgement}

a2351601029ea2df5c44bbc7e00914571c432803

\section{Introduction} \label{sec:intro} The luminous supernova SN~2010jl interacting with a dense circumstellar matter (CSM) is a special case among SNe~IIn due to a large volume of observational data accumulated in different bands including X-ray \citep{Chandra2012,Ofek2014,Chandra2015}, optical \citep{Stoll2011,Zhang2012,Smith2012,Ofek2014,Fransson2014,Gall2014,Borish2015,Jencson2016}, infrared (IR) \citep{Andrews2011,Fransson2014,Gall2014,Borish2015,Dwek2017}, and radio \citep{Chandra2015}. This provide us with an opportunity to get deeper insight into complicated phenomena accompanied the ejecta interaction with a dense CSM. Of particular interest is the suspected dust formation in the postshock cool dense shell (CDS) indicated by the late time hot IR excess \citep{Maeda2013,Gall2014} combined with the line blueshift \citep{Smith2012}. The estimated amount of the dust is of $\approx 10^{-3}\,M_{\odot}$ after about 500 d \citep{Maeda2013,Gall2014,Sarangi2018}. The possibility of the dust formation in the CDS of SNe~IIn was discussed earlier for SN~1998S \citep{Pozzo2004}, SN~2006jc \citep{Smith2008,Mattila2008,Chugai2009}, SN~2005ip \citep{Smith2009}, SN~2007od \citep{Andrews2010}. Yet there are open questions that may cast doubts on this conjecture. Among them is the issue of so called blueshifted Lorentz-like profiles (narrow core plus broad wings). The point is that, if the dusty line-emitting shell is composed of the CDS material only, one would expect then to find a blueshifted broad boxy line profile because in this case the velocity dispersion of the line-emitting material is small compared to the expansion velocity. This issue was emphasised by \citet{Fransson2014} to argue in favour of the CS dust. The concept of the external dust however leaves blueshifted Lorentz-like profiles unexplained. Since the conjecture on the dusty CDS and the line blueshift seems to be closely linked the origin of the Lorentz-like profile requires an explanation before a confident conclusion on the dust location could be made. Furthermore, accepting the idea of the dust formation in the CDS we have not so far answers to a straightforward questions concerning the site and velocity distribution of the line-emitting gas and the dust distribution. It is noteworthy that, apart from the CDS, the dust in SNe~IIn might form also in unshocked ejecta in the same way as it does in SNe~IIP, e.g., SN~1987A \citep{Lucy1989} and SN~1999em \citep{Elmhamdi2003}. Note, the dust formation in the unshocked ejecta of SN~IIn has been already proposed in a context of an infrared excess in the SN~2005ip \citep{Fox2010}. This possibility should be considered as highly probable for SN~2010jl as well. Below I address two principal questions: (i) what is the origin of blueshifted lines in spectra of SN~2010jl; (ii) whether the late time SN radius is consistent with the radius of the IR-emitting dust shell. The answer to these questions will permit us to present at least a qualitative model that accounts for blueshifted line profiles and late IR emission in unique picture lacking at the moment. I start with an explantion of the origin of Lorentz profile and the blueshift modelling (Section \ref{sec:specmod}). This is followed by the modelling of the SN/CSM interaction that provides us with an important tool for the probing the relation between the SN radius and the radius of the IR-emitting shell (Section \ref{sec:dmod}). In the Discussion section the clumpy dust distribution is discussed and the estimate of the dust amount is given. Following \citet{Fransson2014} we adopt the explosion date 2010 October 9 (JD 2455479). \section{Spectral modelling} \label{sec:specmod} \subsection{Preliminaries: Why Lorentz profile?} \label{sec:preview} The interaction of spherically-symmetric freely expanding ($v = r/t$) SN ejecta with a spherically-symmetric CSM occurs via the formation of two shock waves: forward (radius $R_f$) and reverse shock (radius $R_r$) with a shocked SN and CS matter swept-up in a thin dense shell at the contact surface (radius $R_0$). (Henceforth radii are measured in units of $R_0 = 1$, if otherwise not stated explicitly). In an adiabatic case for the typical density distribution of ejecta ($\rho \sim v^{-8}$) and CSM ($\rho \sim r^{-2}$) the shock radii ratios are $R_r/R_0 = 0.98$ and $R_f/R_0 = 1.27$ \citep{Chevalier1982b}. In strongly radiative regime the postshock layers get narrower, so one can adopt in this case $R_r = R_0$. The X-rays from both shocks excite cold unshocked SN ejecta and CSM thus giving rise to optical broad lines and narrow lines respectively. In a realistic situation the picture grow significantly more complicated: the flow between forward and reverse shock turns out three-dimensional due to Rayleigh-Taylor (RT) instability of the thin shell \citep{Chevalier1982a,CheBlo1995,BloEllis2001} and because of a lumpiness of the CSM. Therefore, in reality additional line-emitting sites related to the CDS and shocked CS clouds can appear between forward and reverse shock. Due to the non-trivial velocity spectrum and highly complicated radiation transfer effects the optical radiation spectrum of SNe~IIn cannot be reliably predicted so far. Emission lines of SNe~IIn at the early stage ($t < 1$ yr) usually (but not always) reveal three components: the narrow component (NC), broad (BC), and intermediate (IC), --- all originally have been recognized in the SN~1988Z spectrum by \citet{Filippenko1991}. These components can be identified (Figure \ref{fig:cart}) with the CSM, unshocked SN ejecta in combination with the CDS, and shocked CS clouds, respectively, \citep{Chugai1994}. In SN~2010jl the distinction between BC and IC is not so apparent as in SN~1988Z: hydrogen lines, e.g. H$\alpha$, of SN~2010jl are rather smooth and indeed reminds a "Lorentz" profile \citep{Fransson2014}. We will use this term, keeping in mind that it has nothing to do with the Lorentz distribution of frequencies of the damping oscillator radiation. Fortunately, in He\,I 1.083\,$\mu$m line the component ratio of BC/IC is large, so both components are easily distinguished \citep{Borish2015}, which keeps safe canonical three-component structure of emission lines for SN~2010jl. It should be emphasised that late time Lorentz profiles in SN~2010jl have essentially different origin than emission lines with Lorentz profile in early spectra of some SNe~IIn, identified with emission of the ionized dense preshock CSM, where these lines get broadened by the Thomson scattering; the phenomenon recognized for SN~1998S \citep{Chugai2001}. This type of the profile cannot present at late epochs ($ t \gtrsim 100$ d) since the required Thomson optical depth of a preshock CSM must be large ($\tau_{\mbox{\scriptsize T}} \gtrsim 2$) that is unattainable at that late time. The Lorentz profile in the late time SNe~2010jl can be interpreted following a suggestion invoked for SN~2006jc (type IIn), where the emission lines have been attributed to shocked fragmented CS clouds \citep{Chugai2009}. A typical value of the cloud shock in SNe~IIn is $v_c \sim 10^3$\,km\,s$^{-1}$\ \citep[e.g., SN~1988Z,][]{Chugai1994}, which for the postshock intercloud velocity $v_{ic} \approx 6000$\,km\,s$^{-1}$\ implies the CS cloud density contrast with respect to the intercloud density $\chi \approx (v_{ic}/v_c)^2 \sim$ 30-40. The shocked fragmented CS clouds resides between the forward shock and the CDS (Figure \ref{fig:cart}) while their velocities range between the initial cloud shock velocity ($v_c$) and the final velocity of fragments $v_{max} \approx (0.75...~0.9)v_{ic}$ for $10 <\chi < 100$ \citep{Klein1994}. Note, the velocity the postshock intecloud gas coincides with the CDS velocity $v_{cds}$. The cloud fragmentation and fragments acceleration by the incident shock is well demonstrated by laser experiments and 3D-hydrodynamic simulations \cite{Klein2003} with the relevant physics well understood earlier \citep{Klein1994}. The case of SN~2006jc indicates that the Lorentz profiles can be reproduced with the velocity distribution of the line emissivity $j(v) \propto (v_{max} - v)$ \citep{Chugai2009}. The following toy model illustrates, how such a velocity spectrum could arise. Consider a steady-state flow of CS clouds into the forward shock with the rate $G$ (clouds/s). The cloud in the forward shock experiences crushing by the radiative shock followed by the fragmentation, acceleration of fragments, and fragments mixing with ambient intercloud hot gas \citep{Klein1994}. Let cloud life span in the forward shock be $T$ and the cloud survival probability at the age $\tau$ be the linear function of the age $p(\tau) \propto (1 - \tau/T)$. Assuming that the fragmented cloud accelerates linearly as $v = v_c +a\tau$, which is close to 2D-simulation results \citep{Klein1994}, and taking into account that the clouds age distribution is $dN/d\tau = Gp(\tau)$ one finds the velocity distribution of shocked clouds $dN/dv = (dN/d\tau)(d\tau/dv) \propto (v_{max}-v)$, where $v_{max} = v_c + aT$. With the constant specific emissivity (per gramm) we thus come to the required velocity distribution of the line emissivity. This is only crude illustration and the emissivity velocity distribution may difer from this simple law. \begin{figure} \includegraphics[trim=0 0 0 0,width=1\columnwidth]{fig1.eps} \caption{ Schematic diagram of the SN/CSM interaction showing source location for the narrow, broad and intermediate components involved in the overall line emission. The layer of cold CDS material is partially mixed with the layer of shocked CS clouds ({\em grey}); only a fragment of the latter is shown for clarity. } \label{fig:cart} \end{figure} \begin{figure} \includegraphics[trim=40 160 0 -20,width=0.95\columnwidth]{fig2.eps} \caption{ Effects of the parameters variation on the H$\alpha$ model profile. Panels from {\bf a} to {\bf f} show the fiducial model mod0 ({\em black}) and models by order from mod1 to mod6 with changed parameter value ({\em red}) (Table \ref{fig:demo}). In the panel {\bf a} the combined contribution of broad components of the CDS and ejecta in the mod0 is shown for clarity ({\em blue}). The panel {\bf f} demonstrates effect of a large contribution of the CDS and unshocked ejecta emission ({\em blue}) while the intermediate component ({\em green}) remains unchanged. } \label{fig:demo} \end{figure} \begin{figure*} \includegraphics[trim=60 140 0 -10, width=0.9\textwidth]{fig3.eps} \caption{ The He\,I 1.083 $\mu$m (panel {\bf a}) and Pa$\beta$ (panel {\bf b}) in SN~2010jl on day 178 ({\em magenta}) compared to models. The excess in the red wing of the He\,I 1.083 $\mu$m is due to Pa$\gamma$. Shown also contributions of the intermediate component ({\it green} line) and combined broad components of CDS and unshocked ejecta ({\it red}). } \label{fig:he} \end{figure*} \begin{table} \centering \caption[]{Parameters of demonstration models.} \label{tab:hademo} \begin{tabular}{l|c|c|c|c|c|c} \hline Model& $\tau_{cds}$ & $\tau_{sn}$ & $\omega_{sn}$ & $R_2$ & $f_{cds}$ & $f_{sn}$\\ \hline mod0 & 0.13 & 1 & 0.1 & 1.1 & 0.05 & 0.08\\ mod1 & 0 & 0 & 0.1 & 1.1 & 0.05 & 0.08\\ mod2 & 0 & 1 & 0.1 & 1.1 & 0.05 & 0.08 \\ mod3 & 0.13 & 0 & 0.1 & 1.1 & 0.05 & 0.08 \\ mod4 & 1.13 & 1 & 1 & 1.1 & 0.05 & 0.08\\ mod5 & 0.13 & 1 & 0.1 & 1.2 & 0.05 & 0.08 \\ mod6 & 0.13 & 1 & 0.1 & 1.1 & 0.2 & 0.2 \\ \hline \end{tabular} \label{tab:demo} \end{table} \subsection{Line profile} \label{sec:prof} \subsubsection{Model} \label{sec:mod} A minimal model that captures major structure elements involved in the line profile formation of SN~2010jl includes three components: (i) freely expanding SN ejecta ($v = r/t$ , $r < R_0 = 1$); (ii) perturbed CDS ($R_0 < r < R_1 = 1.1$) with the average velocity $v_{cds}$ and a random component in the range of $\delta v \sim 0.1v$ \citep[e.g.,][]{BloEllis2001}; (iii) line-emitting fragments of shocked CS clouds ($R_0 < r < R_2$) macroscopically mixed with the absorbing CDS material. This component is identified with the line intermediate component, while former two are responsible for the broad component. The narrow component originated in the preshock CSM is not considered in our model. I neglect the hot thin layer in the reverse shock and set the radius of the unshocked SN ejecta equal to the CDS radius $R_0$. In line with arguments of Section \ref{sec:preview} the emissivity velocity distribution for the shocked CS clouds responsible for the intermediate component is approximated by the function \begin{equation} g(v) = \left\{\begin{array}{ccc} \dfrac{v-v_{min}}{v_c-v_{min}}\,, & \quad \textrm{if $v_{min}<v<v_c$} \\ \left[\dfrac{v_{max}-v}{v_{max}-v_c}\right]^q\,, & \quad \textrm{if $v_c<v<v_{max}$} \\ 0\,, & \quad \textrm{otherwise,} \end{array}\right. \label{eq:vpdf} \end{equation} where the index $q$ is a free parameter close to unity. The function $g(v)$ has a maximum at $ v = v_c$ with the linear rise in the range $v_{c,min} < v < v_c$ that makes allowance for the fact that CS clouds are not identical; we adopt $v_{min} = 0.9v_c$. Note that in the case of $v_{min} = v_c$ the line profile would be flat in the range of radial velocities $\pm v_c$ which is not seen in observed profiles. I preliminary explored two radial distribution of the emissivity of shocked clouds: (i) uniform in the range $R_0 < r < R_2$ (i.e., no radius/velocity correlation) and (ii) linear radius/velocity correlation, viz. $r=R_0 + (R_2-R_0)(v-v_{min})/(v_{max}-v_{min})$. The resulting profiles in these cases are indistiguishable, except for the late spectrum on day 804 when the correlation produces slightly better fit. This choice therefore is applied for all the epochs. The broad component is emitted by the CDS and unshocked ejecta with relative fractions $f_{cds}$ and $f_{sn}$, so the intermediate component fraction is $1 - f_{cds} - f_{sn}$. The CDS emissivity is distributed homogeneously in the velocity randomly distributed in the range of $v_{cds} < v < 1.1v_{cds}$. The emissivity distribution in the SN ejecta is described by the "broken" power law $j(v) = (v/v_b)^{k_1}/[1 + (v/v_b)^{k_2}]$ with the velocity turnover $v_b$ and indices $k_1 \approx 2$ and $k_2 \approx 9$ inferred from broad component of the He\,I 1.083 $\mu$m line. The turnover of the $j(v)$ function is expected because the ionizing radiation from the shocks is strongly absorbed in outer layers of ejecta. The optical line emitted by ejecta can be scattered by electrons and dust (if any) and absorbed by the hydrogen and dust. The distribution of absorbers in SN ejecta and CDS is assumed to be homogeneous along the radius. The absorption is specified by the optical depth of the CDS and ejecta ($\tau_{cds}$ and $\tau_{sn}$), and albedo ($\omega$), viz. the ratio of the scattering to extinction coefficient. To reduce the number of free parameters the albedo is set to be the same in the CDS and ejecta, which may not be the case. However this assumption does not affect significantly results because, as we will see, the ejecta dominates in the observable scattering effect. In search for the optimal optical depth and albedo I ignore a possible lumpiness of the dust distribution. However, since effects of clumpy dust distribution will be discussed below, it is instructive to make some remarks concerning this issue. The effective optical depth in the clumpy medium is introduced \citet{Ambar1947} as $\tau_{eff} = \tau_{oc}\langle 1 - \exp{(-\tau)}\rangle$, where $\tau_{oc}$ is the number of clouds along the ray, or the occultation optical depth, $\tau$ is the random cloud optical depth, and the averaging should be performed over cloud cross section and cloud ensemble. In the toy model of homogeneous spherical clouds with of the radius $b$ and the optical radius $\tau =kb$ homogeneously distributed in a sphere of a radius $R$ the average is calculated analytically \citep{Utrobin2015} and reads $\langle...\rangle = (4\tau/3)p(\tau)$, where $p(\tau)$ is the escape probability of photon emitted in the homogeneous sphere of the optical radius $\tau$ \citep[see][]{Osterbrock1989}. In the limit of $\tau \ll 1$ retaining only linear terms we get $\tau_{eff} = (4/3)\tau\tau_{oc}$ that can be easily transformed into $\tau_{eff} = fkR$, where $f$ is the filling factor of clouds. In the oposite case of $\tau \gg 1$ we obtain $\tau_{eff} = \tau_{oc}$ in line with the euristic definition of the optical depth for the ensemble of clouds treated as macroatoms. The optical depth recovered in the homogeneous model in clumpy case should be identified with the effective optical depth $\tau_{eff}$ and in the case of optically thick clouds this will be simply $\tau_{oc}$. The albedo of the homogeneous model ($\omega$) for the clumpy medium should be associated with the albedo of a dusty cloud, which is expected to be lower than single scattering albedo of a grain ($\omega_d$). The relation between $\omega_d$ and $\omega$ can be illustrated by Monte Carlo simulations for the spherical dusty cloud and spherical phase function. In the limit of a large cloud optical radius ($\tau > 10$) for the set of the single scattering albedo (0.1, 0.3, 0.6, 0.9) the asymptotic values of $\omega$ turn out to be 0.023, 0.08, 0.2, and 0.5, respectively. If, e.g., one finds in the homogeneous model $\omega = 0.1$, then in the clumpy case with the large optical depth of the dusty cloud one expects the single scattering albedo to be $\omega_d \approx 0.35$. The apparent drawback of the outlined model is a large number of free parameters even in a homogeneous case. The list includes $R_2$, $v_{sn}$, $v_{cds}$, $v_c$, $v_{max}$, $q$, $\tau_{cds}$, $\tau_{sn}$, $\omega$, $f_{cds}$ and $f_{sn}$, among which four parameters $\tau_{cds}$, $\tau_{sn}$, $\omega$, and $v_{sn}$ are of prime interest for us. The rest of parameters are constrained with different precisions. Particularly, at late time the contribution of the CDS is poorly constrained, so neither $f_{cds}$ and $v_{cds}$ can reliably inferred. Fortunately this does not affect the determination of the principal four parameters. Although an effect of each parameter can be controlled, the problem of the unique set of parameters that defines the best fit model in multi-dimensional parameter space would be unproductive to solve rigorously, because the model is rather crude and ignores a possible deviation from the spherical symmetry that is plausible given a significant intrinsic polarization found in type IIn SN~1997eg \citep{Hoffman2008}, in some aspects reminding SN~2010jl. \subsubsection{Results} In a search for the optimal model I explored an extended volume of parameter space. Fortunately effects of most parameters are rather obvious, although in some cases the effect is less trivial. To provide insight into the role of principal paprameters I present a set of simulated Lorentz-like profiles (Table \ref{tab:demo}, Fig. \ref{fig:demo}). Each panel shows a template model (mod0) that corresponds to the H$\alpha$ model on day 510 and a model with the altered parameter value. The template model is dominated by the emission of shocked CS clouds (intermediate component); ejecta and CDS contribute together 10\% (panel {\bf a}). The unabsorbed line with zero optical depth for both CDS and SN ejecta (panel {\bf a}) shows symmetriic profile with roundish top that is due to the low velocity increasing part of the velocity spectrum $g(v)$ of shocked CS clouds (cf. Section \ref{sec:mod}). The case of a zero CDS absorption (panel {\bf b}) demonstrates that the ejecta absorption is not sufficient to produce required blueshift of the line maximum. On the other hand, the absence of the ejecta absorption (panel {\bf c}) shows that while the CDS absorption is able to describe the blueshift of the maximum, the profile in the red is unlike what is required. Note that the dip in the red part is a limb effect for the line arising in a narrow shell ($\Delta R/R < 1$) with the continuum absorption/scattering \citep[cf.][]{Chugai1991}. The case of $\omega = 1$ corresponding to the Thomson scattering with the thermal broadening taken inro account results in the pronounced red wing (panel {\bf d}), a natural outcome of the photon scattering in the expanding envelope \citep{Auer1972}. This effect is also present in the case when line photons are scattered off a dust with albedo $\omega \sim 0.5$ in SN ejecta \citep{Lucy1989,Bevan2016}. The need for the macroscopic mixing between the layer of shocked CS clouds and the dusty CDS material is demonstrated by model with the outward shift of the line-emitting layer by $\Delta r = 0.1R_0$ (panel {\bf e}): the model shows lower maximum blueshift. Interestingly, the profile in the model mod5 almost exactly repeats the profile in the model mod2. This is not unexpected since in both cases line-emitting shell lies above the absorbing shell. This similarity however cannot be considered as the parameter degeneracy, because in the model with correct blueshift these similar cases cannot appear. The effect of large contribution of the broad components related to CDS and SN ejecta indicates how the broad components can be constrained (panel {\bf f}). The value of the index $q$ affects the intermediate component in a sense that $q \gg 1$ makes profile narrower and more gaussian-like, while $q \ll 1$ makes profile broader and closer to the triangular shape. We now consider an interesting case of the He\,I 1.083 $\mu$m line. At the early stage ($t < 400$ d) He\,I~1.083 $\mu$m and He\,I~2.06 $\mu$m lines show a pronounced broad component unlike smooth hydrogen line profiles \citep{Borish2015}. A reasonable suggestion is that this component is caused by the emission of the SN ejecta \citep{Borish2015}. Figure \ref{fig:he} demonstrates satisfactory fits He\,I 1.083 ${\mu}$m and Pa$\beta$ on day 178 in the models with the same parameters except for the ejecta emission fraction, $f_{sn} = 0.68$ for He\,I 1.083 ${\mu}$m compared to 0.3 for for Pa$\beta$, and the index, $q = 1.5$ for He\,I 1.083 ${\mu}$m compared to $q = 0.9$ for Pa$\beta$. The latter difference presumably reflects different excitation conditions of hydrogen and helium in the intermediate component. Both helium and hydrogen line shows blueshift that is expressed in the inferred values of optical depths $\tau_{cds} = 0.07$, $\tau_{sn} = 0.58$, and albedo $\omega = 0.6$. Since at this epoch the dust formation in the CDS and SN ejecta is unlikely, the extiction is probaly related to the hydrogen photoabsorption and Thomson scattering. It should be emphasised that the blueshift of the broad component of He\,I 1.083 ${\mu}$m is produced entirely by the absorption in the unshocked ejecta. It is not clear whether the different fraction of the broad component for helium and hydrogen lines is related to the different helium abundance in SN ejecta and CSM, or different excitation conditions in the line-emitting regions. Discussion of this issue would require highly complicated modelling, which is currently beyond reach. The recovered parameters of the He\,I ejecta component, viz. $v_b = 5500$\,km\,s$^{-1}$, $k_1 = 2$, and $k_2 = 9$ (cf. Equation \ref{eq:vpdf}) are adopted henceforth for hydrogen lines as well because in hydrogen lines the ejecta component parameters cannot be reliably constrained. \begin{table*} \centering \caption[]{Parameters of H$\alpha$ models.} \label{tab:hamod} \begin{tabular}{l|c|c|c|c|c|c|c|c|c|c|c} \hline Model & $v_{sn}$ & $v_{cds}$ & $v_{c}$ & $v_{max}$ & $q$ & $R_2$ & $\tau_{cds}$ & $\tau_{sn}$ & $\omega$ & $f_{cds}$ & $f_{sn}$ \\ \hline & \multicolumn{4}{c} {km\,s$^{-1}$} & & & & & & & \\ \hline m172 & 8500 & 5600 & 900 & 5300 & 0.9 & 1.13 & 0.07 & 0.58 & 0.6 & 0.05 & 0.3 \\ m400 & 6100 & (4700) & 1100 & 4400 & 0.6 & 1.1 & 0.14 & 1.15 & 0.2 & 0.05 & 0.12 \\ m510 & 5700 & (4200) & 850 & 3500 & 0.4 & 1.1 & 0.12 & 0.99 & 0.08 & 0.05 & 0.08 \\ m804 & (5000) & (3200) & 850 & 3000 & 1.2 & 1.1 & 0.18 & 0.96 & 0.1 & (0.02) & (0.02) \\ \hline \end{tabular} \label{tab:param} \end{table*} The H$\alpha$ line is modelled for spectra taken on 2011 March 31 (172 days), 2011 Nov. 13 (400 days), 2012 Feb. 28 (510 days), and 2012 Dec. 21 (804 days). The data are retrieved from the Weizmann supernova data repository {\em http://wiserep.weizmann.ac.il} \citep{Yaron2012}. Models overplotted on the observed spectra are displayed in Fig. \ref{fig:hamod} with parameters given in the Table \ref{tab:hamod}. Columns one by one contain: the model name (the number is the age in days), terminal velocity of unshocked SN ejecta ($v_{sn}$), CDS velocity ($v_{cds}$), velocity of CS cloud shock ($v_{c}$), maximal velocity of the CS shocked cloud material ($v_{max}$), index $q$ of the emissivity of shocked CS clouds, outer radius of the line-emitting layer of shocked CS clouds ($R_2$); continuum optical depths of the CDS ($\tau_{cds}$) and unshocked SN ejecta ($\tau_{sn}$); albedo ($\omega$) and the relative contribution of the CDS ($f_{cds}$) and unshocked SN ($f_{sn}$) into the H$\alpha$ luminosity. The values in parentheses have large uncertainties. For all the considered epochs the CDS contribution is dominant. The contribution of the broad components both ejecta and the CDS is large only in the spectrum on day 172 and gets small with the age. Remarkably, the CDS optical depth at the late stage (510 and 804 d), when presumably the dust have been formed, is rather small (0.13 and 0.15 respectively), whereas SN ejecta turns out essentially more opaque ($\tau_{sn} = 1$). This indicates that apart from the CDS the dust also forms in the unshocked ejecta and possibly in larger amount. Note that at late time the lower albedo is preferred. This fact is consistent with the dust to be the main source of opacity, since the hydrogen density and hydrogen excitation in the SN ejecta decreses with time and therefore the hydrogen absorption and Thomson scattering gets inefficient. The estimated parameter errors that reflect precision of model fits lie in the range 10-20\% except for those inicated by parentheses in Table \ref{tab:param}. The errors however reflect only the adopted model. In fact, the radial distribution of absorbers and emitters in ejecta and velocity and radial distributions of shocked CS clouds might be more complicated; moreover, the anisotropy is not rulled out. All these can result in some change of values of four principal parameters ($\tau_{cds}$, $\tau_{sn}$, $\omega$, and $v_{sn}$). With this remark, the proposed spectral model should be considered as a reasonable possibility in the absence of alternative models. \begin{figure*} \includegraphics[trim=60 170 0 -20,width=0.95\textwidth]{fig4.eps} \caption{ H$\alpha$ line in SN~2010jl ({\em magenta}) compared to the model profile at four epochs ({\it black} line). The blushifted profiles are well reproduced for adopted parameters (Table \ref{tab:param}). The excess on day 172 around zero velocity is the narrow component, while on day 804 the excess is due to superimposed H\,II region. Shown are also contributions of the dominant intermediate component ({\em green}) and combined broad components of CDS and unshocked ejecta ({\it red}). } \label{fig:hamod} \end{figure*} \section{CS interaction and dusty shell} \label{sec:dmod} The interaction of the SN ejecta with the CSM is analysed using the hydrodynamic model in a thin shell approximation \citep{Giuliani1982,Chevalier1982a}. The central to the model is the equation of motion of the swept up thin shell driven by the dynamic pressure of SN ejecta $\rho_{sn}(R_s/t - v_s)^2$ and dynamic pressure of CSM $\rho_{cs}(v_s - v_w)^2$, where $R_s$ and $v_s$ are the radius and velocity of the thin shell, $v_w$ is the wind velocity taken to be 100\,km\,s$^{-1}$\ \citep{Fransson2014}. The model has been described in details earlier \citep{Chugai2001}, here we recap only key points. We do not distinguish between the swept up shells formed in the reverse and forward shocks: both are presumably merged into a single shell. The optical luminosity of the reverse and forward shocks at the age $t$ are calculated as the shock kinetic luminosities $0.5\rho_{sn}(R_s/t - v_s)^3$ and $0.5\rho_{cs}(v_s - v_w)^3$ multiplied by the radiation efficiency $\eta = t/(t+t_c)$ of a relevant shock, where $t_c$ is the cooling time. The radiation escape from both shocks is assumed to be instant, i.e., weak diffusion effects are ignored. The model does not include an additional luminosity related to the internal energy of an explosion. The cooling time is calculated neglecting the lumpiness of the CSM and adopting postshock density to be equal to four-fold preshock density characteristic of a strong adiabatic shock with the adiabatic index $\gamma=5/3$. This approximation describes the energetically significant initial stage of the radiative shock cooling. The reverse shock in the considered case is always strongly radiative, while the forward shock is partially radiative with $\eta \sim 0.5$ between days 200 and 900. The lumpiness of the CSM could modify the cooling time of the forward shock. The effect however is not easy to estimate without detailed 3D-modelling. At first glance the presence of the dense shocked CS clouds should accelerate the radiative cooling. However, in a collision of forward shock rarefied flow and dense CS cloud most of the dissipated kinetic luminosity is deposited into the rarefied flow which results in the slowdown of the cooling process thus compensating effect of the dense shocked clouds. We therefore do not expect a significant modification of the bolometric light curve. As to the CDS dynamics, it is not affected by the cooling time modification. The adopted density distribution of SN ejecta is \mbox{$\rho = \rho_0/(1 + (v/v_0)^8)$} with $\rho_0$ and $v_0$ determined by the ejecta mass and kinetic energy. The result is not very sensitive to the ejecta mass, nevertheless we consider two cases of ejecta mass: $M_{sn} = 8\,M_{\odot}$, the value used earlier for SN~2010jl \citep{Chandra2015}, and $M_{sn} = 30\,M_{\odot}$. The remaining parameters are fixed by fitting to the bolometric light curve, CDS, ejecta velocities (Table \ref{tab:param}), and the CSM column density. The CSM density is set by a broken power low $\rho \propto r^{-s}$ with $s = 0$ in the inner region $r < 3.1\times10^{15}$ cm, $s \sim 1.6...~2$ in the intermediate zone $r < 4\times10^{16}$ cm, and steep drop ($s = 11$) in the outer region. The model with the ejecta mass of $8\,M_{\odot}$ and kinetic energy of $E = 3.9\times10^{51}$ erg describes the observational bolometric luminosity, ejecta velocities, and CSM column density (Fig. \ref{fig:dyn1}). Note that the observational bolometric light curve \citep{Fransson2014} is modified by the inclusion of the late ($t > 460$ d) infrared luminosity reported by \citet{Fransson2014}. The total mass of the CSM turns out to be $M_{cs} = 3.9~M_{\odot}$. The model with $M_{sn} = 30\,M_{\odot}$ and $E = 6.9\times10^{51}$ erg (Fig. \ref{fig:dyn2}) describes data with the same CSM mass $M_{cs} = 3.8~M_{\odot}$. Both models fit the data well and neither option is preferred. The main outcome of the modelling is the fact that in both models the CDS/SN radius at late time $t > 460$ d (Fig. \ref{fig:dyn1},~\ref{fig:dyn2}) coincides with the radius of the dusty shell reported by \citet{Fransson2014}. This strongly suggests that the dust of CDS and/or unshocked ejecta is responsible for the late IR luminosity. \begin{figure} \includegraphics[trim=70 100 0 -10,width=\columnwidth]{fig5.eps} \caption{ The SN/CSM interaction model for the ejecta mass of $8\,M_{\odot}$ with observational data. Panel {\bf a} shows the bolometric light curve compared to observational data from \citet{Fransson2014} ({\em asterisks}; IR luminosity summed with the optical is shown by {\em triangles}). Inset shows the boundary velocity of the unshocked ejecta ({\it thin} line) compared to data recovered from the H$\alpha$ and the CDS velocity ({\it thick} line) Panel {\bf b} shows the evolution of the CSM column density ahead of the forward shock compared to data ({\em squares}) \citep{Chandra2015}. Inset shows the CDS radius compared to the radius of the IR-emitting shell ({\em diamonds}) \citep{Fransson2014}. } \label{fig:dyn1} \end{figure} \begin{figure} \includegraphics[trim=70 100 0 -10,width=\columnwidth]{fig6.eps} \caption{ The same as Fig. \ref{fig:dyn1} but for the ejecta mass of $30\,M_{\odot}$ } \label{fig:dyn2} \end{figure} \section{Discussion} \label{sec:disc} The aim of the paper has been to test a conjecture that in the late time SN~2010jl the dust forms in the CDS. To this end I propose a model explaining the Lorentz-like line profile as an outcome of hydrogen line emission from from shocked CS clumps in the forward shock. It is found that the macroscopic mixing of CS cloud fragments with the absorbing CDS and a significant absorption in the unshocked ejecta accounts for the observed blueshift. Noteworthy that the blueshift of hydrogen and He\,1.083 $\mu$m lines at the early stage (< 200 d) may not be related with the dust absorption at all; the opacity is likely maintained by the excited hydrogen. The remarkable result suggested by the line profile modelling is that the absorption in the CDS is not sufficient: a significant absorption in the unshocked SN ejecta is needed to account for the whole blueshifted line profile. Moreover, the required optical depth of ejecta is 5-7 times larger compared to the CDS at the epoch 500-800 days. This indicates that the dust should form not only in the CDS but in the unshocked ejecta as well. I found that the SN/CSM interaction model tuned by the fit to the bolometric light curve, ejecta velocities recovered from H$\alpha$, and the CSM column density \citep{Chandra2015} predicts the outer SN radius at late time ($t > 460$\,d) consistent with the radius of the dusty IR-emitting shell recovered by \citet{Fransson2014}. The latter fact combined with the line blueshift analysis strongly suggests that at late time (> 460 d) the dust indeed forms both in the CDS and unshocked ejecta. Thus the conjecture of the dust formation in the CDS \citep{Maeda2013,Gall2014} is only partially true: a significant amount of the dust should form in the unshocked ejecta. Moreover, the low covering factor of the CDS indicated by the small CDS optical depth $\sim 0.2$ at 510\,d and 804\,d suggests that the effective radius of the IR-emitting CDS photosphere turns out to be twice as small compared to the reported radius of the dusty shell \citep {Fransson2014}. This means that the late IR emission is likely primarily related to the more opaque dusty SN ejecta. The radius of the late dust-emitting shell have been obtained \citep{Fransson2014} using black body approximation, which suggests that the IR-emitting ejecta should be optically thick in the near IR band, e.g., at 3\,$\mu$m the optical depth must be $ > 1$. The direct consequence of that is the large optical depth at the H$\alpha$ ($\tau > 10$) assuming the absorption efficiency $Q_a \propto \lambda^{-1.5}$. This optical depth is by an order of magnitude larger than the value inferred from the H$\alpha$ profile between days 500 and 800 (Table \ref{tab:param}). The tension can be resolved assuming that the dust in ejecta is locked in opaque clumps, likewise in the model proposed for the dusty zone of SN~1987A ejecta \citep{Lucy1991}. Remarkably, the dust distribution in opically thick clumps for SN~2010jl has been already proposed by \citet{Maeda2013}. In the case of clumpy dust distribution in unshocked SN ejecta the optical depth recovered from the blueshift assuming homogeneous model should be identified with the occultation optical depth $\tau_{oc}$, i.e., the number of clouds intersecting by the ejecta radius $R$. One can estimate the amount of dust in SN ejecta adopting a simple model of opaque spherical dusty blobs of a radius $b$ homogeneously distributed in SN ejecta. The occultation optical depth is then $\tau_{oc} = \pi b^2n_cR = (3/4)(b/R)^2N_c$, where $n_c = 3N_c/(4\pi R^3)$ is the number density of clouds and $N_c$ is the total number of clouds. The optical radius of the individual cloud is $\tau_c = n_g\langle \sigma_a \rangle b$, where $n_g$ is number density of grains, and $\langle \sigma_a \rangle$ is the absorption cross section of a grain averaged over the grain ensemble assuming grain radii distribution $dn/da \propto a^{-3.5}$ ($a_{min}=3\times10^{-7}$\,cm, $a_{max}=3\times10^{-5}$\,cm) and using absorption efficiency of the carbon and silicate dust according to \citet{Draine1984}. The total amount of the dust in SN ejecta is $M_d =(4\pi/3)b^3n_c\langle m_g\rangle N_c$, where $\langle m_g\rangle$ is the grain mass avaraged over the grain ensemble assuming the grain density 3\,g\,cm$^{-3}$. Expressing $N_c$ via $\tau_{oc}$ and $b$ via $\tau_c$ the dust mass thus reads \begin{equation} M_d = (16\pi/9)R^2\tau_c\tau_{oc}\langle m_g\rangle/\langle \sigma_a \rangle\,. \label{eq:mdust} \end{equation} Inserting $\tau_{oc} = 1$, the value derived from H$\alpha$ on day 804, and the ejecta radius $R = 3.8\times10^{16}$ cm obtained in the interaction model we infer from equation (\ref{eq:mdust}) the dust mass of $7.7\times10^{-4}\tau_b\,M_{\odot}$ for the graphite and $3.9\times10^{-3}\tau_b\,M_{\odot}$ for silicate dust. For optically thick blobs with $\tau_b = 2$ at 3 $\mu$m the amount of dust in ejecta is $8\times10^{-3}\,M_{\odot}$ for the silicate and $1.5\times10^{-3}\,M_{\odot}$ for the graphite dust. The conservative estimate of the dust amount is therefore $> 10^{-3}\,M_{\odot}$. In the IR data reported by \citet{Fransson2014} a signature of the hot dust becomes apparent only on day 465. In this regard, the blueshift at the early epoch (e.g. 172 d) is unlikely related to the dust, rather it is due to the hydrogen opacity. To illustrate the possibility, consider the situation on day 172. Assuming the excitation temperature equal to the spectrophotometric estimate of 7500\,K \citep[cf.][]{Fransson2014} and the CDS mass of $\approx 2\,M_{\odot}$, according to the interaction model (Section \ref{sec:dmod}), the optical depth due to the hydrogen continuum absorption at 6500 \AA\ assuming Saha-Boltzmann population distribution turns out to be $\sim 0.5$, sufficient to account for the CDS optical depth in the model m172. This consideration emphasises the point that the line blueshift in SNe~IIn at early epoch cannot be a reliable argument in favour of the dust formation in CDS and/or ejecta unless supported by the blueshift emergence on the time scale $\ll t$. The observed phenomena in late SN~2010jl, of course, are more complicated than our simplified consideration suggests. Particularly, we did not address the ratio of escaping optical-to-IR fluxes, implicitly assuming that the late IR luminosity is a result of the radiation cascade of the XUV radiation of the forward and reverse shocks. Meanwhile, at late time the optical-to-IR ratio is surprisingly small: around day 800 the difference is a factor of ten \citep{Fransson2014}. That low ratio cannot be explained in the picture in which the primary optical emission is comparable to the IR radiation and only the dust absorption is responsible for the low escaping optical luminosity. Indeed, despite the absorption, almost half of the emitted optical radiation could escape because of the small CDS optical depth ($\sim 0.2$). The low optical-to-IR ratio at late time might be related to the conversion of the XUV radiation in the cool gas directly into the IR radiation avoiding significant emission of the optical radiation. This conjecture could be verified by the detection of strong IR features related to molecular species in supernovae of the similar category at the similar epoch (1.5-2 yr). An alternative possibility is that at late time the forward shock becomes adiabatic which accounts for the low optical luminosity while the high IR luminosity originates from the collisional heating of the newly formed dust mixed with the hot postshock gas. Another puzzling issue is the weakness of the H$\alpha$ broad boxy component related to the CDS material. A plausible answer can be seen from the following arguments. The strong upper limit for the H$\alpha$ luminosity of a thin dense spherical shell with the radius $R$ in the case of very large line optical depth ($\tau > 10^4$) is \begin{equation} L(\mbox{H}\alpha) = 4\pi^2R^2B_{\nu}(T)x_e\Delta \nu_D\,, \label{eq:halum} \end{equation} where $B_{\nu}(T)$ is the black body intensity, $\Delta \nu_D$ is the thermal Doppler width for hydrogen, $x_e = (a\tau/\sqrt{\pi})^{1/3}$ is the local line width in units of the Doppler width in the case of the very opaque layer \citep{Adams1975}, $a =3.3\times10^{-3}$ is the Voigt parameter for the H$\alpha$ for $T \approx 7500$\,K. For SN~2010jl on day 200 the shell radius is $R = 1.4\times10^{16}$ cm, and the CDS mass is $\approx 2\,M_{\odot}$ (according to the interaction model in Section \ref{sec:dmod}). Adopting the excitation temperature of 7500\,K \citep{Fransson2014}, and assuming that the hydrogen is mostly neutral (which majorizes line width $x_e$) for solar hydrogen abundance one gets $L(\mbox{H}\alpha) = 0.9\times10^{41}$ erg\,s$^{-1}$. Note that assumptions made above strongly overestimate the luminosity. Even in that case the obtained value is a factor of ten lower compared to the observed H$\alpha$ luminosity of $\approx 10^{42}$ erg\,$s^{-1}$ \citep{Fransson2014} at that stage. If the CDS is not strongly disturbed, so that the area ratio for the cumulative surface $S$ of the line-emitting CDS material $A = S/4\pi R^2 \sim 1$, then the broad boxy H$\alpha$ related to the CDS turns out relatively weak ($ < 10$\%) and therefore gets overwhelmed by the broad wings of the ejecta component. The above argument have an interesting implication: the dominant contribution of the shocked CS clumps in the H$\alpha$ luminosity suggests that the area ratio of the cumulative surface of the line-emitting CS cloud fragments should be large, of the order of $A \sim 10$. \section{Conclusion} I propose a simple model that accounts for the blueshifted line profiles in spectra of SN~2010jl at late time ($\geq 500$ d). The key elements of this picture are (i) the shocked CS clouds in the forward shock, which are responsible for the dominant intermediate component of line profiles, (ii) the dusty cool dense shell with the small optical depth that accounts for the blueshift of the line maximum, and (iii) dusty unshocked SN ejecta that are responsible for the overall line blueshift. The modelling of the SN interaction with the CSM shows that the SN radius is consistent with the radius of optically thick dusty shell recovered from IR data. This taken together with the line profile analysis suggests that the dust formed in CDS and in unshocked SN ejecta is responsible for both the line blueshift and the IR emission. Most of the dust reside in optically thick clumps of the unshocked ejecta. \section*{Acknowledgements} \bibliographystyle{mnras}

bb7d28d149e238df0eea8e3a04318944e28a7ecd

\section{Introduction}\label{sec:intro} Our ability to construct a dynamic yet coherent perception of the world, despite the numerous parallel sources of information that affect our senses, is to a great extent reliant on the brain's capability to prioritize the processing of task-relevant information over task-irrelevant distractions according to one's goals and desires. This capability, commonly known as goal-driven selective attention (GDSA), has been the subject of extensive research over the past decades. Despite major advances, a fundamental understanding of GDSA and, in particular, how it emerges from the dynamics of the underlying neuronal networks, is still missing. The aim of this work is to reduce this gap by bringing tools and insights from systems and control theory into these questions from neuroscience. In this two-part paper, we propose the novel theoretical framework of Hierarchical Selective Recruitment (HSR) for GDSA. \nnew {As stated in Part I, HSR consists of a novel hierarchical model of brain organization, a set of analytical results regarding the multi-timescale dynamics of this model, and a careful translation between the properties of these dynamics and well known experimental observations about GDSA}. \new{Inspired and supported by extensive experimental research~\cite{ECC:53,AMT:69,JM-RD:85,BCM:93,RD-JD:95,SK-PD-RD-LGU:98,LI-CK:01,MAP-GMD-SK:04,NL:05,JJF-ACS:11,AG-ACN:12,CCR-MRD:14,MG-KH-EN:16}, % HSR relies on four major assumptions about the neuronal mechanisms underlying GDSA. These include (i) the brain's hierarchical organization, so that (cognitively-)higher areas provide control inputs to the activity of lower level ones~\cite{RD-JD:95,LI-CK:01,MAP-GMD-SK:04,NL:05,AG-ACN:12,CCR-MRD:14,MG-KH-EN:16}, % (ii) its sparse coding, so that task-relevant and task-irrelevant stimuli is represented and processed by sufficiently distinct neuronal populations (particularly if the two stimuli differ in major or multiple properties, such as location, sensory modality, etc.)~\cite{JM-RD:85,BCM:93,RD-JD:95,SK-PD-RD-LGU:98,LI-CK:01,MAP-GMD-SK:04,AG-ACN:12,MG-KH-EN:16}, (iii) the distributed and graded nature of GDSA, so that selective attention happens at multiple layers of the hierarchy~\cite{AMT:69,BCM:93,RD-JD:95,SK-PD-RD-LGU:98,LI-CK:01,MAP-GMD-SK:04,JJF-ACS:11,AG-ACN:12,MG-KH-EN:16}, and (iv) the concurrence of the suppression and enhancement of task-irrelevant and task-relevant activity, resp.~\cite{ECC:53,AMT:69,JM-RD:85,BCM:93,RD-JD:95,SK-PD-RD-LGU:98,MAP-GMD-SK:04,NL:05,JJF-ACS:11,AG-ACN:12,CCR-MRD:14,MG-KH-EN:16} % (formulated as \emph{selective inhibition} and \emph{top-down recruitment} in HSR, resp.).} The hierarchical structure of the brain plays a key role in both selective inhibition and top-down recruitment. The position of brain areas along this hierarchy is determined based on the direction in which sensory information and decisions flow, but also by the separation of timescales between the areas. As expected, the timescale of the internal dynamics of the neuronal populations increases (becomes slower) as one moves up the hierarchy~\nnew{\cite{UH-EY-IV-DJH-NR:08,CJH-TT-THD-LJS-CEC-OD-WKD-NR-DJH-UH:12,BG-EE-GH-AG-AK:12,UH-JC-CJH:15,MGM-DAK-SLT-GKA:16,JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14,RC-KK-MG-HK-XW:15}.} Although this hierarchy of timescales is well known in neuroscience, its role in GDSA has remained, to the best of our knowledge, uncharacterized. Using tools from singular perturbation theory, we here reveal the critical role played by this separation of timescales in the top-down recruitment of the task-relevant subnetworks and provide rigorous conditions on the joint structure of all layers that guarantee such recruitment. \subsubsection*{Literature Review} The hierarchical organization of the brain has been recognized for decades~\cite{NT:50,ARL:70,DJF-DCEV:91} and applies to multiple aspects of brain structure and function. These aspects include (i) network topology~\cite{DJF-DCEV:91,AK-ATR-EW-GB-RK:10,GZL-CZ-JK:10,NTM-JV-PC-AF-RQ-CH-CL-PM-PG-SU-PB-CD-KK-HK:14} (where nodes are assigned to layers based on their position on bottom-up and top-down pathways), (ii) encoding properties~\cite{PL:98,DB-MD:09} (where nodes that have larger receptive fields and/or encode more abstract stimulus properties constitute higher layers), (iii) dynamical timescale~\cite{UH-EY-IV-DJH-NR:08,CJH-TT-THD-LJS-CEC-OD-WKD-NR-DJH-UH:12,BG-EE-GH-AG-AK:12,UH-JC-CJH:15,MGM-DAK-SLT-GKA:16,JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14,RC-KK-MG-HK-XW:15,AK-ATR-EW-GB-RK:10,NTM-JV-PC-AF-RQ-CH-CL-PM-PG-SU-PB-CD-KK-HK:14,JPG-MM-JIL:16,CC-HA-DB-YB-SM:14,CAR-EP-SP-CDH:17,SJK-JD-KJF:08,YY-JT:08} (where nodes are grouped into layers according to the timescale of their dynamics), (iv) nodal clustering~\cite{DSB-ETB-BAV-VSM-DRW-AML:08,DM-RL-AF-KE-ETB:09,DM-RL-ETB:10,ZZ-HF-JL:13} (where nodes only constitute the leafs of a clustering tree), and (v) oscillatory activity~\cite{PL-ASS-KHK-IU-GK-CES:05} (where layers correspond to nested oscillatory frequency bands). Note that while hierarchical layers are composed of brain regions in (i)-(iii), this is not the case for (iv) and (v). The hierarchies (i)-(iii) are remarkably similar \new{(in terms of the assignment of brain regions to the layers of the hierarchy)}, and here we particularly focus on (iii) (the timescale separation between hierarchical layers) as it plays a pivotal role in HSR. Studies of timescale separation between cortical regions are more recent. Several experimental works have demonstrated a clear increase in intrinsic timescales as one moves up the hierarchy using indirect measures such as the length of stimulus presentation that elicits a response~\cite{UH-EY-IV-DJH-NR:08,CJH-TT-THD-LJS-CEC-OD-WKD-NR-DJH-UH:12}, resonance frequency~\cite{BG-EE-GH-AG-AK:12}, the length of the largest time window over which the responses to successive stimuli interfere~\cite{UH-JC-CJH:15}, and how quickly the activation level of any brain region can track changes in sensory stimuli~\cite{MGM-DAK-SLT-GKA:16}. Direct evidence for this hierarchical separation of timescales was indeed provided in~\cite{JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14} using the decay rate of spike-count autocorrelation functions. This was shown even more comprehensively in~\cite{RC-KK-MG-HK-XW:15} using linear-threshold rate models and the concept of \emph{continuous hierarchies}~\cite{AK-ATR-EW-GB-RK:10,NTM-JV-PC-AF-RQ-CH-CL-PM-PG-SU-PB-CD-KK-HK:14} (whereby the layer of each node can vary continuously according to its intrinsic timescale, therefore removing the rigidity and arbitrariness of node assignment in classical hierarchical structures). Interestingly, recent studies show that this timescale variability may have roots not only in synaptic dynamics of individual neurons~\cite{JPG-MM-JIL:16}, but also in sub-neuronal genetic factors~\cite{CC-HA-DB-YB-SM:14} as well as supra-neuronal network structures~\cite{CAR-EP-SP-CDH:17}. In terms of applications, computational models of motor control were perhaps the first to exploit this cortical hierarchy of timescales~\cite{SJK-JD-KJF:08,YY-JT:08}. Despite the vastness of the literature on its roots and applications, we are not aware of any theoretical analysis of the effects of this separation of timescales on the hierarchical dynamics of neuronal networks. The accompanying Part I~\cite{EN-JC:18-tacI} proposes the HSR framework, which is strongly rooted in this separation of timescales. Part I analyzes the internal dynamics of the subnetworks at each layer of the hierarchy. Using the class of linear-threshold network models, it characterizes the networks that have a unique equilibrium, are locally/globally asymptotically stable, and have bounded trajectories. In Part I, we also provide a detailed account of feedforward and feedback mechanisms for selective inhibition between any two layers of the hierarchy and show that the internal dynamical properties of the task-relevant subnetwork at each layer is the sole determiner of the dynamical properties achievable under selective inhibition. In this paper, we complete the development of the HSR framework for GDSA by analyzing the mechanisms for top-down recruitment of the task-relevant subnetwork, combining it with the feedforward and feedback mechanisms of selective inhibition, and generalizing the combination to arbitrary number of layers. \new{Top-down recruitment is one of the most experimentally well-documented phenomena in selective attention, see, e.g., \cite{JM-RD:85,BCM:93,RD-JD:95,SK-PD-RD-LGU:98,LI-CK:01,MAP-GMD-SK:04,AG-ACN:12,CCR-MRD:14,MG-KH-EN:16}. While the enhancement (a.k.a. \emph{modulation}) of activity in the task-relevant populations is the simplest form of recruitment, our model is general and thus also allows for more complex observed forms of recruitment, such as changes in the receptive fields% \footnote{\new{The receptive field of each neuron is the area in the stimulus space where the neuron is responsive to the presence of stimuli}.\label{ft:rf}}% ~\cite{YY-MC:98,JBF-ME-SVD-SAS:07,KA-VMS-ST:09}.} \new{In the analysis of top-down recruitment,} we use tools from singular perturbation theory to rigorously leverage this separation of timescales. The classical result on singularly perturbed ODEs goes back to Tikhonov~\cite{ANT:52}, \cite[Thm 11.1]{HKK:02} and has since inspired an extensive literature, see, e.g.~\cite{ABV:94,DN:02,JKK-JDC:12,REO:12}. Tikhonov's result, however, requires smoothness of the vector fields, which is not satisfied by linear-threshold dynamics. Fortunately, several works have sought extensions to non-smooth differential equations and even differential inclusions~\cite{ALD-VMV:83,ALD-IIS:90,MQ:95,AD-TD-IIS:96}, culminating in the work~\cite{VV:97} which we use here. Similar to Tikhonov's original work, \cite{VV:97} only applies to finite intervals. Extensions to infinite intervals exist~\cite{FW:05,GG:05} but, as expected, they require asymptotic stability of the reduced-order model (ROM) which we do not in general have.% \footnote{\new{Recall that in two-timescale dynamics, ROM results from replacing the fast variable with its equilibrium (reducing order to that of the slow variable).}} \subsubsection*{Statement of Contributions} The paper has four main contributions. First, we use the timescale separation in hierarchical brain networks and the theory of singular perturbations to provide an analytic account of top-down recruitment in terms of conditions on the network structure. These conditions guarantee the stability of the task-relevant part of a (fast) linear-threshold subnetwork towards a reference trajectory set by a slower subnetwork. \new{This, in particular, subsumes the most classical enhancement (strengthening) of the activity of task-relevant nodes but is more general and can account for recent, complex observations such as changes in neuronal receptive fields under GDSA~$^{\!\!\ref{ft:rf}}$.} We further combine these results with the results of Part I to allow for simultaneous selective inhibition and top-down recruitment, as observed in GDSA. % Second, we extend this combination to hierarchical structures with an arbitrary number of layers, as observed in nature, to yield a fully developed HSR framework. Here, we also derive an extension of the stability results in Part I that guarantees GES of a multi-layer multiple timescale linear-threshold network. % Third, to validate the proposed HSR framework, we provide a detailed case study of GDSA in real brain networks. Using single-unit recordings from two brain regions of rodents performing a selective listening task, we provide an in-depth analysis of appropriate choices of neuronal populations in each brain region as well as the timescales of their dynamics. We propose a novel hierarchical structure for these populations, tune the parameters of the resulting network % using a novel objective function, and show that the resulting structure conforms to the theoretical results and requirements of HSR while explaining more than $90\%$ of variability in the data. As part of our technical approach, our fourth and final contribution is a novel converse Lyapunov theorem that extends the state of the art on GES for state-dependent switched affine systems. This result only requires continuity of the vector field and guarantees the existence of an infinitely smooth quadratically-growing Lyapunov function if the dynamics is GES. Because of independent interest, we formulate and prove the result for general state-dependent switched affine systems.% \footnote{Throughout the paper, we use the following notation. $\ensuremath{\mathbb{R}}$, $\ensuremath{\mathbb{R}_{\ge 0}}$, $\ensuremath{\mathbb{R}_{\le 0}}$, and $\ensuremath{\mathbb{R}_{>0}}$ denote the set of reals, nonnegative reals, nonpositive reals, and positive reals, resp. $\mathbf{1}_n$, $\mathbf{0}_n$, $\mathbf{0}_{p \times n}$, and $\Ibf_n$ stand for the $n$-vector of all ones, the $n$-vector of all zeros, the $p$-by-$n$ zero matrix, and the identity $n$-by-$n$ matrix, resp. The subscripts are omitted when clear from the context. When a vector $\xbf$ or matrix $\Abf$ are block-partitioned, $\xbf_i$ and $\Abf_{ij}$ refer to the $i$th block of $\xbf$ and $(i, j)$th block of $\Abf$, resp. Given $\Abf \in \ensuremath{\mathbb{R}}^{n \times n}$, its element-wise absolute value and spectral radius are $|\Abf|$ and $\rho(\Abf)$, resp. $\|\cdot\|$ denotes vector $2$-norm. \new{For $x \in \ensuremath{\mathbb{R}}$ and $m \in \ensuremath{\mathbb{R}_{>0}} \cup \{\infty\}$, % $[x]_0^m = \min\{\max\{x, 0\}, m\}$, which is the projection of $x$ onto $[0, m]$. When $\xbf \in \ensuremath{\mathbb{R}}^n$ and $\mbf \in \ensuremath{\mathbb{R}_{>0}}^n \cup \{\infty\}^n$, we similarly define $[\xbf]_\mathbf{0}^\mbf = [[x_1]_0^{m_1} \quad \cdots \quad [x_n]_0^{m_n}]^T$. For $\sigmab \in \zls^n$, $\Sigmab^\ell = \Sigmab^\ell(\sigmab)$ is a diagonal matrix with $\Sigma^\ell_{ii} = 1$ if $\sigma_i = \ell$ and $\Sigma^\ell_{ii} = 0$ otherwise. $\Sigmab^\s$ is defined similarly. We set the convention that $\Sigmab^\s \mbf = \mathbf{0}$ if $\Sigmab^\s = \mathbf{0}$ and $\mbf = \infty \mathbf{1}_n$. For $D \subseteq \ensuremath{\mathbb{R}}^n$ and $\Abf \in \ensuremath{\mathbb{R}}^{p \times n}, \bbf \in \ensuremath{\mathbb{R}}^p$, we let $\Abf D + \bbf = \setdef{\Abf \xbf + \bbf}{\xbf \in D}$.}} \begin{figure}[tbh] \begin{center} \begin{tikzpicture} \node[circle, draw, line width=0.6pt, inner sep=-0.4pt] (1) {\includegraphics[width=16pt]{EI_main_half_gray}}; \draw[-, line width=0.3pt] (1.north) to (1.south); \node[below of=1, yshift=-10pt, circle, draw, line width=0.6pt, inner sep=-0.4pt] (2) {\includegraphics[width=16pt]{EI_main_half_gray}}; \draw[-, line width=0.3pt] (2.north) to (2.south); \node[below of=2, yshift=-10pt, circle, draw, line width=0.6pt, inner sep=-0.4pt] (3) {\includegraphics[width=16pt]{EI_main_half_gray}}; \draw[-, line width=0.3pt] (3.north) to (3.south); \draw[latex-latex, line width=0.3pt, shorten <=1pt, shorten >=1pt] (1.270) to (2.90); \draw[latex-latex, line width=0.3pt, shorten <=1pt, shorten >=1pt] (2.270) to (3.90); \node[below of=3, yshift=10pt] {$\vdots$}; \node[above of=1, yshift=-5pt] {$\vdots$}; \node[left of=2, xshift=-20pt, yshift=40pt, scale=0.8] (i-1) {Subnetwork $i - 1$}; \node[below of=i-1, yshift=-11pt, scale=0.8] (i) {Subnetwork $i$}; \node[below of=i, yshift=-12pt, scale=0.8] (i+1) {Subnetwork $i + 1$}; \node[right of=2, xshift=50pt, circle, draw, line width=1pt, inner sep=-2pt] (i-big) {\includegraphics[width=40pt]{EI_main_half_gray}}; \draw[-] (i-big.north) to (i-big.south); \draw[shorten <=5pt, shorten >=25pt] (2.75) to (i-big.85); \draw[shorten <=5pt, shorten >=25pt] (2.285) to (i-big.275); \node[below of=i-big, xshift=-30pt, yshift=-25pt, scale=0.8] (i1) {\parbox{50pt}{\centering $\mathcal{N}_i^\ssm$ \\ (inhibited, state = $\xbf_i^\ssm$)}}; \node[below of=i-big, xshift=30pt, yshift=-25pt, scale=0.8] (i2) {\parbox{50pt}{\centering $\mathcal{N}_i^\ssp$ \\ (recruited, state = $\xbf_i^\ssp$)}}; \draw[-latex, bend right=0, shorten <=-8pt, shorten >=3pt] (i1.45) to (i-big.250); \draw[-latex, bend right=0, shorten <=-8pt, shorten >=3pt] (i2.135) to (i-big.290); \end{tikzpicture} \end{center} \caption{Hierarchical network structure considered in this work.}\label{fig:multi} \vspace*{-1.5ex} \end{figure} \section{Problem Statement}\label{sec:prob-state} The problem formulation is the same as in Part I~\cite{EN-JC:18-tacI}. We include here a streamlined description for a self-contained exposition. We consider a hierarchical neuronal network $\mathcal{N}$, cf. Figure~\ref{fig:multi}, whereby the nodes in each layer $\mathcal{N}_i$ are further decomposed into a task-irrelevant part $\mathcal{N}_i^\ssm$ and a task-relevant part $\mathcal{N}_i^\ssp$. The state evolution of each layer $\mathcal{N}_i$ is modeled with linear-threshold network dynamics of the form \begin{align}\label{eq:dyn} \notag \tau_i \dot \xbf_i(t) = -\xbf_i(t) + [\Wbf_{i, i} \xbf_i(t) + \dbf_i&(t)]_\mathbf{0}^{\mbf_i}, \ \mathbf{0} \le \xbf_i(0) \le \mbf_i, \\ &\mbf_i \in \ensuremath{\mathbb{R}_{>0}}^n \cup \{\infty\}^n, \!\! \end{align} where $\xbf_i \in \ensuremath{\mathbb{R}}^{n_i}$, $\Wbf_{i, i} \in \ensuremath{\mathbb{R}}^{n_i \times n_i}$, $\dbf_i \in \ensuremath{\mathbb{R}}^{n_i}$, and $\mbf \in \ensuremath{\mathbb{R}_{>0}}^n$ denote the state, internal synaptic connectivity, external inputs, and state upper bounds of $\mathcal{N}_i$, resp.% \footnote{\nnew{We note that this is a standard and widely used model in computational neuroscience, as mentioned in Part I~\cite{EN-JC:18-tacI}. Please see therein for a detailed literature review of its origins and prior analysis.}} The development of HSR is structured in four thrusts: \begin{enumerate}[wide] \item the analysis of the relationship between structure ($\Wbf_{i, i}$) and dynamical behavior % % % for each subnetwork when operating in isolation from the rest of the network ($\dbf_i(t) \equiv \dbf_i$); \item the analysis of the conditions on the joint structure of each two successive layers $\mathcal{N}_i$ and $\mathcal{N}_{i+1}$ that allows for selective inhibition of $\mathcal{N}_{i+1}^\ssm$ by its input from $\mathcal{N}_i$, being equivalent to the stabilization of $\mathcal{N}_{i+1}^\ssm$ to the origin % (inactivity); \item the analysis of the conditions on the joint structure of each two successive layers $\mathcal{N}_i$ and $\mathcal{N}_{i+1}$ that allows for top-down recruitment of $\mathcal{N}_{i+1}^\ssp$ by its input from $\mathcal{N}_i$, being equivalent to the stabilization of $\mathcal{N}_{i+1}^\ssp$ toward a desired trajectory set by $\mathcal{N}_i$ (activity); \item the combination of (ii) and (iii) in a unified framework and its extension to the complete $N$-layer network of networks. \end{enumerate} Problems (i) and (ii) are addressed in Part I~\cite{EN-JC:18-tacI}, while problems (iii) and (iv) are the subject of this work. We let \begin{align}% \dbf_i(t) = \Bbf_i \ubf_i(t) + \tilde \dbf_i(t), \qquad \ubf_i \in \ensuremath{\mathbb{R}_{\ge 0}}^{p_i}, \end{align} where $\ubf_i$ is the top-down control used for inhibition of $\mathcal{N}_i^\ssm$. While in Part I we assumed for simplicity that $\tilde \dbf_i(t)$ is given and constant, we here consider its complete form \begin{align*} \tilde \dbf_i(t) = \Wbf_{i, i-1} \xbf_{i-1}(t) + \Wbf_{i, i+1} \xbf_{i+1}(t) + \cbf_i, \end{align*} where the inter-layer connectivity matrices $\Wbf_{i, i-1}$ and $\Wbf_{i, i+1}$ have appropriate dimensions and $\cbf_i \in \ensuremath{\mathbb{R}}^{n_i}$ captures un-modeled background activity and possibly nonzero activation thresholds. Substituting these into~\eqref{eq:dyn}, the dynamics of each layer $\mathcal{N}_i$ is given by {\interdisplaylinepenalty=10000 \new{\begin{align}\label{eq:dyn-multi} \tau_i \dot \xbf_i(t) = -\xbf_i(t) + [&\Wbf_{i,i} \xbf_i(t) + \Wbf_{i, i-1} \xbf_{i-1}(t) \\ \notag &+ \Wbf_{i, i+1} \xbf_{i+1}(t) + \Bbf_i \ubf_i(t) + \cbf_i]_\mathbf{0}^{\mbf_i}. \end{align}} } \new{Also following Part I, we partition network variables as % \begin{align}\label{eq:wbi} \notag \xbf_i = \begin{bmatrix}[1.3] \xbf_i^\ssm \\ \xbf_i^\ssp \end{bmatrix}, \quad \Wbf_{i, j} &= \begin{bmatrix}[1.3] \Wbf_{i, j}^{\ssm\ssm} & \Wbf_{i, j}^{\ssm\ssp} \\ \Wbf_{i, j}^{\ssp\ssm} & \Wbf_{i, j}^{\ssp\ssp} \end{bmatrix} , \quad \Bbf_i = \begin{bmatrix}[1.3] \Bbf_i^\ssm \\ \mathbf{0} \end{bmatrix}, \\ \cbf_i &= \begin{bmatrix}[1.3] \cbf_i^\ssm \\ \cbf_i^\ssp \end{bmatrix}, \quad \mbf_i = \begin{bmatrix}[1.3] \mbf_i^\ssm \\ \mbf_i^\ssp,\end{bmatrix} \end{align}} where $\xbf_i^\ssm \in \ensuremath{\mathbb{R}}^{r_i}$ for all $i, j \in \until{N}$. By convention, $\Wbf_{1, 0} = \mathbf{0}$, $\Wbf_{N, N+1} = \mathbf{0}$, and $r_1 = 0$ (so $\Bbf_1 = \mathbf{0}$ and the first subnetwork has no inhibited part). We assume that the hierarchical layers have sufficient timescale separation, i.e., \begin{align}\label{eq:tau} \tau_1 \gg \tau_2 \gg \dots \gg \tau_N. \end{align} Finally, let $\epsilonb = (\epsilon_1, \dots, \epsilon_{N-1})$, with \begin{align}\label{eq:eps} \nnew{\epsilon_i = \tau_{i+1}/\tau_i, \qquad i = \until{N-1}.} \end{align} Next, we first develop the main concepts and results for the case of bilayer networks (Section~\ref{sec:bilayer}) and then extend them to the setup with $N$ layers (Section~\ref{sec:multilayer}). \section{Selective Recruitment in Bilayer Networks}\label{sec:bilayer} In this section we tackle the analysis of simultaneous selective inhibition and top-down recruitment in a two-layer network. We consider the same dynamics as in~\eqref{eq:dyn-multi} for the lower-level subnetwork $\mathcal{N}_2$, but temporarily allow the dynamics of $\mathcal{N}_1$ to be arbitrary. This setup allows us to study the key ingredients of selective recruitment without the extra complications that arise from the multilayer interconnections of linear-threshold subnetworks and is the basis for our later developments. Further, by keeping the higher-level dynamics arbitrary, the results presented here are also of independent interest beyond HSR, as they allow for a broader range of external inputs $\dbf(t)$ than those generated by linear-threshold dynamics. This can be of interest in, for example, \new{direct brain stimulation applications where $\xbf_1(t)$ is generated and applied by a computer in order to control the activity $\xbf_2(t)$ of certain areas of the brain. In this view, appropriate stimulation signals $\xbf_1(t)$ may be considered as an augmentation of the natural hierarchy of the brain if they vary slow enough to satisfy~\eqref{eq:tau}.} Section~\ref{sec:multilayer} builds on the insights obtained here to generalize this framework to the multilayer case described in Section~\ref{sec:prob-state}. \new{For any $\Wbf \in \ensuremath{\mathbb{R}}^{n \times n}$, define $h: \ensuremath{\mathbb{R}}^n \rightrightarrows \ensuremath{\mathbb{R}_{\ge 0}}^n$ by \begin{align}\label{eq:h} h(\dbf) = h_{\Wbf, \mbf}(\dbf) \triangleq \setdef{\xbf \in \ensuremath{\mathbb{R}_{\ge 0}}^n}{\xbf = [\Wbf \xbf + \dbf]_\mathbf{0}^\mbf}, \end{align} which maps any constant input~$\dbf \in \ensuremath{\mathbb{R}}^n$ to the corresponding set of the equilibria of~\eqref{eq:dyn}. Due to the switched-affine form of the dynamics, $h$ has the piecewise-affine form \begin{align}\label{eq:h-pa} h(\dbf) &= \setdef{\Fbf_\sigmab \dbf + \fbf_\sigmab}{\Gbf_\sigmab \dbf + \gbf_\sigmab \ge \mathbf{0}, \, \sigmab \in \zls^n}, \\ \notag \Fbf_\sigmab &= (\Ibf - \Sigmab^\ell \Wbf)^{-1} \Sigmab^\ell, \ \ \fbf_\sigmab = (\Ibf - \Sigmab^\ell \Wbf)^{-1} \Sigmab^\s \mbf, \\ \notag \Gbf_\sigmab &= \begin{bmatrix} \Sigmab^\ell + \Sigmab^\s - \Ibf & \Sigmab^\ell & -\Sigmab^\ell & \Sigmab^\s \end{bmatrix}^T \Fbf_\sigmab, \\ \notag \gbf_\sigmab &= \begin{bmatrix} \fbf_\sigmab^T(\Sigmab^\ell \!\!+\!\! \Sigmab^\s \!\!-\!\! \Ibf) & \!\!\fbf_\sigmab^T \Sigmab^\ell & \!\!(\mbf \!\!-\!\! \fbf_\sigmab)^T \Sigmab^\ell & \!\!(\fbf_\sigmab \!\!-\!\! \mbf)^T \Sigmab^\s \end{bmatrix}^T. \end{align}} The existence and uniqueness of equilibria of~\eqref{eq:dyn} precisely corresponds to $h$ being single-valued on $\ensuremath{\mathbb{R}}^n$, in which case we let $h: \ensuremath{\mathbb{R}}^n \to \ensuremath{\mathbb{R}_{\ge 0}}^n$ be an ordinary function. For our subsequent analysis we need $h$ to be Lipschitz, as stated next. The proof of this result is a special case of Lemma~\ref{lem:h-lip-gen} and thus omitted. \begin{lemma}\longthmtitle{Lipschitzness of $h$}\label{lem:h-lip} Let $h$ be as in~\eqref{eq:h} and single-valued% \footnote{It is possible to show, using the same proof technique, that $h$ is Lipschitz in the Hausdorff metric even when it is multiple-valued \new{(recall that the Hausdorff distance between two sets $S_1, S_2 \in \ensuremath{\mathbb{R}}^n$ is defined as $\max\{\sup_{\abf \in S_1} \inf_{\bbf \in S_2} \|\abf - \bbf\|, \sup_{\bbf \in S_2} \inf_{\abf \in S_1} \|\abf - \bbf\|\}$)}.}% on $\ensuremath{\mathbb{R}}^n$. Then, it is globally Lipschitz on $\ensuremath{\mathbb{R}}^n$. \end{lemma} The main result of this section is as follows. \begin{theorem}\longthmtitle{Selective recruitment in bilayer % networks}\label{thm:sp-inhib} \new{Consider the multilayer dynamics~\eqref{eq:dyn-multi} where $N = 2$, $\Wbf_{2, 1}^{\ssm\ssp} = \mathbf{0}$, and $\cbf_2^\ssm = \mathbf{0}$} but $\xbf_1(t)$ is generated by some arbitrary dynamics \begin{align}\label{eq:dtilde} \tau_1 \dot \xbf_1(t) = \gamma(\xbf_1(t), \xbf_2(t), t). \end{align} Let \new{$h_2^\ssp = h_{\Wbf_{2, 2}^{\ssp\ssp}, \mbf_2^\ssp}$} as in~\eqref{eq:h}. If \begin{enumerate} \item $\gamma$ is measurable in $t$, locally bounded, and locally Lipschitz in $(\xbf_1, \xbf_2)$ uniformly in $t$; % \item \eqref{eq:dtilde} has bounded solutions uniformly in $\xbf_2(t)$; \item $p_2 \ge r_2$; % % % % % % \item \new{$\Wbf_{2, 2}^{\ssp\ssp}$ is such that % % $\tau \dot \xbf_2^\ssp = -\xbf_2^\ssp + [\Wbf_{2, 2}^{\ssp\ssp} \xbf_2^\ssp + \dbf_1^\ssp]_\mathbf{0}^{\mbf_2^\ssp}$ is GES towards a unique equilibrium for any constant $\dbf_1^\ssp$}; \end{enumerate} then there exists $\Kbf_2 \in \ensuremath{\mathbb{R}}^{p_2 \times n_2}$ such that by using the feedback control $\ubf_2(t) = \Kbf_2 \xbf_2(t)$, one has \new{\begin{align}\label{eq:tik} \lim\limits_{\epsilon_1 \to 0} \ \sup\limits_{t \in [\underline t , \bar t]} \Big\|\xbf_2(t) - \Big(\mathbf{0}_{r_2}, h_2^\ssp\big(\Wbf_{21}^{\ssp\ssp} \xbf_1(t) + \cbf_2^\ssp\big)\Big)\Big\| = 0, \end{align}} for any $0 < \underline t < \bar t < \infty$, \nnew{with $\epsilon_1$ given in~\eqref{eq:eps}}. Further, if the dynamics of $\xbf_2$ is monotonically bounded% \footnote{See~\cite[Def V.1]{EN-JC:18-tacI}.}, there also exists a feedforward control $\ubf_2(t) \equiv \bar \ubf_2$ such that~\eqref{eq:tik} holds for any $0 < \underline t < \bar t < \infty$ \new{and arbitrary % $\Wbf_{2, 1}^{\ssm\ssp}$ and~$\cbf_2^\ssm$}. \end{theorem} \begin{proof} First we prove the result for feedback control. By~\emph{(iii)}, % % there exists $\Kbf_2 \in \ensuremath{\mathbb{R}}^{p_2 \times n_2}$ \new{almost always} (i.e., for \new{almost all} $(\Wbf_{2, 2}^{\ssm\ssm}, \Wbf_{2, 2}^{\ssm\ssp}, \Bbf_2^\ssm)$) such that \begin{align}\label{eq:w_bk} \Wbf_{2, 2} + \Bbf_2 \Kbf_2 = \begin{bmatrix} \mathbf{0} & \mathbf{0} \\ \Wbf_{2, 2}^{\ssp\ssm} & \Wbf_{2, 2}^{\ssp\ssp} \end{bmatrix}. \end{align} Further, by~\cite[Thm IV.7(ii) \& Thm V.3(ii)]{EN-JC:18-tacI}, all the principal submatrices of $-\Ibf + (\Wbf_{2, 2} + \Bbf_2 \Kbf_2)$ are Hurwitz. Therefore, by~\cite[Thm IV.3 \& Assump 1]{EN-JC:18-tacI}, $h_2^\ssp$ is singleton-valued \new{almost always (i.e., for almost all $\Wbf_{2, 2}$)}. Thus, the dynamics of $\xbf_2$ simplifies to \new{\begin{align}\label{eq:dyn-d-sim} \tau_2 \dot \xbf_2^\ssm &= -\xbf_2^\ssm, \\ \notag \tau_2 \dot \xbf_2^\ssp &= -\xbf_2^\ssp + [\Wbf_{2, 2}^{\ssp\ssm} \xbf_2^\ssm + \Wbf_{2, 2}^{\ssp\ssp} \xbf_2^\ssp + \Wbf_{2, 1}^{\ssp\ssp} \xbf_1 + \cbf_2^\ssp]_\mathbf{0}^{\mbf_2^\ssp}, \end{align}} and has a unique equilibrium for any \emph{fixed} $\xbf_1$. Assumption~\emph{(iv)} and~\cite[Lemma~A.2]{EN-JC:18-tacI} then ensure that~\eqref{eq:dyn-d-sim} is GES relative to \new{$(\mathbf{0}_{r_2}, h_2^\ssp(\Wbf_{21}^{\ssp\ssp} \xbf_1(t) + \cbf_2^\ssp))$} for any fixed~$\xbf_1$. Based on assumption~\emph{(ii)}, let $D \subset \ensuremath{\mathbb{R}}^n$ be a compact set that contains the trajectory of the reduced-order model % \new{$\tau_1 \dot \xbf_1 = \gamma(\xbf_1, (\mathbf{0}_{r_2}, h_2^\ssp(\Wbf_{21}^{\ssp\ssp} \xbf_1(t) + \cbf_2^\ssp)), t)$.} % By assumption~\emph{(i)}, $\gamma$ is Lipschitz in $(\xbf_1, \xbf_2)$ on compacts uniformly in $t$. Let $L_\gamma$ be its associated Lipschitz constant on \new{$D \times \{\mathbf{0}_{r_2}\} \times h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} D + \cbf_2^\ssp)$}. Using~\eqref{eq:h-pa} and Lemma~\ref{lem:h-lip}, \new{for all $\xbf_1, \hat \xbf_1 \in D$, \begin{align*} &\|\gamma(\xbf_1, h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \xbf_1 + \cbf_2^\ssp), t) - \gamma(\hat \xbf_1, h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \hat \xbf_1 + \cbf_2^\ssp), t)\| \\ &\le L_\gamma \|(\xbf_1 - \hat \xbf_1, h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \xbf_1 + \cbf_2^\ssp) - h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \hat \xbf_1 + \cbf_2^\ssp))\| \\ &\le L_\gamma(\|\xbf_1 - \hat \xbf_1\| \!+\! \|h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \xbf_1 + \cbf_2^\ssp) \!-\! h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \hat \xbf_1 + \cbf_2^\ssp)\|) \\ &\le L_\gamma(1 + L_h \|\Wbf_{2, 1}^{\ssp\ssp}\|) \|\xbf_1 - \hat \xbf_1\|, \end{align*} so $\gamma(\cdot, h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \cdot + \cbf_2^\ssp), t): \ensuremath{\mathbb{R}}^{n_1} \to \ensuremath{\mathbb{R}}^{n_1}$ is $L_\gamma(1 + L_h \|\Wbf_{2, 1}^{\ssp\ssp}\|)$-Lipschitz on~$D$.} Using this, Lemma~\ref{lem:h-lip-gen} again, and the change of variables $t' \triangleq t/\tau_1$, the claim follows from~\cite[Prop 1]{VV:97}.% \footnote{\new{\cite[Prop 1]{VV:97} is applicable to singularly perturbed differential inclusions and thus technically involved, but for non-smooth ODEs such as~\eqref{eq:dyn-multi}, its assumptions can be simplified to: 1. Lipschitzness of dynamics uniformly in $t$, 2. Existence, uniqueness, and Lipschitzness of the equilibrium map of fast dynamics, 3. Lipschitzness and boundedness of the reduced-order model, 4. asymptotic stability of the fast dynamics uniformly in $t$ and the slow variable, and 5. global attractivity of fast dynamics for any fixed slow variable.}} Next, we prove the result for constant feedforward control $\ubf_2(t) \equiv \bar \ubf_2$. Based on assumption~\emph{(ii)}, let $\bar \xbf_1 \in \ensuremath{\mathbb{R}_{>0}}^{n_1}$ be the bound on the trajectories of~\eqref{eq:dtilde} and $\bar \ubf_2$ be a solution of \new{\begin{align*} \Bbf_2^\ssm \bar \ubf_2 = -[[\Wbf_{2, 2}^{\ssm\ssm} \ \Wbf_{2, 2}^{\ssm\ssp}]]_\mathbf{0}^\infty \nub(\bar \xbf_1) - [\Wbf_{2, 1}^{\ssm\ssp}]_\mathbf{0}^\infty \bar \xbf_1 + [\cbf_2^\ssm]_\mathbf{0}^\infty, % \end{align*}} where $\nub$ comes from the monotone boundedness of the dynamics of $\xbf_2$. This solution \new{almost always} exists by assumption~\emph{(ii)}. Then, the dynamics of $\xbf_2$ simplifies to~\eqref{eq:dyn-d-sim}, and~\cite[Lemma~A.2]{EN-JC:18-tacI} guarantees that it is GES relative to \new{$(\mathbf{0}_{r_2}, h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \xbf_1 + \cbf_2^\ssp))$} for any \emph{fixed} $\xbf_1$. The claim then follows, similar to the feedback case, from~\cite[Prop 1]{VV:97}. \end{proof} \begin{remark}\longthmtitle{Validity of the assumptions of Theorem~\ref{thm:sp-inhib}.}\label{re:assumptions-reasonable} {\rm \new{Assumption~\emph{(i)} is merely technical and satisfied by all well-known models of neuronal rate dynamics, including the linear-threshold model itself.} Likewise, assumption~\emph{(ii)} is always satisfied in reality, as the firing rates of all biological neuronal networks are bounded by the inverse of the refractory period of their neurons. \nnew{In theory, the verification of this assumption depends clearly on $\gamma$. If a linear-threshold model is used, we can instead use Theorem~\ref{thm:multi} and relax assumption~\emph{(ii)} to a less restrictive one (assumption~\emph{(i)} of Theorem~\ref{thm:multi}), which can in turn be verified using the sufficient condition in Theorem~\ref{thm:multi-EUE-GES}.} \new{Assumption~\emph{(iii)} requires the existence of at least as many inhibitory control channels as the number of nodes in $\mathcal{N}_2$ that are to be inhibited. Indeed, selective inhibition is still possible without this assumption (cf. Theorem~\ref{thm:multi}), but may require excessive inhibitory resources.} \new{The most critical requirement is assumption~\emph{(iv)}, but is not only sufficient but also necessary for inhibitory stabilization (cf.~\cite[Thm IV.8]{EN-JC:18-tacI} for conditions on $\Wbf_{2, 2}^{\ssp\ssp}$ that ensure this assumption as well as~\cite[Thm V.2 \& V.3]{EN-JC:18-tacI} for its necessity for feedforward and feedback inhibitory stabilization)}. } \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol \end{remark} The main conclusion of Theorem~\ref{thm:sp-inhib} is the Tikhonov-type singular perturbation statement in~\eqref{eq:tik}. According to this statement, \new{the tracking error can be made arbitrarily small, i.e.,} for any $\theta > 0$, \begin{align}\label{eq:tikkk} % % % |\xbf_2(t) - (\mathbf{0}_{r_2}, h_2^\ssp(\xbf_1(t)))| &\le \theta \mathbf{1}_{n_2}, \qquad \forall t \in [\underline t, \bar t], \end{align} provided that $\tau_2 / \tau_1$ is sufficiently small. As discussed in Section~\ref{sec:intro}, this timescale separation is characteristic of biological neuronal networks. \new{In general, the smaller the time constant ratio $\tau_2/\tau_1$, the smaller the tracking error $\theta$. As shown in~\cite{JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14}, several pairs of regions along the sensory-frontal pathways have successive time constant ratios between $1/1.5$ and $1/2.5$, which is often (more than) enough in simulations for~\eqref{eq:tikkk} to hold with small enough $\theta$, as shown in Example~\ref{ex:osc} below.} An important observation regarding~\eqref{eq:tikkk} is that the equilibrium map $h_2^\ssp$ does not have a closed-form expression, so the reference trajectory $h_2^\ssp(\xbf_1(t))$ of the lower-level network is only implicitly known for any given~$\xbf_1(t)$. \new{However, if a desired trajectory $\xib_2^\ssp(t) \in \prod_{j = r_2+1}^{n_2} [0, m_{2, j}]$ % for $\xbf_2^\ssp$ is known a priori, one can specify the appropriate $\gamma$ such that $h_2^\ssp(\xbf_1(t)) = \xib_2^\ssp(t)$. To show this, let the dynamics of $\xib_2^\ssp(t)$ be \begin{align*} \tau_1 \dot \xib_2^\ssp(t) = \gamma_\xi(\xib_2^\ssp(t), t). \end{align*} Then, choosing $\xbf_1(t) = (\Wbf_{2, 1}^{\ssp\ssp})^{-1} \big((\Ibf - \Wbf_{2, 2}^{\ssp\ssp}) \xib_2^\ssp(t) - \cbf_2^\ssp\big)$, % \begin{align*} [\Wbf_{2, 2}^{\ssp\ssp} \xib_2^\ssp(t) + \Wbf_{2, 1}^{\ssp\ssp} \xbf_1(t) + \cbf_2^\ssp]_\mathbf{0}^{\mbf_2^\ssp} = [\xib_2^\ssp(t)]_\mathbf{0}^{\mbf_2^\ssp} = \xib_2^\ssp(t) , \end{align*} which, according to~\eqref{eq:h}, implies $\xib_2^\ssp(t) = h_2^\ssp(\xbf_1(t))$.} We use this result to illustrate the core concepts of the bilayer HSR in a synthetic but biologically-inspired example, where a inhibitory subnetwork generates oscillations which are selectively induced on a lower-level excitatory subnetwork. \begin{example}\longthmtitle{HSR of an excitatory subnetwork by inhibitory oscillations}\label{ex:osc} \rm Consider the dynamics~\eqref{eq:dyn-multi} with $N = 2$, a $3$-dimensional excitatory subnetwork at the lower level, a $3$-dimensional inhibitory subnetwork at the higher level, \new{and $\mbf_1 = \mbf_2 = \infty \mathbf{1}_3$} (Figure~\ref{fig:sim-hier}). Let \begin{align}\label{eq:ex-osc} \notag &\Wbf_{1, 1} = \begin{bmatrix} 0 & -0.8 & -1.7 \\ -1 & 0 & -0.5 \\ -0.7 & -1.8 & 0 \end{bmatrix}, \quad \cbf_1 = \begin{bmatrix} 11 \\ 10 \\ 10 \end{bmatrix}, \\ \notag &\Wbf_{2, 2} = \begin{bmatrix} 0 & 0.9 & 1.2 \\ 0.7 & 0 & 1 \\ 0.8 & 0.2 & 0 \end{bmatrix}, \quad \Bbf_2 = \begin{bmatrix} -1 \\ 0 \\ 0 \end{bmatrix}, \quad \cbf_2 = \begin{bmatrix} 2 \\ 3.5 \\ 2.5 \end{bmatrix}, \\ &\Wbf_{1, 2} = \mathbf{0}, \quad \Wbf_{2, 1} = -\Ibf, \quad u_2 = 5. \end{align} This example satisfies all the assumptions of Theorem~\ref{thm:sp-inhib}, so we expect the actual $\xbf_2$-trajectory to be close to the \emph{desired} $\xbf_2$-trajectory $(0, h_2^\ssp(\xbf_1(t))$ provided that $\epsilon_1 \ll 1$. As shown in Figure~\ref{fig:sim-hier}, $\xbf_2(t)$ and $(0, h_2^\ssp(\xbf_1(t))$ are remarkably close even with a mild separation of timescales, $\epsilon_1 = 0.5$. \begin{figure} \centering \begin{tikzpicture} \node(g) {\includegraphics[width=0.13\textwidth]{ei_hierarchy}}; \node[left of=g, xshift=-80pt, yshift=-8pt] (gt) {\includegraphics[width=0.24\textwidth]{sim_hier_traj}}; \node[above right of=g, xshift=35pt, yshift=3pt, rotate=0, scale=0.8] (i) {\parbox{40pt}{\centering Inhibitory \\ $\tau_1$}}; \node[below right of=g, xshift=35pt, yshift=2pt, rotate=0, scale=0.8] (e) {\parbox{45pt}{\centering Excitatory \\ $\tau_2 = 0.5 \tau_1$}}; \end{tikzpicture} \caption{The network structure (right) and trajectories (left) of the two-timescale network in~\eqref{eq:ex-osc} \new{for $\tau_1 = 3.3^\text{ms}$}. The red pyramids and blue circles depict excitatory and inhibitory nodes, resp., and the trajectory colors on the left correspond to node colors on the right. The dashed lines show the desired reference trajectories \new{$\big(0, h_2^\protect\ssp(\Wbf_{2, 1}^{\protect\ssp\protect\ssp} \xbf_1(t) + \cbf_2^\protect\ssp)\big)$}.} % \label{fig:sim-hier} \vspace{-1.5ex} \end{figure} \new{This example further illustrates the complementary roles of selective inhibition and selective recruitment. The complete $\xbf_2$-subsystem is unstable by itself, but any two-dimensional subnetwork of it is stable. Therefore, $\mathcal{N}_1$ can selectively inhibit any single node of $\mathcal{N}_2$ while simultaneously recruiting (e.g., by inducing oscillations in) the remaining two.} \new{Thus, as suggested earlier in~\cite[Rem V.7]{EN-JC:18-tacI}, different ``tasks" can be accomplished at different times by varying the selectively recruited subnetwork over time. Generalizing this to more complex networks allows for more flexible selective recruitment schemes of larger neuronal subnetworks, % as observed in nature. % } \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol \end{example} \begin{remark}\longthmtitle{Biological relevance of Example~\ref{ex:osc}} {\rm In addition to providing a simple illustration of the HSR framework developed here, Example~\ref{ex:osc} has interesting similarities with % well-known aspects of selective attention in the brain. Extensive % studies have demonstrated a robust correlation between oscillatory activity, particularly in the gamma range % ($\sim 30-100^\text{Hz}$), and selective attention% ~\cite{PF-JHR-AER-RD:01,AS-MP-WL-NB:04,SR-EN-SSH-AS-NEC:08,NK-MB-ESM-AS:12}. % Furthermore, gamma oscillations in the cortex are shown to be primarily generated by networks of inhibitory neurons, which then recruit the excitatory populations (see~\cite{JAC-MC-KM-UK-FZ-KD-LT-CIM:09} and the references therein), as captured by the network structure of Figure~\ref{fig:sim-hier}. Interestingly, the oscillations generated by the higher-level inhibitory subnetwork fall within the gamma band by setting $\tau_1 \sim 3^\text{ms}$ which lies within the decay time constant range of $\text{GABA}_\text{A}$ inhibitory receptors% \footnote{See, e.g., the Neurotransmitter Time Constants database of the CNRGlab at the University of Waterloo, \url{http://compneuro.uwaterloo.ca/research/constants-constraints/neurotransmitter-time-constants-pscs.html}.} \new{(the major type of inhibitory synapse in the central nervous system)}. \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol} \end{remark} \section{Selective Recruitment in Multilayer Networks}\label{sec:multilayer} We tackle here the problem of Section~\ref{sec:prob-state} in its general form and consider an $N$-layer hierarchical structure of subnetworks with linear-threshold dynamics. Given~\eqref{eq:dyn-multi}, let \new{\begin{align*} &h_i^\ssp: \cbf_i^\ssp \rightrightarrows \setdef{\xbf_i^\ssp}{\xbf_i^\ssp = [\Wbf_{i, i+1}^{\ssp\ssp} h_{i+1}^\ssp(\Wbf_{i+1, i}^{\ssp\ssp} \xbf_i^\ssp + \cbf_{i+1}^\ssp) \\ &\hspace{92pt}+ \Wbf_{i, i}^{\ssp\ssp} \xbf_i^\ssp + \cbf_i^\ssp]_\mathbf{0}^{\mbf_i^\ssp}}, \ i \!=\! 2, \dots, N\!-\!1, % % \end{align*} with $h_N^\ssp = h_{\Wbf_{N, N}^{\ssp\ssp}, \mbf_N^\ssp}$,} be the recursive definition of the (set-valued) equilibrium maps of the task-relevant parts of the layers $\{2, \dots, N\}$. These maps % play a central role in the multiple-timescale dynamics of~\eqref{eq:dyn-multi}. Therefore, we begin by characterizing their piecewise-affine nature. The proof of the following result can be found in~\ref{app:pf}. \begin{lemma}\longthmtitle{Piecewise affinity of equilibrium maps is preserved along layers of hierarchical linear-threshold network}\label{lem:affinity} Let $h:\ensuremath{\mathbb{R}}^n \to \ensuremath{\mathbb{R}}^n$ be a piecewise affine function, \begin{align*} h(\cbf) = \Fbf_\lambda \cbf + \fbf_\lambda, \qquad &\forall \cbf \in \Psi_\lambda \triangleq \setdef{\cbf}{\Gbf_\lambda \cbf + \gbf_\lambda \ge \mathbf{0}}, \\ &\forall \lambda \in \Lambda, \end{align*} where $\Lambda$ is a finite index set and $\bigcup_{\lambda \in \Lambda} \Psi_\lambda = \ensuremath{\mathbb{R}}^n$. Given matrices $\Wbf_\ell, \ell = 1, 2, 3$ and a vector $\bar \cbf$, assume % \new{\begin{align}\label{eq:gen-eq} \xbf = [\Wbf_1 \xbf + \Wbf_2 h(\Wbf_3 \xbf + \bar \cbf) + \cbf']_\mathbf{0}^\mbf, \end{align}} is known to have a unique solution $\xbf \in \ensuremath{\mathbb{R}}^{n'}$ for all $\cbf' \in \ensuremath{\mathbb{R}}^{n'}$ and let $h'(\cbf')$ be this unique solution. Then, there exists a finite index set $\Lambda'$ and $\{(\Fbf'_{\lambda'}, \fbf'_{\lambda'}, \Gbf'_{\lambda'}, \gbf'_{\lambda'})\}_{\lambda' \in \Lambda'}$ such that \begin{align}\label{eq:h'} \notag h'(\cbf') = \Fbf'_{\lambda'} \cbf' + \fbf'_{\lambda'}, \quad &\forall \cbf' \in \Psi'_{\lambda'} \triangleq \setdef{\cbf'}{\Gbf'_{\lambda'} \cbf' + \gbf'_{\lambda'} \ge \mathbf{0}}, \\ &\forall \lambda' \in \Lambda', \end{align} and $\bigcup_{\lambda' \in \Lambda'} \Psi'_{\lambda'} = \ensuremath{\mathbb{R}}^{n'}$. \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol \end{lemma} A special case of Lemma~\ref{lem:affinity} is when $\Wbf_2 = \mathbf{0}$, where $h'$ becomes, like $h_N^\ssp$, the standard equilibrium map~\eqref{eq:h}. % Next, we characterize the global Lipschitzness property of the equilibrium maps. The proof is in~\ref{app:pf}. \begin{lemma}\longthmtitle{Piecewise affine equilibrium maps are globally Lipschitz}\label{lem:h-lip-gen} Let $h:\ensuremath{\mathbb{R}}^n \to \ensuremath{\mathbb{R}}^n$ be a piecewise affine function of the form \begin{align*} h(\cbf) = \Fbf_\lambda \cbf + \fbf_\lambda, \qquad &\forall \cbf \in \Psi_\lambda \triangleq \setdef{\cbf}{\Gbf_\lambda \cbf + \gbf_\lambda \ge \mathbf{0}}, \\ &\forall \lambda \in \Lambda, \end{align*} where $\Lambda$ is a finite index set and $\bigcup_{\lambda \in \Lambda} \Psi_\lambda = \ensuremath{\mathbb{R}}^n$. Then, $h$ is globally Lipschitz. \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol \end{lemma} We are now ready to generalize Theorem~\ref{thm:sp-inhib} to an $N$-layer architecture while at the same time relaxing several of its simplifying assumptions in favor of generality. \begin{theorem}\longthmtitle{Selective recruitment in multilayer % networks}\label{thm:multi} Consider the dynamics~\eqref{eq:dyn-multi}. If \begin{enumerate} % % % \item The reduced-order model (ROM) \new{\begin{align*} \hspace{-20pt} \tau_1 \dot {\bar \xbf}_1^\ssp = -\bar \xbf_1^\ssp + [\Wbf_{1, 1}^{\ssp\ssp} \bar \xbf_1^\ssp + \Wbf_{1, 2}^{\ssp\ssp} h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \bar \xbf_1^\ssp + \cbf_2^\ssp) + \cbf_1^\ssp]_\mathbf{0}^{\mbf_1^\ssp}, \end{align*}} of the first subnetwork has bounded solutions (recall $\xbf_1 \equiv \xbf_1^\ssp$ since $r_1 = 0$); \item For all $i = 2, \dots, N$, \new{\begin{align*} \hspace{-20pt} \tau_i \dot \xbf_i^\ssp(t) = &-\xbf_i^\ssp(t) + [\Wbf_{i, i}^{\ssp\ssp} \xbf_i^\ssp(t) \\ &+ \Wbf_{i, i+1}^{\ssp\ssp} h_{i+1}^\ssp(\Wbf_{i+1, i}^{\ssp\ssp} \xbf_i^\ssp(t) + \cbf_{i+1}^\ssp) + \cbf_i^\ssp]_\mathbf{0}^{\mbf_i^\ssp}, \end{align*}} is GES towards a unique equilibrium for any $\cbf_{i+1}^\ssp$ and~$\cbf_i^\ssp$; \end{enumerate} then there exists $\Kbf_i \in \ensuremath{\mathbb{R}}^{p_i \times n_i}$ and $\bar \ubf_i: \ensuremath{\mathbb{R}_{\ge 0}} \to \ensuremath{\mathbb{R}_{\ge 0}}^{p_i}, i \in \{2, \dots, N\}$ such that using the feedback-feedforward control \begin{align}\label{eq:u-gen} \ubf_i(t) = \Kbf_i \xbf_i(t) + \bar \ubf_i(t), \qquad i \in \{2, \dots, N\}, \end{align} we have, for any $0 < \underline t < \bar t < \infty$, \begin{subequations}\label{eq:multi-tik} \begin{align} \lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [\underline t, \bar t]} \|\xbf_i^\ssm(t)\| = \mathbf{0}, \qquad \forall i \in \{2, \dots, N\}, \end{align} and \begin{align} &\!\!\!\!\lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [0, \bar t]} \|\xbf_1^\ssp(t) - \bar \xbf_1^\ssp(t)\| = 0, \\ &\!\!\!\!\lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [\underline t, \bar t]} \|\xbf_2^\ssp(t) - h_2^\ssp(\Wbf_{2, 1}^{\ssp\ssp} \xbf_1^\ssp(t) + \cbf_2^\ssp)\| = 0, \\ \notag &\!\!\!\!\qquad \vdots \\ % % % % &\!\!\!\lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [\underline t, \bar t]} \!\!\|\xbf_N^\ssp(t) \!-\! h_N^\ssp(\Wbf_{N, N-1}^{\ssp\ssp} \xbf_{N-1}^\ssp(t) \!+\! \cbf_N^\ssp)\| \!=\! 0. \end{align} \end{subequations} \end{theorem} \begin{proof} For any $2 \times 2$ block-partitioned matrix $\Wbf$, we introduce the convenient notation $\Wbf^{\ell,\ssmp} \triangleq [\Wbf^{\ell\ssm} \ \Wbf^{\ell\ssp}]$ % % and $\Wbf^{\ssmp,\ell} \triangleq [(\Wbf^{\ssm\ell})^T \ (\Wbf^{\ssp\ell})^T]^T$ for $\ell = \ssm, \ssp$. Further, for any $i \in \{2, \dots, N\}$, let $\xbf_{1:i} = [\xbf_1^T \ \dots \ \xbf_i^T]^T$. % To begin with, let $\Kbf_N$ and $\bar \ubf_N$ be such that \begin{subequations}\label{eq:inhib-ineqs} \begin{align} \Bbf_N^\ssm \Kbf_N &\le -\Wbf_{N, N}^{\ssm,\ssmp}, \label{eq:inhib-ineqs1} \\ \bar \ubf_N(t) &\le -\Wbf_{N, N-1}^{\ssm,\ssmp} \xbf_{N-1}(t) - \cbf_N^\ssm, \qquad \forall t, \label{eq:inhib-ineqs2} \end{align} \end{subequations} Note that, if $p_N \ge r_N$, then~\eqref{eq:inhib-ineqs1} can be satisfied with equality. Otherwise, \eqref{eq:inhib-ineqs1} can still be satisfied since all the rows of $\Bbf_N^\ssm$ are nonzero, but may require excessive amounts of inhibition. Also, notice that $\bar \ubf_N$ is set by the subnetwork $N - 1$, which has access to $\xbf_{N-1}(t)$ and can thus fulfill~\eqref{eq:inhib-ineqs2}. As a result, the nodes in $\xbf_N^\ssm$ are fully inhibited and evolve according to % $\tau_N \dot \xbf_N^\ssm = -\xbf_N^\ssm$. % The overall dynamics become \new{\begin{align*} \tau_1 \dot \xbf_1 &\!=\! -\xbf_1 + [\Wbf_{1, 1} \xbf_1 + \Wbf_{1, 2} \xbf_2 + \cbf_1]_\mathbf{0}^{\mbf_1}, \\ &\ \ \vdots \\ \tau_{N-1} \dot \xbf_{N-1} &\!=\! -\xbf_{N-1} + [\Wbf_{N-1, N-1} \xbf_{N-1} \!+\! \Bbf_{N-1} \ubf_{N-1} \\ &\quad \!\!\!\!\!\!\!\!\!\!\!\!+\! \Wbf_{N-1, N} \xbf_N \!+\! \Wbf_{N-1, N-2} \xbf_{N-2} \!+\! \cbf_{N-1}]_\mathbf{0}^{\mbf_{N-1}}\!\!, \\ \epsilon_{N-1} \tau_{N-1} \dot \xbf_N^\ssm &\!=\! -\xbf_N^\ssm, \\ \epsilon_{N-1} \tau_{N-1} \dot \xbf_N^\ssp &\!=\! -\xbf_N^\ssp \!+\! [\Wbf_{N, N}^{\ssp,\ssmp} \xbf_N \!+\! \Wbf_{N, N-1}^{\ssp,\ssmp} \xbf_{N-1} \!+\! \cbf_N^\ssp]_\mathbf{0}^{\mbf_N^\ssp}. \end{align*}} Letting $\epsilon_{N-1} \to 0$, we get our first separation of timescales between $\xbf_N$ and $\xbf_{1:N-1}$, as follows. For any constant $\xbf_{N-1}$, the $\xbf_N$ dynamics is GES by assumption \emph{(ii)} and \cite[Lemma~A.2]{EN-JC:18-tacI}. Further, the equilibrium map $h_N = (\mathbf{0}_{r_N}, h_N^\ssp)$ % % % of the $N$'th subnetwork is globally Lipschitz by Lemmas~\ref{lem:affinity} and~\ref{lem:h-lip-gen}, and the entire vector field of network dynamics is globally Lipschitz due to the Lipschitzness of $[\, \cdot\, ]_0^m$. Therefore, it follows from~\cite[Prop 1]{VV:97} that for any $0 < \underline t < \bar t < \infty$, \begin{align*} &\lim_{\epsilon_{N-1} \to 0} \sup_{t \in [\underline t, \bar t]} \|\xbf_N^\ssm(t)\| = 0, \\ &\lim_{\epsilon_{N-1} \to 0} \sup_{t \in [\underline t, \bar t]} \|\xbf_N^\ssp(t) - h_N^\ssp(\Wbf_{N, N-1}^{\ssp,\ssmp} \xbf_{N-1}(t) + \cbf_N^\ssp)\| = 0, \\ % % % &\lim_{\epsilon_{N-1} \to 0} \sup_{t \in [0, \bar t]} \|\xbf_{1:N-1}(t) - \xbf_{1:N-1}^{(1)}(t)\| = 0. \end{align*} Here, $\xbf_{1:N-1}^{(1)}$ is the solution of the ``first-step ROM" \new{\begin{align*} &\tau_1 \dot \xbf_1^{(1)} = -\xbf_1^{(1)} + [\Wbf_{1, 1} \xbf_1^{(1)} + \Wbf_{1, 2} \xbf_2^{(1)} + \cbf_1]_\mathbf{0}^{\mbf_1}, \\ &\ \ \vdots \\ &\tau_{N-1} \dot \xbf_{N-1}^{(1)} = -\xbf_{N-1}^{(1)} + [\Wbf_{N-1, N-1} \xbf_{N-1}^{(1)} \\ &\qquad \qquad + \Wbf_{N-1, N}^{\ssmp,\ssp} h_N^\ssp(\Wbf_{N, N-1}^{\ssp,\ssmp} \xbf_{N-1}^{(1)}(t) + \cbf_N^\ssp) \\ &\qquad \qquad + \Wbf_{N-1, N-2} \xbf_{N-2}^{(1)} + \Bbf_{N-1} \ubf_{N-1} + \cbf_{N-1}]_\mathbf{0}^{\mbf_{N-1}}, \end{align*}} which results from replacing $\xbf_N$ with its equilibrium value. Except for technical adjustments, the remainder of the proof essentially follows by repeating this process $N - 2$ times. In particular, for $i = N-1, \dots, 2$, let $\Kbf_i$ and $\bar \ubf_i$ be such that {\interdisplaylinepenalty=10000 \begin{align*} \Bbf_i^\ssm \Kbf_i &\le -|\Wbf_{i, i}^{\ssm,\ssmp}| - |\Wbf_{i, i+1}^{\ssm\ssp}| \bar \Fbf_{i+1} |\Wbf_{i+1, i}^{\ssp,\ssmp}|, \\ \bar \ubf_i(t) &\le -\Wbf_{i, i-1}^{\ssm:} \xbf_{i-1}(t) - \cbf_i^\ssm, \qquad \forall t, \end{align*} } where $\bar \Fbf_i \in \ensuremath{\mathbb{R}}^{(n_i - r_i) \times (n_i - r_i)}$ is the entry-wise maximal gain of the map $h_i^\ssp$ over $\ensuremath{\mathbb{R}}^{n_i - r_i}$ (cf. Theorem~\ref{thm:multi-EUE-GES}). This results in the ``$(N - i)$'th-step ROM" \new{\begin{align*} &\tau_1 \dot \xbf_1^{(N-i)} \!=\! -\xbf_1^{(N-i)} \!\!+\! [\Wbf_{1, 1} \xbf_1^{(N-i)} \!\!+\! \Wbf_{1, 2} \xbf_2^{(N-i)} \!\!+\! \cbf_1]_\mathbf{0}^{\mbf_1}, \\ &\qquad \vdots \\ &\tau_{i-1} \dot \xbf_{i-1}^{(N-i)} = -\xbf_{i-1}^{(N-i)} + [\Wbf_{i-1, i-1} \xbf_{i-1}^{(N-i)} \\ &\qquad \qquad \qquad \ \ + \Wbf_{i-1, i} \xbf_i^{(N-i)} + \Wbf_{i-1, i-2} \xbf_{i-2}^{(N-i)} \\ &\qquad \qquad \qquad \ \ + \Bbf_{i-1} \ubf_{i-1} + \cbf_{i-1}]_\mathbf{0}^{\mbf_{i-1}}, \\ &\epsilon_{i-1} \tau_{i-1} \dot \xbf_i^{(N-i)\ssm} = -\xbf_i^{(N-i)\ssm}, \\ &\epsilon_{i-1} \tau_{i-1} \dot \xbf_i^{(N-i)\ssp} = -\xbf_i^{(N-i)\ssp} + [\Wbf_{i, i}^{\ssp,\ssmp} \xbf_i^{(N-i)\ssp} \\ &\qquad \qquad \qquad \quad \ + \Wbf_{i, i+1}^{\ssmp,\ssp} h_{i+1}^\ssp(\Wbf_{i+1, i}^{\ssp,\ssmp} \xbf_i^{(N-i)}(t) \!+\! \cbf_{i+1}^\ssp) \\ &\qquad \qquad \qquad \quad \ + \Wbf_{i, i-1}^{\ssp,\ssmp} \xbf_{i-1}^{(N-i)} + \cbf_i^\ssp]_\mathbf{0}^{\mbf_i^\ssp}. \end{align*}} Similarly to above, invoking~\cite[Prop 1]{VV:97} then ensures that \begin{align*} &\lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [\underline t, \bar t]} \|\xbf_i^{(N-i)\ssm}(t)\| = 0, \\ &\lim_{\epsilonb \to \mathbf{0}} \sup_{t \in [\underline t, \bar t]} \|\xbf_i^{(N-i)\ssp}(t) \!-\! h_i^\ssp(\Wbf_{i, i-1}^{\ssp,\ssmp} \xbf_{i-1}^{(N-i)}(t) \!+\! \cbf_i^\ssp)\| = 0, \\ &\lim_{\epsilonb \to 0} \sup_{t \in [0, \bar t]} \|\xbf_{1:i-1}^{(N-i)}(t) - \xbf_{1:i-1}^{(N-i+1)}(t)\| = 0. \end{align*} Note that, after every invocation of~\cite[Prop 1]{VV:97}, the super-index inside the parenthesis increases by $1$, showing one more replacement of a fast dynamics by its equilibrium state. In particular, after the $(N - 1)$'th invocation of~\cite[Prop 1]{VV:97}, we reach $\xbf_1^{(N-1)\ssp}$, which is the same as $\bar \xbf_1^\ssp$ in the statement. Together, these results (and sufficiently many applications of the triangle inequality and Lemma~\ref{lem:h-lip-gen}) ensure~\eqref{eq:multi-tik}. \end{proof} \new{An instructive difference, by design, between Theorems~\ref{thm:sp-inhib} and~\ref{thm:multi} is the separate treatment of feedforward and feedback inhibition in the former versus the combination of the two in the latter. While the separate treatment in Theorem~\ref{thm:sp-inhib} is conceptually simpler and highlights the theoretical difference between the two inhibitory mechanisms, the combination in Theorem~\ref{thm:multi} results in more flexibility and less conservativeness:} in pure feedforward inhibition, countering local excitations requires monotone boundedness and a sufficiently large $\bar \ubf$ providing inhibition under the worst-case scenario, a goal that is achieved more efficiently using feedback. On the other hand, pure feedback inhibition needs to dynamically cancel local excitations at all times and is also unable to counter the effects of constant background excitation, limitations that are easily addressed when combined with feedforward inhibition. \new{Similar to Theorem~\ref{thm:sp-inhib} (cf. Remark~\ref{re:assumptions-reasonable}), assumption~\emph{(ii)} of Theorem~\ref{thm:multi} is its only critical requirement which we next relate to the joint structure of the subnetworks.}% \begin{theorem}\longthmtitle{Sufficient condition for existence and uniqueness of equilibria and GES in multilayer linear-threshold networks}\label{thm:multi-EUE-GES} Let $h:\ensuremath{\mathbb{R}}^n \to \ensuremath{\mathbb{R}}^n$ be a piecewise affine function of the form \begin{align}\label{eq:h-aff-form} \notag h(\cbf) = \Fbf_\lambda \cbf + \fbf_\lambda, \qquad &\forall \cbf \in \Psi_\lambda \triangleq \setdef{\cbf}{\Gbf_\lambda \cbf + \gbf_\lambda \ge \mathbf{0}}, \\ &\forall \lambda \in \Lambda, \end{align} where $\Lambda$ is a finite index set and $\bigcup_{\lambda \in \Lambda} \Psi_\lambda = \ensuremath{\mathbb{R}}^n$. Further, let $\bar \Fbf \triangleq \max_{\lambda \in \Lambda}|\Fbf_\lambda|$ be the matrix whose elements are the maximum of the corresponding elements from $\{|\Fbf_\lambda|\}_{\lambda \in \Lambda}$. For arbitrary matrices $\Wbf_\ell$, $ \ell = 1, 2, 3$, if $ \rho\big(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|\big) < 1$, then the linear-threshold dynamics \new{\begin{align*} \tau \dot \xbf(t) = -\xbf(t) + [\Wbf_1 \xbf(t) + \Wbf_2 h(\Wbf_3 \xbf(t) + \bar \cbf) + \cbf]_\mathbf{0}^\mbf, \end{align*}} is GES towards a unique equilibrium for all $\bar \cbf$ and $\cbf$. \end{theorem} \begin{proof} We use the same proof technique as in~\cite[Prop. 3]{JF-KPH:96}. For simplicity, assume that $|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|$ is irreducible (i.e., the network topology induced by it is strongly connected)% \footnote{If $|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|$ is not irreducible, it can be ``upper-bounded'' by the irreducible matrix $|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3| + \mu \mathbf{1}_n \mathbf{1}_n^T$, with $\mu > 0$ sufficiently small such that $\rho(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3| + \mu \mathbf{1}_n \mathbf{1}_n^T) < 1$. The same argument can then be employed for this upper bound.}. % Then, the left Perron-Frobenius eigenvector $\alphab$ of $|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|$ has positive entries~\cite[Fact 4.11.4]{DSB:09}, making the map $\|\cdot\|_\alphab: \vbf \to \|\vbf\|_\alphab \triangleq \alphab^T |\vbf|$ a norm on $\ensuremath{\mathbb{R}}^n$. Further, it can be shown, similar to the proof of Lemma~\ref{lem:h-lip-gen}, that for all $\cbf_1, \cbf_2 \in \ensuremath{\mathbb{R}}^n$, $|h(\cbf_1) - h(\cbf_2)| \le \bar \Fbf |\cbf_1 - \cbf_2|$, where the inequality is entrywise. Thus, for any $\xbf, \hat \xbf \in \ensuremath{\mathbb{R}}^n$, \new{\begin{align*} &\big\|[\Wbf_1 \xbf + \Wbf_2 h(\Wbf_3 \xbf + \wbf) + \cbf]_\mathbf{0}^\mbf \\ &\qquad \qquad \qquad - [\Wbf_1 \hat \xbf + \Wbf_2 h(\Wbf_3 \hat \xbf + \wbf) + \cbf]_\mathbf{0}^\mbf\big\|_\alphab \\ &= \alphab^T \big|[\Wbf_1 \xbf + \Wbf_2 h(\Wbf_3 \xbf + \wbf) + \cbf]_\mathbf{0}^\mbf \\ &\qquad \qquad \qquad - [\Wbf_1 \hat \xbf + \Wbf_2 h(\Wbf_3 \hat \xbf + \wbf) + \cbf]_\mathbf{0}^\mbf\big| \\ &\le \alphab^T \big|\Wbf_1(\xbf - \hat \xbf) + \Wbf_2(h(\Wbf_3 \xbf + \wbf) - h(\Wbf_3 \hat \xbf + \wbf))\big| \\ &\le \alphab^T \big(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|\big) |\xbf - \hat \xbf| \\ &= \rho\big(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|\big) \alphab^T |\xbf - \hat \xbf| \\ &= \rho\big(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|\big) \|\xbf - \hat \xbf\|_\alphab. \end{align*}} \new{This proves that $\xbf \mapsto [\Wbf_1 \xbf + \Wbf_2 h(\Wbf_3 \xbf + \wbf) + \cbf]_\mathbf{0}^\mbf$ is a contraction (on $\ensuremath{\mathbb{R}_{\ge 0}}^n$ if $\mbf = \infty \mathbf{1}_n$ or on $\prod_i [0, m_i]$ if $\mbf < \infty \mathbf{1}_n$) and has a unique fixed point, denoted $\xbf^*$, by the Banach Fixed-Point Theorem~\cite[Thm 9.23]{WR:76}.} To show GES, let $\xib(t) \triangleq (\xbf(t) - \xbf^*) e^t$, satisfying \begin{align}\label{eq:xi-dot-2} \tau \dot \xib(t) = \Mbf(t) \Wbf \xib(t), \end{align} where $\Mbf(t)$ is a diagonal matrix with diagonal entries \begin{align*} M_{ii}(t) \triangleq \frac{\big([\Wbf_1 \xbf(t) + \Wbf_2 h(\Wbf_3 \xbf(t) + \wbf) + \cbf]_\mathbf{0}^\mbf - \xbf^*)_i}{\xi_i(t)} & \end{align*} if $\xi_i(t) \neq 0$ and $M_{ii}(t) \triangleq 0$ otherwise. Then \begin{align*} |\Mbf(t)| \le |\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|, \qquad \forall t \ge 0, \end{align*} where the inequality is entry-wise. Then, by using~\cite[Lemma]{KPH-DK:87} (which is essentially a careful application of Gronwall-Bellman's Inequality~\cite[Lemma A.1]{HKK:02} to~\eqref{eq:xi-dot-2}), \begin{align*} &\|\xib(t)\|_\alphab \le \|\xib(0)\|_\alphab e^{\rho(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|) t} \\ &\Rightarrow \|\xbf(t) \!-\! \xbf^*\|_\alphab \!\!\le \|\xbf(0) \!-\! \xbf^*\|_\alphab e^{-(1 - \rho(|\Wbf_1| + |\Wbf_2| \bar \Fbf |\Wbf_3|)) t}, \end{align*} establishing GES by the equivalence of norms on $\ensuremath{\mathbb{R}}^n$. \end{proof} Note that Theorem~\ref{thm:multi-EUE-GES} applies to each layer of~\eqref{eq:dyn-multi} separately. When put together, Theorem~\ref{thm:multi}\emph{(ii)} is satisfied if \begin{align}\label{eq:rec-cond-sim} \notag &\rho\big(|\Wbf_{2, 2}^{\ssp\ssp}| + |\Wbf_{2, 3}^{\ssp\ssp}| \bar \Fbf_3^\ssp |\Wbf_{3, 2}^{\ssp\ssp}|\big) < 1, \\ \notag &\qquad \vdots \\ \notag &\rho\big(|\Wbf_{N-1, N-1}^{\ssp\ssp}| + |\Wbf_{N-1, N}^{\ssp\ssp}| \bar \Fbf_N^\ssp |\Wbf_{N, N-1}^{\ssp\ssp}|\big) < 1, \\ &\rho\big(|\Wbf_{N, N}^{\ssp\ssp}|\big) < 1, \end{align} where $\bar \Fbf_i^\ssp, i = 2, \dots, N$ is the matrix described in Theorem~\ref{thm:multi-EUE-GES} corresponding to $h_i^\ssp$, and the affine form~\eqref{eq:h-aff-form} of $h_i^\ssp$ is computed recursively using Lemma~\ref{lem:affinity}. \nnew{Moreover, if $\mbf^\ssp_1 = \infty \mathbf{1}_{r_1}$, then $\rho\big(|\Wbf_{1, 1}^{\ssp\ssp}| + |\Wbf_{1, 2}^{\ssp\ssp}| \bar \Fbf_2^\ssp |\Wbf_{2, 1}^{\ssp\ssp}|\big) < 1$ serves as a sufficient condition for Theorem~\ref{thm:multi}\emph{(i)} (which is trivial if $\mbf^\ssp_1 < \infty \mathbf{1}_{r_1}$).} \section{Case Study: Selective Listening in Rodents} We present an application of our framework to a specific real-world example of goal-driven selective attention using measurements of single-neuron activity in the brain. Beyond the conceptual illustration of our results in Example~\ref{ex:osc} above, we argue that the cross-validation of theoretical results with real data performed here is a necessary step to make a credible case for neuroscience research and significantly enhances the relevance of the developed analysis. We have been fortunate to have access to data from a novel and carefully designed experimental paradigm~\cite{CCR-MRD:14,CCR-MRD:14-crcns} that involves goal-driven selective listening in rodents and displays the key features of hierarchical selective recruitment noted~here. \subsection{Description of Experiment and Data} A long standing question in neuroscience involves our capability to selectively listen to specific sounds in a crowded environment~\cite{ECC:53,AWB:15}. % To understand the neuronal basis of this phenomena, the work~\cite{CCR-MRD:14} has rats simultaneously presented with two sounds and trains them to selectively respond to one sound while actively suppressing the distraction from the other. % In each trial, the animal simultaneously hears a white noise burst and a narrow-band warble. The noise burst may come from the left or the right while the warble may have low or high pitch, both chosen at random. Which of the two sounds (noise burst or warble) is relevant and which is a distraction depends on the ``rule'' of the trial: in ``localization'' (LC) and ``pitch discrimination'' (PD) trials, the animal has to make a motor choice based on the location of the noise burst (left/right) or the pitch of the warble (low/high), resp., to receive a reward. Each rat performs several blocks of LC and PD trials during each session (with each block switching randomly between the 4 possible stimulus pairs), requiring it to quickly switch its response following the rule changes. While the rats perform the task, spiking activity of single neurons is recorded in two brain areas: the primary auditory cortex (A1) and the medial prefrontal cortex (PFC). A1 is the first region in the cortex that receives auditory information (from subcortical areas and ears), thus forming a (relatively) low level of the hierarchy. PFC is composed of multiple regions that form the top of the hierarchy, and serve functions such as imagination, planning, decision-making, and attention~\cite{JF:15}. Spike times of 211 well-isolated and reliable neurons are recorded in $5$ rats, $112$ in PFC and $99$ in A1, see~\cite{CCR-MRD:14-crcns}. \new{Using statistical analysis, it was shown in~\cite{CCR-MRD:14} that} (i) the rule of the trial and the stimulus sounds are more strongly encoded by PFC and A1 neurons, resp., (ii) electrical disruption of PFC significantly impairs task performance, and (iii) PFC activity temporally precedes A1 activity. These findings are all consistent with a model where PFC controls the activity of A1 based on the trial rule in order to achieve GDSA. We next build on these observations to define an appropriate network structure and rigorously analyze it using HSR. \subsection{Choice of Neuronal Populations} To form meaningful populations among the recorded neurons, we perform three classifications of them: \begin{enumerate}[wide] \item first, we classify the neurons into excitatory and inhibitory. The standard procedure for this classification is based on the spike waveform of each neuron: excitatory neurons have slower and wider spikes while inhibitory neurons have faster and narrower ones~\cite{RMB-DJS:02}. \new{Using standard k-means clustering on the $24$-dimensional spike waveform time-series, we identify $174$ excitatory and $36$ inhibitory neurons% \footnote{The type of one neuron could not be identified with confidence and was discarded from further analysis.} (Figure~\ref{fig:ei}(a)).} These clusters conform with spike width difference of excitatory and inhibitory neurons (Figure~\ref{fig:ei}(b)) and the \new{common expectation that about $80\%$ of mammalian cortical neurons are excitatory}. \begin{figure} \subfloat[]{ \includegraphics[width=0.5\linewidth]{tsne_ei_clear} } \subfloat[]{ \parbox{0.4\linewidth}{ \vspace*{-124pt} \includegraphics[width=0.95\linewidth]{waveforms_ei_clear} }} \caption{\new{Excitatory/inhibitory classification of neurons. (a) Clusters of spike waveforms. For illustration, clusters are shown in the two-dimensional space arising from t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction. (b) The spike waveforms of clustered neurons. As expected, the inhibitory neurons have faster and narrower spikes. }}\label{fig:ei} \vspace*{-1.5ex} \end{figure} \item Second, we classify the PFC neurons based on their rule-encoding (RE) property. This classification was also done in~\cite{CCR-MRD:14}, so we briefly review the method for completeness. A neuron is said to have a RE property if its firing rate is significantly different during the LC and PD trials \emph{before the stimulus onset}. % In the absence of stimulus, any such difference is attributable to the animal's knowledge of the task rule (i.e., which upcoming stimulus it has to pay attention to in order to get the reward). Thus, it is standard to assess neurons' RE property during the \emph{hold period}, namely, the time interval between the initiation of each trial and the stimulus onset of that trial. Therefore for each PFC neuron, we calculate its mean firing rate during the hold period of each trial and then statistically compare the results for LC and PD trials ($p < 0.05$, \new{randomization test}). Among the $112$ neurons in PFC, $40$ encoded for LC while $44$ encoded for PD (the remaining PFC neurons with no RE property are discarded). \item Finally, we classify the A1 neurons based on their evoked response (ER) property. In contrast to RE, a neuron has an ER property if its firing rate is significantly different in response to the white noise (LC stimulus) and warble (PD stimulus) \emph{after the stimulus onset}. % Since the white noise and warble are always presented simultaneously, it is not possible to make such a distinction based on normal trials. However, before each LC or PD block, the animal is only presented with the respective stimulus for a few \emph{cue trials} (which is how the animal realizes the rule change). Thus, for each A1 neuron, we compare its mean firing rate during the \emph{listening period} of each cue trial (namely, the interval between the stimulus onset and the time that the animal commits to a decision) and statistically compare the distribution of the results for LC and PD cue trials ($p < 0.05$, \new{randomization test}). Among the $99$ A1 neurons, $21$ had an ER for LC while another $21$ had an ER for PD (the remaining A1 neurons with no ER property are discarded from further analysis). % \end{enumerate} \begin{remark}\longthmtitle{RE vs. ER detection} {\rm It is noteworthy that a smaller fraction of PFC and A1 neurons also have ER and RE properties, resp. However, it is expected from systems neuroscience that these properties arise from the PFC-A1 interaction, as auditory and attention/decision making information disseminate from A1 and PFC, resp. This motivates our classification of A1 and PFC neurons based on ER and RE, resp., and their reciprocal connection in the proposed network structure below. Further, we note that our ER detection has a difference with respect to~\cite{CCR-MRD:14}. In~\cite{CCR-MRD:14}, the difference between the post-stimulus and pre-stimulus firing rates (the latter being RE) is used for ER detection, with the motivation of removing the portion of post-stimulus firing rate that is due to RE (and thus independent of stimulus). However, this relies on the strong assumption that the RE and ER responses superimpose linearly, which we found likely not to be true based on the statistical analysis of the present dataset, perhaps as RE may have driven neurons close to their maximum firing rate, leaving little room for \emph{additional} ER. We thus use the complete post-stimulus firing rate for ER detection, as above. \relax\ifmmode\else\unskip\hfill\fi\oprocendsymbol} \end{remark} As a result of the classifications described above, we group the neurons into $8$ populations based on the PFC/A1, excitatory/inhibitory, and LC/PD classifications. The firing rate of each population (as a function of time) is then calculated as follows. For each neuron and each trial, the interval $[-10, 10]$ (with time $0$ corresponding to stimulus onset) is decomposed into $100^\text{ms}$-wide bins and the firing rate of each bin (spike count divided by bin width) is assigned to the bin's center time. This time series is then averaged over all trials with the same stimulus pair and all the neurons within each population, and finally smoothed with a Gaussian kernel with $1^\text{s}$ standard deviation. This results in one firing rate time series for each neuron and each stimulus pair. We limit our choice of stimulus pairs as follows. Recall that each of LC and PD blocks contains 4 stimulus pairs (left-low, left-high, right-low, right-high). % In each block, these 4 pairs are divided into two \emph{go} and two \emph{no-go} pairs. When the animal hears a go stimulus pair, his correct response is to go to a nearby food port to receive his reward. In no-go trials, the correct response is simply inaction (action is punished by a delay before the animal can do the next trial). Due to strong motor and reward-consumption artifacts in go trials (cf.~\cite[Fig. S4]{CCR-MRD:14}), we limit our analysis here to no-go trials. Further, we also discard the no-go stimulus pair that is shared between LC and PD blocks, since the correct decision (no-go) is independent of the block and thus does not require selective attention. Hence, our analysis only involves one firing rate time series for each neuronal population in each~block. \subsection{Network Binary Structure} We next describe our proposed network binary structure% \footnote{We here make a distinction between the binary structure of the network, composed of only the connectivity pattern among nodes, and its full structure, that also includes the connection weights.}. In each of the two regions (PFC and A1), the $4$ populations are connected to each other according to the following physiological properties (see~\cite{PSG:95,AFTA-MJW-CDP:12,PS-GT-RL-EHB:98} and~\cite{GKW-RA-BL-HWT-LIZ:08,HKK-SKA-JSI:17,PS-GT-RL-EHB:98} for evidence of these properties in PFC and A1, resp.): \begin{enumerate} \item each excitatory population projects to (i.e., makes synapses on) the inhibitory population with the same LC/PD preference (RE in PFC or ER in A1); \item neurons in each excitatory population project to each other (captured by the excitatory self-loops in Figure~\ref{fig:struct}). \item each inhibitory population projects to the populations (both excitatory and inhibitory) with \emph{opposite} LC/PD preference (the so-called \emph{lateral inhibition} property); \end{enumerate} While within-region connections are both excitatory and inhibitory, between-region connections in the cortex (including PFC and A1) are almost entirely excitatory, completing the binary structure shown in Figure~\ref{fig:struct}. \begin{figure} \begin{center} \begin{tikzpicture} \node[scale=0.86] (all) { \begin{tikzpicture} \node (fig) {\includegraphics[width=0.5\linewidth]{net_struct}}; \node[left of=fig, xshift=-40pt, yshift=-70pt] (mgn-r) {}; \node[left of=mgn-r, xshift=-25pt] (mgn-l) {}; \draw[dashed, black] (mgn-l) to (mgn-r); \node[above of=mgn-l, yshift=45pt] (a1-r) {}; \node[above of=mgn-r, yshift=45pt] (a1-l) {}; \draw[dashed, black] (a1-l) to (a1-r); \node[right of=fig, xshift=40pt, yshift=-70pt] (n4-l) {}; \node[right of=n4-l, xshift=25pt] (n4-r) {}; \draw[dashed, black] (n4-l) to (n4-r); \node[above of=n4-l, yshift=10pt] (n3-l) {}; \node[above of=n4-r, yshift=10pt] (n3-r) {}; \draw[dashed, black] (n3-l) to (n3-r); \node[above of=n3-l, yshift=40pt] (n2-l) {}; \node[above of=n3-r, yshift=40pt] (n2-r) {}; \draw[dashed, black] (n2-l) to (n2-r); % \node[above of=fig, xshift=-25pt, yshift=63pt] (lc-b) {}; % \node[above of=lc-b, yshift=5pt] (lc-t) {}; % \draw[dashed, black] (lc-b) to (lc-t); % \node[above of=fig, xshift=25pt, yshift=63pt] (pd-b) {}; % \node[above of=pd-b, yshift=5pt] (pd-t) {}; % \draw[dashed, black] (pd-b) to (pd-t); % \node[left of=fig, xshift=-67pt, yshift=-85pt, scale=0.9] (mgn) {Thalamus}; % \node[above of=mgn, yshift=25pt, scale=0.9] (a1) {A1}; % \node[above of=a1, yshift=57pt, scale=0.9] (pfc) {PFC}; % \node[right of=fig, xshift=67pt, yshift=-85pt, scale=0.9] (n4) {$\mathcal{N}_4$}; % \node[above of=n4, yshift=5pt, scale=0.9] (n3) {$\mathcal{N}_3$}; % \node[above of=n3, yshift=25pt, scale=0.9] (n2) {$\mathcal{N}_2$}; % \node[above of=n2, yshift=40pt, scale=0.9] (n1) {$\mathcal{N}_1$}; % \node[above of=fig, xshift=-46pt, yshift=80pt, scale=0.9] (lc) {LC}; % \node[right of=lc, xshift=17pt, scale=0.9] (t) {Time}; % \node[right of=t, xshift=17pt, scale=0.9] (pd) {PD}; % \node[right of=pd, xshift=45pt, yshift=0pt, draw, line width=0.5pt, inner sep=3pt, scale=0.7] (leg) { \begin{tikzpicture} \node (e) {\includegraphics[width=10pt]{E_node}}; \node[right of=e, xshift=35pt, yshift=-2pt, inner sep=0pt] {\parbox{110pt}{\flushleft Manifest Excitatory Node}}; \node[below of=e, yshift=15pt] (i) {\includegraphics[width=8pt]{I_node}}; \node[right of=i, xshift=35pt, yshift=3pt, inner sep=0pt] {\parbox{110pt}{\flushleft Manifest Inhibitory Node}}; \node[below of=i, yshift=16pt] (u) {\includegraphics[width=10pt]{E_node_hollow}}; \node[right of=u, xshift=5pt] {Input Node}; \end{tikzpicture} }; \end{tikzpicture}}; \end{tikzpicture} \end{center} \caption{Proposed network binary structure. % % % The physiological region, hierarchical layer, and encoding properties of nodes are indicated on the left, right, and above the figure, resp.}\label{fig:struct} \vspace*{-1.5ex} \end{figure} \subsubsection*{Hierarchical Structure} To apply the HSR framework to the network of Figure~\ref{fig:struct}, we still need to assign the nodes to hierarchical layers. This assignment is in general arbitrary except for two critical requirements, (i) the existence of timescale separation between layers and (ii) the existence of both excitatory and inhibitory projections from any layer to the layer below (to allow for simultaneous inhibition and recruitment). The trivial choice here is to consider each region as a layer, which also satisfies (i) (since PFC has slower dynamics than A1) but not (ii) (since there would be no inhibitory connection between regions). We thus propose an alternative 3-layer choice, as shown in Figure~\ref{fig:struct}.% \footnote{The bottom-most layer $\mathcal{N}_4$ represents ``external'' inputs from sub-cortical areas. Since we have no recordings from these areas, we do not consider any dynamics for $\mathcal{N}_4$ and accordingly do not include it in HSR analysis.} This choice clearly satisfies (ii), and we next show that it also satisfies (i). \subsubsection*{Computation of Timescales} To assess the intrinsic timescales of each population, we employ the common method in neuroscience based on the decay rate of the correlation coefficient~\cite{JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14,RC-KK-MG-HK-XW:15}. In brief, for each neuron $\ell$, we partition the time window \emph{before} the stimulus onset% \footnote{In general, the time interval used for timescale estimation should not include stimulus presentation in order to reduce the effects of external factors on the internal neuronal dynamics.} into small bins ($200^\text{ms}$-wide here) and compute the smoothed mean firing rate of this neuron during each bin and each trial. This yields a set $\{r_{i, k}^\ell\}_{i, k, \ell}$, where $r_{i, k}^\ell$ denotes the mean firing rate of neuron $\ell$ in the $k$'th time bin of trial $i$. The Pearson correlation coefficient between two time bins $k_1$ and $k_2$ is estimated~as \begin{align*} \rho^\ell_{k_1, k_2} = \frac{\sum_i (r_{i, k_1}^\ell - \bar r_{k_1}^\ell) (r_{i, k_2}^\ell - \bar r_{k_2}^\ell)}{\sqrt{\sum_i (r_{i, k_1}^\ell - \bar r_{k_1}^\ell)^2 \sum_i (r_{i, k_2}^\ell - \bar r_{k_2}^\ell)^2}} \in [-1, 1], \end{align*} where $\bar r_k^\ell$ is the average of $r_{i, k}^\ell$ across all the trials for neuron $\ell$. Let $\rho_k^\ell$ be the average of $\rho_{k_1, k_2}^\ell$ over all $k_1, k_2$ such that $|k_1 - k_2| = k$ and $\bar \rho_k^p$, for any population $p$, be the average of $\rho_k^\ell$ for all the neurons $\ell$ in the population $p$. Figure~\ref{fig:taus} shows this function for populations of excitatory and inhibitory neurons in PFC and A1 (we do not split the neurons based on their LC/PD preference because it is not relevant for timescale separation). Fitting $\bar \rho_k^p$ by an exponential function of the form $A e^{-k/\tau}$ gives an estimate of the intrinsic timescale $\tau$ of this population, which becomes exact for spikes generated by a Poisson point process under certain regularity conditions~\cite{JDM-AB-DJF-RR-JDW-XC-CP-TP-HS-DL-XW:14}. Here, we use the range of $k$ values for which the decay of $\bar \rho_k^p$ is approximately exponential for calculating the fit. As seen in Figure~\ref{fig:taus}, there is a clear timescale separation between the layer of A1 excitatory neurons, the layer of A1 inhibitory and PFC excitatory neurons, and the layer of PFC inhibitory neurons, satisfying the requirement (i) above.% \footnote{Note that this method inherently underestimates the timescale separation between layers due to the mutual dynamical interactions between them.} \begin{figure} \centering \includegraphics[width=0.9\linewidth]{taus} \caption{Timescale separation among the layers $\mathcal{N}_1$, $\mathcal{N}_2$, and $\mathcal{N}_3$ in Figure~\ref{fig:struct}. The circles illustrate the values of the average auto-correlation coefficient $\bar \rho_k^p$ as a function of time lag $k$, whereas the lines represent the best exponential fit over the range of time lags where each $\bar \rho_k^p$ decays exponentially (note the logarithmic scale on the y-axis).}\label{fig:taus} \vspace*{-1.5ex} \end{figure} \subsubsection*{Exogenous Inputs and Latent Nodes} The last step in specifying the binary structure of the network involves the exogenous inputs to the prescribed neuronal populations (nodes). Clearly, nodes at the bottom layer (layer $3$) receive auditory inputs from subcortical areas which we represent as two input signals $x_1^4$ and $x_2^4$ coming from layer $4$ and corresponding to the white noise and warble, resp. Both these signals are constructed by smoothing a square pulse that equals $1$ during stimulus presentation and $0$ otherwise with the same Gaussian window used for smoothing the firing rate time-series. The choice of the inputs to the PFC populations is more intricate. PFC is itself composed of a complex network of several regions, each involved in some aspects of high-level cognitive functions. The RE properties of the recorded PFC populations is only one outcome of such complex PFC dynamics that also host the animal's overall understanding of how the task works, his perception of time, etc. In order to capture the effects of such unrecorded PFC dynamics, we consider 3 additional excitatory PFC populations, as follows. Two input populations $x_3^1$ and $x_4^1$ simply encode the rule of each block% \footnote{Note that this static response is different from, and much simpler than, the RE of the recorded PFC neurons, which is greatly dynamic.}% : \begin{align*} x_3^1 \equiv \begin{cases} 1, & \text{if in LC block}, \\ 0, & \text{if in PD block}, \end{cases} \quad x_4^1 \equiv \begin{cases} 0, & \text{if in LC block}, \\ 1, & \text{if in PD block}. \end{cases} \end{align*} Populations with such a sustained constant activity only as a function of task parameters are indeed observed during GDSA in PFC~\cite{NPB-MTH-EMD-RD:15}. The third additional PFC population encodes the time relative to the stimulus onset, which is critical for the functioning of the recorded PFC populations. Among the various forms of encoding time, we consider a population $x_5^1$ with firing rate \begin{align*} x_5^1(t) = \begin{cases} |t_0| - t & t \in [t_0, 0), \\ 0 & t \in (0, t_f], \end{cases} \end{align*} where $[t_0, t_f] = [-7, 7]$ is the duration of each trial, since populations with such activity patterns have been observed in PFC~\cite{AM-HM-KS-YM-JT:09}.% \footnote{Even though both~\cite{NPB-MTH-EMD-RD:15} and~\cite{AM-HM-KS-YM-JT:09} involve primates, populations with similar activity patterns are expected to exist in rodents.} % Since these three populations have very slow dynamics but are excitatory, following the same logic as before, we position them in the layer 1 together with the recorded inhibitory PFC populations $x_1^1$, $x_2^1$. % Finally, to capture the effects of the large populations of neurons whose activity is not recorded, we consider one \emph{latent} node for each of the $8$ \emph{manifest} nodes in the network% \footnote{A node is called \emph{manifest} if its activity is recorded during the experiment and \emph{latent} otherwise.} % with the same in- and out-neighbors as their respective manifest node (latent nodes are not plotted in Figure~\ref{fig:struct} to avoid cluttering the network structure). We let $\{x_{1, j}\}_{j = 6, 7}$, $\{x_{2, j}\}_{j = 5}^8$, and $\{x_{3, j}\}_{j = 3, 4}$ denote these nodes in $\mathcal{N}_1$, $\mathcal{N}_2$, and $\mathcal{N}_3$, resp. \subsection{Identification of Network Parameters} Having established the binary structure of the network, we next seek to determine its unknown parameters $\Wbf^{i, j}$. While there are physiological methods for measuring the synaptic weight between a pair of neurons in vitro, they are not applicable in vivo and thus not available for our dataset. Also, our nodes consist of several neurons, making their aggregate synaptic weight an abstract quantity. Therefore, we resort to system identification/machine learning techniques to ``learn'' the structure of the network given its input-output signals. For this purpose, the choice of objective function is crucial, for which we propose \begin{align}\label{eq:f} f(z) &= f_\text{SSE}(z) + \gamma_1 f_\text{corr}(z) + \gamma_2 f_\text{var}(z), \\ \notag f_\text{SSE}(z) &= \sum_{\ell = 1}^2 \sum_{i = 1}^3 \sum_{j = 1}^{n_{m,i}} \sum_k (\hat x_{i, j}(kT; \ell) - x_{i, j}(kT; \ell))^2, \\% \notag f_\text{corr}(z) &= 1 - \frac{1}{2 n_m} \sum_{\ell = 1}^2 \frac{1}{n_m} \sum_{i = 1}^3 \sum_{j = 1}^{n_{m,i}} \frac{1}{K - 1} \\ \notag &\qquad \times \sum_{k = 1}^K \frac{(\hat x_{i, j}(kT; \ell) - \hat \mu_{i, j, \ell}) (x_{i, j}(kT; \ell) - \mu_{i, j, \ell})}{\hat \sigma_{i, j, \ell} \sigma_{i, j, \ell}}, \\ \notag f_\text{var}(z) &= \Big(\sum_{\ell = 1}^2 \sum_{i = 1}^3 \sum_{j = 1}^{n_{m,i}} (\hat \sigma_{i, j, \ell} - \sigma_{i, j, \ell})^4 \Big)^{1/4}, \end{align} where, \begin{itemize}[wide] \item[--] $z$ is the vector of all unknown network parameters consisting of not only the synaptic weights but also the time constants $\tau_i$, the background inputs $\cbf_i$, and the initial states $\xbf_i(0), i = 1, 2, 3$; \item[--] $n_{m, i}$ is the number of manifest nodes in layer $i$ (so $n_{m, 1} = 2, n_{m, 2} = 4, n_{m, 3} = 2$) and $n_m = 8$ is the total number of manifest nodes; \item[--] $x_{i, j}(t; \ell)$ is the measured state of $j$'th node in the $i$'th layer in response to the $\ell$'th stimulus at time $t$ (where $\ell = 1$ indicates the LC block and $\ell = 2$ the PD block) and $\hat x_{i, j}(t; \ell)$ is its model estimate; \item[--] $T = 0.1$ is the sampling time \new{and $K$ is the total number of samples of each signal}; and \item[--] $\mu_{i, j, \ell}, \sigma_{i, j, \ell}, \hat \mu_{i, j, \ell}, \hat \sigma_{i, j, \ell}$ are the means and standard deviations of $x_{i, j}(\cdot; \ell)$ and $\hat x_{i, j}(\cdot; \ell)$, resp. \end{itemize} The rationale behind% ~\eqref{eq:f} is as follows. $f_\text{SSE}(z)$ is the standard sum of squared error (SSE). In HSR, an important property of nodal state trajectories is the sign of their derivatives, which \emph{transiently} indicate recruitment (positive derivative) or inhibition (negative derivative). This is captured by the average correlation coefficient $f_\text{corr}(z)$, which is added to $f_\text{SSE}(z)$ to enforce similar recruitment and inhibition patterns between measured states and their estimates. Nevertheless, correlation coefficient between a pair of signals is invariant to the amount of variation in them, requiring us to add the third term $f_\text{var}(z)$. The use of $4$-norm in $f_\text{var}(z)$ particularly weights the nodes with large standard deviation mismatches. \new{Appropriate weights $\gamma_1 = 250$ and $\gamma_2 = 150$ were found via trial and error.} The objective function $f$ is highly nonconvex and we thus use the \matlab{GlobalSearch} algorithm from the MATLAB Optimization Toolbox to minimize it. Figure~\ref{fig:x} shows the manifest nodal states as well as their best model estimates. In order to quantify the similarity between these states and their estimates, we use the standard $R^2$ measure given by \new{\begin{align*} R^2 = 1 - \frac{\sum_{\ell, i, j, k} (x_{i, j}(kT; \ell) - \hat x_{i, j}(kT; \ell))^2}{\sum_{\ell, i, j, k} (x_{i, j}(kT; \ell) - \mu_{i, j, \ell})^2} \simeq 93.6\%. \end{align*}} This high value is indeed remarkable, especially given the very small size of the network and the limited availability of measurements in the experiment \new{($2240$ data points, $175$ parameters)}. \begin{figure} \centering \includegraphics[width=0.9\linewidth]{x_rand_good10} \caption{State trajectories of manifest nodes in the network of Figure~\ref{fig:struct} (blue: measured, red: model estimate). $t = 0$ indicates stimulus onset. Solid and dashed lines correspond to LC and PD blocks, resp. The description of each node is indicated above its corresponding panel. The LC/PD in the legend refers to the trial rule, while the LC/PD above each panel refers to the preference of that particular node.}\label{fig:x} \vspace*{-1.5ex} \end{figure} \subsection{Concurrence of the Identified Network with Analysis} To conclude, we verify here whether the identified network structure satisfies the requirements of the HSR framework in terms of timescale separation and stability. Regarding the former, the identified time constants are given by \begin{align*} \new{\tau_1 = 3.36, \qquad \tau_2 = 1.68, \qquad \tau_3 = 0.70,} \end{align*} yielding an almost twofold separation of timescales conforming to Figure~\ref{fig:taus}. Regarding stability, we have to consider the LC and PD blocks separately (as the definition of task-relevant ($^\ssp$) and task-irrelevant ($\ssm$) nodes changes according to the block). In the LC block, the (manifest) LC nodes are task-relevant and the (manifest) PD nodes are task-irrelevant. Therefore, under this condition, \new{ \begin{align*} W_{3, 3}^{\ssp\ssp} &= 0.01, \qquad \qquad \ \ \Wbf_{3, 2}^{\ssp\ssp} = \begin{bmatrix} 0.01 & 0 \end{bmatrix}, \\ \Wbf_{2, 2}^{\ssp\ssp} &= \begin{bmatrix} 0.83 & 0 \\ 0.76 & 0 \end{bmatrix}, \qquad \Wbf_{2, 3}^{\ssp\ssp} = \begin{bmatrix} 0.04 \\ 0.58 \end{bmatrix}. \end{align*}} It is then straightforward to see that \begin{align*} h_3^\ssp(c_3^\ssp) = \begin{cases} 0 &; \quad c_3^\ssp \le 0 \\ c_3^\ssp/(1 - W_{3, 3}^{\ssp\ssp}) &; \quad c_3^\ssp \ge 0 \end{cases} \Rightarrow \bar F_3^\ssp = \frac{1}{1 - W_{3, 3}^{\ssp\ssp}}. \end{align*} Therefore, \new{ \begin{align*} &\rho(|W_{3, 3}^{\ssp\ssp}|) = 0.01 < 1, \\ &\rho\big(|\Wbf_{2, 2}^{\ssp\ssp}| + |\Wbf_{2, 3}^{\ssp\ssp}| \bar F_3^\ssp |\Wbf_{3, 2}^{\ssp\ssp}|\big) = \rho\Big(\begin{bmatrix} 0.83 & 0 \\ 0.77 & 0 \end{bmatrix}\Big) = 0.83 < 1, \end{align*}} satisfying the sufficient conditions for GES in~\eqref{eq:rec-cond-sim}. Similarly, in the PD block, we have \new{ \begin{align*} &W_{3, 3}^{\ssp\ssp} = 0.01 < 1, \qquad \Wbf_{3, 2}^{\ssp\ssp} = \begin{bmatrix} 4.7 \times 10^{-3} & 0 \end{bmatrix}, \\ &\Wbf_{2, 2}^{\ssp\ssp} = \begin{bmatrix} 0.12 & 0 \\ 0.56 & 0 \end{bmatrix}, \qquad \Wbf_{2, 3}^{\ssp\ssp} = \begin{bmatrix} 0.39 \\ 0.02 \end{bmatrix}, \\ &\rho\big(|\Wbf_{2, 2}^{\ssp\ssp}| + |\Wbf_{2, 3}^{\ssp\ssp}| \bar F_3^\ssp |\Wbf_{3, 2}^{\ssp\ssp}|\big) = \rho\Big(\begin{bmatrix} 0.12 & 0 \\ 0.56 & 0 \end{bmatrix}\Big) = 0.12 < 1, \end{align*}} also satisfying the GES conditions of~\eqref{eq:rec-cond-sim}. \new{While this concurrence is promising, its robustness to the choice of dataset and data processing steps is critical. A comprehensive robustness analysis requires access to multiple datasets and experimental re-design, which is beyond the scope of this case study. However, we repeated our entire analysis with Mann-Whitney-Wilcoxon rank-sum test (used originally in~\cite{CCR-MRD:14}) and also with varying significance thresholds $0.001 \le \alpha \le 0.05$ and observed that, despite the change in the neural populations, our theoretical conditions remained satisfied.} % Given the concurrence between the identified network structure and the hypotheses of our results, Theorems~\ref{thm:sp-inhib} and~\ref{thm:multi} provide strong analytical support to explain the conclusions drawn in~\cite{CCR-MRD:14,CCR-MRD:14-crcns} from experimental data and statistical analysis. We believe HSR constitutes a rigorous framework for the analysis of the multiple-timescale network interactions underlying GDSA, complementing the conventional statistical and computational analyses in neuroscience. \section{Conclusions and Future Work} We have proposed hierarchical selective recruitment as a framework to explain several fundamental components of goal-driven selective attention. HSR consists of an arbitrary number of neuronal subnetworks that operate at different timescales and are arranged in a hierarchy according to their intrinsic timescales. In this paper, we have resorted to control-theoretic tools to focus on the top-down recruitment of the task-relevant nodes. We have derived conditions on the structure of multi-layer networks guaranteeing the convergence of the state of the task-relevant nodes in each layer towards their reference trajectory determined by the layer above in the limit of maximal timescale separation between the layers. In doing so, we have characterized the piecewise affinity and global Lipschitzness properties of the equilibrium maps and unveiled their key role in the multiple-timescale dynamics of the network. Combined with the results of Part I, these contributions provide conditions for the simultaneous GES of the state of task-irrelevant nodes of all layers to the origin (inhibition) as well as the GES of the state of task-relevant nodes towards an equilibrium that moves at a slower timescale as a function of the state of the subnetwork at the layer above (recruitment). To demonstrate that applicability to brain networks, we have presented a detailed case study of GDSA in rodents and showed that a network with a binary structure based on HSR and parameters learned using a carefully designed optimization procedure can achieve remarkable accuracy in explaining the data while conforming to the theoretical requirements of HSR. Our technical treatment has also established a novel converse Lyapunov theorem for continuous GES switched affine systems with state-dependent switching. Future work will include the extension of this framework to selective inhibition using output feedback and cases where the recruited subnetworks are asymptotically stable towards more complex attractors such as limit cycles. \new{Also of paramount importance is the study of the robustness of network trajectories as well as the theoretical conditions of HSR to network parameters, disturbances, and experimental variations (inter-subject variability, different tasks, measurement noise, etc.).} Other topics of relevance to the understanding of GDSA that we plan to explore are the analysis of the information transfer along the hierarchy, the controllability and observability of linear-threshold networks, and the optimal sensor and actuator placement in hierarchical interconnections of these networks. \section*{Acknowledgments} We would like to thank Dr. Erik J. Peterson for piquing our interest with his questions on dimensionality control in brain networks and for introducing us to linear-threshold modeling in neuroscience. We are also indebted to Drs. Michael R. DeWeese and Chris C. Rodgers for the public release of their dataset~\cite{CCR-MRD:14-crcns} and their subsequent discussion of its details. This work was supported by NSF Award CMMI-1826065 (EN and JC) and ARO Award W911NF-18-1-0213~(JC).

1c69748a46c9fbb2cb07a7a9e35f9c7c0e40b087

\@startsection{subsection}{1}\z@{.7\linespacing\@plus\linespacing{\@startsection{subsection}{1}\z@{.7\linespacing\@plus\linespacing} {.5\linespacing}{\normalfont\scshape\centering}}\makeatother \newcounter{claimcount} \newcounter{thmcount} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{proposition}[theorem]{Proposition} \newtheorem*{theorem*}{Theorem} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{question}[theorem]{Question} \newtheorem{example}[theorem]{Example} \newtheorem*{examples}{Examples} \newtheorem{remark}[theorem]{Remark} \newenvironment{claim}[1]{\stepcounter{claimcount}\par\noindent\underline{Claim \theclaimcount:}\space#1}{} \newenvironment{claimproof}[1]{\par\noindent\underline{Proof:}\space#1} {\leavevmode\unskip\penalty9999\hbox{}\nobreak\hfill\quad\hbox{$\diamondsuit$}\vspace{2mm}} \newcommand{yes}{yes} \newsavebox{\commentbox}\newenvironment{com}{\ifthenelse{\equal{yes}{yes}} {\footnotemark\begin{lrbox}{\commentbox}\begin{minipage}[t]{1in}\raggedright\sffamily\tiny\footnotemark[\arabic{footnote}]}{\begin{lrbox}{\commentbox}}}{\ifthenelse{\equal{yes}{yes}}{\end{minipage}\end{lrbox}\marginpar{\usebox{\commentbox}}}{\end{lrbox}}\hspace{-1.25mm}} \newcommand*{\numberedtheorem}[3]{\theoremstyle{plain}\newtheorem*{makethm\thethmcount}{#1} \begin{makethm\thethmcount}[#2]#3\end{makethm\thethmcount}\stepcounter{thmcount}} \newenvironment{shortenum}{\begin{enumerate} \vspace{-\parskip}\setlength{\itemsep}{1pt}\setlength{\parskip}{0pt}\setlength{\parsep}{0pt}}{\end{enumerate}} \newenvironment{shortenumletters}{\begin{enumerate}[label=(\alph*)] \vspace{-\parskip}\setlength{\itemsep}{1pt}\setlength{\parskip}{0pt}\setlength{\parsep}{0pt}}{\end{enumerate}} \newenvironment{shortitem}{\begin{itemize} \vspace{-\parskip}\setlength{\itemsep}{1pt}\setlength{\parskip}{0pt}\setlength{\parsep}{0pt}}{\end{itemize}} \renewcommand{\bar}{\overline} \newcommand*{\mathfrak{a}}{\mathfrak{a}} \DeclareMathOperator{\adjoint}{ad} \newcommand*{\mathbb{C}}{\mathbb{C}} \DeclareMathOperator{\der}{der} \newcommand*{\mathbb{F}}{\mathbb{F}} \newcommand*{\mathfrak{g}_2}{\mathfrak{g}_2} \newcommand*{\mathbb{H}}{\mathbb{H}} \newcommand*{\mathfrak{h}}{\mathfrak{h}} \DeclareMathOperator{\imag}{Im} \newcommand*{\mathbb{K}}{\mathbb{K}} \newcommand*{\mathfrak{M}}{\mathfrak{M}} \newcommand*{\mathbb{N}}{\mathbb{N}} \newcommand*{\mathbb{O}}{\mathbb{O}} \DeclareMathOperator{\pure}{Pu} \newcommand*{\mathbb{Q}}{\mathbb{Q}} \newcommand*{\mathbb{R}}{\mathbb{R}} \DeclareMathOperator{\real}{Re} \newcommand*{\mathbb{Z}}{\mathbb{Z}} \newcommand*{\mathfrak{S}}{\mathfrak{S}} \newcommand*{\mathfrak{sl}}{\mathfrak{sl}} \DeclareMathOperator{\SL}{SL} \newcommand*{\mathfrak}{\mathfrak} \newcommand*{\mathcal}{\mathcal} \newcommand*{\lhalf}[1]{\overleftarrow{#1}} \newcommand*{\rhalf}[1]{\overrightarrow{#1}} \newcommand*{\sqsubset}{\sqsubset} \newcommand*{\sqsubsetneq}{\sqsubsetneq} \newcommand*{\sqsupset}{\sqsupset} \newcommand*{\sqsupsetneq}{\sqsupsetneq} \newcommand*{\pitchfork}{\pitchfork} \DeclareMathOperator{\Aut}{Aut} \DeclareMathOperator{\bigset}{Big} \DeclareMathOperator{\diam}{diam} \DeclareMathOperator{\dist}{\mathsf{d}} \DeclareMathOperator{\hull}{hull} \DeclareMathOperator{\link}{Lk} \DeclareMathOperator{\mcg}{MCG} \newcommand{\hspace{1mm}\vcentcolon\hspace{1mm}}{\hspace{1mm}\vcentcolon\hspace{1mm}} \newcommand{\lfrac}[2]{{} _{#1}\backslash^{#2}} \newcommand{\ignore}[2]{\left\{\kern-.7ex\left\{#1\right\}\kern-.7ex\right\}_{#2}} \tikzset{symbol/.style={draw=none,every to/.append style={edge node={node [sloped, allow upside down, auto=false]{$#1$}}}}} \begin{document} \maketitle \begin{abstract} It is well-known that the exceptional Lie algebras $\mathfrak{f}_4$ and $\mathfrak{g}_2$ arise from the octonions as the derivation algebras of the $3\times3$ hermitian and $1\times1$ antihermitian matrices, respectively. Inspired by this, we compute the derivation algebras of the spaces of hermitian and antihermitian matrices over an octonion algebra in all dimensions. \end{abstract} \section{Introduction} In \cite{benkartosborn:derivations}, Benkart and Osborn calculate the derivation algebra for the algebra of $n\times n$ matrices with entries in an arbitrary unital algebra, under the standard matrix product, the commutator product, and the anticommutator product. In the case that the unital algebra is an octonion algebra over a field $\mathbb{F}$, their results show that for both the standard product and the anticommutator, the derivation algebra is $\mathfrak{g}_2\oplus\mathfrak{gl}_n(\mathbb{F})$; while in the commutator case it is the direct sum of this with $\mathbb{F}$. The exceptional Lie algebra $\mathfrak{f}_4$ can be constructed as the derivation algebra of the exceptional Jordan algebra, which is the set of $3\times3$ hermitian matrices with entries in an octonion algebra, under the anticommutator product. If we increase the size of these matrices then we lose the Jordan property but still get well defined algebras. It is then natural to ask what the corresponding derivation algebras are, and to do the same for antihermitian (or skew-hermitian) matrices. When the characteristic of $\mathbb{F}$ is not two, our answers are: \begin{theorem*} If $n\geq4$ then $\der(\mathfrak{h}_n(\mathbb{O}))=\mathfrak{g}_2\oplus\mathfrak{so}_n(\mathbb{F})$. \end{theorem*} \begin{theorem*} $\der(\mathfrak{a}_n(\mathbb{O}))=\mathfrak{g}_2\oplus\mathfrak{so}_n(\mathbb{F})$ for all natural numbers $n$. \end{theorem*} This is strongly reminiscent of Benkart and Osborn's results, so since every matrix decomposes as the sum of a hermitian matrix with an antihermitian matrix, one might hope to use their methods. In practice, however, many of the tools they use break down in our case. This is mostly because, for us, entries on the diagonal come from a subspace of the octonion algebra. I would like to thank Dmitriy Rumynin for many useful discussions, and without whom the present work would not have been possible. I would also like to thank Ivan Shestakov for valuable information. \section{Set-up} Let $\mathbb{F}$ be a field of characteristic not two, and let $\mathbb{O}$ be an octonion algebra over $\mathbb{F}$. That is, $\mathbb{O}$ is a unital, alternative, 8-dimensional $\mathbb{F}$-algebra with a nondegenerate quadratic form $|\cdot|^2\vcentcolon\mathbb{O}\rightarrow\mathbb{F}$ which is multiplicative in the sense that $|zw|^2=|z|^2|w|^2$. We call elements of $\mathbb{O}$ \emph{octonions}. The nucleus of $\mathbb{O}$ is $\mathbb{F}$. More than being alternative, the nonzero octonions form a Moufang loop under multiplication. In particular they satisfy the left Moufang law \begin{align} z(w(zu))=(zwz)u. \label{moufanglaw} \end{align} We distinguish two cases for $\mathbb{O}$. We say that $\mathbb{O}$ is \emph{Type~I} if it has an orthonormal basis $1,e_1,\dots,e_7$ such that $e_i^2=-1$ for all $i$, and \emph{Type~II} otherwise. From work of Jacobson \cite[Sect. 3]{jacobson:composition}, if $\mathbb{O}$ is Type~II then it is split and has an orthonormal basis $1,e_1,\dots,e_7$ such that \hspace{1mm} i) $e_i^2=-1$ for $i\leq3$ and $e_i^2=1$ otherwise; and \hspace{1mm} ii) the $\mathbb{F}$-span of $1,e_1,e_2,e_3$ is isomorphic to $_\mathbb{F}\mathbb{H}$, the quaternions over $\mathbb{F}$. These basis elements anticommute, and consequently \begin{align} e_ie_je_i=-e_ie_ie_j=\pm e_j. \label{bimultiplication} \end{align} Note that being Type~I does not mean that $\mathbb{O}$ is a division algebra -- consider $\mathbb{F}=\mathbb{C}$, for example (for a classification of octonion algebras, see \cite{serre:cohomologie}). In any case, the algebra $\der(\mathbb{O})$ is simple of type $\mathfrak{g}_2$ \cite[Thm. 6]{jacobson:composition}. We denote conjugation in $\mathbb{O}$ by a bar: if $z=z_0+\sum_{i=1}^7z_ie_i$ then $\bar{z}=z_0-\sum_{i=1}^7z_ie_i$. We write $\real(z)=z_0$ and call it the \emph{real part} of $z$, even when the base field is not $\mathbb{R}$. Likewise, we call $\imag(z)=\sum_{i=1}^7z_ie_i$ the imaginary part of $z$, and we denote the set of all such purely imaginary octonions by $\pure(\mathbb{O})$. For more on octonions, see \cite{baez:octonions,conwaysmith:onquaternions,springerveldkamp:octonions}. We are interested in certain spaces of matrices with entries in $\mathbb{O}$. Such a matrix $x$ has a conjugate, $x^*$, which is obtained from $x$ by taking the transpose and conjugating all the entries. If $x^*=x$ then we call it \emph{hermitian}, and we denote the set of hermitian $n\times n$ matrices with entries in $\mathbb{O}$ by $\mathfrak{h}_n(\mathbb{O})$. The anticommutator product $x\circ y=xy+yx$ makes $\mathfrak{h}_n(\mathbb{O})$ into an $\mathbb{F}$-algebra. Similarly, if $x^*=-x$ then we say $x$ is \emph{antihermitian}, and we write $\mathfrak{a}_n(\mathbb{O})$ for the set of antihermitian matrices, which is made into an $\mathbb{F}$-algebra by the commutator product $[x,y]=xy-yx$. We write $E_{ij}$ for the matrix with a 1 in the $ij^\mathrm{th}$ place and zeros everywhere else. Our aim is to calculate the derivation algebras $\der(\mathfrak{h}_n(\mathbb{O}))$ and $\der(\mathfrak{a}_n(\mathbb{O}))$. Note that $\mathfrak{h}_1(\mathbb{O})=\mathbb{F}$ and $\mathfrak{a}_1(\mathbb{O})=\pure(\mathbb{O})$, so $\der(\mathfrak{h}_1(\mathbb{O}))=0$ and $\der(\mathfrak{a}_1(\mathbb{O}))=\mathfrak{g}_2$. Let $\mathfrak{so}_n(\mathbb{F})$ be the algebra of antisymmetric $n\times n$ matrices with entries in $\mathbb{F}$ under the commutator product. We use this to state an important lemma. \begin{lemma} \label{g2son} If $n>1$ then both $\der(\mathfrak{h}_n(\mathbb{O}))$ and $\der(\mathfrak{a}_n(\mathbb{O}))$ have a subalgebra isomorphic to $\mathfrak{g}_2\oplus\mathfrak{so}_n(\mathbb{F})$. \end{lemma} \begin{proof} An element of $\mathfrak{g}_2$ gives a derivation by acting on a matrix $x$ entrywise. The action of $\mathfrak{so}_n(\mathbb{F})$ is the adjoint action: if $A\in\mathfrak{so}_n(\mathbb{F})$ then $\adjoint_A\vcentcolon x\mapsto[A,x]$ is a derivation because $A$ has entries in $\mathbb{F}$, meaning that $A^*=A^T=-A$. It is easy to check that the two actions commute, whence the direct summation. \end{proof} \section{Hermitian} For hermitian matrices, in dimensions 2 and 3 Jacobson tells us \cite[Thm. 14]{jacobson:some}: \begin{theorem} If the characteristic of $\mathbb{F}$ is not two or three then $\der(\mathfrak{h}_2(\mathbb{O}))=\mathfrak{so}_9(\mathbb{F})$ and $\der(\mathfrak{h}_3(\mathbb{O}))=\mathfrak{f}_4$. \end{theorem} This result extends earlier work of Chevalley and Schafer over algebraically closed fields of characteristic zero \cite{chevalleyschafer:exceptional}, and is the main motivation for the present work. We henceforth consider $n\geq4$ only in this section, and allow the characteristic of $\mathbb{F}$ to be three. There are five types of nonzero product in $\mathfrak{h}_n(\mathbb{O})$: \begin{gather} E_{ii}\circ E_{ii}=2E_{ii} \label{hiiii} \\ E_{ii}\circ (zE_{ij}+\bar{z}E_{ji}) =zE_{ij}+\bar{z}E_{ji} \label{hiiij} \\ E_{jj}\circ (zE_{ij}+\bar{z}E_{ji}) =zE_{ij}+\bar{z}E_{ji} \label{hjjij} \\ (zE_{ij}+\bar{z}E_{ji})\circ(wE_{ij}+\bar{w}E_{ji}) =2\real(z\bar{w})(E_{ii}+E_{jj}) \label{hijij} \\ (zE_{ij}+\bar{z}E_{ji})\circ(wE_{jk}+\bar{w}E_{kj}) =zwE_{ik}+\bar{w}\bar{z}E_{ki}. \label{hijjk} \end{gather} By applying a derivation $\partial$ to these we can obtain constraints that $\partial$ must satisfy. We first do this for a special subset of derivations. \begin{proposition} \label{htraceless} If $n\geq4$ then the subalgebra of derivations $\partial\vcentcolon\mathfrak{h}_n(\mathbb{O})\rightarrow\mathfrak{h}_n(\mathbb{O})$ such that $\partial(E_{ii})=0$ for all~$i$ is isomorphic to $\mathfrak{g}_2$. \end{proposition} \begin{proof} From \eqref{hiiij} and \eqref{hjjij} we have \[ E_{ii}\circ\partial(zE_{ij}+\bar{z}E_{ji}) =\partial(zE_{ij}+\bar{z}E_{ji}) =E_{jj}\circ\partial(zE_{ij}+\bar{z}E_{ji}), \] and hence there are linear maps $\alpha^{ij}\vcentcolon\mathbb{O}\rightarrow\mathbb{O}$ such that \begin{align*} \partial(zE_{ij}+\bar{z}E_{ji})=\alpha^{ij}(z)E_{ij}+\overline{\alpha^{ij}(z)}E_{ji}. \end{align*} Since $\partial(E_{ii})=0$ for all $i$, the $\alpha^{ij}$ determine $\partial$. Note in particular that \begin{align} \alpha^{ij}(z)=\overline{\alpha^{ji}(\bar{z})}. \label{alphaijalphaji} \end{align} Applying $\partial$ to \eqref{hijij}, we find in the $ii^\mathrm{th}$ place the equality \[ \alpha^{ij}(z)\bar{w} +w\overline{\alpha^{ij}(z)} +z\overline{\alpha^{ij}(w)} +\alpha^{ij}(w)\bar{z}=0, \] which we can restate as \begin{align} \real(\alpha^{ij}(z)\bar{w})+\real(\alpha^{ij}(w)\bar{z})=0. \label{realalphazw} \end{align} Letting $z=w$ run through the standard basis of $\mathbb{O}$ in \eqref{realalphazw} we get that both the real part of $\alpha^{ij}(1)$ and the $e_k^\mathrm{th}$ part of $\alpha^{ij}(e_k)$ are zero. By taking $z=e_k\neq e_l=w$ in \eqref{realalphazw} we get that the $e_k^\mathrm{th}$ part of $\alpha^{ij}(e_l)$ is the negative of the $e_l^\mathrm{th}$ part of $\alpha^{ij}(e_k)$. With $z=1$, $w=e_k$ in \eqref{realalphazw} we get two cases. If $\mathbb{O}$ is Type~I then the $e_k^\mathrm{th}$ part of $\alpha^{ij}(1)$ is the negative of the real part of either $\alpha^{ij}(e_k)$, and hence $\alpha^{ij}\in\mathfrak{so}_8(\mathbb{F})$. On the other hand, if $\mathbb{O}$ is Type~II then the $e_k^\mathrm{th}$ part of $\alpha^{ij}(1)$ is equal to the real part of either $\alpha^{ij}(e_k)$ or its negative, depending on whether $k>3$ or $k\leq3$, respectively. Thus the matrix of $\alpha^{ij}$ has the form \begin{align} \left( \begin{array}{c|cc} 0 & -v_1^T & v_2^T \\ \hline v_1 & & \\ v_2 & \multicolumn{2}{c}{\smash{\raisebox{.4\normalbaselineskip}{$A$}}} \end{array} \right) \in \mathbb{F}^{8\times8}, \label{TypeIImatrix} \end{align} where $v_1\in\mathbb{F}^3$, $v_2\in\mathbb{F}^4$, and $A\in\mathfrak{so}_7(\mathbb{F})$. We now apply $\partial$ to \eqref{hijjk}, and find in the $ik^\mathrm{th}$ place the equality \begin{align} \alpha^{ik}(zw)=\alpha^{ij}(z)w+z\alpha^{jk}(w). \label{hprederivation} \end{align} In particular, if $e_t\neq e_r$ then \[ \alpha^{ik}(e_te_r)=\alpha^{ij}(e_t)e_r+e_t\alpha^{jk}(e_r) \hspace{5mm} \mathrm{and} \hspace{5mm} \alpha^{ij}(e_te_r)=\alpha^{ik}(e_t)e_r+e_t\alpha^{kj}(e_r). \] Comparing $e_t^\mathrm{th}$ parts, it follows from the form \eqref{TypeIImatrix} of $\alpha^{ij}$ that if $\mathbb{O}$ is Type~II and $r>3$ then $\real(\alpha^{jk}(e_r))=\real(\alpha^{kj}(e_r))$, and so is zero by \eqref{alphaijalphaji}. In particular, irrespective of whether $\mathbb{O}$ is Type~I or Type~II, we have \begin{align} \alpha^{ij}\in\mathfrak{so}_8(\mathbb{F}). \label{soeight} \end{align} Returning to \eqref{hprederivation}, the maps $\alpha^{ik}, \alpha^{ij}, \alpha^{jk}$ are said to be in triality, and in light of \eqref{soeight}, any one uniquely determines the other two \cite[p.42]{springerveldkamp:octonions}. We use these trialities to show that all $\alpha^{ij}$ are equal. If $j>2$ then we have trialities $\alpha^{12},\alpha^{1j},\alpha^{j2}$. Since these share the same first map we have $\alpha^{1j}=\alpha^{13}$ and $\alpha^{2j}=\alpha^{23}$ whenever $j>2$. If $k>j>2$ then the trialities $\alpha^{1j},\alpha^{1k},\alpha^{kj}$ share the same first map, so all $\alpha^{jk}$ with $k>j>2$ are equal to $\alpha^{34}$. The two trialities $\alpha^{12},\alpha^{14},\alpha^{42}$ and $\alpha^{13},\alpha^{14},\alpha^{43}$ share the same second map, so $\alpha^{12}=\alpha^{13}$ and $\alpha^{24}=\alpha^{34}$. Hence if $k>j>1$ then all $\alpha^{1j}$ are equal to $\alpha^{12}$, and all $\alpha^{jk}$ are equal to~$\alpha^{23}$. Finally, the two trialities $\alpha^{13},\alpha^{14},\alpha^{43}$ and $\alpha^{23},\alpha^{24},\alpha^{43}$ share the same third map, so $\alpha^{14}=\alpha^{24}$. Hence all the $\alpha^{ij}$ are equal to $\alpha^{12}$. Writing $\alpha=\alpha^{12}$, we can now read \eqref{hprederivation} as \begin{align*} \alpha(zw)=\alpha(z)w+z\alpha(w). \end{align*} That is, $\alpha\in\der\mathbb{O}=\mathfrak{g}_2$, and $\partial$ is given by applying $\alpha$ to each entry. \end{proof} \begin{theorem} \label{hermitianderivations} If $n\geq4$ and the characteristic of $\mathbb{F}$ is not two, then $\der(\mathfrak{h}_n(\mathbb{O}))=\mathfrak{g}_2\oplus\mathfrak{so}_n(\mathbb{F})$. \end{theorem} \begin{proof} Let $\partial$ be a derivation. Our strategy is to show that $\partial$ differs from one of the derivations of Proposition \ref{htraceless} by the adjoint action of an element of $\mathfrak{so}_n(\mathbb{F})$. Applying $\partial$ to \eqref{hiiii} we find that there are constants $\mu_{ik}^i\in\mathbb{O}$ for $k\neq i$ such that \begin{align*} \partial(E_{ii})=\sum_{k\neq i} (\mu_{ik}^iE_{ik}+\overline{\mu_{ik}^i}E_{ki}), \end{align*} and $\partial$ applied to $E_{ii}\circ E_{jj}=0$ yields \begin{align} \mu_{ij}^i=-\overline{\mu_{ji}^j}. \label{muantisymmetry} \end{align} The $\partial(E_{ii})$ are thus determined by the choice of $\mu_{ij}^i$ with $j>i$. Now let \[ \partial(zE_{ij}+\bar{z}E_{ji})=\sum_{k,l}\alpha_{kl}^{ij}(z)E_{kl}. \] Similarly to in the proof of Proposition \ref{htraceless}, applying $\partial$ to \eqref{hiiij} and \eqref{hjjij} leads to \begin{align*} \partial(zE_{ij}+\bar{z}E_{ji}) =2E_{jj}&\real(z\mu_{ij}^i)-2E_{ii}\real(\mu_{ij}^i\bar{z}) +\alpha_{ij}^{ij}(z)E_{ij}+\overline{\alpha_{ij}^{ij}(z)}E_{ji} \\ +&\sum_{t\neq i,j}\big( z\mu_{jt}^jE_{it} +\overline{\mu_{jt}^j}\bar{z}E_{ti}+\overline{\mu_{it}^i}zE_{tj} +\bar{z}\mu_{it}^iE_{jt}\big). \nonumber \end{align*} In particular, if the pair $(k,l)$ is not equal to either $(i,j)$ or $(j,i)$ then $\alpha_{kl}^{ij}=0$. We therefore abbreviate $\alpha_{ij}^{ij}$ to just $\alpha^{ij}$, and note that $\partial$ is determined by the $\mu_{ij}^i$ and the $\alpha^{ij}$ with $j>i$. Now apply $\partial$ to both sides of \eqref{hijjk}. If $t\neq i,j,k$ then in the $it^\mathrm{th}$ place we find the equality \begin{align*} (zw)\mu_{kt}^k=z(w\mu_{kt}^k). \end{align*} Thus $\mu_{kt}^k$ lies in the nucleus of $\mathbb{O}$, which is $\mathbb{F}$. By varying $i,j,k$ we find that all $\mu_{ij}^i$ lie in~$\mathbb{F}$. Let $A=(\mu_{ij}^i)_{ij}$. Since all $\mu_{ij}^i$ lie in $\mathbb{F}$, equation \eqref{muantisymmetry} tells us that $A\in\mathfrak{so}_n(\mathbb{F})$, so $\adjoint_A\in\der(\mathfrak{h}_n(\mathbb{O}))$. Moreover, \[ \adjoint_A(E_{ii})=\sum_{k\neq i}(\mu_{ik}^iE_{ik}+\overline{\mu_{ik}^i}E_{ki})=\partial(E_{ii}). \] Hence $\partial-\adjoint_A$ is a derivation which maps all $E_{ii}$ to zero, so by Proposition \ref{htraceless}, $\partial-\adjoint_A$ is given by an element of $\mathfrak{g}_2$, and by Lemma \ref{g2son} we are done. \end{proof} The exceptional Lie algebra $\mathfrak{e}_6$ can be constructed as \[ \mathfrak{e}_6=\der(\mathfrak{h}_3(\mathbb{O}))+\{L_x\hspace{1mm}\vcentcolon\hspace{1mm} x\in\mathfrak{h}_3(\mathbb{O}), \hspace{1mm} \mathrm{tr}(x)=0\}, \] where $L_x$ denotes left multiplication by $x$. This is due to Chevalley and Schafer \cite{chevalleyschafer:exceptional} (see also \cite[Sect. 4.4]{schafer:introduction}). A natural question therefore arises from Theorem~\ref{hermitianderivations}: \begin{question} How does this construction of $\mathfrak{e}_6$ generalise to $\mathfrak{h}_n(\mathbb{O})$? \end{question} One barrier to generalisation is that the commutator of two left multiplications may fail to be a derivation, for while in the $3\times3$ case the derivation algebra has dimension $\dim(\mathfrak{f}_4)=52$, in the $4\times4$ case its dimension is only $\dim(\mathfrak{g}_2\oplus\mathfrak{so}_4(\mathbb{F}))=20$. One remedy would be to include products of multiplication maps, and some work in this vein is done in \cite{petyt:special}. \section{Antihermitian} Here we compute the algebras $\der(\mathfrak{a}_n(\mathbb{O}))$ for all $n$. We find this to be more fiddly than the hermitian case. Again there are five types of nonzero product in $\mathfrak{a}_n(\mathbb{O})$: \begin{gather} [e_iE_{tt},e_jE_{tt}]=2e_ie_jE_{tt} \hspace{5mm} i\neq j \label{atttt} \\ [e_iE_{tt},zE_{tr}-\bar{z}E_{rt}]=e_izE_{tr}+\bar{z}e_iE_{rt} \label{atttr} \\ [e_iE_{rr},zE_{tr}-\bar{z}E_{rt}]=-ze_iE_{tr}-e_i\bar{z}E_{rt} \label{arrtr} \\ [zE_{tr}-\bar{z}E_{rt},wE_{tr}-\bar{w}E_{rt}]= 2\imag(w\bar{z})E_{tt}+2\imag(\bar{w}z)E_{rr} \label{atrtr} \\ [zE_{tr}-\bar{z}E_{rt},wE_{rs}-\bar{w}E_{sr}]=zwE_{ts}-\bar{w}\bar{z}E_{st}. \nonumber \end{gather} and again we find restrictions on a derivation by applying it to (the first four of) these. \begin{theorem} If the characteristic of $\mathbb{F}$ is not two, then $\der(\mathfrak{a}_n(\mathbb{O}))=\mathfrak{g}_2\oplus\mathfrak{so}_n(\mathbb{F})$ for all natural numbers $n$. \end{theorem} \begin{proof} By Lemma \ref{g2son} it suffices to bound the dimension of $\der(\mathfrak{a}_n(\mathbb{O}))$ above by $14+\frac{n(n-1)}{2}$. Applying a derivation $\partial$ to both sides of \eqref{atttt}, we find that for $k\neq t$ there are linear maps $a_{tk}^t\vcentcolon\pure(\mathbb{O})\rightarrow\mathbb{O}$ and $a_{tt}^t\vcentcolon\pure(\mathbb{O})\rightarrow\pure(\mathbb{O})$ such that \begin{align*} \partial(e_iE_{tt})=a_{tt}^t(e_i)E_{tt}+ \sum_{k\neq t} \big(a_{tk}^t(e_i)E_{tk}-\overline{a_{tk}^t(e_i)}E_{kt}\big), \end{align*} and moreover, if $k\neq t$ and $i\neq j$ then \begin{align} 2a_{tk}^t(e_ie_j)=e_ia_{tk}^t(e_j)-e_ja_{tk}^t(e_i). \label{aij} \end{align} Applying $\partial$ to $[e_iE_{tt},e_iE_{rr}]=0$ we get that $a_{tr}^t(e_i)e_i+e_i\overline{a_{rt}^r(e_i)}=0$, and hence \begin{align} a_{rt}^r(e_i)=\pm e_i\overline{a_{tr}^t(e_i)}e_i. \label{artrtrt} \end{align} Now let $\partial(zE_{tr}-\bar{z}E_{rt})=\sum\beta_{kl}^{tr}(z)E_{kl}$. Applying $\partial$ to \eqref{atttr}, we get in positions $rr$, $tr$, $tk$, $rk$, $kl$ (with $k,l,r,t$ pairwise distinct) the following respective equalities: \begin{gather} \beta_{rr}^{tr}(e_iz)=2\imag(\bar{z}a_{tr}^t(e_i)) \label{tbrr} \\ \beta_{tr}^{tr}(e_iz)=e_i\beta_{tr}^{tr}(z)+a_{tt}^t(e_i)z \label{tbtr} \\ \beta_{tk}^{tr}(e_iz)=e_i\beta_{tk}^{tr}(z) \label{tbtk} \\ \beta_{rk}^{tr}(e_iz)=\bar{z}a_{tk}^t(e_i) \label{tbrk} \\ \beta_{kl}^{tr}(e_iz)=0. \label{tbkl} \end{gather} \begin{claim}{} \label{alambda} The $a_{tr}^t$ are scalar multiples of the identity map $I\vcentcolon \pure(\mathbb{O})\rightarrow\pure(\mathbb{O})$. \end{claim} \begin{claimproof} Taking $z=e_i$ in \eqref{tbrr} gives $\beta_{rr}^{tr}(1)=\pm2\imag(e_ia_{tr}^t(e_i))$, depending on whether $\mathbb{O}$ is Type~I or Type~II and on the value of $i$, and it follows that all but the $e_i^\mathrm{th}$ part of $a_{tr}^t(e_i)$ is determined by $a_{tr}^t(e_1)$. Comparing $(e_ie_j)^\mathrm{th}$ parts in \eqref{aij} we find that the $(e_ie_j)^\mathrm{th}$ part of $a_{tk}^t(e_ie_j)$ is half the sum of the $e_j^\mathrm{th}$ part of $a_{tk}^t(e_j)$ with the $e_i^\mathrm{th}$ part of $a_{tk}^t(e_i)$. Cycling $e_i\mapsto e_j\mapsto e_ie_j\mapsto e_i$, we get that the $e_i^\mathrm{th}$ part of $a_{tk}^t(e_i)$ is the same for all~$i$. Thus, if we set $c_{tr}^t=a_{tr}^t+\real(e_1a_{tr}^t(e_1))I$ then $c_{tr}^t$ is simply $a_{tr}^t$ except that the $e_i^\mathrm{th}$ part of $c_{tr}^t(e_i)$ is zero for all $i$, and by \eqref{tbrr} we have $2e_ic_{tr}^t(e_i)=\beta_{rr}^{tr}(1)$. In particular, \begin{align} c_{tr}^t(e_j)=\left\{\begin{array}{@{}l@{}l} \multirow{2}{*}{$-e_j(e_ic_{tr}^t(e_i))$} &\hspace{5mm} \text{if }\mathbb{O} \text{ is Type I} \\ &\hspace{7mm} \text{or }\mathbb{O} \text{ is Type II and }j\leq3; \\[1mm] e_j(e_ic_{tr}^t(e_i)) &\hspace{5mm} \text{if }\mathbb{O} \text{ is Type II and }j>3. \end{array}\right. \label{citoj} \end{align} Since $c_{tr}^t$ differs from $a_{tr}^t$ only by a multiple of the identity, \eqref{aij} holds for $c_{tr}^t$. If $\mathbb{O}$ is Type~II and either $i\leq3<j$ or $j\leq3<i$, then combining \eqref{aij} with \eqref{citoj} gives \[ 2c_{tr}^t(e_ie_j)=e_ic_{tr}^t(e_j)-e_jc_{tr}^t(e_i)= \pm e_i(e_j(e_ic_{tr}^t(e_i)))-e_jc_{tr}^t(e_i)=0, \] where the last equality holds by the left Moufang law \eqref{moufanglaw} and equation \eqref{bimultiplication}. On the other hand, if $i,j\leq3$, $i,j>3$, or $\mathbb{O}$ is Type~I then \begin{align} 2c_{tr}^t(e_ie_j)=e_ic_{tr}^t(e_j)-e_jc_{tr}^t(e_i) =-2e_i(e_j(e_ic_{tr}^t(e_i))) \label{typeiiandijparity} \end{align} But \eqref{citoj} tells us that $2c_{tr}^t(e_ie_j)=-2(e_ie_j)(e_ic_{tr}^t(e_i))$, so the associator $[e_i,e_j,e_ic_{tr}^t(e_i)]=0$, and hence $c_{tr}^t(e_i)$ lies in the subalgebra generated by $e_i$ and $e_j$. If $\mathbb{O}$ is Type~I then this holds for all choices of $e_j$, so $c_{tr}^t(e_i)$ lies in the complex subalgebra generated by $e_i$. But we constructed $c_{tr}^t$ so that the $e_i^\mathrm{th}$ part of $c_{tr}^t(e_i)$ is zero, and hence $c_{tr}^t(e_i)\in\mathbb{F}$. Now comparing real parts in \eqref{aij}, we find that the real part of $c_{tr}^t(e_ie_j)$ is zero. Hence $c_{tr}^t=0$ if $\mathbb{O}$ is Type~I. Similarly, if $\mathbb{O}$ is Type~ II and $i,j,k>3$ then $c_{tr}^t(e_i)$ lies in both the subalgebra generated by $e_i$ and $e_j$ and in the subalgebra generated by $e_i$ and $e_k$, so it lies in the subalgebra generated by $e_i$, and hence is an element of $\mathbb{F}$. If $i\leq 3$ then we can partition $\{e_4,e_5,e_6,e_7\}$ into two pairs $e_{j_1}, e_{k_1}$ and $e_{k_2},e_{k_2}$ such that $e_i=e_{j_l}e_{k_l}$. Then by \eqref{typeiiandijparity}, $c_{tr}^t(e_i)=-e_{j_l}(e_{k_l}(e_{j_l}c_{tr}^t(e_{j_l})))$, which since $j_l>3$ lies in the $\mathbb{F}$-span of $e_{k_l}$ for both $l=1$ and $l=2$. Hence $c_{tr}^t=0$ if $\mathbb{O}$ is Type~II as well. It follows from the construction of $c_{tr}^t$ that there exist constants $\lambda_{tr}^t\in\mathbb{F}$ such that \begin{align} a_{tr}^t=\lambda_{tr}^tI\hspace{1mm}\vcentcolon\hspace{1mm} \pure(\mathbb{O})\longrightarrow\pure(\mathbb{O}), \end{align} which proves the claim. \end{claimproof} Applying $\partial$ to \eqref{arrtr}, we get in positions $tt$, $tr$, $tk$, $rk$ (with $k\neq r,t$) the following respective equalities: \begin{gather} \beta_{tt}^{tr}(ze_i)=-2\imag(\lambda_{tr}^te_i\bar{z}) \label{rbtt} \\ \beta_{tr}^{tr}(ze_i)=za_{rr}^r(e_i)+\beta_{tr}^{tr}(z)e_i \label{rbtr} \\ \beta_{tk}^{tr}(ze_i)=za_{rk}^r(e_i) \label{rbtk} \\ \beta_{rk}^{tr}(ze_i)=-e_i\beta_{rk}^{tr}(z). \label{rbrk} \end{gather} Taking $z=1$ in \eqref{tbrk} and \eqref{rbtk} and using Claim 1 gives \begin{gather*} \beta_{rk}^{tr}\vert_{\pure(\mathbb{O})}=a_{tk}^t=\lambda_{tk}^tI, \hspace{5mm} \beta_{tk}^{tr}\vert_{\pure(\mathbb{O})}=a_{rk}^r=\lambda_{rk}^rI, \end{gather*} and then by taking $z=e_i$ in \eqref{tbtk} and \eqref{rbrk} we conclude that \begin{gather} \beta_{tk}^{tr}=\lambda_{rk}^rI, \hspace{5mm} \beta_{rk}^{tr}=\lambda_{tk}^tI. \label{brklambda} \end{gather} Now, taking $z=1$ in \eqref{tbrr}, \eqref{tbtr}, \eqref{rbtt}, \eqref{rbtr} and $z=e_i$ in \eqref{tbrr} and \eqref{rbtt} gives: \begin{gather} \beta_{rr}^{tr}(e_i)=2\lambda_{tr}^te_i, \hspace{4mm} \beta_{rr}^{tr}(1)=0 \label{brrlambda} \\ a_{tt}^t(e_i)=\beta_{tr}^{tr}(e_i)-e_i\beta_{tr}^{tr}(1) \label{attt} \\ \beta_{tt}^{tr}(e_i)=-2\lambda_{tr}^te_i, \hspace{4mm} \beta_{tt}^{tr}(1)=0 \label{bttlambda} \\ a_{rr}^r(e_i)=\beta_{tr}^{tr}(e_i)-\beta_{tr}^{tr}(1)e_i. \label{arrr} \end{gather} \begin{claim}{} There is an element $\beta\in\mathfrak{g}_2$ such that $a_{tt}^t=\beta_{tr}^{tr}=\beta$ for all $t$ and $r$. \end{claim} \begin{claimproof} Applying $\partial$ to \eqref{atrtr}, in the $tt^\mathrm{th}$ place we find the equality \[ 2a_{tt}^t(e_i)=\beta_{tr}^{tr}(1)e_i+ e_i\overline{\beta_{tr}^{tr}(1)}+ \beta_{tr}^{tr}(e_i)-\overline{\beta_{tr}^{tr}(e_i)}. \] Combining this with \eqref{attt} gives \[ 2\beta_{tr}^{tr}(e_i) -2e_i\beta_{tr}^{tr}(1) =\beta_{tr}^{tr}(1)e_i +e_i\overline{\beta_{tr}^{tr}(1)} +2\imag(\beta_{tr}^{tr}(e_i)), \] and comparing $e_i^\mathrm{th}$ parts we find that \begin{align} \real(\beta_{tr}^{tr}(1))=0. \label{realbtr} \end{align} Taking $z=e_i$ in \eqref{rbtr} and using \eqref{arrr} leads to \[ \pm\beta_{tr}^{tr}(1)=\beta_{tr}^{tr}(e_i)e_i +e_i\beta_{tr}^{tr}(e_i) -e_i\beta_{tr}^{tr}(1)e_i, \] and by \eqref{realbtr}, if we compare real parts in this then we get that the $e_i^\mathrm{th}$~part of $\beta_{tr}^{tr}(e_i)$ is zero. Combining \eqref{tbtr} with \eqref{attt} and \eqref{rbtr} with \eqref{arrr} we get, respectively: \begin{gather} \beta_{tr}^{tr}(e_ie_j) =e_i\beta_{tr}^{tr}(e_j)+\beta_{tr}^{tr}(e_i)e_j-(e_i\beta_{tr}^{tr}(1))e_j \label{beiej} \\ \beta_{tr}^{tr}(e_ie_j) =e_i\beta_{tr}^{tr}(e_j)-e_i(\beta_{tr}^{tr}(1)e_j)+\beta_{tr}^{tr}(e_i)e_j, \nonumber \end{gather} and hence $(e_i\beta_{tr}^{tr}(1))e_j=e_i(\beta_{tr}^{tr}(1)e_j)$ for all $i$ and $j$, so $\beta_{tr}^{tr}(1)$ is in the nucleus of $\mathbb{O}$, which is $\mathbb{F}$. By \eqref{realbtr} we now have $\beta_{tr}^{tr}(1)=0$, and it follows from \eqref{beiej} that \begin{align*} \beta_{tr}^{tr}(e_ie_j)=e_i\beta_{tr}^{tr}(e_j)+\beta_{tr}^{tr}(e_i)e_j. \end{align*} That is $\beta_{tr}^{tr}\in\der(\mathbb{O})=\mathfrak{g}_2$. Using \eqref{attt} and \eqref{arrr} then varying $r$ and $t$ completes the proof of the claim. \end{claimproof} Putting together \eqref{tbkl}, \eqref{brklambda}, \eqref{brrlambda}, \eqref{bttlambda} with Claims 1 and 2 we find that $\partial$ is completely determined by the choice of $\beta\in\mathfrak{g}_2$ and the $\lambda_{tk}^t$ with $k\neq t$. By \eqref{artrtrt} we only need $k>t$, so $\dim(\der(\mathfrak{a}_n(\mathbb{O})))\leq14+\frac{n(n-1)}{2}$, and by Lemma \ref{g2son} we are done. \end{proof} \bibliographystyle{alpha}

f1c4bfb1acb70f848e0f76955ac35fb88bb4cd5c

\section{Introduction} \label{sec:introduction} In contrast to conventional supervised learning, where a single target variable needs to be predicted on the basis of a set of features describing an instance, multi-target prediction (MTP) is concerned with the simultaneous prediction of multiple target variables of possibly different type, such as binary, nominal, ordinal, or real-valued. Applications of multi-target prediction are omnipresent in the digitalized society of the 21st century. Classical applications that are often studied in machine learning papers include image tagging in computer vision, document cateogorization in text mining, and product recommendation in online advertising. Besides, MTP problems arise in many other application domains as well. In medicine, one is interested in predicting several clinical outcomes of patients at the same time. In climate sciences, one would like to forecast extreme weather events for many regions in the world, which are related through complex physical and geological processes. In biology, one would like to unravel the different biological functions that a gene might express. In chemistry, one would like to know which molecules might be potential drugs to cure a given disease. In social networks, one intends to predict which users interact with a given user. In ecology, one constructs species distribution models, which describe the prevalence of different types of species in a given habitat. Applications of that kind have resulted in novel research questions, and a need for developing specialized multi-target prediction methods. Often, several target variables are related to each other, for example because they obey certain constraints (for instance, if variables are positions in a ranking, they must be mutually exclusive), they correspond to nodes of a graph, or they provide evidence of parent-child relationships. In other situations, specific properties of the targets are known, such as molecular structures or feature representations. Obviously, knowledge of that kind could be used to improve predictive performance. However, in many applications, neither target relations nor representations are known a priori. Instead, they need to be discovered from the data. The main credo of research in MTP is the conviction that, compared to the most obvious approach of learning an individual prediction model for each target variable independently of the others, the exploitation of target dependencies will lead to better performance. In this review-style article, we present a unifying view of MTP problems and methods. When speaking about multi-target prediction, machine learning researchers often allude to either multi-label classification, multi-output regression (a.k.a.\ multivariate regression in statistics), or multi-task learning problems. Moreover, if additional side information in the form of target relations or target representations is accessible, those three settings further extend to multi-target prediction scenarios that are known as dyadic prediction, hierarchical multi-label classification, and zero-shot learning in the literature. Besides, there is a close connection to matrix completion and network inference, which becomes obvious when representing the relationship between instances and targets in the form of a matrix or a graph. Despite strong commonalities, there is little interaction between the different sub-communities. Moreover, there are several problems that have been studied in different communities under different names. Sometimes there is even terminological confusion within the same community. As a first contribution, we provide a structured overview of the multitude of MTP problems. To this end, we present a formal framework for multi-target prediction in Section~2. By identifying a set of characteristic properties, we subdivide MTP problems into a number of well-known settings. Particular attention will be paid to the formal definitions of problems that characterize the fields of multi-label classification, multivariate regression, multi-task learning, zero-shot learning, and matrix completion. We also discuss a number of related settings, such as structured output prediction and multi-class classification, and argue why those settings should not be covered by the umbrella of MTP. Subsequently, in Section~3, we present a unifying view of MTP methods. Here, we intend to unravel a number of general mechanisms that are essential for obtaining state-of-the-art predictive performance in MTP. As will be seen, the applicability or usefulness of a method strongly depends on properties of the problem setting, such as whether or not side information is available for targets, and how this side information looks like. Another important question is whether or not one intends to generalize to novel targets and/or instances. The overall goal of this paper is to provide insight into the vast literature on MTP problems and methods, especially for readers who are new to the field. The paper is not a typical review paper, however, and does not lay claim on being comprehensive in this regard---in light of the breadth of the field, that appears to be an impossible endeavor. Instead, we intend to focus on some general principles that might be helpful in identifying the right approach for a given problem. In Section~4, we conclude with a couple of remarks and challenges for future research, and briefly discuss some important aspects of multi-target prediction that are less emphasized in this paper. \section{A unifying view on MTP problems} \subsection{A general framework for multi-target prediction} In this section, we establish links between different MTP problems. We start by describing a general framework that covers both simple and more advanced MTP problems. Formally, our framework is defined as follows. \begin{definition} {\bf (Multi-target prediction)} A multi-target prediction setting is characterized by instances $\vec{x} \in \mathcal{X}$ and targets $\vec{t} \in \mathcal{T}$ with the following properties: \begin{itemize} \item[P1.] A training dataset $\mathcal{D}$ consists of triplets $(\vec{x}_i,\vec{t}_j,y_{ij})$, where $y_{ij} \in \mathcal{Y}$ denotes a score that characterizes the relationship between the instance $\vec{x}_i$ and the target $\vec{t}_j$. \item[P2.] In total, $n$ different instances and $m$ different targets are observed during training, with $n$ and $m$ finite numbers. Thus, the scores $y_{ij}$ of the training data can be arranged in an $n \times m$ matrix $Y$, which is in general incomplete, i.e., $Y$ has missing values. \item[P3.] The score set $\mathcal{Y}$ is one-dimensional. It consists of nominal, ordinal or real values. \item[P4.] The goal consists of predicting scores for any instance-target couple $(\vec{x},\vec{t}) \in \mathcal{X} \times \mathcal{T}$. \end{itemize} \end{definition} The above definition accommodates the availability of side information for targets. To keep notation simple, we stick to vector representations in our definitions, i.e., we identify targets with feature vectors $\vec{t}$. We remark, however, that other types of side knowledge, such as structured representations or relations, could also be considered. The examples to be discussed below will make this point more clear. In what follows, we show that various multi-target prediction problems are recovered as specific instantiations of the above framework. As already said, if side information for targets is available, we assume it can be encoded in the vector representation $\vec{t}$. If no side information is available, $\vec{t}$ will be an uninformative vector (e.g., merely consisting of an identifying number). \subsection{Conventional multi-target prediction settings} We start by explaining what we call conventional multi-target prediction problems. For those problems, side information for targets is normally not available. Multi-label classification, multivariate regression, and multi-task learning are the most well-known subfields of machine learning that can be mentioned as examples. Multivariate regression and multi-label classification consider the simultaneous prediction of multiple binary or real-valued targets, respectively~\citep{Demb2012a,Tsoumakas_and_Katalos_2007}. Multi-task learning then unifies those subfields, and further extends them to problems where not all targets are relevant for all instances~\citep{Caruana1997}. To make this point clearer, we discuss three prototypical examples. \begin{figure} \begin{center} \includegraphics[width=6cm]{Slide1} \includegraphics[width=6cm]{Slide4} \\ \includegraphics[width=6cm]{Slide2} \includegraphics[width=6cm]{Slide5} \\ \includegraphics[width=6cm]{Slide3} \includegraphics[width=6cm]{Slide6} \end{center} \caption{A visualization of six prototypical multi-target prediction problems: (1) drug design as an example of a multivariate regression problem, (2) document categorization as an example of a multi-label classification problem, (3) student grading as an example of a multi-task learning problem, (4) drug design with a similarity measure for targets, (5) document categorization with a hierarchy for targets, (6) student grading with target features. See Examples~1 to 6 in the text for more details.} \label{fig:basicexamples} \end{figure} \begin{example} Consider the problem of predicting whether a protein will bind to a set of experimentally developed small molecules. This is an important application in the area of drug design, where machine learning methods can help in screening interesting novel chemical compounds. Using multi-target prediction methods, high-potential compounds can be selected and analyzed further during (more rigorous) wet-lab experimentation. Figure~\ref{fig:basicexamples}-1 illustrates the data that might be generated in an application of that kind. For a set of training proteins (first five rows, in green), the binding affinity with a set of small molecules (represented as columns) has been verified experimentally. Furthermore, we assume that we have additional information about the proteins available, in the form of a feature representation or a more structured representation such as a graph (shown as red puzzle pieces in the picture). One can use this data to train a multi-target prediction model that is able to predict, on the same set of small molecules, binding affinities for new proteins that have not yet been analyzed in the wet-lab. As experimental verification of binding affinities delivers continuous scores that represent binding strength, one arrives at a multivariate regression problem. \end{example} \begin{example} In the area of document categorization, assigning appropriate category tags to documents is important but laborious and time consuming. Hence, one often applies multi-target prediction methods to automate this tagging process. As shown in Figure~\ref{fig:basicexamples}-2, one might have asked a human annotator to provide a training dataset with relevant tags being assigned to a set of documents. Using a feature representation of those documents, such as a bag-of-words representation, one can then train a multi-target prediction method that will be able to assign tags to new documents outside the training dataset. More than one tag might be relevant for a particular document, so automated document categorization boils down to solving a multi-label classification problem. \end{example} \begin{example} We consider the application of predicting student marks in the final exam for a typical high-school course, let's say mathematics. Forecasting those marks at an early stage, prior to the start of a course, might be useful to give students advise with respect to study directions and perspectives. Using historical student records that might be stored by schools, one can train a multi-target prediction method that estimates the exam marks of newly entering students. To this end, one would typically construct a dataset that looks like Figure~\ref{fig:basicexamples}-3, where columns represent different courses or different schools. If one aims for estimating the marks for different courses or different schools simultaneously, one ends up with solving a multi-task learning problem. \end{example} Examples 1--3 discuss basic multi-target prediction settings in which no additional knowledge about the targets is known. Example~1 is a multivariate regression problem, as one aims to predict the values of continuous variables, whereas Example~2 is a multi-label classification problem with binary target variables. Example~3 depicts an application that can only be tackled with multi-task learning methods. As students usually attend only one school, only one label will be observed for per student. The other labels are unknown and typically irrelevant---they are represented by grey cells in Figure~\ref{fig:basicexamples}-3. Likewise, in the prediction phase, it is meaningful to restrict predictions to the school of the student, or at least a subset of schools for which the student is considering an enrolment. In most multi-task learning problems, only one task or a subset of tasks is relevant for a given instance. We therefore introduce the following formal definitions to include multivariate regression, multi-label classification, and multi-task learning in our framework. \begin{definition} {\bf (Multivariate regression)} A multivariate regression problem is a specific instantiation of the general framework, which exhibits the following additional properties: \begin{enumerate} \item[P5.] The cardinality of $\mathcal{T}$ is $m$. This implies that all targets are observed during training. \item[P6.] No side information is available for targets. Without loss of generality, we can hence assign the numbers $1$ to $m$ as identifiers to targets, such that the target space is $\mathcal{T} = \{1,...,m\}$. \item[P7.] The score matrix $Y$ has no missing values. \item[P8.] The score set is $\mathcal{Y} = \mathbb{R}$. \end{enumerate} \end{definition} \begin{definition} {\bf (Multi-label classification)} A multi-label classification problem is a specific instantiation of the general framework, which exhibits the following additional properties: \begin{enumerate} \item[P5.] The cardinality of $\mathcal{T}$ is $m$; this implies that all targets are observed during training. \item[P6.] No side information is available for targets. Again, without loss of generality, we can hence identify targets with natural numbers, such that the target space is $\mathcal{T} = \{1,...,m\}$. \item[P7.] The score matrix $Y$ has no missing values. \item[P8*.] The score set is $\mathcal{Y} = \{0,1\}$. \end{enumerate} \end{definition} \begin{definition} {\bf (Multi-task learning)} A multi-task learning problem is a specific instantiation of the general framework, which exhibits the following additional properties: \begin{enumerate} \item[P5.] The cardinality of $\mathcal{T}$ is $m$; this implies that all targets are observed during training. \item[P6.] No side information is available for targets. Again, the target space can hence be taken as $\mathcal{T} = \{1,...,m\}$. \item[P8{**}.] The score set is homogenous across columns of $Y$, e.g., $\mathcal{Y} = \{0,1\}$ or $\mathcal{Y} = \mathbb{R}$. \end{enumerate} \end{definition} In the above definitions, we adopt a matrix view for multi-task learning. This means we consider a matrix of targets, where rows are indexed by the number of training instances, and columns are indexed by the different tasks. A cell $(i,j)$ is only filled with a value $y_{ij}$ if instance $i$ is contained in the training dataset of task $j$. We will use the letter $Y$ to denote the resulting sparsely filled matrix. Figure~1-3 depicts an example of such a matrix. From this matrix viewpoint, multivariate regression and multi-label classification arise as special cases (when the matrix $Y$ has no missing values, like in the examples in Figure~\ref{fig:basicexamples}). Multi-task learning methods can usually process multivariate regression and multi-label classification datasets with no problems. In fact, many authors of multi-task learning papers turn out to be analyzing multivariate regression or multi-label classification datasets in their experimental studies, because such datasets are widely available. However, albeit not always visible in such experimental studies, multi-task learning methods are usually more general than multivariate regression and multi-label classification methods. Multivariate regression and multi-label classification can be further generalized to targets with other types of values, such as nominal, ordinal, or mixed. This setting basically omits Property $8^{**}$. It is sometimes referred to as multi-dimensional classification, even though it has been rarely investigated in the machine learning literature---see e.g.\ \citep{Bielza2011,Read2013}. Another problem that can be seen as an instantiation of our framework is label ranking, where each instance is associated with a ranking (total order) of the targets \citep{Huellermeier_et_al_2008}. Thus, the score $y_{ij} \in \{1,...,m\}$ for a pair $(\vec{x}_i, \vec{t}_j)$ is the position of $\vec{t}$ in the ranking associated with $\vec{x}$, i.e., each row of the score matrix $Y$ is a permuation of $\{1,...,m\}$. \subsection{Problems that involve side information for targets} Let us now extend multivariate regression, multi-label classification, and multi-task learning to settings where additional side information about the target space is available. To this end, we adjust the three examples from above slightly. \begin{example} In the drug design application of Example~1, we treated proteins as instances and small molecules that can potentially bind to those proteins as targets. Let us assume that, in addition to the graph-based representation for the proteins, a representation for the target molecules is also available, as shown in Figure~\ref{fig:basicexamples}-4. This is a common situation in research on protein-ligand prediction, where the representation for the targets is either graph-based or feature-based. The resulting machine learning setting can be interpreted as a dyadic prediction problem. \end{example} \begin{example} Let us return to the document categorization application that was described in Example~2. Often document categories are organized in a hierarchy that describes the degree of relatedness among document tags. An example is shown in Figure~\ref{fig:basicexamples}-5. Problems of that kind, where the target space is equipped with a hierarchical structure, are referred to as hierarchical multi-label classification problems. Naturally, state-of-the-art algorithms will try to exploit the structure for better prediction. \end{example} \begin{example} In the student mark forecasting application that was considered before, one might also assume that additional information about the targets is given. One might be able to collect all sorts of variables about schools and courses, such as geographical location, qualifications of the teachers, reputation of the school, etc. Figure~\ref{fig:basicexamples}-6 visualizes such a situation with binary features. One might be able to improve multi-task learning algorithms by taking the features into account. This side information forms a key element for tackling transfer learning and zero-shot learning problems. \end{example} Examples 4-6 further extend basic MTP settings to situations where various types of side information about targets is available. Example~4 illustrates this with the case of a structured representation in the form of a molecular graph. Example~5 assumes a hierarchy that describes relations among targets in a specific manner, and Example~6 considers feature representations for targets. Applications of that kind lead to more complex multi-target prediction problems that are often referred to as dyadic prediction, link prediction, or network inference settings---see e.g. \citep{Menon2010,Schafer2015}. In this area, one can distinguish algorithms that model vector representations or structured target representations, as well as methods that model target relations. This will be further discussed in Section~3. Dyadic prediction, link prediction, and network inference are general terms that cover a wide range of problems. Generally speaking, they cover problems that obey the four properties listed in Definition~1. The labels $y_{ij}$ can be arranged in a matrix $Y$, which is often sparsely filled. Thus, one may argue that dyadic prediction is nothing else than multi-task learning with task features. However, the multi-task learning terminology is rarely used in the dyadic prediction literature. Dyadic prediction problems emerge in a variety of application domains, including product recommendation, social network analysis, drug design, various bioinformatics applications, and game playing \citep{Basilico2004,park2009pairwise,Stock2014,Benhur2005,Kashima2009,Pelossof2015,Jacob2008a,pahikkala2010reciprocalkm,pahikkala2013efficientcondrank}. \subsection{Inductive versus transductive learning problems} One may argue that the problems analyzed in Examples~1-3 are inductive w.r.t.\ instances and transductive w.r.t.\ targets. Predictions need be be generated for novel instances, whereas the set of targets is known beforehand and observed during the training phase. Side information is of crucial importance for generalizing to novel targets that are unobserved during the training phase, such as a novel target molecule in the drug design example, a novel tag in the document annotation example, or a novel course in the student grading example. \begin{figure} \includegraphics[width=6cm]{Slide7} \hspace{0.5cm} \includegraphics[width=6cm]{Slide8} \caption{Visualization of the different types of prediction problems that arise when prior knowledge of the target space is modelled. Left: an extension of the drug design example, where predictions have to be made for a novel protein (Setting B in blue), a new ligand (Setting C in yellow) or a combination of a new protein and ligand (Setting D in red). Right: a general overview of the four settings that can be distinguished. Setting A depicts the situation where missing values need to be imputed in the matrix. See text for details.} \label{fig:foursettings} \end{figure} This is visualized in Figure~\ref{fig:foursettings}, where the red part of the dataset alludes to a protein-drug pair that was not observed during the training phase. In the multi-task and transfer learning literature, this type of problem setting is known as zero-shot learning \citep{Larochelle2008}. Since no training data is available for the novel target, the problem is intrinsically more difficult than those analyzed in Section~2.2 \citep{park2012flaws}. In this work, we define zero-shot learning as follows. \begin{definition} {\bf (Zero-shot learning)} A zero-shot learning problem is a specific instantiation of the general framework with the following additional property: \begin{enumerate} \item[P5*.] The cardinality $m^*$ of the set $\mathcal{T}$ is bigger than $m$. Some targets are hence not observed during training, but may nevertheless appear at prediction time. \end{enumerate} \end{definition} By substituting P5 with P5*, one now tackles problems that are inductive instead of transductive w.r.t.\ targets. The same subdivision can be made for instances. As pointed out by \citet{Pahikkala2014}, in total four different settings can be distinguished in the presence of side information about instances and targets. In Figure~\ref{fig:foursettings}, Setting D resembles the zero-shot learning setting, whereas Setting B coincides with the problems that we discussed in Section~2.2. In addition, one can also think of Setting C (predicting scores for a novel target on an instance that was contained in the training dataset) and Setting A (predicting scores for an instance-target combination, where the target and the instance were already seen during the training phase). Combinations of those settings also exist. For example, one can distinguish zero-shot learning settings where only targets that are unknown during training need to be predicted in the test phase (thus Setting D), versus settings where both training and test targets have to be predicted for novel instances (a combination of Setting B and D). The latter, for example, happens in image classification tasks, where it is referred to as generalized zero-shot learning \citep{Rohrbach2011a,Xian2017}. Setting C is inductive w.r.t.\ targets and transductive w.r.t.\ instances. Setting A is transductive w.r.t.\ instances and targets. Setting C is in theory identical to Setting B. When side information is available for both instances and targets, the two settings become interchangeable. In Example~4, we referred to proteins as instances and small molecules as targets, but in principle, one could transpose the matrix without changing the algorithms. However, interchanging the rows and columns of a particular dataset and applying a particular multi-target prediction algorithm on the two scenarios will in practice lead to different results. Often, one of the two sources of side information is more informative than the other, and many algorithms take this implicitly into account. Take the document categorization example, for which textual descriptors on the document level are much more informative and richer than the hierarchical structure on the tag level. Therefore, modelling the side information on the document level is of uttermost importance, while a reasonable performance might still be expected even if the side information on the tag level is ignored. In contrast, Setting A results in yet a different type of problem, namely that of matrix completion. Both the targets and the instances are observed, albeit not for all instance-target combinations. In principle, one can solve this setting with methods that ignore side information of the target and the instance space. In such cases, latent representations of instance and target space are both deduced using only the instance-target interaction matrix. In light of Examples~1-3, this means that the feature representations of proteins, documents, and students are not used. In fact, they do not even have to be available. As a result, we formally define matrix completion problems as follows. \begin{definition} {\bf (Matrix completion)} A matrix completion problem is a specific instantiation of the general framework with the following additional properties: \begin{enumerate} \item[P5.] The cardinality of $\mathcal{T}$ is $m$. This implies that all targets are observed during training. \item[P6.] No side information is available for targets. Without loss of generality, we can hence assign identifiers to targets from the set $\{1,...,m\}$ such that the target space is $\mathcal{T} = \{1,...,m\}$. \item[P9.] The cardinality of $\mathcal{X}$ is $n$. This implies that all instances are observed during training. \item[P10.] No side information is available for instances. Without loss of generality, we can hence assign identifiers to instances from the set $\{1,...,n\}$, such that the instance space is $\mathcal{X} = \{1,...,n\}$. \end{enumerate} \end{definition} Matrix completion methods are extensively used in areas such as recommender systems, social network analysis, and biological network inference. They are therefore also known as collaborative filtering and link prediction methods. One can distinguish different versions of matrix completion. One may consider matrices that contain continuous, ordinal or binary values, but situations with presence-only data are also common. Observed entries in the matrix then correspond to known positive interactions between the items in the rows and the columns, and the goal consists of predicting more positive interactions from the large pool of missing entries. The assumption made in such cases is that the overwhelming majority of the missing entries correspond to negative interactions, with the exception of a few positive ones that need to be found. Properties 5 and 9 characterize the transductiveness of matrix completion methods. Properties 6 and 10 characterize the difference between a pure matrix completion setting, and a hybrid setting that simultaneously exploits the structure of the matrix and side information to generate predictions. \subsection{Beyond MTP: Problems that are not covered} Our formal framework is rather generic and covers several machine learning problems as special cases, including those discussed above. In principle, every prediction problem with (original) output space $\mathcal{Y}$ could be seen as a special case, simply by taking $\mathcal{T} = \mathcal{Y}$ as the set of tasks and $\{0,1\}$ as a score set. This amounts to treating each candidate output as a target, and reinterpreting the task to predict an output from $\mathcal{Y}$ as predicting, for each candidate value, whether or not the sought output corresponds to that candidate. Consequently, a consistent prediction has to obey strong (deterministic) dependencies between the targets: It should assign the score 1 to exactly one target and 0 to all the others. As two examples, consider multi-class classification and structured output prediction (SOP), which can be formalized, respectively, as specializations of multi-label classification and zero-shot learning. \begin{definition} {\bf (Multi-class classification)} A multi-class classification problem is a specific instantiation of the general framework with the following additional properties: \begin{enumerate} \item[P5.] The cardinality of $\mathcal{T}$ is $m$. This implies that all targets are observed during training. \item[P6.] No side information is available for targets. Without loss of generality, we can hence assign identifiers to targets from the set $\{1,...,m\}$ such that the target space is $\mathcal{T} = \{1,...,m\}$. \item[P7.] The score matrix $Y$ has no missing values. \item[P8*.] The score set is $\mathcal{Y} = \{0,1\}$. \item[P11.] Each row in $Y$ has a single ``positive" entry, and all other entries are zero. \end{enumerate} \end{definition} \begin{definition} {\bf (Structured output prediction)} A structured output prediction problem is a specific instantiation of the general framework with the following additional properties: \begin{enumerate} \item[P5*.] The cardinality $m^*$ of the set $\mathcal{T}$ is bigger than $m$. Some targets are hence not observed during training, but they appear at test time. \item[P7.] The score matrix $Y$ has no missing values. \item[P8*.] The score set is $\mathcal{Y} = \{0,1\}$. \item[P11.] Each row in $Y$ has a single ``positive" entry, and all other entries are zero. \end{enumerate} \end{definition} Note that these encodings of the problems correspond to what is known as a 1-versus-rest decomposition, a specific reduction technique that makes prediction problems with $|\mathcal{Y}| > 2$ amenable to binary classification. Obviously, other reduction techniques such as all-pairs or, more generally, error correcting output codes (ECOC) lead to similar representations: The problem of predicting the output for an instance $\vec{x} \in \mathcal{X}$ is decomposed into a set of binary prediction problems, and each of these problems is seen as a task. While formally possible, we are not in favor of considering problems such as multi-class classification and structured output prediction as special cases of MTP. The reasons include both conceptual and algorithmic aspects. Conceptually, the view of each candidate prediction as a separate target appears to be rather artificial. Actually, one is still interested in a single prediction, not multiple ones. To comply with the corresponding consistency constraints, a kind of post-processing (like the decoding step in ECOC) is normally required. Algorithmically, it is also clear that the multi-target perspective is not typical of methods for SOP. Instead, such methods are specifically tailored for output spaces in SOP, which are often huge but equipped with a strong structure. In this paper, we will also exclude other (multi-target) prediction problems where the ground truth cannot be represented in a matrix format with optional side information for rows and columns. This includes problems that involve multi-instance learning representations \citep{Zhou2007} and dyadic feature representations \citep{Vanpeer2017}. The latter scenario occurs, for example, in bio-informatics and recommender systems applications, when features are available that describe an interaction between an instance and a target, e.g., the day of the week when a user clicked a specific item on a website. \section{A unifying view on MTP methods} In this section, we present a unifying view on MTP methods. To this end, we categorize methods according to some of the properties that were discussed in Section~2. In particular, we believe that the properties characterizing Settings A, B, C and D are specifically important in this regard, and suitable to distinguish MTP methods. We therefore put a main emphasis on those properties, while some other properties of Section~2 are not further considered. As a result, our categorization of MTP methods does not yield a one-to-one mapping with MTP problems. Table~\ref{tab:sections} gives an overview of the subdivision we propose. Most of the methods to be discussed can be applied to Settings B and C. Some of them are also applicable to Setting D, while Setting A normally calls for more specifilized methods. \begin{large} \begin{center} \begin{table}[h] \caption{Categorization of MTP methods.} \label{tab:sections} \begin{tabular}{ll} \hline Group of methods & Applicable setting \\ \hline \hline Section \ref{sec:methods-that-do-not-use-domain-knowledge}: Similarity-enforcing methods & B and C \\ Section \ref{sec:methods-that-use-target-relations}: Relation-exploiting methods & B, C and D \\ Section \ref{sec:methods-that-learn-target-relations}: Relation-constructing methods & B and C \\ Section \ref{sec:methods-that-use-target-representations}: Representation-exploiting methods & B, C and D \\ Section \ref{sec:methods-that-learn-target-representations}: Representation-constructing methods & B and C \\ Section \ref{sec:matrix-completion-and-hybrid-methods}: Matrix completion and hybrid methods & A \\ \hline \end{tabular} \end{table} \end{center} \end{large} In what follows, we focus on the average squared prediction error over all targets as a loss function to be minimized. Formally, for a given pair $(\vec{x}, \vec{t})$, a predictor $f$ yields a prediction $\hat{y}= f(\vec{x}, \vec{t})$ of an actual outcome (score) $y$. Let $\vec{y}$ be a vector of actual output values of size $m$ (or $m^*$ in the case of zero-shot learning), and let $\hat{\vec{y}}$ be a vector of the corresponding predictions. Then, the squared-error loss is given by \begin{equation} L(\vec{y},\hat{\vec{y}}) = \sum_{i=1}^{m} (y_i - \hat{y}_i)^2 \,. \label{eqn:squared-error-loss} \end{equation} In spite of obvious shortcomings, both in regression and classification settings, this loss is somewhat representative in the sense of being used in the majority of algorithms for multivariate regression and multi-task learning, and therefore suitable for explaining general concepts. It has the additional advantage of being applicable for both numerical and binary target variables, again making it a natural choice for the unifying view that we intend to provide. All the methods presented below can be tailored for minimizing this loss function. In multi-label classification, many methods optimize loss functions that are more complex and not decomposable over targets. Because those methods only appear in the MLC literature, an in-depth discussion is beyond the scope of this paper. Instead, we refer the interested reader to \citet{Demb2012a}. We start our discussion on multi-target methods by introducing a simple baseline that does not exploit any target dependencies. Subsequently, we present more and more complex models that tackle the problem of multi-target prediction in different manners. \subsection{Independent models} \label{sec:independent-models} The most straight-forward approach to solving MTP problems consists of constructing one model $f_i$ for every target independently, and to concatenate the predictions of these models into the sought multi-target prediction. In the multi-label classification community, this approach has an own name and is known as \emph{binary relevance learning}~\citep{Tsoumakas_and_Katalos_2007}. In order to introduce the approach more formally, let $f_i(\vec{x})$ denote the score (continuous or discrete) assigned to the $i$th target for instance $\vec{x}$. Using linear basis function models, the model for the $i$th target can be represented as \begin{equation} f_i(\vec{x}) = \vec{a}_i^\intercal \phi(\vec{x}) \,, \label{eq:binrel} \end{equation} where $\vec{a}_i$ is a target-specific parameter vector and $\phi(\vec{x})$ a feature representation that is either given explicitly or first obtained from another representation in a preprocessing step. Multivariate ridge regression is probably the most basic model of this kind. This approach fits a linear model in a regularized least-squares fashion, while ignoring potential dependencies among targets \citep{Hastie_et_al_2007}. More formally, multivariate ridge regression boils down to minimizing the following objective function: \begin{equation} \label{eq:multiridge} \min_A ||Y - XA ||^2_F + \sum_{i=1}^m \lambda_i \, ||\vec{a}_i||^2 \,, \end{equation} where $Y$ is the $n \times m$ target matrix with $n$ the number of observations and $m$ the number of targets. Denoting the number of features by $p$, $X$ and $A$ are $n \times p$ and $p \times m$ matrices, respectively, which are constructed as follows: \begin{equation} \label{eq:notation} X = \begin{bmatrix} \phi(\vec{x}_1)^T \\ \vdots \\ \phi(\vec{x}_n)^T \end{bmatrix} \qquad A = [\vec{a}_1 \quad \cdots \quad \vec{a}_m] \,. \end{equation} $||.||_F$ is the Frobenius-norm, so that the left-hand side corresponds to the sum of squared error losses (\ref{eqn:squared-error-loss}) on the $m$ targets. The $\lambda_i$ are regularization parameters. Optimization problem (2) can be solved by matrix inversion in a least-squares algebraic fashion. This is efficient because the inversion needs to be done only once for all targets. The formulation assumes a linear statistical model, but independent models can be constructed using any well-known method for regression (in the case of continuous variables as targets) or classification (in the case of binary variables as targets). Can one perform better with more sophisticated approaches that exploit dependencies among targets? Unsurprisingly, the answer is affirmative. Many of the approaches that were introduced over the last decade in each of the MTP subcommunities are able to outperform the simple baseline that treats every target as an independent problem. Generally speaking, these methods put soft constraints on the values that targets can take, so they assume that not all combinations of values are equally likely to occur: some combinations of target values have a higher probability of occurrence than others, yielding a dependency among targets in a statistical sense. Authors introducing a novel multi-target prediction method often emphasize the exploitation of target dependencies as the key to improved performance. Without questioning explanations of that kind, we like to note that improvements over independent predictions can be achieved even for problems without any statistical dependence between the targets. This insight is illustrated, for example, by the famous James-Stein paradox, showing that, under certain conditions, maximum likelihood estimation is not optimal (or, using statistical terminology, not admissible) for estimating the mean of a multivariate Gaussian \citep{James1961}. Since the James-Stein estimation principle is less known among machine learning scholars, we provide a short description in Appendix~A. \subsection{Similarity-enforcing methods} \label{sec:methods-that-do-not-use-domain-knowledge} We start our overview of MTP methods by explaining a few simple methods that do not consider any particular domain knowledge about targets. Such methods can be found in multivariate regression, multi-label classification, and multi-task learning, and they are usually only applicable to Settings B and C. The central idea in those simple methods is that models for different targets should behave \emph{similar}. This can be achieved, for example, by enforcing similarity of the parameterizations of models for different targets. To this end, one usually defines a mathematical objective that incorporates a specific regularizer. In early work on multi-task learning, \citet{Evgeniou_2004}, \citet{Evgeniou2005}, and \citet{Jalali2010} solve the following optimization problem: \begin{equation} \label{eq:meanreg} \min_A ||Y - XA ||^2_F + \lambda \sum_{i=1}^m ||\vec{a}_i - \frac{1}{m} \sum_{j=1}^m \vec{a}_j||^2 \, , \end{equation} with $A$ the parameter matrix, $X$ the feature matrix, and $Y$ the target matrix, as defined before. In this way, one penalizes for deviations from the ``mean target". The parameter estimates are biased to the mean, in an attempt to reduce the variance and prevent overfitting. Another common way of obtaining restricted models with less flexibility (compared to modelling each target independently) consists of implementing feature selection strategies that retain the same features for different targets. Joint feature selection leads to a model with less parameters compared to binary relevance, and this may in turn result in performance gains. The idea has been particularly popular in the area of multi-label classification \citep{Obzinski2010,Gu:2011,Kong2012,Spolaor:2016}. Even though most of those authors put the emphasis on computational efficiency as a motivation, selecting a subset of identical features for different targets can also improve the predictive performance. Similar ideas are appearing in the literature on multi-task learning \citep{Zhou2011} and in joint regression of binary and continuous response variables \citep{Zhang2012}. An embedded feature selection strategy can be obtained, for example, by a combination of L1 and L2 norms: $$ \min_A ||Y - XA ||^2_F + \lambda \sum_{j=1}^p ||\vec{a}_j||^2 $$ In contrast to (\ref{eq:meanreg}), the sum is taken over the features. Unlike (\ref{eq:notation}), the vectors $\vec{a}_j$ now represent the rows of matrix $A$: $$A = \begin{bmatrix} \vec{a}_1^T \\ \vdots \\ \vec{a}_p^T \end{bmatrix}$$ Thus, $\vec{a}_j$ specifies the set of parameters for the $j$th feature, across different targets. The combination of the two types of norms results in group sparsity, so that a set of identical features is selected for different targets. Yet another straight-forward approach to exploiting dependencies among targets is stacking \citep{Wolpert1992}. This approach was initially introduced as an ensemble learning technique in conventional classification and regression settings. Later on, it has been extended for multi-label classification \citep{Godbole_Sarawagi_2004,Cheng_Hullermeier_2009}. In the statistical literature, a slightly more advanced method has shown to improve the predictions for multivariate regression \citep{Breiman_Friedman_1997}. Although one could easily think of extensions for multi-task learning problems, too, the method has been less popular in that area. Stacking implements a two-step procedure. In the first step, a model is fitted for every target individually, as discussed in Section~\ref{sec:independent-models}. Then, in a second step, the predictions obtained by each model are used as a feature representation to train a second series of models, again one for each target. Obviously, these second-order models are able to capture dependencies between different targets, as they seek to represent the prediction of one target as a function of the (predictions of) the others. Thus, the goal of the second step is to discover dependencies among targets, and to identify, for each target in turn, a set of other targets that help improve predictions. More formally, with the shorthand-notation $$\vec{f}(\vec{x}) = \begin{bmatrix} f_1(\vec{x}) \\ f_2 (\vec{x}) \\ \vdots \\ f_m(\vec{x}) \end{bmatrix} $$ for the independent models of (\ref{eq:binrel}), the general scheme of stacking can be expressed as follows: \begin{equation} \label{eqn:slp_1} \vec{f}^*(\vec{x}) = \mathbf{b} (\mathbf{f}(\vec{x}), \vec{x})\, , \end{equation} with $\vec{f}^*(\cdot)$ the predictions made by stacking and $\mathbf{b}(\cdot)$ a second-stage model that shrinks or regularizes the solution of the initial models. \citet{Breiman_Friedman_1997} show formally that stacking prevents overfitting. As soon as some regularization is present in the second-order models, one introduces a bias by encouraging models for different targets to learn similar parameters. The same authors also reveal some connections with James-Stein estimation. Two appealing properties of stacking are its ease of implementation and its generality, as it can be applied in tandem with any type of classifier or regressor. In addition, the second-level model $\mathbf{b}$ can be trained on the first-level predictions $\mathbf{h}(\vec{x})$ alone or in concatenation with the original features $\vec{x}$. In the case of classification, there is also a choice between feeding the second-level model with binary predictions or with continuous scores obtained from scoring functions or probabilities, if such outputs are delivered by the classifier. The idea of enforcing that models for different targets behave similarly can also be found in neural networks, starting from classical papers such as \citep{Caruana1997}, till modern deep learning architectures, such as deep convolutional nets for multi-label classification \citep{Wei2016}. Such methods are based on weight sharing in the final network layer, which usually includes $m$ nodes for a multi-target prediction setting with $m$ targets. Likewise, the common idea of pre-training a deep neural network on an auxiliary task with a lot of training data, and further optimizing it for another task with limited training data, can be interpreted as weight sharing---see e.g.\ \citep{Donahue2014,Girshick2014,Oquab2014,Razavian2014,Sermanet2014,Gong2014b}. \subsection{Relation-exploiting methods} \label{sec:methods-that-use-target-relations} In this section, we discuss MTP methods using side information in the form of target relations, such as hierarchies, graph structures, decision rules, and correlation matrices. This side information might be useful for at least three purposes. First, it might help obtain performance gains for Settings B and C, compared to the methods that were discussed in Section~\ref{sec:methods-that-do-not-use-domain-knowledge}. Second, it can be considered as a key element when generalizing to novel targets in Setting D. Third, it may also effect the computational scalability of algorithms in a positive way. When speaking about target relations, it is important to make a distinction between deterministic relations, which are guaranteed to hold for each observation, and probabilistic relations, which only give an indication of the likelihood of occurrence. In multi-label classification, for example, one could observe deterministic relations such as implications, subsumptions, and mutual exclusions between relevant labels of an instance. In label ranking, the values of target variables are positions in a ranking, which means that all target values must be different (each position can only be occupied once). In multivariate regression, targets may represent consecutive time steps, for which constraints such as smoothness or monotonicity over consecutive values might apply. For MLC problems such as Example 5 in Section~2, relationships between labels are often represented in the form of a tree or hierarchy, resulting in a set of problems that is known as hierarchical multi-label classification. Many authors have shown that such a hierarchy can be integrated in specific methods \citep{Rousu2006,Barutcuoglu2006,Vens_et_al_2008,Silla2010,Gopal2012,Nam2015}. The hierarchy not only helps to improve predictions, but also allows for defining extended prediction settings that are characterized by specific performance measures. For example, when a classifier is unsure about certain classes, it could be allowed to return intermediate nodes from the tree as prediction instead of the leaf nodes that correspond to single labels \citep{Bi2012}. A few authors have presented more general approaches for incorporating deterministic relations that cannot be represented as a hierarchy. \citet{Gopal2013} and \citet{Deng2014} both consider graph structures. When targets follow a chronological order, chain-based graph structures might be a natural choice. For example, \citet{Zhou:2012} study disease progression prediction as a multi-task learning problem, where tasks follow a chronological order, leading to a formulation where a specific temporal regularizer in the optimization problem is introduced. This regularizer can be seen as a special case of the graph-based regularization of \citet{Gopal2013}: \begin{equation*} \min_A ||Y - XA ||^2_F + \lambda \sum_{i=1}^m \sum_{j \in \mathcal{N}(i)} ||\vec{a}_i - \ \vec{a}_j||^2 \, , \end{equation*} where $\mathcal{N}(i)$ denotes the set of targets that are related to the $i$th target. This formulation shows that methods using target relations are actually very related to those exploiting similarity (cf.\ Section~\ref{sec:methods-that-do-not-use-domain-knowledge}). The available side information essentially allows one to define prior knowledge concerning statistical dependencies that are likely to occur. In addition to deterministic relations, there are also methods that incorporate probabilistic relations. In the context of kernel methods, it is natural to represent target similarities by means of so-called output kernels, which can be interpreted as expressing some sort of correlation between targets. Output kernels have been mainly developed for multivariate regression, under the framework of vector-valued kernel functions \citep{Caponnetto2008,baldassarre2012multioutput,Alvarez2012}. Similar kernel-based formulations have been proposed for multi-label classification \citep{Hariharan_et_al_2010}. All the above methods essentially assume that the provided information about the target space results in a good representation for modelling target dependencies. This is clearly a necessary condition for improvement---a few authors have also reported performance drops in certain applications, when the given target relations lead to an incorrect representation of target dependencies. \subsection{Relation-constructing methods} \label{sec:methods-that-learn-target-relations} When target relations to be used as side information are not available in the form of prior knowledge, one can try to construct such relations from the training data. As an important consequence of this approach, note that it excludes a generalization to novel targets. Thus, unlike the methods discussed in the previous section, methods that learn target relations are not applicable to Setting D; instead, they are mainly useful for settings B and C. Learning target relations can be useful for several reasons. First of all, domain knowledge is often not available, so that target relations cannot be modelled as prior information. Second, even if such knowledge can in principle be provided, target relations specified by domain experts might be incorrect and hence misleading for machine learning algorithms. A third motivation is computational efficiency, since learning target relations paves the way for methods that scale sublinearly in the number of targets. Like in the case where target relations are given instead of being learned, one can distinguish methods w.r.t.\ the type of relations: hierarchies, more general graph structures, and correlation matrices. In hierarchical multi-label classification, where the correctness of a hierarchy could be questioned, one can find several methods that adjust the existing hierarchy in the course of the learning procedure. In that sense, one can distinguish level-flattening algorithms, node removal algorithms, hierarchy modification algorithms, and hierarchy generation algorithms; see \citep{Rangwala2017} for an overview. In those methods, the focus is on improving the predictive performance. Conversely, in extreme multi-label classification, i.e., problems that involve thousands or millions of labels, hierarchies are often learned to obtain a training or prediction time that is logarithmic in the number of labels \citep{Agrawal2013,Weston2013,Prabhu2014,Demb2016}. Graphs more general than hierarchies might be considered. This idea is popular when applying graphical models in multi-target prediction settings, see e.g.\ \citep{Guo2011,Papagiannopoulou2015}. Using directed graphs, one can infer asymmetric relationships among targets \citep{Lee2016}, while similar results might be obtained with specific rule-based systems \citep{Park2008,Loza2016}. Let us remark that, for many of those methods, it becomes more difficult to say whether the inferred target relations are deterministic or probabilistic, because the learning process usually contains both deterministic and probabilistic elements. In some applications, one may cluster targets into a number of non-overlapping groups, such that interactions between targets need to be considered only inside the same group. This can be established by using a cluster-norm in the optimization \citep{jacob2008,Wang2009}. Similarly, clustered multi-task learning was considered in a Bayesian setting by \citet{Bakker2003} via a mixture of Gaussians instead of a single Gaussian prior, and by \citet{Xue2007} via a Dirichlet process prior. Furthermore, \citet{Gong2012} identify outlier targets, resulting in a structure where only information is shared among the inliers. Similarly, clustering methods have been considered in multi-label classification, where both performance gains in prediction and computational complexity have been reported. Finally, one can also learn target relations in the form of similarity or correlation matrices. Unsurprisingly, this idea has been very popular in the field of kernel methods. The older methods rather assume that the output kernel is given as domain knowledge, whereas more recent methods try to discover target similarities by learning the output kernel in a convex formulation \citep{Dinuzzo2011,Dinuzzo2013,Jawanpuria2015}. Similarly, one can learn the covariance function among targets in a Bayesian treatment, by assuming a multivariate Gaussian distribution as prior \citep{Zhang2010}. \subsection{Representation-exploiting methods} \label{sec:methods-that-use-target-representations} Instead of using relations between targets, one can also consider to use target representations as side information. Examples~4 and 6 in Section~2 describe two applications where this type of side information is available in the form of a graph representation and a vector representation, respectively. Just like for target relations, this side information can be useful for several purposes. It might boost the predictive performance in Settings B and C, and is essential for generalizing to novel targets in Setting~D. Compared to target relations, target representations still require the learning algorithm to discover which elements of the representations are useful and which are not. For instance, in the student grading application of Example~6, one still has to figure out which of the school features are useful. Likewise, in the drug design application of Example~4, one would need specialized methods that can handle structured data such as molecular graphs. In contrast, when working with target relations, the side information is usually available in a more direct and a less noisy format. Historically, kernel methods have played an important role in modelling target representations. Such methods can easily process vectorial representations, as well as structured data. Moreover, they also establish a clear connection with methods that model target relations: a transformation from a primal formulation to a dual formulation in fact implies a transformation from a target representation to a target relation. In the primal formulation, information about the instance space and the target space is combined by means of a joint feature representation, whereas in the dual a joint kernel is modelled. Such representations typically yield prediction models of the following form: \begin{equation} \label{eq:pairwise} f(\vec{x},\vec{t}) = \vec{w}^T \Psi(\vec{x},\vec{t}) = \sum_{(\bar{\vec{x}},\bar{\vec{t}}) \in \mathcal{D}} \alpha_{(\bar{\vec{x}},\bar{\vec{t}})} \Gamma((\vec{x},\vec{t}),(\bar{\vec{x}},\bar{\vec{t}})) \, , \end{equation} where $\mathcal{D}$ represents the training data as defined before, $\vec{w}$ and $\alpha_{(\bar{\vec{x}},\bar{\vec{t}})}$ are primal and dual parameters, $\Psi(\vec{x},\vec{t})$ is a joint feature representation in the primal formulation, and $$\Gamma((\vec{x},\vec{t}),(\bar{\vec{x}},\bar{\vec{t}})) = \Psi(\vec{x},\vec{t})^T \Psi(\bar{\vec{x}},\bar{\vec{t}})$$ is a joint kernel in the dual formulation, with $\vec{t}$ and $\bar{\vec{t}}$ representations of two targets. The most commonly used pairwise kernel is the Kronecker product pairwise kernel \citep{Basilico2004,oyama2004using,Benhur2005,park2009pairwise,Hayashi2012,Bonilla2007,pahikkala2013efficientcondrank}, resulting in the following mathematical model: $$\Psi(\vec{x},\vec{t}) = \phi(\vec{x}) \otimes \psi(\vec{t}) \,, \qquad \Gamma((\vec{x},\vec{t}),(\bar{\vec{x}},\bar{\vec{t}})) = k(\vec{x},\bar{\vec{x}}) \cdot g(\vec{t},\bar{\vec{t}}) \,,$$ with $\phi$ and $\psi$ feature maps in primal form, and $k(\vec{x},\bar{\vec{x}})$ and $g(\vec{t},\bar{\vec{t}})$ any kernel on the instance and target space in dual form. For example, for the drug design application of Example~4, one would typically develop domain-specific kernels for both $k(\vec{x},\bar{\vec{x}})$ and $g(\vec{t},\bar{\vec{t}})$. In the student grading application of Example~6, one would more likely apply conventional kernels for vectors on the given feature representations. For the document categorization application of Example~5, one would perhaps prefer to use other methods from the domain of hierarchical multi-label classification, but one could also apply a kernel that mimics the hierarchical structure, such as the shortest path kernel. As an aggregation of the instance kernel $k(\vec{x},\bar{\vec{x}})$ and the target kernel $g(\vec{t},\bar{\vec{t}})$, the Kronecker product pairwise kernel might be a reasonable first choice in experimental studies, as it is known to exhibit universal approximation properties \citep{Stock2016}. Variants also exist for modelling additional domain knowledge concerning relationships between instances and targets \citep{Vert2007,pahikkala2010reciprocalkm,Waegeman2012,pahikkala2013efficientcondrank}). Pairwise kernels are much more general building blocks than the output kernels that were discussed in Sections~\ref{sec:methods-that-use-target-relations} and~\ref{sec:methods-that-learn-target-relations}. Pairwise kernels model a similarity score based on target representations, whereas no target representations are available for output kernels. Many pairwise kernels can be easily plugged into standard kernel methods for binary classification or regression, such as support vector machines or kernel ridge regression. Approaches of that kind are often critisized for computational reasons, and claimed to be infeasible for large sample studies, but efficient implementations based on algebraic shortcuts exist \citep{VanLoan2000,Kashima2009,Raymond2010scalable,Alvarez2012,pahikkala2013efficientcondrank}. More recently, \citet{Pahikkala2014} and \citet{Romera-paredes2015} independently proposed algorithms that avoid the explicit construction of pairwise kernels. Those algorithms have computational advantages w.r.t.\ cross-validation and online training. Most importantly, pairwise learning methods are capable of generalizing to zero-shot problems, such as a novel target molecule in the drug design example, a novel tag in the document annotation example, or a novel course in the student grading example. In recent years, specific zero-shot learning methods based on deep learning have become extremely popular in image classification applications. The central idea in all those methods is to construct semantic feature representations for class labels, for which various techniques might work. One class of methods constructs binary vectors of visual attributes \citep{Lampert2009,Palatucci2009,Liu2011,Fu2013}. Figure~\ref{fig:semantic} shows an example of two such vectors for two images. Another class of methods rather considers continuous word vectors that describe linguistic context of images \citep{Mikolov2013,Frome2013,Socher2013}. For the same two images of Figure~\ref{fig:semantic}, one could for example look up ``zebra" and ``whale" on Wikipedia, and extract a Word2Vec representation from this source. \begin{figure} \begin{center} \includegraphics[scale=0.4]{semanticclasses} \end{center} \caption{Two examples of labeled images where the labels ``zebra" and ``whale" are described by four visual attributes. Examples taken from the CVPR 2016 Tutorial on Zero-shot learning for Computer Vision.} \label{fig:semantic} \end{figure} Many zero-shot learning methods for image classification adopt principles that originate in kernel methods. The model structure can often be formalized as follows: \begin{equation} \label{eq:zero-shot-pairwise} f(\vec{x},\vec{t}) = \vec{w}^T \big(\phi(\vec{x}) \otimes \psi(\vec{t})\big) \end{equation} This model in fact coincides with the primal formulation of (\ref{eq:zero-shot-pairwise}) with a Kronecker-based choice for $\Psi(\vec{x},\vec{t})$. Different optimization problems with this model have been proposed \citep{Frome2013,Akata2015,Akata2016}, and related methods provide nonlinear extensions \citep{Socher2013,Xian2016b}. Most of these optimization problems do not minimize squared error loss, and according to the definitions of Section~2, they should rather be seen as structured output prediction methods. Indeed, a representation such as (\ref{eq:pairwise}) is in fact commonly used in structured output prediction. SOP methods additionally have inference procedures that allow for finding the best-scoring targets in an efficient manner. Some of the zero-shot learning methods from computer vision also turn out to be useful for the related field of text classification. For documents, it is natural to model a latent representation for both the (document) instances and class labels in a joint space \citep{Nam2016}. \subsection{Representation-constructing methods} \label{sec:methods-that-learn-target-representations} In this section, we discuss methods that construct vector representations for targets without assuming side information as prior knowledge. So, in contrast to the methods that we discussed in Section \ref{sec:methods-that-use-target-representations}, no target representations need to be available. Instead, such representations are learned from scratch in the training phase\footnote{Let us remark that our notion of ``representation-constructing" differs substantially from the notion of ``representation learning" as commonly used in the area of deep neural networks. Here, we consider the construction of vector representations for targets. This is something that is not commonly done in multi-target prediction extensions of deep architectures.}. As a result, these methods are mainly useful for Settings B and C, while not being applicable to Setting D. In principle, target representations can also be learned for Setting A. However, since this involves additional complications, methods for Setting A will be discussed in Section~3.7. Methods that learn vector representations from scratch are in fact embedding-based methods. They all proceed from the assumption that the targets can be embedded in a vector space with a dimension lower than $m$. Principal component analysis (PCA), canonical correlation analysis (CCA), and partial least-squares (PLS) are simple techniques of that kind. In the case of PCA, one first embeds the targets in a low-dimensional space that is spanned by the most important principal components. In a second stage, one constructs regression models with the original features and the retained principal components as response variables. Predictions for new observations are then obtained by feeding them to the regression models and applying the inverse transformation to the original space. CCA and PLS both solve a different formulation, in which targets and features are embedded jointly in a lower-dimensional space~\citep{Shawe-Taylor2004,Yu_et_al_2006}. One can also obtain an embedding in another way, by introducing a rank constraint on the parameter matrix $A$ during optimization: $$\min_A ||Y - XA ||^2_F + \lambda \, \mathrm{rank}(A) $$ For example, \citet{Ando2005} and \citet{Chen2009} have shown that such an approach yields performance improvements in multi-task learning. Similar formulations for multivariate regression can be found in the statistical literature, starting in the seventies with well-known methods such as reduced-rank regression (RRR)~\citep{Izenman_1975} and FICYREG~\citep{Merwe_Zidek_1980}. According to~\citet{Breiman_Friedman_1997}, these methods have the same generic form: $$ \vec{f}^*(\vec{x}) = (T^{-1}GT)A\vec{x}\, , $$ where $T$ is the matrix of sample canonical coordinates, the solution of canonical correlation analysis (CCA), and the diagonal matrix $G$ contains the shrinkage factors for scaling the solutions of ordinary linear regression $A$. Based on this equation, \citet{Breiman_Friedman_1997} also establish a close connection with stacking, which was discussed in Section~\ref{sec:methods-that-do-not-use-domain-knowledge}. While embedding-based methods originated in multivariate regression, they have been used extensively for multi-label classification (as well as structured output prediction) in recent years. Nowadays they are very popular in the area of extreme multi-label classification, when the number of labels $m$ is extremely large. Then, embedding-based methods are needed for computational reasons, namely to obtain reasonable training and prediction times. They mainly differ in the choice of compression or reduction technique used. One can distinguish methods based on PCA \citep{Weston_et_al_2002}, CCA \citep{Rai2009}, compressed sensing \citep{Hsu_et_al_2009}, singular value decompostion \citep{Tai_Lin_2010,Tai2012}, output codes \citep{Zhang2011}, landmark labels \citep{Balasu2012,Bi2013}, Bloom filters \citep{Cisse2013}, auto-encoders \citep{Wicker2016}, etc. In spite of their success, there is also evidence that embedding-based methods may have important shortcomings. In many extreme multi-label classification datasets, the low-rank assumption is violated as a result of the presence of many so-called tail labels. These are labels that appear very rarely in the dataset, sometimes less than five times in total, and recovering them with low-rank matrix approximation methods is very difficult \citep{Bhatia2015,Yen2016}. \subsection{Matrix completion and hybrid methods} \label{sec:matrix-completion-and-hybrid-methods} In this section, we review matrix completion methods. In Section~2, such methods were claimed to be useful for an MTP setting with partially-observed matrices $Y$---in Figure~\ref{fig:foursettings} referred to as Setting A. Both the targets and the instances are observed, but not for all instance-target combinations. In Setting A, side information about instances or targets is not required per se. We hence distinguish between methods that ignore side information and methods that also exploit such information, in addition to analyzing the matrix $Y$. Inspired by the Netflix challenge in 2006, the former type of methods has been mainly popular in the area of recommender systems. Those methods often impute missing values by computing a low-rank approximation of the sparsely-filled matrix $Y$, and many variants exist in the literature, including algorithms based on nuclear norm minimization \citep{Candes2008}, Gaussian processes \citep{Lawrence2009}, probabilistic methods \citep{Shan2010}, spectral regularization \citep{Mazumder2010}, non-negative matrix factorization \citep{Gaujoux2010}, and alternating least-squares minimization \citep{Jain2013}. In addition to recommender systems, matrix factorization methods are commonly applied to social network analysis \citep{Menon2010}, biological network inference \citep{Gonen2012,Liu2015}, and travel time estimation in car navigation systems~\citep{Dembczynski_et_al_2013}. Let us stress that those methods differ substantially from the methods that were discussed in Section~\ref{sec:methods-that-learn-target-representations}. Here, matrix factorization is used to impute missing values in $Y$, while in Section~\ref{sec:methods-that-learn-target-representations} it was needed to reduce the dimensionality of the target space. In other words, here we are discussing methods that have been proposed for Setting A, whereas the methods from Section~\ref{sec:methods-that-learn-target-representations} are applicable to Setting B. In addition to matrix factorization, a few other methods exist for Setting A. Historically, memory-based collaborative filtering has been popular, and corresponding methods are very easy to implement. They make predictions for the unknown cells of the matrix by modelling a similarity measure between either rows or columns---see e.g.\ \citep{Takacs2008}. For example, when rows and columns correspond to users and items, respectively, then one can predict novel items for a particular user by searching for other users with similar interests. In the resulting nearest neighbor search, the cosine similarity often outperforms other similarity measures. Many variants of matrix factorization and other collaborative methods have been presented, in which side information of rows and columns is considered during learning, in addition to exploiting the structure of the matrix $Y$---see e.g.\ \citep{Basilico2004,Abernethy2008,Adams2010,Fang2011,Zhou2011a,Menon2011,Zhou2012a}. One simple but effective method is to extract latent feature representations for instances and targets in a first step, and combine those latent features with explicit features in a second step, using any of the more conventional approaches that were discussed in the previous paragraphs \citep{Volkovs2012}. In light of Examples~1--3, when feature representations are available for proteins, documents, and students, it would be pointless to ignore them. Hybrid methods seek to combine the best of both worlds, by simultaneously modeling side information and the structure of $Y$. In addition to Setting A, they can often be applied to Settings B and C, which coincide, respectively, with a novel user and a novel item in recommender systems. In that context, one often speaks about cold-start recommendations. \section{Conclusion and future perspectives} In this paper, we provided a unifying view on MTP problems and methods. In Section 2, we presented a general MTP framework and explained how well-known subfields of machine learning can be seen as specific instantiations of this framework. In Section 3, we gave an overview of MTP methods, categorized according to the prediction setting of interest and the domain knowledge that is available. For researchers who are new to the field of multi-target prediction, this categorization might help identify which method to use for a given MTP problem. As an overall conclusion, it should be clear that no MTP method is applicable, let alone optimal, under all conditions. It is up to the data scientist to choose the right method for a given problem, and we hope that this article will be helpful in this regard. There are several topics related to multi-target prediction that are not discussed in this paper. One important issue is the loss one seeks to optimize. As briefly discussed in Section~3, we would like to emphasize that the methods reviewed in this paper are mainly useful for optimizing loss functions that are decomposable over targets---our example of mean squared error is exactly of that kind. For more complex loss functions, which mainly seem to appear in the multi-label classification literature, other algorithms will be needed. We refer to our previous work for an overview on this topic, with a particular focus on the subset zero-one loss, the rank loss, and the F1-measure \citep{Demb2012a,Waegeman2014}. We believe that research on multi-target prediction will be further intensified in the coming years, because novel applications are constantly emerging. Besides, several research questions still remain unanswered. In general, one can observe a trend towards analyzing larger target spaces, which require sophisticated algorithms that scale linearly or sublinearly in the number of targets. One also observes a trend towards extending classical machine learning paradigms for an MTP context---including semi-supervised learning, multi-instance learning, time series classification, data stream mining, network analysis, etc. We are convinced that extensions of that kind will be a driving force of fundamental machine learning research in the coming years. Finally, there is also a need for further theoretical research in multi-target prediction. Many of the recent algorithms rely on a purely heuristic or intuitive motivation, which is then tested (and usually confirmed) in empirical studies on classical benchmark datasets. Most of the time, however, a deeper understanding of those algorithms and a theoretical explanation of their behavior are missing. Needless to say, an understanding of that kind is needed to mature the field of multi-target prediction, and to unify the different branches discussed in this paper. \section*{Appendix A: James-Stein Estimation} \label{sec:James-Stein-Estimation} In the late sixties, James and Stein discovered that the best estimator of the mean of a multivariate Gaussian distribution is not necessarily the maximum likelihood estimator. More formally, assume that $\theta$ is the unknown mean of a multivariate Gaussian distribution with dimension $m>2$ and a diagonal covariance matrix. Consider a single observation $\vec{y}$ randomly drawn from that distribution: $$\vec{y} \sim \mathcal{N}(\theta, \sigma^2 I) \,.$$ Using only this observation, the maximum-likelihood estimator for $\theta$ would be $\hat{\theta}_{ML} = \vec{y}$. James and Stein discovered that the maximum likelihood estimator is suboptimal in terms of mean squared error $$\mathbb{E} \big[||\theta - \hat{\theta}||^2 \big] \,, $$ where the expectation is over the distribution of $\vec{y}$. (In general, the expectation is taken over all samples that contain a single observation $\vec{y}$. Later on we will shortly discuss a situation in which we draw more than one observation to compute the value of the estimator). An estimator with lower squared error can be obtained by applying a regularizer to the maximum likelihood estimator. In case $\sigma^2$ is known, the James-Stein estimator is defined as follows: $$\hat{\theta}_{JS} = \left(1 - \frac{(m-2)\sigma^2}{||\vec{y}||^2} \right) \vec{y} \,.$$ From a machine learning perspective, a regularizer is introduced that shrinks the estimate towards the zero vector, and hence reduces variance at the cost of introducing a bias. It has been shown that this biased estimator outperforms the maximum likelihood estimator in terms of mean squared error. The result even holds when the covariance matrix is non-diagonal, but in view of the discussion concerning target dependence, it is most remarkable for diagonal covariance matrices. In fact, in the latter case, it means that joint target regularization will be beneficial even if targets are intrinsically independent. This is somewhat in contradiction with what is commonly assumed in the machine learning literature. Let us notice, however, that the advantage of the James-Stein estimate over the maximum likelihood estimate will vanish for larger samples (of more than one observation). In the second term in parentheses, $\sigma^2$ is then divided by the size of the sample, so that the James-Stein estimate converges to the maximum likelihood estimate when the sample size grows to infinity. The James-Stein paradox analyzes a very simple estimation setting, for which suboptimality of the maximum likelihood estimator can be proved analytically, but the principle extends to various multi-target prediction settings. By interpreting each component of $\theta$ as an individual target (and omitting the instance space, or reducing it to a single point), the maximum likelihood estimator coincides with independent model fitting, whereas the James-Stein estimator adopts a regularization mechanism that is very similar to most of the regularization techniques used in the machine learning literature. For some specific multivariate regression models, connections of that kind have been discussed in the statistical literature \citep{Breiman_Friedman_1997}. As long as mean squared error is considered as a loss function and errors follow a Gaussian distribution, one can immediately extend the James-Stein paradox to multivariate regression settings by assuming that target vectors $\vec{y}$ are generated according to the following statistical model: $$\vec{y} \sim \mathcal{N}(\theta(\vec{x}), \sigma^2 I) \, , $$ where the mean is now conditioned on the input space. For other loss functions, we are not aware of any formal analysis of that kind, but it might be expected that similar conclusions can be drawn. \bibliographystyle{spbasic}

8cbe345642b1383efae7c05a4cdf9ff4261ad891

\section{0pt}{12pt plus 3pt minus 3pt}{6pt plus 0pt minus 3pt} \newcommand{\addreferencesection}{ \phantomsection \addcontentsline{toc}{section}{References} } \newcommand{\Rucha}[1]{{\color{green} #1}} \newcommand{{network-coordination game}\xspace}{{network-coordination game}\xspace} \newcommand{{network-coordination games}\xspace}{{network-coordination games}\xspace} \newcommand{\alert}[1]{} \newcommand{\rhl}[1]{{\textcolor{black}{#1}}} \newcommand{{\mathrm{payoff}}}{{\mathrm{payoff}}} \newcommand{{{\bm \sigma}}}{{{\bm \sigma}}} \newcommand{{{\mbox{argmax}}}}{{{\mbox{argmax}}}} \newcommand{{{\mathscr P}}}{{{\mathscr P}}} \newcommand{{{\mathscr Q}}}{{{\mathscr Q}}} \newcommand{{{\mathscr V}}}{{{\mathscr V}}} \newcommand{{{\mathscr A}}}{{{\mathscr A}}} \newcommand{{transformation vector}\xspace}{{transformation vector}\xspace} \newcommand{{transformation set}\xspace}{{transformation set}\xspace} \newcommand{{\mathscr{L}}}{{\mathscr{L}}} \newcommand{\rnote}[1]{{\color{blue}\bf ** Ruta: #1}} \title{Smoothed Efficient Algorithms and Reductions for Network Coordination Games} \author{Shant Boodaghians} \author{Rucha Kulkarni} \author{Ruta Mehta\footnote{This work was supported by NSF grant CCF-1750436}} \affil{Department of Computer Science,\\ University of Illinois at Urbana-Champaign,\\ \texttt{\{boodagh2,ruchark2\}@illinois.edu, [email protected]} } \date{} \newtheorem{thm}{Theorem}[section] \newtheorem{lem}[thm]{Lemma} \newtheorem{claim}[thm]{Claim} \newtheorem{prop}[thm]{Proposition} \newtheorem{cor}[thm]{Corollary} \newtheorem{conj}{Conjecture}[section] \newtheorem*{claim*}{Claim} \theoremstyle{definition} \newtheorem{exmp}{Example}[section] \newtheorem{rem}{Remark} \newtheorem{defn}{Definition}[section] \newcommand{\ip}[1]{\left\langle #1 \right\rangle} \newcommand{\mathbb R}{\mathbb R} \newcommand{\{1,\,\dotsc,\,k\}}{\{1,\,\dotsc,\,k\}} \newcommand{\textsuperscript}{\textsuperscript} \newcommand{{NetCoordNash}\xspace}{{NetCoordNash}\xspace} \newcommand{{FlipMaxCut}\xspace}{{FlipMaxCut}\xspace} \newcommand{{Congestion}\xspace}{{Congestion}\xspace} \usepackage{tikz} \usepackage[mathscr]{euscript} \usepackage{natbib} \bibliographystyle{plainnat} \begin{document} \thispagestyle{empty} \maketitle \begin{abstract} Worst-case hardness results for most equilibrium computation problems have raised the need for {\em beyond-worst-case} analysis. To this end, we study the smoothed complexity of finding pure Nash equilibria in Network Coordination Games, a PLS-complete problem in the worst case. This is a potential game where the {\em sequential-better-response} algorithm is known to converge to a pure NE, albeit in exponential time. First, we prove polynomial (resp. quasi-polynomial) smoothed complexity when the underlying game graph is a complete (resp. arbitrary) graph, and every player has constantly many strategies. We note that the complete graph case is reminiscent of perturbing {\em all parameters}, a common assumption in most known smoothed analysis results. \vspace*{2pt} Second, we define a notion of {\em smoothness-preserving reduction} among search problems, and obtain reductions from $2$-strategy network coordination games to local-max-cut, and from $k$-strategy games (with arbitrary $k$) to local-max-cut up to two flips. The former together with the recent result of \cite{BCC18} gives an alternate $O(n^8)$-time smoothed algorithm for the $2$-strategy case. This notion of reduction allows for the extension of smoothed efficient algorithms from one problem to another. For the first set of results, we develop techniques to bound the probability that an (adversarial) better-response sequence makes slow improvements on the potential. Our approach combines and generalizes the local-max-cut approaches of \cite{ER14,A+17} to handle the multi-strategy case: it requires a careful definition of the matrix which captures the increase in potential, a tighter union bound on adversarial sequences, and balancing it with good enough rank bounds. We believe that the approach and notions developed herein could be of interest in addressing the smoothed complexity of other potential and/or congestion games. \end{abstract} \newpage \thispagestyle{empty} \setcounter{tocdepth}{2} \tableofcontents \newpage \pagenumbering{arabic} \section{Introduction}\label{sec:introduction} \input{Section-1-Introduction} \section{Game Model, Smoothed Analysis, and Statement of Results} \input{Section-2-Model-and-Prelims} \section{Overview of Our Results and Techniques} \input{Section-3-Overview} \paragraph{Acknowledgment.} We would like to thank Pravesh Kothari for the insightful discussions in the initial stages of this work. \addreferencesection \bibliographystyle{ACM-Reference-Format} \subsection Smoothed Algorithms: Challenges and Techniques} \label{sec:challenge} Our results follow a framework which is common to past work on smoothed algorithms for local max cut~\cite{ER14,A+17,BCC18} (see Section~\ref{sec:common}). The goal is to show that with high probability {\em every} sufficiently long sequence of improving moves (of the local-search algorithm) is very likely to significantly improve the potential. This is shown via a two-step process hinging on Lemma~\ref{lem:probability}: first, we represent the potential improvement in every step as a linear combination of the input parameters and consider the corresponding matrix for a subsequence, and second, show a tight union bound on the number of different sequences and relevant initial configurations, paired with a high rank-bound for the matrix. Several obstacles were encountered while trying to apply the general framework to the $k$-{NetCoordNash}\xspace problem. The first was how to correctly represent each move in a better-response (BR) sequence. Specifying only the moving player is insufficient, as every player has more than $2$ strategies to choose from. A move is entirely specified by the triple (player, strategy-from, strategy-to). This, however, is too descriptive and the union bound is too large. Labelling as (player, to-strategy) suffices, and strikes the right balance between rank and union bound. This definition, however, muddies the proof technique further on. \begin{itemize} \item \textbf{Rank analysis:} When the BR sequence represents moves as (player, strategy) pairs, within a sequence, some {\em pairs} may be repeated, and some {\em players} may be repeated, but always playing distinct strategies. Therefore, notions of repeating and non-repeating players need to be carefully defined. Secondly, we would want the nodes in the directed-influence-graph arguments of Section~\ref{sec:cyclic}, used for showing rank bounds in terms of repeat-moves, to also be labeled as (player,strategy) pairs. This would have multiple nodes in the graph corresponding to the same players, each influencing multiple players. This makes for messy analysis and poor rank bounds. We instead label the nodes of these influence graphs as simply the players, unlike the improving sequence itself. This causes rank bounds to depend on the number of players with repeating moves, rather than the number of repeating moves. Thus, notions of {\em critical subsequences} and {\em separated blocks} need to be carefully adapted from past notions. \item \textbf{Defining \textit{critical subsequences} and \textit{separated blocks}:} We show that moves by non-repeating players allow us to ``separate'' the sequence and combine the rank bounds from both sides. This leads to the notion of {\em separated blocks,} which requires a careful selection of the boundary moves. To combine these bounds, we use the idea of a {\em critical block} from~\cite{A+17}, which was very important to their poly-time bound, and is similarly helpful in our result. It would have been preferable to use any non-repeated move as the separators, but it is not possible to do so, as the same player may also repeat other moves. These notions must be adapted carefully so that the rank and union bounds balance each other. This, overall, loosens the rank bound. \item \textbf{Union bounds:} The union bound analysis must now bear the brunt of the looser rank bounds above, and must be made tighter through properties of the $k$-{NetCoordNash}\xspace problem. Eliminating the influence of the inactive players allows us to avoid having to take their strategies into account. Taking a sum of repeated moves does not suffice in our setting, and we define the notion of ``cyclic sums'' to handle this. \end{itemize} \subsection{Related Work}\label{sec:related-work} The works most related to ours are \cite{ER14} and \cite{A+17} analyzing smoothed complexity of local-max-cut; see Section \ref{sec:challenge} for a detailed comparison. Independently, Bibaksereshkeh, Carlson, and Chandrasekaran \cite{BCC18} improved the running-time for the local-max-cut algorithm, and obtained smoothed polynomial and quasi-polynomial algorithms for local-$3$-cut and local-$k$-cut with constant $k$ respectively. Reduction of Figure \ref{fig:maxcut-to-netco} easily extends to reduce local-$k$-cut to $k$-strategy {network-coordination games}\xspace, implying that the latter significantly generalizes the former. However, the reduction is not {\em smoothness-preserving}, and hence our smoothed efficient algorithm is not directly applicable to solve the local-$k$-cut problem. \paragraph{Beyond-worst case complexity of NE computation.} The smoothed-efficient algorithm for local-max-cut directly gives one for party affiliation games~\cite{FPT04}. For two-player games which are non-potential in general, Chen, Deng, and Teng \cite{CDT06Smooth} ruled out polynomial smoothed complexity unless RP=PPAD. While, towards average case analysis B\'ar\'any, Vempala, and Vetta \cite{BVV07} showed that a game picked uniformly at random has a NE with support size two for both the players whp. The average case complexity of a random potential game was shown to be polynomial in the number of players and strategies by Durand and Gaujal \cite{DG16}. \paragraph{Worst-case analysis.} Potential games and equivalently congestion games have been studied at length (e.g., \cite{Rosenthal73, MS96, RT02, FPT04, CMN05}), capturing routing and traffic situations (e.g., \cite{Smith79, DN84, Rgarden07, HS10, TMMA13,nw1}), and resource allocation under strategic agents (e.g., \cite{JT04,FT12,cong1}). Unlike general games, existence of the potential function ensures that these games always have a pure NE \cite{Rosenthal73}. Finding pure NE is typically PLS-complete \cite{FPT04,CD11}, while finding any NE, mixed or pure, is known to be in CLS \cite{DP11}, a class in the intersection of PPAD and PLS. A remarkable collection of work studies the loss in welfare at NE through the notions of Price-of-Anarchy and Price-of-Stability (e.g., \cite{KP99,RT02,CK05,ADGMS06,ADKTWR08,spoa,poa1}). Our approach should help provide ways to obtain smoothed efficient algorithms for these games. Worst case complexity of NE computation in general non-potential games has been studied extensively. The computation is typically PPAD-complete, even for various special cases (e.g., \cite{AKV05,CDT06,Mehta14,inbal}) and approximation (e.g., \cite{CDT06Smooth,Rub16}). On the other hand efficient algorithms have been designed for interesting sub-classes (e.g., \cite{KT07,TS,Imm11,AGMS11,CD11,CCDP15,ADHLMS16,BB17,B18}), exploiting the structure of NE for the class to either enumerate, or through other methods such as parameterized LP and binary search. For two-player games, Lipton, Mehta, and Markakis gave a quasi-polynomial time algorithm to find a constant approximate Nash equilibrium \cite{LMM03}. Recently, Rubinstein \cite{Rub17ETH} showed this to be the best possible assuming exponential time hypothesis for PPAD, and Kothari and Mehta \cite{KM18} showed a matching unconditional hardness under the powerful algorithmic framework of Sum-of-Squares with oblivious rounding and enumeration. These results are complemented by communication \cite{RubCom1,RubCom2} and query complexity lower bounds \cite{Que2,Que1,Que3}. Lower bounds in approximation under well-accepted assumptions have been studied for the decision versions \cite{GZ,CS,HK,BKW,DFS}. \begin{comment} There has been extensive work on various potential games, equivalently congestion games (e.g., \cite{Rosenthal73, MS96, RT02, FPT04, CMN05}), capturing routing and traffic situations (e.g., \cite{Smith79, DN84, Rgarden07, HS10, TMMA13}), and resource allocation under strategic agents (e.g., \cite{JT04,FT12}). Unlike general games, existence of the potential function ensures that these games always have a pure NE \cite{Rosenthal73}. Finding pure NE is typically PLS-complete \cite{FPT04,CD11}, while finding any NE, mixed or pure, is known to be in CLS \cite{DP11}, a class in the intersection of PPAD and PLS. The average case complexity of a random potential game was shown to be polynomial in the number of players and strategies by Durand and Gaujal recently \cite{DG16}. Another series of remarkable works have studied the loss in welfare at NE through the notions of Price-of-Anarchy and Price-of-Stability (e.g., \cite{KP99,RT02,CK05,ADGMS06,ADKTWR08}). To the best of our knowledge no smoothed complexity results are known for these games, and our approach to analyze potential increase may provide ways to obtain smoothed efficient algorithms for other potential games. The only known smoothed complexity results for finding NE are for party affiliation games\cite{FPT04}, implied by the polynomial smoothed complexity of local-max-cut\cite{A+17}, and for two-player games which are non-potential games in general. Chen, Deng, and Teng \cite{CDT06Smooth} showed that this problem does not have polynomial smoothed complexity unless RP=PPAD. While, towards average case analysis B\'ar\'any, Vempala, and Vetta \cite{BVV07} showed that a game picked uniformly at random has a NE with support size two for both the players whp. Worst case complexity of NE computation in general non-potential games has been studied extensively. The computation is typically PPAD-complete, even for various special cases (e.g., \cite{AKV05,CDT06,Mehta14}) and approximation (e.g., \cite{CDT06Smooth,Rub16}). On the other hand efficient algorithms have been designed for interesting sub-classes (e.g., \cite{KT07,TS,Imm11,AGMS11,CD11,CCDP15,ADHLMS16,BB17,B18}), exploiting the structure of NE for the class to either enumerate, or through other methods such as parameterized LP and binary search. For two-player games, Lipton, Mehta, and Markakis gave a quasi-polynomial time algorithm to find a constant approximate Nash equilibrium \cite{LMM03}. Recently, Rubinstein \cite{Rub17ETH} showed this to be the best possible assuming exponential time hypothesis for PPAD, and Kothari and Mehta \cite{KM18} showed a matching unconditional hardness under the powerful algorithmic framework of Sum-of-Squares with oblivious rounding and enumeration. Lower bounds in approximation under well-accepted assumptions, like NP-hardness, SETH, and planted-clique, have also been studied for the decision versions of the problem, {\em e.g.,} if there exist more than one equilibrium, equilibrium with payoff at least $h$, etc. \cite{GZ,CS,HK,BKW,DFS}. The works most related to ours are \cite{ER14} and \cite{A+17}. As discussed in the previous section, both study the smoothed complexity of the FLIP algorithm for local-max-cut, which is also PLS-complete. The first gives a quasi-polynomial time analysis for FLIP in general (when applied on local-max-cut on arbitrary graphs), and the second, a polynomial time analysis in the case of complete graphs. Since the {network-coordination game}\xspace problem significantly generalizes the local-max-cut problem, more novel techniques and ideas are needed for the analysis to go through. However, a common theme among these results and ours, as well as the works of Arthur et al for the ICP algorithm for k-means method \cite{AV09} and Englert et al for the $2$-OPT algorithm for the travelling salesman problem \cite{ERV14}, is that they all track improvement in some function which acts as a measure of progress. Independently, Bibaksereshkeh, Carlson, and Chandrasekaran \cite{BCC18} improved the running-time analysis for the local-max-cut algorithm, and obtained smoothed polynomial and quasi-polynomial algorithms for local-$3$-cut and local-$k$-cut respectively. In local-$k$-cut the goal is to find a $k$-partition of the vertices such that the total weight of edges crossing any partition can not be improved by moving any one vertex. It is easy to see that local-$k$-cut for any $k$ reduces to $k$-strategy {network-coordination games}\xspace, where the weight $w_{uv}$ for some edge $uv$ translates to a payoff matrix $A_{uv}$ in the game, where $A_{uv}(i,j)=0$ if $i=j$, and $w_{uv}$ otherwise. Thus, {network-coordination games}\xspace are at least as hard as the significantly generalized local-$k$-cut problem. On the other hand, this reduction is not {\em smoothness-preserving}, and hence our smoothed efficient algorithm is not directly applicable to solve the local-$k$-cut problem. \end{comment} \subsection{Nash Equilibria in Network Coordination Games}\label{sec:PrelGame} A game with two players, where each player has $k$ strategies, can be defined by two $k \times k$ payoff matrices $(A,B)$, one for each player. It is called a {\em coordination game} if $A=B$. We assume without loss of generality that every player has the same number of strategies. \paragraph{$k$-{network-coordination game}\xspace.} A {network-coordination game}\xspace is a multi-player extension of coordination games. The game is specified by an underlying undirected graph $G=(V,E)$, where the nodes are players, and each edge represents a two-player coordination game between its endpoints. A $k$-{network-coordination game}\xspace is where each player has $k$ strategies, and the edge games are represented by $k\times k$ matrices $A_{uv}$. If $u$ plays $i\in[k]$ and $v$ plays $j\in [k]$ then both get payoff $A_{u,v}(i,j)$ on this edge; we will sometimes denote this by $A((u,i)(v,j))$ to disambiguate. Nash equilibria are invariant to shifting and scaling of the payoffs, so $w.l.o.g.$ we assume every entry of $A_{uv}$ is contained in $[-1,1]$. Let $n=|V|$. A {\em strategy profile} is a vector ${{\bm \sigma}}\in[k]^n$ where each player chooses a strategy from $[k]$. The payoff of player $u$ is then: \[ {\mathrm{payoff}}_u({{\bm \sigma}}) := \textstyle\sum_{v:\: uv\in E} A_{uv}(\sigma_u,\sigma_v) \] \paragraph{Nash Equilibrium.} At a Nash equilibrium (NE) no player gains by deviating unilaterally. In general, NE strategy profiles may be randomized. However, a NE where every player chooses a strategy deterministically is called {\em pure Nash equilibrium} (PNE). Formally, strategy profile ${{\bm \sigma}}$ is a PNE, if and only if $ \forall u \in V,\ \ {\mathrm{payoff}}_u({{\bm \sigma}}) \ge {\mathrm{payoff}}_u(\sigma'_u,{{\bm \sigma}}_{-u}),\ \ \forall \sigma'_u \in [k], $ where ${{\bm \sigma}}_{-u}$ denotes the strategies of all the players in ${{\bm \sigma}}$ except $u$. In {\em Potential Games} \cite{Rosenthal73}, PNE's are known to always exist. Such a game admits a {\em potential function} which encodes the individual ``progress'' of the players, {\em i.e.} $\exists g:[k]^n \rightarrow \mathbb R$ such that for all ${{\bm \sigma}}\in[k]^n$, $ g({{\bm \sigma}}) - g(\sigma'_u,{{\bm \sigma}}_{-u}) = {\mathrm{payoff}}_u({{\bm \sigma}}) - {\mathrm{payoff}}_u(\sigma'_u,{{\bm \sigma}}_{-u}),\ \ \ \forall u, \forall \sigma'_u \in [k] $ Clearly, every local-maximum of $g$, {\em i.e.} ${{\bm \sigma}}$ such that $g({{\bm \sigma}}) \ge g(\sigma'_u,{{\bm \sigma}}_{-u}),\ \forall u, \forall \sigma'_u\in[k]$, is a pure NE. \begin{lem}[\cite{CD11}] Network Coord.{ }Games are potential games with potential function \begin{equation}\label{eq:pot} {\mathrm{payoff}}({{\bm \sigma}}) =\textstyle \sum_{(u,v) \in E} A_{uv}(\sigma_u,\sigma_v) = \frac12\sum_{u\in V} {\mathrm{payoff}}_u({{\bm \sigma}}) \end{equation} \end{lem} Our goal is to find a pure NE for a given $k$-{network-coordination game}\xspace. \paragraph{Better-Response Algorithm (BR alg., or BRA).} Another immediate consequence of being a potential game is that, for any strategy profile ${{\bm \sigma}}$, if some player $u$ can deviate to $\sigma'_u$ and improve her payoff, then the move $\sigma_u\mapsto\sigma'_u$ is termed a {\em better-response (BR) move} for player $u$ from strategy profile ${{\bm \sigma}}$. Clearly, under such a BR move, the potential function value increases, {\em i.e.,} ${\mathrm{payoff}}(\sigma'_u,\bm \sigma_{-u})-{\mathrm{payoff}}(\bm \sigma) = {\mathrm{payoff}}_u(\sigma'_u,\bm \sigma_{-u}) -{\mathrm{payoff}}_u(\bm \sigma)>0.$ Note that ${\mathrm{payoff}}({{\bm \sigma}})$ may only take $k^n$ possible values. Hence, if a BR move is made whenever possible, the players must converge to a local optimum of the potential function, or equivalently, to a pure NE of the game. This gives a local-search based {\em better-response} algorithm to solve $k$-{NetCoordNash}\xspace. \subsection{Smoothed Analysis and Reductions}\label{sec:PrelSmooth} The notion of smoothed analysis was introduced by Spielman and Teng \cite{ST04} to bridge the gap between average- and worst-case analysis. For a search problem ${{\mathscr P}}$ let $(I,X)$ be an instance where $I$ is possibly discrete information, and $X$ is a real-valued vector whose dimension depends on $I$. For example, in the case of a $k$-{network-coordination game}\xspace, $I$ consists of the game graph $G$ and the number of strategies $k$, and $X$ is the payoff vector $A$. \begin{defn}[Polynomial Smoothed Complexity $w.h.p.$ or in expectation]\label{def:Smooth} Let $(I,X)$ be a random instance of ${{\mathscr P}}$, where $I$ is chosen arbitrarily, and $X$ is a random real-valued vector whose entries are independent and have density at most $\phi$. If there exists an algorithm which solves arbitrary instances of ${{\mathscr P}}$ in finite time, and for all $I$, solves the random input $(I,X)$ in time at most $(\phi\cdot |I|\cdot |X|)^c$ for some $c>0$ with probability at least $1-1/poly(\phi,|I|,|X|)$, where $|X|$ denotes the number of entries in $X$, then ${{\mathscr P}}$ is said to have {\em polynomial smoothed complexity $w.h.p.$} If the same holds in expectation, then ${{\mathscr P}}$ is said to have {\em polynomial smoothed complexity in expectation.} \end{defn} Standard (Turing) reductions between two search problems ${{\mathscr P}}$ and ${{\mathscr Q}}$ are well-defined, and used to extend an algorithm of ${{\mathscr Q}}$ to solve instances of ${{\mathscr P}}$, or to imply hardness for ${{\mathscr Q}}$ given hardness for ${{\mathscr P}}$. We extend this notion to define {\em smoothness preserving reductions}. \begin{defn}[Strong and Weak Smoothness-Preserving Reductions]\label{def:SmoothRed} A {\em randomized, smoothness-preserving reduction} from a search problem ${{\mathscr P}}$ to ${{\mathscr Q}}$ is defined by poly-time computable functions $f_1$, $f_2$, and $f_3$, and a real probability space $\Omega \subseteq \mathbb R^d$, such that, \begin{itemize} \item For any $(I,X)\in {{\mathscr P}}$, and for arbitrary $R\in \Omega$, $\big(f_1(I),\, f_2(X,R)\big)$ is an instance of ${{\mathscr Q}}$, such that all (locally optimal) solutions ${{\bm \sigma}}$ map to a solution $f_3({{\bm \sigma}})$ of $(I,X)$. \item Whenever the entries of $X$ and $R$ are drawn independently at random from distributions with density at most $\phi$, then $f_2(X,R)$ has entries which are independent random variables with density at most $poly(\phi,|X|,|R|)$. This is called a {\em strong} reduction. \item If the entries of $f_2(X,R)$ instead of being independent are linearly independent combinations of entries of $X$ and $R$, and the density is similarly bounded, then call it a {\em weak} reduction. \end{itemize} \end{defn} \paragraph{$d$-{FlipMaxCut}\xspace.}\label{sec:PrelLMC} Extending the notion of local-max-cut \cite{SY91}, we define the $d$-{FlipMaxCut}\xspace problem, given by an undirected graph $G=(V,E)$ with edge weights $w_{uv}$ for all $uv\in E$. The goal is to find a non-empty subset $S\subsetneq V$ of vertices, such that $\delta(S) := \sum_{uv\in E:u\in S,v\notin S} w_{uv}$ is a local optimum up to $d$-flips, {\em i.e.,} $\delta(S)\geq \delta(S')$ for all $S'\subset V$ such that $S$ and $S'$ differ by at most $d$ vertices, or $|(S\setminus S') \cup (S'\setminus S)|\le d$. Note that, $1$-{FlipMaxCut}\xspace is the usual local-max-cut problem. For a constant $d$, the $d$-{FlipMaxCut}\xspace problem admits a natural {\em FLIP algorithm}, which like the BR algorithm, will check whether there exists a local improvement, and move the candidate solution to the improved solution until no local improvement is possible. \subsection{A Common Framework for Local-Improvement Algorithms}\label{sec:common} Observe that the potential function for $k$-{NetCoordNash}\xspace, as given in~\eqref{eq:pot}, is an integer linear combination of the payoff values, and is actually a 0--1 combination. The following framework can be applied whenever this holds, as long as the potential function's range is also polynomially bounded. It hinges on the following lemma: \begin{lem}[\cite{Rog08}]\label{lem:main-praobability} Let $X\in \mathbb R^d$ be a vector of $d$ independent random variables where each $X_i$ has density bounded by $\phi$. Let $\alpha_1,\,\dotsc,\,\alpha_r$ be $r$ linearly independent vectors in $\mathbb{Z}^d$. then the joint density of $(\ip{\alpha_i,X})_{i\in [r]}$ is bounded by $\phi^r$, and for any given $b_1,b_2,\dotsc \in \mathbb{R}$ and $\epsilon>0$,\\[-.5em] \begin{equation}\label{eq:prob-bound-intro} \Pr\Big[\textstyle\bigwedge_{i=1}^r\ip{\alpha_i,X}\in [b_i,b_i+\epsilon]\Big] \leq (\phi\epsilon)^r \end{equation} \end{lem} Here, $X$ constitutes the random inputs to the smoothed problem, and the $\alpha$ vectors represent the change in the potential function. Formally, if a potential function $\Phi(\bm \sigma)$ is given for some problem, and $\bm \sigma_1,\bm \sigma_2,\bm \sigma_3,\dotsc$ denotes the walk through the state space given by some local-improvement algorithm, then we have vectors $\alpha_i$ such that $\ip{\alpha_i,X}=\Phi(\bm \sigma_i)-\Phi(\bm \sigma_{i-1})$ for all $i$. Setting $b_i=0$, the above lemma upper bounds the probability that {\em every} step of the local-improvement algorithm is only a small improvement \mbox{($\ip{\alpha,X}<\epsilon$)}, while still being an improvement \mbox{($\ip{\alpha,X}>0$)}. Let $\mathcal E$ be the event in the probability statement, that is, $\textstyle\bigwedge_{i=1}^r\ip{\alpha_i,X}\in [0,0+\epsilon]$. If $\mathcal E$ does not hold, and the sequence is indeed an improving one, then at least one of the improvements must be at least $\epsilon$. If $\mathcal E$ does not hold for {\em any} sequence of $\Omega(n)$ moves, then we can bound the running time of the iterative algorithm by \[ \frac n \epsilon\cdot \left(\max_{\sigma} \Phi(\bm \sigma) - \min_{\sigma}\Phi(\bm \sigma)\right) \] Finally, if $\Phi$ is bounded in a (quasi)polynomially-sized range, and $\epsilon$ is taken to be $1/(quasi)poly(n)$, then we conclude that the procedure runs in (quasi)polynomial time. \paragraph{Rank Bound vs Union Bound.} Note that in order to get (quasi)polynomial running time with high probability, we must first upper-bound the probability of event $\mathcal E$, simultaneously for {\em all} sequences of $\Omega(n)$ local-improvement moves, for which we simply take the union bound. To counteract this large union bound, we must lower-bound the rank of the matrix $[\alpha_i]_{i=1}^n$. It remains then to choose $\epsilon$ correctly to counteract the union bound, and proving the best rank in general. This highlights the main technical challenge when applying the common framework: labelling the moves. If a move's label is too descriptive --- {\em e.g.} a full state vector --- then the union bound will be much too large. However, if a move's label is too vague --- {\em e.g.} denoting a move only by the player who is moving, not the strategy --- then the rank bound will not be large enough. We introduce, in the next section, those parameters which work in our setting. At a high-level, our analysis follows the framework of the previous local-max-cut papers~\cite{ER14,A+17}. However, since players have multiple strategies, this poses some challenges for the technical details. We keep notation consistent whenever possible, to allow making analogies to past approaches. \subsection{Notation}\label{sec:notation} Recall the problem of $k$-{NetCoordNash}\xspace defined above, and the better-response algorithm (BRA) discussed in Section~\ref{sec:PrelGame}. We represent each better-response (BR) move by a player-strategy pair $(u,i)$, denoting that player $u$ is replacing strategy $\sigma_u$ by $i$, (assuming $i\neq \sigma_u$). We also denote as~${{\bm \sigma}}^t$ the strategy profile after the $t$\textsuperscript{th} BR move. Formally, ${{\bm \sigma}}^0$ is the initial strategy profile, and if the move at time $t$ is given by $(u_t,i_t)$, then \( {{\bm \sigma}}^{t} := (i_t,{{\bm \sigma}}^{t-1}_{-u_t}). \) The change in the potential function at this step is then given by ${\mathrm{payoff}}({{\bm \sigma}}^t) - {\mathrm{payoff}}({{\bm \sigma}}^{t-1})$, which is clearly an integer linear combination of the $A_{uv}(i,j)$ payoff values. Since the combinations have integer coefficients, and the payoff values have density bounded by $\phi$, then the total improvement (a random variable) has density at most $\phi$ as well. For any fixed BR sequence $S$ of length $2nk$, we define these linear combinations as the set of vectors ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$, where $L_t\in \{-1,0,1\}^{(|E|\times k^2)},\forall t\in [2nk]$, with entries indexed by each of the payoff values. The values of its entries are chosen as follows: \begin{equation*} L_t((u,i),(v,j))=\left\{\begin{array}{llll} 1 & \text{if: ~~}u_t\in \{u,v\}&\text{and ~~}\sigma^t_u=i&\text{and ~~}\sigma^t_v=j.\\ -1 & \text{if: ~~}u_t\in \{u,v\}&\text{and ~~}\sigma^{t-1}_u=i&\text{and ~~}\sigma^{t-1}_v=j.\\ 0 & \text{otherwise.} \end{array} \right. \end{equation*} That is, every entry signifies whether the corresponding payoff value remains unchanged ($0$), or gets added ($1$) or removed ($-1$) from the potential function. The inner product $\ip{L_t,A}$ gives the change in the total payoff of the player who makes a move at time $t$, and thus, the increase in potential due to the $t$\textsuperscript{th} move. Each of the inner products $\ip{L_t,A}$ can be shown to be unlikely to take values in the range $(0,\epsilon]$ by the assumption of bounded density. To argue that ${\mathscr{L}}$ has high rank, we partition all players who make a move in the sequence into two sets: those players who never play the same strategy twice throughout the whole sequence (non-repeating players), and those who do (repeating players). We will denote these quantities as $p_1$ and $p_2$ respectively, and define $p=p_1+p_2$. Furthermore, since we will sometimes have to refer to players, and other times to moves, we denote as $d$ the number of distinct (player,strategy) pairs which appear in the sequence, and let $q_0$ denote the number of players which play a ``return move,'' that is, moves where a player returns to their original strategy. Note that for any sequence of moves $S$, we have \[ p(S)\leq d(S)\leq k\cdot p(S),\qquad q_0(S)\leq p_2(S), \qquad q_0(S)\leq d(S)/2. \] We also introduce the quantity $d_1$, which is the number of (distinct) moves by all non-repeating players, so $p_1\leq d_1\leq k\cdot p_1$. \subsection{Smoothed Polynomial Complexity, Rank Bounds, and Union Bounds} \label{sec:smoothed-poly-overview} Recall the definition of the smoothed $k$-{NetCoordNash}\xspace defined above. The first main contribution of this paper is to show the following result: \begin{thm}\label{thm:poly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} on a complete game graph, and with an arbitrary initial strategy profile, then any execution of BRA, where improvements are chosen arbitrarily, will converge to a PNE in at most $(nk\phi)^{O(k)}$ steps, with probability \mbox{$1-1/poly(n,k,\phi)$}. \end{thm} \vspace*{-1ex} We also have convergence in $(nk\phi)^O({k})$ moves in expectation, as shown in theorem~\ref{thm:expectation}, discussed below. Note that the probability value in the above statement is over the possible choices of payoff values for the network-coordination game, and not over executions of the BR algorithm. This statement holds true regardless of how BRA is implemented, even adversarially. The {\em complete game graph} condition requires that any two players in this game share an edge in the game graph, and the payoff matrix for a non-existing game edge is not a fixed, all-zeros matrix, but is instead a random payoff matrix like all other edges. This completeness technicality fits the model of most known smoothed polynomial-time algorithms ($e.g.$ \cite{ST04,SST06,ERV14,A+17,BCC18}) which require {\em every} parameter to be perturbed. We will later discuss results in the case of incomplete graphs, where missing edges are assumed to be 0-payoff games and are unperturbed. Theorem~\ref{thm:poly} is shown using the ``common framework'' from Section~\ref{sec:common}. The random input is the set of (random) payoff matrices $\{A_{uv}\}_{u,v\in G}$, and the ``$\alpha$'' vectors are the columns of the $\mathcal L$ matrix as defined in Section~\ref{sec:notation} for a BR sequence of length $2nk$. If all of the $n$ players appear in the sequence $S$, then as shown in Corollary \ref{cor:all-active}, $\mathcal L$ has rank at least $\left(1-\tfrac 1n\right)\big(d(S)-q_0(S)\big) \geq n-1$. Therefore, any sequence where every player is present is ``good'' with probability $(\phi\epsilon)^{n-1}$, where a sequence is ``good'' if it contains either a non-improving move, or a move which improves the potential by at least~$\epsilon$. \paragraph{Case I. All Players Active.} Recall from Section~\ref{sec:common} that we wish to take the union bound over all sequences of length $\Omega(n)$. In fact, for this result, we will consider sequences of length $2nk$. Since there are $n$ players, there are $k^n$ possible initial configurations of the players, and $(nk)^{2nk}$ possible sequences, so with probability $1-(nk)^{O(nk)}(\phi\epsilon)^{\Omega(n)}$, all linear-length sequences are ``good.'' Setting $\epsilon= 1/\phi(nk)^{O(k)}$ suffices to have the probability of success be $1-1/poly(n,k)$. Since $-n^2\leq {\mathrm{payoff}}({{\bm \sigma}})\leq n^2$, then with probability $1-1/poly(n)$, the BRA must terminate in at most $2n^2/\epsilon = \phi\cdot (nk)^{O(k)}$ many iterations, as desired. \paragraph{Case II. Few Players Active.} However, it is not always the case that there are $n$ {\em active} players, or even $\Omega(n)$ active players in a given sequence of length $2nk$. A player is {\em active} if they appear in the sequence, and otherwise, {\em inactive}. We will show, in the following sections, rank bounds which depend on the $p_1$, $p_2$, $q_0$, and $d$ values, as defined in Section~\ref{sec:notation}. As we will discuss, these ranks will not be sufficiently large to handle the na\"ive union bound described above. The following table summarizes the bound we show in each case, and the resulting success probability, under the assumption $p(S)\leq \ell = 2(d(S)-q(S))\leq k\cdot p(S)$, where $\ell$ is sequence length, and $p_1(S)\leq d_1(S)$. \begin{center} \renewcommand{1}{1.4} \begin{tabular}{c|ccc} Case&Rank Bound&Union Bound&Probability of Success\\\hline $p_1\geq p_2$&$d(S)-q_0(S)+d_1(S)/2$&$k^{p(S)}(4n/\epsilon)^{d(S)-q_0(S)}(nk)^\ell$& $1-(nk\phi)^{O(k\cdot p(S))} \epsilon^{p(S)/4}$ \\[2pt] $p_2\geq p_1$&$p_2(S)/2$&$k^{p(S)}(nk)^\ell$&$1-(nk)^{O(k\cdot p(S))}(\ell \phi\epsilon)^{p(S)/4}$ \end{tabular} \end{center} And so, setting $\epsilon = 1/(nk\phi)^{O(k)}$ suffices in both cases for good success probability. We later show why $\ell= 2(d(S)-q_0(S))$ suffices. The other inequalities follow by definition. \subsubsection{Mostly Non-Repeating Players, and a Union Bound via Bucketing} Recall, a {\em non-repeating} player is one who plays each strategy at most once throughout the better-response sequence, including the initial strategy at the beginning of the sequence. Non-repeating players are key to showing rank bounds in the following sense: \paragraph{Rank Bounds through Separated Blocks.} Let $v$ be some non-repeating player, and suppose the $\tau$-th move is $(v,\sigma)$. Let $\tau'$ be the next move of $v$. Then we must have that $((v,\sigma)(*,*))$ entries of $L_t$ can only be nonzero for $\tau\leq t\leq \tau'$, and at least one of these entries must be nonzero. Therefore, if the submatrix consisting of columns $\{L_t:\tau<t<\tau'\}$ restricted to rows of the form $((v,\sigma)(*,*))$ can be shown to have large rank, then in some sense we may isolate this submatrix and inductively show a large rank for the rest of the matrix. Furthermore, if there is some player $w$ which is inactive in the sequence, playing strategy $\sigma_w$ from the start, then the submatrix restricted to $((v,*)(w,\sigma_w))$ rows, and columns indexed by moves of player $v$, can again be isolated and shown to be upper-triangular. These observations can be extended to hold simultaneously for all intervals between two consecutive moves of the non-repeating players: Let $P_1$ be the set of non-repeating players, and let $T=\{\tau_v:v\in P_1\}$ be the set of all moves for non-repeating players. Suppose $T=\{t_1<t_2<\dotsc<t_m\}$, and say $t_0=0$, $t_{m+1}=|S|$. Let $S_i$ for $i=0,\,1,\,\dotsc,\, m$ be the subsequence of $S$ between $t_i$ and $t_{i+1}$ excluding endpoints. We call these $S_i$ the {\em separated blocks} of the sequence $S$. If the move at time $t_i$ is $(v_i,\sigma^i)$, then it can be seen that \[ rank(S)\geq |T| + \textstyle\sum_{i=0}^m rank\big|_{((v_i,\sigma^i)(*,*))} (S_i) \] where $rank\big|_{C}(S)$ is the rank of the submatrix given by $S$ restricted to entries from $C$. This is shown in Lemma~\ref{lem:rank-d-q} by applying the above observations, and sorting the blocks in increasing order. \paragraph{Extension via Critical Subsequences.} As argued in Lemma~\ref{lem:rank-d-q}, it is not hard to show that for any sequence $S$, $rank(S)\geq d(S)-q_0(S)$. In fact, we show that for a separable block $S_i$, $rank\big|_{((v_i,\sigma^i)(*,*))}(S_i)\geq d(S_i)-q_0(S_i)$. However, this is not enough to immediately give the desired rank bound of $d(S)-q_0(S)+d_1(S)/2$. For this, we introduce the notion of a {\em critical subsequence}, based on the notion of a critical block introduced in~\cite{A+17}. Let $\ell(S)$ denote the length of a sequence, and call a contiguous subsequence $S'\subseteq S$ {\em critical} if $\ell(S')\geq 2(d(S')-q_0(S'))$, but for every sub-subsequence $S''\subsetneq S'$, $\ell(S'')< 2(d(S'')-q_0(S''))$. We show in Claim~\ref{lem:crit} that for any $S$ with $|S|\geq 2nk$, $S$ must contain some critical subsequence $S'$ which satisfies $\ell(S')=2(d(S')-q_0(S'))$. Consider, now, the rank bounds due to separated blocks applied to a critical subsequence $S'$: since every separated block $S_i$ of $S'$ is a strict subsequence of $S'$, $d(S_i)-q_0(S_i)> \ell(S_i)/2$, but $\ell(S') = d_1(S') + \sum_{S_i\text{ separated}} \ell(S_i)$, so \[ rank(S') \ \geq\ d_1(S') + \textstyle\sum_{S_i\text{ sep.}} d(S_i)-q_0(S_i) \ > \ d_1(S')/2 + \ell(S')/2 \] Our desired rank bound follows from recalling that $\ell(S')=2d(S')-2q_0(S')$. \paragraph{Union Bound via Bucketing.} Working within a critical subsequence allows us to find high-rank subsequences. However, when the critical subsequence is too small, issues may arise when taking the union bound over all sequences: there are $k^n(nk)^\ell$ sequences of length $\ell$, and the rank of a sequence is at most its length. Thus, we get success probability $1-k^n (nk)^\ell (\phi\epsilon)^{\ell}$ in the best case. If $\ell$ is small, the $k^n$ term dominates the probability bound, and $\epsilon$ may need to be exponentially small for good results. The issue at hand is that the initial strategies of inactive players contribute too much to the union bound. For the case of $p_1\geq p_2$, we separate out their effect on the potential function, and simply keep track of the effect size. The first part of this method was introduced in~\cite{A+17}, and allows us to reduce the $k^n$ term to a $k^{p(S)}$ term, while paying in the exponent of $\epsilon$. The idea is as follows: let $P_0$ be the inactive players, and $P_1$ be the active players. then \[ {\mathrm{payoff}}({{\bm \sigma}}) = \sum_{uv\in E}A_{uv}(\sigma_u,\sigma_v) = \sum_{u,v\in P_1}\!\! A_{uv}(\sigma_u,\sigma_v) + \sum_{u\in P_1} \sum_{w\in P_0}A_{uw}(\sigma_u,\sigma_w) + \sum_{w,w'\in P_0}\!\!A_{ww'}(\sigma_w,\sigma_{w'}) \] The left terms depend only on the strategies of the active players, the right term is constant, and the middle terms can be separated into $|P_1|$ constant terms, one per active player, per strategy played. These constant terms may be rounded to the nearest multiple of $\epsilon$, and lie in the range $[-n,n]$. Therefore, we need only keep track of $k^{p(S)}(2n/\epsilon)^{d(S)}(nk)^\ell$ values. Unfortunately, this is shy of our goal, as we can only show rank $d(S)-q_0(S)+\Omega(p_1)$, rather than $d(S)+\Omega(p_1)$. Shifting the sum to cancel the initial payoff of all players, however, allows us to reduce the union bound: \[ {\mathrm{payoff}}({{\bm \sigma}}^t)-{\mathrm{payoff}}({{\bm \sigma}}^0) = \sum_{u,v\in P_1}\!\! \widetilde A_{uv}(\sigma^t_u,\sigma^t_v) + \sum_{u\in P_1 \sum_{w\in P_0}\widetilde A_{uw}(\sigma^t_u,\sigma^t_w) + \sum_{w,w'\in P_0}\!\!\widetilde A_{ww'}(\sigma^t_w,\sigma^t_{w'} \] where $\widetilde A_{uv}(\sigma,\sigma') = A_{uv}(\sigma,\sigma')-A_{uv}(\sigma^0_u,\sigma^0_v)$. This has the effect of cancelling out $q_0(S)$ distinct middle terms, getting the desired union bound. \subsubsection{Mostly Repeating Players, and Cyclical Sums}\label{sec:cyclic} Recall the table of bounds from Section~\ref{sec:smoothed-poly-overview}. The previous section showed the bounds and analyses for the first row of the table, in the case $p_1\geq p_2$. The previous section's analysis only works when restricted to a critical subsequence $S'$ with $p_1(S')\geq p_2(S')$. Thus, we must also restrict ourselves to a critical subsequence, where here $p_2(S')\geq p_1(S')$. This means we must again find ways to control the union bound terms due to inactive players. The fundamental concept here is the notion of a {\em cyclical sum}. These are vectors which are sums of vectors from $\mathcal L$, which all have zero entries in rows for inactive players. Suppose player $u$ moved to strategy $i$ twice, and let $\tau_0$ be the first occurrence of $(u,i)$ in the BR sequence (possibly t=0), and let $\tau_1,\tau_2,\dotsc,\tau_k$ be all subsequent appearances of $u$ in the sequence, playing any strategy. Suppose $\tau_m$ is the second occurrence of $(u,i)$ in the BR sequence. Let $w$ be some inactive player, who is always playing strategy $\sigma_w$. Then the sum $L_{\tau_1}+L_{\tau_2}+\dotsm+L_{\tau_m}$ cancels out all $((u,*)(w,\sigma_w))$ entries, since each one gets added in some $L_{\tau_i}$, and and removed in $L_{\tau_{i+1}}$, where $\tau_{m+1}=\tau_1$. These sums, therefore, do not depend on the initial configurations of inactive players. Thus, working with the cyclic sums gives a $k^{p(S)}(nk)^\ell$-sized union bound. However, the main lemma of the common framework does not directly apply, since we are bounding the rank of these cyclic sums instead of $\mathcal L$. We use the fact that $\Pr[\ip{L_t,A}\in(0,\epsilon)\text{ for }t=\tau_1,\tau_2,\dotsc]\leq \Pr[\ip{L_{\tau_1}+L_{\tau_2}+\dotsm+L_{\tau_m},\,A}\in (0,\ell\epsilon)]$ and bound this instead. \paragraph{A Rank Bound for Cyclic Sums.} It remains to show that these cyclical sums have large rank. This rank bound will be applicable in the case of non-complete graphs as well as complete. A cyclical sum must contain a non-zero entry, as otherwise the sequence can not be improving. This allows us to form an auxiliary digraph, where the nodes are the active players, and we add an edge from a repeating player $u$ to any other player $v$, if {\em some} cyclic sum for $u$ contains a non-zero entry with $v$. We show in Lemma~\ref{lem:rank-p2} that there exists a way to bi-partition the nodes of this graph such that for one of the two halves contains at least half of the repeating players, and each node in this half has an out-neighbor in the other half. These edges, and their associated matrix entries, allow us to form an upper-triangular sub-matrix of rank $p_2/2$, giving us the desired bound. \subsection{Smoothed Quasipolynomial Complexity of the BRA on General Graphs} We have shown above that for complete game graphs, the BRA terminates in polynomial time with high probability, and in expectation. We discuss here how the cyclic-sum interpretation of Section~\ref{sec:cyclic} immediately gives quasi-polynomial smoothed complexity for arbitrary game graphs. Lemma~3.4 in~\cite{ER14} proves that any sequence of $\Omega(n)$ {\em improving} moves must contain a subsequence $S'$ with at least $\Omega(\ell(S')/\log n)$ distinct repeated moves. For $k$-{NetCoordNash}\xspace, this implies that a sequence of $\Omega(nk)$ player-strategy pairs in a BRA sequence must contain a subsequence $S'$ with at least $\Omega(\ell(S')/\log(nk))$ distinct recurring pairs. Therefore, $p_2\geq \Omega(\ell/k\log(nk))$, since each player can only appear in $k$ distinct pairs. This fact, along with the discussion in Section~\ref{sec:cyclic}, allow us to show the following: \begin{thm}\label{thm:qpoly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of a BR algorithm where improvements are chosen arbitrarily will converge to a PNE in at most $\phi\cdot (nk)^{O(k\log(nk))}$ steps, with probability $1-1/poly(n,k)$. \end{thm} \vspace*{-1ex} Our results also include a meta-theorem stating that, for any problem in PLS with bounded total improvement, a $w.h.p.$ smoothed complexity bound implies an expected-time bound (see Section~\ref{sec:exp}, Theorem~\ref{thm:expected-time}). This, together with Theorem \ref{thm:poly}, imply: \begin{thm}\label{thm:expectation} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of BRA where improvements are chosen arbitrarily will converge to a PNE in $O(\phi)\cdot (nk)^{O(k\log(nk))}$ steps in expectation. Furthermore, if the game graph is complete, it converges in at most $(nk\phi)^{O(k)}$ steps in expectation. \end{thm} \subsection{Smoothness-Preserving Reduction to 1- and 2-{FlipMaxCut}\xspace}\label{sec:overview-reduction} Our second set of results analyzes {\em smoothness-preserving} reductions for the $k$-{NetCoordNash}\xspace problem, to give alternate algorithms, and to introduce a reduction framework for smoothed problems. The authors of this paper do not know of any prior notion of smoothness-preserving reduction. Recall the {\em smoothness-preserving} reductions defined in Section \ref{sec:PrelSmooth}. In this section, we discuss how these reduction may be applied (Theorem \ref{thm:meta-red}), and then give an overview of how to obtain such reductions for the $k$-{NetCoordNash}\xspace problem and prove Theorems \ref{thm:kstrat} and \ref{thm:2strat} below. All the details and proofs are in Section \ref{sec:red}. First, if problem ${{\mathscr P}}$ admits a {\em strong} smoothness-preserving reduction to problem ${{\mathscr Q}}$, and ${{\mathscr Q}}$ has smoothed polynomial algorithm {\em w.h.p.}, then so does ${{\mathscr P}}$. Since such reductions allow use of extra randomness, we need to ensure that it does not affect the {\em high-probability} statement by much. We do this using Markov's inequality and careful interpretation of the extra randomness. Second, if problem ${{\mathscr P}}$ admits a {\em weak} smoothness-preserving reduction to {\em local-max-cut} on (in)-complete graphs, then ${{\mathscr P}}$ has smoothed (quasi)-polynomial complexity. This crucially requires a rank-based analysis like \cite{ER14,A+17,BCC18} for local-max-cut. These are formalized in the following result: \begin{thm}\label{thm:meta-red} Suppose problem ${{\mathscr Q}}$ has (quasi-)polynomial smoothed complexity. Then, if problem ${{\mathscr P}}$ admits a strong smoothness-preserving reduction to ${{\mathscr Q}}$, then ${{\mathscr P}}$ also has (quasi-)polynomial smoothed complexity. If instead, ${{\mathscr P}}$ admits a weak smoothness-preserving reduction to local-max-cut on an (arbitrary) complete graph, then ${{\mathscr P}}$ again has a (quasi)polynomial smoothed complexity. \end{thm} \vspace*{-1ex} This allows us to extend smoothed efficient algorithms for one problem to others. Ideally we would like to show {\em strong} reductions. However, we manage to show {\em weak} reductions from $k$-{NetCoordNash}\xspace to $1$- or $2$-{FlipMaxCut}\xspace. Note that the smoothed complexity of $2$-{FlipMaxCut}\xspace is not known yet, but we believe that the $2$-FLIP algorithm may admit a similar rank-based analysis as FLIP. This would imply a smoothed efficient algorithm for $k$-{NetCoordNash}\xspace for non-constant $k$. \paragraph{$k$-{NetCoordNash}\xspace to $2$-{FlipMaxCut}\xspace.} We wish to show the following result: \begin{thm}\label{thm:kstrat} $k$-{NetCoordNash}\xspace admits a {\em weak} smoothness-preserving reduction to 2-{FlipMaxCut}\xspace. \end{thm} \vspace*{-1ex} The idea for the reduction is, given an instance of a Network Coordination Game, construct a graph whose locally optimal cuts can be mapped to strategy profiles, and cut values can be interpreted as the total payoff of the network game. To do this, we construct a graph with $nk+2$ nodes, including two terminals $s$ and $t$, and $nk$ nodes indexed by player-strategy pairs $(u,i)$. All nodes are connected to $s$ and $t$, any two $(u,*)$ nodes are connected, and any two $(u,*),\,(v,*)$ nodes are connected if $u$ and $v$ share an edge in the game graph. Thus, the cut graph is complete if and only if the game graph is. Call a cut $S,T$ valid if $s\in S$, $t\in T$, and $S$ contains at most one $(u,*)$ for each $u$. Any such cut naturally maps to a strategy profile as follows: If player $u$ appears in $S$ paired with strategy $i$, then set $\sigma_u=i$. Otherwise, set $\sigma_u=0$, a ``dummy'' strategy with bad payoff. Call this profile ${{\bm \sigma}}(S)$. We select edge weights such that all local-max cuts are valid, by ensuring that any non-valid cut can always be improved by removing any redundant node. We also get that for any valid cut $S$, the total cut value is equal to ${\mathrm{payoff}}({{\bm \sigma}}(S))$. Furthermore, updating a player's strategy in a valid cut amounts to removing one $(u,i)$ vertex from $S$, and adding some $(u,i')$. Therefore, a unilateral deviation is equivalent to a 2-FLIP step. Since any deviation amounts to one step of 2-FLIP, it follows that ${{\bm \sigma}}(S)$ is a PNE if and only if $S$ is a local-max-cut up to 2 flips. We also show that edge-weights are linearly independent combinations of the random inputs, and that if the network coordination game instance satisfies the smoothed-inputs condition, then the edge weights do too, proving Theorem \ref{thm:kstrat}. \paragraph{\bf $2$-{NetCoordNash}\xspace to $1$-{FlipMaxCut}\xspace.} Here, we take a slightly different reduction, where there are $n+2$ nodes in the graph, and any $s$-$t$ cut is interpreted as follows: if $u$ is on the same side as $s$ of the cut, $\sigma_u=1$, otherwise, $\sigma_u=2$. The same analysis goes through, but now, locally max cuts up to one flip are Nash Equilibria, which provides a weak smoothness-preserving reduction from 2-{NetCoordNash}\xspace to 1-{FlipMaxCut}\xspace, showing the following result: \begin{thm}\label{thm:2strat} 2-{NetCoordNash}\xspace admits a {\em weak} smoothness-preserving reduction to 1-{FlipMaxCut}\xspace. \end{thm} \vspace*{-1ex} The FLIP algorithm for $1$-{FlipMaxCut}\xspace has smoothed quasi-polynomial running time in general \cite{ER14}, and smoothed polynomial running time if $G$ is a complete graph \cite{A+17}. A recent result \cite{BCC18} has improved the running time of the latter, so this reduction allows us to provide better bounds on the performance of BRA in the case $k=2$. These local-max-cut results, together with Theorems \ref{thm:meta-red} and \ref{thm:2strat}, give an alternate smoothed efficient algorithm for the $2$-{NetCoordNash}\xspace problem. \subsection{A Common Framework for Local-Improvement Algorithms}\label{sec:common} To highlight the framework common to our problem and local-max-cut, we observe that in both cases, the potential function for $k$-{NetCoordNash}\xspace, as given in~\eqref{eq:pot}, is an integer linear combination of the payoff values. In the case of our problem and local-max-cut, it is a 0--1 combination. The framework can be applied whenever this holds, as long as the potential function's range is also polynomially bounded. It hinges on the following lemma: \begin{lem}[{\color{red}\bf Give an Older Citation}]\label{lem:main-praobability} Let $X\in \mathbb R^d$ be a vector of $d$ independent random variables where each $X_i$ has density bounded by $\phi$. Let $\alpha_1,\,\dotsc,\,\alpha_r$ be $r$ linearly independent vectors in $\mathbb{Z}^d$. then the joint density of $(\ip{\alpha_i,X})_{i\in [r]}$ is bounded by $\phi^r$, and for any given $b_1,b_2,\dotsc \in \mathbb{R}$ and $\epsilon>0$, \begin{equation}\label{eq:prob-bound-intro} \Pr\Big[\textstyle\bigwedge_{i=1}^r\ip{\alpha_i,X}\in [b_i,b_i+\epsilon]\Big] \leq (\phi\epsilon)^r \end{equation} \end{lem} Here, the $X$ vector denotes the random input values to the smoothed problem, and the $\alpha$ vectors represent some combination of changes in the potential function. Formally, given a potential function $\Phi(\bm \sigma)$ for some problem, and letting $\bm \sigma_1,\bm \sigma_2,\bm \sigma_3,\dotsc$ be the walk through the state space given by some local-improvement algorithm, then we have vectors $\alpha_i$ such that $\ip{\alpha_i,X}=\Phi(\bm \sigma_i)-\Phi(\bm \sigma_{i-1})$ for all $i$. Setting $b_i=0$, the above lemma bounds the probability that {\em every} step of the local-improvement algorithm be a small improvement \mbox{($\ip{\alpha,X}<\epsilon$)}, while still being an improvement \mbox{($\ip{\alpha,X}>0$)}. Let $\mathcal E$ be the event in the probability statement, that is, $\textstyle\bigwedge_{i=1}^r\ip{\alpha_i,X}\in [0,0+\epsilon]$. If $\mathcal E$ does not hold, and the sequence is indeed an improving one, then at least one of the improvements must be at least $\epsilon$. If $\mathcal E$ does not hold for {\em any} sequence of $\Omega(n)$ moves, then we can bound the running time of the iterative algorithm by \[ \frac n \epsilon\cdot \left(\max_{\sigma} \Phi(\bm \sigma) - \min_{\sigma}\Phi(\bm \sigma)\right) \] Finally, if $\Phi$ is bounded in a (quasi)polynomially-sized range, and $\epsilon$ is taken to be $1/(quasi)poly(n)$, then we conclude that the procedure runs in (quasi)polynomial time. \paragraph{Rank Bound vs Union Bound.} Note that in order to get (quasi)polynomial running time with high probability, first we need to upper bound the probability of event $\mathcal E$. This in turn requires proving a ``high'' rank $r$ on matrix $[\alpha_i]_{i=1}^r$. Furthermore, for $\mathcal E$ to not occur on {\em any} sequence of $\Omega(n)$ local-improvement moves, we must take the union bound over all sequences of length $\Theta(n)$. It remains then to choose $\epsilon$ correctly to counteract the union bound, for the best rank $r$ which can be shown in general. This highlights the main technical challenge when applying the common framework: labelling the moves. If a move's label is too descriptive --- {\em e.g.} if it contains a state vector --- then the union bound will be much too large. However, if a move's label is too vague --- {\em e.g.} defining a move in $k$-{NetCoordNash}\xspace by the player who is moving, but not the strategy they are moving to --- then the rank bound will not be large enough. We introduce, in the next section, those parameters which work in our setting. {\color{red}\bf decide if we need the following} At a high-level our analysis builds on \cite{}. However, handling the game framework where players have multiple strategies throws significant challenges. To keep the exposition relatable to previous approaches we try to use similar notations and terminology as far as possible. \subsection{Notation} Recall the problem of $k$-{NetCoordNash}\xspace defined above, and the better-response algorithm (BRA) discussed in Section \ref{}. We represent each better-response (BR) move by a player-strategy pair $(u,i)$, denoting that player $u$ is replacing strategy $\sigma_u$ by $i$, (assuming $i\neq \sigma_u$). We also denote as ${{\bm \sigma}}^t$ the strategy profile after the $t$\textsuperscript{th} BR move. Formally, ${{\bm \sigma}}^0$ is the initial strategy profile, and if the move at time $t$ is given by $(u_t,i_t)$, then \( {{\bm \sigma}}^{t} := (i_t,{{\bm \sigma}}^{t-1}_{-u_t}). \) The change in the potential function at this step is then given by ${\mathrm{payoff}}({{\bm \sigma}}^t) - {\mathrm{payoff}}({{\bm \sigma}}^{t-1})$, which is clearly an integer linear combination of the $A_{uv}(i,j)$ payoff values. Since the combinations have integer coefficients, and the payoff values have density bounded by $\phi$, then the total improvement (a random variable) has density at most $\phi$ as well. For any fixed BR sequence $S$ of length $2nk$, we define these linear combinations as the set of vectors ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$, where $L_t\in \{-1,0,1\}^{(|E|\times k^2)},\forall t\in [2nk]$, with entries indexed by each of the payoff values. The values of its entries are chosen as follows: \begin{equation*} L_t((u,i),(v,j))=\left\{\begin{array}{llll} 1 & \text{if: ~~}u_t\in \{u,v\}&\text{and ~~}\sigma^t_u=i&\text{and ~~}\sigma^t_v=j.\\ -1 & \text{if: ~~}u_t\in \{u,v\}&\text{and ~~}\sigma^{t-1}_u=i&\text{and ~~}\sigma^{t-1}_v=j.\\ 0 & \text{otherwise.} \end{array} \right. \end{equation*} That is, every entry signifies whether the corresponding payoff value gets added, removed, or remains unchanged, from the potential function. The inner product $\ip{L_t,A}$ gives the change in the total payoff of the player who makes a move at time $t$, and thus, the increase in potential due to the $t$\textsuperscript{th} move. Each of the inner products $\ip{L_t,A}$ can be shown to be unlikely to take values in the range $(0,\epsilon]$ by the assumption of bounded density. To argue that ${\mathscr{L}}$ has high rank, we partition all players who make a move in the sequence into two sets: those players who never play the same strategy twice throughout the whole sequence (non-repeating players), and those players who do (repeating players). We will denote these quantities as $p_1$ and $p_2$ respectively, and define $p=p_1+p_2$. Furthermore, since we will sometimes have to refer to players, and other times to moves, we denote as $d$ the number of distinct (player,strategy) pairs which appear in the sequence, and let $q$ denote the number of those moves which are ``return moves,'' that is, moves where a player returns to their original strategy. Note that for any sequence of moves $S$, we have $p(S)\leq d(S)\leq k\cdot p(S)$. {\color{red}\bf Is $q$ the same things as $p_2$??} \rnote{No! But it seems we should be able to merge these two, no? Like if we define notation to count repeat (player,strategy) pairs starting with $\sigma^0$?} \subsection{Critical Blocks and Small Union Bounds}\label{sec:union-critical} The notion of {\em critical block}, introduced in \cite{A+17}, allows us to translate rank and union bounds in terms of $p$ and $d$ values, into bounds which are linear in the sequence size. Formally, let $\ell(S)$ be the length of the sequence, and call a sequence $S$ ``diverse'' if $\ell(S)< 2(d(S)-q(S))$. A sequence is diverse if, for its size, it has many different moves. A {\em critical block} is a contiguous sub-sequence which is, in some sense, maximally diverse. That is, a block $B\subseteq S$ such that $\ell(B)\geq 2(d(S)-q(S))$, and for every $B'\subseteq B$, $B's$ is diverse. To see that such a block must exist, note that any sequence of length 1 must be diverse, and any sequence of length $2nk$ can not be diverse. Therefore, if we begin with a sequence of length $2nk$, it must contain a ``maximally diverse'' subsequence. It then becomes clear that, restricting our attention to critical blocks will allow us to show $\Theta(\ell)$ bounds by instead showing $\Theta(p)$ bounds, if we take $k$ to be a constant. However, these critical blocks might be significantly smaller than the original $2nk$-sized sequence, which introduces some difficulties for the union bound. Now, recall that we must consider {\em any} sequence of $\ell$ player-strategy pairs. This na\"ively have $k^n$ possible initial strategy configurations, and $(nk)^\ell$ choices for the $\ell$ moves, and so we have a union bound of size $k^n (nk)^\ell$. If we can show a rank lower-bound of $r$, then we get success probability \[ 1-k^n (nk)^\ell (\phi\epsilon)^{r} \] Note that we must have $r\leq \ell$, since $\ell$ is the number of columns whose rank we are bounding. If we had $r\geq \Omega(n)$, then to get polynomially decaying probability of failure, we can set $1/\epsilon = \phi(nk)^{O(k)}$, which gives an overall run-time of $(nk\phi)^{O(k)}$. This runtime does indeed match the results that we prove for the case of complete game graphs. However, when $\ell$, and subsequently $r$, are much smaller than $n$, then the $k^n$ term dominates the probability bound, and $\epsilon$ is needed to be significantly smaller. Observe that the $k^n$ term is the effect of the initial configuration of those players which do not move in the sequence, but do have an effect on the potential function. There are two possible options for getting around this issue. The one, introduced in~\cite{A+17}, isolates the effect of these players, and ``buckets'' their influence. Thus, we need not keep track of each player's state, but instead, of the value of their contributions. The other, introduced in~\cite{ER14}, and extended in~\cite{BCC18}, uses the notion of a ``cyclic sum,'' which altogether eliminates the effect of non-moving (or {\em inactive}) players' strategy from the linear combinations. It does, however, change the vectors being considered, and makes the analysis more subtle. \paragraph{Union Bounds via Bucketing.} Note, the potential, $\sum_{uv\in E}A_{uv}(\sigma_u,\sigma_v)$ can be separated into terms between two players present in the sequence ({\em active}), which we can na\"ively union-bound, terms between two inactive players, which remain constant, and terms between an active and an inactive player. This latter value can be split up into different sums for each active player, of their payoffs with inactive players. These values can then be rounded to the nearest multiple of $\epsilon$, and instead of recording the states of the non-moving players, we just record which of the $2n\epsilon$ possible values it has been rounded to, and take the additive error into account when considering the probability bound. This leaves a union bound of the form $k^p (n/\epsilon)^d$, one term for each of the distinct moves of the active players. Some shifting of the weight of the initial strategies of the active players into the constant term allows us to instead get a bound of the form $k^p (n/\epsilon)^{d-q} (nk)^\ell$. These exponents are more promising than the $k^n$ term of the na\"ive bound. However, note that the negative exponent on the $\epsilon$ term requires the rank bound to at least match it. We will later see how a rank bound of the form $r\geq d-q+\Omega(p_1)$ can be shown, which allows the probability bound to look more like $k^p (nk)^{kp} (\phi\epsilon)^p$, as desired. \paragraph{Union Bounds via Cyclic Sums} This union bound is not for the number of possible sequences of the $L$ vectors as above, but instead, is simply the na\"ive bound on a difference set of vectors, which allow us to show similar results. These vectors are sums of a subset of the $L$ vectors, and are only defined when {\em repeat moves} are made. That is, either repeat occurrences of the same (player,strategy) pair, or a return move. These sums allow us to cancel out the effect of inactive players. Let $L_t$ be the $t$-th column of $L$, and suppose player $u$ moved to strategy $i$ twice. Let $\tau_0$ be the first occurrence of $(u,i)$ in the BR sequence, and let $\tau_1,\tau_2,\dotsc,\tau_k$ be all subsequent appearances of $u$ in the sequence, playing any strategy. Suppose $\tau_m$ is the second occurrence of $(u,i)$ in the BR sequence. Let player $v$ be some player who starts in strategy $j$, and never deviates. Then the sum $L_{\tau_1}+L_{\tau_2}+\dotsm+L_{\tau_m}$ cancels out all $(uv,\cdot,\cdot)$ entries, since each one gets added in one column, and removed from the next (cyclically). These cyclical sums, therefore, do not depend on the initial configurations of the players who do not appear in the sequence. Thus, taking these summed vectors, we need not consider the initial configurations of the inactive players, and can take a union bound of the form $O(k^\ell(nk)^\ell)$. However, it does affect the probability computation a bit: to bound the probability that $\ip{L_t,A}$ lie in $(0,\epsilon)$ for all $t$, it suffices to bound the probability that $\ip{L_{\tau_1}+L_{\tau_2}+\dotsm+L_{\tau_m},\,A}$ lie in $(0,n\epsilon)$. It remains to show how the above rank bounds can be attained, and conclude that the BRA terminates in polynomial time with high probability, and in expectation. \subsection{Smoothed Polynomial Complexity of the BRA} \label{sec:complete-graph-rank} We wish to show, for the two above union bounds to be useful, a bound of the form $r\geq d(S)-q(S)+\Omega(p_1(S))$ on the usual $L$ vectors, and a bound of the form $r\geq \Omega(p_2(S))$ for the cyclic-sum vectors. The ``common framework'' of section~\ref{sec:common} allows us then to show. \begin{thm}\label{thm:poly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} on a complete game graph, and with an arbitrary initial strategy profile, then any execution of BRA where improvements are chosen arbitrarily will converge to a PNE in at most $(nk\phi)^{O(k)}$ steps, with probability \mbox{$1-1/poly(n,k,\phi)$}. \end{thm} \vspace*{-1ex} We note that the probability value in the above statement is over the possible choices of payoff values for the Network game, and not over executions of the BR algorithm. This statement holds true regardless of how BRA is implemented, even adversarially. By complete game graph, we simply mean that all payoff values are random variables, and any two players must participate in some ``edge game.'' We can not simply take missing edges to be 0-payoff games, as that would not be perturbed. This fits the model of most known smoothed polynomial-time algorithms ($e.g.$ \cite{ST04,SST06,ERV14,A+17,BCC18}) which do require {\em every} parameter to be perturbed. We will later discuss results in the incomplete-graph cases, where missing edges are assumed to be 0-payoff games, unperturbed. We also present a meta-theorem which abstracts a pre-existing result, and state that for any problem in PLS with bounded total improvement, a $w.h.p.$ smoothed complexity bound implies an expected time bound (see Section \ref{sec:exp}). This, together with theorem \ref{thm:poly}, implies \begin{thm}\label{thm:poly-expectation} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of BRA where improvements are chosen arbitrarily will converge to a PNE in at most $(nk\phi)^{O(k)}$ steps in expectation, for a complete game graph. \end{thm} Thus it remains to show the two rank bounds described above. Our rank bounds will be shown by identifying sufficiently large upper-triangular submatrices of $\mathcal L$. As discussed in section~\ref{sec:common}, the bounds shown are sensitive to the column labels, since we may only use these when talking about worst-case $\mathcal L$ matrices. In fact, we will see that the $d-q+\Omega(p_1)$ bound can not be raised to a $d+\Omega(p_1)$ bound, which is why it was essential to get the ``bucketing'' union bound to need only $d-q$ terms. \paragraph{Rank Bounds on $\mathcal L$.} If there are no inactive players, then as proved in Lemma \ref{lem:all-active}, it is easy to argue a minimum rank of $n-1$ for vectors of ${\mathscr{L}}$ and pair it with the na\"ive union bound. Thus, we restrict our attention to the case where there are many inactive players. To bound the rank, we first divide the sequence into blocks, where the endpoints of each block are the {\em last} occurrence of a non-repeating player in the BR sequence. (The blocks do not include these players' final moves). These blocks are termed \textit{transition blocks}, and the vectors inside and outside the transition block can be ``separated,'' by whether they contain a non-zero values on the associated payoff entries. Let $v$ be the non-repeating player who appears for the last time, and let $j$ be her final strategy. Let $u$ be some other player, then vectors inside the transition blocks will have 0 coefficient for the $(uv,*,j)$ term, while the vectors outside will have non-zero. Let $T$ be some transition block, then it can be shown that the rank of the moves of $T$ is at least $d(T)-q(T)$. Recall that we are restricting ourselves to some critical block $B$: this implies that the rank of the transition blocks is at least half their length, since $T\subsetneq B$, and there $\ell(T)< 2(d(T)-q(T))$, by criticality. We can also show that the final moves of the non-repeating players are themselves independent of all other columns, which adds a factor $p_1$ to the rank. This, and the transition-block observation, gives us our desired bounds. \paragraph{Rank Bounds on cyclic sums.} {\em\color{red} This kinda depends on how the proof goes..... I'll write what I guess to be the correct form, but it needs double-checking for sure.} This rank bound will be applicable in the case of non-complete graphs as well as complete. Recall, from section~\ref{sec:union-crit}, the notion of the ``cyclic sum'' vector, the sum of all moves of some repeating player between two instances of the same move, excluding the first, including the last. Clearly, this sum must form a non-zero vector, as otherwise, the sequence is returning to the same state as the original time the move was played, which makes it non-improving. This allows us to form an auxiliary digraph, where the nodes are the active players. Create an edge from a repeating player $u$ to any other player $v$, if {\em some} cyclic sum for $u$ contains a non-zero entry with $v$. We show that there exists a way to bi-partition the nodes of this graph such that for one of the two partitions, 1. it contains at least half of the repeating players, and 2. each node in this partition has an out-neighbour outside of the partition. These edges, and their associated matrix entries, allow us to form an upper-triangular submatrix of rank $\Omega(p_2)$, on the set of cyclic sums. \subsection{Smoothed Quasipolynomial Complexity of the BRA on General Graphs} We have shown, above, that for complete game graphs, the BRA terminates in polynomial time with high probability, and as discussed, in expectation. We discuss here how the cyclic-sum interpretation immediately gives quasi-polynomial smoothed complexity for arbitrary game graphs (Theorem~\ref{thm:qpoly}), based on a {\color{red}\bf a lemma of~\cite{ER14}}. This lemma states that any sequence of $\Omega(n)$ {\em improving} moves must have at least $\Omega(n/\log n)$ distinct repeated moves. Therefore, a sequence of $\Omega(nk)$ player-strategy pairs in a BRA sequence must contain at least $\Omega(nk/\log(nk))$ distinct recurring pairs, and therefore, $p_2\geq \Omega(n/\log(nk))$, since each player can only appear in $k$ distinct pairs. This fact, combined with the discussion from section~\ref{sec:complete-graph-rank}, allows us to show the following: \begin{thm}\label{thm:qpoly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of a BR algorithm where improvements are chosen arbitrarily will converge to a PNE in at most $\phi\cdot (nk)^{O(k\log(nk))}$ steps, with probability $1-1/poly(n,k)$. \end{thm} \vspace*{-1ex} Along with the same meta-theorem that was used to show polynomial expected-time convergence of the BRA for complete graphs, we also show \begin{thm}\label{thm:poly-expectation} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of BRA where improvements are chosen arbitrarily will converge to a PNE in at most $O(\phi)\cdot (nk)^{O(k\log(nk))}$ steps in expectation. \end{thm} \subsection{Smoothness-Preserving Reduction to 1- and 2-{FlipMaxCut}\xspace}\label{sec:overview-reduction} We have shown, above, that the BRA converges quickly for $k$-{NetCoordNash}\xspace. Our second set of results analyzes {\em smoothness-preserving} reductions for the $k$-{NetCoordNash}\xspace problem, to give alternate algorithms, and to introduce a reduction framework for smoothed problems. The authors of this paper have not found any prior notion of smoothness-preserving reduction. Recall the {\em smoothness-preserving} reductions defined in Section \ref{sec:PrelSmooth}. In this section, we discuss a result explaining how these reduciton notions may be applied (Theorem \ref{thm:meta-red}), and then give an overview of how to obtain such reductions for the $k$-{NetCoordNash}\xspace problem and prove Theorems \ref{thm:2strat} and \ref{thm:kstrat} below. All the details and proofs are in Section \ref{sec:red}. First, if problem ${{\mathscr P}}$ admits a {\em strong} smoothness-preserving reduction to problem ${{\mathscr Q}}$, and ${{\mathscr Q}}$ has smoothed polynomial algorithm in terms of {\em whp}, then so does ${{\mathscr P}}$. Since such reductions allow use of extra randomness, we need to ensure that it does not affect the {\em high-probability} statement by much. We do this using the Markov's inequality and careful interpretation of the extra randomness. Second, if problem ${{\mathscr P}}$ admits a {\em weak} smoothness-preserving reduction to the {\em local-max-cut} with (arbitrary) complete graphs, then ${{\mathscr P}}$ has smoothed (quasi-polynomial) polynomial complexity. Here, the analysis builds on the first result, but also crucially depends on the rank-based analysis of \cite{ER14} and \cite{A+17} for the local-max-cut. These are formalized in the following result: \begin{thm}\label{thm:meta-red} Suppose problem ${{\mathscr Q}}$ has (quasi-)polynomial smoothed complexity. Then, if problem ${{\mathscr P}}$ admits a strong smoothness-preserving reduction to ${{\mathscr Q}}$, then ${{\mathscr P}}$ also has (quasi-)polynomial smoothed complexity. If instead, ${{\mathscr P}}$ admits a weak smoothness-preserving reduction to local-max-cut on an (arbitrary) complete graph, then ${{\mathscr P}}$ again has a (quasi)polynomial smoothed complexity. \end{thm} \vspace*{-1ex} \medskip This allows us to extend smoothed efficient algorithm from one problem to another. Ideally we would like to show {\em strong} reductions. However, we manage to show {\em weak} reductions from $k$-{NetCoordNash}\xspace to $1$- or $2$-{FlipMaxCut}\xspace. Note that, smoothed complexity of $2$-{FlipMaxCut}\xspace is not known yet, but we believe that $2$-FLIP algorithm may admit a similar rank-based analysis as FLIP. And that will give a smoothed efficient algorithm for $k$-{NetCoordNash}\xspace with arbitrary $k$. \paragraph{$k$-{NetCoordNash}\xspace to $2$-{FlipMaxCut}\xspace.} We wish to show the following result: \begin{thm}\label{thm:kstrat} $k$-{NetCoordNash}\xspace admits a {\em weak} smoothness-preserving reduction to 2-{FlipMaxCut}\xspace. \end{thm} \vspace*{-1ex} Though this result does not immediately provide an algorithm to solve $k$-{NetCoordNash}\xspace, we show that it does raise the following implication: If we can show that the 2-FLIP algorithm solves smoothed instances of \mbox{2-{FlipMaxCut}\xspace} in (quasi-)polynomial time through rank-based analysis similar to FLIP, then {NetCoordNash}\xspace has (quasi-)polynomial smoothed complexity for $k$ in the input, rather than fixed $k$. We leave the smoothed analysis of 2-{FlipMaxCut}\xspace as an open problem. The idea for the reduction is, given an instance of a Network Coordination Game, construct a graph whose locally optimal cuts can be mapped to strategy profiles, and the total cut value of said cuts is equal to the total payoff of the associated strategy profile. To do this, we construct a graph with $nk+2$ nodes, two terminals $s$ and $t$, and $nk$ nodes indexed by player-strategy pairs $(u,i)$. Nodes $s$ and $t$ are connected to every other node in the graph, and for each player $u$ and strategy $i$, there is an edge from node $(u,i)$ to node $(u,j)$ for all $j\neq i$. Furthermore, if player $u$ and $v$ share an edge in the game graph ($i.e.$ play together), there is also an edge from $(u,i)$ to $(v,j)$ for all $1\leq i,j\leq k$. Therefore, the cut graph is complete if and only if the game graph is. Call a cut $S,T$ valid if it is an $s$-$t$ cut with $s\in S$ and $t\in T$, and if $S$ contains at most one node $(u,i)$ for each player $u$. Now, for any valid $s$-$t$ cut, we can interpret this cut as determining a strategy profile as follows: If player $u$ appears in $S$ the cut paired with strategy $i$, then set $\sigma_u=i$. Otherwise, set $\sigma_u=0$, a ``dummy'' strategy with bad payoff. Call this profile ${{\bm \sigma}}(S)$. We wish to choose edge weights such that all locally maximal cuts are valid cuts, and also such that for any valid cut $S$, the total cut value is equal to ${\mathrm{payoff}}({{\bm \sigma}}(S))$. To show that this reduction is indeed a smoothness-preserving reduction, it suffices to show that ${{\bm \sigma}}(S)$ is a PNE if and only if $S$ is a local-max-cut up to 2 flips, and that for random payoffs with bounded density, the edge weights in the reduction also have bounded density, and are full-rank combinations of independent random variables. This last condition is guaranteed by choosing the edge weights as a full-rank, integer, linear combination of the payoff values. It remains to show that this is possible while ensuring the first condition. Indeed, we ensure that any non-valid cut can always be improved by flipping at least one node. Furthermore, we ensure that any valid cut has total cut value equal to the total payoff of the associated strategy profile. Since replacing $\sigma_u$ with $\sigma'_u$ is equivalent to dropping $(u,\sigma_u)$ from the cut and adding $(u,\sigma'_u)$, then any locally optimal cut up to two flips must be equivalent to a Nash Equilibrium. Thus, we achieve a weak reduction, proving Theorem \ref{thm:kstrat}. \paragraph{\bf $2$-{NetCoordNash}\xspace to $1$-{FlipMaxCut}\xspace.} Here, we take a slightly different reduction, where there are $n+2$ nodes in the graph, and any $s$-$t$ cut is interpreted as follows: if $u$ is on the same side as $s$ of the cut, $\sigma_u=1$, otherwise, $\sigma_u=2$. The same analysis goes through, but now, locally max cuts up to one flip are Nash Equilibria, which provides a weak smoothness-preserving reduction from 2-{NetCoordNash}\xspace to 1-{FlipMaxCut}\xspace, showing the following result: \begin{thm}\label{thm:2strat} 2-{NetCoordNash}\xspace admits a {\em weak} smoothness-preserving reduction to 1-{FlipMaxCut}\xspace. \end{thm} \vspace*{-1ex} The FLIP algorithm for $1$-{FlipMaxCut}\xspace has smoothed quasi-polynomial running time in general \cite{ER14}, and smoothed polynomial running time if $G$ is a complete graph \cite{A+17}. A recent result \cite{BCC18} seems to have improved the running time of the latter, and so this reduction allows us to provide better bounds on the performance of BRA in the case $k=2$. These local-max-cut results, together with Theorems \ref{thm:meta-red} and \ref{thm:2strat}, give an alternate smoothed efficient algorithm for the $2$-{NetCoordNash}\xspace problem. \subsection{Smoothed Polynomial Complexity for Complete Game Graphs} Our analysis begins with the case where the game graph is complete. The completeness of the game graph will allow us to single out a vertex, and use edges to that vertex to show large rank. We wish to show that the BRA algorithm will terminate in $poly(n^k,k,\phi)$ steps with high probability. We begin by handling a simple case: \medskip \noindent\textbf{Informal Lemma.} \textit{Suppose we could guarantee that every $2nk$-length sequence had all players active, then any execution of BRA must terminate in $(nk\phi)^{O(k)}$ steps with probability $1-1/(nk)^{O(1)}$.} \begin{proof} From Corollary~\ref{cor:all-active}, which we will show below, if $S$ has all $n$ players active, then the transformation set ${\mathscr{L}}$ of $S$ has rank at least $\big(1-\tfrac 1n\big)\big(d(S)-q_0(S)\big)\geq n-1$. Thus, any such sequence $S$ is either $\epsilon$-improving (has improvement at least $\epsilon$) or non-improving (non-positive improvement) with probability $1-(\phi \epsilon)^{n-1}$. However, we want this to hold for every sequence. For any fixed $\ell$, there are $(nk)^\ell$ BR sequences of length $\ell$, and $k^n$ possible initial configurations. Thus, the probability that {\em every} sequence of length $2nk$ is neither non-improving nor $\epsilon$-improving, for any initial configuration, is at most $k^n(nk)^{2nk}(\phi\epsilon)^{n-1}$. For this value to be vanishingly small, {\em i.e.} $1/(nk)^{O(1)}$, it suffices to set $\epsilon= 1/\phi(nk)^{O(k)}$. As discussed above, this implies that with probability $1-1/poly(n,k)$, the BRA will terminate in time $2n^3k/\epsilon = (nk\phi)^{O(k)}$. \end{proof} The above lemma, however, relies on a condition which cannot be guaranteed: not every sufficiently long sequence has $n$ active players. In what follows, we lower bound the rank of the {transformation set}\xspace for sufficiently long BR sequences, and pair them with nontrivial union bounds, to get our desired results. \subsubsection{Case I: Mostly Non-Repeating Players} We will be showing rank bounds which depend on the $p_1$ and $p_2$ parameters defined above. Splitting this analysis into the cases $p_1\geq p_2$ and $p_2\geq p_1$ allows us to combine these and get bounds in terms of $p$. We consider first the case $p_1\geq p_2$, and the following definition: \begin{defn}[Separated Blocks]\label{def:separated} Let $P_1(S)$ be the set of non-repeating players in a BR sequence, and for any $u\in P_1$, let $T_u$ be the set of indices where the moving player is $u$. Let $T=\bigcup_{u\in P_1}T_u$, the set of indices of all non-repeating-player moves, and suppose $T=\{t_1<t_2<\dotsm<t_m\}$ We will show below how the $t_i$'s ``separate'' the sequence $S$. To this end, let $S_i$ for $i=0,\,1,\,\dotsc,\,m$ be the subsequences of $S$ from time $t_i$ to $t_{i+1}$ excluding boundaries, respectively, where $t_0=0$ and $t_{m+1}=|S|$. Then these $S_i$'s are the {\em separated blocks} of $S$. Denote their collection as $\mathbb S = \{S_0,\,S_1,\,\dotsc,\,S_m\}$. Furthermore, denote $|T|$ as $d_1(S)$. \end{defn} The following lemma allows us to take advantage of this notion of separated block, to break up the rank bounds into simpler subproblems. \begin{lem}\label{lem:rank-d-q} Let $S$ be a BR sequence with at least one inactive player, and let ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$ be its {transformation set}\xspace. Then ${\mathscr{L}}$ contains at least $d_1(S)+\sum_{S'\in \mathbb S}d(S')-q_0(S')$ linearly independent vectors, where $\mathbb S$ is the collection of separated blocks of $S$. \end{lem} \begin{proof} Let $w$ be some inactive player, which we have assumed exists. Let $T=\{t_1<t_2<\dotsm<t_m\}$ be the endpoints of the separated blocks, as in Definition~\ref{def:separated} above. For $i=0,\,1,\,\dotsc,\,m$, let $D_i$ be the set of distinct (player,strategy) moves which occur in $S_i$, which are not {\em return moves} of $S_i$, {\em i.e.} $(u,{{\bm \sigma}}^{t_i}_u)$ moves. For all $i$, the move at $t_i$ must be some non-repeating player of $S$, which we denote $v_i$, and call the strategy it moves to as $\sigma^i$. (Take $v_0=w$, $\sigma^0={{\bm \sigma}}^0_w$). For all $(u,\sigma)\in D_i$, let $\tau^i_{(u,\sigma)}$ be the time of the {\em first} occurrence of $(u,\sigma)$ in the subsequence~$S_i$, and let $H_i=\{\tau^i_{(u,\sigma)}:(u,\sigma)\in D_i\}$. Let $H=\bigcup_{i=0}^{|\mathbb S|-1} H_i \cup \{t_1,\,\dotsc,\,t_{|\mathbb S|-1}\}$. For each $t\in H$, if $t=\tau^i_{(u,\sigma)}\in H_i$ for some $i,\,u,\,\sigma$, then {\em associate} to $L_t$ the row $((u,\sigma)(v_i,\sigma^i))$. If, instead, $t=t_i$ for some $i$, then {\em associate} to $L_t$ the row $((v_i,\sigma^i)(w,\sigma^0))$. Consider the submatrix of ${\mathscr{L}}$ consisting of all columns $\{L_{t}:t\in H\}$, sorted in ``chronological'' order, and all of their associated rows, in the same order as their respectively associated columns. We claim that this matrix is upper-triangular, and its diagonal entries are non-zero. For each column $L_t$, the diagonal entry in the submatrix is the entry for the associated row, which we have chosen to be nonzero. Furthermore, if $t=t_i\in H$, then $v_i$ no $((v_i,\sigma^i)(*,*))$ entry can have been non-zero, since $v_i$ is non-repeating. Thus, $L_{t_i}$ is the first column where the associated row has a nonzero entry. If, instead, $t=\tau^i_{(u,\sigma)}\in H_i$, then the associated row $((u,\sigma)(v_i,\sigma^i))$ must have been 0 up until column $L_{t_i}$ as described above. Furthermore, since $\tau^i_{(u,\sigma)}$ is the first occurrence of $(u,\sigma\neq {{\bm \sigma}}^{t_i}_u)$ after time $t_i$, we must have had the row $((u,\sigma)(v_i,\sigma^i))$ be 0 before the $\tau^i_{(u,\sigma)}$-th column. These observations imply that our $|H|\times|H|$ submatrix, with the given row-ordering, must be upper-triangular with nonzero diagonal terms. Therefore, it must be full-rank. Since $|H_i| = d(S_i)-q_0(S_i)$, then $|H|=d_1(S)+\sum_{S'\in \mathbb S} d(S')-q_0(S')$, and we conclude the desired bound. \end{proof} \begin{cor}\label{cor:all-active} Let $S$ be a BR sequence where all players are active, and let ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$ be its {transformation set}\xspace. Then ${\mathscr{L}}$ contains at least $\left(1-\tfrac 1n\right)(d(S)-q_0(S))$ linearly independent vectors. \end{cor} \begin{proof} Consider the above proof method with $|T|=0$, and $S_0=S$. Note that now, $H=D_0$. It is still correct if some arbitrary player is chosen to be the $w$ player, and all $((u,\sigma)(v_0,\sigma^0))$ terms are replaced with $((u,\sigma)(w,{{\bm \sigma}}^{\tau^0_{(u,\sigma)}}))$ terms. We must further restrict $H$ not to contain any moves of player $w$. Suppose we choose, as our $w$ player, the player which appears the least number of times in $H$, then we suffer a $\left(\tfrac 1n\right)$-fraction loss in the size of $H$, concluding the proof. \end{proof} Now, the rank bound of Lemma~\ref{lem:rank-d-q} is not in a very usable form, as it requires too much structural information about the sequence $S$ to use. To this end, we turn to the definition of {\em critical subsequence}, closely based on the definition of a critical block in~\cite{A+17}. \begin{defn}[Critical Subsequence]\label{def:critical-blocks} For every contiguous subsequence $B$ of $S$, let $\ell(B)$, $d(B)$, and $q_0(B)$ be length, number of distinct pairs, and number of return moves, in $B$, respectively. Such a subsequence is termed {\em critical} if $\ell(B)\geq 2\big(d(B)-q_0(B)\big)$, but for every $B'\subseteq B$, $\ell(B')< 2\big(d(B')-q_0(B')\big)$. \end{defn} Note that a return move for a subsequence $B$ which starts at time $t_B$ is a move $(u,{{\bm \sigma}}^{t_B}_u)$, as opposed to a $(u,{{\bm \sigma}}^0_u)$ move. We show here that critical subsequences always exist. \begin{claim}\label{lem:crit} A critical subsequence always exists in any sequence $S$ of length $2nk$. Furthermore, if $B$ is a critical subsequence, then $\ell(B) = 2(d(B)-q_0(B))$. \end{claim} \begin{proof} As there are at most $nk$ distinct player-strategy pairs possible, the entire sequence $S$ satisfies the relation $\ell(S)\geq 2d(S)\geq 2(d(S)-q_0(S))$. Conversely, for every subsequence $B$ of length 1 ($i.e.$ a single move), $d(B)=1,q_0(B)=0 \Rightarrow 1=\ell(B)<2(d(B)-q_0(B))=2$. Thus, it suffices to take an inclusion-minimal subsequence which satisfies $\ell(B)\geq 2(d(B)-q_0(B))$ and obtain a critical subsequence. It remains to show that for $B$ critical $\ell(B) = 2d(B)-2q_0(B)$. Suppose not, then it is strictly larger. Let $B'$ be obtained from $B$ by dropping the last column. Then, \[ \ell(B') \ = \ \ell(B)-1 \ \geq \ 2d(B) - 2 q_0(B) + 1 - 1 \] Now, we claim $d(B)-q_0(B)\geq d(B')-q_0(B')$. Clearly $d(B)-1\leq d(B')\leq d(B)$, and $q_0(B)-1\leq q_0(B') \leq q_0(B)$. However, if $q_0(B')= q_0(B)-1$, then we must also have $d(B')=d(B)-1$. Thus, in all cases, $d(B)-q_0(B)\geq d(B')-q_0(B')$. This implies $\ell(B')\geq 2(d(B')-q_0(B'))$, contradicting the criticality of $B$. \end{proof} The tight bound $\ell(B)=2d(B)-q_0(B)$ is the final key in proving the main rank lemma from this section, below. Since we prove that any critical subsequence has good rank, and any sequence has a critical subsequence, then this shows that any length-$2nk$ sequence has a high-rank subsequence. \begin{lem}\label{lem:crit-rank} Let $S$ be a BR sequence of length $2nk$ which has at least one inactive player. Let $B$ be some critical subsequence of $S$ and let ${\mathscr{L}}$ be $B$'s {transformation set}\xspace. Then ${\mathscr{L}}$ contains at least $\tfrac12d_1(B) + d(B)-q_0(B)$ linearly independent vectors. \end{lem} \begin{proof} Since $S$ has an inactive player, then so must $B$. Therefore, Lemma~\ref{lem:rank-d-q} applies. Recall, Lemma~\ref{lem:rank-d-q} shows that ${\mathscr{L}}$ contains at least $d_1(B)+\sum_{S'\in \mathbb S(B)} d(S')-q_0(S')$ linearly independent vectors. If $p_1(B)=d_1(B)=0$, then we are done. Otherwise, since $B$ is critical, then for all $S'\in \mathbb S(B)$, $\ell(S')<2(d(S')-q_0(S'))$. Hence, \[ rank({\mathscr{L}}) \ \geq\ d_1(B)+\sum_{S'\in \mathbb S(B)}d(S')-q_0(S') \ > \ \tfrac12 d_1(B) + \tfrac12 d_1(B) + \sum_{S'\in \mathbb S(B)} \tfrac12 \ell(S') \] However, $\ell(B) = d_1(B) + \sum_{S'\in \mathbb S(B)} \ell(S')$, and so this implies $rank({\mathscr{L}})\geq \tfrac12 d_1(B)+\tfrac12 \ell(B)$. By criticality and Claim~\ref{lem:crit}, $\ell(B)\geq 2(d(B)-q_0(B))$, giving us our desired bound. \end{proof} Since we have bounded the total increase on critical subsequences, though, rather than long sequences, we introduce the following notation: \begin{defn}\label{defn:Delta-p} Recall the notation $\Delta(\ell)$ from Definition \ref{def:Delta-l}. We similarly define $\Delta'(p)$ as the minimum total increase due to any critical subsequence with exactly $p$ active players, where the initial strategy profile is arbitrary. We also denote as $\overline{\Delta}(p)$ the minimum total increase taken over critical subsequences where $p_1\geq p_2$, and $\underline \Delta(p)$, the converse. Thus, $\Delta'(p) = \min\{\overline \Delta(p),\,\underline\Delta(p)\}$. \end{defn} Our first main result is to show that the probability of $\overline\Delta(p)$ being small is vanishing. It does not suffice, however, to follow the same structure as the ``informal lemma'' above: critical subsequences may be very short, and for any sequence $S$, its rank can not be more than $\ell(S)$. A probability bound of the form $k^n(nk)^\ell(\phi\epsilon)^\ell$ will require $\epsilon$ far too small if $\ell\ll n$, since the $k^n$ term will dominate. The following proof technique illustrates that no $k^n$ term is needed in a union bound when $p$ is small. \begin{thm}\label{thm:prob-p1} \( \Pr\big[\:\overline \Delta(p)\in (0,\epsilon)\big] \leq \left(\left(20\phi^2n^3k^3\right)^k\epsilon^{1/4}\right)^p. \) \end{thm} \begin{proof} We have shown in Lemma~\ref{lem:crit-rank} that a critical subsequence has high rank. However, as discussed above, this does not suffice to counteract simple union bounds. Recall the potential function ${\mathrm{payoff}}(\bm \sigma)$ from \eqref{eq:pot}, which represents the sum of the payoffs on all game edges at strategy $\bm \sigma$. For simplicity of notation, let $H(t)$ denote ${\mathrm{payoff}}({{\bm \sigma}}^t)$. Note that analyzing the change in $H(t)$ through any BR sequence is unaffected by shifting $H$ by a constant. We will the initial payoff as this constant: let $H'(t):= H(t)-H(0)$. Our goal is to bound the rate of change of $H'(t)$ over all possible BR sequence which form critical subsequences. Let $P$ be the set of all active players, and $Q$, the set of inactive players. Note that \begin{align*} H'(t)&= \sum_{u=1}^{n-1} \sum_{v=u+1}^n A((u,{{\bm \sigma}}^t_u)(v,{{\bm \sigma}}^t_v))-A((u,{{\bm \sigma}}^0_u)(v,{{\bm \sigma}}^0_v)) \end{align*} For simplicity of notation, denote $\widetilde A((u,\sigma_u)(v,\sigma_v)) = A((u,\sigma_u)(v,\sigma_v))-A((u,{{\bm \sigma}}^0_u)(v,{{\bm \sigma}}^0_v))$. Then \begin{align*} H'(t):= \sum_{u,v\in P} \widetilde A((u,{{\bm \sigma}}^t_u)(v,{{\bm \sigma}}^t_v)) + \sum_{w,w'\in Q} \widetilde A((w,{{\bm \sigma}}^t_w)(w',{{\bm \sigma}}^t_{w'})) + \sum_{u\in P} \sum_{w\in Q} \widetilde A((u,{{\bm \sigma}}^t_u)(w,{{\bm \sigma}}^t_{w})) \end{align*} Now, for $w\in Q$, ${{\bm \sigma}}^t_w={{\bm \sigma}}^0_w$, so the second term is 0. Furthermore, the inner-sum of the 3rd terms are in fact constants which depend on the strategy of the active player. Thus, define $C(u,\sigma):= \sum_{w\in Q} \widetilde A((u,\sigma)(w,{{\bm \sigma}}^0_w))$. Then the above sum can be expressed as \begin{align*} H'(t):= \sum_{u,v\in P} \widetilde A((u,{{\bm \sigma}}^t_u)(v,{{\bm \sigma}}^t_v)) + 0 + \sum_{u\in P} C(u,{{\bm \sigma}}^t_u) \end{align*} Also, note that $C(u,{{\bm \sigma}}^0_u)=0$, since the $\widetilde A$ terms cancel. This reduction to $C$ terms is the key to our analysis: To bound $\overline\Delta$, we must bound $H'(t)-H'(t-1)$ for all $t\geq 1$, and these values, in turn, depend only on the vector $A$, the initial strategies of the active players, the improving sequence, and the $C(u,\sigma)$ values for each of the pairs $(u,\sigma)$ which appear in the sequence. As noted, $C(u,{{\bm \sigma}}^0_u)=0$ for all $u$, so we need not consider them for initial strategies. We can not, however, enumerate all the possible $C$ values in our union bound. Note, however, that approximating the $C$ values simply approximated the $H$ values. Thus, we may round the $C$ values to the nearest multiple of $\epsilon$, as was first introduced in~\cite{A+17}. Let $C'(u,\sigma)$ be the nearest multiple of $\epsilon$ to $C(u,\sigma)$. Since $C(u,\sigma)\in [-n,n]$ for all $u$ and $\sigma$, then there are $2n/\epsilon$ possible choices for $C'(u,\sigma)$. Furthermore, for any time $t$, note that $H(t)-H(t-1)$ depends only on two $C$ terms, namely $C(u,{{\bm \sigma}}^{t}_u)$ and $C(u,{{\bm \sigma}}^{t-1}_u)$. Thus, replacing these with $C'$ terms modifies the value of $H(t)-H'(t)$ by at most $2\epsilon/2$, and so the event $H'(t)-H'(t-1)\in (0,\epsilon)$ is less likely than $\widetilde H'(t)-\widetilde H'(t-1)\in (-\epsilon,2\epsilon)$, where $\widetilde H'$ is simply the approximation due to replacing $C$ with $C'$. It remains, then, to apply Lemma~\ref{lem:probability}: for any one critical subsequence $S$ on $p$ players, if $p_1\geq p_2$, then by Lemma~\ref{lem:crit-rank}, the improvement of each step of the {\em approximate} potential along the sequence will lie in $(-\epsilon,2\epsilon)$ with probability $(3\phi\epsilon)^{d(S)-q_0(S)+p(S)/4}$. Taking a union bound over {\em all} approximated sequences, this event holds with probability $k^{p(S)}(nk)^{\ell(S)}(2n/\epsilon)^{d(S)-q_0(S)}(3\phi\epsilon)^{d(S)-q_0(S)+p(S)/4}$, since there are only $d(S)-q_0(S)$ $C$ values to approximate, and the rest of the value depends only on the initial configurations of the active players. Thus, noting that $d(S)-q_0(S)\leq k\cdot p(S)$, and since $S$ is critical, $\ell(S)\leq 2d(S)-2q_0(S)\leq 2kp(S)$, we have \begin{align*} \Pr\big[\:\overline\Delta \in (0,\epsilon)\big]&\leq k^{p(S)}(nk)^{\ell(S)}(2n/\epsilon)^{d(S)-q_0(S)}(3\phi\epsilon)^{d(S)-q_0(S)+p(S)/4}\\ &\leq 20^{k\cdot p(S)}(nk\phi)^{2k\cdot p(S)}(nk)^{k\cdot p(S)} \epsilon^{p(S)/4}\\ &= \left((20 n^3k^3\phi^2)^{k}\epsilon^{1/4}\right)^{p(S)} \end{align*} as desired. \end{proof} \subsubsection{Case II: Mostly Repeating Players}\label{sec:poly-case-p2} We have shown in the previous section how to bound the probability of $\overline \Delta$ being small. In this case, we handle $\underline \Delta$, that is, the minimum improvement of critical subsequences when $p_2\geq p_1$. The proof method in this case is very different from the converse case, but as above, we must still use work around the $k^n$ term in the standard union bound. This is done, in this case, by combining the columns of ${\mathscr{L}}$ into vectors which have zero entries for all payoff values with inactive players, and bounding their rank instead. \begin{lem}\label{lem:rank-p2} Let $S$ be a BR sequence, and let ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$ be its {transformation set}\xspace. Then the span of ${\mathscr{L}}$ contains at least $p_2(S)/2$ linearly independent vectors $V_1,\,\dotsc,\,V_{p_2/2}$, such that for all~$j$, \begin{enumerate} \item[(i)] the vector $V_j$ is a 0-1 combination of the $L_i$'s, and \item[(ii)] the value $\ip{V_j,A}$ does not depend on the strategies of the inactive players. \end{enumerate} \end{lem} \begin{proof} Fix a repeating player $u$, and denote one of its repeating strategies as $\sigma$. Let $t_0$ be the index of the first occurrence of $(u,\sigma)$ in $S$, setting $t_0=0$ if $i={{\bm \sigma}}^0_u$. Let $t_0<t_1<t_2<\dotsm$ be all occurrences of player $u$ in the sequence $S$ starting with $t_0$, and suppose $t_\sigma$ is the index of the second occurrence of $(u,\sigma)$. Formally, if $S=(u_1,i_1),(u_2,i_2),\dotsm$, then $u_{t_j}=u$ for all $j$, $i_{t_0}=i_{t_s}=\sigma$, and $i_{t_j}\neq \sigma$ for all $0< j<s$. Consider, then, the vector $V(u) = \sum_{j=1}^s L_{t_j}$. Vector $V(u)$ satisfies condition $(i)$ by construction. We will show that it satisfies $(ii)$, and that at least $p_2/2$ of them must be linearly independent. Let $w$ be any inactive player. For simplicity of notation, denote the strategy $i_{t_j}$ as $i_j$. Consider the inner product $\ip{V(u),A}$ when restricted to the rows indexed by $((u,*)(w,*))$:\\[-20pt] \begin{align*} \ip{V(u),A}_{\big|((u,*)(w,*))}&=\sum_{j=1}^s A((u,{i_j})(w,{{\bm \sigma}}^0_w)) - A((u,{i_{j-1}})(w,{{\bm \sigma}}^0_w))\\ &= A((u,{i_s})(w,{{\bm \sigma}}^0_w))-A((u,{i_0})(w,{{\bm \sigma}}^0_w))\\ &=0 \end{align*} Therefore, the inner product $\ip{V(u),A}$ is independent of the value of ${{\bm \sigma}}^0_w$. Since this holds for all inactive $w$, we have proved part $(ii)$. Now, it suffices to argue that some collection of $p_2/2$ many $V$ vectors are linearly independent. We begin by constructing an auxiliary directed graph $G'=(V, E')$, where $V$ is the set of players, and $E'$ will be defined as follows: let $u$ be some repeating player, and define $V(u)$ as above, for the repeat move $(u,\sigma)$. The $V(u)$ vector can not be entirely 0, as this would imply that ${{\bm \sigma}}^{t_s}$ and ${{\bm \sigma}}^{t_0}$ are the same, and so the sequence could not have been strictly improving. Then, for {\em every} player $w$ such that $V(u)$ for $(u,\sigma)$ has a non-zero $((u,\sigma)(w,*))$ entry, add the edge $(u,w)$ to $E'$. Consider the following procedure: pick an arbitrary vertex $r_1\in P_2$, and let $T_1$ be the BFS arborescence rooted at $r_1$ which spans all nodes reachable from $r_1$. Then delete $V(T_1)$ from $G'$ and repeat, picking an arbitrary root vertex $r_2\in P_2\setminus V(T_1)$, and get the arborescence $T_2$ on everything reachable from $r_2$. We may continue this until every vertex of $P_2$ is covered by some arborescence. For each $i=1,2,\dotsc$, let $T_i^0$ and $T_i^1$ be the set of nodes of $T_i$ which are of even or odd distance from $r$ along $T_i$, respectively. Let $P'_i$ be the larger of $V(T_i^0)\cap P_2$ and $V(T_i^1)\cap P_2$, and $P_2':= \bigcup_{i=1}^\infty P_i'$. We must have that $|P_2'|\geq |P_2|/2 = p_2/2$. We wish to show that the collection $\mathcal V:=\{V(u):u\in P_2'\}$ is independent. Every $u\in P_2'$ must have some out-neighbour $w$. If $u$ was not a leaf of the arborescence it was selected in, then it must have some out-neighbour along the arborescence, and we may choose this neighbour. This out-neighbour can not also be in $P_2'$. In this case, $V(u)$ will be the only vector from $\mathcal V$ to contain a non-zero $((u,*)(w,*))$ entry, since $w$ was not taken in $P_2'$. If, instead, $u$ was a leaf of its arborescence, then its out-neighbours must be in previously constructed arborescences. Let $w$ be any such neigbour, then $V(w)$ can not contain a non-zero $((u,*)(w,*))$ entry, as otherwise $u$ would have been in the other arborescence. Therefore, $V(u)$ is the only vector in $\mathcal V$ to contain a nonzero $((u,*)(w,*))$ entry. Thus, $\mathcal V$ must contain a $|\mathcal V|\times |\mathcal V|$ diagonal submatrix, and therefore has rank at least $|\mathcal V|\geq p_2/2$, as desired. \end{proof} Therefore, we have shown that the transformation set of any sequence must have large rank if $p_2\geq p_1$. However, the vectors of the transformation set depend on the strategies of inactive players, whereas these $V$ vectors do not. We will show that this issue can be circumvented, and prove our second main result. Recall that $\underline\Delta(p)$ is the (random) minimum improvement of any critical subsequence $S$ with $p$ players and $p_2(S)\geq p_1(S)$. \medskip \begin{thm}\label{thm:prob-p2} $\Pr\big[\underline\Delta(p)\in(0,\epsilon)]\leq\left(2(nk)^{2k}k^{5/4}(n\phi\epsilon)^{1/4}\right)^p$. \end{thm} \begin{proof} Let ${\mathscr{L}}$ be the transformation set of some critical subsequence $S$, and $\mathcal V$ be a collection of $p_2(S)/2$ independent $V$ vectors from Lemma~\ref{lem:rank-p2} applied to $S$. If $V\in \mathcal V$ is given by $V=\sum_{i=1}^m L_{t_i}$ for some collection of indices $t_i$, then we have that $\Pr[\bigwedge_{i=1}^m\ip{L_{t_i},A}\in(0,\epsilon)]\leq \Pr[\ip{V,A}\in (0,m\epsilon)]$. Since $m\leq \ell$, then taking the collection of all $V$ vectors and applying Lemma~\ref{lem:probability}, we have \begin{align*} \Pr\left[\textstyle\bigwedge_{t=1}^{\ell(S)}\ip{L_{t},A}\in(0,\epsilon)\right]&\leq \Pr\left[\textstyle\bigwedge_{V\in \mathcal V}\ip{V,A}\in (0,\ell\epsilon)\right] \leq (\ell\phi\epsilon)^{p_2(S)/2} \end{align*} Note that to construct the $V$ vectors, it suffices to have the initial strategies of the active players, and the BR sequence. Thus, there are at most $k^{p(S)}(nk)^{\ell(S)}$ possible collections $\mathcal V$. Since we are restricting ourselves to critical subsequences, we have $\ell(S)=2(d(S)-q_0(S))\leq k\cdot p(S)$. Therefore, we have \begin{align*} \Pr\big[\underline\Delta(p)\in(0,\epsilon)]&\leq k^{p(S)}(nk)^{\ell(S)}(\ell\phi\epsilon)^{p_2(S)/2}\\ &\leq n^{2k\cdot p(S)}k^{(2k+1)p(S)}(2kp)^{p(S)/4}(\phi\epsilon)^{p(S)/4}\\ &\leq \left(2(nk)^{2k} k^{5/4}(n\phi\epsilon)^{1/4}\right)^{p(S)} \end{align*} as desired. \end{proof} \subsubsection{Combining Both Cases} We have shown above that $\overline\Delta(p)$ and $\underline\Delta(p)$ have vanishing probability of lying in $(0,\epsilon)$. In this section, we use these results to show that the BRA will terminate in time polynomial in $n^k$, $k$ and $\phi$, with high probability, when the game graph is complete. The following lemma combines our two previous results: \begin{lem}\label{poly:sequence-of-length-2nk} Given an improving sequence of length $2nk$, the minimum improvement after performing all moves in the sequence is at least $\epsilon=(nk\phi)^{-O(k)}$ with probability $1-1/poly(n,k)$. \end{lem} \begin{proof} We will perform a case analysis based on the values of $p_1(S)$ and $p_2(S)$, with cases for $p(S)=n$, $p(S)<n$ and $p_1(S)\geq p_2(S)$, and $p_2(S)\geq p_1(S)$. If $p(S)=n$, we apply the rank bound of Corollary~\ref{cor:all-active} and take a union bound over all initial strategy profiles, and all possible sequences to get \begin{align*}\label{poly-delta-fixed-n} \Pr[\Delta'(n)\in (0,\epsilon)] &\leq k^n(nk)^{2nk} (\phi\epsilon)^{n-1}\\ &\leq \left(k^{3k}n^{2k}\phi\epsilon\right)^n\Big/\phi\epsilon \end{align*} This union bound over-counts the number of sequences with $p(S)=n$, but this isn't a problem. Setting $\epsilon = \phi^{-1}\left(n^2k^3\right)^{-2k}$ gives, for $n$ sufficiently large, \( \Pr[\Delta'(n)\in (0,\epsilon)]\leq\left(\frac 1{n^2k^3}\right)^{n}. \) In the converse case, we combine Theorems~\ref{thm:prob-p1} and~\ref{thm:prob-p2}, then take a union bound over all possible values of $p$ to bound the probability for any sequence of the given length. As defined previously, $\Delta'(p) = \min\{\overline\Delta(p),\,\underline\Delta(p)\}$ and so, \begin{align} \Pr[\Delta'(p)\in (0,\epsilon)] \ \leq \ \left( (20\phi^2n^3k^3)^k\epsilon^{1/4} \right)^p + \left( 2(\phi\epsilon)^{1/4}n^{2k+1/4}k^{2k+5/4} \right)^p \ \leq\ 2\left( (20\phi^2n^3k^3)^k\epsilon^{1/4} \right)^p \end{align}\label{poly-delta-fixed-p} Since any sequence of length $2nk$ must contain a critical subsequence, it suffices to set $\epsilon=\left(20\phi^2n^3k^3\right)^{-4k-4}$, and taking the union bound over all choices of $p$, we get \begin{align*} \Pr[\Delta'\in(0,\epsilon)]&\leq \sum_{p=1}^n \left(20\phi^2n^3k^4\right)^{-p} \ \leq\ \frac{1}{(20\phi^2n^3k^4)-1} \end{align*} Combining the two cases of $p=n$ and $p<n$ gives us our desired result. \end{proof} It remains to conclude that any execution of the BRA will find a PNE in polynomial time with high probability. \begin{thm}\label{thm:whp-poly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} on a complete game graph, and with an arbitrary initial strategy profile, then any execution of BRA where improvements are chosen arbitrarily will converge to a PNE in at most $(nk\phi)^{O(k)}$ steps, with probability \mbox{$1-1/poly(n,k,\phi)$}. \end{thm} \begin{proof} Lemma \ref{poly:sequence-of-length-2nk} directly implies the theorem. As outlined in the ``common framework'' (Section~\ref{sec:common}), we begin by partitioning the BRA sequence into blocks of length $2nk$. Each such block must contain a critical subsequences, and therefore with probability $1-1/poly(n,k,\phi)$ every block in the partition increases the potential by at least $\epsilon=(nk\phi)^{-O(k)}$ Since the total improvement is at most $2n^2$, since there are only $\binom n2$ games, this implies that the BR algorithm can only make at most $2n^2(nk\phi)^{O(k)}$ moves. Since making one move takes time polynomial in $n$ and $k$, we are done. \end{proof} \subsection{Smoothed Quasi-polynomial Complexity for Arbitrary Graphs} \label{sec:quasipoly} In this section we show the quasi-polynomial running time when the game graph $G$ is incomplete, and thus prove Theorem \ref{thm:qpoly}. The analysis mostly uses the lemmas from Section~\ref{sec:poly-case-p2}, paired with the following definition and lemma, from~\cite{ER14}: \begin{defn}\label{def:Delta-l} Recall the random variable $\Delta$ from Definition~\ref{def:Delta}. Call a sequence of length $\ell$ log-repeating if it contains at least $\ell/(5\log(nk))$ repeating moves (pairs). We denote as $\Delta(\ell)$ the minimum total potential-improvement after any log-repeating BR sequence of length exactly $\ell$. \end{defn} \begin{lem}[From \cite{ER14}, Lemma 3.4]\label{lem:min-repeating-pairs} Let $\Delta_N$ and $\Delta(\ell)$ be as above. Then $\Delta_{5nk}:= \min_{1\leq \ell\leq 5nk}\Delta(\ell)$ \end{lem} The proof of the above lemma proves that any sequence on $5nk$ pairs must contain some contiguous sub-sequence which is log-repeating. Thus, for the remainder of the analysis, it suffices to bound $\Delta(\ell)$. Since a sequence captured by $\Delta(\ell)$ must have at least $\ell/(5\log(nk))$ repeated terms, it must have $p_2\geq \ell/(5k\log(nk))$. Therefore, as we have shown in the proof of Theorem~\ref{thm:prob-p2}, we have $\Pr[\Delta(\ell)\in (0,\epsilon)]\leq k^{\ell}(nk)^\ell (\ell\phi\epsilon)^{\ell/10k\log(nk)}$. It suffices, then to simply take the union bound over all possible values of $\ell$. \begin{thm}\label{thm:in-text-qpoly} Given a smoothed instance of {\em $k$-{NetCoordNash}\xspace} with an arbitrary initial strategy profile, then any execution of a BR algorithm where improvements are chosen arbitrarily will converge to a PNE in at most $\phi\cdot (nk)^{O(k\log(nk))}$ steps, with probability $1-1/poly(n,k)$. \end{thm} \begin{proof}As discussed above, \begin{align}\label{quasi-delta-fixed} Pr[\Delta(\ell)\in (0,\epsilon)]&\leq k^{\ell}(nk)^{\ell}(\ell\phi\epsilon)^{\ell/10k\log(nk)}\nonumber \\ &\leq \left(k^2n(5nk\phi\epsilon)^{1/(10k\log(nk))}\right)^{\ell} & (\ell\leq 5nk)\nonumber\\ &\leq \left(2k^3n^2 (\phi\epsilon)^{1/(10k\log(nk))}\right)^\ell. &(5^{1/10}\leq 2) \end{align} Setting $\epsilon = \phi^{-1}(2n^2k^3)^{-2\cdot 10k\log(nk)}$, this gives \[ Pr[\Delta(\ell)\in (0,\epsilon)] \leq\left(\frac{1}{2n^2k^3}\right)^\ell \] Let $\Delta_{5nk}$ be the improvement in potential in any length $5nk$ BR sequence. Then using Lemma \ref{lem:min-repeating-pairs}, and taking the union bound over all choices of $\ell$, we have, \[\Pr[\Delta_{5nk}\in (0,\epsilon)]\leq \sum_{\ell=1}^{5nk}\Pr[\Delta(\ell)\in(0,\epsilon)]\le \sum_{\ell=1}^{5nk} (2n^2k^3)^{-\ell}\leq \frac{(2n^2k^3)^{-1}}{1-(2n^2k^3)^{-1}}=\frac{1}{2n^2k^3 -1} \] Hence, with probability $1-1/poly(n,k)$ (over the draw of payoff vector $A$), {\em all} BR sequences of length $5nk$ will have total improvement at least $\epsilon$. In that case, any execution of BR algorithm makes an improvement of at least $\epsilon$ every $5nk$ moves. Since the total improvement is at most $2n^2$, we conclude that the total number of steps is at most $5nk\cdot 2n^2/\epsilon = 10n^3k(2n^2k^3)^{20k\log(nk)}\cdot \phi= \phi\cdot (nk)^{O(k\log(nk))}$, and this occurs with probability $1-1/poly(n,k)$. \end{proof} This completes our analysis of the smoothed performance of BRA for finding pure Nash equilibria in network coordination games. In the next section, we show that this result indeed holds in expectation, and then go on to show a notion of smoothness-preserving reduction which allows us to prove alternative, conditional, algorithms for this problem. \section{Expected Smoothed Time Complexity}\label{sec:exp} The analysis in the previous section establishes smoothed complexity of {network-coordination games}\xspace with respect to the with high probability notion. Another aspect of smoothed analysis is to analyze the expected time of completion of the algorithm. In this section, we provide a theorem to obtain expected time results from the with high probability bounds. The results are presented in a general form to allow application to any problem in PLS that has a bounded total improvement in potential value. \begin{thm}\label{thm:expected-time} Given a PLS problem with input size $N$, potential function range $[-N^{r_1}, N^{r_2}]$, and a local-search algorithm ${{\mathscr A}}$ to solve it, let $d$ be the number of distinct choices the algorithm has in each step and let $\Lambda$ be the total size of the search space of the algorithm. For an instance $I$ drawn at random with maximum density $\phi$, suppose the probability that any length-$N^\beta$ sequence of improving moves of ${{\mathscr A}}$ results in total improvement in the potential value at most~$\epsilon$, is at most $\sum_{q=1}^{N^\beta}((\phi N)^{f(N)}(\phi\epsilon)^{1/g(N)})^q$. Then the expected running-time of the algorithm is $O(N^{\beta+r}\cdot g(N)\cdot (\phi N)^{f(N)g(N)}\cdot \ln\Lambda)$. Here, $f(N)$ and $g(N)$ are functions of $N$. \end{thm} \begin{proof} The proof is from \cite{ER14}. As we have stated it in a more general form, the analysis is included for completeness. The maximum improvement possible before ${{\mathscr A}}$ terminates is the maximum change in the potential function value, given by $N^{r_2}+N^{r_1}$. For any integer $t\geq 1$, if the algorithm requires more than $t$ steps to terminate, then there must exist some subsequence of length $N^{\beta}$ that results in an improvement in the potential value of less than $N^{\beta}(N^{r_2}+N^{r_1})/t\leq 2N^{\beta+{\max\{r_2,r_1\}}}/t$. We denote $r:=\max\{r_1,r_2\}$. We define a random variable $T$ as the number of steps ${{\mathscr A}}$ requires to terminate. Using the notation $\Delta(N^\beta)$ to denote the minimum total improvement in a length-$N^\beta$ sequence of the algorithm ${{\mathscr A}}$, this gives the probability of ${{\mathscr A}}$ running for more than $t$ steps as: \begin{equation*} \Pr[T\geq t]\leq Pr[\Delta(N^\beta)\in (0,N^{r+\beta}/t)] \leq \sum\limits_{q=1}^{N^\beta} \left((\phi N)^{f(N)}\left(\phi\cdot \frac{N^{\beta+r}}{t}\right)^{1/g(N)}\right)^q. \end{equation*} We define $t=\gamma i$, for $\gamma=(\phi N)^{f(N)g(N)}(\phi N^{r+\beta})=\phi^{f(N)g(N)+1}N^{f(N)g(N)+\beta+r}$, and compute the probability of $T\geq \gamma i$ for any integer $i$: \begin{equation*} \Pr[T\geq \gamma i]\leq \sum\limits_{q=1}^{N^\beta} \left((\phi N)^{f(N)}\left(\phi\cdot\frac{N^{r+\beta}}{\gamma i}\right)^{1/g(N)}\right)^q \leq \sum\limits_{q=1}^\infty \left(\frac{1}{i}\right)^{q/g(N)} \leq g(N)\sum_{q'=1}^\infty \left(\frac 1i\right)^{q'}\leq \frac{g(N)}{i-1}. \end{equation*} We now sum over all values of $t$, by using that $\Pr[T\geq t]\leq \Pr[T\geq t\cdot\lceil t/\gamma \rceil]$, and compute the expected time steps as: \begin{align*} \mathbb E[T]&=\sum_{t=1}^\Lambda \Pr[T\geq t] \leq \sum_{i=1}^{\Lambda/\gamma} \sum_{t=1}^\gamma \Pr[T\geq (i+1)\gamma] \leq \sum_{i=2}^{\Lambda/\gamma} \frac{g(N)\gamma}{i-1} = O(g(N)\cdot\gamma\cdot\ln\Lambda) \end{align*} Thus, replacing the value for $\gamma$, the expected runtime is at most $O(N^{\beta+r}g(N)(\phi N)^{f(N)g(N)}\ln\Lambda)$ \end{proof} \begin{cor}\label{corr:expected-quasi} The smoothed expected time for BR to terminate for all {network-coordination games}\xspace is polynomial in $(n^{(k\log(nk))},\phi)$. \end{cor} \begin{proof} From ~\eqref{quasi-delta-fixed} in Theorem \ref{thm:in-text-qpoly}, we know that the probability that the minimum improvement in a fixed BR sequence of length $5nk$ is at most $\epsilon$, is at most $\sum_{\ell=1}^{5nk}\left(2n^2k^3(\phi \epsilon)^{1/(10k\log(nk))}\right)^l$. Applying Theorem \ref{thm:expected-time}, for $N=nk$ and $\Lambda \leq k^n$, we get $f(N)=O(1)$, $N^{r+\beta}\leq N^3$, and $g(N)=O(k\log(nk))$, and the result follows. \end{proof} \begin{cor}\label{corr:expected-poly} For complete graphs, the smoothed expected time for BR to terminate for {network-coordination games}\xspace is polynomial in $(n^k,\phi)$. \end{cor} \begin{proof} From \eqref{poly-delta-fixed-n} in Lemma \ref{poly:sequence-of-length-2nk}, for the case of complete graphs when a BR sequence has all active players, we have: \begin{equation*} \Pr[\Delta(p)\in (0,\epsilon)]\leq \left(k^{3k}n^{2k}\phi\epsilon\right)^n\Big/\phi\epsilon \leq \sum_{i=1}^n\left(k^{3k}n^{2k}\phi^{1/2}\epsilon^{1/2}\right)^i\Big/\phi\epsilon. \end{equation*} Similarly, from \eqref{poly-delta-fixed-p} in Lemma \ref{poly:sequence-of-length-2nk}, the probability that the minimum improvement in a BR sequence of length $2nk$ is at most $\epsilon$, is given by: \begin{equation*} \Pr[\Delta(p)\in (0,\epsilon)] \leq \sum_{p=1}^n 2\left((100\phi^2n^3k^4)^k\epsilon^{1/4}\right)^p \end{equation*} Combining these sums, we get the probability that a BR sequence of length $2nk$ has improvement at most $\epsilon$ is: \begin{align*} \Pr[\Delta(p)\in (0,\epsilon)] &\leq \max\left\{\sum_{p=1}^n 2\left((100\phi^2n^3k^5)^k\epsilon^{1/4}\right)^p, \sum_{i=1}^n\left(k^{3k}n^{2k}\phi^{1/2}\epsilon^{1/2}\right)^i\right\}\\ &\leq \sum_{j=1}^n ((\phi^c_1(nk)^{c_2k}(\phi\epsilon)^{1/c_3})^j. \end{align*} Applying Theorem \ref{thm:expected-time}, for $N=nk$, $N^{r+\beta}\leq N^3$, and $\Lambda \leq k^n$, we get $f(N)=O(1)$ and $g(N)=O(1)$, and the result follows. \end{proof} \subsection{Notation} \begin{defn} A \textit{congestion game} among $n$ players is defined by a set of resources $E$ of size $r$, an explicitly described strategy set $S_i\subseteq 2^E$ for all players$i\in [n]$, and for each resource $e\in E$, a monotone cost vector $c_e\in\mathbb R^{n+1}$, where we assume $c_e(i)-c_e(i-1)\in [0,1]$ for all $i\geq 1$. A strategy profile ${{\bm \sigma}}=(s_1,\,s_2,\,\dotsc,\,s_{n-1},\,s_n\}$, with $s_i\in S_i$ for all $i\in [n]$ denotes a particular outcome of the game. Let $n_e({{\bm \sigma}})$ denote the number of strategies $s_i$ in ${{\bm \sigma}}$ which include resource $e$, {\em i.e.} $n_e:=|\{i\in[n] : e\in s_i\}|$, then the cost of resource $e$ at profile ${{\bm \sigma}}$ is $c_e(n_e({{\bm \sigma}}))$. Thus, the cost player $i$ pays is $\sum_{e\in s_i} c_e(n_e({{\bm \sigma}}))$. \end{defn} The objective is to find a {\em Nash equilibrium} of these games, which is defined naturally. We include the formal definition below for completeness. \begin{defn} A \textit{Nash equilibrium} of a congestion game is a strategy profile whose cost cannot be decreased by unilateral deviation of any player. Formally, \[ \sum_{e\in s_i} c_e(n_e(s_i,{{\bm \sigma}}_{-i})) \leq \sum_{e\in s_i'} c_e(n_e(s_i',{{\bm \sigma}}_{-i}))\quad \text{ ~ for all ~ }i\in [n]\text{ ~ and ~ }s_i'\in S_i \] $(s_{i}',{{\bm \sigma}}_{-i})$ denotes a strategy profile where every player except $i$ chooses the same strategy as in~${{\bm \sigma}}$. \end{defn} In fact, similarly to the network coordination games, this game also admits a {\em potential function} \begin{equation} \Phi({{\bm \sigma}}):= \textstyle\sum_{e\in E}\sum_{i=0}^{n_e({{\bm \sigma}})} c_e(i) \end{equation} It can be shown that ${{\bm \sigma}}$ is a Nash equilibrium if for any player $i$ and strategy $s_i\in S_i$, $\Phi(s_i,{{\bm \sigma}}_{-i})\leq \Phi(s_i',{{\bm \sigma}}_{-i})$. In fact, if $\psi_i({{\bm \sigma}})$ is the cost payed by player $i$ at profile ${{\bm \sigma}}$, then \[\psi_i(s_i,{{\bm \sigma}}_{-i}) - \psi_i(s_i',{{\bm \sigma}}_{-i}) = \Phi_i(s_i,{{\bm \sigma}}_{-i}) - \Phi_i(s_i',{{\bm \sigma}}_{-i})\] This observation, and others, may be found at~\cite{RoughLectureCongestion} If two strategy profiles are said to be as {\em adjacent} when they differ in the strategy choice of a single player, then a PNE of a congestion game is a profile which minimizes $\Phi$ in its neighbourhood. To find a local optimum of $\Phi$, the better-response algorithm (BRA) for the game starts at an arbitrary strategy profile ${{\bm \sigma}}^0$, and recursively checks if the total cost can be improved by unilateral deviation of some player, and moves to an arbitrary profile with non zero improvement. That is, for all $t\geq 0$, if ${{\bm \sigma}}$ is adjacent to ${{\bm \sigma}}^t$ with $\Phi({{\bm \sigma}})>\Phi({{\bm \sigma}}^t)$, then set ${{\bm \sigma}}^{t+1}\gets {{\bm \sigma}}$. This process continues until we reach a PNE of the game. Although the aim of a congestion game is to (locally) \textit{minimize} the total cost, to use the technical lemmas of $k$-{NetCoordNash}\xspace we modify the input by inverting the signs of all entries of the cost vector, and aim to (locally) maximize the potential $-\Phi({{\bm \sigma}})=-\sum_{e\in E}\sum_{i\le n_e({{\bm \sigma}})}c_e(i)$. Hence forth, we abuse notation slightly and assume the given cost vector is this sign inverted-vector. \subsection{Model} We define a $(k,\ell)$-congestion game, as a congestion game where the number of distinct strategies a player can choose from is at most $k$ ($|S_i|\leq k$ for all $i$), and any resource belongs in at most $\ell$ strategies overall ($|\{s\in \bigcup_{i=1}^n S_i:e\in s\}|\leq \ell$ for all $e\in E$). We denote the problem of finding a PNE in a $(k,\ell)$-congestion game as $(k,\ell)$-{Congestion}\xspace. A move in the game is a tuple $(u,s_u)$, denoting that player $u$ deviates to strategy $s_u$. We represent every move by an $r\cdot (\ell+1)$-length vector $L$, where the $(e,i)^{th}$ entry is $+1$ if $n_e$ went from $i\pm 1$ to $i$ due to the move, $-1$ if $n_e$ went $i$ to $i\pm 1$ due to the move, and $0$ if the move leaves $n_e$ unchanged. Let $C$ be the $r(\ell+1)$-dimensional vector whose $(e,i)$-th entry is $\sum_{j=1}^i c_e(j)$, for all $e\in E$ and all $0\leq i\leq \ell$. Note, then, that given a move vector $L_t$, the inner product $\ip{L_t,C}=\Phi({{\bm \sigma}}^t)-\Phi({{\bm \sigma}}^{t-1})$. Recall, from section~\ref{sec:notation}, the definitions of {transformation set}\xspace (${\mathscr{L}}$), {\em active} and {\em inactive} players, and {\em repeating} and {\em non-repeating} players, whose quantities are denoted as $p_2$ and $p_1$, respectively. A sequence of BR moves thus generates a {transformation set}\xspace as before, ${\mathscr{L}}=\{L_1,L_2,..\}$, where $L_t$ is defined just above. The main difference between this model and that of $k$-{NetCoordNash}\xspace is that, instead of the rows of the vectors being denotes by pairs of moves (``$((u,i)(v,j))$''), they are denoted by resource-utilization amounts, (``$(e,i)$''). In most other regards, the matrix arguments are similar. \paragraph{Smoothed $(k,\ell)$-congestion games.} To analyze the smoothed complexity of this problem, we must define a smoothed model of congestion games. Since the real-valued input to the problem is the cost vectors of resource utilization, the random input must be the cost vectors. However, we have required $c_e(i)$ to be monotone in $i$. To this end, define the differential cost $d_e(i) := c_e(i)-c_e(i-1)$, with $d_e(0):= c_e(0)$. Then, we may restrict $d_e(i)$ to be a random variable supported on $[0,1]$ with density at most $\phi$, and then the cost function $c_e(i) = \sum_{j=0}^i d_e(i)$ will be monotone. However, the $C$ vector neither consists of the $c_e$ values or the $d_e$ values. Instead, let $\bar c_e(i) = \sum_{j=0}^i c_e(i) = \sum_{j=0}^i \sum_{m=0}^j d_e(m)$. Note, however, that the $c_e$'s are a full-rank combination of the $d_e$'s, and the $\bar c_e$'s are a full-rank combination of the $c_e$'s, and so there must exist a square, full-rank matrix $U$ such that $C= UD$, where $D$ is the vector of $d_e$ values. An instance of a {\em smoothed $(k,\ell)$-congestion game} is then given by a ground set $E$, $n$ sets of strategies $S_i\subseteq 2^E$, and a random vector $D$ of differential payoff values. We wish to show: \begin{thm}\label{thm:CGquasipoly} Given a smoothed instance of {\em $(k,\ell)$-{Congestion}\xspace} with an arbitrary initial strategy profile, then any execution of a BR algorithm where improvements are chosen arbitrarily will converge to a PNE in at most $\phi r(nk)^{O(\ell k\log(nk))}$ steps, with probability $1-1/poly(n,k)$. \end{thm} The idea is, again, to prove that any arbitrary $O(nk)$-length BR sequence of sufficiently high rank has polynomially large improvement {\em w.h.p.}, then applying a union bound over all choices for a sequence, to prove that the any execution of BRA over a sufficiently long sequence will have good improvement with good probability. This will allow us to prove the above result. We first show: \begin{lem}\label{lem:rank-p2-congestion} Let $S$ be a BR sequence in a $(k,\ell)$-congestion game, and let ${\mathscr{L}}=\{L_1,\,L_2,\,\dotsc\}$ be its {transformation set}\xspace. Then the span of ${\mathscr{L}}$ contains at least $p_2(S)/\ell$ linearly independent vectors $V_1,\,\dotsc,\,V_{p_2/\ell}$ that are all 0-1 combinations of the $L_i$'s. \end{lem} \begin{proof} This proof follows the outline of that of lemma~\ref{lem:rank-p2}. Fix a repeating player $u$, and denote one of its repeating strategies as $s_u$. Let $t_u$ be the first occurrence of $(u,s_u)$ in the sequence, setting $t_u=0$ if $i_u={{\bm \sigma}}^0_u$. Let $t_{u}'$ be the first occurrence of $(u,s_u)$ after $t_u$. Formally, if $S=\{(v_1,s_1),(v_2,s_2),\dotsm\}$, then $v_{t_u}=v_{t'_{u}}=u$, $s_{t_u}=s_{t'_{u}}=s_u$, and $(v_t,s_t)\neq (u,s_u),\ \forall t_u<t<t'_{u}$. Consider the vector $V(u) = \sum_{\tau= t_u}^{t'_u}L_{\tau}$. Vector $V(u)$ is a 0-1 combination of the $L_i's$ by construction. We will show that at least $p_2/\ell$ of these must be linearly independent. \hrule Note that, the sum of all move vectors between two consecutive repeating moves of a player is a vector with entry $1$ in a row $(e,j)$ if the number of players using resource $e$ is $j$ at the end of the second appearance of the move, and $-1$ in the row $(e,j+1)$ or $(e,j-1)$, denoting that prior to this move the number of such players was different by $1$. The inner product $\ip{V(u),C}$ represents the total change in the cost of the game because of the sequence of moves between times $t_u$ and $t_{u_r}$. Thus, by definition of an improving sequence, at least one entry in every $V(u)$ must be non zero. Consider a matrix $M_1$ where the columns are all vectors $V(u)$ defined above, arranged in chronological order of the second occurrence of the repeated move. We rename the players in this matrix by $1,2,\dotsc$ for notational convenience. Every column in $M_1$ must have a $1$ entry. As a resource $e_i$ can belong in at most $\ell$ strategies at once, it is non zero in at most $\ell$ of the column vectors $V(i)$s. Denote the set of column vectors where $(e_i,n_i)\ne 0$ by $V_{(e_i,n_i)}$. We create a matrix $M$ by taking exactly one vector from every group of vectors $V_{(e_i,n_i)}$, and removing the rest from $M_1$. This leaves a set of at least $p_2/\ell$ vectors, that each have at least one non zero ($1$) entry, that is zero in all the rest. Thus, by a suitable rearrangement, we can form a diagonal sub-matrix of size $p_2/\ell$, thereby proving the claim. \end{proof} \begin{lem}\label{lem:cg:prfixedseq} The probability that a BR sequence $S$ of length $\lambda$ has total improvement in potential at most $\epsilon$ is at most $(\lambda\phi\epsilon)^{p_2/\ell}$, where $p_2$ is the number of repeating players in $S$. \end{lem} \begin{proof} Similar to the proof of Lemma \ref{thm:in-text-qpoly}, the probability of total improvement of the sequence $S$ being at most $\epsilon$, is at most $$Pr[\sum\limits_{i=1}^\lambda \ip{L_i,C}\in (0,\epsilon)|\ip{L_i,C}>0,\ \forall i]\le Pr[\bigwedge_{i=1}^{p_2/\ell} \ip{V_j,C} \in (0,\epsilon)]\le (\lambda\phi\epsilon)^{p_2/\ell}.$$ \end{proof} We use Definition \ref{def:Delta-l} and Lemma \ref{lem:min-repeating-pairs} to prove Theorem \ref{thm:CGquasipoly}. For an arbitrary BR sequence $S$ of $\lambda$ moves where at least $\lambda/(5log(nk))$ distinct moves appear at least twice, the number of repeating players is at least $\lambda/(5klog(nk))$. By Lemma \ref{lem:cg:prfixedseq} the probability that the total improvement due to $S$ is at most $\epsilon$, is at most $(\lambda\phi\epsilon)^{\lambda/\ell klog(nk)}$. To find the probability that any sequence is at most $\epsilon$-improving (gives an improvement of at most $\epsilon$), we take a union bound over all possible initial strategy profiles, and all $(nk)^\lambda$ choices of a BR sequence, to get the probability bound \begin{equation*} \begin{split} Pr[\Delta(\lambda)\in (0,\epsilon)]&\le k^\lambda(nk)^\lambda(\lambda\phi\epsilon)^{p_2/\ell}\\ &\le \left(k(nk)(5nk\phi\epsilon)^{1/5\ell klog(nk)}\right)^{\lambda}\\ &\le \left(2k^3n^2(\phi\epsilon)^{1/5\ell klog(nk)}\right)^{\lambda} \end{split} \end{equation*} We set $\epsilon=\phi^{-1}(2n^2k^3)^{-2\cdot \ell k log(nk)}$ to get $$Pr[\Delta(\lambda)\in (0,\epsilon)]\le \left(\frac{1}{2n^2k^3}\right)^{\lambda}.$$ Using Lemma \ref{lem:min-repeating-pairs}, we now take a union bound over all choices of $\lambda$ to get: $$Pr[\Delta_{5nk}\in (0,\epsilon)]\le \sum\limits_{i=1}^{5nk}(2n^2k^3)^{-\lambda}\le \frac{1}{2n^2k^3-1}.$$ Hence, with probability $1-1/poly(n,k)$ (over the draw of cost vector $C$), {\em all} BR sequences of length $5nk$ will have total improvement at least $\epsilon$. In that case, any execution of BR algorithm makes an improvement of at least $\epsilon$ every $5nk$ moves. Since the total improvement is at most $2r$, where $r=|E|$, we conclude that the total number of steps is at most $5nk\cdot 2r/\epsilon = 10nkr(2n^2k^3)^{2\ell k\log(nk)}\cdot \phi= \phi\cdot r(nk)^{O(\ell k\log(nk))}$, and this occurs with probability $1-1/poly(n,k)$.

26908ba63e01b49c06b094910532561275e6d7d9

\section{Introduction} Let ${\mathcal P}({\mathbb R}^d)$ be the space of all probability measures over $({\mathbb R}^d,{\mathcal B}({\mathbb R}^d))$, which is endowed with the weak convergence topology. Consider the following distribution dependent stochastic differential equation (abbreviated as DDSDEs): \begin{align}\label{SDE} {\mathord{{\rm d}}} X_t=b_t(X_t,\mu_{X_t}){\mathord{{\rm d}}} t+\sigma_t(X_t,\mu_{X_t}){\mathord{{\rm d}}} W_t, \end{align} where $b:{\mathbb R}_+\times{\mathbb R}^d\times{\mathcal P}({\mathbb R}^d)\to{\mathbb R}^d$ and $\sigma:{\mathbb R}_+\times{\mathbb R}^d\times{\mathcal P}({\mathbb R}^d)\to{\mathbb R}^d\otimes{\mathbb R}^d$ are two Borel measurable functions, $W$ is a $d$-dimensional standard Brownian motion on some filtered probability space $(\Omega,{\mathscr F},({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P})$, and $\mu_{X_t}:={\mathbf P}\circ X^{-1}_t$ is the time marginal of $X_t$ at time $t$, By It\^o's formula, it is easy to see that $\mu_{X_t}$ satisfies the following non-linear Fokker-Planck equation (abbreviated as FPE) in the distributional sense: \begin{align}\label{Non} \partial_t\mu_{X_t}=({\mathscr L}_t^{\sigma^X})^*\mu_{X_t}+\mathord{{\rm div}}(b^X_t\mu_{X_t}), \end{align} where $\sigma^X_t(x):=\sigma_t(x,\mu_{X_t})$, $b^X_t(x):=b_t(x,\mu_{X_t})$, and $({\mathscr L}_t^{\sigma^X})^*$ is the adjoint operator of the following second order partial differential operator \begin{align}\label{ET2} {\mathscr L}^{\sigma^X}_t f(x):=\frac{1}{2}\sum_{i,j,k=1}^d(\sigma^{ik}_t\sigma^{jk}_t)(x,\mu_{X_t})\partial_i\partial_jf(x). \end{align} We note that if $$ \sigma^X_t(x)=\int_{{\mathbb R}^d}\sigma_t(x,y)\mu_{X_t}({\mathord{{\rm d}}} y),\ \ b^X_t(x)=\int_{{\mathbb R}^d}b_t(x,y)\mu_{X_t}({\mathord{{\rm d}}} y), $$ then DDSDE \eqref{SDE} is also called mean-field SDE or McKean-Vlasov SDE in the literature, which naturally appears in the studies of interacting particle systems and mean-field games (see \cite{Ka,Mc,Sz,Ca-De, Ca-Gv-Pa-Sc}, in particular, \cite{Ca-De1} and references therein). \medskip Up to now, there are numerous papers devoted to the study of this type of nonlinear FPEs and DDSDE \eqref{SDE}. In \cite{Fu}, Funaki showed the existence of martingale solutions for \eqref{SDE} under broad conditions of Lyapunov's type and also the uniqueness under global Lipschitz assumptions. His method is based on a suitable time discretization. Thus, the well-posedness of FPE \eqref{Non} is also obtained. More recently, under some one-side Lipschitz assumptions, Wang \cite{Wa} showed the strong well-posedness and some functional inequalities to DDSDE \eqref{SDE}. In \cite{Ha-Si-Sz}, Hammersley, Sitsa and Szpruch proved the existence of weak solutions to SDE \eqref{SDE} on a domain $D\subset{\mathbb R}^d$ with continuous and unbounded coefficients under Lyapunov-type conditions. Moreover, uniqueness is also obtained under some functional Lyapunov conditions. Notice that all the above results require the continuity of coefficients. In \cite{Ch}, Chiang obtained the existence of weak solutions for time-independent SDE \eqref{SDE} with drifts that have some discontinuities. When the diffusion matrix is uniformly non-degenerate and $b,\sigma$ are only measurable and of at most linear growth, by using the classical Krylov estimates, Mishura and Veretennikov \cite{Mi-Ve} showed the existence of weak solutions. The uniqueness is also proved when $\sigma$ does not depend on $\mu$ and is Lipschitz continuous in $x$ and $b$ is Lipschitz continuous with respect to $\mu$ with Lipschitz constant linearly depending on $x$. It should be noted that by Schauder's fixed point theorem and Girsanov's theorem, Li and Min \cite{Li-Mi} also obtained the existence and uniqueness of weak solutions when $b$ is bounded measurable and $\sigma$ is nondegenerate and Lipschitz continuous. On the other hand, by a purely analytic argument, Manita and Shaposhnikov \cite{Ma-Sh} and Manita, Romanov and Shaposhnikov \cite{Ma-Ro-Sh} showed the existence and uniqueness of solutions to the nonlinear FPE \eqref{Non} under quite general assumptions. As observed in \cite{Ba-Ro}, by a result of Trevisan \cite{Tr} (see Theorem \ref{Th23} below), one in fact can obtain the well-posedness of DDSDE \eqref{SDE} from \cite{Ma-Sh} and \cite{Ma-Ro-Sh}. In \cite{Ba-Ro}, a technique is developed to prove weak existence of solutions to \eqref{SDE} by first solving \eqref{Non} which works also for coefficients whose dependence on $\mu_{X_t}$ is of ``Nemytskii-type'', i.e., are not continuous in $\mu_{X_t}$ in the weak topology. \medskip In this work we are interested in extending Krylov-R\"ockner's result \cite{Kr-Ro} to the singular distribution dependent case, that is not covered by all of the above results. More precisely, we want to show the well-posedness of the following DDSDE: \begin{align}\label{SDE1} {\mathord{{\rm d}}} X_t=\left(\int_{{\mathbb R}^d}b_t(X_t,y)\mu_{X_t}({\mathord{{\rm d}}} y)\right){\mathord{{\rm d}}} t+\sqrt{2}{\mathord{{\rm d}}} W_t, \end{align} where $b:{\mathbb R}_+\times{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}^d$ is a Borel measurable function and satisfies \begin{enumerate}[{\bf (H$^b$)}] \item $|b_t(x,y)|\leqslant h_t(x-y)$ for some $h\in L^q_{loc}({\mathbb R}_+; \widetilde L^p({\mathbb R}^d))$, where $p,q\in(2,\infty)$ satisfy $\frac{d}{p}+\frac{2}{q}<1$, and $\widetilde L^p({\mathbb R}^d)$ is the localized $L^p$-space defined by \eqref{Ck} below. \end{enumerate} Here the advantage of using the localized space $\widetilde L^p({\mathbb R}^d)$ is that for any $1\leqslant p\leqslant p'\leqslant\infty$, $$ L^\infty({\mathbb R}^d)+L^{p'}({\mathbb R}^d)\subset\widetilde L^{p'}({\mathbb R}^d)\subset \widetilde L^p({\mathbb R}^d)\subset_{p>d}{\mathbb K}_{d-1}, $$ where ${\mathbb K}_{d-1}$ is the usual Kato's class defined by $$ {\mathbb K}_{d-1}:=\left\{f:\lim_{\varepsilon\to 0}\sup_{x\in{\mathbb R}^d}\int_{|x-y|\leqslant\varepsilon}|x-y|^{1-d}f(y){\mathord{{\rm d}}} y=0\right\}. $$ We note that the above DDSDE is not covered by Huang and Wang's recent results \cite{Hu-Wa} since $\mu\mapsto\int_{{\mathbb R}^d}b_t(x,y)\mu({\mathord{{\rm d}}} y)$ is not weakly continuous. In fact, if we let \begin{align}\label{BB} B_t(x,\mu):=\int_{{\mathbb R}^d}b_t(x,y)\mu({\mathord{{\rm d}}} y),\ \mu\in{\mathcal P}({\mathbb R}^d), \end{align} then by $|b_t(x,y)|\leqslant h_t(x-y)$, we only have \begin{align}\label{TTV} |\mspace{-3mu}|\mspace{-3mu}| B_t(\cdot,\mu)-B_t(\cdot,\mu')|\mspace{-3mu}|\mspace{-3mu}|_p\leqslant|\mspace{-3mu}|\mspace{-3mu}| h_t|\mspace{-3mu}|\mspace{-3mu}|_p\|\mu-\mu'\|_{TV}, \end{align} where $\|\cdot\|_{TV}$ is the total variation distance, and $|\mspace{-3mu}|\mspace{-3mu}|\cdot|\mspace{-3mu}|\mspace{-3mu}|_p$ is defined by \eqref{Ck} below. \medskip Throughout this paper we assume $d\geqslant 2$. One of the main results of this paper is stated as follows (but see also section 4 for corresponding results when the diffusion matrix $\sigma$ is non-degenerate, but not constant): \begin{theorem}\label{Th11} Under {\bf (H$^b$)}, for any $\beta>2$ and initial random variable $X_0$ with finite $\beta$-order moment, there is a unique strong solution to SDE \eqref{SDE1}. Moreover, the following assertions hold: \begin{enumerate}[{\rm (i)}] \item The time marginal law $\mu_t$ of $X_t$ uniquely solves the following nonlinear FPE in the distributional sense: \begin{align}\label{FK} \partial_t\mu_t=\Delta\mu_t+\mathord{{\rm div}}\left(\mu_t(b_t(x,\cdot))\mu_t\right),\ \ \lim_{t\downarrow 0}\mu_t({\mathord{{\rm d}}} y)={\mathbf P}\circ X^{-1}_0({\mathord{{\rm d}}} y) \end{align} in the class of all measures such that $t\mapsto \mu_t$ is weakly continuous and $$ \int^T_0\!\!\int_{{\mathbb R}^d}\!\int_{{\mathbb R}^d}|b_t(x,y)|\mu_t({\mathord{{\rm d}}} y)\mu_t({\mathord{{\rm d}}} x){\mathord{{\rm d}}} t<\infty,\ \ \forall T>0. $$ \item $\mu_t({\mathord{{\rm d}}} y)=\rho^X_t(y){\mathord{{\rm d}}} y$ and $(t,y)\mapsto\rho^X_t(y)$ is continuous on $(0,\infty)\times{\mathbb R}^d$ and satisfies the following two-sided estimate: for any $T>0$, there are constants $\gamma_0, c_0\geqslant 1$ such that for all $t\in(0,T]$ and $y\in{\mathbb R}^d$, $$ c_0^{-1}P_{t/\gamma_0}\mu_0(y)\leqslant \rho^X_t(y)\leqslant c_0P_{\gamma_0 t}\mu_0(y), $$ where $P_t\mu_0(y):=(2\pi t)^{-d/2}\int_{{\mathbb R}^d}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-|x-y|^2/(2t)}\mu_0({\mathord{{\rm d}}} x)$ is the Gaussian heat semigroup. \item If $\mathord{{\rm div}} b=0$, then for each $t>0$, $\rho^X_t(\cdot)\in C^1({\mathbb R}^d)$ and we have the following gradient estimate: for any $T>0$, there are constants $\gamma_1, c_1\geqslant 1$ such that for all $t\in(0,T]$ and $y\in{\mathbb R}^d$, $$ |\nabla\rho^X_t(y)|\leqslant c_1 t^{-1/2}P_{\gamma_1 t}\mu_0(y). $$ \end{enumerate} \end{theorem} \begin{Examples} Let $b_t(x,y):=a_t(x,y)/|x-y|^\alpha$ for some $\alpha\in[1,2)$, where $a_t(x,y):{\mathbb R}_+\times{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}^d$ satisfies that for some $\kappa>0$, $$ |a_t(x,y)|\leqslant \kappa |x-y|. $$ Then it is easy to see that $b$ satisfies {\bf (H$^b$)} for some $p>d$ and $q=\infty$. \end{Examples} \begin{remark} Here an open question is to show the following propagation of chaos (see \cite{Sz}): Given $N\in{\mathbb N}$, let $X^{N,j}, j=1,\cdots, N$ solve the following SDEs $$ {\mathord{{\rm d}}} X^{N,j}_t=\frac{1}{N}\sum_{i=1}^Nb_t(X^{N,j}_t,X^{N,i}_t){\mathord{{\rm d}}} t+\sqrt{2}{\mathord{{\rm d}}} W^j_t,\ \ j=1,\cdots, N, $$ where $W^j_\cdot, j=1,\cdots,N$ are $N$-independent $d$-dimensional Brownian motion. Let $X$ be the unique solution of SDE \eqref{SDE1} in Theorem \ref{Th11}. Is it possible to show that $$ X^{N,1}_\cdot\to X_\cdot\mbox{ in distribution as $N\to\infty$?} $$ Even for bounded measurable $b$, the above question seems to be still open. \end{remark} To show the existence of a solution to DDSDE \eqref{SDE1}, by the well-known result for bounded measurable drift $b$ obtained in \cite{Mi-Ve} (see also \cite{Li-Mi}and \cite{Zh1}), for each $n\in{\mathbb N}$, there is a solution to the following distribution dependent SDE: \begin{align}\label{SDE11} {\mathord{{\rm d}}} X^n_t=\left(\int_{{\mathbb R}^d}b^n_t(X^n_t,y)\mu_{X^{n}_t}({\mathord{{\rm d}}} y)\right){\mathord{{\rm d}}} t+\sqrt{2}{\mathord{{\rm d}}} W_t,\ \ X^n_0=X_0, \end{align} where $b^n_t(x,y):=(-n)\vee b_t(x,y)\wedge n.$ By the well-known results in \cite{Xi-Xi-Zh-Zh}, one can show the following uniform Krylov estimate: For any $p_1,q_1\in(1,\infty)$ with $\frac{d}{p_1}+\frac{2}{q_1}<2$ and $T>0$, there is a constant $C>0$ such that for any $f\in\widetilde{\mathbb L}^{p_1}_{q_1}(T)$, \begin{align}\label{Kr0} \sup_n{\mathbf E}\left(\int^T_0 f_t(X^n_t){\mathord{{\rm d}}} t\right)\leqslant C_T|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(T)}. \end{align} By this estimate and Zvonkin's technique, we can further show the tightness of $X^n_\cdot$ in the space of continuous functions. However, since $b$ is allowed to be singular, it is not obvious by taking the limit $n\to\infty$ to obtain the existence of a solution. Indeed, one needs the following Krylov estimate: for suitable $p_0,q_0\in(1,\infty)$ and any $f:{\mathbb R}_+\times{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}_+$, $$ \sup_n{\mathbf E}\left(\int^t_0f_s(X^n_s,\tilde X^{n}_s){\mathord{{\rm d}}} s\right)\leqslant |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_0}_{q_0}(T)}, $$ where $\tilde X^{n}_\cdot$ is an independent copy of $X^n$. When $b$ is bounded measurable, such an estimate is easy to get by considering $(X^n, \tilde X^{n})$ as an ${\mathbb R}^{2d}$-dimensional It\^o process and using the classical Krylov estimates (see \cite{Mi-Ve}). While for singular $b$, such simple observation fails in order to obtain best integrability index $p$. We overcome this difficulty by a simple duality argument (see Lemma \ref{Le27} below). Moreover, concerning the uniqueness, under assumption \eqref{TTV}, we shall employ Girsanov's transformation as usual. \medskip This paper is organized as follows: In Section 2, we prepare some well-known results and tools for later use. In Section 3, we show the existence of weak and strong solutions to DDSDE \eqref{SDE} when the drift satisfies {\bf (H$^b$)}, and the diffusion coefficient is uniformly nondegenerate and bounded H\"older continuous. In Section 4, we prove the uniqueness of weak and strong solutions to \eqref{SDE} in two cases: the coefficients $b$ and $\sigma$ are Lipschitz continuous in the third variable with respect to the Wasserstein metric; drift $b$ is Lipschitz continuous in the third variable with respect to the total variation distance and the diffusion coefficient does not depend on the distribution. In Section 5, we present some applications to nonlinear FPE \eqref{Non} and prove Theorem \ref{Th11}. \medskip Finally we collect some frequently used notations and conventions for later use. \begin{itemize} \item For $\theta>0$, ${\mathcal P}_\theta({\mathbb R}^d):=\left\{\mu\in{\mathcal P}({\mathbb R}^d): \int_{{\mathbb R}^d}|x|^\theta\mu({\mathord{{\rm d}}} x)<\infty\right\}$. \item For $R>0$, set $B_R:=\{x\in{\mathbb R}^d: |x|<R\}$. \item For a function $f:{\mathbb R}^d\to{\mathbb R}$, ${\mathcal M}_R f(x):=\sup_{r\in(0,R)}\frac{1}{|B_r|}\int_{B_r}|f|(x+y){\mathord{{\rm d}}} y$. \item Let ${\mathbf S}_{\rm toch}$ be the set of all measurable stochastic processes on $(\Omega,{\mathscr F},{\mathbf P})$ that are stochastically continuous. \item Let $b:{\mathbb R}_+\times{\mathbb R}^d\times{\mathcal P}({\mathbb R}^d)\to{\mathbb R}^d$ be a measurable vector field. For $X\in{\mathbf S}_{\rm toch}$, define \begin{align}\label{No1} b^X_t(x):=b_t(x,\mu_{X_t}),\ \ \mu_{X_t}:={\mathbf P}\circ X_t^{-1}. \end{align} \item For a signed measure $\mu$, we denote by $\|\mu\|_{TV}:=\sup_{\|f\|_\infty\leqslant 1}|\mu(f)|$ the total variation of $\mu$. \item For $j=1,2$, we introduce the index set ${\mathscr I}_j$ as following: \begin{align}\label{In} {\mathscr I}_j:=\Big\{(p,q)\in(1,\infty): \tfrac{d}{p}+\tfrac{2}{p}<j\Big\}. \end{align} \item For a matrix $\sigma$, we use $\|\sigma\|_{HS}$ to denote the Hilbert-Schmidt norm of $\sigma$. \item We use $A\lesssim B$ (resp. $\asymp$) to denote $A\leqslant CB$ (resp. $C^{-1}B\leqslant A\leqslant CB$) for some unimportant constant $C\geqslant 1$, whose dependence on the parameters can be traced from the context. \end{itemize} \section{Preliminaries} In this section we recall some well-known results. We first introduce the following spaces and notations for later use. For $(\alpha,p)\in{\mathbb R}_+\times(1,\infty)$, the usual Bessel potential space $H^{\alpha,p}$ is defined by $$ H^{\alpha,p}:=\big\{f\in L^1_{loc}({\mathbb R}^d): \|f\|_{\alpha,p}:=\|({\mathbb I}-\Delta)^{\alpha/2}f\|_p<\infty\big\}, $$ where $\|\cdot\|_p$ is the usual $L^p$-norm, and $({\mathbb I}-\Delta)^{\alpha/2}f$ is defined by Fourier transform $$ ({\mathbb I}-\Delta)^{\alpha/2}f:={\mathcal F}^{-1}\big((1+|\cdot|^2)^{\alpha/2}{\mathcal F} f\big). $$ Notice that for $n\in{\mathbb N}$, an equivalent norm in $H^{n,p}$ is given by $$ \|f\|_{n,p}=\|f\|_p+\|\nabla^n f\|_{p}. $$ For $T>S\geqslant 0$, $p,q\in(1,\infty)$ and $\alpha\in{\mathbb R}_+$, we introduce space-time function spaces $$ {\mathbb L}^p_q(S,T):=L^q\big([S,T];L^p\big),\ \ {\mathbb H}^{\alpha,p}_q(S,T):=L^q\big([S,T];H^{\alpha,p}\big). $$ Let $\chi\in C^\infty_c({\mathbb R}^d)$ be a smooth function with $\chi(x)=1$ for $|x|\leqslant 1$ and $\chi(x)=0$ for $|x|>2$. For $r>0$ and $z\in{\mathbb R}^d$, define \begin{align}\label{CHI} \chi^z_r(x):=\chi((x-z)/r). \end{align} Fix $r>0$. We introduce the following localized $H^{\alpha,p}$-space: \begin{align}\label{Ck} \widetilde H^{\alpha,p}:=\Big\{f\in H^{\alpha,p}_{loc}({\mathbb R}^d),|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\alpha,p}:=\sup_z\|f\chi^z_r\|_{\alpha,p}<\infty\Big\}, \end{align} and the localized space-time function space $\widetilde{\mathbb H}^{\alpha,p}_q(S,T)$ with norm \begin{align}\label{GG1} |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(S,T)}:=\sup_{z\in{\mathbb R}^d}\|\chi^z_r f\|_{{\mathbb H}^{\alpha,p}_q(S,T)}<\infty. \end{align} For simplicity we shall write $$ \widetilde{\mathbb H}^{\alpha,p}_q(T):=\widetilde{\mathbb H}^{\alpha,p}_q(0,T),\ \ \widetilde{\mathbb L}^{p}_q(T):=\widetilde{\mathbb H}^{0,p}_q(0,T), $$ and $$ \widetilde{\mathbb H}^{\alpha,p}_q:=\cap_{T>0}\widetilde{\mathbb H}^{\alpha,p}_q(T),\ \ \widetilde{\mathbb L}^{p}_q:=\cap_{T>0}\widetilde{\mathbb L}^{p}_q(T). $$ The following lemma list some easy properties of $\widetilde{\mathbb H}^{\alpha,p}_q$ (see \cite{Zh-Zh2} and \cite{Xi-Xi-Zh-Zh}). \begin{proposition}\label{Pr41} Let $p,q\in(1,\infty)$, $\alpha\in{\mathbb R}_+$ and $T>0$. \begin{enumerate}[{\rm(i)}] \item For $r\not=r'>0$, there is a $C=C(d,\alpha,r,r',p,q)\geqslant 1$ such tha \begin{align}\label{GT1} C^{-1}\sup_{z}\|f\chi^{z}_{r'}\|_{{\mathbb H}^{\alpha,p}_q(T)}\leqslant \sup_{z}\|f\chi^{z}_r\|_{{\mathbb H}^{\alpha,p}_q(T)}\leqslant C \sup_{z}\|f\chi^{z}_{r'}\|_{{\mathbb H}^{\alpha,p}_q(T)}. \end{align} In other words, the definition of $\widetilde {\mathbb H}^{\alpha,p}_q$ does not depend on the choice of $r$. \item Let $\alpha>0$ , $p,q\in[1,\infty)$ and $p'\in[p,\tfrac{pd}{d-p\alpha}{\mathbf{1}}_{p\alpha<d}+\infty\cdot{\mathbf{1}}_{p\alpha>d}]$. It holds that for some $C=C(d,\alpha,p,p')>0$, \begin{align}\label{Sob} |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p'}_q(T)}\leqslant C|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(T)}. \end{align} \item For any $k\in{\mathbb N}$, there is a constant $C=C(d,k,\alpha,p,q)\geqslant 1$ such tha $$ C^{-1} |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha+k,p}_q(T)}\leqslant |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(T)}+|\mspace{-3mu}|\mspace{-3mu}| \nabla^k f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(T)}\leqslant C |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha+k,p}_q(T)}. $$ \item Let $(\rho_\varepsilon)_{\varepsilon\in(0,1)}$ be a family of mollifiers in ${\mathbb R}^d$ and $f_\varepsilon(t,x):=f(t,\cdot)*\rho_\varepsilon(x)$. For any $f\in\widetilde {\mathbb H}^{\alpha,p}_q$, it holds that $f_\varepsilon\in L^q_{loc}({\mathbb R}; C^\infty_b({\mathbb R}^d))$ and for some $C=C(d,\alpha,p,q)>0$, \begin{align}\label{GT2} |\mspace{-3mu}|\mspace{-3mu}| f_\varepsilon|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(T)}\leqslant C|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_q(T)},\ \forall \varepsilon\in(0,1), \end{align} and for any $\varphi\in C^\infty_c({\mathbb R}^{d})$, \begin{align}\label{GT3} \lim_{\varepsilon\to0}\|(f_\varepsilon-f)\varphi\|_{{\mathbb H}^{\alpha,p}_q(T)}=0. \end{align} \item For $r=p/(p-1)$ and $s=q/(q-1)$, \begin{align} &|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\asymp |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)} =\sup_{ |\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)}\leqslant 1}\left|\int^T_0\!\!\!\int_{{\mathbb R}^d} f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t\right|,\label{KH01}\\ &\qquad\mbox{\rm and }|\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)} =\sup_{|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)} \leqslant 1}\left|\int^T_0\!\!\!\int_{{\mathbb R}^d} f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t\right|,\label{KH02} \end{align} where $|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)}:=\sup_{z\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z} f\|_{{\mathbb L}^p_q(T)}$ and $|\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)}:=\sum_{z\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z} g\|_{{\mathbb L}^{r}_{s}(T)}$, \begin{align}\label{QQZ} Q_z:=\Pi_{i=1}^d(z_i, z_i+1],\quad z=(z_1,\cdots,z_d)\in{\mathbb Z}^d. \end{align} \end{enumerate} \end{proposition} \begin{proof} The first four conclusions can be found in \cite[Proposition 4.1]{Zh-Zh2}. We only prove (v). The equivalence between $|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}$ and $ |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)}$ is obvious by definition. Concerning the others, we note that by H\"older's inequality, \begin{align}\label{KH3} \begin{split} &\int^T_0\!\!\!\int_{{\mathbb R}^d} f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t=\sum_{z\in{\mathbb Z}^d}\int^T_0\!\!\!\int_{{\mathbb R}^d}{\mathbf{1}}_{Q_z}(x) f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t\\ &\qquad\leqslant \sum_{z\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z}f\|_{{\mathbb L}^p_q(T)}\|{\mathbf{1}}_{Q_z} g\|_{{\mathbb L}^{r}_{s}(T)}\leqslant|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)}|\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)}. \end{split} \end{align} On the other hand, assume that $z_n$ is a sequence in ${\mathbb Z}^d$ so that for $Q_n:=Q_{z_n}$, \begin{align} \lim_{n\to\infty}\|{\mathbf{1}}_{Q_{n}} f\|_{{\mathbb L}^p_q(T)}=|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)}.\label{KH33} \end{align} If we take $$ g_t(x):=\frac{{\mathbf{1}}_{Q_n}(x) |f_t(x)|^{p-1}}{\|{\mathbf{1}}_{Q_n}f_t\|^{p-q}_{p}}\left(\int^T_0\|{\mathbf{1}}_{Q_n} f_t\|^{q}_{p}{\mathord{{\rm d}}} t\right)^{1/{q}-1} $$ with the convention $0/0=0$, then by easy calculations, we have $|\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)}=1$ and \begin{align*} \int^T_0\!\!\!\int_{{\mathbb R}^d}f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t=\left(\int^T_0\|{\mathbf{1}}_{Q_n}f_t\|^{q}_{p}{\mathord{{\rm d}}} t\right)^{1/{q}}=\|{\mathbf{1}}_{Q_n} f\|_{{\mathbb L}^p_q(T)}, \end{align*} which together with \eqref{KH3} and \eqref{KH33} yields \eqref{KH01}. Similarly, if we take $$ f_t(x):=\sum_{z\in{\mathbb Z}^d}\frac{{\mathbf{1}}_{Q_z}(x) |g_t(x)|^{r-1}}{\|{\mathbf{1}}_{Q_z}g_t\|^{r-s}_{r}}\cdot \left(\int^T_0\|{\mathbf{1}}_{Q_z} g_t\|^{s}_{r}{\mathord{{\rm d}}} t\right)^{1/{s}-1}, $$ then $|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|'_{\widetilde{\mathbb L}^p_q(T)}=1$ and \begin{align*} \int^T_0\!\!\!\int_{{\mathbb R}^d}f_t(x)g_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t=\sum_{z\in{\mathbb Z}^d}\left(\int^T_0\|{\mathbf{1}}_{Q_z}g_t\|^{s}_{r}{\mathord{{\rm d}}} t\right)^{1/{s}}=|\mspace{-3mu}|\mspace{-3mu}| g|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r}_{s}(T)}, \end{align*} which together with \eqref{KH3} yields \eqref{KH02}. \end{proof} We now recall the following result about $L^q(L^p)$-solvability of PDE (see \cite{Xi-Xi-Zh-Zh}). \begin{theorem}\label{pde} Let $(p,q)\in{\mathscr I}_1$ (see \eqref{In}) and $T>0$. Assume that $\sigma_t(x,\mu)=\sigma_t(x)$ and $b_t(x,\mu)=b_t(x)$ are independent of $\mu$, and satisfy that for some $c_0\geqslant 1$, $\gamma\in(0,1]$ and for all $t\geqslant 0$, $x,y,\xi\in{\mathbb R}^d$, \begin{align}\label{SI} c_0^{-1}|\xi|\leqslant |\sigma_t(x)\xi|\leqslant c_0|\xi|,\ \ \|\sigma_t(x)-\sigma_t(y)\|_{HS}\leqslant c_0|x-y|^\gamma, \end{align} and $|\mspace{-3mu}|\mspace{-3mu}| b|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}\leqslant\kappa_0$ for some $\kappa_0>0$, Then for any $\lambda\geqslant 1$ and $f\in \widetilde{\mathbb L}^p_q(T)$, there exists a unique solution $u\in \widetilde{\mathbb H}^{2,p}_q(T)$ to the following backward parabolic equation: \begin{align} \partial_tu+({\mathscr L}^\sigma_t-\lambda)u+b\cdot\nabla u=f,\quad u(T,x)=0.\label{pide55} \end{align} Moreover, letting $\Theta:=(\gamma,c_0,d,p,q,\kappa_0, T)$, we have the following: \begin{enumerate}[{\rm (i)}] \item For any $\alpha\in[0,2-\frac{2}{q})$, there is a $c_1=c_1(\alpha,\Theta)>0$ such that for all $\lambda\geqslant 1$, \begin{align}\label{Max} \lambda^{1-\frac{\alpha}{2}-\frac{1}{q}}|\mspace{-3mu}|\mspace{-3mu}| u|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_\infty(T)} +|\mspace{-3mu}|\mspace{-3mu}| u|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{2,p}_q(T)}\leqslant C|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}. \end{align} \item Let $(\sigma',b',f')$ be another set of coefficients satisfying the same assumptions as $(\sigma,b,f)$ with the same parameters $(\gamma,c_0,\kappa_0)$. Let $u'$ be the solution of \eqref{pide55} corresponding to $(\sigma',b',f')$. For any $\alpha\in[0,2-\frac{2}{q})$, there is a constant $c_2=c_2(\alpha,\Theta)>0$ such that for all $\lambda\geqslant 1$, \begin{align}\label{es10} \begin{split} &\lambda^{1-\frac{\alpha}{2}-\frac{1}{q}}|\mspace{-3mu}|\mspace{-3mu}| u-u'|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_\infty(T)}\leqslant c_2|\mspace{-3mu}|\mspace{-3mu}| f-f'|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\\ &\quad+ c_2\|f\|_{\widetilde{\mathbb L}^p_q(T)}\big(\|\sigma-\sigma'\|_{ {\mathbb L}^\infty(T)}+|\mspace{-3mu}|\mspace{-3mu}| b-b'|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}\big). \end{split} \end{align} \end{enumerate} \end{theorem} \begin{proof} The existence and uniqueness of $u\in\widetilde{\mathbb H}^{2,p}_q(T)$ as well as the first conclusion are proved in \cite[Theorem 3.1]{Xi-Xi-Zh-Zh}. We only show (ii). Let $w=u'-u$. Then $$ \partial_t w+({\mathscr L}^{\sigma'}_t-\lambda) w+b'\cdot\nabla w=({\mathscr L}^{\sigma}_t-{\mathscr L}^{\sigma'}_t)u+(b-b')\cdot\nabla u+f'-f. $$ By \eqref{Max} and H\"older's inequality we have \begin{align*} &\lambda^{1-\frac{\alpha}{2}-\frac{1}{q}}|\mspace{-3mu}|\mspace{-3mu}| w|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_\infty(T)} \lesssim |\mspace{-3mu}|\mspace{-3mu}|({\mathscr L}^{\sigma}_t-{\mathscr L}^{\sigma'}_t)u+(b-b')\cdot\nabla u+f'-f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\\ &\qquad\lesssim \|\sigma'-\sigma\|_{{\mathbb L}^\infty(T)}|\mspace{-3mu}|\mspace{-3mu}|\nabla^2u|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}+|\mspace{-3mu}|\mspace{-3mu}| b'-b|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)} \cdot\|\nabla u\|_{{\mathbb L}^\infty(T)}+|\mspace{-3mu}|\mspace{-3mu}| f'-f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}. \end{align*} Estimate \eqref{es10} now follows by Sobolev's embedding \eqref{Sob} due to $\frac{d}{p}+\frac{2}{q}<1$ and \eqref{Max}. \end{proof} \begin{remark} It should be noted that if $b$ is bounded measurable, then the assertions in Theorem \ref{pde} holds for all $p,q\in(1,\infty)$. \end{remark} The following stochastic Gronwall inequality for continuous martingales was proved by Scheutzow \cite{Sc}, and for general discontinuous martingales in \cite{Xi-Zh}. \begin{lemma}[Stochastic Gronwall's inequality]\label{im} Let $\xi(t)$ and $\eta(t)$ be two nonnegative c\`adl\`ag ${\mathscr F}_t$-adapted processes, $A_t$ a continuous nondecreasing ${\mathscr F}_t$-adapted process with $A_0=0$, $M_t$ a local martingale with $M_0=0$. Suppose that \begin{align}\label{Gron} \xi(t)\leqslant\eta(t)+\int^t_0\xi(s){\mathord{{\rm d}}} A_s+M_t,\ \forall t\geqslant 0. \end{align} Then for any $0<q<p<1$ and $\tau>0$, we have \begin{align} \big[{\mathbf E}(\xi(\tau)^*)^{q}\big]^{1/q}\leqslant \Big(\tfrac{p}{p-q}\Big)^{1/q}\Big({\mathbf E} \text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{pA_{\tau}/(1-p)}\Big)^{(1-p)/p}{\mathbf E}\big(\eta(\tau)^*\big), \label{gron} \end{align} where $\xi(t)^*:=\sup_{s\in[0,t]}\xi(s)$. \end{lemma} We also recall the following result about maximal functions (see \cite[Lemma 2.1]{Xi-Xi-Zh-Zh}). \begin{lemma}\label{Le2} \begin{enumerate}[{\rm (i)}] \item For any $R>0$, there exists a constant $C=C(d,R)$ such that for any $f\in L^\infty({\mathbb R}^d)$ with $\nabla f\in L^1_{loc}({\mathbb R}^d)$ and Lebesgue-almost all $x,y\in {\mathbb R}^d$, \begin{align} |f(x)-f(y)|\leqslant C |x-y|({\mathcal M}_R|\nabla f|(x)+{\mathcal M}_R|\nabla f|(y)+\|f\|_\infty),\label{ES2} \end{align} where ${\mathcal M}_R$ is defined at the end of the introduction. \item For any $p>1$ and $R>0$, there is a constant $C=C(R,d,p)$ such that for any $T>0$ and all $f\in\widetilde{\mathbb L}^p_q(T)$, \begin{align}\label{GW1} |\mspace{-3mu}|\mspace{-3mu}|{\mathcal M}_R f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\leqslant C|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}. \end{align} \end{enumerate} \end{lemma} We introduce the following notion about Krylov's estimates. \begin{definition} Let $p,q\in(1,\infty)$ and $T,\kappa>0$. We say a stochastic process $X\in{\mathbf S}_{\rm toch}$ satisfies Krylov's estimate with index $p,q$ and constant $\kappa$ if for any $f\in \widetilde{\mathbb L}^p_q(T)$, \begin{align}\label{Kry} {\mathbf E}\left(\int^T_0f_t(X_t){\mathord{{\rm d}}} t\right)\leqslant \kappa|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}. \end{align} The set of all such $X$ will be denoted by ${\bf K}^{p,q}_{T,\kappa}$. \end{definition} \iffalse\begin{remark} Let $X\in{\bf K}^{p,q}_{T,\kappa}$. By the interpolation theorem, for any $p'\in[p,\infty)$ and $q'\in(1,q]$ with $\tfrac{p'}{q'}-p'=\tfrac{p}{q}-p$, we have $X\in{\bf K}^{p',q'}_{T,\kappa'}$ for some $\kappa'>0$. In fact, note that by the interpolation theorem (see \cite[Theorem 5.1.2]{Be-Lo}), $$ (\widetilde{\mathbb L}^\infty_1(T), \widetilde{\mathbb L}^p_q(T))_{[\theta]}=\widetilde{\mathbb L}^{p'}_{q'}(T), $$ where $\theta\in(0,1)$, $\frac{1}{q'}=1-\theta+\frac{\theta}{q}$ and $\frac{1}{p'}=\frac{\theta}{p}$, $(\cdot,\cdot)_{[\theta]}$ stands for the complex interpolation. On the other hand, since obviously, $$ {\mathbf E}\left(\int^T_0f_t(X_t){\mathord{{\rm d}}} t\right)\leqslant|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^\infty_1(T)}. $$ By the interpolation theorem, we get $X\in{\bf K}^{p',q'}_{T,\kappa'}$. \end{remark} \f For a space-time function $f_t(x,y):{\mathbb R}_+\times{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}$ and $p_1,p_2,q_0\in[1,\infty]$, we also introduce the norm $$ |\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1,p_2}_{q_0}(T)}:=\sup_{z,z'\in{\mathbb R}^d}\left(\int^T_0\left(\int_{Q_{z'}}\|{\mathbf{1}}_{Q_z}f_t(\cdot,y)\|^{p_2}_{p_1}{\mathord{{\rm d}}} y\right)^{\frac{q_0}{p_2}}\right)^{\frac{1}{q_0}}. $$ The following lemma is an easy consequence of Proposition \ref{Pr41} (v). \begin{lemma}\label{Le27} Let $p_1,p_2, q_0,q_1,q_2\in(1,\infty)$ with $\frac{1}{q_1}+\frac{1}{q_2}=1+\frac{1}{q_0}$ and $T,\kappa_1,\kappa_2>0$. Let $X\in{\mathbf K}^{p_1,q_1}_{T,\kappa_1}$ and $Y\in{\mathbf K}^{p_2,q_2}_{T,\kappa_2}$ be two independent processes. Then for any $f_t(x,y)\in\widetilde{\mathbb L}^{p_1,p_2}_{q_0}(T)$, \begin{align}\label{GQ2} {\mathbf E}\left(\int^T_0f_t(X_t, Y_t){\mathord{{\rm d}}} t\right)&\leqslant \kappa_1\kappa_2|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1,p_2}_{q_0}(T)}. \end{align} \end{lemma} \begin{proof} Let $Z^1=X$ and $Z^2=Y$. First of all, by Krylov's estimate \eqref{Kry}, for each $i=1,2$, there is a function $\rho^{Z^i}\in{\mathbb L}^{r_i}_{s_i}(T)$ with $r_i=\frac{p_i}{p_i-1}$, $s_i=\frac{q_i}{q_i-1}$ so that $$ \int^T_0\!\!\!\int_{{\mathbb R}^d}f_t(x)\rho^{Z^i}_t(x){\mathord{{\rm d}}} x{\mathord{{\rm d}}} t={\mathbf E}\left(\int^T_0f_t(Z^i_t){\mathord{{\rm d}}} t\right)\leqslant \kappa_i|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_i}_{q_i}(T)}\leqslant \kappa_i\|f\|_{{\mathbb L}^{p_i}_{q_i}(T)}. $$ By Proposition \ref{Pr41} (v), we further have $$ |\mspace{-3mu}|\mspace{-3mu}|\rho^{Z^i}|\mspace{-3mu}|\mspace{-3mu}|^*_{\widetilde{\mathbb L}^{r_i}_{s_i}(T)}:=\sum_{z\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z}\rho^{Z^i}\|_{{\mathbb L}^{r_i}_{s_i}(T)}\leqslant \kappa_i,\ \ i=1,2, $$ where $Q_z$ is defined by \eqref{QQZ}. Now by the independence of $X,Y$ and H\"older's inequality, we have \begin{align*} &{\mathbf E}\left(\int^T_0f_t(X_t, Y_t){\mathord{{\rm d}}} t\right)=\int^T_0\!\!\!\int_{{\mathbb R}^d}\!\int_{{\mathbb R}^d} f_t(x, y)\rho^X_t(x)\rho^Y_t(y){\mathord{{\rm d}}} x{\mathord{{\rm d}}} y{\mathord{{\rm d}}} t\\ &\qquad=\sum_{z\in{\mathbb Z}^d}\sum_{z'\in{\mathbb Z}^d}\int^T_0\!\!\!\int_{{\mathbb R}^d}\!\int_{{\mathbb R}^d}{\mathbf{1}}_{Q_z}(x){\mathbf{1}}_{Q_{z'}}(y) f_t(x, y)\rho^X_t(x)\rho^Y_t(y){\mathord{{\rm d}}} x{\mathord{{\rm d}}} y{\mathord{{\rm d}}} t\\ &\qquad\leqslant\sum_{z\in{\mathbb Z}^d}\sum_{z'\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z\times Q_{z'}} f\|_{{\mathbb L}^{p_1,p_2}_{q_0}(T)}\|{\mathbf{1}}_{Q_z}\rho^X\|_{{\mathbb L}^{r_1}_{s_1}(T)}\|{\mathbf{1}}_{Q_{z'}}\rho^Y\|_{{\mathbb L}^{r_2}_{s_2}(T)}\\ &\qquad\leqslant \kappa_1\kappa_2\sup_{z, z'\in{\mathbb Z}^d}\|{\mathbf{1}}_{Q_z\times Q_{z'}} f\|_{{\mathbb L}^{p_1,p_2}_{q_0}(T)}=\kappa_1\kappa_2|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1,p_2}_{q_0}(T)}, \end{align*} which gives \eqref{GQ2}. The proof is complete. \end{proof} Now we prove the following convergence lemmas, which have independent interest and will be crucial for showing the existence of solutions in Section 3. \begin{lemma}\label{Le32} Let $X^n,Y^n, X, Y\in{\mathbf S}_{\rm toch}$ be such that for each $t\geqslant 0$, $X^n_t$ converges to $X_t$ almost surely and $Y^n_t$ converges to $Y_t$ in distribution. Let $p,q>1$ and $T,\beta,\kappa>0$. Suppose that $X^n\in{\mathbf K}^{p,q}_{T,\kappa}$ for each $n\in{\mathbb N}$, and for some $C_1>0$, \begin{align}\label{Kry1} \sup_n\sup_{t\in[0,T]}{\mathbf E}|X^n_t|^\beta\leqslant C_1. \end{align} If for each $(t,x)$, $\mu\mapsto b_t(x,\mu)$ is continuous with respect to the weak convergence topology and for some $\gamma>1$, $C_2>0$ and all $Z\in{\mathbf S}_{\rm toch}$, \begin{align}\label{JW1} |\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|_{{\mathbb L}^{\gamma p}_{\gamma q}(T)}\leqslant C_2, \end{align} where $b^Z$ is defined by \eqref{No1}, then \begin{align}\label{ER4} \lim_{n\to\infty}{\mathbf E}\left(\int^T_0|b^{Y_n}_t(X^n_t)-b^Y_t(X_t)|{\mathord{{\rm d}}} t\right)=0. \end{align} \end{lemma} \begin{proof} To prove \eqref{ER4}, it suffices to show the following: \begin{align} \lim_{n\to\infty}{\mathbf E}\left(\int^T_0\left|b^{Y^n}_t(X^n_t)-b^Y_t(X^n_t)\right|{\mathord{{\rm d}}} t\right)=0,\label{ER1}\\ \lim_{n\to\infty}{\mathbf E}\left(\int^T_0\left|b^Y_t(X^n_t)-b^Y_t(X_t)\right|{\mathord{{\rm d}}} t\right)=0.\label{ER2} \end{align} We first look at \eqref{ER1}. Since $\mu_{Y^n_t}$ weakly converges to $\mu_{Y_t}$ for each $t\geqslant 0$, by the assumption we have \begin{align}\label{LK1} b^{Y_n}_t(x)\stackrel{n\to\infty}{\to} b^Y_t(x),\ \ \forall (t,x)\in{\mathbb R}_+\times{\mathbb R}^d. \end{align} For fixed $R,M>0$, since $X^n\in{\mathbf K}^{p,q}_{T,\kappa}$ (see \eqref{Kry}), by the definitions we have \begin{align*} &{\mathbf E}\left(\int^T_0{\mathbf{1}}_{B_R}(X^n_t)\left|b^{Y_n}_t(X^n_t)-b^Y_t(X^n_t)\right|{\mathord{{\rm d}}} t\right) \leqslant \kappa|\mspace{-3mu}|\mspace{-3mu}|{\mathbf{1}}_{B_R}(b^{Y_n}-b^Y)|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\\ &\quad\lesssim\left\|{\mathbf{1}}_{B_R}(b^{Y_n}-b^Y){\mathbf{1}}_{|b^{Y_n}-b^Y|\leqslant M}\right\|_{{\mathbb L}^p_q(T)} +\left\|{\mathbf{1}}_{B_R}(b^{Y_n}-b^Y){\mathbf{1}}_{|b^{Y_n}-b^Y|>M}\right\|_{{\mathbb L}^p_q(T)}\\ &\quad\leqslant \left\|{\mathbf{1}}_{B_R}(b^{Y_n}-b^Y){\mathbf{1}}_{|b^{Y_n}-b^Y|\leqslant M}\right\|_{{\mathbb L}^p_q(T)} +\left\|{\mathbf{1}}_{B_R}|b^{Y_n}-b^Y|^\gamma\right\|_{{\mathbb L}^p_q(T)}/M^{\gamma-1}. \end{align*} By the dominated convergence theorem and \eqref{LK1}, the first term converges to zero as $n\to\infty$ for each $M>0$. By \eqref{JW1}, the second term converges to zero uniformly in $n$ as $M\to\infty$. Thus, we obtain that for any $R>0$, \begin{align}\label{Lim1} \lim_{n\to\infty}{\mathbf E}\left(\int^T_0{\mathbf{1}}_{B_R}(X^n_t)\left|b^{Y_n}_t(X^n_t)-b^Y_t(X^n_t)\right|{\mathord{{\rm d}}} t\right)=0. \end{align} On the other hand, by H\"older and Chebyshev's inequalities and \eqref{Kry1}, we have \begin{align*} &{\mathbf E}\left(\int^T_0{\mathbf{1}}_{B^c_R}(X^n_t)\left|b^{Y_n}_t(X^n_t)-b^Y_t(X^n_t)\right|{\mathord{{\rm d}}} t\right)\\ &\leqslant \int^T_0{\mathbf P}(|X^n_t|>R)^{\frac{\gamma-1}{\gamma}}\left({\mathbf E}\left|b^{Y_n}_t(X^n_t)-b^Y_t(X^n_t)\right|^\gamma\right)^{\frac{1}{\gamma}}{\mathord{{\rm d}}} t\\ &\leqslant\sup_{t\in[0,T]}{\mathbf P}(|X^n_t|>R)^{\frac{\gamma-1}{\gamma}} T^{\frac{\gamma-1}{\gamma}} \left(\int^T_0{\mathbf E}\left|b^{Y_n}_t(X^n_t)-b^Y_t(X^n_t)\right|^\gamma{\mathord{{\rm d}}} t\right)^{\frac{1}{\gamma}}\\ &\leqslant\left(\frac{C_1 T}{R^\beta}\right)^{\frac{\gamma-1}{\gamma}} \kappa^{\frac{1}{\gamma}}|\mspace{-3mu}|\mspace{-3mu}| b^{Y_n}-b^Y|\mspace{-3mu}|\mspace{-3mu}|_{{\mathbb L}^{\gamma p}_{\gamma q}(T)} \stackrel{\eqref{JW1}}{\leqslant} \left(\frac{C_1 T}{R^\beta}\right)^{\frac{\gamma-1}{\gamma}} \kappa^{\frac{1}{\gamma}}\cdot 2C_2. \end{align*} Combining this with \eqref{Lim1}, we obtain \eqref{ER1}. Next we show \eqref{ER2}. Let $b^{Y,\varepsilon}_t(x):=b^Y_t(\cdot)*\varrho_\varepsilon(x)$ be a mollifying approximation of $b^Y$. By Proposition \ref{Pr41} (iv) and \eqref{Kry1}, as above one can derive that \begin{align}\label{ET3} \lim_{\varepsilon\to 0}\sup_{n\in{\mathbb N}\cup\{\infty\}}{\mathbf E}\left(\int^T_0\left|b^{Y,\varepsilon}_t(X^n_t)-b^Y_t(X^n_t)\right|{\mathord{{\rm d}}} t\right)=0, \end{align} where we have used the convention $X^\infty:=X$. On the other hand, since by \eqref{Kry}, $$ \sup_n{\mathbf E}\left(\int^T_0\left|b^{Y,\varepsilon}_t(X^n_t)-b^{Y,\varepsilon}_t(X_t)\right|^\gamma{\mathord{{\rm d}}} t\right)\leqslant C|\mspace{-3mu}|\mspace{-3mu}| b^{Y,\varepsilon}|\mspace{-3mu}|\mspace{-3mu}|^\gamma_{\widetilde{\mathbb L}^{\gamma p}_{\gamma q}(T)}, $$ and for fixed $\varepsilon>0$ and any $t>0$, $x\mapsto b^{Y,\varepsilon}_t(x)$ is continuous, by the dominated convergence theorem, we have $$ \lim_{n\to\infty}{\mathbf E}\left(\int^T_0\left|b^{Y,\varepsilon}_t(X^n_t)-b^{Y,\varepsilon}_t(X_t)\right|{\mathord{{\rm d}}} t\right)=0, $$ which together with \eqref{ET3} yields \eqref{ER2}. \end{proof} There are, of course, many examples where the {\it weak} continuity assumption of $\mu\mapsto b_t(x,\mu)$ in the above lemma is not satisfied, as in the following interesting case: \begin{align}\label{LK2} b_t(x,\mu)=\int_{{\mathbb R}^d}{\bar b}_t(x,y)\mu({\mathord{{\rm d}}} y), \end{align} where ${\bar b}:{\mathbb R}_+\times{\mathbb R}^d\times{\mathbb R}^d\to{\mathbb R}$ is a bounded measurable function. Obviously the weak continuity of $\mu\mapsto b(t,x,\mu)$ does not hold. However, in this case we still have the following limiting result. \begin{lemma}\label{Le33} Let $X^n, Y^n, X, Y\in{\mathbf S}_{\rm toch}$ be such that for each $t\geqslant 0$, $X^n_t$ converges to $X_t$ almost surely and $Y^n_t$ converges to $Y_t$ in distribution. Let $p_1,p_2, q_0,q_1,q_2\in(1,\infty)$ with $\frac{1}{q_1}+\frac{1}{q_2}=1+\frac{1}{q_0}$ and $T,\beta,\kappa>0$. Suppose that $X^n\in{\mathbf K}^{p_1,q_1}_{T,\kappa}$ and $Y^n\in{\mathbf K}^{p_2,q_2}_{T,\kappa}$ for each $n\in{\mathbb N}$, and that there is a constant $C_1>0$ such that \begin{align}\label{LK3} \sup_n\sup_{t\in[0,T]}{\mathbf E}\left(|X^n_t|^\beta+|Y^n_t|^\beta\right)\leqslant C_1. \end{align} Let $\gamma>1$. Then for any $\bar b\in \widetilde{\mathbb L}^{\gamma p_1,\gamma p_2}_{\gamma q_0}(T)$, we have \begin{align}\label{ER44} \lim_{n\to\infty}{\mathbf E}\left(\int^T_0|b^{Y^n}_t(X^n_t)-b^{Y}_t(X_t)|{\mathord{{\rm d}}} t\right)=0. \end{align}\end{lemma} \begin{proof} Let ${\mathbb N}_\infty:={\mathbb N}\cup\{\infty\}$ and $Y^\infty:=Y, X^\infty:=X$. Since $b^{Y^n}$ only depends on the distribution of $Y^n$, by Skorohod's representation, without loss of generality we may assume that $(X^n)_{n\in{\mathbb N}_\infty}$ and $(Y^n)_{n\in{\mathbb N}_\infty}$ are independent, and $(X^n_t,Y^n_t)\to (X_t,Y_t)$ a.e. as $n\to\infty$ for each $t$. Notice that by the assumptions and \eqref{GQ2}, \begin{align}\label{KG1} \sup_{n\in{\mathbb N}_\infty}{\mathbf E}\left(\int^T_0|\bar b_t(X^n_t, Y^n_t)|^\gamma{\mathord{{\rm d}}} t\right) \leqslant \kappa^2|\mspace{-3mu}|\mspace{-3mu}| \bar b|\mspace{-3mu}|\mspace{-3mu}|^\gamma_{\widetilde{\mathbb L}^{\gamma p_1,\gamma p_2}_{\gamma q_0}(T)}<\infty. \end{align} Let $\bar b^\varepsilon_t(x,y)=\bar b_t*\varrho_\varepsilon(x,y)$ be a mollifying approximation of $\bar b$. As in the proof of \eqref{ER1}, we have \begin{align}\label{QA1} \lim_{\varepsilon\to 0}\sup_{n\in{\mathbb N}_\infty}{\mathbf E}\left(\int^T_0|\bar b^\varepsilon_t(X^n_t, Y^n_t)-\bar b_t(X^n_t, Y^n_t)|{\mathord{{\rm d}}} t\right)=0. \end{align} Thus, to prove \eqref{ER44}, it suffices to show that for fixed $\varepsilon\in(0,1)$, \begin{align*} &\lim_{n\to\infty}{\mathbf E}\left(\int^T_0|{\bar b}^\varepsilon_t(X^n_t,Y^n_t)-{\bar b}^\varepsilon_t(X^n_t,Y_t)|{\mathord{{\rm d}}} t\right)=0,\\ &\lim_{n\to\infty}{\mathbf E}\left(\int^T_0|{\bar b}^\varepsilon_t(X^n_t,Y_t)-{\bar b}^\varepsilon_t(X_t,Y_t)|{\mathord{{\rm d}}} t\right)=0, \end{align*} which follows by \eqref{KG1} and the dominated convergence theorem. \end{proof} \section{Existence of weak and strong solutions} In this section we show the weak existence and strong existence of DDSDEs with singular drifts. First of all we recall the notions of martingale solutions and weak solutions for \eqref{SDE}. Let ${\mathbb C}$ be the space of all continuous functions from ${\mathbb R}_+$ to ${\mathbb R}^d$, which is endowed with the usual Borel $\sigma$-field ${\mathcal B}({\mathbb C})$. The set of all probability measures on $({\mathbb C},{\mathcal B}({\mathbb C}))$ is denoted by ${\mathcal P}({\mathbb C})$. Let $w_t$ be the coordinate process over ${\mathbb C}$, that is, $$ w_t(\omega)=\omega_t,\ \ \omega\in{\mathbb C}. $$ For $t\geqslant 0$, let ${\mathcal B}_t({\mathbb C})=\sigma\{w_s: s\leqslant t\}$ be the natural filtration. For a probability measure ${\mathbb P}\in{\mathcal P}({\mathbb C})$, the expectation with respect to ${\mathbb P}$ will be denoted by ${\mathbb E}$ if there is no confusion. \begin{definition}[Martingale solutions]\label{Def2} We call a probability measure ${\mathbb P}\in{\mathcal P}({\mathbb C})$ a martingale solution of DDSDE \eqref{SDE} with initial distribution $\nu\in{\mathcal P}({\mathbb R}^d)$ if ${\mathbb P}\circ w_0^{-1}=\nu$ and for any $f\in C^\infty({\mathbb R}^d)$, $$ \int^{t}_0|{\mathscr L}^{\sigma^{\mathbb P}}_s f|(w_s){\mathord{{\rm d}}} s+\int^t_0|b^{\mathbb P}_s\cdot\nabla f|(w_s){\mathord{{\rm d}}} s<\infty,\ \ {\mathbb P}-a.s,\ \ \forall t>0, $$ where $\sigma^{\mathbb P}_t(x):=\sigma_t(x,\mu^{\mathbb P}_t)$ and $b^{\mathbb P}_t(x):=b_t(x,\mu^{\mathbb P}_t)$, $\mu^{\mathbb P}_t:={\mathbb P}\circ w_t^{-1}$, and \begin{align}\label{EQ1} M^f_t:=f(w_{t})-f(w_0)-\int^{t}_0({\mathscr L}^{\sigma^{\mathbb P}}_s f)(w_s){\mathord{{\rm d}}} s-\int^t_0 (b^{\mathbb P}_s\cdot\nabla f)(w_s){\mathord{{\rm d}}} s, \end{align} is a continuous local ${\mathcal B}_t({\mathbb C})$-martingale under ${\mathbb P}$. All the martingale solutions of DDSDE \eqref{SDE} with coefficients $\sigma,b$ and initial distribution $\nu$ are denoted by ${\mathscr M}^{\sigma,b}_\nu$. \end{definition} \begin{definition}[Weak solutions] \label{Def99} Let $(X,W)$ be two ${\mathbb R}^d$-valued continuous adapted processes on some filtered probability space $(\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P})$. We call $$ (\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X,W) $$ a weak solution of DDSDE \eqref{SDE} with initial distribution $\nu\in{\mathcal P}({\mathbb R}^d)$ if \begin{enumerate}[{\rm (i)}] \item ${\mathbf P}\circ X^{-1}_0=\nu$ and $W$ is a $d$-dimensional standard ${\mathscr F}_t$-Brownian motion. \item For all $t>0$, it holds that $$ \int^t_0|b_s|(X_s,\mu_{X_s}){\mathord{{\rm d}}} s+\int^t_0\|\sigma_s\sigma^*_s\|_{HS}(X_s,\mu_{X_s}){\mathord{{\rm d}}} s<\infty,\ \ {\mathbf P}-a.s. $$ and \begin{align}\label{Def9} X_t=X_0+\int^t_0b_s(X_s,\mu_{X_s}){\mathord{{\rm d}}} s+\int^t_0\sigma_s(X_s,\mu_{X_s}){\mathord{{\rm d}}} W_s,\ \ {\mathbf P}-a.s. \end{align} \end{enumerate} \end{definition} \begin{remark}\rm It is well known that weak solutions and martingale solutions are equivalent (cf. \cite{St-Va}), which means that for any ${\mathbb P}\in{\mathscr M}^{\sigma,b}_\nu$, there is a weak solution $$ (\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X,W) $$ to DDSDE \eqref{SDE} with initial distribution $\nu\in{\mathcal P}({\mathbb R}^d)$ such that $$ {\mathbb P}={\mathbf P}\circ X^{-1}. $$ \end{remark} Now we make the following assumptions about $\sigma$ and $b$: \begin{enumerate}[{\bf (H$^{\sigma,b}$)}] \item For each $t,x$, the mapping $\mu\mapsto \sigma_t(x,\mu)$ is weakly continuous, and there are $c_0\geqslant 1$ and $\gamma\in(0,1]$ such that for all $t\geqslant 0$, $x,x',\xi\in{\mathbb R}^d$ and $\mu\in{\mathcal P}({\mathbb R}^d)$, \begin{align}\label{SIG} \qquad\quad c_0^{-1}|\xi|\leqslant |\sigma_t(x,\mu)\xi|\leqslant c_0|\xi|,\ \|\sigma_t(x,\mu)-\sigma_t(x',\mu)\|_{HS}\leqslant c_0|x-x'|^\gamma. \end{align} The drift $b$ satisfies one of the following conditions: \begin{enumerate}[{\rm (i)}] \item For each $t,x$, the mapping $\mu\mapsto b_t(x,\mu)$ is weakly continuous, and for some $(p,q)\in{\mathscr I}_1$ and $\kappa_0>0$, \begin{align}\label{GR2} \sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}\leqslant\kappa_0<\infty. \end{align} \item $b$ has the form \eqref{LK2} with $\bar b$ satisfying {\bf (H$^b$)}. \end{enumerate} \end{enumerate} It should be noticed that under {\bf (H$^{b}$)}, \eqref{GR2} holds. Indeed, by definition we have \begin{align*} |\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|^q_{\widetilde{\mathbb L}^{p}_{q}(T)} &=\sup_{z\in{\mathbb R}^d}\int^T_0\left\|\chi^z_r\int_{{\mathbb R}^d}\bar b_s(\cdot,y)\mu_{Z_s}({\mathord{{\rm d}}} y)\right\|_p^q{\mathord{{\rm d}}} s\\ &\leqslant\sup_{z\in{\mathbb R}^d}\int^T_0\left\|\chi^z_r\int_{{\mathbb R}^d}h_s(\cdot-y)\mu_{Z_s}({\mathord{{\rm d}}} y)\right\|_p^q{\mathord{{\rm d}}} s\\ &\leqslant\sup_{z\in{\mathbb R}^d}\int^T_0\!\!\int_{{\mathbb R}^d}\left\|\chi^{z-y}_r h_s\right\|_p^q\mu_{Z_s}({\mathord{{\rm d}}} y){\mathord{{\rm d}}} s\\ &\leqslant\int^T_0\sup_{z\in{\mathbb R}^d}\left\|\chi^{z}_r h_s\right\|_p^q\mu_{Z_s}({\mathbb R}^d){\mathord{{\rm d}}} s=\int^T_0|\mspace{-3mu}|\mspace{-3mu}| h_s|\mspace{-3mu}|\mspace{-3mu}|_p^q{\mathord{{\rm d}}} s. \end{align*} To show the existence of weak solutions, we first establish the following apriori estimates. \begin{lemma}\label{Le35} Let $\beta>0$. Under {\bf (H$^{\sigma,b}$)}, for any $\nu\in{\mathcal P}_\beta({\mathbb R}^d)$ and $Z\in{\mathbf S}_{\rm toch}$, there is a unique weak solution $(\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X,W)$ to the following SDE: $$ {\mathord{{\rm d}}} X_t=b^Z_t(X_t){\mathord{{\rm d}}} t+\sigma^Z_t(X_t){\mathord{{\rm d}}} W_t,\ \ {\mathbf P}\circ X^{-1}_0=\nu. $$ Moreover, letting $\Theta=(d,p,q,c_0,\gamma, \kappa_0,\beta)$, we have \begin{enumerate}[{\rm (i)}] \item For any $T>0$, there is a $C_1=C_1(\Theta,T)>0$ such that \begin{align}\label{PK10} {\mathbf E}\left(\sup_{t\in[0,T]}|X_t|^\beta\right)\leqslant C_1({\mathbf E}|X_0|^{\beta}+1), \end{align} and for any $\delta<T$, \begin{align}\label{PK1} {\mathbf E}\left(\sup_{t\in[0,T-\delta]}|X_{t+\delta}-X_t|^\beta\right)\leqslant C_1\delta^{\beta/2}. \end{align} \item For any $(p_1,q_1)\in{\mathscr I}_2$ and $T>0$, there is a constant $C_2=C_2(p_1,q_1,\Theta,T)>0$ such that for all $0\leqslant t_0<t_1\leqslant T$ and $f\in\widetilde{\mathbb L}^{p_1}_{q_1}(t_0,t_1)$, \begin{align}\label{Kry2} {\mathbf E}\left(\int^{t_1}_{t_0} f_s(X_s){\mathord{{\rm d}}} s\Big|{\mathscr F}_{t_0}\right)\leqslant C_2|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(t_0,t_1)}. \end{align} \end{enumerate} \end{lemma} \begin{proof} The proof of this lemma is essentially contained in \cite{Zh-Zh}. For the reader's convenience, we sketch the proofs below. We use Zvonkin's transformation to kill the drift $b^Z$. For $\lambda, T>0$, consider the following backward PDE: $$ \partial_t u+({\mathscr L}^{\sigma^Z}_t-\lambda) u+b^Z\cdot\nabla u+b^Z=0,\ u(T,x)=0. $$ Since $b^Z\in\widetilde{\mathbb L}^{p}_{q}(T)$ with $(p,q)\in{\mathscr I}_1$, by Theorem \ref{pde}, for $\lambda\geqslant 1$, there is a unique solution $u\in\widetilde{\mathbb H}^{2,p}_{q}(T)$ solving the above PDE. Moreover, for any $\alpha\in[0,2-\frac{2}{q})$, there is a constant $c_1=c_1(\alpha,\Theta,T)>0$ such that for all $\lambda\geqslant 1$, \begin{align}\label{HF11} \lambda^{1-\frac{\alpha}{2}-\frac{2}{q}}|\mspace{-3mu}|\mspace{-3mu}| u|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_\infty(T)} +|\mspace{-3mu}|\mspace{-3mu}| u|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{2,p}_{q}(T)}\leqslant c_1|\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}. \end{align} In particular, since $\frac{d}{p}+\frac{2}{q}<1$, by \eqref{Sob} we can choose $\lambda$ large enough so that $$ \|u\|_{{\mathbb L}^{\infty}(T)}+\|\nabla u\|_{{\mathbb L}^{\infty}(T)}\leqslant 1/2. $$ Now if we define $$ \Phi_t(x):=x+u_t(x), $$ then it is easy to see that \begin{align}\label{LK6} |x-y|/2\leqslant |\Phi_t(x)-\Phi_t(y)|\leqslant 2|x-y| \end{align} and \begin{align}\label{ER5} \partial_t\Phi+{\mathscr L}^{\sigma^Z}_t\Phi+b^Z\cdot\nabla\Phi=\lambda u. \end{align} By the generalized It\^o formula and \eqref{ER5}, we have \begin{align} Y_t:=\Phi_t(X_t)&=\Phi_0(X_0)+\lambda\int^t_0u_s(X_s){\mathord{{\rm d}}} s+\int^t_0(\sigma^Z_s\cdot\nabla\Phi_s)(X_s){\mathord{{\rm d}}} W_s,\nonumber\\ &=\Phi_0(X_0)+\int^t_0\tilde b_s(Y_s){\mathord{{\rm d}}} s+\int^t_0\tilde\sigma_s(Y_s){\mathord{{\rm d}}} W_s,\label{LK7} \end{align} where $$ \tilde\sigma:=(\sigma^Z\cdot\nabla\Phi)\circ\Phi^{-1},\ \ \tilde b:=\lambda u\circ\Phi^{-1}. $$ Moreover, by \eqref{HF11}, \eqref{LK6} and the Sobolev embedding \eqref{Sob}, it is easy to see that for some $c_2=c_2(\Theta, T)>0$ and $\gamma_0=\gamma_0(\gamma,p,q)\in(0,1)$, \begin{align}\label{HF2} c_2^{-1}|\xi|\leqslant |\tilde\sigma_t(x)\xi|\leqslant c_2|\xi|,\ \ \|\tilde\sigma_t(x)-\tilde\sigma_t(y)\|_{HS}\leqslant c_2|x-y|^{\gamma_0}, \end{align} and \begin{align}\label{HF20} \|\tilde b\|_{{\mathbb L}^\infty(T)}+\|\nabla\tilde b\|_{{\mathbb L}^\infty(T)}\leqslant 4\lambda. \end{align} By well-known results, SDE \eqref{LK7} admits a unique weak solution (cf. \cite{St-Va}). Moreover, as in \cite{Zh-Zh}, one can check that $X_t:=\Phi^{-1}_t(Y_t)$ solves the original SDE. (i) Let $\beta>0$. By \eqref{HF2} and \eqref{HF20}, estimate \eqref{PK10} directly follows by BDG's inequality. We prove \eqref{PK1}. Fix $\delta\in(0,T)$. Let $\tau$ be any stopping time less than $T-\delta$. By equation \eqref{LK7} and BDG's inequality, we have \begin{align*} {\mathbf E}|Y_{\tau+\delta}-Y_\tau|^\beta&\lesssim {\mathbf E}\left|\int^{\tau+\delta}_{\tau}\tilde b_s(X_s){\mathord{{\rm d}}} s\right|^\beta +{\mathbf E}\left|\int^{\tau+\delta}_{\tau}\tilde\sigma_s(X_s){\mathord{{\rm d}}} W_s\right|^\beta\\ &\lesssim\|\tilde b\|^\beta_{{\mathbb L}^\infty(T)}\delta^\beta+\|\tilde \sigma\|^\beta_{{\mathbb L}^\infty(T)}\delta^{\beta/2}\leqslant C\delta^{\beta/2}, \end{align*} which yields \eqref{PK1} by \cite[Lemma 2.7]{Zh-Zh1} and \eqref{LK6}. (ii) It was proved in \cite[Theorem 2.1]{Zh0} (see also \cite[Theorem 5.7]{Xi-Zh}) that for any $(p_1,q_1)\in{\mathscr I}_2$, there is a constant $C_2=C_2(p_1,q_1,\Theta,T)>0$ such that for all $0\leqslant t_0<t_1\leqslant T$ and $f\in\widetilde{\mathbb L}^{p_1}_{q_1}(t_0,t_1)$, $$ {\mathbf E}\left(\int^{t_1}_{t_0} f_s(Y_s){\mathord{{\rm d}}} s\Big|{\mathscr F}_{t_0}\right)\leqslant C_2|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(t_0,t_1)}. $$ By a change of variable and \eqref{LK6} again, we obtain \eqref{Kry2}. \end{proof} \begin{remark} An important conclusion of (ii) above is the following Khasminskii's type estimate (see \cite[Lemma 3.5]{Xi-Zh}): For any $\lambda, T>0$ and $f\in\widetilde{\mathbb L}^{p_1}_{q_1}(T)$ with $(p_1,q_1)\in{\mathscr I}_2$, \begin{align}\label{Kh} {\mathbf E}\exp\left(\lambda\int^T_0|f_s(X_s)|{\mathord{{\rm d}}} s\right)\leqslant C_3, \end{align} where $C_3$ only depends on $\lambda,\Theta,p_1,q_1,T$ and $|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(T)}$. \end{remark} Now we can show the following weak existence result. \begin{theorem} Let $\beta>2$. Under {\bf (H$^{\sigma,b}$)}, for any $\nu\in{\mathcal P}_\beta({\mathbb R}^d)$, there exists a weak solution $(\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X,W)$ to DDSDE \eqref{SDE} with ${\mathbf P}\circ X^{-1}_0=\nu$. \end{theorem} \begin{proof} Let $X^0_t\equiv X_0$. For $n\in{\mathbb N}$, consider the following approximating SDE: \begin{align} X^n_t=X^n_0 +\int^t_0b^n_s(X^n_s, \mu_{X^n_s}){\mathord{{\rm d}}} s+\int^t_0\sigma_s(X^n_s, \mu_{X^n_s}){\mathord{{\rm d}}} W_s,\label{SDE0} \end{align} where $$ b^n_s(x,\mu):=(-n)\vee b_s(x,\mu)\wedge n,\ \ \bar b^n_s(x,y):=(-n)\vee \bar b_s(x,y)\wedge n. $$ Since $b^n$ is bounded measurable, by \cite{Mi-Ve} or \cite[Theorem 1.2]{Zh1}, there is a weak solution $$ (\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X^n,W) $$ to DDSDE \eqref{SDE0} with ${\mathbf P}\circ (X^n_0)^{-1}=\nu$. Moreover, since $$ \sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}| b^{n,Z}|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}\leqslant\sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}| b^{Z}|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}\leqslant\kappa_0, $$ by Lemma \ref{Le35}, the following uniform estimates hold: \begin{enumerate}[{\rm (i)}] \item For any $T>0$, there is a constant $C_1>0$ such that $$ \sup_n{\mathbf E}\left(\sup_{t\in[0,T]}|X^n_t|^\beta\right)\leqslant C_1({\mathbf E}|X_0|^{\beta}+1), $$ and for all $\delta\in(0,T)$, $$ \sup_n{\mathbf E}\left(\sup_{t\in[0,T-\delta]}|X^n_{t+\delta}-X^n_t|^\beta\right)\leqslant C_1\delta^{\beta/2}. $$ \item Let $(p_1,q_1)\in{\mathscr I}_2$. For any $T>0,$ there is a $C_2>0$ such that for all $f\in\widetilde{\mathbb L}^{p_1}_{q_1}(T)$, $$ \sup_n{\mathbf E}\left(\int^T_0 f_s(X^n_s){\mathord{{\rm d}}} s\right)\leqslant C_2|\mspace{-3mu}|\mspace{-3mu}| f|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(T)}. $$ \end{enumerate} Now by (i), the laws ${\mathbb Q}^n$ of $(X^n, W)$ in ${\mathbb C}\times{\mathbb C}$ are tight. Let ${\mathbb Q}$ be any accumulation point of ${\mathbb Q}^n$. Without loss of generality, we assume that ${\mathbb Q}^n$ weakly converges to some probability measure ${\mathbb Q}$. By Skorokhod's representation theorem, there are a probability space $(\tilde\Omega,\tilde{\mathscr F},\tilde{\mathbf P})$ and random variables $(\tilde X^n , \tilde W^n)$ and $(\tilde X,\tilde W)$ defined on it such that \begin{align}\label{FD4} (\tilde X^n , \tilde W^n)\to (\tilde X,\tilde W),\ \ \tilde{\mathbf P}-a.s. \end{align} and \begin{align}\label{FD5} \tilde{\mathbf P}\circ(\tilde X^n , \tilde W^n)^{-1}={\mathbb Q}^n={\mathbf P}\circ(X^n, W)^{-1},\quad \tilde{\mathbf P}\circ(\tilde X, \tilde W)^{-1}={\mathbb Q}. \end{align} Define $\tilde{\mathscr F}^n _t:=\sigma(\tilde W^n _s, \tilde X^n_s;s\leqslant t)$. We note that \begin{align*} &{\mathbf P}(W _t-W _s\in\cdot |{\mathscr F} _s)={\mathbf P}(W _t-W _s\in\cdot)\\ &\Rightarrow \tilde{\mathbf P}(\tilde W^n _t-\tilde W^n _s\in\cdot |\tilde {\mathscr F}^n _s)=\tilde{\mathbf P}(\tilde W^n _t-\tilde W^n _s\in\cdot). \end{align*} In other words, $\tilde W^n $ is an $\tilde{\mathscr F}_t^n $-Brownian motion. Thus, by \eqref{SDE0} and \eqref{FD5} we have $$ \tilde X^n _t=\tilde X^n_0 +\int^t_0b^n_s(\tilde X^n_s,\mu_{\tilde X^n_s}){\mathord{{\rm d}}} s+\int^t_0\sigma_s(\tilde X^n _s,\mu_{\tilde X^n_s}){\mathord{{\rm d}}} \tilde W^n _s. $$ By (ii), \eqref{FD4}, Lemmas \ref{Le32}, \ref{Le33} and \cite[Theorem 6.22, p383]{Ja-Sh}, one can take limits as $n\to\infty$ to obtain $$ \tilde X_t=\tilde X_0+\int^t_0b_s(\tilde X_s,\mu_{\tilde X_s}){\mathord{{\rm d}}} s+\int^t_0\sigma_s(\tilde X_s,\mu_{\tilde X_s}){\mathord{{\rm d}}} \tilde W_s. $$ Here we only check that the assumptions of Lemma \ref{Le33} are satisfied in the case that $b$ takes the form \eqref{LK2} with $\bar b$ satisfying {\bf (H$^b$)}. Clearly, by (ii) above, for any $(p_1,q_1)\in{\mathscr I}_2$, there is a $\kappa>0$ such that for each $n\in{\mathbb N}$, $$ \tilde X^n\in{\mathbf K}^{p_1,q_1}_{T,\kappa}. $$ We note that $|\bar b_t(x,y)|\leqslant h_t(x-y)$, where for some $(p,q)\in{\mathscr I}_1$, $h\in L^q_{loc}({\mathbb R}_+;\widetilde L^p({\mathbb R}^d))\subset\widetilde{\mathbb L}^p_q.$ One can choose $\gamma>1$ so that $\tfrac{d\gamma }{p}+\tfrac{\gamma}{q}<1.$ Now if we take $p_1=p_2=\frac{p}{\gamma}$, $q_0=\frac{q}{\gamma}$, $q_1=q_2=\frac{2q}{q+\gamma}$, then it is easy to see that $(p_1,q_1)\in{\mathscr I}_2$ and $$ \bar b\in\widetilde{\mathbb L}^{p,\infty}_{q}=\widetilde{\mathbb L}^{\gamma p_1,\infty}_{\gamma q_0}\subset\cap_{p'\geqslant 1}\widetilde{\mathbb L}^{\gamma p_1,p'}_{\gamma q_0}. $$ Thus one can apply Lemma \ref{Le33} to conclude that $$ \lim_{n\to\infty}\tilde{\mathbf E}\left(\int^t_0|b_s(\tilde X^n_s,\mu_{\tilde X^n_s})-b_s(\tilde X_s,\mu_{\tilde X_s})|{\mathord{{\rm d}}} s\right)=0. $$ Moreover, as in showing \eqref{QA1}, we also have $$ \lim_{m\to\infty}\sup_n\tilde{\mathbf E}\left(\int^t_0|\bar b^m_s-\bar b_s|(\tilde X^n_s,Y^n_s){\mathord{{\rm d}}} s\right)=0, $$ where $Y^n_\cdot$ is an independent copy of $\tilde X^n_\cdot$. The proof is thus complete. \end{proof} About the existence of strong solutions, we have \begin{corollary}\label{Co38} Let $\beta>2$. Under {\bf (H$^{\sigma,b}$)}, if for some $(p_1,q_1)\in{\mathscr I}_1$, $$\sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}|\nabla\sigma^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{q_1}_{p_1}(T)}<\infty,$$ then for any initial random variable $X_0$ with finite $\beta$-order moment, there exists a strong solution to DDSDE \eqref{SDE}. \end{corollary} \begin{proof} Let $(\Omega,{\mathscr F}, ({\mathscr F}_t)_{t\geqslant 0}, {\mathbf P}; X,W)$ be a weak solution of DDSDE \eqref{SDE}. Define $$ b^X_t(x):=b_t(x,\mu_{X_t}),\ \ \sigma^X_t(x):=\sigma_t(x,\mu_{X_t}),\ \mu_{X_t}:={\mathbf P}\circ X^{-1}_t. $$ Consider the following SDE: $$ {\mathord{{\rm d}}} Z_t=b^X_t(Z_t){\mathord{{\rm d}}} t+\sigma^X_t(Z_t){\mathord{{\rm d}}} W_t. $$ Under the assumption of the theorem, it has been shown in \cite{Xi-Xi-Zh-Zh} that there is a unique strong solution to this equation. Since $X$ also satisfies the above equation, by strong uniqueness, we obtain that $X=Z$ is a strong solution. \end{proof} \begin{remark} Although we have shown the existence of strong or weak solutions, the uniqueness of strong solutions or weak solutions is a more difficult problem. \end{remark} \section{Uniqueness of strong and weak solutions} In this section we study the uniqueness of strong and weak solutions. We introduce the following assumptions about the dependence on third variable $\mu$: \begin{enumerate}[\bf (A$^{\sigma,b}_\theta$)] \item Let $(p,q), (p_1,q_1)\in{\mathscr I}_1$ and $\theta\geqslant 1$. It holds that $$ \sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_q(T)}<\infty,\ \ \sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}|\nabla\sigma^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(T)}<\infty, $$ and there are $\ell\in L^q_{loc}({\mathbb R}_+)$ and a constant $c_0\geqslant 1$ such that for any two random variables $X,Y$ with finite $\theta$-order moments, \begin{align}\label{DG2} \begin{split} &|\mspace{-3mu}|\mspace{-3mu}| b_t(\cdot,\mu_X)-b_t(\cdot,\mu_Y)|\mspace{-3mu}|\mspace{-3mu}|_p\leqslant \ell_t\|X-Y\|_{\theta},\ \\ &\|\sigma_t(\cdot,\mu_X)-\sigma_t(\cdot,\mu_Y)\|_\infty\leqslant c_0\|X-Y\|_{\theta}, \end{split} \end{align} where $\|\cdot\|_\theta$ stands for the $L^\theta$-norm in the probability space $(\Omega,{\mathscr F},{\mathbf P})$. \end{enumerate} Notice that \eqref{DG2} is equivalent to that for all $\mu,\mu'\in{\mathcal P}_\theta({\mathbb R}^d)$, \begin{align*} \begin{split} &\|b_t(\cdot,\mu)-b_t(\cdot,\mu')\|_p\leqslant \ell_t{\mathcal W}_\theta(\mu,\mu'),\\ &\|\sigma_t(\cdot,\mu)-\sigma_t(\cdot,\mu')\|_\infty\leqslant c_0{\mathcal W}_\theta(\mu,\mu'), \end{split} \end{align*} where ${\mathcal W}_\theta$ is the usual Wasserstein metric of $\theta$-order. For convenience, we would like to use \eqref{DG2} rather than introducing the Wasserstein metric. \begin{remark} We note that in \cite{Hu-Wa}, \eqref{DG2} is assumed to hold for $p=\infty$. \end{remark} We first show the following strong uniqueness result. \begin{theorem}\label{Th51} Let $\theta\geqslant 1$ and $\beta>2\vee\theta$. Under {\bf (H$^{\sigma,b}$)} and {\bf (A$^{\sigma,b}_\theta$)}, for any initial random variable $X_0$ with finite $\beta$-order moment, there is a unique strong solution to DDSDE \eqref{SDE}. \end{theorem} \begin{proof} Below we fix $p,q\in{\mathscr I}_1$, and without loss of generality, we consider the time interval $[0,1]$ and assume that for some $\gamma>1$, \begin{align}\label{GG} \|\ell\|_{L^{\gamma q}(0,1)}+\sup_{Z\in{\mathbf S}_{\rm toch}}\|b^Z\|_{{\mathbb L}^{\gamma q}_{\gamma p}(1)}<\infty. \end{align} Otherwise, we may choose $\gamma>1$ so that $\frac{2\gamma}{q}+\frac{d\gamma}{p}<1$ holds and replace $(p,q)$ with $(p/\gamma,q/\gamma)$. The existence of strong solutions has been shown in Corollary \ref{Co38}. We only need to prove the pathwise uniqueness. Let $X,Y$ be two strong solutions defined on the same probability space with same starting points $X_0=Y_0$ a.s. We divide the proof into three steps and use the convention that all the constants below will be independent of $T\in[0,1]$. (i) Let $T\in(0,1)$ and $\lambda>0$. We consider the following backward PDE: \begin{align}\label{ET4} \partial_t u^X+({\mathscr L}^{\sigma^X}_t-\lambda) u+b^X\cdot\nabla u^X+b^X=0,\ u^X(T,x)=0. \end{align} By Theorem \ref{pde}, for $\lambda\geqslant 1$, there is a unique solution $u^X\in\widetilde{\mathbb H}^{2,p}_{q}(T)$ solving the above PDE. Moreover, for any $\alpha\in[0,2-\frac{2}{q})$, there is a constant $c_1>0$ such that for all $\lambda\geqslant 1$ and $T\in[0,1]$, \begin{align}\label{HF1} \lambda^{1-\frac{\alpha}{2}-\frac{2}{q}}|\mspace{-3mu}|\mspace{-3mu}| u^X|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{\alpha,p}_\infty(T)} +|\mspace{-3mu}|\mspace{-3mu}| u^X|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb H}^{2,p}_{q}(T)}\leqslant c_1|\mspace{-3mu}|\mspace{-3mu}| b^X|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_{q}(T)}. \end{align} In particular, since $\frac{d}{p}+\frac{2}{q}<1$, by \eqref{Sob}, we can choose $\lambda$ large enough so that \begin{align}\label{ER8} \|u^X\|_{{\mathbb L}^{\infty}(T)}+\|\nabla u^X\|_{{\mathbb L}^{\infty}(T)}\leqslant 1/2,\ \ \forall T\in[0,1]. \end{align} Below we shall fix such a $\lambda$ and define $$ \Phi^X_t(x):=x+u^X_t(x). $$ It is easy to see that $$ \partial_t\Phi^X+{\mathscr L}^{\sigma^X}_t\Phi^X+b^X\cdot\nabla\Phi^X=\lambda u^X. $$ (ii) By the generalized It\^o formula, we have \begin{align}\label{KJ1} \tilde X_t:=\Phi^X_t(X_t)=\Phi^X_0(X_0)+\lambda\int^t_0u^X_s(X_s){\mathord{{\rm d}}} s+\int^t_0\tilde\sigma^X_s(X_s){\mathord{{\rm d}}} W_s, \end{align} where $$ \tilde\sigma^X:=\sigma^X\cdot\nabla\Phi^X. $$ Similarly, we define $\tilde Y_t:=\Phi^Y_t(Y_t)$, and for simplicity write $$ \xi_t:=X_t-Y_t,\ \ \tilde \xi_t:=\tilde X_t-\tilde Y_t. $$ Noting that by \eqref{ER8}, $$ |x-y|\leqslant 2|\Phi^X_t(x)-\Phi^X_t(y)|\leqslant 2|\Phi^X_t(x)-\Phi^Y_t(y)|+2\|u^X-u^Y\|_{{\mathbb L}^\infty(T)} $$ and $$ |\Phi^X_t(x)-\Phi^Y_t(y)|\leqslant 2|x-y|+\|u^X-u^Y\|_{{\mathbb L}^\infty(T)}, $$ we have \begin{align}\label{KJ2} |\xi_t|\leqslant 2|\tilde\xi_t|+2\|u^X-u^Y\|_{{\mathbb L}^\infty(T)},\ \ |\tilde\xi_t|\leqslant 2|\xi_t|+\|u^X-u^Y\|_{{\mathbb L}^\infty(T)}. \end{align} By \eqref{KJ1} and again It\^o's formula, we have for any $\beta\geqslant 1$, \begin{align*} |\tilde\xi_t|^{\beta}&=|\tilde\xi_0|^{\beta}+\beta\lambda\int^t_0|\tilde\xi_s|^{\beta-2}\<\tilde\xi_s,u^X_s(X_s)-u^Y_s(Y_s)\>{\mathord{{\rm d}}} s\\ &\quad+\beta\int^t_0|\tilde\xi_s|^{\beta-2}\<(\tilde\sigma^X_s(X_s)-\tilde\sigma^Y_s(Y_s))^*\tilde\xi_s, {\mathord{{\rm d}}} W_s\>\\ &\quad+\beta\Big(\tfrac{\beta}{2}-1\Big)\int^t_0|\tilde\xi_s|^{\beta-4}|(\tilde\sigma^X_s(X_s)-\tilde\sigma^Y_s(Y_s))^*\tilde\xi_s|^2{\mathord{{\rm d}}} s\\ &\quad+\frac{\beta}{2}\int^t_0|\tilde\xi_s|^{\beta-2}\|\tilde\sigma^X_s(X_s)-\tilde\sigma^Y_s(Y_s)\|^2_{HS}{\mathord{{\rm d}}} s\\ &:=I_1+I_2+I_3+I_4+I_5. \end{align*} Since by \eqref{ER8}, \begin{align*} |u^X_t(x)-u^Y_t(y)|\leqslant |x-y|+\|u^X-u^Y\|_{{\mathbb L}^\infty(T)}, \end{align*} by Young's inequality we obtain \begin{align*} I_2&\lesssim \int^t_0|\tilde\xi_s|^{\beta}{\mathord{{\rm d}}} s+\lambda\int^t_0|u^X_s(X_s)-u^Y_s(Y_s)|^\beta{\mathord{{\rm d}}} s\\ &\lesssim \int^t_0(|\tilde\xi_s|^{\beta}+\lambda |\xi_s|^{\beta}){\mathord{{\rm d}}} s +\lambda^{\beta}T\|u^X-u^Y\|_{{\mathbb L}^\infty(T)}^{\beta}. \end{align*} Let $$ g^X_s(x):=|\nabla^2 u^X_s(x)|+|\nabla\sigma^X_s(x)|+\|\nabla u^X\|_{{\mathbb L}^\infty(T)}+\|\sigma^X\|_{{\mathbb L}^\infty(T)}. $$ By the definition of $\tilde\sigma^X$, we also have that \begin{align*} &|\tilde\sigma^X_s(x)-\tilde\sigma^Y_s(y)|\\ &\leqslant \|\sigma^Y\|_{{\mathbb L}^\infty(T)}|\nabla\Phi^X_s(x)-\nabla\Phi^Y_s(y)| +|\sigma^X_s(x)-\sigma^Y_s(y)|\cdot\|\nabla\Phi^X\|_{{\mathbb L}^\infty(T)}\\ &\leqslant \|\sigma^Y\|_{{\mathbb L}^\infty(T)}\Big(|\nabla u^X_s(x)-\nabla u^X_s(y)|+|\nabla u^X_s(y)-\nabla u^Y(s,y)|\Big)\\ &\quad+\Big(|\sigma^X_s(x)-\sigma^X_s(y)|+|\sigma^X_s(y)-\sigma^Y_s(y)|\Big)\cdot\|\nabla\Phi^X\|_{{\mathbb L}^\infty(T)}\\ &\stackrel{\eqref{ES2}}{\lesssim} |x-y|\Big({\mathcal M}_1 g^X_s(x)+{\mathcal M}_1 g^X_s(y)\Big)+\|\nabla u^X-\nabla u^Y\|_{{\mathbb L}^\infty(T)}+ \|\sigma^X_s-\sigma^Y_s\|_\infty. \end{align*} Hence, \begin{align*} I_4+I_5&\lesssim \int^t_0\Big(|\xi_s|^\beta+|\tilde\xi_s|^\beta\Big)\Big({\mathcal M} g^X_s(X_s)+{\mathcal M} g^X_s(Y_s)\Big)^2{\mathord{{\rm d}}} s\\ &\quad+T\|\nabla u^X-\nabla u^Y\|^{\beta}_{{\mathbb L}^\infty(T)}+\int^t_0\|\sigma^X_s-\sigma^Y_s\|_\infty^{\beta}{\mathord{{\rm d}}} s. \end{align*} Combining the above calculations and noting that $|\tilde\xi_0|\leqslant \|u^X_0-u^Y_0\|_\infty$, we obtain \begin{align}\label{KH1} \begin{split} |\tilde\xi_t|^\beta&\lesssim\|u^X-u^Y\|^\beta_{{\mathbb H}^{1,\infty}_\infty(T)} +\int^t_0\Big(|\tilde\xi_s|^\beta+|\xi_s|^\beta+\|\xi_s\|_\theta^\beta\Big){\mathord{{\rm d}}} s\\ &\quad+\int^t_0\Big(|\xi_s|^\beta+|\tilde\xi_s|^\beta\Big)\Big({\mathcal M}_1 g^X_s(X_s)+{\mathcal M}_1 g^X_s(Y_s)\Big)^2{\mathord{{\rm d}}} s+M_t, \end{split} \end{align} where $M_t$ is a continuous local martingale. (iii) Now we define $$ A_t:=t+\int^t_0\Big({\mathcal M}_1 g^X_s(X_s)+{\mathcal M}_1 g^X_s(Y_s)\Big)^2{\mathord{{\rm d}}} s. $$ By \eqref{KH1} and \eqref{KJ2}, we obtain that for all $t\in[0,T]$, \begin{align*} |\xi_s|^\beta+|\tilde\xi_s|^\beta\lesssim \|u^X-u^Y\|^\beta_{{\mathbb H}^{1,\infty}_\infty(T)}+\int^t_0\|\xi_s\|_\theta^\beta{\mathord{{\rm d}}} s +\int^t_0\Big(|\xi_s|^\beta+|\tilde\xi_s|^\beta\Big){\mathord{{\rm d}}} A_s+M_t. \end{align*} Note that by the assumption and \eqref{GW1}, $$ (s,x)\mapsto ({\mathcal M}_1 |\nabla^2 u^X_s(x)|)^2\in \widetilde{\mathbb L}^{p/2}_{q/2}(T), $$ and $$ (s,x)\mapsto ({\mathcal M}_1 |\nabla\sigma^X_s(x)|)^2\in \widetilde{\mathbb L}^{p_1/2}_{q_1/2}(T). $$ Since $(\frac{p}{2},\frac{q}{2}), (\frac{p_1}{2},\frac{q_1}{2})\in{\mathscr I}_2$, by Khasminskii's estimate \eqref{Kh}, we have $$ {\mathbb E}\exp{\gamma A_T}<\infty,\ \ \forall \gamma>0,\ \ \forall T\in[0,1]. $$ Thus we can use the stochastic Gronwall inequality \eqref{gron} to derive that \begin{align}\label{ET6} \sup_{s\in[0,T]}\|\xi_s\|_\theta^\beta=\left(\sup_{s\in[0,T]}{\mathbf E} |\xi_s|^{\theta}\right)^{\beta/\theta}&\lesssim \|u^X-u^Y\|^\beta_{{\mathbb H}^{1,\infty}_\infty(T)}+\int^T_0\|\xi_s\|_\theta^\beta{\mathord{{\rm d}}} s. \end{align} Noticing that by \eqref{DG2}, $$ |\mspace{-3mu}|\mspace{-3mu}| b^X-b^Y|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\leqslant \left(\int^T_0\ell^q_t\|X_t-Y_t\|^q_\theta{\mathord{{\rm d}}} t\right)^{1/q} \leqslant\|\ell\|_{L^q(0,T)}\sup_{t\in[0,T]}\|\xi_t\|_\theta, $$ and $$ \|\sigma^X-\sigma^Y\|_{{\mathbb L}^\infty(T)}\leqslant c_0\sup_{t\in[0,T]}\|X_t-Y_t\|_\theta=c_0\sup_{t\in[0,T]}\|\xi_t\|_\theta, $$ we have by \eqref{es10}, \begin{align*} &\|u^X-u^Y\|_{{\mathbb H}^{1,\infty}_\infty(T)}\\ &\lesssim |\mspace{-3mu}|\mspace{-3mu}| b^X-b^Y|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}+|\mspace{-3mu}|\mspace{-3mu}| b^X|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\left(\|\sigma^X-\sigma^Y\|_{{\mathbb L}^\infty(T)} +|\mspace{-3mu}|\mspace{-3mu}| b^X-b^Y|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\right)\\ &\lesssim \left(\|\ell\|_{L^q(0,T)}+|\mspace{-3mu}|\mspace{-3mu}| b^X|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\right)\sup_{t\in[0,T]}\|\xi_t\|_\theta \stackrel{\eqref{GG}}{\lesssim}T^{\frac{\gamma-1}{\gamma q}}\sup_{t\in[0,T]}\|\xi_t\|_\theta. \end{align*} Substituting this into \eqref{ET6}, we obtain $$ \sup_{s\in[0,T]}\|\xi_s\|_\theta^\beta\leqslant CT^{\frac{\beta(\gamma-1)}{\gamma q}}\sup_{t\in[0,T]}\|\xi_t\|^\beta_\theta,\ \ T\in(0,1), $$ where $C$ does not depend on $T\in(0,1)$. By choosing $T$ small enough, we get $\|\xi_t\|^\beta_\theta=0$ for all $t\in[0,T]$. By shifting the time $T$, we obtain the uniqueness. \end{proof} It is obvious that $b$ defined in \eqref{LK2} does not satisfy \eqref{DG2}. Below we shall relax it to the weighted total variation norm by Girsanov's transformation. The price we have to pay is that we need to assume that the diffusion coefficient does not depend on the time marginal law of $X$. For $\theta\geqslant 1$, let $$ \phi_\theta(x):=1+|x|^\theta. $$ We assume \begin{enumerate}[{\bf ($\widetilde{\bf A}^{\sigma,b}_\theta$)}] \item Let $(p,q), (p_1,q_1)\in{\mathscr I}_1$ and $\theta\geqslant 1$ and $\sigma_t(x,\mu)=\sigma_t(x)$. It holds that $$ \sup_{Z\in{\mathbf S}_{\rm toch}}|\mspace{-3mu}|\mspace{-3mu}| b^Z|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p}_q(T)}<\infty,\ \ |\mspace{-3mu}|\mspace{-3mu}|\nabla\sigma|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^{p_1}_{q_1}(T)}<\infty, $$ and there is an $\ell\in L^q_{loc}({\mathbb R}_+)$ such that for all $\mu,\mu'\in{\mathcal P}({\mathbb R}^d)$ and $t\geqslant 0$, \begin{align}\label{CC1} \begin{split} |\mspace{-3mu}|\mspace{-3mu}| b(t,\cdot,\mu)-b(t,\cdot,\mu')|\mspace{-3mu}|\mspace{-3mu}|_p\leqslant \ell_t\|\phi_\theta\cdot (\mu-\mu')\|_{TV}. \end{split} \end{align} \end{enumerate} It should be noted that \cite[Theorem 6.15]{Vi} implies, $$ {\mathcal W}_\theta(\mu,\mu')\leqslant c\|\phi_\theta\cdot (\mu-\mu')\|^{1/\theta}_{TV}. $$ \begin{theorem}\label{Th43} Let $\theta\geqslant 1$ and $\beta>2\theta$. Under {\bf (H$^{\sigma,b}$)} and {\bf ($\widetilde{\bf A}^{\sigma,b}_\theta$)}, for any initial random variable $X_0$ with finite $\beta$-order moment, there is a unique weak solution to DDSDE \eqref{SDE}, which is also a unique strong solution. \end{theorem} \begin{proof} We use the Girsanov transform in the same way asin \cite{Mi-Ve} to show the weak uniqueness, and so also the strong uniqueness. Since under the assumptions of the theorem, weak solutions are also strong solutions (see Corollary \ref{Co38}), without loss of generality, let $X^{(i)}, i=1,2$ be two solutions of SDE \eqref{SDE} defined on the same probability space $(\Omega,{\mathscr F},{\mathbf P})$ and with the same Brownian motion and starting point $\xi$. That is, \begin{align}\label{EQ0} {\mathord{{\rm d}}} X^{(i)}_t=\sigma_t(X^{(i)}_t){\mathord{{\rm d}}} W_t+b_t(X^{(i)}_t,\mu^{(i)}_t){\mathord{{\rm d}}} t,\ \ X^{(i)}_0=\xi, \end{align} where $\mu^{(i)}_t={\mathbf P}\circ (X^{(i)}_t)^{-1}$. We want to show $\mu^{(1)}_t=\mu^{(2)}_t$. Since $\sigma_t(x,\mu)=\sigma_t(x)$ satisfies \eqref{SI} under our assumptions, it is well known that there is a unique weak solution to SDE $$ {\mathord{{\rm d}}} Z_t=\sigma_t(Z_t){\mathord{{\rm d}}} W_t,\ \ Z_0=\xi. $$ Let $\beta>2\theta$. Since $\sigma$ is bounded, it is easy to see that \begin{align}\label{Mo1} \sup_{t\in[0,T]}{\mathbf E}|Z_t|^\beta\leqslant C\Big({\mathbf E}|\xi|^\beta+1\Big). \end{align} Define $$ \tilde b^{(i)}_s(x):=\sigma^{-1}_s(x)\cdot b^{X^{(i)}}_s(x),\ \ \tilde W^{(i)}_t:=W_t-\int^t_0\tilde b^{(i)}_s(Z_s){\mathord{{\rm d}}} s, $$ and $$ {\mathscr E}^{(i)}_T:=\exp\left\{\int^T_0\tilde b^{(i)}_s(Z_s)\cdot{\mathord{{\rm d}}} W_s-\frac{1}{2}\int^T_0|\tilde b^{(i)}_s(Z_s)|^2{\mathord{{\rm d}}} s\right\}. $$ Since $|\mspace{-3mu}|\mspace{-3mu}|\tilde b^{(i)}|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}\leqslant|\mspace{-3mu}|\mspace{-3mu}| b^{X^{(i)}}|\mspace{-3mu}|\mspace{-3mu}|_{\widetilde{\mathbb L}^p_q(T)}<\infty$ for some $(p,q)\in{\mathscr I}_1$, by Khasminskii's estimate \eqref{Kh}, we have \begin{align}\label{ET5} {\mathbf E}\exp\left\{\gamma\int^T_0|\tilde b^{(i)}_s(Z_s)|^2{\mathord{{\rm d}}} s\right\}\leqslant C_{T,\gamma},\ \ \forall \gamma>0, \end{align} and for any $\gamma\in{\mathbb R}$, \begin{align}\label{ET50} {\mathbf E} ({\mathscr E}^{(i)}_T)^\gamma\leqslant C_{T,\gamma}<\infty. \end{align} Hence, for each $i=1,2$, ${\mathbf E}{\mathscr E}^{(i)}_T=1$, and $\tilde W^{(i)}$ is still a Brownian motion under ${\mathscr E}^{(i)}_T\cdot{\mathbf P}$, and $$ {\mathord{{\rm d}}} Z_t=\sigma_t(Z_t){\mathord{{\rm d}}}\tilde W^{(i)}_t+b^{X^{(i)}}_t(Z_t){\mathord{{\rm d}}} t,\ \ Z_0=\xi. $$ Since the above SDE admits a unique strong solution (see also \eqref{EQ0}), we have $$ ({\mathscr E}^{(i)}_T{\mathbf P})\circ Z_{T}^{-1}={\mathbf P}\circ (X^{(i)}_{T})^{-1}=\mu^{(i)}_T, \ \ i=1,2. $$ Therefore, for $\delta=\frac{\beta}{\beta-\theta}<2$, by H\"older's inequality, we get \begin{align} &\|\phi_\theta\cdot(\mu^{(1)}_T-\mu^{(2)}_T)\|_{TV}=\|\phi_\theta\cdot (({\mathscr E}^{(1)}_T{\mathbf P})\circ Z_{T}^{-1}-({\mathscr E}^{(2)}_T{\mathbf P})\circ Z_{T}^{-1})\|_{TV}\nonumber\\ &\qquad\leqslant{\mathbf E}\Big(\phi_\theta(Z_T)|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T|\Big)\leqslant\|\phi_\theta(Z_T)\|_{\delta/(\delta-1)}\|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T\|_{\delta}\nonumber\\ &\qquad=\|1+|Z_T|^\theta\|_{\beta/\theta}\|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T\|_{\delta} \stackrel{\eqref{Mo1}}{\leqslant} C\|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T\|_{\delta}.\label{LK9} \end{align} Noting that $$ {\mathord{{\rm d}}} {\mathscr E}^{(i)}_t={\mathscr E}^{(i)}_t\tilde b^{(i)}_t(Z_t)\cdot{\mathord{{\rm d}}} W_t, $$ we have $$ {\mathord{{\rm d}}} ({\mathscr E}^{(1)}_t-{\mathscr E}^{(2)}_t)=({\mathscr E}^{(1)}_t\tilde b^{(1)}_t(Z_t)-{\mathscr E}^{(2)}_t\tilde b^{(2)}_t(Z_t))\cdot{\mathord{{\rm d}}} W_t. $$ By It\^o's formula, we have \begin{align*} &{\mathord{{\rm d}}} |{\mathscr E}^{(1)}_t-{\mathscr E}^{(2)}_t|^2=|{\mathscr E}^{(1)}_t\tilde b^{(1)}_t(Z_t)-{\mathscr E}^{(2)}_t\tilde b^{(2)}_t(Z_t)|^2{\mathord{{\rm d}}} t+M_t,\\ &\quad\leqslant 2|{\mathscr E}^{(1)}_t-{\mathscr E}^{(2)}_t|^{2}|\tilde b^{(1)}_t(Z_t)|^2{\mathord{{\rm d}}} t+2|{\mathscr E}^{(2)}_t(\tilde b^{(1)}_t(Z_t)-\tilde b^{(2)}_t(Z_t))|^2{\mathord{{\rm d}}} t+M_t, \end{align*} where $M$ is a continuous local martingale. Since $\delta<2$, by the stochastic Gronwall inequality \eqref{gron} and \eqref{ET5}, we obtain $$ \|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T\|_{\delta}^2\lesssim \int^T_0{\mathbf E} |{\mathscr E}^{(2)}_t(\tilde b^{(1)}_t(Z_t)-\tilde b^{(2)}_t(Z_t))|^2{\mathord{{\rm d}}} t. $$ Since $(p,q)\in{\mathscr I}_1$, one can choose $\gamma\in(1,1/(d/p+2/q))$ so that $$(p/(2\gamma),q/(2\gamma))\in{\mathscr I}_2.$$ Thus by H\"older's inequality and Krylov's estimate \eqref{Kry2}, we further have \begin{align*} \|{\mathscr E}^{(1)}_T-{\mathscr E}^{(2)}_T\|_{\delta}^2&\stackrel{\eqref{ET50}}{\lesssim} \left(\int^T_0{\mathbf E} |\tilde b^{(1)}_t(Z_t)-\tilde b^{(2)}_t(Z_t)|^{2\gamma}{\mathord{{\rm d}}} t\right)^{\frac{1}{\gamma}}\\ &\lesssim |\mspace{-3mu}|\mspace{-3mu}| |\tilde b^{(1)}-\tilde b^{(2)}|^{2\gamma}|\mspace{-3mu}|\mspace{-3mu}|^{1/\gamma}_{\widetilde{\mathbb L}^{p/(2\gamma)}_{q/(2\gamma)}(T)} =|\mspace{-3mu}|\mspace{-3mu}|\tilde b^{(1)}-\tilde b^{(2)}|\mspace{-3mu}|\mspace{-3mu}|^2_{\widetilde{\mathbb L}^p_q(T)}\\ &\lesssim \left(\int^T_0|\mspace{-3mu}|\mspace{-3mu}| b_t(\cdot,\mu^{(1)}_t)-b_t(\cdot,\mu^{(2)}_t)|\mspace{-3mu}|\mspace{-3mu}|^{q}_p{\mathord{{\rm d}}} t\right)^{\frac{2}{q}}\\ &\stackrel{\eqref{CC1}}{\lesssim} \left(\int^T_0\ell_t^{q}\|\phi_\theta\cdot (\mu^{(1)}_t-\mu^{(2)}_t)\|^{q}_{TV}{\mathord{{\rm d}}} t\right)^{\frac{2}{q}}, \end{align*} which together with \eqref{LK9} yields $$ \|\phi_\theta\cdot (\mu^{(1)}_T-\mu^{(2)}_T)\|^q_{TV}\leqslant C\int^T_0\ell_t^{q}\|\phi_\theta\cdot (\mu^{(1)}_t-\mu^{(2)}_t)\|^{q}_{TV}{\mathord{{\rm d}}} t. $$ By Gronwall's inequality, we obtain $$ \|\phi_\theta\cdot (\mu^{(1)}_T-\mu^{(2)}_T)\|^{q}_{TV}=0\Rightarrow\mu^{(1)}_T=\mu^{(2)}_T. $$ The proof is thus complete. \end{proof} \section{Application to nonlinear Fokker-Planck equations} In this section we present some applications to nonlinear Fokker-Planck equations. First of all we recall the following superposition principle: one-to-one correspondence between DDSDE \eqref{SDE} and nonlinear Fokker-Planck equation \eqref{Non}, which was first proved in \cite{Ba-Ro, Ba-Ro1}, and is based on a result for linear Fokker-Planck equations due to Trevisan \cite{Tr} (see also \cite{Fi} for the special linear case where the coefficients are bounded). We repeat the argument from \cite{Ba-Ro, Ba-Ro1} here. \begin{theorem}[Superposition principle]\label{Th23} Let $\mu_t:{\mathbb R}_+\to{\mathcal P}({\mathbb R}^d)$ be a continuous curve such that for each $T>0$, \begin{align}\label{ET1} \int^T_0\!\!\!\int_{{\mathbb R}^d}\Big(|(\sigma^{ik}_t\sigma^{jk}_t)(x,\mu_t)|+|b_t(x,\mu_t)|\Big)\mu_t({\mathord{{\rm d}}} x){\mathord{{\rm d}}} t<\infty. \end{align} Then $\mu_t$ solves the nonlinear Fokker-Planck equation \eqref{Non} in the distributional sense if and only if there exists a martingale solution ${\mathbb P}\in{\mathscr M}^{\sigma,b}_\nu$ to DDSDE \eqref{SDE} so that for each $t>0$, $$ \mu_t={\mathbb P}\circ w^{-1}_t. $$ In particular, if there is at most one element in ${\mathcal M}^{\sigma,b}_\nu$ with time martingale $\mu_t:=\mu_{X_t}, t\geqslant 0$, satisfying \eqref{ET1}, then there is at most one solution to \eqref{Non} satisfying \eqref{ET1}. \end{theorem} \begin{proof} If ${\mathbb P}\in {\mathscr M}^{\sigma,b}_\nu$ and $\mu_t={\mathbb P}\circ w^{-1}_t$, then by \eqref{ET1} and It\^o's formula, it is easy to see that $\mu_t$ solves \eqref{Non}. Now we assume $\mu_t$ solves \eqref{Non}. Consider the following linear Fokker-Planck equation: $$ \partial_t\tilde\mu_t=({\mathscr L}_t^{\sigma^\mu})^*\tilde\mu_t+\mathord{{\rm div}}(b^\mu_t\cdot\tilde\mu_t), $$ where $b^\mu_t(x):=b_t(x,\mu_t)$ and $\sigma^\mu_t(x):=\sigma_t(x,\mu_t)$. Since $\mu_t$ is a solution of the above linear Fokker-Planck equation, by \cite[Theorem 2.5]{Tr}, there is a martingale solution ${\mathbb P}\in{\mathscr M}^{\sigma^\mu, b^\mu}_\nu$ so that $$ \mu_t={\mathbb P}\circ w_t^{-1}. $$ In particular, ${\mathbb P}\in{\mathscr M}^{\sigma, b}_\nu$. The last assertion is then obvious and thus the proof is complete. \end{proof} \iffalse We have the following useful consequence. \begin{corollary} For $\nu\in{\mathcal P}({\mathbb R}^d)$, let ${\mathbb M}_\nu$ be the set of all continuous curves $\mu_t:{\mathbb R}_+\to{\mathcal P}({\mathbb R}^d)$ with property \eqref{ET1} and $\mu_0=\nu$. Then there is exactly one solution for \eqref{Non} in the class $\Theta_\nu$ for all $\nu\in{\mathcal P}({\mathbb R}^d)$ if and only if there is exactly one ${\mathbb P}\in{\mathscr M}^{\sigma,b}_\nu$ so that $t\mapsto{\mathbb P}\circ w^{-1}_t\in{\mathbb M}_\nu$ for all $\nu\in{\mathcal P}({\mathbb R}^d)$. \end{corollary} \begin{proof} The backward direction ``($\Leftarrow$)'' follows from the last assertion in Theorem \ref{Th23}. The other direction follows from a well-known fact in the theory of martingale problems (see \cite{St-Va}). \end{proof} \begin{remark} By \cite{Ma-Sh} or \cite[Theorem 6.7.8]{Bo-Kr-Ro-Sh}, \cite[Theorem 3.1]{Ma-Ro-Sh} and the above corollary, one sees that for a large class of functions $\sigma$ and $b$, there exists a unique martingale solution to DDSDE \eqref{SDE}. For example, if $b$ is bounded measurable and $\sigma$ satisfies \eqref{SI}, then there is a unique weak solution for DDSDE \eqref{SDE}. However, the results in \cite{Ma-Ro-Sh} does not apply to \eqref{BB} with $b\in \widetilde{\mathbb L}^p_q(T)$ since in this case the Lyapunov condition is not satisfied. \end{remark} \fi From the above superposition principle and our well-posedness results, we can obtain the following wellposedness result about the nonlinear Fokker-Planck equations. \begin{theorem} In the situations of Theorems \ref{Th51} and \ref{Th43}, there is a unique continuous curve $\mu_t$ solving the nonlinear Fokker-Planck equation \eqref{Non}. \end{theorem} Now we turn to the proof of Theorem \ref{Th11}. \begin{proof}[Proof of Theorem \ref{Th11}] The existence and uniqueness of solutions to the nonlinear FPE \eqref{FK} are consequences of Theorem \ref{Th43} and Theorem \ref{Th23}. We now aim to show the existence and smoothness of the density $\rho^X_t(y)$. Let $\mu_t$ be the solution of the Fokker-Planck equation \eqref{FK}. We consider the following SDE: \begin{align}\label{SDE2} {\mathord{{\rm d}}} X_t=b^\mu_t(X_t){\mathord{{\rm d}}} t+\sqrt{2}{\mathord{{\rm d}}} W_t, \ X_0=\xi, \end{align} where $b^\mu_t(x):=\int_{{\mathbb R}^d}b_t(x,y)\mu_t({\mathord{{\rm d}}} y)$. Since $b^\mu\in\widetilde{\mathbb L}^p_q$, where $\frac{d}{p}+\frac{2}{q}<1$, it is well known that the operator $\Delta+b^\mu\cdot\nabla$ admits a heat kernel $\rho_{b^\mu}(s,x;t,y)$ (see \cite[Theorems 1.1 and 1.3]{Ch-Hu-Xi-Zh}), which is continuous in $(s,x;t,y)$ on $\{(s,x;t,y): 0\leqslant s<t<\infty, x,y\in{\mathbb R}^d\}$ and satisfies the following two-sided estimate: For any $T>0$, there are constants $c_0,\gamma_0>1$ such that for all $0\leqslant s<t\leqslant T$ and $x,y\in{\mathbb R}^d$ $$ c_0^{-1}(t-s)^{-d/2}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-\gamma_0 |x-y|^2/(t-s)}\leqslant \rho_{b^\mu}(s,x;t,y)\leqslant c_0(t-s)^{-d/2}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-|x-y|^2/(\gamma_0(t-s))}, $$ and the gradient estimate: for some $c_1,\gamma_1>1$, $$ |\nabla_x\rho_{b^\mu}(s,x;t,y)|\leqslant c_1(t-s)^{-(d+1)/2}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-|x-y|^2/(\gamma_1(t-s))}. $$ If $\mathord{{\rm div}} b\equiv 0$, then $\rho_{b^\mu}(s,x;t,y)=\rho_{-b^\mu}(s,y;t,x)$, and so in this case, $$ |\nabla_y\rho_{b^\mu}(s,x;t,y)|\leqslant c_1(t-s)^{-(d+1)/2}\text{\rm{e}}} \def\ua{\underline a} \def\OO{\Omega} \def\oo{\omega^{-|x-y|^2/(\gamma_1(t-s))}. $$ In particular, the density of the law of $X_t$ is just given by $$ \rho^X_t(y)=\int_{{\mathbb R}^d}\rho(0,x;t,y)({\mathbf P}\circ X^{-1}_0)({\mathord{{\rm d}}} x). $$ Strong uniqueness of SDE \eqref{SDE2} ensures that $\rho^X_t(y){\mathord{{\rm d}}} y=\mu_t({\mathord{{\rm d}}} y)$. The desired estimates now follow from the above estimates. \end{proof} {\bf Acknowledgement:} The authors thank Dr. Xing Huang for pointing out an error in the earlier version.

b488aa082fe1ac2f3bbcc932a2377495341ef6f0

\section{Introduction} In this paper, we study methods for calculating the degree distance and the Gutman index of a graph. These graph invariants are important degree- and distance-based topological indices. The \textit{degree distance} of a connected graph $G$, denoted by $DD(G)$, is defined as $$DD(G) = \sum_{\lbrace u,v \rbrace \subseteq V(G)}(\textrm{deg}(u) + \textrm{deg}(v))d(u,v)$$ and was firstly introduced in 1994 \cite{dobrynin}. However, a similar concept was invented five years earlier by H. Schultz (see \cite{schultz}) and therefore, the degree distance is sometimes referred to as the Schultz index. The \textit{Gutman index} of a connected graph $G$, denoted by $Gut(G)$, was introduced in \cite{gutman_index} and it is defined in the following way: $$Gut(G) = \sum_{\lbrace u,v \rbrace \subseteq V(G)}\textrm{deg}(u)\textrm{deg}(v)d(u,v).$$ \noindent Both topological indices are special variations of the well known Wiener index, which was introduced in 1947 by H. Wiener \cite{wiener} as $$W(G)= \sum_{\lbrace u,v \rbrace \subseteq V(G)}d(u,v).$$ In fact, the Gutman index and the degree distance are special cases of weighted Wiener indices. Therefore, our results are stated in a general way. The degree-distance and the Gutman index were extensively studied in the past. For example, some recent research on the degree-distance can be found in \cite{hua,khalifeh-2009,zhu} and on the Gutman index in \cite{knor,mazo}. Moreover, both indices were studied and compared in \cite{das,ghala,guo,gutman1}. In addition, paper \cite{khalifeh-2010} provides some methods for computing these two indices on partial cubes. Furthermore, recently some generalizations using the Steiner distance were introduced \cite{gutman,mao,mao1}. The paper is organized in six sections. The following section states some basic definitions and preliminary results. We continue with a cut method for computing the degree-distance of a graph. Note that a cut method is a powerful tool for investigating distance-based topological indices \cite{klavzar-2015}. In \cite{glavni} a cut method for computing the degree-distance of a benzenoid system was developed. Therefore, in section 3 we generalize this result such that it can be applied for any connected graph by using quotient graphs with respect to a partition coarser than $\Theta^*$-partition. In addition, our result holds for the Wiener index of a double vertex-weighted graph and is analogous to Theorem \ref{rac_wie} from \cite{nad_klav}, which can be applied to calculate the Wiener index of a weighted graph. In section 4, the mentioned theorems are used for computing the degree-distance and the Gutman index of phenylenes. More precisely, these indices of an arbitrary phenylene are expressed by using its hexagonal squeeze and inner dual. Such result is already known for the Wiener index of a phenylene \cite{klav_gut-1997}. Moreover, it is shown how these two indices of a phenylene can be obtained from the four quotient trees. In \cite{redukcija} some reduction theorems for calculating the Wiener index of a weighted graph were proposed. In section 5, we apply a similar method to obtain such theorems also for the Wiener index of a double vertex-weighted graph, which can then be used also for the degree-distance. As an example, we calculate the degree-distance and the Gutman index of an infinite family of graphs that were firstly considered in \cite{nad_klav}. Finally, in section 6 we show an inequality for the Wiener index of a weighted graph and apply it for the Gutman index. In particular, the equality holds if and only if the graph is a partial Hamming graph. Note that similar results are known for the Wiener index \cite{klavzar-2006} and for the degree-distance \cite{glavni}. \section{Preliminaries} All the graphs considered in this paper are simple and finite. For a graph $G$, the set of all the vertices is denoted by $V(G)$ and the set of edges by $E(G)$. Moreover, we define $d_G(u,v)$ to be the usual shortest-path distance between vertices $u, v\in V(G)$ and also sometimes write $d(u,v)$ instead of $d_G(u,v)$ if there is no confusion. Furthermore, for any $u \in V(G)$ the {\it open neighbourhood} $N(u)$ is defined as the set of all the vertices that are adjacent to $u$. The \textit{degree} of $u$, denoted by $\textrm{deg}(u)$, is defined as the cardinality of the set $N(u)$. Finally, we set $N[u]=N(u) \cup \lbrace u \rbrace$. \smallskip \noindent If $G$ is a graph, we say that a function $w: V(G) \rightarrow \mathbb{R}^+$ is a \textit{weight}. A pair $(G,w)$ is then called a \textit{vertex-weighted graph} or shortly a \textit{weighted graph}. Moreover, if $a,b$ are weights, then the triple $(G,a,b)$ is called a \textit{double vertex-weighted graph}. \smallskip \noindent The Wiener index of a connected weighted graph $(G,w)$ can be defined in two different ways (see \cite{khalifeh-2010}): $$W_{*}(G,w) = \sum_{\lbrace u,v \rbrace \subseteq V(G)}w(u)w(v)d(u,v),$$ $$W_{+}(G,w) = \sum_{\lbrace u,v \rbrace \subseteq V(G)}(w(u) + w(v))d(u,v).$$ However, since the first definition is more common, we usually write $W(G,w)$ instead of $W_{*}(G,w)$ and call it the Wiener index of a weighted graph. Obviously, the Gutman index and the degree distance are just special cases of these two weighted indices. \smallskip \noindent The Wiener index of a connected double vertex-weighted graph $(G,a,b)$, introduced in \cite{glavni}, is defined as: $$W(G,a,b)= \sum_{\lbrace u,v \rbrace \subseteq V(G)}\left(a(u)b(v) + a(v)b(u) \right)d(u,v).$$ Obviously, $W_{+}(G,a)$ is exactly $W(G,a,b)$ if we take $b\equiv 1$. \smallskip \noindent Two edges $e_1 = u_1 v_1$ and $e_2 = u_2 v_2$ of graph $G$ are in relation $\Theta$, $e_1 \Theta e_2$, if $$d(u_1,u_2) + d(v_1,v_2) \neq d(u_1,v_2) + d(u_1,v_2).$$ Note that this relation is also known as Djokovi\' c-Winkler relation. The relation $\Theta$ is reflexive and symmetric, but not necessarily transitive. We denote its transitive closure (i.e.\ the smallest transitive relation containing $\Theta$) by $\Theta^*$. It is easy to observe that any two edges in an odd cycle are in $\Theta^*$ relation. Moreover, any two diametrically opposite edges in an even cycle are in $\Theta$ relation. Let $ \mathcal{E} = \lbrace E_1, \ldots, E_k \rbrace$ be the $\Theta^*$-partition of the set $E(G)$. Then we say that a partition $\lbrace F_1, \ldots, F_r \rbrace$ of $E(G)$ is \textit{coarser} than $\mathcal{E}$ if each set $F_i$ is the union of one or more $\Theta^*$-classes of $G$. \smallskip \noindent Suppose $G$ is a graph and $F \subseteq E(G)$. The \textit{quotient graph} $G / F$ is a graph whose vertices are connected components of the graph $G \setminus F$, such that two components $C_1$ and $C_2$ are adjacent in $G / F$ if some vertex in $C_1$ is adjacent to a vertex of $C_2$ in $G$. \smallskip \noindent The following theorem gives a method for computing the Wiener index of a weighted graph. \begin{theorem} \cite{nad_klav} \label{rac_wie} Let $G$ be a connected graph. If $\lbrace F_1, \ldots, F_k \rbrace$ is a partition coarser than the $\Theta^*$-partition, then $$W(G,w) = \sum_{i=1}^k W(G / F_i, w_i),$$ where $w_i: V(G/F_i) \rightarrow \mathbb{R}^+$ is defined by $w_i(C)= \sum_{x \in V(C)} w(x)$ for all connected components $C$ of the graph $G \setminus F_i$. \end{theorem} \noindent The \textit{Cartesian product} $G_1 \Box \cdots \Box G_k$ of graphs $G_1, \ldots, G_k$ has the vertex set $V(G_1)\times \cdots \times V(G_k)$, two vertices $(u_1,\ldots,u_k)$ and $(v_1,\ldots, v_k)$ being adjacent if they differ in exactly one position, say in $i$th, and $u_iv_i$ is an edge of $G_i$. \smallskip \noindent Let $H$ and $G$ be two graphs. A function $\ell: V(H) \rightarrow V(G)$ is called an \textit{embedding of $H$ into $G$} if $\ell$ is injective and, for any two vertices $u,v \in V(H)$, $\ell(u)\ell(v) \in E(G)$ if $uv \in E(H)$. If such a function $\ell$ exists, we say that $H$ can be \textit{embedded} in $G$. An embedding $\ell$ of graph $H$ into graph $G$ is called an \textit{isometric embedding} if for any two vertices $u,v \in V(H)$ it holds $d_H(u,v) = d_G(\ell(u),\ell(v))$. Moreover, subgraph $H$ of a graph $G$ is called an \textit{isometric subgraph} if for each $u,v \in V(H)$ it holds $d_H(u,v) = d_G(u,v)$. \smallskip \noindent A \textit{Hamming graph} is the Cartesian product of complete graphs and a \textit{partial Hamming graph} is any isometric subgraph of a Hamming graph. In the particular case where all the factors are $K_2$'s we speak of hypercubes and partial cubes, respectively. Partial cubes constitute a large class of graphs with a lot of applications and includes, for example, many families of chemical graphs (benzenoid systems, trees, phenylenes, cyclic phenylenes, polyphenyls). Partial Hamming graphs and in particular partial cubes were studied and characterized in many papers (for example, see \cite{klavzar-book}). \smallskip \noindent For an edge $ab$ of a graph $G$, let $W_{ab}$ be the set of vertices of $G$ that are closer to $a$ than to $b$. We write $\langle S \rangle$ for the subgraph of $G$ induced by $S \subseteq V(G)$. Moreover, a subgraph $H$ of $G$ is called {\it convex} if for arbitrary vertices $u,v \in V(H)$ every shortest path between $u$ and $v$ in $G$ is also contained in $H$. The following theorem proved by Djokovi\' c and Winkler puts forth two fundamental characterizations of partial cubes: \begin{theorem} \cite{klavzar-book} \label{th:partial-k} For a connected graph $G$, the following statements are equivalent: \begin{itemize} \item [(i)] $G$ is a partial cube. \item [(ii)] $G$ is bipartite, and $\langle W_{ab} \rangle $ and $\langle W_{ba} \rangle$ are convex subgraphs of $G$ for all $ab \in E(G)$. \item [(iii)] $G$ is bipartite and $\Theta = \Theta^*$. \end{itemize} \end{theorem} Is it also known that if $G$ is a partial cube and $E$ is a $\Theta$-class of $G$, then $G \setminus E$ has exactly two connected components, namely $\langle W_{ab} \rangle $ and $\langle W_{ba} \rangle$, where $ab \in E$. For more information about partial cubes see \cite{klavzar-book}. \smallskip \noindent The \textit{canonical embedding} of a connected graph $G$ is defined as follows: Let the $\Theta^*$-classes of $G$ be $E_1,E_2,\ldots,E_k$. Let $\alpha_i : V(G)\rightarrow V(G / E_i)$ be the map sending any $u \in V(G)$ to the component of $G \setminus E_i$ that contains it. The canonical embedding $\alpha : V(G) \rightarrow G / E_1 \Box \cdots \Box G / E_k$ is defined by $$\alpha(u) = (\alpha_1(u), \ldots, \alpha_k(u))$$ for any $u \in V(G)$. It is well known that the canonical embedding is irredundant isometric embedding, see \cite{klavzar-book}. Here irredundant means that every factor graph $G / E_i$, $i \in \lbrace 1,\ldots,k \rbrace $, has at least two vertices and that each vertex of $G / E_i$ appears as a coordinate of some vertex $\alpha(u)$, where $u \in V(G)$. \baselineskip=16pt \section{The degree distance via unification of $\Theta^*$-classes} The main goal of this section is to find a formula for calculating the degree distance of a graph. However, we present a result in a more general setting, i.e.\,by proving it for the Wiener index of a double vertex-weighted graph. Some chemical applications are given in the next section. Let $G$ be a connected graph and $\lbrace F_1, \ldots, F_r \rbrace$ a partition coarser than the $\Theta^*$-partition. For any $u \in V(G)$ and $i \in \lbrace 1, \ldots, r \rbrace$ we denote by $\ell_i(u)$ the connected component of the graph $G \setminus F_i$ which contains $u$. The result of the following lemma was proved in \cite{nad_klav}. A complete proof can also be found in \cite{tratnik}. \begin{lemma} \cite{nad_klav,tratnik} \label{distance} Let $G$ be a connected graph. If $\lbrace F_1, \ldots, F_r \rbrace$ is a partition coarser than the $\Theta^*$-partition, then for any $u,v \in V(G)$ it holds $$d_G(u,v) = \sum_{i=1}^r d_{G / F_i}(\ell_i(u),\ell_i(v)).$$ \end{lemma} \noindent The following theorem is the main result of the section. \begin{theorem} \label{splosen} If $(G,a,b)$ is a connected double vertex-weighted graph with a partition $\{F_1,...,F_r\}$ coarser than $\Theta^*$-partition, then $$W(G,a,b)=\sum_{i=1}^r W(G/F_i, a_i, b_i),$$ where $a_i: V(G/F_i) \rightarrow \mathbb{R}^+$ is defined by $a_i(C)= \sum_{x \in V(C)} a(x)$ and $b_i: V(G/F_i) \rightarrow \mathbb{R}^+$ is defined by $b_i(C)= \sum_{x \in V(C)} b(x)$ for all connected components $C$ of the graph $G \setminus F_i$. \end{theorem} \begin{proof} By the definition and Lemma \ref{distance} we obtain \begin{eqnarray*} W(G,a,b) & = & \sum_{\{u,v\} \in V(G)} (a(u)b(v)+a(v)b(u))d_G(u,v) \\ & = & \sum_{\{u,v\} \in V(G)}(a(u)b(v)+a(v)b(u))\left( \sum_{i=1}^r d_{G/F_i}(\ell_i(u),\ell_i(v))\right) \\ & = & \sum_{i=1}^r \left(\sum_{\{u,v\} \in V(G)} (a(u)b(v)+a(v)b(u)) d_{G/F_i}(\ell_i(u),\ell_i(v)) \right). \end{eqnarray*} \noindent Let $X,Y$ be two arbitrary distinct connected components of $G \setminus F_i$. Obviously, for any $x,x' \in V(X)$ and $y, y' \in V(Y)$ it holds $d_{G/F_i}(\ell_i(x),\ell_i(y))=d_{G/F_i}(\ell_i(x'),\ell_i(y'))$. Moreover, \begin{eqnarray*} \sum_{x \in V(X)} \sum_{y \in V(Y)} \left(a(x)b(y)+a(y)b(x) \right)& = & \sum_{x \in V(X)}a(x)\sum_{y \in V(Y)}b(y) + \sum_{y \in V(Y)}a(y)\sum_{x \in V(X)}b(x) \\ & = & a_i(X)b_i(Y) + a_i(Y)b_i(X). \\ \end{eqnarray*} \noindent Therefore, \begin{eqnarray*} W(G,a,b)& = & \sum_{i=1}^r \left(\sum_{\{X,Y\} \in V(G/F_i)} (a_i(X)b_i(Y) + a_i(Y)b_i(X)) d_{G/F_i}(X,Y) \right) \\ & = & \sum_{i=1}^r W(G/F_i, a_i, b_i), \\ \end{eqnarray*} \noindent which finishes the proof. \hfill $\square$ \bigskip \end{proof} In the rest of the section we present some consequences of the previous theorem. Firstly, the obtained result can be used to compute $W_+(G,a)$ for a connected weighted graph. \begin{corollary} Let $(G,a)$ be a connected weighted graph with a partition $\{F_1,...,F_r\}$ coarser than $\Theta^*$-partition. It holds $$W_+(G,a)=\sum_{i=1}^r W(G/F_i, a_i,b_i)$$ where $a_i: V(G/F_i) \rightarrow \mathbb{R}$ is defined by $a_i(C)= \sum_{x \in V(C)} a(x)$ and $b_i: V(G/F_i) \rightarrow \mathbb{R}^+$ is defined by $b_i(C)= |V(C)|$ for all connected components $C$ of the graph $G \setminus F_i$. \end{corollary} \begin{proof} If we set $b\equiv1$ in Theorem \ref{splosen}, we quickly get the obtained formula. \hfill $\square$ \bigskip \noindent \end{proof} \noindent Finally, we are able to calculate the degree distance of an arbitrary graph $G$. \begin{corollary} \label{rac_degree} If $G$ is a connected graph with a partition $\{F_1,...,F_r\}$ coarser than $\Theta^*$-partition, then $$DD(G)=\sum_{i=1}^r W(G/F_i, a_i, b_i),$$ where $a_i: V(G/F_i) \rightarrow \mathbb{R}$ is defined by $a_i(C)= \sum_{x \in V(C)} {\rm deg}(x)$ and $b_i: V(G/F_i) \rightarrow \mathbb{R}^+$ is defined by $b_i(C)= |V(C)|$ for all connected components $C$ of the graph $G \setminus F_i$. \end{corollary} If $G$ is a partial cube and $E_1, \ldots, E_k$ are the $\Theta$-classes of $G$, we know that the graph $G \setminus E_i$ has exactly two connected components for any $i \in \lbrace 1,\ldots, k \rbrace$ and these two components will be denoted by $C_i^1$ and $C_i^2$. Moreover, for any $i \in \lbrace 1, \ldots, k \rbrace$ and $j \in \lbrace 1,2 \rbrace$ we define $A_j(E_i) = \sum_{x \in V(C_i^j)} a(x)$ and $B_j(E_i) = \sum_{x \in V(C_i^j)} b(x)$. The next result follows directly from Theorem \ref{splosen} and generalizes Lemma 4.2 from \cite{glavni} where it was stated just for trees. \begin{corollary} \label{rac_delne} If $(G,a,b)$ is a double vertex-weighted partial cube and if $E_1,\ldots,E_k$ are the $\Theta$-classes of $G$, then $$W(G,a,b)=\sum_{i=1}^k \left( A_1(E_i)B_2(E_i) + A_2(E_i)B_1(E_i) \right).$$ \end{corollary} \section{Applications to phenylenes} In this section, we apply Corollary \ref{rac_degree} and Theorem \ref{rac_wie} to obtain the relationships between the degree-distance and the Gutman index of a phenylene with the weighted Wiener indices of its hexagonal squeeze and the inner dual. Moreover, it is described how the two indices can be computed from four weighted quotient trees which enables us to compute the indices in linear time. Firstly, we need to introduce some additional notation. Let ${\cal H}$ be the hexagonal (graphite) lattice and let $Z$ be a cricuit on it. Then a {\em benzenoid system} is induced by the vertices and edges of ${\cal H}$, lying on $Z$ and in its interior. Let $B$ be a benzenoid system. A vertex shared by three hexagons of $B$ is called an \textit{internal} vertex of $B$. A benzenoid system is said to be \textit{catacondensed} if it does not possess internal vertices. Otherwise it is called \textit{pericondensed}. Two distinct hexagons with a common edge are called \textit{adjacent}. The \textit{inner dual} of a benzenoid system $B$, denoted as $ID(B)$, is a graph which has hexagons of $B$ as vertices, two being adjacent whenever the corresponding hexagons are adjacent. Obviously, the inner dual of a catacondensed benzenoid system is always a tree. Let $B$ be a catacondensed benzenoid system. If we add rectangles between all pairs of adjacent hexagons of $B$, the obtained graph $G$ is called a \textit{phenylene}. We then say that $B$ is a \textit{hexagonal squeeze} of $G$ and denote it by $HS(G)=B$, see Figure \ref{ben_phe}. \begin{figure}[h] \begin{center} \includegraphics[scale=0.7]{ben_phe.eps} \end{center} \caption{\label{ben_phe} A benzenoid system $B$ with its inner dual and a phenylene $G$ such that $B=HS(G)$.} \end{figure} By Theorem \ref{th:partial-k} it follows that any benzenoid system or a phenylene is a partial cube. An \textit{elementary cut} $C$ of a benzenoid system or a phenylene $G$ is a line segment that starts at the center of a peripheral edge of $G$, goes orthogonal to it and ends at the first next peripheral edge of $G$. By $C$ we usually also denote the set of edges that are intersected by the corresponding elementary cut. Elementary cuts in benzenoid systems have been described and illustrated by numerous examples in several earlier articles. The main insight for our consideration is that every $\Theta$-class of a benzenoid system or a phenylene coincides with exactly one of its elementary cuts. Let $G$ be a phenylene, $B=HS(G)$, and $T=ID(B)$. Firstly, we introduce four different weights in the following way. The weights $\omega_1,\omega_2 : V(B) \rightarrow \mathbb{R}^+$ are defined as $$w_1(u)= 4{\rm deg}(u) - 6,$$ $$w_2(u)= {\rm deg}(u) - 1$$ for any $u \in V(B)$. Moreover, the weights $\omega_3,\omega_4 : V(T) \rightarrow \mathbb{R}^+$ are defined as $$w_3(x)= 2{\rm deg}(x) + 12,$$ $$w_4(x)= 6$$ for any $x \in V(T)$. \begin{theorem} If $G$ is a phenylene, $B=HS(G)$, and $T=ID(B)$, then it holds $$DD(G)= W(B,w_1,w_2) + W(T,w_3,w_4),$$ $$Gut(G)= W(B,w_1)+W(T,w_3).$$ \end{theorem} \begin{proof} Let $F_1$ be the set of all the edges of $G$ that correspond to the edges of the hexagonal squeeze $B$ (the edges of all the hexagons of $G$). Moreover, let $F_2 = E(G) \setminus F_1$. Since the sets $F_1,F_2$ are both unions of elementary cuts of $G$, it is obvious that the partition $\lbrace F_1, F_2 \rbrace$ is a partition coarser than $\Theta$-partition. Therefore, by Corollary \ref{rac_degree} we have $$DD(G) = W(G / F_1, a_1,b_1) + W(G / F_2, a_2, b_2),$$ \noindent where $a_i$, $i \in \lbrace 1,2 \rbrace$, represents the sum of all the degrees in the corresponding connected components of the graph $G \setminus F_i$, and $b_i$, $i \in \lbrace 1,2 \rbrace$, represents the number of vertices in the corresponding connected components of the graph $G \setminus F_i$. We also notice that $G / F_1 \cong B$ and $G / F_2 \cong T$. Furthermore, the connected components of $G \setminus F_1$ are either vertices (the vertices of degree 2 in $G$, which correspond to vertices of degree 2 in $B$) or edges (composed of two vertices of degree 3 in $G$, which correspond to vertices of degree 3 in $B$). Therefore, $$a_1(u) = \begin{cases} 6; & {\rm deg}(u) = 3\\ 2; & {\rm deg}(u) = 2 \end{cases}, \quad b_1(u)=\begin{cases} 2; & {\rm deg}(u) = 3\\ 1; & {\rm deg}(u) = 2 \end{cases}$$ for any $u \in V(B)$. Obviously, $a_1 \equiv w_1$ and $b_1 \equiv w_2$. On the other hand, the connected components of $G \setminus F_2$ are the hexagons of $G$, each of them corresponds to exactly one vertex from $T$. Therefore, $$a_2(x) = \begin{cases} 18; & {\rm deg}(x) = 3\\ 16; & {\rm deg}(x) = 2\\ 14; & {\rm deg}(x) = 1 \end{cases}, \quad b_2(x)=6 $$ for any $x \in V(T)$. Obviously, $a_2 \equiv w_3$ and $b_2 \equiv w_4$, which completes the proof for the degree distance. For the Gutman index we consider the same partition $\lbrace F_1, F_2 \rbrace$ of the set $E(G)$. By Theorem \ref{rac_wie} we obtain $$Gut(G) = W(G / F_1, a_1) + W(G / F_2, a_2),$$ \noindent where $a_i$, $i \in \lbrace 1,2 \rbrace$, represents the sum of all the degrees in the corresponding connected components of the graph $G \setminus F_i$. Since $a_1 \equiv w_1$ and $a_2 \equiv w_3$, the proof is complete. \hfill $\square$ \bigskip \end{proof} Another way to compute the degree distance and the Gutman index of a phenylene is by using four weighted quotient trees, which are defined in the following way. Let $G$ be a phenylene and $B$ the hexagonal squeeze of $G$. The edge set of $B$ can be naturally partitioned into sets $E_1'$, $E_2'$, and $E_3'$ of edges of the same direction. Denote the sets of edges of $G$ corresponding to the edges in $E_1'$, $E_2'$, and $E_3'$ by $E_1, E_2$, and $E_3$, respectively. Moreover, let $E_4 = E(G) \setminus (E_1 \cup E_2 \cup E_3)$ be the set of all the edges of $G$ that do not belong to $B$. The quotient graph $T_i$, $1\le i\le 4$, is then defined in the standard way as the graph $G / E_i$. In a similar way we can define the quotient graphs $T_1', T_2', T_3'$ of the hexagonal squeeze $B$. It is known that for any benzenoid system its quotient graphs are trees, see \cite{chepoi-1996}. Obviously, a tree $T_i'$ is isomorphic to $T_i$ for $i=1,2,3$ and $T_4$ is isomorphic to the inner dual of $B$. Therefore, all quotient graphs $T_1,T_2,T_3,T_4$ are trees. Now we extend a quotient tree $T_i$, $i \in \lbrace 1,2,3,4 \rbrace$, to a double vertex-weighted tree $(T_i,a_i,b_i)$ as follows: \begin{itemize} \item for $C \in V(T_i)$, let $a_i(C)$ be the sum of all the degrees of vertices in the connected component $C$ of $G \setminus E_i$; \item for $C \in V(T_i)$, let $b_i(C)$ be the number of vertices in the component $C$ of $G \setminus E_i$. \end{itemize} \noindent Everything is prepared for the following theorem. \begin{theorem} If $G$ is a phenylene and $(T_i,a_i,b_i)$, $i \in \lbrace 1,2,3,4 \rbrace$, are the corresponding double vertex-weighted quotient trees, then it holds $$DD(G)= W(T_1,a_1,b_1) + W(T_2,a_2,b_2) + W(T_3,a_3,b_3) + W(T_4,a_4,b_4),$$ $$Gut(G)= W(T_1,a_1) + W(T_2,a_2) + W(T_3,a_3) + W(T_4,a_4).$$ \end{theorem} \begin{proof} It is clear the each set $E_i$, $i \in \lbrace 1,2,3,4 \rbrace$, is the union of some elementary cuts of $G$. Therefore, the partition $\lbrace E_1, E_2, E_3, E_4 \rbrace$ is a partition coarser than $\Theta$-partition. The theorem now follows by Corollary \ref{rac_degree} and Theorem \ref{rac_wie}. \hfill $\square$ \bigskip \end{proof} By the same reasoning as in \cite{chepoi-1996} we can show that the double vertex-weighted quotient trees of a phenylene can be computed in linear time with respect to the number of vertices. Moreover, it was shown in \cite{chepoi-1997,glavni} that the Wiener index of any weighted tree or double vertex-weighted tree can be computed in linear time with respect to the number of vertices of a tree. Therefore, we easily obtain the following corollary. \begin{corollary} If $G$ is a phenylene with $n$ vertices, then the degree distance of $G$ and the Gutman index of $G$ can be computed in $O(n)$ time. \end{corollary} For an example, we consider phenylene $G$ from Figure \ref{ben_phe}. The double vertex-weighted quotient trees for $G$ are depicted in Figure \ref{Wiener_trees}. \begin{figure}[h] \begin{center} \includegraphics[scale=0.6]{Wiener_trees.eps} \end{center} \caption{\label{Wiener_trees} The double vertex-weighted quotient trees of phenylene $G$.} \end{figure} \noindent We next compute the quantities (note that this can be done by using Corollary \ref{rac_delne}): $$W(T_1, a_1,b_1) = 5208, W(T_2,a_2,b_2) = 2976, W(T_3,a_3,b_3) = 4416, W(T_4,a_4,b_4) = 5784.$$ \noindent Therefore, the degree distance of $G$ is the sum $$DD(G)=5208+2976+4416+5784=18384.$$ \noindent Analogously, to compute the Gutman index we first calculate: $$W(T_1, a_1) = 6484, W(T_2,a_2) = 3600, W(T_3,a_3) = 5520, W(T_4,a_4) = 7252.$$ \noindent Finally, the Gutman index of $G$ is the sum $$Gut(G)=6484+3600+5520+7252=22856.$$ \section{Reduction theorems for connected double vertex-weighted graphs} We prove results analogous to the result from \cite{redukcija} for calculating the weighted Wiener index by a special reduction. To state the theorems, some additional definitions are needed. \noindent If $G$ is a graph, then vertices $x$ and $y$ are in relation $R$ if $N(x) = N(y)$. Obviously, $R$ is an equivalence relation on $V(G)$. The $R$-equivalence class containing $x$ will be denoted with $[x]_R$. \noindent Let $G$ be a connected graph, $c \in V (G)$ and $C = [c]_R$. We define a new graph $G'$ with $G' = G \setminus (C \setminus \{c \})$. For any weight $w: V(G) \rightarrow \mathbb{R}^+$ we define $w': V(G') \rightarrow \mathbb{R}^+$ in the following way: $\omega'(c)= \sum_{x \in C}\omega(x)$ and $\omega'(x) = w(x)$ for any $x \notin C$. The next theorem was recently obtained. \begin{theorem} \cite{redukcija} \label{posebna_redukcija0} Let $(G,w)$ be a connected weighted graph, $c \in V(G)$, and $C=[c]_R = \{c_1,\ldots,c_k\}$. Then $$W(G,w)= W(G',w') + \sum_{\{c_i,c_j\} \subseteq C}2w(c_i)w(c_j).$$ \end{theorem} \noindent Finally, we can state the main theorem of the section. \begin{theorem} \label{posebna_redukcija} Let $(G,a,b)$ be a connected double vertex-weighted graph, $c \in V(G)$, and $C=[c]_R = \{c_1,\ldots,c_k\}$. Then $$W(G,a,b)= W(G',a',b') + \sum_{\{c_i,c_j\} \subseteq C}2(a(c_i)b(c_j)+a(c_j)b(c_i)).$$ \end{theorem} \begin{proof} If $|C|=1$, $(G',\omega')=(G,\omega)$, hence the result is trivial. Let $c_1=c$ and $k \geq 2$. Then, we obtain \begin{itemize} \item [$(i)$] $d_G(c_i,x)=d_G(c_j,x)$ holds for any $c_i, c_j \in C$ and $x \notin C$, \item [$(ii)$] $d_G(x,y)=d_{G'}(x,y)$ holds for any vertices $x,y \in V(G) \setminus C$, \item [$(iii)$] $d_G(c_i,c_j)=2$ holds for any $c_i, c_j \in C$, $i \neq j$. \end{itemize} Using these facts we can compute the Wiener index of $(G,a,b)$ as follows: \begin{eqnarray*} W(G,a,b)&=& \sum_{\{x,y\} \subseteq V(G)}(a(x)b(y)+a(y)b(x))d_G(x,y) \\ &=& \sum_{x \notin C}\sum_{i=1}^k \left( a(c_i)b(x) + a(x)b(c_i) \right) d_G(c_i,x) \\ &+& \sum_{\{x,y\} \subseteq V(G)\setminus C}(a(x)b(y)+a(y)b(x))d_G(x,y) \\ &+& \sum_{\{c_i,c_j\} \subseteq C}(a(c_i)b(c_j)+a(c_j)b(c_i))d_G(c_i,c_j) \\ &=& \sum_{x \notin C} \left( b(x)\sum_{i=1}^k a(c_i)d_G(c_i,x)+ a(x)\sum_{i=1}^k b(c_i)d_G(c_i,x) \right) \\ &+& \sum_{\{x,y\} \subseteq V(G)\setminus C}(a(x)b(y)+a(y)b(x))d_G(x,y) \\ &+& \sum_{\{c_i,c_j\} \subseteq C}(a(c_i)b(c_j)+a(c_j)b(c_i))d_G(c_i,c_j). \end{eqnarray*} Therefore, applying properties $(i)$, $(ii)$, and $(iii)$ we deduce \begin{eqnarray*} W(G,a,b) &=& \sum_{x \in V(G') \setminus \lbrace c \rbrace} \left( b'(x)a'(c) + a'(x)b'(c) \right) d_{G'}(c,x) + \\ &+& \sum_{\{x,y\} \subseteq V(G')\setminus \lbrace c \rbrace}(a'(x)b'(y)+a'(y)b'(x))d_{G'}(x,y) + \\ &+& \sum_{\{c_i,c_j\} \subseteq C}2(a(c_i)b(c_j)+a(c_j)b(c_i)) \\ & = & W(G',a',b') + \sum_{\{a_i,a_j\} \subseteq C}2(a(a_i)b(a_j)+a(a_j)b(a_i)) \end{eqnarray*} and the proof is complete. \hfill $\square$ \bigskip \end{proof} \begin{corollary} \label{posledica_redukcije} Let $(G,a,b)$ be a connected double vertex-weighted graph and let $a(c_i)=k_1$ and $b(c_i)=k_2$ for all $c_i \in C$, $k_1, k_2 \in \mathbb{R^+}$. Then $$W(G,a,b)= W(G',a',b') + 2k_1k_2|C|(|C|-1).$$ \end{corollary} In a similar way as before, we can introduce relation $S$ in the following way, see \cite{redukcija}. If $G$ is a graph, then vertices $x$ and $y$ are in relation $S$ if $N_G[x] = N_G[y]$. Obviously, $S$ is an equivalence relation on $V(G)$ and the $S$-equivalence class containing $x$ will be denoted with $[x]_S$. Analogously as before we can define $(G',a',b')$ for any double vertex-weighted graph $(G,a,b)$ and $c \in V(G)$. The arguments in the proof of the next theorem are parallel to the proof of Theorem \ref{posebna_redukcija}, the only difference is that $d_G(c_i, c_j )$ equals one whenever $c_i,c_j$ are two distinct vertices from $C=[c]_S$. \begin{theorem} Let $(G,a,b)$ be a connected double vertex-weighted graph, $c \in V(G)$, and $C=[c]_S = \{c_1,\ldots,c_k\}$. Then $$W(G,a,b)= W(G',a',b') + \sum_{\{c_i,c_j\} \subseteq C}(a(c_i)b(c_j)+a(c_j)b(c_i)).$$ \end{theorem} In the rest of the section we use the obtained results on a family of graphs $G_n$, $n \geq 2$, where $n$ is the number of vertical layers, see Figure \ref{gn}. These graphs where introduced in \cite{nad_klav}, where the Wiener index was computed. Later, the obtained result was corrected in \cite{redukcija}. First, we determine the $\Theta^*$-classes of $G_n$, which are denoted by $E_1, \ldots, E_{n-1}$, see Figure \ref{gn}. \begin{figure}[h] \begin{center} \includegraphics[scale=0.6]{gn.eps} \end{center} \caption{\label{gn} Graph $G_n$, $n \geq 2$, and the subgraphs $G_n \setminus E_i$, $i \in \lbrace 1, \ldots, n-1 \rbrace$.} \end{figure} \noindent Moreover, let $F_1=E_1$ and $F_2 = \bigcup_{i=2}^{n-1}E_i$. The graph $G_n \setminus F_2$ is depicted in Figure \ref{gn_zdruzeni_razred}. In addition, Figures \ref{kvocientniE1} and \ref{kvocientniE2} show quotient graphs $(G_n / F_1,a_1,b_1)$ and $(G_n/F_2,a_2,b_2)$, where the weights are defined as in Corollary \ref{rac_degree}. \begin{figure}[h] \begin{center} \includegraphics[scale=0.8]{gn_zdruzeni_razred.eps} \end{center} \caption{\label{gn_zdruzeni_razred} The graph $G_n \setminus F_2$, where $F_2 = \bigcup_{i=2}^{n-1}E_i$.} \end{figure} \begin{figure}[h] \begin{center} \includegraphics[scale=0.65]{kvocientniE1.eps} \end{center} \caption{\label{kvocientniE1} The graph $(G_n / F_1,a_1,b_1)$.} \end{figure} \begin{figure}[h] \begin{center} \includegraphics[scale=0.65]{kvocientniE2.eps} \end{center} \caption{\label{kvocientniE2} The graph $(G_n / F_2,a_2,b_2)$, where $F_2 = \bigcup_{i=2}^{n-1}E_i$.} \end{figure} \noindent The Wiener indices of graphs $(G_n / F_1,a_1,b_1)$ and $(G_n / F_1,a_1)$ can be computed directly. The results are $$W(G/F_1,a_1,b_1) = 84 n^2 + 354n -152,$$ $$W(G/F_1,a_1) = 110 n^2 + 404n -196.$$ \noindent To simplify the calculation of the Wiener indices of graphs $(G_n / F_2,a_2,b_2)$ and $(G_n / F_2,a_2)$, we use Corollary \ref{posledica_redukcije} exactly $(n-2)$-times. Thus, we get the double vertex-weighted path on $(2n-3)$ vertices and we denote it by $(P,a^{(n-2)},b^{(n-2)})$, see Figure \ref{redukcijaE2}. \begin{figure}[h] \begin{center} \includegraphics[scale=0.65]{redukcijaE2.eps} \end{center} \caption{\label{redukcijaE2} Path $(P,a^{(n-2)},b^{(n-2)})$ obtained by performing special reductions.} \end{figure} \noindent To compute $W(P,a^{(n-2)},b^{(n-2)})$, we take into account separately the contributions of pairs of vertices such that \begin{itemize} \item both vertices have weights $(6,3)$, \item both vertices have weights $(10,3)$, \item one vertex has weights $(6,3)$ and the other $(10,3)$, \item one vertex has weights $(21,8)$ and the other $(6,3)$ or $(10,3)$ or $(21,8)$. \end{itemize} \noindent Therefore, we get the next result \begin{eqnarray*} W(P,a^{(n-2)},b^{(n-2)}) & = & (6 \cdot 3 + 3\cdot 6)\sum_{i=1}^{n-3} \sum_{j=1}^{i} (2j) + (10 \cdot 3 + 3\cdot 10)\sum_{i=1}^{n-4} \sum_{j=1}^{i} (2j) \\ & + & 2 \cdot (6 \cdot 3 + 3\cdot 10)\sum_{i=1}^{n-3} \sum_{j=1}^{i} (2j-1) \\ & + & 2 \cdot (21 \cdot 3 + 8 \cdot 6)\sum_{j=1}^{n-2}(2j-1) + 2 \cdot (21 \cdot 3 + 8 \cdot 10)\sum_{j=1}^{n-3}(2j) \\ & + & (21 \cdot 8 + 8 \cdot 21)(2n-4) \\ & = & 64n^3 + 16n^2 - 402n + 228. \end{eqnarray*} \noindent Hence, by Corollary \ref{posledica_redukcije} we obtain $$ W(G_n/F_2,a_2,b_2) = W(P,a^{(n-2)},b^{(n-2)}) + 24(n-2) = 64n^3 + 16n^2 - 378n + 180.$$ \noindent Finally, by Corollary \ref{rac_degree} one can calculate $$DD(G_n) = W(G_n/F_1,a_1,b_1) + W(G_n/F_2,a_2,b_2) = 64n^3 + 100n^2-24n+28.$$ Similarly, one can quickly get the Gutman index of $G_n$. We obtain \begin{eqnarray*} W(P,a^{(n-2)}) & = & 6^2 \sum_{i=1}^{n-3} \sum_{j=1}^{i} (2j) + 10^2\sum_{i=1}^{n-4} \sum_{j=1}^{i} (2j) \\ & + & 2 \cdot (6 \cdot 10)\sum_{i=1}^{n-3} \sum_{j=1}^{i} (2j-1) \\ & + & 2 \cdot (21 \cdot 6)\sum_{j=1}^{n-2}(2j-1) + 2 \cdot (21 \cdot 10)\sum_{j=1}^{n-3}(2j) \\ & + & 21^2 \cdot (2n-4) \\ & = & \frac{2}{3} \cdot(128n^3-731n+438). \end{eqnarray*} \noindent Hence, by Theorem \ref{posebna_redukcija0} we get $$ W(G_n/F_2,a_2) = W(P,a^{(n-2)}) + 24 \cdot (n-2) = \frac{2}{3} \cdot (128n^3-695n+366).$$ \noindent Lastly, by Theorem \ref{rac_wie} we obtain $$Gut(G_n) = W(G_n/F_1,a_1) + W(G_n/F_2,a_2) = \frac{2}{3} \cdot (128n^3+165n^2-89n +72).$$ \section{The Wiener index of weighted partial Hamming graphs} In this section, we show how the Wiener index of a weighted partial Hamming graph can be easily computed. Naturally, the obtained result can be applied also for the Gutman index. The following well known result will be needed and it was already used in \cite{klavzar-2006}. For the sake of completeness, we give the proof anyway. \begin{lemma} \label{hamming_lema} If $G$ is a connected graph, then $G$ is a partial Hamming graph if and only if all the quotient graphs with respect to $\Theta^*$-classes of $G$ are complete. \end{lemma} \begin{proof} The backward implication is obvious and it follows from the definition of the canonical isometric embedding. Let $G$ be a partial Hamming graph and let $\gamma : G \rightarrow \prod_{j=1}^m {K_{i_j}}$ be an isometric embedding into the Cartesian product of complete graphs $K_{i_j}$. Discard all the factors which contain only one vertex and the unused vertices in each factor. We get an irredundant embedding of a graph $G$ into a product of complete graphs. Therefore, by Theorem 5.3 from \cite{klavzar1993}, this embedding is the canonical isometric embedding, so the quotient graphs of $G$ with respect to $\Theta^*$-classes of $G$ are all complete. \hfill $\square$ \bigskip \end{proof} We are ready to prove the main theorem of this section, which generalizes a similar result from \cite{klavzar-2006}. In the rest of the section, the $\Theta^*$-classes of a connected graph $G$ will be denoted by $E_1, \ldots, E_k$, $k \in \mathbb{N}$. Moreover, for any $i \in \lbrace 1,\ldots, k \rbrace$, we denote the connected components of the graph $G \setminus E_i$ by $C_{i}^1, \ldots, C_{i}^{r_i}$. \begin{theorem} \label{hamming_graf_izrek} If $(G,w)$ is a connected weighted graph, then $$W(G,w) \geq \frac{1}{2}\sum_{i=1}^k \sum_{j=1}^{r_i} w(C_i^j)w^c(C_i^j),$$ where $w(C_i^j) = \sum_{x \in V(C_i^j)}w(x)$ and $w^c(C_i^j) = \sum_{x \in V(G) \setminus V(C_i^j)}w(x)$ for any $i \in \lbrace 1,\ldots,k \rbrace, j \in \lbrace 1, \ldots, r_i\rbrace$. Moreover, the equality holds if and only if $G$ is a partial Hamming graph. \end{theorem} \begin{proof} Let $\alpha = (\alpha_1, \ldots, \alpha_k)$ be the canonical isometric embedding for graph $G$. By the definition and Lemma \ref{distance} we have \begin{eqnarray*} W(G,w) & = & \sum_{\lbrace u,v \rbrace \subseteq V(G)} w(u)w(v)d_G(u,v) \\ & = & \sum_{\lbrace u,v \rbrace \subseteq V(G)} w(u)w(v) \left( \sum_{i=1}^k d_{G / E_i}(\alpha_i(u),\alpha_i(v)) \right) \\ & = & \sum_{i=1}^k \sum_{\lbrace u,v \rbrace \subseteq V(G)} w(u)w(v)d_{G / E_i}(\alpha_i(u),\alpha_i(v)). \end{eqnarray*} \noindent In the rest of the proof, we introduce notation $$w^c(u) =\sum_{\substack{x \in V(G) \\ \alpha_i(u) \neq \alpha_i(x)}}w(x)$$ for any $u \in V(G)$. Since $d_{G / E_i}(\alpha_i(u),\alpha_i(v)) \geq 1$ for all $u,v \in V(G)$ with $\alpha_i(u) \neq\alpha_i(v)$, $i \in \lbrace 1,\ldots, k \rbrace$, we deduce \begin{eqnarray*} W(G,w) & \geq & \sum_{i=1}^k \sum_{\substack{\lbrace u,v \rbrace \subseteq V(G) \\ \alpha_i(u) \neq \alpha_i(v)}} w(u)w(v) \\ & = & \frac{1}{2}\sum_{i=1}^k \sum_{u \in V(G)} w(u)w^c(u) \\ & = & \frac{1}{2}\sum_{i=1}^k \sum_{j=1}^{r_i} \sum_{u \in C_i^j} w(u)w^c(u). \end{eqnarray*} \noindent Obviously, for any $u,v \in V(C_i^j)$ it holds $w^c(u)=w^c(v)$. Hence, \begin{eqnarray*} W(G,w) & \geq & \frac{1}{2}\sum_{i=1}^k \sum_{j=1}^{r_i} w(C_i^j)w^c(C_i^j). \end{eqnarray*} Obviously, the equality holds if and only if $d_{G / E_i}(\alpha_i(u),\alpha_i(v)) = 1$ for all $u,v \in V(G)$ with $\alpha_i(u) \neq\alpha_i(v)$, $i \in \lbrace 1,\ldots, k \rbrace$, which is satisfied when all the quotient graphs with respect to $\Theta^*$ relation are complete graphs. By Lemma \ref{hamming_lema}, this is fulfilled if and only if $G$ is a partial Hamming graph. \hfill $\square$ \bigskip \end{proof} \noindent It is worth mentioning that the previous result can also be derived from Theorem \ref{rac_wie} and Lemma \ref{hamming_lema}. \begin{corollary} \label{hamming_graf_posl} If $G$ is a connected graph, then $$Gut(G) \geq \frac{1}{2}\sum_{i=1}^k \sum_{j=1}^{r_i} {\rm deg}(C_i^j){\rm deg}^c(C_i^j),$$ where ${\rm deg}(C_i^j) = \sum_{x \in V(C_i^j)}{\rm deg}(x)$ and ${\rm deg}^c(C_i^j) = \sum_{x \in V(G) \setminus V(C_i^j)}{\rm deg}(x)$ for any $i \in \lbrace 1,\ldots,k \rbrace, j \in \lbrace 1, \ldots, r_i\rbrace$. Moreover, the equality holds if and only if $G$ is a partial Hamming graph. \end{corollary} For an example, consider the family of graphs $H_n, n \geq 2$, where $n$ denotes the number of inner faces, see Figure \ref{hiska}. It is clear that $|V(H_n)|= 2n+1$ holds for any $n \geq 2$. Figure \ref{hiska} also illustrates all $n$ graphs $H_n$ with removed $\Theta^*$-classes and in Figure \ref{quotientHn} we can see their quotient graphs. \begin{figure}[h] \begin{center} \includegraphics[scale=0.8]{hiska.eps} \end{center} \caption{\label{hiska} Graph $H_n$ and its subgraphs $H_n \setminus E_i, i \in \{ 1,2,...,n\}$.} \end{figure} \begin{figure}[h] \begin{center} \includegraphics[scale=0.8]{quotientHn.eps} \end{center} \caption{\label{quotientHn} Quotient graphs $H_n / E_i, i \in \{ 1,2,...,n\}$.} \end{figure} All the quotient graphs with respect to $\Theta^*$ relation are complete, and therefore, by Lemma \ref{hamming_lema}, $H_n$ is a partial Hamming graph for any $n \geq 2$. Moreover, using Corollary \ref{hamming_graf_posl}, we can calculate the closed formula for the Gutman index of $H_n$ in the following way. \begin{eqnarray*} Gut(H_n) & = & \frac{1}{2}\sum_{i=1}^n \sum_{j=1}^{r_i} {\rm deg}(C_i^j){\rm deg}^c(C_i^j)\\ & = & \frac{1}{2} \big[ (3n-1)(3n+1) + (3n-1)(3n+1) + 2(6n-2) \big]\\ & + & \sum_{i=1}^{n-1} \big[ (2+6i)(6(n-i-1)+4) \big] \\ & = & 6n^3 + 9n^2 - 4n + 1. \end{eqnarray*} \section*{Acknowledgment} \noindent The author Niko Tratnik acknowledge the financial support from the Slovenian Research Agency (research core funding No. P1-0297 and J1-9109).

025b316eb906ca217e2bda3c79967e0e5a8ef1b0

\subsection{Security of multi-party protocols in the semi-honest model}\label{subsec:appendix_security} Definition~\ref{def:view} and Definition~\ref{def:coalition_view} state that the view of any of the parties participating in the protocol, on each possible set of inputs, can be simulated based only on its own input and output. The intuition behind these definitions can also be viewed as follows: consider a \mbox{black-box} protocol that receives the inputs of Problem~\eqref{eq:problem} and solves it, not exchanging any messages in order to achieve the solution. Such a \mbox{black-box} protocol satisfies Definitions~\ref{def:view} and~\ref{def:coalition_view}. In other words, meeting the requirements of the above definitions is equivalent to not revealing any other information than what is already known to any of the parties, i.e. inputs, prescribed outputs, if any, and previously known side information, meaning details about the optimization algorithm. \begin{remark}\label{remark:aux}(\cite[Ch.~7]{Goldreich04foundations||},\cite[Ch.~4]{Goldreich03foundations|}) \textbf{Auxiliary inputs}, which are inputs that capture additional information available to each of the parties (e.g. local configurations, side-information), are implicit in Definition~\ref{def:view} and Definition~\ref{def:coalition_view}. \end{remark} \begin{remark}~\label{remark:leak} Satisfying Definitions~\ref{def:view} and~\ref{def:coalition_view} is a stronger requirement than guaranteeing that an adversary cannot uniquely retrieve the data of the honest parties, i.e. Definition~\ref{def:non-unique}. \end{remark} Revealing sensitive information does not always lead to a unique retrieval of the private data. Nevertheless, any piece of information revealed by the execution of the protocol, that cannot be obtained only from its inputs and outputs, leads to the violation of Definitions~\ref{def:view},~\ref{def:coalition_view}, even if the private data cannot be singled out with this information. \subsection{Paillier cryptosystem}\label{subsec:appendix_paillier} In this part, we will provide the rest of the details regarding the implementation of the Paillier cryptosystem, introduced in Section~\ref{sec:encryption_sch}. The pair of keys corresponding to this cryptosystem is $(pk,sk)$, where the public key is $pk = (N,g)$ and the secret key is $sk = (\gamma,\delta)$. $N$ is called the modulus and is the product of two large prime numbers $p,q$, and $g$ is an element of order $N$ in the multiplicative group $\mathbb Z^\ast_{N^2}$. Furthermore, \[\gamma = lcm(p-1,q-1),~\delta = ((g^\gamma~\text{mod}~N^2 -1)/N)^{-1}~\text{mod}~N.\] A simpler variant to generate the pair of keys when $p$ and $q$ have the same number of bits is to choose $g = N+1$, $\gamma = \phi(N) = (p-1)(q-1)$, where $\phi$ is Euler's totient function and $\delta = \phi(N)^{-1}\bmod N$. For a plaintext $a\in \mathbb Z_N$, the Paillier encryption is: \begin{equation} [[a]] = g^ar^N\bmod N^2, \end{equation} where $r$ is a random non-zero element in $\mathbb Z_N$, which makes Paillier a probabilistic encryption scheme. Therefore, the ciphertexts do not preserve the order relation between the plaintexts. The decryption primitive is the following: \begin{equation} a = (([[a]]^\gamma\bmod N^2-1)/N)\delta\bmod N, \end{equation} which uses the fact that $(1+N)^a = 1+Na \bmod N^2$. In the Paillier scheme, in order to obtain addition between plaintexts, the operation between ciphertexts is modular multiplication, which was denoted by $\oplus$ in the text: \begin{align} \begin{split} [[a]]\cdot[[b]] &= g^ar^N\bmod N^2 \cdot g^br'^N\bmod N^2 \\ &= g^{a+b}(rr')^N\bmod N^2 \\ &= [[a + b]]\bmod N^2. \end{split} \end{align} Negation is achieved by modular inverse: \begin{align} \begin{split} [[a]]^{-1} &= g^{-a}(r^{-1})^N\bmod N^2 \\ &= [[-a]]\bmod N^2. \end{split} \end{align} The multiplication between a plaintext value $c$ and an encrypted value $[[a]]$, which was denoted by $c\otimes[[a]]$ in the text, is obtained in the following way: \begin{equation} [[a]]^{c} = g^{ac}(r^c)^N\bmod N^2 = [[ca]]. \end{equation} \begin{remark} The Paillier cryptosystem is malleable, which means that a party that does not have the private key can alter a ciphertext such that another valid ciphertext is obtained. Malleability is a desirable property in order to achieve cloud-outsourced computation on encrypted data, but allows ciphertext attacks. In this work, we assume that the parties have access to authenticated channels, therefore an adversary cannot alter the messages sent by the honest parties. \end{remark} \subsection{Semantic security of a cryptosystem} Proving security in the \mbox{semi-honest} model involves the concepts of semantic security of an encryption scheme. Under the assumptions of decisional composite residuosity~\cite{Paillier99} and hardness of factoring, the partially homomorphic encryption schemes Paillier and DGK are semantically secure and have indistinguishable encryptions, which, in essence, means that an adversary that has the plaintext messages $a$ and $b$ cannot distinguish between the encryption $[[a]]$ and encryption $[[b]]$. \begin{definition}\label{def:semantic}(Semantic Security~\cite[Ch.~5]{Goldreich04foundations||}) An encryption scheme is \textbf{semantically secure} if for every probabilistic polynomial-time algorithm, A, there exists a probabilistic polynomial-time algorithm $A'$ such that for every two polynomially bounded functions $f,h:\{0,1\}^\ast\rightarrow \{0,1\}^\ast$ and for any probability ensemble $\{X_n\}_{n\in\mathbb N}$, $|X_n|\leq poly(n)$, for any positive polynomial $p$ and sufficiently large $n$: \begin{align*} \text{Pr}& \left[ A(E(X_n),h(X_n),1^n) = f(X_n)\right] < \\ \text{Pr}& \left[ A'(h(X_n),1^n) = f(X_n)\right] + 1/p(n), \end{align*} where $E(\cdot)$ is the encryption primitive. \end{definition} The security parameter $n$ is given as input in the form $1^n$ to $A$ and $A'$ to ensure they can run in $\text{poly}(n)$-time. \subsection{Additively Homomorphic Encryption}\label{subsec:ahe} Let $E(\cdot)$ define a generic encryption primitive, with domain the space of private data, called plaintexts, and codomain the space of encrypted data, called ciphertexts. For probabilistic encryption schemes, the encryption primitive also takes as input a random number. The decryption primitive $D(\cdot)$ is defined on the space of ciphertexts and takes values in the space of plaintexts. Additively homomorphic schemes satisfy the property that there exists an operator $\oplus$ defined on the space of ciphertexts such that: \begin{equation}~\label{eq:abstract_PHE} E(a)\oplus E(b) \subset E(a+b), \end{equation} for any plaintexts $a,b$ supported by the scheme. Here, we use set inclusion instead of equality because the encryption of a message is not unique in probabilistic cryptosystems. Intuitively, equation~\eqref{eq:abstract_PHE} means that performing this operation on the two encrypted messages, we obtain a ciphertext that is equivalent to the encryption of the sum of the two plaintexts. Formally, the decryption primitive $D(\cdot)$ is a homomorphism between the group of ciphertexts with the operator $\oplus$ and the group of plaintexts with addition $+$, which justifies the name of the scheme. It is immediate to see that if a scheme supports addition between encrypted messages, it will also support subtraction, by adding the additive inverse, and multiplication between an integer plaintext and an encrypted message, obtained by adding the encrypted messages for the corresponding number of times. In this paper, we use the popular Paillier encryption, but any other additively homomorphic encryption scheme can be employed. The \textbf{Paillier cryptosystem}~\cite{Paillier99} is an asymmetric additively homomorphic encryption scheme. Asymmetric or public key cryptosystems involve a pair of keys: a public key that is disseminated publicly, and which is used for the encryption of the private messages, and a private key which is known only to its owner, used for the decryption of the encrypted messages. In this section, we will only present the notation for the encryption and for the allowed operations between the encrypted messages and postpone the implementation details to Appendix~\ref{subsec:appendix_paillier}. In the Paillier encryption scheme, the private data are elements of the ring of integers modulo $N$, denoted by $\mathbb Z_N$, where $N$ is a large integer of $\sigma$ bits, called the Paillier modulus. The encrypted messages take values in the positive integers modulo $N^2$, denoted by $\mathbb Z^\ast_{N^2}$. Paillier is a probabilistic encryption scheme, which means that the encryption primitive also takes as an argument a random number which, for simplicity, we avoid in this notation. For a plaintext message $a$, we denote the Paillier encryption by $[[a]]:=E(a)$, for an instance of the random number. For readability, we will use throughout the paper the following abstract notation for the operations on the encrypted space: \begin{align}\label{eqref:hom_prop} \begin{split} [[a]]\oplus[[b]] &\stackrel{d}{=} [[a+b]]\\ b\otimes[[a]] &\stackrel{d}{=} [[ba]],~~ \text{ for plaintexts } a,b\in\mathbb Z_N, \end{split} \end{align} where $\stackrel{d}{=}$ means that the equality holds after applying the decryption primitive on both sides~\cite{Paillier99}. The expressions for the abstracted operations are described in Appendix~\ref{subsec:appendix_paillier}. We will slightly abuse these notations to denote encryptions of vectors and additions and multiplication by vectors and matrices. In our protocol, the target node $\mathcal T$ is the owner of a pair of Paillier keys. Hence, for everything that follows, $[[\cdot]]$ denotes the encryption with the target node's public key $pk_{\mc T}$. \emph{Example: Private unconstrained quadratic optimization.} In order to illustrate the usage of the Paillier encryption scheme for computations on encrypted data, consider the following example of an unconstrained optimization problem: \begin{equation}~\label{eq:unconstrained} \min\limits_{x} f(x) = \frac{1}{2}x^\intercal Q_{\mc C} x + c_{\mc A}^\intercal x, \end{equation} where the variables have the same meaning as in Problem~\eqref{eq:problem} described in Section~\ref{sec:problem_setup}. The cloud has access to the matrix $Q_{\mc C}$. If the cloud also had access to private data $c_{\mc A}$ from the agents, the gradient method could be employed to obtain the optimal unencrypted solution of Problem~\eqref{eq:unconstrained}. An iteration of the gradient method has the form: \[x_{k+1} = x_k - \eta\nabla f(x_k) = x_k - \eta Q_{\mc C} x_k - \eta c_{\mc A},\] where $\eta>0$ is the step-size chosen by the cloud and $k=0,\ldots,K-1$ for a number of iterations $K$. However, in order to protect their private data and the privacy of the solution $x^\ast$, the agents send the value $c_{\mc A}$ encrypted to the cloud as $[[c_{\mc A}]]$. It is very important to notice that \textit{for a quadratic objective function, only linear operations in the private data $x_k$ and $c_{\mc A}$ are required in order to compute one iteration of the gradient descent algorithm}. Hence, by taking advantage of the additively homomorphic property~\eqref{eqref:hom_prop}, the cloud can locally compute the optimal solution by the gradient descent in the encrypted domain, for all $k = 0,\ldots,K-1$: \begin{equation}\label{eq:unconstrained_iter} [[x_{k+1}]] = (I-\eta Q_{\mc C})\otimes[[x_k]]\oplus(-\eta)\otimes [[c_{\mc A}]], \end{equation} and send the final point $[[x_K]]$ to the target node to decrypt and obtain the desired result. Such a protocol satisfies the desired security according to Definition~\ref{def:view} provided in Section~\ref{sec:privacy_goals} by the fact that the cloud only has access to data encrypted by a semantically secure encryption scheme, such as Paillier~\cite{Paillier99}. More details are provided in Section~\ref{sec:security_proof}. However, the optimization problem considered in this paper is a quadratic problem with constraints, which introduce nonlinear operations in the optimization algorithm and cannot be handled inherently by the Paillier cryptosystem. We will describe in Section~\ref{sec:gradient_ascent} how we leverage communication between the parties in order to achieve privacy-preserving nonlinear operations. \subsection{Symmetric encryption scheme}\label{subsec:symmetric} Apart from the (public key) partially homomorphic cryptosystem above, we will also use a specific symmetric key cryptosystem for reasons that will become apparent in the protocol, related to hiding the data from the target node. This symmetric key cryptosystem is similar to the well-known one-time pad~\cite{Vernam1926,Bellovin11}, and can be thought of as additively blinding a private value by noise. In contrast to public key cryptography, symmetric key algorithms perform the encryption and decryption with the same key. For messages of $l$ bits, the key is generated as $sk\in \mathbb Z$ with length of $\lambda+l$ bits, where $\lambda$ is the security parameter. The encryption primitive is $E'(a) = a + sk$, with $a\in [0,2^l)\cap \mathbb Z$, and the decryption is obtained as $a = D'(E'(a)) = E'(a) - sk$. This is also called two-out-of-two additive secret sharing~\cite{Dodis2007lecture} when the addition is on an abelian group. The security of this scheme lies on generating a uniformly random key and on using this key for encryption only once, which yields that the distribution of $E'(a)$ is statistically indistinguishable from a random number sampled of $l+\lambda+1$ bits. For this reason, $\lambda$ is called the statistical security parameter. We also notice that \textit{this symmetric cryptosystem is compatible with the Paillier cryptosystem, in the sense that the two encryptions commute}: $D'(E(E'(a)))) = E(a)$ by using $E(sk)$ instead of $sk$ for decryption, where $E'(a)$ is performed on the message space of the Paillier cryptosystem. Such a scheme is commonly employed for blinding messages, for instance in~\cite{Veugen10,Jeckmans13,Bost15}. For simplicity, in the protocols described in Section~\ref{sec:gradient_ascent}, we will directly use the summation with a random number to depict the above symmetric key cryptosystem, respectively the difference by the same random number for decryption. The strength of a cryptosystem relies on the computational intractability of retrieving the private key from the public information -- an adversary holding the public information cannot find the private key by brute force computations. Public key cryptosystems like the Paillier cryptosystem involve a security parameter $\sigma$, which is usually the length of the modulus of the public key, and symmetric key cryptosystems involve a security parameter $\lambda$, which is the bit-length of the key. As is common in practice, increasing the size of the security parameters to values larger than, say 1024 bits for $\sigma$ and 80 bits for $\lambda$, increases the security of the system against any computationally efficient inference attack. \subsection{Projected gradient ascent on encrypted data}\label{sec:encrypted_gradient} As stated in Section~\ref{sec:problem_setup}, we aim to solve an optimization problem outsourced to the cloud on private distributed data from the agents and send the result to the target node. To protect the agents' data, we use an encryption scheme that allows the cloud to perform linear manipulations on encrypted data, as described in Section~\ref{sec:encryption_sch}. To this end, the target node generates a pair of keys $(pk_{\mc T}, sk_{\mc T})$ and distributes the public key to the agents and the cloud, enabling them to encrypt their data, which only the target node will be able to decrypt, using the private key. We consider that all the data is represented on integers of~$l$ bits and comment on this further in Section~\ref{subsec:quantization}. The main difficulty in performing the projected gradient ascent on encrypted data is performing iteration~\eqref{eq:projection}. We have already seen in the example in Section~\ref{subsec:ahe} that the update of the iterate in the direction of the gradient ascent can be computed locally by the cloud directly on the encrypted data~\eqref{eq:unconstrained_iter}. However a first challenge lies in performing the comparison with zero. Due to the probabilistic nature of the Paillier encryption scheme, \textit{the order on the plaintext space is not preserved on the ciphertext space} and comparison on encrypted data cannot be performed locally by the cloud. Moreover, after the comparison is performed, the update of the encrypted iterate~\eqref{eq:projection} has to be done in a private way, so that the result of the maximum operation is not revealed to any of the parties involved (the cloud and the target node). These two steps are the main computational bottleneck in the protocol we propose, as both require secure communication between the cloud and the target node. A preliminary version of this solution was presented in~\cite{Alexandru17}. We can privately achieve the two steps mentioned above in three stages. First, the cloud has to randomize the order of the two encrypted variables it wants to compare (Protocol~\ref{alg:rand_step}). Second, the cloud and target engage in an interactive comparison protocol that takes as inputs the two randomized variables and outputs the result of the comparison to the target node (Protocol~\ref{alg:comp_DGKV}). Third, the update of the dual iterate is achieved through an interactive protocol between the cloud and target node, which takes as inputs the two randomized variables and the result of the comparison and outputs to the cloud the updated iterate (Protocol~\ref{alg:update_step}). Throughout this paper, by comparison we mean element-wise comparison, since the variable $\mu$ is a vector. \subsection{Secure comparison protocol}\label{subsec:comparison} In order to privately compute~\eqref{eq:projection}, i.e., hide the result from \textit{all} the parties involved, we want to keep the result of the comparison of the updated iterate $\mu_k+\eta\nabla g(\mu_k)$ with zero unknown to both the cloud and the target node. The comparison protocol will reveal the result of the comparison between the two inputs to the target node $\mathcal T$. However, if we introduce an additional step where $\mathcal C$ randomizes the order of the two values that it wants to compare, then $\mathcal T$ does not learn any information by knowing the result of the comparison. \begin{protocol}[Randomization step] \label{alg:rand_step} \begin{algorithmic}[1] \small \Require{$\mathcal C$: $[[\bar \mu]],[[0]]$, where $\bar\mu:= \mu + \eta\nabla g(\mu)$} \Ensure{$\mathcal C$: $[[a]],[[b]]$} \State $\mathcal C$: choose a random permutation $\pi$ on two elements \State $\mathcal C$: output $[[a]],[[b]]\leftarrow \pi([[0]],[[\bar\mu]])$ \end{algorithmic} \end{protocol} Next, we will demonstrate how the comparison protocol works via the following instantiation. Damg{\aa}rd, Geisler and Kr{\o}igaard introduced a protocol in~\cite{DGK07,DGK09correction} for secure comparison of two private inputs of different parties, which we will call the DGK protocol. To this end, they also propose an additively homomorphic encryption scheme, which has the property that checking if the value zero is encrypted is more efficient than simply decrypting, which is useful for comparisons and working with bits. An extension of this protocol to the case where none of the parties knows the two values to be compared, which is of interest to us, and some improvements in terms of efficiency were proposed in~\cite{Veugen12}. \begin{protocol}[Protocol for secure \mbox{two-party} comparison with two encrypted inputs using DGK \cite{DGK07,Veugen12}] \label{alg:comp_DGKV} \begin{algorithmic}[1] \small \Require{$\mathcal C$: $[[a]],[[b]]$; $\mathcal T$: $sk_{\mc T},sk_{DGK}$} \Ensure{$\mathcal T$: bit $t$: $(t=1)\Leftrightarrow (a \leq b)$} \State $\mathcal C$: choose random number $\rho$ \State $\mathcal C$: $[[z]] \gets [[b]]\oplus(-1)\otimes[[a]]\oplus[[2^l+\rho]]$, send $[[z]]$ to $\mathcal T$ \Statex \Comment{$z\gets b-a+2^l +\rho$} \State $\mathcal T$: decrypt $[[z]]$ \State $\mathcal C$: $\alpha\gets \rho\mod 2^l$ \State $\mathcal T$: $\beta\gets z\mod 2^l$ \State $\mathcal C$,$\mathcal T$: perform a comparison protocol, e.g, DGK, such that $\mathcal C$ gets $\delta_{\mc C}$ and $\mathcal T$ gets $\delta_{\mc T}$ with $\delta_{\mc C}\veebar\delta_{\mc T} = (\alpha\leq\beta)$ \Comment $\veebar$ denotes the exclusive or operation \State $\mathcal T$: encrypt $[[z\div 2^l]]$ and $[[\delta_{\mc T}]]$ and send them to $\mathcal C$ \State $\mathcal C$: $[[t']] \leftarrow \begin{cases} [[\delta_{\mc T}]] &\text{ if } \delta_{\mc C} = 1 \\ [[1]]\oplus(-1)\otimes[[\delta_{\mc T}]] &\text{ otherwise} \end{cases}$ \State $\mathcal C$: $[[t]]\gets [[z\div 2^l]]\oplus(-1)\otimes ([[\rho\div 2^l]]\oplus [[t']])$ \Statex \Comment{$t \gets z\div 2^l - \rho\div 2^l - t'$} \State $\mathcal C$: send $[[t]]$ to $\mathcal T$ \State $\mathcal T$: decrypts $[[t]]$ \end{algorithmic} \end{protocol} The comparison protocol that will be used in our optimization protocol is as follows. Let $\mathcal C$ have two encrypted values under the Paillier scheme $[[a]]$ and $[[b]]$ that it obtained after running Protocol~\ref{alg:rand_step}, and let $\mathcal T$ have the decryption key. Furthermore, let $\mathcal T$ also have the decryption key of the DGK homomorphic encryption scheme, which we describe in Appendix~\ref{app:comparison}. At the end of the protocol, $\mathcal T$ will have the result of the comparison in the form of one bit $t$ such that $(t=1)\Leftrightarrow (a\leq b)$. Let $l$ denote the number of bits of the unencrypted inputs $a,b$. Protocol~\ref{alg:comp_DGKV} is based on the fact that the most significant bit of $(b-a+2^l)$ is the bit that indicates if $(a\leq b)$. The security of this protocol is proved in \cite{DGK07,DGK09correction,Veugen12} and the proof is omitted here. \begin{proposition}\label{prop:comp_DGK} Protocol~\ref{alg:comp_DGKV} is secure in the \mbox{semi-honest} model, according to Definition~\ref{def:view}. \end{proposition} \subsection{Secure update protocol}\label{subsec:update} Moreover, we need to ensure that when the cloud $\mathcal C$ updates the value of the dual iterate at iteration $k+1$ in equation~\eqref{eq:projection}, it does not know the new value. The solution is to make the cloud blind the values of $[[a]]$ and $[[b]]$ and send them to the target node in this order, where the latter selects the value accordingly to the comparison result and then sends it back to the cloud. However, there are two important issues that have to be addressed in order for the update step to not leak information about the sign of the iterate: the blinding should be additive and effectuated with different random values, and the ciphertexts should be refreshed. The reasons are the following: if the blinding is multiplicative, by decrypting the product, the target knows which one of the values is zero. Moreover, if the two values are additively blinded with the same random value, the target can subtract them and reveal at least if the value is zero. Re-randomization of the encryptions is necessary so that the cloud cannot simply compare $[[a]]$ and $[[b]]$ with the received value. This can be done by adding an encryption of zero or by decryption followed by encryption. Protocol~\ref{alg:update_step} is the solution to the update problem: \begin{protocol}[Secure update of the dual variable] \label{alg:update_step} \begin{algorithmic}[1] \small \Require{$\mathcal C$: $[[a]],[[b]]$; $\mathcal T$: $t$ such that $(t=1)\Leftrightarrow(a\leq b)$} \Ensure{$\mathcal C$: $[[\mu]]$} \State $\mathcal C$: choose two random numbers $r,s$ \State $\mathcal C$: $[[\bar a]] \leftarrow [[a]]\oplus [[r]], [[\bar b]] \leftarrow [[b]]\oplus [[s]]$ \State $\mathcal C$: send $[[\bar a]]$ and $[[\bar b]]$ to $\mathcal T$ \If{$t=0$} {~$\mathcal T$: $[[v]]\leftarrow \widetilde{[[\bar a]]} = [[\bar a]] + [[0]]$}\Else {~$\mathcal T$: $[[v]]\leftarrow \widetilde{[[\bar b]]}=[[\bar b]] + [[0]]$}\EndIf \Comment Refresh the ciphertext \State $\mathcal T$: send $[[v]]$ and $[[t]]$ to $\mathcal C$ \State $\mathcal C$: $[[\mu]]\gets [[v]]\oplus r\otimes[[t]]\oplus [[-r]] \oplus (-s)\otimes [[t]]$ \Statex \Comment{$\mu\gets v + r(t-1) - st$} \end{algorithmic} \end{protocol} The intuition behind the protocol is that if $a\leq b$, then $t=1$ and we obtain $\mu = \bar b - s = b$, and otherwise, $t=0$ and we obtain $\mu = \bar a - r = a$. In both cases the cloud correctly updates the dual variable with the projected value. \subsection{Protocol for solving strictly-convex quadratic problems}\label{subsec:protocol} Having defined these protocols, we can now build a protocol that represents one iteration \eqref{eq:projection} of the dual projected gradient ascent method. \begin{protocol}[Secure iteration of the dual projected gradient ascent method] \label{alg:iteration} \begin{algorithmic}[1] \small \Require{$\mathcal C$: $A_{\mc C}\in \mathbb R^{m\times n}, Q_{\mc C}\in \mathbb S^n_{++}, [[b_{\mc A}]], [[c_{\mc A}]],\eta > 0, [[\mu_k]]$; $\mathcal T$: $sk_{\mc T} $} \Ensure{$\mathcal C$: $[[\mu_{k+1}]]$} \State $\mathcal C$: $[[\nabla g(\mu_k)]] \gets (-A_{\mc C} Q_{\mc C}^{-1} A_{\mc C}^\intercal)\otimes[[\mu_k]] \oplus (-A_{\mc C} Q_{\mc C}^{-1})\otimes[[c_{\mc A}]]\oplus (-1)\otimes [[b_{\mc A}]]$ \Statex \Comment Compute the encrypted gradient as in~\eqref{eq:grad} \State $\mathcal C$: $[[\bar \mu_k]] \gets [[\mu_k]] \oplus \eta\otimes [[\nabla g(\mu_k)]]$ \Comment Update the value in the ascent direction \State $\mathcal C, \mathcal T$ truncate $[[\bar \mu_k]]$ to $l$ bits \State $\mathcal C$ execute Protocol~\ref{alg:rand_step}: $\mathcal C$ gets $[[a_k]],[[b_k]]$ \Comment Randomly assign $[[a_k]],[[b_k]]$ with values of $[[\bar \mu_k]], [[0]]$ \State $\mathcal C, \mathcal T$ execute Protocol~\ref{alg:comp_DGKV} element-wise on inputs $[[a_k]],[[b_k]]$: $\mathcal T$ gets $t_k$ \Comment Secure comparison protocol \State $\mathcal C, \mathcal T$ execute Protocol~\ref{alg:update_step}: $\mathcal C$ obtains $[[\mu_{k+1}]]$ \Comment Secure update protocol that ensures $\mu_{k+1} = \max\{\bar \mu_k, 0\}$ \end{algorithmic} \end{protocol} Line 3 ensures that the updated iterate has the required number of bits for the comparison protocol. This step is achieved by an exchange between the cloud and target node: the cloud additively blinds the iterate by a random number, sends it to the target node, which decrypts and truncates the sum and sends it back, where the cloud then subtracts the truncated random number. The proof of security in the \mbox{semi-honest} model follows similar steps as in the argmax protocol in~\cite{Bost15} and we will address it in Appendix~\ref{subsec:proof_thm1}. \begin{proposition}\label{prop:iteration} Protocol~\ref{alg:iteration} is secure in the \mbox{semi-honest} model, according to Definition~\ref{def:view}. \end{proposition} Using the building blocks described above, we can finally assemble the protocol that privately solves the constrained quadratic optimization problem \eqref{eq:problem} with private data and sends the optimal solution to the target node. The public key $pk_{\mc T}$ and bit-length $l$ are known by all the parties, hence we omit them from the inputs. \begin{protocol}[Privacy preserving algorithm for solving strictly-convex quadratic optimization problems] \label{alg:main_alg} \begin{algorithmic}[1] \small \Require{$\mathcal A_{i=1,\ldots,p}$: $b_{\mc A}= \{b_j\}_{j=1,\ldots,m}, c_{\mc A}= \{c_j\}_{j=1,\ldots,n}$; $\mathcal C$: $A_{\mc C}\in \mathbb R^{m\times n}, Q_{\mc C}\in \mathbb S^n_{++},K$; $\mathcal T$: $sk_{\mc T},K $} \Ensure{$\mathcal T$: $x^\ast$} \For{i=1,\ldots,p} \State $\mathcal A_i:$ encrypt the private information $msg_i \gets$ $ ([[b_i]]$, $[[c_i]])$ \State $\mathcal A_i:$ send the encrypted messages to $\mathcal C$ \EndFor \State $\mathcal C$: Construct the vectors $[[b_{\mc A}]]$ and $[[c_{\mc A}]]$ from the messages \State $\mathcal C$: $\eta \gets 1/\lambda_{max}(A_{\mc C} Q_{\mc C}^{-1}A_{\mc C}^\intercal)$ \State $\mathcal C$: Choose a random positive initial value $\mu_0$ for the dual variable and encrypt it: $[[\mu_0]]$ \For {each $k=0,\ldots,K-1$} \State $\mathcal C,\mathcal T$ execute Protocol~\ref{alg:iteration}: $\mathcal C$ gets $[[\mu_{k+1}]]$ \Comment{$\mathcal C,\mathcal T$ securely effectuate an iteration of the dual projected gradient ascent} \EndFor \State $\mathcal C$: $[[x^\ast]] \gets (-Q_{\mc C}^{-1} A_{\mc C} ^\intercal)\otimes[[\mu_K]] \oplus (-Q_{\mc C}^{-1})\otimes[[c_{\mc A}]]$ and send it to $\mathcal T$ \Comment Compute the primal optimum from the optimal dual solution as in~\eqref{eq:KKT1} \State $\mathcal T$: Decrypt $[[x^\ast]]$ and output $x^\ast$ \end{algorithmic} \end{protocol} \subsection{Fixed-point arithmetic}~\label{subsec:quantization} The optimization problem~\eqref{eq:problem} is defined on real variables, whereas the Paillier encryption scheme is defined on integers. To address this issue, we adopt a \mbox{fixed-point} arithmetic setting, where we allow for a number to have a fixed number of fractional bits. First, we consider numbers that have the magnitude between $-2^{l_i-1}< x < 2^{l_i-1}$. Second, we consider a value having $l_i$ bits for the integer part and $l_f$ bits for the fractional part. Therefore, by multiplying the real values by $2^{l_f}$ and truncating the result, we obtain integers. We choose $l=l_i+l_f$ large enough such that the loss in accuracy is negligible and assume that there is no overflow. For ease of exposition, we consider this data processing done implicitly in the protocols described. The random numbers used for blinding the sensitive values (in Protocols~\ref{alg:comp_DGKV} and~\ref{alg:update_step}) are sampled uniformly from the integers in $\mathbb Z_N$ of $l+\lambda$ bits, where $\lambda$ is the statistical security parameter, as already explained in Section~\ref{sec:problem_setup}, chosen such that brute-forcing the solution is intractable. In order to guarantee correctness of the comparison protocol, no overflow must take place, so we must impose $log_2 N > l+ \lambda + 1$. The errors in the solution caused by the fixed-point arithmetic operations necessary for the encryption can be analyzed with the same tools as in~\cite{Jerez14,Rubagotti2016real,Alexandru18}. The round-off errors can be regarded as states in a stable dynamical system with bounded disturbances, and hence, have a bounded norm that offers a guide on how to choose the number of fractional bits $l_f$ for the fixed-point representation. On the other hand, the overflow and quantization errors depend on the magnitude of the dual iterates. We considered feasible and bounded problems -- the dual problem~\eqref{eq:dual} has a finite solution -- therefore, one can select the number of integer bits $l_i$ in the representation such that no overflow occurs. \subsection{Related work} Homomorphic encryption has been recently used to design encrypted controllers, with partially homomorphic encryption schemes for linear controllers in~\cite{Kogiso15,Farokhi17}, model predictive control applications in~\cite{Darup18,Alexandru18}, with fully homomorphic encryption in~\cite{Kim16}, and to design observers in~\cite{Alanwar17}. Unconstrained optimization problems are commonly used in machine learning applications, e.g., for linear regression, and several works have addressed gradient methods with partially homomorphic encryption~\cite{Han10grad,Hardy17} or with ADMM in a distributed setting~\cite{Zhang17privacy}. However, adding constraints substantially complicates the optimization problem by restricting the feasible domain. In addition, constraints introduce nonlinear operations in the optimization algorithm, which cannot be handled inherently by the partially homomorphic cryptosystems. In our prior work~\cite{Shoukry16,Alexandru17}, we showed how to overcome these difficulties by using blinded communication in a centralized setup. Recently,~\cite{Lu18} proposed a distributed approach to the constrained optimization problem with homomorphic encryption, where each party locally performs the projection on the feasible domain on unencrypted data. Similarly,~\cite{Zheng2019helen} proposed a distributed secure multi-party solution to regularized linear model training that also allows for malicious adversaries. Examples of works that tackle private linear programming problems or machine learning applications with optimization subroutines using garbled circuits, secret sharing, and hybrid approaches between the above are~\cite{Li2006secure,Xie16,Mohassel17,Gascon17,Chase17} and the references within. In control applications, works such as~\cite{leNy14differentially,Dong2015differential,Cortes2016differential,Wang2017differential} employ differential privacy to protect dynamic data. Furthermore, examples such as~\cite{Huang15,Han17,Nozari18} make use of differential privacy techniques in optimization algorithms. Other lines of work solve optimization problems with \mbox{non-cryptographic} methods: using transformations, e.g.~\cite{Vaidya2009privacy,Weeraddana2013per,Wang2015efficient,Salinas2016efficient}, which guarantee an uncertainty set around the private data, or using arbitrary functions for obfuscating the objective functions in a special graph topology~\cite{Gade16}. \subsection{Contributions} In this paper we consider a privacy-preserving cloud-based constrained quadratic optimization. In control theoretic applications and machine learning, linear and quadratic optimization problems arise frequently -- e.g., state estimation under minimum square error, model predictive control, support vector machines -- which require privacy guarantees of the data involved in the computation. We develop a new tractable optimization protocol to privately solve constrained quadratic problems. Our protocol relies on cryptographic tools called encryption schemes. To solve the optimization problem on encrypted data, we use an additively homomorphic encryption scheme, where, in short, \textit{addition commutes with the encryption function}. Thus, a party can process encrypted data without having direct access to the data. The novelty is how to handle in a privacy-preserving manner the constraints of the problem that introduce non-linearities that cannot be supported by additively homomorphic encryption schemes. We show that a projected gradient method that operates on the Lagrange dual problem can alleviate this problem and can be efficiently run on encrypted data by exploiting communication between the participating parties. The main contributions of our work are the following: \begin{itemize}[topsep=0pt,wide, labelwidth=!, labelindent=0pt] \item We formally state and prove computational security guarantees for such protocol. The proof relies on applying cryptographic tools to the specific optimization algorithm that runs over multiple iterations. \item We propose an alternative protocol which sacrifices some privacy but involves less communication overhead. \item We implement the protocols and show the computational~and communication complexity produce reasonable running times. \end{itemize} Furthermore, we emphasize the difference between the computational security guarantee with the \textit{non-unique retrieval} guarantee that is important in such optimization problems. We finally point out that the present manuscript provides detailed security proofs and analyses not available in previous versions of the papers~\cite{Shoukry16,Alexandru17}. \emph{Organization.} The rest of the paper is organized as follows: we formulate the problem in Section~\ref{sec:problem_setup} and formally introduce the computational security guarantees in Section~\ref{sec:privacy_goals}. We describe the cryptographic tools we use in Section~\ref{sec:encryption_sch}, more specifically, the partially homomorphic scheme, along with its properties and the symmetric encryption used. In Section~\ref{sec:gradient_ascent}, we describe the optimization theory behind the proposed protocol and justify the choice of the gradient method. Furthermore, we present the main protocol and the subroutines that compose it and in Section~\ref{sec:security_proof}, we show that the protocol achieves privacy of the agent's data and of the target node's result. We discuss possible relaxations of the security guarantees and propose a more efficient protocol under these weaker conditions in Section~\ref{sec:relaxation_security}. In Section~\ref{sec:discussion}, we provide a privacy analysis of the problem concerning the input-output relation. We present the complexity analysis of the protocols and show that the experimental results point out reasonable running times in Section~\ref{sec:complexity_time}. Additional tools necessary for proving the security of the proposed protocol are introduced in Appendix~\ref{app:cryptographic_prel} and further details on the protocols used are given in Appendix~\ref{app:comparison}. Finally, the detailed privacy proofs are given in Appendix~\ref{app:proof}. \emph{Notation.} We denote matrices by capital letters. $\mathbb R^{m\times n}$ represents the set of $m\times n$ matrices with real elements and $\mathbb S^n_{++}$ represents the set of symmetric positive definite $n\times n$ matrices. Element-wise inequalities between vectors are denoted by $\preceq$. The notation $x\leftarrow X$ means that $x$ is uniformly drawn from a distribution $X$. $\text{Pr}$ denotes the probability taken over a specified distribution. We refer to algorithms that are run interactively by multiple parties as protocols. \section{Introduction}\label{sec:introduction} \input{introduction_section.tex} \section{Problem Setup}\label{sec:problem_setup} \input{problem_setup_section.tex} \section{Privacy goals}\label{sec:privacy_goals} \input{privacy_goals_section.tex} \section{Encryption scheme}\label{sec:encryption_sch} \input{encryption_scheme_section.tex} \section{Secure Constrained Quadratic Optimization}\label{sec:gradient_ascent} \input{gradient_ascent_section.tex} \section{Privacy of Quadratic Optimization Protocol}\label{sec:security_proof} \input{security_proof_section.tex} \section{Alternative quadratic optimization protocol without secure comparison}\label{sec:relaxation_security} \input{relaxation_security.tex} \section{Privacy discussion}\label{sec:discussion} \input{discussion_section.tex} \section{Implementation}\label{sec:complexity_time} \input{computational_complexity_section.tex} \section*{Acknowledgment} This work was partially sponsored by the NSF awards 1739816, 1705135, CNS-1505799 and the Intel-NSF Partnership for Cyber-Physical Security and Privacy, ONR N00014-17-1-2012, and by TerraSwarm, one of six centers of STARnet, a Semiconductor Research Corporation program sponsored by MARCO and DARPA. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of NSF, or the U.S. Government. Andreea Alexandru would like to thank Brett Hemenway for helpful discussions. \ifCLASSOPTIONcaptionsoff \newpage \fi \bibliographystyle{IEEEtran} \subsection{Motivating Examples} Quadratic optimization is a class of problems frequently employed in control system design and operation. As a first motivating example, consider the problem of estimating the state of a dynamical system from privacy-sensitive sensor measurements. For instance, such a problem arises in smart houses where the temperature or energy readings of the sensors are aggregated by a cloud controller and can reveal whether the owners are in the building. In particular, let the system dynamics and sensor measurements be described by: \begin{align} \begin{split} x_{t+1} &= A \, x_t + w_t \\ y_t &= C \, x_t + v_t, \end{split} \end{align} for $t =0,\ldots, T-1$, where $w_t, v_t$ are process and measurement noise respectively. The system and sensor parameters $A \in \mathbb{R}^{n \times n},C\in \mathbb{R}^{p \times n}$ can be thought as publicly available information while the sensor measurements $y_0, \dots, y_{T-1}$ are privacy-sensitive data. The untrusted cloud has to collect the measurements and output an initial state estimate $x_0$ of the system to a target agent, while maintaining the privacy of the sensor data and final output. A simple state estimate may be found as the solution to the least squares problem: \begin{align}\label{eq:estimation_example} &\underset{x_0 \in \mathbb{R}^n }{\text{min}} \quad \frac{1}{2} \sum_{t=0}^\intercal \| y_t - C A^t\, x_0 \|^2_2 = \frac{1}{2} \| {y} - \mathcal{O}\, x_0 \|^2_2, \end{align} where \begin{equation} {y} = \left[ \begin{array}{c} y_0 \\ y_1\\ \vdots \\ y_{T-1} \end{array}\right], \; \mathcal{O} = \left[ \begin{array}{c} C\\ CA\\ \vdots\\ CA^{T-1} \end{array}\right] \end{equation} and $\mathcal{O} \in \mathbb{R}^{Tp \times n}$ is the well-known system observability matrix. More general state estimation problems may also include constraints, e.g., the initial state may lie in a polyhedral set $D x_0 \preceq b$ where the shape of the polyhedron captured by the matrix $D$ is publicly known but its position and size captured by the vector $b$ is private information. As a second example, consider a control problem with privacy-sensitive objectives. Suppose we are interested in steering a dynamical system: \begin{align}\label{eq:system} \begin{split} x_{t+1} &= A \, x_t +B \, u_t \\ y_t &= C \, x_t, \end{split} \end{align} starting from a private initial position $x_0$ while guaranteeing safety state constraints of the form: \begin{align} \underline{x}_t \preceq x_t \preceq \overline{x}_t, t =1,\ldots, T, \end{align} where $\underline{x}_t , \overline{x}_t$ for $ t =1,\ldots, T $ are private values. Such problems arise when exploring vehicles are deployed in uncertain or hazardous environments or when different users compete against each other and want to hide their tactics. Denote by $x^r$ and $u^r$ the private values such that the system is guaranteed to track a reference~$r$. Then, we need to solve the problem: \begin{align}\label{eq:control_example_1} \begin{split} \underset{x_k \in \mathbb{R}^n, u_k \in \mathbb{R}^m }{\text{min}}~&~\frac{1}{2} \sum_{t=0}^{T-1} \|y_t-C\,x^r_t\|^2_Q + \|u_t-u^r_t\|^2_R \\ % \text{s.t.}~&~ x_{t+1} = A \, x_t +B \, u_t, \quad y_t = C\, x_t \\ &~\underline{x}_t \preceq x_t \preceq \overline{x}_t, \quad t =1,\ldots, T, \end{split} \end{align} where $Q,R$ are positive definite cost matrices and we used the quadratic norm notation $\|z\|^2_P := z^\intercal P z$. The publicly known system dynamics \eqref{eq:system} are equality constraints which can be eliminated to obtain a control input design problem with only linear inequality constraints: \begin{align}\label{eq:control_example_2} \begin{split} \underset{x_k \in \mathbb{R}^n, u_k \in \mathbb{R}^m }{\text{min}}~&~ \frac{1}{2}\sum_{t=0}^{T-1} u_t^\intercal \, H_t\,u_t + [\begin{array}{ccc} x_0 & x^r_t & u^r_t \end{array}] \,F_t \, u_t \\ % \text{s.t.}~&~\underline{x}_t \preceq A^t\, x_0 +\sum_{j=0}^{t-1} A^{t-1-j} B\, u_j \preceq \overline{x}_t, \\ &~t =1,\ldots, T, \end{split} \end{align} with $H_t $ and $F_t$ are matrices that depend on the costs $Q,R$ and the system's dynamics~\eqref{eq:system} appropriately computed from~\eqref{eq:control_example_1}. \subsection{Problem statement} The above examples can be modeled as constrained quadratic optimization problems with distributed private data. We consider three types of parties involved in the problem: a number of agents~$\mathcal A_i$, $i=1,\ldots,p$, a cloud server~$\mathcal C$ and a target node~$\mathcal T$. The purpose of this setup is to solve an optimization problem with the data provided by the agents and the computation performed on the cloud and send the result to the target node. The architecture is presented in Figure~\ref{fig:setup}. Let us consider a strictly-convex quadratic optimization problem, which we assume to be feasible: \begin{align}\label{eq:problem} \begin{split} x^\ast ~= ~\argmin\limits_{x\in\mathbb R^n}&~\frac{1}{2} x^\intercal \, Q _{\mc C} \, x + c_{\mc A}^\intercal x\\ s.t.&~ A_{\mc C} \, x\preceq b_{\mc A}, \end{split} \end{align} where the variables and the parties to which they belong to are described as follows:\\ \noindent\textbf{Agents $\mathcal A = (\mathcal A_1,\ldots,\mathcal A_p)$}: The agents are parties with low computational capabilities that possess the private information $b_{\mc A}$ and $c_{\mc A}$. The private information is decomposed across the agents as: $b_{\mc A} = (b_1,\ldots,b_p)$ and $c_{\mc A} = (c_1,\ldots,c_p)$, with $b_i\in \mathbb R^{m_i}$ and $c_i\in \mathbb R^{n_i}$ being the private data of agent $i$ such that $\sum_{i=1}^p m_i = m$ and $\sum_{i=1}^p n_i = n$.\\ \noindent\textbf{Cloud $\mathcal C$}: The cloud is a party with high computational capabilities that has access to the matrices $Q_{\mc C}\in \mathbb S^n_{++}$ and $A_{\mc C}\in\mathbb R^{m\times n}$. When the computation is sophisticated and/or involves proprietary algorithms, $Q_{\mc C}$ and $A_{\mc C}$ are private data of the cloud. In order to capture a greater number of problems, we will also consider the case where $Q_{\mc C}$ or $A_{\mc C}$ are public. \\ \noindent\textbf{Target Node $\mathcal T$}: The target node is a party with more computational capabilities than the agents that is interested in the optimal solution $x^\ast$ of the problem. The target node can be either one of the agents or a separate cloud server. \begin{figure}[h] \centering \includegraphics[width=0.38\textwidth]{setup.pdf} \caption{ Architecture of the problem: Agents are low-resource parties that have private data that they outsource to a powerful server, called the cloud. The cloud has to solve an optimization problem on the private data of the agents and send the result to a party called the target node, which will help with the computations.} \label{fig:setup} \end{figure} Note that in the first motivating example~\eqref{eq:estimation_example}, the matrix $Q_{\mc C}$ in~\eqref{eq:problem} corresponds to the publicly known matrix $\mathcal{O}^\intercal \mathcal{O}$ and the vector $c_{\mathcal{A}}$ corresponds to the private vector $-\mathcal{O}^\intercal \, {y}$. For the objective to be strongly convex, we require that $\text{rank}(\mathcal{O})=n$, which also implies standard system observability. Similarly, in the second example~\eqref{eq:control_example_2}, the matrix $Q_{\mc C}$ is composed from the regulating cost matrices, the cost vector $c_{\mathcal{A}}$ is formed from the private initial conditions and steady-state solution for the reference tracking, mixed by the system's dynamics, and the constraints vector $b_{\mathcal{A}}$ depends on the private state bounds and initial condition. \subsection{Adversarial model} In most cloud applications, the service provider has to deliver the contracted service or otherwise the clients would switch to another service provider. This means that the cloud cannot alter the data it receives. Moreover, the clients' interest is to obtain the correct result from the service they pay for, hence, the agents and target node will also not alter the data. However, the parties can locally process the data they receive in any fashion they want. This model is known as \mbox{semi-honest}, which is defined as follows: \begin{definition}(\mbox{Semi-honest model}) A party is \mbox{semi-honest} if it does not deviate from the steps of the protocol but may store the transcript of the messages exchanged and process the data received in order to learn more information than stipulated by the protocol. \end{definition} This model also holds when considering eavesdroppers on the communication channels. An adversarial model that only considers eavesdroppers as adversaries differs from the \mbox{semi-honest} model by the fact that, apart from the untrusted channels, the parties that perform the computations are also not trusted. Malicious and active adversaries -- that diverge from the protocols or tamper with the messages -- are not considered in this paper. The purpose of the paper is to solve Problem~\eqref{eq:problem} using a secure multi-party computation protocol for semi-honest parties. This protocol takes as inputs the private data of the agents, as well as the cloud's data, and involves the parties in exchanging messages and participating in some specified computations, and eventually outputs to the target node the solution of the optimization problem. This protocol should guarantee \textit{computational security}. More specifically, the cloud cannot obtain any information about the private inputs of the agents and the output of the target node and similarly, the target node cannot obtain any information about the private inputs of the agents and the cloud, other than what they can compute using their inputs and outputs and public information, by running a computationally efficient algorithm after the execution of the protocol. \subsection{Proof of Theorem~\ref{thm:main_alg}}\label{subsec:proof_thm1} For simplicity of the exposition, we avoid mentioning the public data (public keys $pk_{\mc T}, pk_{DGK}$, number of iterations $K$ and number of bits $l$) in the views, since they are public. In what follows, we will successively discuss the views of each type of party participating in the protocol: agents, cloud and target node. As mentioned in Definitions~\ref{def:view} and~\ref{def:coalition_view}, the views of the parties during the execution of Protocol~\ref{alg:main_alg} are constructed on all the inputs of the parties involved and symbolize the real values the parties get in the execution of the protocol. We will denote by $\mathcal I$ the inputs of all the parties: \[\mathcal I = \{b_{\mc A},c_{\mc A},A_{\mc C},Q_{\mc C},sk_{\mc T},sk_{DGK}\}.\] Furthermore, in order to construct a simulator that simulates the actions of a party, we need to feed into it the inputs and outputs of that respective party. \subsubsection{Simulator for one agent $\mathcal A_i$} Agent $\mathcal A_i$, for every $i=1,\ldots,p$, has the following inputs $I_{\mathcal A_i} = (\{b_j\}_{j=1,\ldots,m_i},$ $\{c_j\}_{j=1,\ldots,n_i})$. To avoid cluttering, we will drop the subscripts $j=1,\ldots,m_i$ and ${j=1,\ldots,n_i}$. Then agent $\mathcal A_i$ has the following view: \[V_{\mathcal A_i}(\mathcal I):= (b_j, c_j, [[b_j]], [[c_j]],\text{coins}), \] where coins represent the random values used for encryption. The agents are only online to send their encrypted data to the cloud, and they do not have any role and output afterwards. Hence, a simulator $S_{\mathcal A_i}$ would simply generate the random values necessary to encrypt its inputs as specified by the protocol and output the view obtained such: \[S_{\mathcal A_i}:= (b_j, c_j, \widetilde{[[b_j]]}, \widetilde{[[c_j]]}, \widetilde{\text{coins}}), \] where by $\widetilde{(\cdot)}$ we denote the quantities obtained by the simulator, which are different than the quantities of the agents, but follow the same distribution. Hence, it is trivial to see that the protocol is secure in the semi-honest model from the point of view of the interaction with the agents. Next, we want to prove the privacy of the protocol from the point of view of the interactions with the cloud and the target node. We will construct a sequence of algorithms such that we obtain that the view of the real parties after the execution of $K$ iterations is the same as the view of simulators that simply execute $K$ iterations with random exchanged messages. \subsubsection{Simulator for the cloud $\mathcal C$}\label{subsubsec:cloud} For the simplicity of the exposition, we will treat $\mu$, $b_{\mc A}$ and $c_{\mc A}$ as scalars. The same steps are repeated for every element in the vectors. For clarity, we will form the view of the cloud in steps, using pointers to the lines in Protocol~\ref{alg:main_alg}. The view of the cloud during the execution of lines 5-6 is: \begin{equation}\label{eq:view_cloud_-1} V_{\mc C}^{-1}(\mathcal I) = \big ( A_{\mc C},Q_{\mc C},\eta,[[b_{\mc A}]],[[c_{\mc A}]],[[\mu_0]],\text{coins}\big)=:I_{\mc C}^{-1}, \end{equation} where coins represent the random values generated for the Paillier encryption. Furthermore, we construct the view of the cloud at iteration $k=0,\ldots,K-1$ during the execution of an instance of Protocol~\ref{alg:iteration}, which will be constructed on the inputs of all parties: the inputs $\mathcal I$ and the data the parties had at iteration $k-1$. We denote the view of the cloud at iteration $k-1$ by $I_{\mc C}^{k-1}$ which, along with $\mathcal I$ and the view of the target node at iteration $k-1$, denoted by $I_{\mc T}^{k-1}$~\eqref{eq:view_target_k}, will be what the view is constructed on at iteration $k$. \begin{align*} &\bar{\mathcal I}^{k-1} := \mathcal I \cup I_{\mc C}^{k-1} \cup I_{\mc T}^{k-1},\numberthis \label{eq:inputs_k}\\ &I_{\mc C}^{k}:=V_{\mc C}^k(\bar{\mathcal I}^{k-1}) = \big(I_{\mc C}^{k-1},[[\mu_k]],[[\bar\mu_k]],\underbrace{\pi_k,\text{coins}_{1k}}_{Protocol~\ref{alg:rand_step}}, \numberthis \label{eq:view_cloud_k}\\ &\underbrace{\rho_k,[[t_k]],\text{m}^{\text{comp}_k},\text{coins}_{2k}}_{Protocol~\ref{alg:comp_DGKV}}, \underbrace{r_k,s_k,[[v_k]],[[\mu_{k+1}]],\text{coins}_{3k}}_{Protocol~\ref{alg:update_step}}\big), \end{align*} where $\text{m}^{\text{comp}_k}$ are messages exchanged in the comparison protocol, and $\text{coins}_{jk}$, for $j=1,2,3$ are the random numbers generated in Protocol~$j$. Finally, the view of the cloud after the execution of line 11 in Protocol~\ref{alg:main_alg} is: \begin{equation}\label{eq:view_cloud_K} V_{\mc C}^K(\bar{\mathcal I}^{K-1}) = \big (I^{K-1}_{\mc C},[[x^\ast]]). \end{equation} Therefore, the view of the cloud during the whole execution of Protocol~\ref{alg:main_alg} is: \[V_{\mc C}(\mathcal I) := V_{\mc C}^K(\bar{\mathcal I}^{K-1}).\] We first construct a simulator on the inputs $I_{\mc C} = \{A_{\mc C}, Q_{\mc C}\}$ that mimics $V^{-1}_{\mc C}(\mathcal I)$ in~\eqref{eq:view_cloud_-1}: \begin{itemize}[wide, labelwidth=!, labelindent=0pt] \item Generate $n+m$ random numbers of $l$ bits $\widetilde {b_{\mc A}},\widetilde{c_{\mc A}}$; \item Generate a random positive initial value $\widetilde{\mu_0}$; \item Generate $n+m+1$ uniformly random numbers for the Paillier encryption and denote them $\widetilde{\text{coins}}$; \item Compute $[[\widetilde {b_{\mc A}}]],[[\widetilde{c_{\mc A}}]],[[\widetilde{\mu_0}]]$; \item Compute $\eta$ following line 6; \item Output $\widetilde I_{\mc C}^{-1} := S^{-1}_{\mc C}(I_{\mc C}) = \big (A_{\mc C}, Q_{\mc C},[[\widetilde{b_{\mc A}}]],[[\widetilde{c_{\mc A}}]],\eta,[[\widetilde{\mu_0}]],$ $\widetilde{\text{coins}}\big)$. \end{itemize} Since Protocol~\ref{alg:iteration} is secure in the semi-honest model (Proposition~\ref{prop:iteration}), we know that there exists a probabilistic polynomial-time (ppt) simulator for the functionality of Protocol~\ref{alg:iteration} on inputs $(A_{\mc C},Q_{\mc C},[[b_{\mc A}]],[[c_{\mc A}]],\eta,[[\mu_k]])$ and output $[[\mu_{k+1}]]$. However, we need to show that we can simulate the functionality of consecutive calls of Protocol~\ref{alg:iteration}, or, equivalently, on one call of Protocol~\ref{alg:iteration} but on the augmented input that contains the data of the cloud in the previous iterations. Call such a simulator $S^k_{\mc C}$, that on the input $I^{k-1}_{\mc C}$ mimics $V^k_{\mc C}(\bar{\mathcal I}^{k-1})$ in~\eqref{eq:view_cloud_k}, for $k=0,\ldots,K-1$: \begin{itemize}[wide, labelwidth=!, labelindent=0pt] \item Compute $[[\nabla g(\mu_k)]]$ and $[[\bar \mu_k]]$ as in lines 1-2 of Protocol~\ref{alg:iteration} from $[[\mu_k]],[[b_{\mc A}]],[[c_{\mc A}]]$ which are included in $I^{k-1}_{\mc C}$; \item Generate a random permutation $\widetilde{\pi_k}$ and apply it on $([[0]],[[\bar \mu_k]])$ as in Protocol~\ref{alg:rand_step}; \item Follow Protocol~\ref{alg:comp_DGKV} and replace the incoming messages by DGK encryptions of appropriately sized random values to obtain $\widetilde{\rho_k},\widetilde{\text{m}^{\text{comp}_k}}$; \item Generate random bit $\widetilde{t_k}$ and its encryption $[[\widetilde{t_k}]]$; \item Generate random values $\widetilde{r_k}$ and $\widetilde{s_k}$ as in line 1 in Protocol~\ref{alg:update_step} and their encryptions $[[\widetilde{r_k}]],[[\widetilde{s_k}]]$; \item Obtain $[[\widetilde{v_k}]]$ by choosing between the elements of $\widetilde{\pi_k}([[0]],[[\bar\mu_k]])+(\widetilde{r_k},\widetilde{s_k})$ according to the generated $\widetilde{t_k}$; \item Compute $[[\widetilde{\mu_{k+1}}]]$ as in line 8 of Protocol~\ref{alg:update_step}; \item Denote the rest of the random values used for encryption and blinding by $\widetilde{\text{coins}_k}$; \item Output $\widetilde I_{\mc C}^k :=S^k_{\mc C}(I^{k-1}_{\mc C})=(I^{k-1}_{\mc C},[[\bar \mu_k]],[[\widetilde{\pi_k}]],[[\widetilde{z_k}]],$ $\widetilde{\text{m}^{\text{comp}_k}},[[\widetilde{t_k}]],[[\widetilde{r_k}]],[[\widetilde{s_k}]],[[\widetilde{v_k}]],[[\widetilde{\mu_{k+1}}]],\widetilde{\text{coins}_k})$. \end{itemize} Finally, a trivial simulator $\widetilde I_{\mc C}^K:= S^K_{\mc C}(I^{K-1}_{\mc C})$ for $V^K_{\mc C}(\bar{\mathcal I}^{K-1})$ is obtained by simply performing line 11 on the inputs. \begin{proposition}\label{prop:sim_cloud} $S_{\mc C}^{k}(I_{\mc C}^{k-1})\stackrel{c}{\equiv}V^{k}_{\mc C}(\bar{\mathcal I}^{k-1})$, for \mbox{$k=-1,\ldots,K$}, where $I_{\mc C}^{k-2}:=I_{\mc C}$. \end{proposition} \begin{proof} For $k=-1$, due to the fact that the coins and the initial iterate are generated by the simulator via the same distributions as specified in the protocol, and due to the semantic security of the Paillier encryption, which guarantees indistinguishability of encryptions, we obtain that $V^{-1}_{\mc C}(\mathcal I)\stackrel{c}{\equiv}S^{-1}_{\mc C}(I_{\mc C})$. For $0\leq k\leq K-1$, $(\text{coins}_{1k},\text{coins}_{2k},\text{coins}_{3k})\stackrel{s}{\equiv}\widetilde{\text{coins}_k}$ because they are generated from the same distributions. Similarly, the distributions of $(\widetilde{\pi_k},\widetilde{\pi_k}([[0]],[[\bar \mu_k]]))$ and $(\pi_k,\pi_k([[0]],[[\bar \mu_k]]))$ are the same and $\widetilde{\pi_k},\pi_k$ are independent of the other parameters. The quantities from the DGK protocol $\widetilde{\rho_k}\stackrel{s}{\equiv}\rho_k$ and $\widetilde{\text{m}^{\text{comp}_k}}\stackrel{c}{\equiv}\text{m}^{\text{comp}_k}$ either due to the semantic security of the DGK protocol, e.g., $[\widetilde{t'}],[\widetilde{\delta_{\mc T}}]$ or due to having the same distributions, e.g., $\widetilde{\rho_k}, \widetilde{\delta_{\mc C}},\widetilde{\alpha}$. The values in the update step $\widetilde{r_k},\widetilde{s_k}$ are sampled from the same distribution as $r_k,s_k$, and, finally, $[[\widetilde{t_k}]],[[\widetilde v_k]],[[\widetilde{\mu_{k+1}}]]\stackrel{c}{\equiv}[[t_k]],[[v_k]],[[\mu_{k+1}]]$ due to the semantic security of the Paillier encryption. Thus, $S^{k}(I_{\mc C}^{k-1})\stackrel{c}{\equiv}V^{k}_{\mc C}(\bar{\mathcal I}^{k-1})$. For $k=K$, $S_{\mc C}^{K}(I_{\mc C}^{K-1})\stackrel{c}{\equiv}V^{K}_{\mc C}(\bar{\mathcal I}^{K-1})$ follows from the fact that the simulator simply executes the last part of the protocol and from the Paillier scheme's semantic security. \end{proof} Hence, we obtained that $I^k_{\mc C} \stackrel{c}{\equiv} \widetilde I^k_{\mc C}$, for $k=-1,\ldots,K$. The essence of the proof of Proposition~\ref{prop:sim_cloud} is that all the messages the cloud receives are encrypted. Then, thanks to the semantic security of the Paillier and DGK schemes, the extra information included in $\bar{\mathcal I}^{k-1}$ from the previous iterations cannot be used to extract other information about the values at iteration $k$. From this, we also have the next corollary: \begin{corollary}\label{corr:consec_steps} $S^{k}_{\mc C}(\widetilde I^{k-1}_{\mc C})\stackrel{c}{\equiv}S^{k}_{\mc C}(I_{\mc C}^{k-1})$, equivalent to $S^{k+1}_{\mc C}(S^{k}_{\mc C}( I_{\mc C}^{k-1}))\stackrel{c}{\equiv}S^{k+1}_{\mc C}(V^{k}_{\mc C}(\bar{\mathcal I}^{k-1}))$, for any $k=0,\ldots,K-1$. \end{corollary} Finally, we construct a simulator $S_{\mc C}(I_{\mc C})$ for the execution of Protocol~\ref{alg:main_alg} and we will show that its view will be computationally indistinguishable from $V_{\mc C}(\mathcal I)$. To this end, we define the following sequence of views -- obtained as hybrids between the real views and the views of the simulators: \begin{align*} V_{\mc C}(\mathcal I)=H_{-1}(\mathcal I) &= V^K_{\mc C} (\bar {\mathcal I}^{K-1})\\ H_0(\mathcal I) &= S^K_{\mc C} \circ V_{\mc C}^{K-1}(\bar {\mathcal I}^{K-2})\\ H_1(\mathcal I) &= S^K_{\mc C} \circ S_{\mc C}^{K-1} \circ V_{\mc C}^{K-2}(\bar {\mathcal I}^{K-3})\\ &~\vdots\\ H_{K}(\mathcal I) &= S^K_{\mc C}\circ S^{K-1}_{\mc C}\circ \ldots\circ S_{\mc C}^0\circ V^{-1}_{\mc C}(\mathcal I)\\ S_{\mc C}(I_{\mc C})=H_{K+1}(I_{\mc C}) &= S^K_{\mc C}\circ S^{K-1}_{\mc C}\circ \ldots\circ S_{\mc C}^0\circ S^{-1}_{\mc C}(I_{\mc C}). \end{align*} By transitivity, $H_{-1}$ and $H_{K+1}$ are computationally indistinguishable if: \[ H_{-1} \stackrel{c}{\equiv} H_0\stackrel{c}{\equiv} \ldots \stackrel{c}{\equiv} H_k \stackrel{c}{\equiv} H_{k+1} \stackrel{c}{\equiv} H_{k+2} \stackrel{c}{\equiv} \ldots \stackrel{c}{\equiv} H_{K+1}. \] This result follows from induction on Corollary~\ref{corr:consec_steps}. In conclusion, we obtain that $S_{\mc C}(I_{\mc C})\stackrel{c}{\equiv} V_{\mc C}(\mathcal I)$, which verifies that Protocol~\ref{alg:main_alg} achieves privacy with respect to the cloud. \subsubsection{Simulator for the target node $\mathcal T$}\label{subsubsec:target} We proceed with the same steps in order to show that the consecutive $K$ iterations form a protocol that is secure in the semi-honest model from the point of view of the target node. We will use the secrecy of the one-time pad variant used for blinding. The symmetric encryption used in this paper, as discussed in Section~\ref{subsec:symmetric}, to which we refer as blinding, guarantees that a value of $l$ bits additively blinded by a random value of $l+\lambda$ bits is statistically indistinguishable from a random value of $l+\lambda+1$ bits. The inputs and output of the target node in Protocol~\ref{alg:main_alg}~are $I_{\mc T} = (sk_{\mc T},sk_{DGK},x^\ast)$. As in~\eqref{eq:inputs_k}, $\bar{\mathcal I}^{k-1}$ represents the~inputs of all the parties at iteration $k$, with $\bar{\mathcal I}^{-1} = \mathcal I$. Then, the view of the target node during iteration $k=0,\ldots,K-1$ is: \begin{align*} I_{\mc T}^{k}:=&V_{\mc T}^k(\bar{\mathcal I}^{k-1}) = (sk_{\mc T},sk_{DGK},\underbrace{z_k,t_k,\text{m}^{\text{comp}_k},\text{coins}_{2k}}_{Protocol~\ref{alg:comp_DGKV}},\\ & \underbrace{\bar a_k,\bar b_k, v_k,\text{coins}_{3k}}_{Protocol~\ref{alg:update_step}}). \numberthis\label{eq:view_target_k} \end{align*} The view of the target node during the last step of the protocol is: \begin{equation}V_{\mc T}^K(\bar{\mathcal I}^{K-1}) := (I_{\mc T}^{K-1},[[x^\ast]]).\end{equation} As before, the view of the target during the execution of Protocol~\ref{alg:main_alg} is: \[V_{\mc T}(\mathcal I) := V_{\mc T}^K(\bar{\mathcal I}^{K-1}). \] In order to be able to construct a simulator with indistinguishable view from the view of the target node, we need to show that the target node is not capable of inferring new relevant information about the private data (other than what can be inferred solely from its inputs and outputs) although it knows the optimal solution $x^\ast$ and has access to the messages from multiple iterations. Apart from the last message, which is the encryption of the optimization solution $[[x^\ast]]$, and the comparison results $t_k$, all the values the target node receives are blinded, with different sufficiently large values at each iteration. The target node knows that $Q_{\mc C} x^\ast = A_{\mc C}^\intercal \mu_K - c_{\mc A}$ and that $\mu_K = v_K - \bar r_K$, for some random value $\bar r_K$. However, from these two equations, although it has access to $x^\ast$ and $v_K$, the target node cannot infer any new information about the private data of the agents or about $\mu_K$, even when $Q_{\mc C},A_{\mc C}$ are public, thanks to the large enough randomization. Let $a_k,b_k = \pi_k(0,\bar\mu_{k})$, where $\pi_k$ is a random permutation. From Protocol~\ref{alg:comp_DGKV}, the target receives $z_k$ which is the additively blinded value of $b_k-a_k+2^l$, and other blinded values, denoted in~\eqref{eq:view_target_k} as $\text{m}^{\text{comp}_k}$. Hence, provided the blinding noises are refreshed at every iteration, the target node cannot infer any information, as follows from Section~\ref{subsec:symmetric}. At the end of the protocol, it receives the bit $t_k$ which is 1 if $a_k\leq b_k$ and 0 otherwise. However, due to the uniform randomization between the order of $\bar\mu_k$ and $0$ at every iteration for assigning the values of $a_k$ and $b_k$, described in Protocol~\ref{alg:rand_step}, the target node cannot identify the sign of $\bar \mu_k$, and by induction, the sign of $\bar \mu_{K-1}$ and magnitude of $\mu_K$. This means that having access to $x^\ast = x_K$ does not bring more information about the blinded values from the intermediary steps. We now investigate the relation between the messages from consecutive iterations. From the update protocol~\ref{alg:update_step}, we know: \begin{align*} v_k &= (a_k+r_k)(1-t_k) + (b_k+s_k)t_k \\ \mu_{k+1} &= v_k - r_k(1-t_k) - s_kt_k = a_k(1-t_k) + b_kt_k, \end{align*} where the variables are vectors. The target node knows the values of $v_k$ and $t_k$, but $\pi_k,\bar\mu_k,r_k,s_k$ are unknown to it. Furthermore, let $\pi_{k+1}$ be the permutation applied by the cloud at step $k+1$ in Protocol~\ref{alg:rand_step}, unknown to the target node. Take for example the case when $t_k = 0$ and $t_{k+1} = 0$. Let $\tilde Q = \mathbf I - \eta A_{\mc C} Q_{\mc C}^{-1} A_{\mc C}^\intercal$ and $\mathbf E = [\mathbf I~ \mathbf 0]$. Then: \begin{align*} v_{k+1} &= a_{k+1} + r_{k+1}= \mathbf E \pi_{k+1} (0, \tilde Q(v_k - r_k) -\eta A_{\mc C} Q_{\mc C}^{-1}c_{\mc A} -\\ &- \eta b_{\mc A} ) + r_{k+1}. \end{align*} The above equation shows the target node cannot construct $v_{k+1}$ from $v_{k}$. Similar equations arise when considering the other values for $t_k$ and $t_{k+1}$. This guarantees that an integer $\tilde v_k^i$ obtained by selecting uniformly at random from $(2^{l+\lambda},2^{l+\lambda+1})\cap\mathbb Z_N$ will have the distribution statistically indistinguishable from $v_k^i$. Moreover: \begin{align*} &\mu_{k+2} = \max\{0,\tilde Q\mu_{k+1} -\eta A_{\mc C} Q_{\mc C}^{-1}c_{\mc A} - \eta b_{\mc A}\} \\ &= \max\{0, \tilde Q(v_k - r_k(1-t_k) - s_kt_k) -\eta A_{\mc C} Q_{\mc C}^{-1}c_{\mc A} - \eta b_{\mc A}\} \\ &= v_{k+1} - r_{k+1}(1-t_{k+1}) - s_{k+1}t_{k+1} . \end{align*} Since the blinding noise is different at each iteration and uniformly sampled, the target node cannot retrieve $\mu_{k+2}$ from $v_k$ and $v_{k+1}$. In short, if the target node receives some random values $\widetilde{\bar a_k},\widetilde{\bar b_k}\in (2^{l+\lambda},2^{l+\lambda+1})\cap \mathbb Z_N$ instead of $a_k + r_k$ and $b_k+s_k$ respectively, it would not be able to distinguish the difference. Similar arguments hold for the blinded messages $c_i$ from the comparison protocol~\ref{alg:plain_DGKV}. Hence, by processing multiple iterations, the target node can only obtain functions of the private data that involve at least one large enough random value, which does not break privacy. We now build a simulator $S_{\mc T}$ that applies the steps of the protocol on randomly generated values. As before, since Protocol~\ref{alg:iteration} is secure in the semi-honest model (Proposition~\ref{prop:iteration}), we know that there exists a ppt simulator for the functionality of Protocol~\ref{alg:iteration} on inputs $I_{\mc T}^{-1} = \{sk_{\mc T},sk_{DGK}\}$. However, we need to show that we can simulate the functionality of consecutive calls of Protocol~\ref{alg:iteration}, or, equivalently, on one call of Protocol~\ref{alg:iteration} but on inputs $(I_{\mc T}^k, x^\ast)$. Call such a simulator $S^k_{\mc T}$, that on the inputs $(I_{\mc T}^k, x^\ast)$ should output a view that is statistically indistinguishable from $V^k_{\mc T}(\bar{\mathcal I}^{k-1})$ in~\eqref{eq:view_target_k}, for $k=0,\ldots,K-1$. We already showed that although the target node has access to the output $x^\ast$ and to blinded messages from all the iterations of Protocol~\ref{alg:main_alg}, it cannot extract information from them or correlate the messages to the iteration they arise from, so $x^\ast$ will only be relevant in the last simulator. \begin{itemize}[wide, labelwidth=!, labelindent=0pt] \item Generate a $\lambda+l$-length random integer $\widetilde{\rho_k}$ and add $2^l$ and obtain $\widetilde{z_k}$; \item Generate a random bit $\widetilde{t_k}$; \item Choose a random bit $\widetilde{\delta_{\mc T}}$. If it is 0, then generate $l$ non-zero values of $2t$ bits, otherwise generate $l-1$ non-zero random values and one 0 value. Those will be the $\widetilde{\text{m}^{\text{comp}_k}}$ (see Protocol~\ref{alg:plain_DGKV}); \item Generate random integers of length $l+ \lambda+1$ $\widetilde{\bar a_k}$ and $\widetilde{\bar b_k}$; \item Compute $\widetilde{v_k}$ according to $\widetilde{t_k}$; \item Denote all Paillier and DGK generated coins by $\widetilde{\text{coins}}$; \item Output $\widetilde I_{\mc T}^{k}:= S_{\mc T}^k(I_{\mc T}^{k-1},x^\ast)=(I_{\mc T}^{k-1},\widetilde{z_k},\widetilde{t_k},\widetilde{\text{m}^{\text{comp}_k}},\widetilde{\bar a_k},$ $\widetilde{\bar b_k},\widetilde{v_k},$ $\widetilde{\text{coins}})$. \end{itemize} Finally, a trivial simulator $\widetilde I_{\mc T}^{K}:=S^K_{\mc T}(I^{K-1}_{\mc T},x^\ast)$ for $V^K_{\mc T}(\bar{\mathcal I}^{K-1})$ is obtained by simply generating an encryption of $x^\ast$ and outputting: $(I^{K-1}_{\mc T},\widetilde{[[x^\ast]]})$. \begin{proposition}\label{prop:sim_target} $S_{\mc T}^{k}(I_{\mc T}^{k-1},x^\ast)\stackrel{c}{\equiv}V^{k}(\bar{\mathcal I}^{k-1})$, for \mbox{$k=0,\ldots,K$}. \end{proposition} \begin{proof} For $0\leq k\leq K-1$, $(\text{coins}_{2k},\text{coins}_{3k})\stackrel{s}{\equiv}\widetilde{\text{coins}_k}$ because they are generated from the same distributions. Similarly, $\widetilde{z_k}\stackrel{s}{\equiv}z_k$ because $\widetilde{\rho_k}\stackrel{s}{\equiv}b_k-a_k+\rho_k$. From the discussion above, the same holds for $\widetilde{\text{m}^{\text{comp}_k}}$ and their counterparts $\text{m}^{\text{comp}_k}$. Furthermore, $(\widetilde{t_k},\widetilde{\bar a_k},\widetilde{\bar b_k},\widetilde{v_k})$ are statistically indistinguishable from $(t_k,\bar a_k, \bar b_k,v_k)$ due to the way they are generated, and $\widetilde{v_k}$ being consistent with $\widetilde{t_k}$. Thus, $S^{k}(I_{\mc T}^{k-1},x^\ast)\stackrel{c}{\equiv}V^{k}(I_{\mc T}^{k-1}\cup \mathcal I)$. For $k=K$, $S^{K}(I_{\mc T}^{K-1},x^\ast)\stackrel{c}{\equiv}V^{K}(\bar{\mathcal I}^{K-1})$ trivially follows from the fact that both $[[x^\ast]]$ and $\widetilde{[[x^\ast]]}$ are decrypted in $x^\ast$. \end{proof} Hence, we obtained that $\widetilde I_{\mc T} ^k \stackrel{c}{\equiv} I_{\mc T}^k$, for $k=0,\ldots,K$. The essence of the proof of Proposition~\ref{prop:sim_target} is that all the messages the target receives are blinded with large enough random noise and the comparison bits are randomized. This results in the messages being statistically indistinguishable from random values of the same length, which means that the extra information included in $(I^{k-1}_{\mc T},x^\ast)$ from the previous iterations cannot be used to extract other information about the current values at iteration~$k$. The next corollary then follows from Proposition~\ref{prop:sim_target}: \begin{corollary}\label{corr:consec_steps_target} $S_{\mc T}^k(\widetilde I_{\mc T}^{k-1},x^\ast)\stackrel{c}{\equiv} S_{\mc T}^k(I_{\mc T}^{k-1},x^\ast)$, equivalent to $S_{\mc T}^{k+1}(S_{\mc T}^{k}(I_{\mc T}^{k-1},x^\ast),x^\ast)\stackrel{c}{\equiv}S^{k+1}_{\mc T}(V^{k}(\bar{\mathcal I}^{k-1}),x^\ast)$, for any $k=1,\ldots,K-1$. \end{corollary} Finally, we construct a simulator $S_{\mc T}(I_{\mc T})$ for the execution of Protocol~\ref{alg:main_alg} and we show that its view will be statistically indistinguishable from $V_{\mc T}(\mathcal I)$. To this end, we define the following sequence, from which we drop the input $x^\ast$ to the simulators to not burden the notation: \begin{align*} V_{\mc T}(\mathcal I)=H_0(\mathcal I) &= V^K_{\mc T}(\bar{\mathcal I}^{K-1})\\ H_1(\mathcal I,x^\ast) &= S^K_{\mc T}\circ V^{K-1}_{\mc T}(\bar{\mathcal I}^{K-2})\\ H_2(\mathcal I,x^\ast) &= S^K_{\mc T}\circ S^{K-1}_{\mc T} \circ V^{K-2}_{\mc T}(\bar{\mathcal I}^{K-3})\\ &~\vdots\\ H_{K}(\mathcal I,x^\ast) &= S^K_{\mc T}\circ S^{K-1}_{\mc T}\circ \ldots\circ S^1_{\mc T}\circ V^0_{\mc T}(\mathcal I)\\ S_{\mc T}(I_{\mc T})=H_{K+1}(I_{\mc T}) &= S^K_{\mc T}\circ S^{K-1}_{\mc T}\circ \ldots\circ S^1_{\mc T}\circ S^0_{\mc T}(I_{\mc T}) \end{align*} By transitivity, $H_0$ and $H_{K+1}$ are statistically indistinguishable if: \[ H_0 \stackrel{c}{\equiv} H_1\stackrel{c}{\equiv} \ldots \stackrel{c}{\equiv} H_k \stackrel{c}{\equiv} H_{k+1} \stackrel{c}{\equiv} H_{k+2} \stackrel{c}{\equiv} \ldots \stackrel{c}{\equiv} H_{K+1}. \] The result follows from induction on Corollary~\ref{corr:consec_steps_target}. In conclusion, we obtain that $S_{\mc T}(I_{\mc T})\stackrel{c}{\equiv} V_{\mc T}(\mathcal I)$ which verifies that Protocol~\ref{alg:main_alg} achieves privacy with respect to the target node. The proof of Theorem~\ref{thm:main_alg} is now complete. \hfill \qed \subsection{Proof of Theorem~\ref{thm:main_alg_coal}}\label{subsec:proof_thm2} We will now show that any coalition consistent with the setup of Propositions~\ref{prop:assum_1},~\ref{prop:assum_2} and with the assumption that the cloud and target node do not collude does not gain any new information about the private data of the honest agents, other than what can be inferred solely from the inputs and outputs of the coalition. As mentioned in Section~\ref{sec:security_proof}, the agents in a coalition only add their inputs to the view of the coalition, but do not have any messages in the protocol. \subsubsection{Simulator for the cloud $\mathcal C$ and $\bar p$ agents $\mathcal A_i$} Consider the coalition between a set of agents $\mathcal A_{i=1,\ldots,\bar p}$ and the cloud $\mathcal C$, which has the inputs $(\{b_i\}_{i=1,\ldots,\bar m},\{c_i\}_{i=1,\ldots,\bar n},A_{\mc C},Q_{\mc C})$ and no output from the execution of Protocol~\ref{alg:main_alg}. Since $\bar p < p$, the coalition is not able to compute $\mu_1$ by itself, and the semantic security of the Paillier cryptosystem is again enough to ensure the privacy of the rest of the sensitive data. A simulator for this coalition can be constructed following the same steps in Section~\ref{subsubsec:cloud} on the augmented inputs defined above, from the fact that every value the cloud receives is encrypted by a semantically secure cryptosystem. \subsubsection{Simulator for the target node $\mathcal T$ and $\bar p$ agents $\mathcal A_i$} Consider the coalition between a set of agents $\mathcal A_{i=1,\ldots,\bar p}$ and the target node $\mathcal T$, which has the inputs $(\{b_i\}_{i=1,\ldots,\bar m},\{c_i\}_{i=1,\ldots,\bar n},sk_{\mc T},sk_{DGK})$ and output $(x^\ast)$ from the execution of Protocol~\ref{alg:main_alg}. If both matrices $Q_{\mc C}, A_{\mc C}$ are public, if there exists $i\in\{1,\ldots,\bar p\}$ such that $a_i^\intercal x^\ast < b_i$, the coalitions finds out $\mu^\ast_i=0$, which comes from public knowledge in the KKT conditions~\eqref{eq:KKT4}. From this, the coalition is able to find some coins of the cloud: $r_K$ and $s_K$ associated to element $i$. However, these values are independent from the rest of the parameters and do not reveal any information about the private inputs of the parties. Apart from this, the coalition is not able to compute any private data of the honest parties from the execution of the protocol, due to the secure blinding. A simulator for this coalition can be build by following the same steps as described in Section~\ref{subsubsec:target} on the augmented inputs defined above. The proof is now complete.\hfill \qed \subsection{Sketch proof of Theorem~\ref{thm:alternative}}\label{subsec:proof_thm3} Following similar steps as in the proof of the main protocol in Appendix~\ref{subsec:proof_thm1}, simulators for the involved parties can be constructed from their inputs and outputs. A simulator for the cloud randomly generates the messages and the indistinguishability between the views is guaranteed by the indistinguishability of the Pailler encryptions. The outputs of Protocol~\ref{alg:iteration_alternative} to the target node are different than the original Protocol~\ref{alg:iteration}. This means that the outputs for the target node in Protocol~\ref{alg:main_alg} with the iterations as in Protocol~\ref{alg:iteration_alternative} are $(x^\ast, \{(r_k)_i(\bar\mu_k)_i\}_{i=1,\ldots,m,~k=0,\ldots,K-1})$. In this case, a simulator for the target node simply outputs its inputs as messages in the view, and indistinguishability follows trivially. For the multi-party privacy, simulators for the allowed coalitions can be easily constructed by expanding the simulators for the cloud and the target node with the auxiliary inputs from the agents in the coalition.\hfill \qed \subsection{Proof of Proposition~\ref{prop:assum_1}}\label{subsec:proof_assum_1} The coalition has access to the following data, which is fixed: $A_{\mc C}, Q_{\mc C}, x^\ast, \{b_i\}_{1,\ldots,\bar m}, \{c_i\}_{i,\ldots, \bar n}$. Proof of (1). We need to address two cases: the non-strict satisfaction of the constraints and the equality satisfaction of the constraints. \noindent(I) Suppose there exists a solution $\{b_i\}_{\bar{m}+1, \ldots, m}$ and $\{c_i\}_{\bar{n}+1, \ldots, n}$ and $\mu$ to the KKT conditions such that $a_i^\intercal x^\ast < b_i$ for some ${\bar{m}+1} \leq i \leq m$. In particular this implies that $\mu_i = 0$. Then define $c' := c_{\mc A}$, $\mu' := \mu$ and $b'$ such that $b'_j := b_j$ for all $j \neq i$ and $b_i'$ to take any value $b_i' \geq a_i^\intercal x^\ast$. The new set of points $(b', c', \mu')$ is also a solution to the KKT conditions, by construction. \noindent(II) Alternatively, suppose there exists a solution $\{b_i\}_{\bar{m}+1, \ldots, m}$ and $\{c_i\}_{\bar{n}+1, \ldots, n}$ and $\mu$ to the KKT conditions such that $a_j^\intercal x^\ast = b_j$ for all $j={\bar{m}+1}, \ldots, m$. Consider there exists a vector $\delta$ that satisfies $\delta\succeq 0$ and $A_{21}^\intercal \delta = 0$. Compute $\epsilon \geq 0$ as: $\epsilon = \min \left(\frac{\mu_k}{\delta_k}\right)_{\delta_k>0, k=\bar m+1,\ldots,m}$. Then, we construct $\mu' := \mu - \epsilon \left[ \begin{array}{c} 0 \\ \delta \end{array}\right]$ that satisfies $\mu'\succeq 0$ and $\mu'_i =0$ for some ${\bar{m}+1} \leq i \leq m$ that is the argument of the above minimum. Furthermore, define $c' := c_{\mc A} + \epsilon\left[ \begin{array}{c} 0\\ A_{22}^\intercal \delta \end{array}\right] $ and $b'$ such that $b'_j := b_j$ for all $j \neq i$ and $b'_i $ to be any value $b'_i > b_i$. Then $(b', c', \mu')$ is also a solution to the KKT conditions. More specifically, the complementarity slackness condition holds for all $j={\bar{m}+1}, \ldots, m$: \begin{equation} \begin{aligned} &~\mu'_j ~ (a_j^\intercal x^\ast - b'_j) =&& \mu_j' (\underbrace{a_j^\intercal x^\ast - b_j}_{=0}) = 0, ~~ j \neq i\\ & \underbrace{\mu'_i}_{=0} (a_i^\intercal x^\ast - b'_i)=&& 0, ~~ j = i. \end{aligned} \end{equation} Then we can check the gradient condition: \begin{align} \begin{split} &Q_{\mc C} x^\ast + A_{\mc C}^\intercal \mu' + c' =\\ &Q_{\mc C} x^\ast + A_{\mc C}^\intercal \left( \mu - \epsilon \left[ \begin{array}{c} 0 \\ \delta \end{array}\right] \right) + \left(c + \epsilon\left[ \begin{array}{c} 0\\ A_{22}^\intercal \delta \end{array}\right] \right) \\ &=Q_{\mc C} x^\ast + A_{\mc C}^\intercal \mu + c_{\mc A} = 0. \end{split} \end{align} Hence, $b'\neq b_{\mc A}$ satisfies the KKT conditions and the coalition cannot uniquely determine $b_{\mc A}$. Proof of (2). Consider a solution $\{b_i\}_{\bar{m}+1, \ldots, m}$ and $\{c_i\}_{\bar{n}+1, \ldots, n}$ and $\mu$ to the KKT conditions. For some $\epsilon >0$ define $\mu' := \mu + \epsilon \left[ \begin{array}{c} 0 \\ \delta \end{array}\right]$ and $c' := c_{\mc A} - \epsilon\left[ \begin{array}{c} 0\\ A_{22}^\intercal \delta \end{array}\right] $. Define $b'$ such that for all $j$, it holds that $b'_j = a_j^\intercal x^\ast$. Then $(b', c', \mu')$ is also a solution to the KKT conditions. Specifically, it follows that $\mu'\geq 0$. Moreover the complementarity slackness condition holds by construction of $b'$, and as before the gradient condition holds. Hence, $c'\neq c_{\mc A}$ satisfies the KKT solution, and the coalition cannot uniquely determine $c_{\mc A}$. \hfill \qed

1fb0a10d2ba3dd194761e0e8ffb549a08276e0d7

\section*{Results and analysis} \subsection*{Proteins with entangled loops are not rare} We use the contact GE parameter ${G'_c}$ to find protein domains with at least one loop $\gamma_i$ intertwined with a ``thread'' $\gamma_j$, which is another portion of the protein (Fig.~\ref{fig:sketch}d-f). More precisely, we can associate ${G'_c}(i,j)$ to a given loop-thread pair by using the Gauss double integral described in {\it Materials and Methods}. By maximizing $|{G'_c}(i,j)|$ over all the possible threads $\gamma_j$ we assign an entanglement score ${G'_c}(i)$ to the loop and, by further maximizing $|{G'_c}(i)|$ over all loops $\gamma_i$, we find the entanglement ${G'_c}$ of the protein. At variance with similar quantities defined for closed curves, ${G'_c}$ is a real number. Yet, we define a loop $\gamma_i$ in a configuration as in Fig.~\ref{fig:sketch}d to be {\it entangled} if $|{G'_c}(i)|\ge1$. Such a threshold is natural because a linking number $|L|=1$ is the minimum value that guarantees that two closed curves are linked~\cite{ricca2011}. In a data set of $16968$ protein domains, $5375$, the $31.7$\%, host at least one entangled loop. We also monitor the value ${L'}$ of the {\it linking entanglement} (LE) for a single protein, defined as ${G'_c}$ for two subchains that are both loops, as in Fig.~\ref{fig:sketch}c. In Fig.~\ref{fig:ex} we show five examples of ``entangled'' protein domains, along with their respective values of ${G'_c}$ and ${L'}$. \begin{figure}[t] \includegraphics[width=\linewidth]{Figure-3-Seno.pdf} \caption{{(A)} Plot of ${L'}$ vs ${G'_c}$ for each protein in the CATH database; the five proteins shown in Fig.~\ref{fig:ex}a-f are highlighted with the corresponding letter. {(B)} Smoothed histogram of data with significant linking ($|{L'}|>1/2$). The highest probability is around ${G'_c}\simeq {L'}\simeq 1$. The data with the values of ${L'}$ and ${G'_c}$ computed for each protein in the CATH database are available at http://researchdata.cab.unipd.it/id/eprint/123.} \label{fig:GL} \end{figure} The non trivial entanglement features of protein structures, when analyzed with GE and LE, are apparent in Fig.~\ref{fig:GL}a, where protein domains are represented in the ${L'}$ vs ${G'_c}$ space. All the points lie in the region $|{L'}| \leq |{G'_c}|$ because the latter quantity is defined as an extremum over a wider subset. A typical example with ${G'_c} \simeq {L'} \simeq 1$ is shown in Fig.~\ref{fig:ex}a. As expected, however, there are cases of proteins with at least one entangled loop ($|{G'_c}|\ge1$) and all pairs of loops with negligible $|{L'}|$. These proteins corresponds to the conformation sketched in Fig.~\ref{fig:sketch}d and to the natural protein represented in Fig.~\ref{fig:ex}b. In other cases, the difference between $|{G'_c}|$ and $|{L'}|$ is large, even in the presence of linked loops. This is due to the behavior of the protein portion which, after threading the first loop, forms a second loop linked with it, and then continues to wind around it without further looping, see Fig.~\ref{fig:ex}c. It is interesting to observe that in several other cases the GE has a different sign with respect to the LE. This may take place if the chain winds around itself with opposite chiralities in different portions of the same protein. An example is shown in Fig.~\ref{fig:ex}d-e. One of the most entangled structures found in the database, with ${G'_c} \simeq {L'} \simeq -3$, is shown in Fig.~\ref{fig:ex}f. Fig.~\ref{fig:GL}a shows that the GE is distributed over a broad spectrum of values and that the threshold $|{G'_c}|\ge1$ for entangled loops is conservative enough. Clusters emerge in the density plot of Fig.~\ref{fig:GL}b, where the majority low LE points are removed by excluding data with $|{L'}|<1/2$ (see also Fig.~S1, which is an enlargement of Fig.~\ref{fig:GL}a). The clusters are found around ${L'}\simeq\pm 1$ vs ${G'_c}\simeq\pm 1$, in particular the most populated region has ${G'_c}\simeq {L'}\simeq 1$. For further analysis, we consider only the GE indicator, which captures more varieties of entangled motifs than the LE (e.g.~winding without linking, see Fig.~\ref{fig:ex}b). \subsection*{Entangled loops are found more frequently on the C-terminal side of the corresponding intertwining segment} In the definition of ${G'_c}(i,j)$, the role of the loop $\gamma_i$ cannot be exchanged with that of the other chain portion $\gamma_j$. This feature of the GE may be exploited to detect possible asymmetries in the respective location along the backbone of $\gamma_i$ and $\gamma_j$. The score ${G'_c}(i)$ associates to a given loop $\gamma_i$ the open arm with which it is mostly entangled. We consider separately the following two cases: when the threading arm is between the N-terminus and the loop (N-terminal thread, see Fig.~\ref{fig:sketch}e) or when it is between the loop and the C-terminus (C-terminal thread, Fig.~\ref{fig:sketch}f). \begin{figure*}[tb] \includegraphics[width=\linewidth]{Figure-4-Seno.pdf} \caption{(a) For four cases (see legend), distributions of the loop-thread sequence separation $s$. Error bars are based on the effectively independent countings determined through the clustering procedure. (b) For the separate cases of N- and C-terminal threads (see legend), tails of the distributions of the loop entanglement for $|{G'_c}(i)|>1/2$. Error bars are based on the effectively independent countings determined through the clustering procedure.} \label{fig:as} \end{figure*} We now focus on the properties of entangled loops ($3.75\%$ of the total) and we count how many of them are classified as N-terminal threads or C-terminal threads. In principle, there is no reason to expect one of the two classes to be more populated than the other. In order to discuss carefully the statistical significance of possible asymmetries, we need to take into account that some degree of correlation occurs in the counting of entangled loops. In fact, different loops can belong to essentially the same topological configuration, for example when a protein arm intertwines both with the loop formed between amino acids $i_1$ and $i_2$, and with the one formed between $i_1+1$ and $i_2$. Thus, we employ a clustering procedure based on a pairwise distance defined between loops (see {\it Materials and Methods} for details). By using the effective countings given by the clustering procedure, we find that the fraction of N-terminal (C-terminal) threads within entangled loops is $0.55$ ($0.45$). The highlighted bias in favor of the N-terminal threads (Fig.~\ref{fig:sketch}e) against the C-terminal ones (Fig.~\ref{fig:sketch}f) is statistically significant at the level of $14$ standard deviations. A somewhat similar result was found by studying topological barriers in protein folding~\cite{norcross2006}. One may ask whether N-terminal threads are favored, simply, because entangled loops are by themselves closer to the C-terminus, without the need of considering the $\left|{G'_c}\right|$ maximization that selects the intertwining thread. In order to check this, we consider a random reference case, whereby one putative threads is sampled randomly for each entangled loop, with uniform probability across all possible segments non overlapping with it. As a matter of fact, putative N-terminal threads are not favored in the reference case. We find instead a small bias in the opposite direction; namely, the fraction of putative N-terminal (C-terminal) threads within entangled loops is $0.487$ ($0.513$). This small but statistically significant ($3.5$ standard deviations) imbalance, suggests that entangled loops are slightly, if any, located closer to the N-terminus, thereby highlighting even more that the favored placement of the intertwining thread to the entangled loop N-terminal side is a genuine effect. \subsection*{Entangled loops favor positive chirality only for N-terminal threads} The formation of an entangled structure is not simple, as it requires a non local concerted organization of the amino acids in space, where a crucial role is played by the order of formation of different native structural elements along the folding pathway~\cite{sulkowska2009}. A misplaced nucleation event in the early stages of the folding pathway might prevent the protein from folding correctly. Dealing with spontaneous ``in vitro'' refolding, there is no reason to expect the folding order of different elements to be related to a preferential specific direction along the chain. Nevertheless, an asymmetry can be envisaged if a protein folds cotranslationally, according to the following argument. For the C-terminal thread, the loop might be formed in the early folding stages, making it difficult for the rest of the protein to entangle with it and thus to reach the native conformation. Conversely, for the N-terminal thread, the loop could wrap more easily around the open threading arm, already folded in its native conformation, after ejection from the ribosome. If confirmed, this picture would explain the asymmetry we observe between N- and C-terminal threads. The latter could be anyway interpreted as a possible fingerprint of an evolutionary process, intimately related to entanglement regulation driven by cotranslational folding. Such conclusion is corroborated by looking, separately for C- and N-terminal threads, at the normalized distributions of loop-thread sequence separations $s$, plotted in Fig.~\ref{fig:as}a. The distributions for the random reference case (empty symbols) are very similar for N-terminal (triangles) and C-terminal (circles) threads, showing again that if the putative thread is chosen randomly no asymmetry is present. One would expect a uniform reference distribution for loops at a fixed distance from the relevant chain terminus (N for N-threads). The regular decay observed for increasing $s$ is due instead to the fact that different protein domains have different lengths and different loops are located differently along the backbone. On the other hand, the $|{G'_c}(i,j)|$ maximization leading to ${G'_c}(i)$ selects preferentially arms that start just after (or before) the loop, at a distance of one or few amino acids. This is similar to what already observed for pokes~\cite{khatib2009}, and reflects the fact that a rapid turning of the protein chain is the simplest way for maximizing the mutual winding between two subchains. However, loop-thread pairs that are one unit distance apart are significantly more favored for C-terminal (squares) than for N-terminal (diamonds) threads (notice the logarithmic scale and the associated statistical errors), showing again a genuine asymmetry between the two cases. A Kolmogorov-Smirnov test yields $P=4\cdot10^{-58}$ for the null hypothesis that the two distributions are the same. Interestingly, the resulting enhancement at intermediate separations ($5<s<20$) allows N-terminal threads to follow closely the reference decay. Consistently with cotranslational folding, N-terminal threads could allow for more complex topological structures with on average larger separations, when compared to C-terminal threads. Accordingly, the distribution of ${G'_c}(i)$ values for both the N- and C-terminal threads, shown in Fig.~\ref{fig:as}b, highlights that the values around ${G'_c}(i)\approx 1$ are more probable in the former case. Strikingly, this happens only for positive ${G'_c}(i)$ values, whereas for negative ones there is a small bias favoring C-terminal threads. As a matter of fact, we find C-thread entangled loops to be balanced between positive and negative chiralities within the $0.3\%$, whereas N-thread entangled loops are highly biased (74\%) in favor of positive chiralities, ${G'_c}(i)>0$. \subsection*{Natural proteins are less entangled than protein-like compact conformations} In the ensemble of the CATH domains there are $3617208$ loops, of which $135530$ ($3.75\%$) are entangled. To assess whether this fraction is small or large we compare it with an analogous quantity computed in an unbiased reference state formed by a set of putative alternative compact conformations (i.e. rich in secondary structures) that a protein could in principle adopt. This ensemble is found in a poly-valine ``VAL60'' database~\cite{cossio2010}, obtained with an all atom simulation that accurately sampled the configurational space of a homopolypeptide formed by $60$ valine amino acids (see {\it Materials and Methods} for details). For a proper comparison with VAL60, we restrict our CATH database only to the proteins of comparable length, filtering out $772$ proteins with length $n$ from $n=55$ to $n=64$ amino acids. In this reduced ``CATH60'' ensemble of natural proteins there are $47954$ loops, of which $138$ ($0.3\%$) are entangled. There are $19$ proteins ($2.46\%$) hosting at least one entangled loop. These values are of course lower than those for the full CATH ensemble, in which longer proteins can host more entanglement. In VAL60 there are $2284693$ loops, of which $57577$ are entangled ($2.52\%$), a fraction ten times larger than for natural proteins of CATH60. Similarly, $3560$ out of the $30064$ VAL60 structures host at least one entangled loop ($11.8\%$), a fraction five times larger than for natural proteins. \begin{figure}[tb] \includegraphics[width=\linewidth]{Figure-5-Seno.pdf} \caption{For both natural protein domains of length $n$ in the range $55\le n \le 64$ from the CATH database, and the VAL60 ensemble of homopolypeptides, we plot the normalized histogram of ${G'_c}(i)$ for loops of length $m$ in the intervals $20\le m\le 24$ {(a)},$30\le m\le 34$ {(b)}, and $40\le m\le 44$ {(c)}. {(d)} For natural protein domains and the VAL60 ensemble, root mean squared ${G'_c}(i)$ as a function of the loop length $m$.} \label{fig:histo} \end{figure} However, it is known that, presumably for kinetic reasons~\cite{cossio2010}, VAL60 is characterized by loops on average longer than those of natural proteins. Consequently, to avoid any possible bias in the comparison, we divide loops in classes of homogeneous length $m$. For some classes, the normalized histogram of the GE for CATH60 and VAL60 datasets are plotted in Fig.~\ref{fig:histo}a-c. In all cases it is apparent that the range of ${G'_c}(i)$ is wider for the VAL60 homopolypeptides than for the natural proteins. The deep difference between the two distributions can be appreciated in Fig.~\ref{fig:histo}d, where the root mean squared ${G'_c}(i)$ is plotted as a function of the loop length: the values for VAL60 are always significantly higher than those for natural proteins. Note that the root mean squared ${G'_c}(i)$ increases with $m$ only up to half of the protein length. From there on, the remaining subchain starts getting too short to entangle. In conclusion, we have a clear statistical evidence that entangled loops occur less frequently in natural proteins with respect to random compact protein-like structures. \subsection*{Amino acids at the ends of entangled loops are frustrated} In the preceding sections we provided two independent evidences that, although entangled loops are not rare in natural protein structures, their occurrence and position along the backbone chain are kept under control. A possible reason is the need to limit potential kinetic traps in the folding process brought about by entangled loops, for example by deferring their formation to the latter stages of the folding pathway. Thus, we expect to detect a related fingerprint in the specific amino acids found in contact with each other at the end of entangled loops (``entangled contacts''). We check whether such amino acids share the same statistical properties of the amino acids forming any possible contact (``normal contacts''). The frequency with which two amino acids are in contact is typically employed to estimate knowledge based potentials~\cite{samudrala1998,lazaridis2000}. In a nutshell, if two amino acids $a$ and $b$ occur to be in contact more frequently than on average, they are expected to manifest a mutual attraction and are therefore characterized by a negative effective interaction energy $E_{\rm norm}(a,b)$ (see {\it Materials and Methods}). If effective interaction energies are computed by restricting the analysis only to the entangled contacts, a new set of entangled contact potentials $E_{\textrm{GE}}(a,b)$ can be derived. The discrepancies between such potentials and the normal ones can be conveniently captured by an enrichment score $\Delta E_{\rm enr}(a,b)$. A negative enrichment score $\Delta E_{\rm enr}(a,b)<0$ implies that $(a,b)$ are more frequently in contact when they are at the ends of entangled loops, and vice-versa for positive scores. Fig.~\ref{fig:correl} shows that $\Delta E_{\rm enr}(a,b)$ anticorrelates with $E_{\rm norm}(a,b)$. This correlation is statistically significant. The Pearson correlation coefficient is $r=-0.31$, with a $P$-value of $2\times 10^{-6}$. The Spearman rank correlation is $\rho=-0.23$ with a $P$-value of $6 \times 10^{-4}$. \begin{figure}[tb] \includegraphics[width=\linewidth]{Figure-6-Seno.pdf} \caption{Scatter plot of the enrichment score $\Delta E_{\textrm{enr}}(a,b)$ vs normal contact potential $E_{\textrm{norm}}(a,b)$. Each point is for an amino acid pair $(a,b)$ and is colored according to amino acid types: black for pairs of aromatic residues (HIS, PHE, TRP, TYR); magenta for CYS-CYS; green for the rest. The dashed line is a linear fit with slope $-0.12$. Error bars are computed with a boostrapping procedure and we plot only errors for $\Delta E_{\textrm{enr}}$ as those for $E_{\textrm{norm}}$ are smaller than the symbol size.} \label{fig:correl} \end{figure} The anticorrelation of Fig.~\ref{fig:correl} has an important consequence: pairs of amino acids that in a globular protein interact strongly ($E_{\rm norm}(a,b)<0$, mainly hydrophobic amino acids) are present less often ($\Delta E_{\rm enr}(a,b)>0$) in entangled contacts, while amino acids that typically interact weakly ($E_{\rm norm}(a,b)> 0$, mainly polar and hydrophilic amino acids) are instead more abundant ($\Delta E_{\rm enr}(a,b)<0$) at the ends of entangled loops. We checked that this result is not trivially due to entangled contacts being preferentially located on the protein surface, finding that residues involved in entangled contacts are even slightly more buried in the protein interior than those involved in normal contacts (see Fig.~S2). The deep difference between the two set of scores $E_{\rm norm}(a,b)$ and $E_{\textrm{GE}}(a,b)$ emerges clearly from the graphical representations in Fig.~\ref{fig:table} of $E_{\rm norm}(a,b)$ and $\Delta E_{\rm enr}(a,b)$, in which positive and negative values are marked red and blue, respectively, whereas white boxes mark scores that are not significant within the related statistical uncertainty. The blue spots in Fig.~\ref{fig:table}a represent interactions between amino acids that interact frequently with each other (mainly hydrophobic pairs), whereas the red area is populated by amino acids that are rarely in contact (mainly polar pairs). In Fig.~\ref{fig:table}b, the blue spots highlight amino acids that have decreased their energy score and which are therefore more present at the ends of the entangled loops than in normal contacts. These include mainly polar amino acids. Note that proline is particularly enriched at the end of entangled loops. The red spots in Fig.~\ref{fig:table}b indicate amino acids which are less present at the ends of the entangled loops than in normal contacts. These include mainly hydrophobic ones. The case of cysteine self-interaction is pedagogical: the strongest attractive interaction between amino acids turns out to be the more diminished one at the end of entangled loops (see also Fig.~\ref{fig:correl}), consistently with the very low number of linked loops closed by disulphide bonds (Fig.~\ref{fig:sketch}c) that was found in the PDB~\cite{dabrowski2017topological}. \begin{figure*}[t] \includegraphics[width=\linewidth]{Figure-7-Seno.pdf} \caption{ {(a)} Normal contact potential $E_{\textrm{norm}}$; amino acids are ranked from left to right (top to bottom) with increasing average $E_{\textrm{norm}}$ (over row/column). {(b)} Enrichment score $\Delta E_{\textrm{enr}}$ for entangled contacts. Different backgrounds are used for highlighting negative and positive values: blue for $E< -E_0$, light blue for $-E_0\le E \le0 $, pink for $0<E\le E_0$, and red for $E>E_0$ with $E_0=35$. White is used for scores that differ from zero less than the corresponding statistical uncertainty, computed by means of a bootstrapping procedure.} \label{fig:table} \end{figure*} Interestingly, the four aromatic amino-acids (HIS, PHE, TRP, TYR) violate the general trend. Interactions between aromatic pairs are found in the bottom-left quadrant of Fig.~\ref{fig:correl}. Despite being very frequent in normal contacts (all their mutual entries are dark blue in Fig.~\ref{fig:table}a), they become even more abundant when at the ends of entangled loops (still blue in Fig.~\ref{fig:table}b), highlighting a special role likely played by aromatic rings in such complex structures. Fig.~\ref{fig:correl} and Fig.~\ref{fig:table}b provide clear evidence for the existence of an evolutionary pressure shaping the amino acid sequences. This natural bias weakens energetically the contacts which close entangled loops, consistently with the argument that a too early stable formation of the loop could prevent the correct folding of the full protein. These results are very robust to changes in the ${G'_c}$ threshold used to define entangled contacts, in the minimum length $m_0$ of the considered subchains, and to the introduction of a minimum loop-thread separation $s_0$, see Figs.~S3-S5. \section*{Discussion and Conclusions} With the notion of Gaussian entanglement we extend the measure of mutual entanglement between two loops to any pair of open subchains of a protein structure. This allows us to perform an unprecedented large scale investigation of the self entanglement properties of protein native structures, through which we identify and locate a large variety of entangled motifs (Fig.~\ref{fig:ex}), by focusing on the notion of ``entangled loop'', a loop intertwining with another subchain (Fig.~\ref{fig:sketch}d). Different entangled motifs can coexist in the same protein domain, even with opposite chiralities, and few domains exhibit a pair of loops intertwining even thrice around each other (see the examples in Fig.~\ref{fig:ex}c and Fig.~\ref{fig:ex}f, and points in Fig.~\ref{fig:GL}). Gaussian entanglement could be used to improve the classification of existing protein folds~\cite{Kolodny2013}, as previously done with Gauss integrals computed over the whole protein chain~\cite{Rogen2003}. Our analysis shows unequivocally (Fig.~\ref{fig:histo}) that, although entangled motifs are present in a remarkably high fraction, $32$\%, of protein domains, these host a lower amount of entangled loops than protein-like decoys produced with molecular dynamic simulations~\cite{cossio2010}. The question is then why natural folds avoid overly entangled conformations with otherwise plausible secondary structure elements. Are entangled loops obstacles for the folding process? If yes, how does Nature cope with them when they are present? To answer these questions, we recall that an efficient folding of proteins is fundamental for sustaining the biological machinery of cell functioning. The rate and the energetics of the protein folding process, which are defined by its energy landscape, are encoded in the amino acid sequence. Over the course of evolution, this landscape was shaped to allow and stabilize protein folding, avoiding possible slowdowns. We find indeed two clear hallmarks suggesting that the entangled loops in proteins are kept under control: (i) a statistically significant asymmetry in their positioning with respect to the other intertwining chain portion, which is consistent with cotranslational folding promoting the presence of entangled loops on the C-terminal side of the intertwining thread (see Fig.~\ref{fig:as}a); (ii) weak non optimized interactions between the amino acids in contact at the end of entangled loops, an example of energetic frustration (see Fig.~\ref{fig:correl} and Fig.~\ref{fig:table}). Both these findings suggest that the late formation of entangled loops along the folding pathway could be a plausible control mechanism to avoid kinetic traps. On the other hand, the additional stability that one expects to be provided by entangled and knotted structures can compensate for the presence of such weak interactions. Interestingly, interactions between aromatic amino acid pairs are promoted at the end of entangled loops (see Fig.~\ref{fig:table}b), suggesting that their presence could be related to the protein biological function. Whether entangled loops may have specific biological functions is an intriguing open question, as in the case of knots in protein domains~\cite{Dabrowski-Tumanski2016,jackson2017}. Finally, we detect a remarkable bias, favoring positive chiralities, that is present only for entangled loops on the C-terminal side of the intertwining thread (see Fig.~\ref{fig:as}b). This suggests that the observed chirality bias arises in the context of cotranslational folding. A simple possibility is that loop winding of the C-terminal part of the chain may have a preferred orientation when just released from the ribosome. Further work will be needed to test this speculation and to fully rationalize the chirality bias. As a matter of fact, the ribosome can discriminate the chirality of amino acids during protein synthesis~\cite{englander2015}. Stemming from works on glassy transitions~\cite{derrida1980,gross1984}, the concept of minimal frustration between the conflicting forces driving the folding process is a well established paradigm~\cite{bryngelson1987,frauenfelder1991,ferreiro2014} in protein physics. It has been further argued~\cite{ferreiro2014} that frustration is an essential feature for the folding dynamics and that it can give surprising insights into how proteins fold or misfold. Is it possible to reconcile the frustration detected at the ends of entangled loops with the minimal frustration principle? Let us assume that a non optimal ordering of the events along the folding pathways (for example, the formation of a loop which has then to be threaded by another portion of the protein to form an entangled structure) is highly deleterious. In order to prevent this, it could indeed be preferable to select suitable sub-optimal interactions. In fact, this would be a remarkable example of minimal frustration in action, having to compromise between topological and energetic frustration. Obviously, other data will be needed to confirm this proposed mechanism for the folding process, from both simulations and experiments. In either case, a simple protocol could consist in mutating into cysteine both residues at the ends of an entangled loop, provided no other cysteines are present in the sequence, and in assessing whether the folding is then hindered by the formation of a disulfide bridge in oxidizing conditions. In the context of knotted proteins, single molecule force spectroscopy techniques were shown to be particularly useful in controlling the topology of the unfolded state~\cite{rief2016}. Similarly, both ``in vivo'' folding experiments~\cite{hegner2017} and appropriate simulation protocols~\cite{cieplak2015,hoang2016,obrien2016} could be employed to test the possible role of cotranslational folding in determining the patterns detected for entangled motifs: double cysteine mutants would then be predicted to be more deleterious for the folding of C-terminal threads with respect to N-terminal threads. In all cases, it is essential to gather statistics over several different proteins before validating or rejecting our hypothesis; the signals that we reveal in this contribution are statistical in nature; therefore we do not expect all entangled loops to form late in the folding process nor all C-terminal threads to be cotranslationally disfavored. For example, it has been recently proposed, in the context of deeply knotted proteins, that loops formed by a synthesized earlier portion of the same protein can be actively threaded by nascent chains at the ribosome~\cite{dabrowski2018protein}. However, this is not in contradiction with our findings, since knotted proteins are much less frequent than the general entangled motifs discussed here. \section*{Materials and Methods} \subsection*{CATH database} We use the v4.1 release of the CATH database for protein domains, with a non-redundancy filter of 35\% homology~\cite{CATH}. To avoid introducing entanglement artificially for proteins with big gaps in their experimental native structures, we do not consider any protein in the CATH database that presents a distance $>10$\AA~ between subsequent C$_\alpha$ atoms in the coordinate file. We find that this selection keeps $N_{\textrm{prot}}=16968$ out of the available $21155$ proteins. CATH domain names such as 2bjuA02 refer to the 2nd domain from chain A with PDB code 2bju. The CATH database is available at http://download.cathdb.info/cath/releases/all-releases/v4\_1\_0/. \subsection*{Poly-valine database} The VAL60 database is an ensemble of $30064$ structures obtained by an exhaustive exploration of the conformational space of a 60 amino acid poly-valine chain described with an accurate all-atom interaction potential~\cite{cossio2010}. The exploration was performed with molecular dynamics simulations using the AMBER03 force field~\cite{duan2003} and the molecular dynamics package GROMACS~\cite{lindahl2001} and by exploiting a bias exchange metadynamics approach~\cite{piana2008} with 6 replicas. The simulation was performed in vacuum at a temperature of $400$ K. The conformations have been selected as local minima of the potential energy with a secondary structure content of at least $30\%$ and a small gyration radius. The protein-like character of VAL60 conformations was successfully tested by using different criteria commonly employed to assess the quality of protein structures~\cite{cossio2010}. The stability of a small subset of VAL60 structures was successfully tested even after mutation of all residues to Alanines. Crucially, it was observed that the VAL60 database contains almost all the natural existing folds of similar length~\cite{cossio2010}. However, these known folds form a rather small subset of the full ensemble, which can be thought as an accurate representation of the universe of all possible conformations physically attainable by polypeptide chains of length around $60$. A repository for the VAL60 database is available at http://dx.doi.org/10.5061/dryad.1922. \subsection*{Mathematical definition of the linking number and its computational implementation} The linking number between two closed oriented curves $\gamma_i=\{\vec r^{(i)}\}$ and $\gamma_j=\{\vec r^{(j)}\}$ in $\mathbb{R}^3$ may be computed with the Gauss double integral \begin{equation} \label{Gauss-int} {G} \equiv \frac{1}{4 \pi} \oint_{\gamma_i}\oint_{\gamma_j} \frac{\vec r^{(i)} -\vec r^{(j)}}{\left| \vec r^{(i)} - \vec r^{(j)}\right|^3} \cdot (d \vec r^{(i)} \times d \vec r^{(j)}) \end{equation} It is an integer number and a topological invariant~\cite{ricca2011}. If computed for open curves, it becomes a real number $G'$ (the GE) that quantifies the mutual entanglement between the curves~\cite{Doi1988,Panagiotou_JPA_2010,Panagiotou_PRE_2013,Baiesi_et_al_SciRep_2016,Baiesi_et_al_JPA_2017}. In proteins, piece-wise linear curves join the coordinates of subsequent C$_\alpha$ atoms. In particular, $\gamma_i$ is an open subchain joining C$_\alpha$ atoms from index $i_1$ to $i_2$ and similarly $\gamma_j$ is another nonoverlapping subchain from $j_1$ to $j_2$. We specialize to the configurations studied in Ref.~\cite{Baiesi_et_al_JPA_2017}, in which $i_1$ and $i_2$ amino acids are required to be in contact. In this study, the contact is present if any of the heavy (non hydrogen) atoms of residue $i_1$ is near any of the heavy atoms of residue $i_2$, namely they are at a distance at most $d=4.5$\AA. The ``contact'' Gaussian entanglement of these configurations (sketched in Fig.~\ref{fig:sketch}d-f) is named ${G'_c}(i,j)$. Since proteins are thick polymers and bonds joining C$_\alpha$ atoms are quite far from each other (compared to their length), we may approximate the integral [\ref{Gauss-int}] with a discrete sum. Given the coordinates $\vec r_i$ of C$_\alpha$'s, the average bond positions $\vec R_i \equiv \frac 1 2 ( \vec r_i + \vec r_{i+1} )$ and the bond vectors $\vec {\Delta R}_i = \vec r_{i+1} - \vec r_{i}$ enter in the estimate of ${G'_c}(i,j)$ for $\gamma_i$ and $\gamma_j$, \begin{equation} \label{Gij} {G'_c}(i,j) \equiv \frac{1}{4 \pi} \sum_{i=i_1}^{i_2-1} \sum_{j=j_1}^{j_2-1} \frac{\vec R_i - \vec R_j}{ \left|\vec R_i - \vec R_j\right|^3} \cdot ( \vec{\Delta R}_i \times \vec{\Delta R}_j) . \end{equation} We then associate a contact entanglement ${G'_c}(i)$ to a ``loop'' $\gamma_i$ as the extreme (i.e.~with largest modulus) ${G'_c}(i,j)$, for all ``threads'' $\gamma_j$, with $j_2-j_1\ge m_0$ ($m_0=10$). Finally, the contact entanglement ${G'_c}$ of a protein is the extreme of ${G'_c}(i)$ for all loops of length $m = i_2-i_1\ge m_0$. The linking entanglement ${L'}$ is equal to ${G'_c}$ for configurations with two loops as in Fig.~\ref{fig:sketch}c. It is not exactly the linking number $L$ because the two closures between contacts are not performed. \subsection*{Clustering procedure for counting effectively independent loops} Each entangled loop $\gamma$ is characterized by five numbers, its two indices ($i_1, i_2$), the indices of the threading portion ($j_1, j_2$), and the corresponding Gaussian entanglement ${G'_c}(i,j)$. It is thus natural to define a distance between two entangled loops $\gamma^A$ and $\gamma^B$ as \begin{widetext} \begin{equation} \label{distance} d_{AB} = \sqrt{\left(i^A_{1}-i^B_{1}\right)^2 + \left(i^A_{2}-i^B_{2}\right)^2 + \left(j^A_{1}-j^B_{1}\right)^2 + \left(j^A_{2}-j^B_{2}\right)^2 + w_g \left[{G'_c}(i^A,j^A)-{G'_c}(i^B,j^B)\right]^2}\;. \end{equation} \end{widetext} where $w_g$ is a weight to be defined. In order to count the effectively independent loops we used the following procedure within each protein in the CATH database: first we selected the loop with the largest number of neighbors, namely with the largest number of loops at a distance smaller than a threshold $d^*$. We assign the selected loop and all its neighbors to the same cluster, removing them from the running list of loops. We iterate this procedure until the running list is empty, so that each loop $\gamma_i$ belongs to a cluster with a given number of members $N_{C_i}$. Each loop is then included in all the statistics and distributions reported in the main text, with an effective counting weight $1/N_{C_i}$. By using $d^*=20$, $w_g=10^4$, we find an overall effective counting of 18041 independent entangled loops, a $13\%$ of the original 135530 countings. The results reported in the main text are qualitatively robust against reasonable variations of the $d^*$, $w_g$ parameters. The data with the values of ${G'_c}$ computed for each loop in each protein in the CATH database, and grouped after clustering, are available at http://researchdata.cab.unipd.it/id/eprint/123. \subsection*{Inference of statistical potentials} In order to estimate effective interactions between amino acids in protein structures, we use an established knowledge based approach~\cite{lazaridis2000}. Pairwise potentials can be obtained by analyzing databases of know protein conformations~\cite{Miyazawa1985}. These potentials are derived measuring the probability of an observable, such as the formation of a contact, relative to a reference unbiased state~\cite{samudrala1998}. The conversion of the probability in an energy is done by employing Boltzmann's law~\cite{sippl1990}. The first step includes characterizing the reference null space of possible pairs of amino acids. All amino acid pairs within each protein sum up to a grand total of $N$ generic {\it pairs} (i.e.~just combinatorial pairings not necessarily related to a spatial contact) in our ensemble of protein structures. In the same way, given two amino acid kinds $a$ and $b$, one sums up the occurrence of $a$-$b$ pairs within each protein to a grand total of $N(a,b)$ pairs in the ensemble. To quantify energies of ``normal'' contacts $E_{\rm norm}(a,b)$ between amino acids of type $a$ and $b$, we consider two amino acids to be in contact if any inter-residue pair of their side chain heavy atoms is found at a distance lower than $4.5$\AA. By considering only the ensemble of amino acids which are in contact within each protein, their total counting results in $N_c$ generic contacts. Similarly, the specific contacts between amino acids of kind $a$ and $b$ are summed up to a total $N_c(a,b)$. The statistical potentials for normal contacts are defined by comparing the frequencies~\cite{samudrala1998,cossio2012} \begin{align} f(a,b) = N(a,b) / N \,, \qquad f_c(a,b) = N_c(a,b) / N_c \,, \end{align} within the ensemble of all pairs or the ensemble of contacts, respectively. If $f_c(a,b)$ is relatively high compared to $f(a,b)$, it means that chemistry and natural selection favored the organization of native protein structures toward configurations where $a$ and $b$ types were in contact. Thus, the argument is that a lower potential energy should be associated to such contacts; the normal contact potentials are therefore given by \begin{align} E_{\rm norm}(a,b) = -\tau \log\frac{f_c(a,b)}{f(a,b)} \,, \end{align} where we introduced a parameter $\tau=100$ for the convenience of rescaling numbers and rounding them off to readable integers. We can compute a similar kind of potentials $E_{\textrm{GE}}(a,b)$ for ``entangled'' contacts, just restricting the analysis to the subset of contacts between amino acids that are at the end of an ``entangled loop'', defined as a loop $\gamma_i$ for which $|{G'_c}(i)|\ge 1$, that is a loop for which at least one thread $\gamma_j$ exists such that the corresponding $|{G'_c}(i,j)|\ge 1$. Within all proteins, in total we count $N_c^{G}$ of such contacts while the specific ones are $N_c^{G}(a,b)$, and hence \begin{align} f_c^{G}(a,b) & = N_c^{G}(a,b) / N_c^{G}\,,\\ E_{\textrm{GE}}(a,b) & = -\tau \log\frac{f_c^{G}(a,b)}{f(a,b)} \,. \end{align} To easily capture dissimilarities between $E_{\rm norm}(a,b)$ and $E_{\textrm{GE}}(a,b)$, we introduce an enrichment score defined as \begin{align} \Delta E_{\rm enr}(a,b) = -\tau \log\frac{f_c^{G}(a,b)}{f_c(a,b)} = E_{\textrm{GE}}(a,b) -E_{\rm norm}(a,b) \end{align} In all cases we have imposed a constraint on the pairs of amino acid considered: the two amino acids $i_1$ and $i_2$ in contact must have indices difference $i_2-i_1\ge m_0=10$. This threshold $m_0$ removes any eventual bias in comparing potentials due to the entanglement constraint, which requires entangled loops and threads of a minimal length to be present. We computed all statistical potentials, together with the related uncertainties, by using a bootstrapping procedure with 101 independent resamplings. The countings and the statistical scores obtained for each amino-acid pair are available at http://researchdata.cab.unipd.it/id/eprint/123. \vspace{1.0cm} \paragraph*{Acknowledgments} MB acknowledges support from Progetto di Ricerca Dipartimentale BIRD173122/17. FS and AT acknowledge R.~Battistuta and G.~Zanotti for fruitful discussions. We thank A.~Kabak{\c c}{\i}o{\u g}lu for a careful reading of the manuscript.

134de4542d1cadaded0a410c5955eee9282bf1b5

\section*{Acknowledgments}\label{sec:ack} We would like to thank Yu Feng, Kostas Ferles, Jiayi Wei, Greg Anderson, and the anonymous OOPSLA'18 reviewers for their thorough and helpful comments on an earlier version of this paper. This material is based on research sponsored by NSF Awards \#1712067, \#1811865, \#1646522, and AFRL Award \#8750-14-2-0270. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the U.S. Government. \section{Complexity Analysis} The complexity of our inductive synthesis algorithm is parametrized by 1) the number of functions to synthesize, 2) the size of ground relational specification, 3) the size of largest FTA in the HFTA, and 4) the bound of AST size for programs under consideration. Specifically, the ground relational specification can be relaxed in linear time $\mathcal{O}(k)$, where $k$ is the size of ground relational specification. Since the HFTA may contain at most $k$ FTAs, it can be constructed in $\mathcal{O}(km)$ time, where $m$ is the size of largest FTA in the HFTA. In particular, $m = \Sigma_{\delta \in \Delta} |\delta|$ where $|\delta| = w+1$ for a transition $\delta$ of the form $\sigma(q_1, \ldots, q_w) \to q$. Observe that \textsc{FindProgs} is a backtracking search procedure, its worst case complexity is $\mathcal{O}(m^{bn})$ where $n$ is the number of functions to synthesize, and $b$ is the bound of AST size for all programs under consideration. Therefore, the overall complexity is $\mathcal{O}(km + m^{bn})$. \section{Conclusion} In this paper, we have introduced \emph{relational program synthesis} ---the problem of synthesizing one or more programs that collectively satisfy a relational specification--- and described its numerous applications in real-world programming scenarios. We have also taken a first step towards solving this problem by presenting a CEGIS-based approach with a novel inductive synthesis component. In particular, the key idea is to construct a relational version space (in the form of a hierarchical finite tree automaton) that encodes all tuples of programs that satisfy the original specification. We have implemented this technique in a relational synthesis framework called \textsc{Relish}\xspace which can be instantiated in different application domains by providing a suitable domain-specific language as well as the relevant relational specifications. Our evaluation in two different application domains (namely, encoders/decoders and comparators) demonstrate that \textsc{Relish}\xspace can effectively synthesize pairs of closely related programs (i.e., inverses) as well as individual programs that must obey non-trivial $k$-safety specifications (e.g., transitivity). Our evaluation also shows that \textsc{Relish}\xspace significantly outperforms \textsc{EUSolver}\xspace, both in terms of the efficiency of inductive synthesis as well as its generalization power per CEGIS iteration. \section{Evaluation} We evaluate \textsc{Relish}\xspace by using it to automatically synthesize (1) string encoders and decoders for program inversion tasks collected from prior work~\cite{inversion-loris} and (2) string comparators to solve sorting problems posted on StackOverflow. The goal of our evaluation is to answer the following questions: \begin{itemize} \item How does \textsc{Relish}\xspace perform on various relational synthesis tasks from two domains? \item What is the benefit of using HFTAs for relational program synthesis? \end{itemize} \paragraph{\textbf{Experimental setup.}} To evaluate the benefit of our approach over a base-line, we compare our method against \textsc{EUSolver}\xspace~\cite{eusolver}, an enumeration-based synthesizer that won the General Track of the most recent SyGuS competition~\cite{sygus}. Since \textsc{EUSolver}\xspace only supports synthesis tasks in linear integer arithmetic, bitvectors, and basic string manipulations by default, we extend it to encoder/decoders and comparators by implementing the same DSLs described in Section~\ref{sec:impl}. Additionally, we implement the same CEGIS loop used in \textsc{Relish}\xspace for \textsc{EUSolver}\xspace and use the same bounded verifier in our evaluation. All experiments are conducted on a machine with Intel Xeon(R) E5-1620 v3 CPU and 32GB of physical memory, running the Ubuntu 14.04 operating system. Due to finite computational resources, we set a time limit of 24 hours for each benchmark. \subsection{Results for String Encoders and Decoders} In our first experiment, we evaluate \textsc{Relish}\xspace by using it to simultaneously synthesize Unicode string encoders and decoders, which are required to be inverses of each other. \paragraph{\textbf{Benchmarks.}} We collect a total of ten encoder/decoder benchmarks, seven of which are taken from a prior paper on program inversion~\cite{inversion-loris}. Since the 14 benchmarks from prior paper~\cite{inversion-loris} are essentially seven pairs of encoders and decoders, we have covered all their encoder/decoder benchmarks. The remaining three benchmarks, namely, \emph{Base32hex, UTF-32, and UTF-7}, are also well-known encodings. Unlike previous work on program inversion, our goal is to solve the considerably more difficult problem of simultaneously synthesizing the encoder \emph{and} decoder from input-output examples rather than inverting an existing function. For each benchmark, we use 2-3 input-output examples taken from the documentation of the corresponding encoders. We also specify the relational property $\forall x.$ \texttt{decode(encode(x))=x} and use the encoder/decoder DSLs presented in Section~\ref{sec:encoder}. \paragraph{\textbf{Main results.}} Our main experimental results are summarized in Table~\ref{tab:result-codec}, where the first two columns (namely, ``Enc Size'' and ``Dec Size'') describe the size of the target program in terms of the number of AST nodes.~\footnote{We obtain this information by manually writing a simplest DSL program for achieving the desired task.} The next three columns under {\bf \textsc{Relish}\xspace} summarize the results obtained by running \textsc{Relish}\xspace on each of these benchmarks, and the three columns under {\sc \bf \textsc{EUSolver}\xspace} report the same results for \textsc{EUSolver}\xspace. Specifically, the column labeled ``Iters'' shows the total number of iterations inside the CEGIS loop, ``Total'' shows the total synthesis time in seconds, and ``Synth'' indicates the time (in seconds) taken by the inductive synthesizer (i.e., excluding verification). If a tool fails to solve the desired task within the 24 hour time limit, we write T/O to indicate a time-out. Finally, the last column labeled ``Speed-up'' shows the speed-up of \textsc{Relish}\xspace over \textsc{EUSolver}\xspace for those benchmarks where neither tool times out. \input{tab-result-codec} As shown in Table~\ref{tab:result-codec}, \textsc{Relish}\xspace can correctly solve all of these benchmarks~\footnote{We manually inspected the synthesized solutions and confirmed their correctness.} and takes an average of 17.9 seconds per benchmark. In contrast, \textsc{EUSolver}\xspace solves half of the benchmarks within the 24 hour time limit and takes an average of approximately 12 \emph{minutes} per benchmark that it is able to solve. For the five benchmarks that can be solved by both tools, the average speed-up of \textsc{Relish}\xspace over \textsc{EUSolver}\xspace is 46.5x.~\footnote{Here, we use geometric mean to compute the average since the arithmetic mean is not meaningful for ratios.} These statistics clearly demonstrate the advantages of our HFTA-based approach compared to enumerative search: Even though both tools use the same DSLs, verifier, and CEGIS architecture, \textsc{Relish}\xspace unequivocally outperforms \textsc{EUSolver}\xspace across all benchmarks. Next, we compare \textsc{Relish}\xspace and \textsc{EUSolver}\xspace in terms of the number of CEGIS iterations for the five benchmarks that can be solved by both tools. As we can see from Table~\ref{tab:result-codec}, \textsc{Relish}\xspace takes 2.8 CEGIS iterations on average, whereas \textsc{EUSolver}\xspace needs an average of 3.4 attempts to find the correct program. This discrepancy suggests that our HFTA-based method might have better generalization power compared to enumerative search. In particular, our method first generates a version space that contains \emph{all} tuples of programs that satisfy the relational specification and then searches for the best program in this version space. In contrast, \textsc{EUSolver}\xspace returns the \emph{first} program that satisfies the current set of counterexamples; however, this program may not be the best (i.e., lowest-cost) one in \textsc{Relish}\xspace's version space. Finally, we compare \textsc{Relish}\xspace against \textsc{EUSolver}\xspace in terms of synthesis time per CEGIS iteration: \textsc{Relish}\xspace takes an average of 4.5 seconds per iteration, whereas \textsc{EUSolver}\xspace takes 208.4 seconds on average across the five benchmarks that it is able to solve. To summarize, these results clearly indicate the advantages of our approach compared to enumerative search when synthesizing string encoders and decoders. \paragraph{\textbf{Memory usage.}} We now investigate the memory usage of \textsc{Relish}\xspace on the encoder/decoder benchmarks (see column labeled Mem in Table~\ref{tab:result-codec}). Here, the memory usage varies between 284MB and 1843MB, with the average being 685MB. As we can see from Table~\ref{tab:result-codec}, the memory usage mainly depends on (1) the number of CEGIS iterations and (2) the size of programs to be synthesized. Specifically, as the CEGIS loop progresses, the number of function occurrences in the ground relational specification increases, which results in larger HFTAs. For example, the impact of the number of CEGIS iterations on memory usage becomes apparent by comparing the ``Base64xml'' and ``UTF-32'' benchmarks, which take 6 and 2 iterations and consume 1843 MB and 284 MB respectively. In addition to the number of CEGIS iterations, the size of the target program also has an impact on memory usage. Intuitively, the larger the synthesized programs, the larger the size of the individual FTAs; thus, memory usage tends to increase with program size. For example, the impact of program size on memory usage is illustrated by the difference between the ``UU'' and ``Base32'' benchmarks. \subsection{Results for String Comparators} In our second experiment, we evaluate \textsc{Relish}\xspace by using it to synthesize string comparators for sorting problems obtained from StackOverflow. Even though our goal is to synthesize a single {\tt compare} function, this problem is still a relational synthesis task because the generated program must obey two 2-safety properties (i.e., anti-symmetry and totality) and one 3-safety property (i.e., transitivity). Thus, we believe that comparator synthesis is also an interesting and relevant test-bed for evaluating relational synthesizers. \paragraph{\textbf{Benchmarks.}} To perform our evaluation, we collected 20 benchmarks from StackOverflow using the following methodology: First, we searched StackOverflow for the keywords ``\emph{Java string comparator}''. Then, we manually inspected each of these posts and retained exactly those that satisfy the following criteria: \begin{itemize} \item The question in the post should involve writing a \texttt{compare} function for sorting strings. \item The post should contain a list of sample strings that are sorted in the desired way. \item The post should contain a natural language description of the desired sorting task. \end{itemize} The relational specification $\Psi$ for each benchmark consists of the following three parts: \begin{itemize} \item Universally-quantified formulas reflecting the three relational properties that \texttt{compare} has to satisfy (i.e., anti-symmetry, transitivity, and totality). \item Another quantified formula that stipulates reflexivity (i.e., $\forall x. \ ${\tt compare}($x,x$) = 0) \item Quantifier-free formulas that correspond to the input-output examples from the StackOverflow post. In particular, given a sorted list $l$, if string $x$ appears before string $y$ in $l$, we add the examples \texttt{compare(x,y)} = \texttt{-1} and \texttt{compare(y,x)} = \texttt{1}. \end{itemize} Among these benchmarks, the number of examples range from 2 to 30, with an average of 16. \paragraph{\textbf{Main results.}} \input{tab-result-comparator} Our main results are summarized in Table \ref{tab:result-comp}, which is structured in the same way as Table~\ref{tab:result-codec}. The main take-away message from this experiment is that \textsc{Relish}\xspace can successfully solve 95\% of the benchmarks within the 24 hour time limit. Among these 19 benchmarks, \textsc{Relish}\xspace takes an average of 94.7 seconds per benchmark, and it solves 55\% of the benchmarks within 1 minute and 75\% of the benchmarks within 2 minutes. In contrast to \textsc{Relish}\xspace, \textsc{EUSolver}\xspace solves considerably fewer benchmarks within the 24 hour time-limit. In particular, \textsc{Relish}\xspace solves 46\% more benchmarks than \textsc{EUSolver}\xspace (19 vs. 13) and outperforms \textsc{EUSolver}\xspace by 26x (in terms of running time) on the common benchmarks that can be solved by both techniques. Furthermore, similar to the previous experiment, we also observe that \textsc{Relish}\xspace requires fewer CEGIS iterations (7.0 vs. 9.9 on average), again confirming the hypothesis that the HFTA-based approach might have better generalization power. Finally, we note that \textsc{Relish}\xspace is also more efficient than \textsc{EUSolver}\xspace per CEGIS iteration (12 seconds vs. 492 seconds). \paragraph{\textbf{Memory usage.}} We also investigate the memory usage of \textsc{Relish}\xspace on the comparator benchmarks. As shown in the Mem column of Table~\ref{tab:result-comp}, memory usage varies between 12 MB and 18367MB, with an average memory consumption of 3382MB. Comparing these statistics with Table~\ref{tab:result-codec}, we see that memory usage is higher for comparators than the encoder/decoder benchmarks. We believe this difference can be attributed to the following three factors: First, as shown in Table~\ref{tab:result-comp}, the size of the synthesized programs is larger for the comparator domain. Second, the relational specification for comparators is more complex and involves multiple properties such as reflexivity, anti-symmetry, totality, and transitivity. Finally, most benchmarks in the comparator domain require more CEGIS iterations to solve and therefore result in larger HFTAs. \paragraph{\textbf{Analysis of failed benchmarks.}} We manually inspected the benchmark ``comparator-20'' that \textsc{Relish}\xspace failed to synthesize within the 24 hour time limit. In particular, we found this benchmark is not expressible in our current DSL because it requires comparing integers that are obtained by concatenating all substrings that represent integers in the input strings. \paragraph{\textbf{Summary.}} In summary, this experiment demonstrates that \textsc{Relish}\xspace can successfully synthesize non-trivial string comparators that arise in real-world scenarios. This experiment also demonstrates the advantages of our new relational synthesis approach compared to an existing state-of-the-art solver. While the comparator synthesis task involves synthesizing a \emph{single} function, the enumeration-based approach performs considerably worse than \textsc{Relish}\xspace because it does not use the relational (i.e., $k$-safety) specification to prune its search space. \section{Hierarchical Finite Tree Automata} As mentioned in Section~\ref{sec:intro}, our relational synthesis technique uses a version space learning algorithm that is based on the novel concept of \emph{hierarchical finite tree automata (HFTA)}. Thus, we first introduce HFTAs in this section before describing our relational synthesis algorithm. A hierarchical finite tree automaton (HFTA) is a tree in which each node is annotated with a finite tree automaton (FTA). More formally, an HFTA is a tuple $\mathcal{H} = (V, \Omega, v_r, \Lambda)$ where \begin{itemize} \item $V$ is a finite set of nodes. \item $\Omega$ is a mapping that maps each node $v \in V$ to a finite tree automaton. \item $v_r \in V$ is the root node. \item $\Lambda \subseteq \emph{States}(\emph{Range}(\Omega)) \times \emph{States}(\emph{Range}(\Omega))$ is a set of inter-FTA transitions. \end{itemize} Intuitively, an HFTA is a tree-structured (or hierarchical) collection of FTAs, where $\Lambda$ corresponds to the edges of the tree and specifies how to transition from a state of a child FTA to a state of its parent. { For instance, the left-hand side of Fig.~\ref{fig:hfta-ex} shows an HFTA, where the inter-FTA transitions $\Lambda$ (indicated by dashed lines) correspond to the edges $(v_3, v_1)$ and $(v_3, v_2)$ in the tree. } Just as FTAs accept trees, HFTAs accept \emph{hierarchical trees}. Intuitively, as depicted in the right-hand side of Fig.~\ref{fig:hfta-ex}, a hierarchical tree organizes a collection of trees in a tree-structured (i.e., hierarchical) manner. More formally, a hierarchical tree is of the form $\mathcal{T} = (V, \Upsilon, v_r, E)$, where: \begin{itemize} \item $V$ is a finite set of nodes. \item $\Upsilon$ is a mapping that maps each node $v \in V$ to a tree. \item $v_r \in V$ is the root node. \item {$E \subseteq V \times V$ is a set of edges. } \end{itemize} \begin{figure}[!t] \centering \includegraphics[scale=0.42]{pics/hfta-ex2.pdf} \vspace{-7pt} \caption{Example HFTA and hierarchical trees. Left-hand side shows HFTA $\mathcal{H} = (V, \Omega, v_3, \Lambda)$, where nodes $V = \set{v_1, v_2, v_3}$ are represented by rectangles, and annotated FTAs are represented as hypergraphs inside the corresponding rectangles. Specifically, circles in the hypergraph correspond to FTA states, double circles indicate final states, and labeled (hyper-)edges correspond to transitions. Inter-FTA transitions $\Lambda$ are represented as dashed lines. An example hierarchical tree $\mathcal{T}_1 = (V, \Upsilon, v_3, E)$ that is \emph{accepted} by $\mathcal{H}$ is shown in the middle, where nodes $V = \set{v_1, v_2, v_3}$ are represented by rectangles, the tree annotation $\Upsilon(v_i)$ is depicted inside the corresponding rectangle $v_i$, and edges are $E = \set{(v_3, v_1), (v_3, v_2)}$. ``FTA runs'' shows the runs of the FTAs for each $v_i$ in $\mathcal{H}$ on the corresponding tree in $\mathcal{T}_1$. On the right-hand side, we show another example of a hierarchical tree $\mathcal{T}_2$ that is \emph{not accepted} by $\mathcal{H}$. } \label{fig:hfta-ex} \vspace{-5pt} \end{figure} We now define what it means for an HFTA to \emph{accept} a hierarchical tree: \begin{definition}\label{def:hfta-accept} A hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ is accepted by an HFTA $\mathcal{H} = (V, \Omega, v_r, \Lambda)$~iff: \begin{itemize} \item For any node $v \in V$, the tree $\Upsilon(v)$ is accepted by FTA $\Omega(v)$. \item For any edge $(v, v') \in E$, there is an accepting run $\pi$ of FTA $\Omega(v)$ on tree $\Upsilon(v)$ and an accepting run $\pi'$ of $\Omega(v')$ on $\Upsilon(v')$ such that there exists a unique leaf node $l \in \Upsilon(v)$ where we have $\pi' \big(\emph{Root}(\Upsilon(v')) \big) \rightarrow \pi(l) \in \Lambda$. \end{itemize} \end{definition} In other words, a hierarchical tree $\mathcal{T}$ is accepted by an HFTA $\mathcal{H}$ if the tree at each node $v$ of $\mathcal{T}$ is accepted by the corresponding FTA $\Omega(v)$ of $\mathcal{H}$, and the runs of the individual FTAs can be ``stitched'' together according to the inter-FTA transitions of $\mathcal{H}$. The language of an HFTA $\mathcal{H}$, denoted $\mathcal{L}(\mathcal{H})$, is the set of all hierarchical trees accepted by $\mathcal{H}$. \begin{example} Consider the HFTA $\mathcal{H} = (V, \Omega, v_3, \Lambda)$ shown on the left-hand side of Fig.~\ref{fig:hfta-ex}. This HFTA has nodes $V = \set{v_1, v_2, v_3}$ annotated with FTAs as follows: \[ \Omega = \set{v_1 \mapsto \mathcal{A}_1, v_2 \mapsto \mathcal{A}_3, v_3 \mapsto \mathcal{A}_3}\] where: $\mathcal{A}_1 = (\set{q_2, q_3, q_5, q_6}, \set{2, 3, +, \times}, \set{q_5, q_6}, \Delta_1)$ with transitions $\Delta_1$: \[\small \begin{array}{llll} 2 \to q_2 & 3 \to q_3 & +(q_2, q_3) \to q_5 & \times(q_2, q_3) \to q_6 \\ \end{array} \] $\mathcal{A}_2 = (\set{r_2, r_3, r_5, r_6}, \set{2, 3, +, \times}, \set{r_5, r_6}, \Delta_2)$ with transitions $\Delta_2$: \[\small \begin{array}{lllll} 2 \to r_2 & 3 \to r_3 & +(r_3, r_2) \to r_5 & \times(r_3, r_2) \to r_6 \\ \end{array} \] $\mathcal{A}_3 = (\set{u_5, u_6, w_5, w_6, s_\top, s_\bot}, \set{x_1, x_2, =}, \set{s_\top}, \Delta_3)$ with transitions $\Delta_3$: \[\small \begin{array}{llll} x_1 \to u_5 & x_1 \to u_6 & =(u_5, w_5) \to s_\top & =(u_6, w_5) \to s_\bot \\ x_2 \to w_5 & x_2 \to w_6 & =(u_5, w_6) \to s_\bot & =(u_6, w_6) \to s_\top \\ \end{array} \] Finally, $\Lambda$ includes the following four transitions: \[\small \begin{array}{llll} q_5 \to u_5 \ \ \ \ \ \ \ & q_6 \to u_6 \ \ \ \ \ \ \ & r_5 \to w_5 \ \ \ \ \ \ \ & r_6 \to w_6 \end{array} \] Intuitively, this HFTA accepts ground first-order formulas of the form $t_1 = t_2$ where $t_1$ (resp. $t_2$) is a ground term accepted by $\mathcal{A}_1$ (resp. $\mathcal{A}_2$) such that $t_1$ and $t_2$ are equal. For example, $t_1 = 2 + 3$ and $t_2 = 3 + 2$ are collectively accepted. However, $t_1 = 2 + 3$ and $t_2 = 3 \times 2$ are not accepted because $t_1$ is not equal to $t_2$. Fig.~\ref{fig:hfta-ex} shows a hierarchical tree $\mathcal{T}_1$ that is accepted by $\mathcal{H}$ and another one $\mathcal{T}_2$ that is rejected by $\mathcal{H}$. \end{example} \section{Implementation} \label{sec:impl} We have implemented the proposed relational program synthesis approach in a framework called \textsc{Relish}\xspace. To demonstrate the capabilities of the \textsc{Relish}\xspace framework, we instantiate it on the first two domains from Section~\ref{sec:motivating-ex}, namely string encoders/decoders, and comparators. In this section, we describe our implementation of \textsc{Relish}\xspace and its instantiations. \subsection{Implementation of \textsc{Relish}\xspace Framework} While the implementation of the \textsc{Relish}\xspace framework closely follows the algorithm described in Section~\ref{sec:synth}, it performs several important optimizations that we discuss next. \vspace{-0.05in} \paragraph{\textbf{Lazy enumeration.}} Recall that Algorithm~\ref{algo:find} needs to (lazily) enumerate all hierarchical trees accepted by a given HFTA. Furthermore, since we want to find a program with the greatest generalization power, our algorithm should enumerate more promising programs first. Based on these criteria, we need a mechanism for predicting the generalization power of a hierarchical tree using some heuristic cost metric. In our implementation, we associate a non-negative cost with every DSL construct and compute the cost of a given hierarchical tree by summing up the costs of all its nodes. Because hierarchical trees with lower cost are likely to have better generalization power, our algorithm lazily enumerates hierarchical trees according to their cost. In our implementation, we reduce the problem of enumerating hierarchical trees accepted by an HFTA to the task of enumerating B-paths in a hypergraph~\cite{min-bpath}. In particular, we first flatten the HFTA into a standard FTA by combining the individual tree automata at each node via the inter-FTA transitions. We then represent the resulting flattened FTA as a hypergraph where the FTA states correspond to nodes and a transition $\sigma(q_1, \ldots, q_n) \to q$ corresponds to a B-edge $(\set{q_1, \ldots, q_n}, q)$ with weight $\emph{cost}(\sigma)$. Given this representation, the problem of finding the lowest-cost hierarchical tree accepted by an HFTA becomes equivalent to the task of finding a minimum weighted B-path in a hypergraph, and our implementation leverages known algorithms for solving this problem~\cite{min-bpath}. \vspace{-0.05in} \paragraph{\textbf{Verification.}} Because our overall approach is based on the CEGIS paradigm, we need a separate verification step to both check the correctness of the programs returned by the inductive synthesizer and find counterexamples if necessary. However, because heavy-weight verification can add considerable overhead to the CEGIS loop, we \emph{test} the program against a large set of inputs rather than performing full-fledged verification in each iteration. Specifically, we generate a set of validation inputs by computing all possible permutations of a finite set up to a bounded length $k$ and check the correctness of the candidate program against this validation set. In our experiments, we use over a million test cases in each iteration, and resort to full verification only when the synthesized program passes all of these test cases. \subsection{Instantiation for String Encoders and Decoders}\label{sec:encoder} \input{fig-cfg-codec} While \textsc{Relish}\xspace is a generic framework that can be used in various application domains, one needs to construct suitable DSLs and write relational specifications for each different domain. In this section, we discuss our instantiation of \textsc{Relish}\xspace for synthesizing string encoders and decoders. Since we have already explained the relational property of interest in Section~\ref{sec:motivating-ex}, we discuss two simple DSLs, presented in Fig.~\ref{fig:cfg-codec}, for implementing encoders and decoders respectively. \vspace{-0.05in} \paragraph{\textbf{DSL for encoders}} We designed a DSL for string encoders by reviewing several different encoding mechanisms and identifying their common building blocks. Specifically, this DSL allows transforming a Unicode string (or binary data) to a sequence of restricted ASCII characters (e.g., to fulfill various requirements of text-based network transmission protocols). At a high-level, programs in this DSL first transform the input string to an integer array and then to a byte array. The encoded text is obtained by applying various kinds of mappers to the byte array, padding it, and attaching length information. In what follows, we informally describe the semantics of the constructs used in the encoder DSL. \begin{itemize}[leftmargin=*] \item {\bf \emph{Code point representation:}} The \emph{codePoint} function converts string $s$ to an integer array $I$ where $I[j]$ corresponds to the Unicode code point for $s[j]$. \item {\bf \emph{Mappers:}} The \emph{encUTF8/16/32} functions transform the code point array to a byte array according to the corresponding standards of UTF-8, UTF-16, UTF-32, respectively. The other mappers \emph{encX} can further transform the byte array into a sequence of restricted ASCII characters based on different criteria. For example, the \emph{enc16} function is a simple hexadecimal mapper that can convert binary data \texttt{0x6E} to the string ``6E''. \item {\bf \emph{Padding:}} The \emph{padToMultiple} function takes an existing character array $E$ and pads it with a sequence of extra \emph{char} characters to ensure that the length of the padded sequence is evenly divisible by \emph{num}. \item {\bf \emph{Header:}} The \emph{header(E)} function prepends the ASCII representation of the length of the text to $E$. \item {\bf \emph{Reshaping:}} The function \emph{reshape(B, num)} first concatenates all bytes in $B$, then regroups them such that each group only contains \emph{num} bits (instead of 8), and finally generates a new byte array where each byte is equal to the value of corresponding group. For example, \emph{reshape}\texttt{([0xFF],4) = [0x0F,0x0F]} and \emph{reshape}\texttt{([0xFE],2) = [0x03,0x03,0x03,0x02]}. \end{itemize} \vspace{-0.05in} \paragraph{\textbf{DSL for decoders}} The decoder DSL is quite similar to the encoder one and is structured as follows: Given an input string $x$, programs in this DSL first process $x$ by removing the header and/or padding characters and then transform it to a byte array using a set of pre-defined mappers. The decoded data can be either the resulting byte array or a Unicode string obtained by converting the byte array to an integer Unicode code point array. In more detail, the decoder DSL supports the following built-in operators: \begin{itemize}[leftmargin=*] \item \emph{\textbf{Unicode conversion:}} The \emph{asUnicode} function converts an integer array $I$ to a string $s$, where $s[j]$ corresponds to the Unicode symbol of code point $I[j]$. \item \emph{\textbf{Mappers:}} The \emph{decUTF8/16/32} functions transform a byte array to a code point array based on standards of UTF-8, UTF-16, UTF-32, respectively. The other mappers \emph{decX} transform a sequence of ASCII characters to byte arrays according to their standard transformation rules. \item \emph{\textbf{Character removal:}} The \emph{removePad} function removes all trailing characters \emph{char} from a given string $C$. The \emph{substr} function takes a string $C$ and an index \emph{num}, and returns the sub-string of $C$ from index \emph{num} to the end. \item \emph{\textbf{Reshaping:}} The \emph{invReshape} function takes a byte array $B$, collects \emph{num} bits from the least significant end of each byte, concatenates the bit sequence and regroups every eight bits to generate a new byte array. For instance, \emph{invReshape}\texttt{([0x0E,0x0F],4) = [0xEF]}. \end{itemize} \begin{example} The desired {\tt encode/decode} functions from Example~\ref{ex:codec} can be implemented in our DSL as follows: \[ \small \begin{BVerbatim} encode(x) : padToMultiple (enc64 (reshape (encUTF8 (codePoint (x)), 6)), 4, '=') decode(x) : asUnicode (decUTF8 (invReshape (dec64 (removePad (x, '=')), 6))) \end{BVerbatim} \] \end{example} \subsection{Instantiation for Comparators} \begin{figure}[!t] \centering \small \[ \begin{array}{llcl} \emph{Comparator} & C & := & B \ | \ \emph{chain}(B, C) \\ \emph{Basic} & B & := & \emph{intCompare}(I_1, I_2) \ | \ \emph{strCompare}(S_1, S_2) \\ \emph{Integer} & I & := & \emph{countChar}(S, c) \ | \ \emph{length}(S) \ | \ \emph{toInt}(S) \\ \emph{String} & S & := & \emph{substr}(v, P_1, P_2) \\ \emph{Position} & P & := & \emph{pos}(v, t, k, d) \ | \ \emph{constPos}(k) \\ \end{array} \] \[ v \in \{ \emph{ x, y } \} \qquad c \in \emph{Characters} \qquad t \in \emph{Tokens} \qquad k \in \emph{Integers} \qquad d \in \{ \emph{ Start, End } \} \] \vspace{-0.1in} \caption{Context-free grammar for comparators.} \label{fig:cfg-comparators} \vspace{-0.2in} \end{figure} We have also instantiated the \textsc{Relish}\xspace framework to enable automatic generation of custom string comparators. As described in Section~\ref{sec:motivating-ex}, this domain is an interesting ground for relational program synthesis because comparators must satisfy three different relational properties (i.e., anti-symmetry, transitivity, and totality). In what follows, we describe the DSL from Fig.~\ref{fig:cfg-comparators} that \textsc{Relish}\xspace uses to synthesize these comparators. In more detail, programs in our comparator DSL take as input a pair of strings \emph{x, y}, and return -1, 0, or 1 indicating that \emph{x} precedes, is equal to, or succeeds \emph{y} respectively. Specifically, a program is either a \emph{basic comparator} $B$ or a \emph{comparator chain} of the form $\emph{chain}(B_1, \ldots, B_n)$ which returns the result of the \emph{first} comparator that does not evaluate to zero. The DSL allows two basic comparators, namely \emph{intCompare} and \emph{strCompare}. The integer or string inputs to these basic comparators can be obtained using the following functions: \begin{itemize}[leftmargin=*] \item \emph{\textbf{Substring extraction:}} The \emph{substring} function is used to extract substrings of the input string. In particular, for a string $v$ and positions $P_1, P_2$, it returns the substring of $v$ that starts at index $P_1$ and ends at index $P_2$. \item \emph{\textbf{Position identifiers:}} A position $P$ can either be a constant index ($\emph{constPos}(k)$) or the (start or end) index of the $k$'th occurrence of the match of token $t$ in the input string ($\emph{pos}(v, t, k, d)$).~\footnote{Tokens are chosen from a predefined universe of regular expressions.} For example, we have $\emph{pos}(\text{``} 12ab \text{''}, \texttt{Number}, 1, \texttt{Start}) = 0 $ and $\emph{pos}(\text{``} 12ab \text{''}, \texttt{Number}, 1, \texttt{End}) = 2$ where \texttt{Number} is a token indicating a sequence of digits. \item \emph{\textbf{Numeric string features:}} The DSL allows extracting various numeric features of a given string $S$. In particular, \emph{countChar} yields the number of occurrences of a given character $c$ in string $S$, \emph{length} yields string length, and \emph{toInt} converts a string representing an integer to an actual integer (i.e., $\emph{toInt}(``123") = 123$ but \emph{toInt}(``abc") throws an exception). \end{itemize} \begin{example} Consider Example~\ref{ex:comparator} where the user wants to sort integers based on the number of occurrences of the number \texttt{5}, and, in the case of a tie, sort them based on the actual integer values. This functionality can be implemented by the following simple program in our DSL: \[\small \begin{BVerbatim} chain (intCompare (countChar (x, '5'), countChar (y, '5')), intCompare (toInt (x), toInt (y)) ) \end{BVerbatim} \] \end{example} \section{Introduction}\label{sec:intro} \emph{Relational properties} describe requirements on the interaction between multiple programs or different runs of the same program. Examples of relational properties include the following: \begin{itemize} \item \emph{Equivalence:} Are two programs $P_1, P_2$ observationally equivalent? (i.e., $\forall \vec{x}.\ P_1(\vec{x}) = P_2(\vec{x})$) \item \emph{Inversion:} Given two programs $P_1, P_2$, are they inverses of each other? (i.e., $\forall x. \ P_2(P_1(x)) = x)$ \item \emph{Non-interference:} Given a program $P$ with two types of inputs, namely \emph{low} (public) input $\vec{l}$ and \emph{high} (secret) input $\vec{h}$, does $P$ produce the same output when run on the same low input $\vec{l}$ but two different high inputs $\vec{h_1}, \vec{h_2}$? (i.e., $\forall \vec{l}, \vec{h_1}, \vec{h_2}. \ P(\vec{l}, \vec{h_1}) = P(\vec{l}, \vec{h_2}))$ \item \emph{Transitivity:} Given a program $P$ that returns a boolean value, does $P$ obey transitivity? (i.e., $\forall x,y,z. \ P(x,y) \land P(y,z) \Rightarrow P(x, z)$) \end{itemize} Observe that the first two properties listed above relate different programs, while the latter two relate multiple runs of the same program. Due to their importance in a wide range of application domains, relational properties have received significant attention from the program verification community. For example, prior papers propose novel \emph{program logics} for verifying relational properties~\cite{rhl,prhl,chl,qchl,relsep} or transform the relational verification problem to standard safety by constructing so-called \emph{product programs}~\cite{product1,product-asymmetric,product3}. In this paper, we consider the dual \emph{synthesis} problem of relational verification. That is, given a relational specification $\Psi$ relating runs of $n$ programs $P_1, \ldots, P_n$, our goal is to automatically {synthesize} $n$ programs that satisfy $\Psi$. This \emph{relational synthesis} problem has a broad range of practical applications. For example, we can use relational synthesis to automatically generate comparators that provably satisfy certain correctness requirements such as transitivity and anti-symmetry. As another example, we can use relational synthesis to solve the \emph{program inversion} problem where the goal is to generate a program $P'$ that is an inverse of another program $P$~\cite{inversion-loris,inversion-swarat}. Furthermore, since many automated repair techniques rely on program synthesis~\cite{repair1,repair2}, we believe that relational synthesis could also be useful for repairing programs that violate a relational property like non-interference. Solving the relational synthesis problem introduces new challenges over its non-relational counterpart due to the combinatorially larger search space. In particular, a naive algorithm that simply enumerates combinations of programs and then checks their correctness with respect to the relational specification is unlikely to scale. Instead, we need to design novel \emph{relational} synthesis algorithms to {efficiently} search for \emph{tuples of programs} that collectively satisfy the given specification. We solve this challenge by introducing a novel synthesis algorithm that learns \emph{relational version spaces (RVS)}, a generalization of the notion of \emph{version space} utilized in prior work~\cite{vsa,flashfill,dace}. Similar to other synthesis algorithms based on counterexample-guided inductive synthesis (CEGIS)~\cite{cegis}, our method also alternates between inductive synthesis and verification; but the counterexamples returned by the verifier are relational in nature. Specifically, given a set of \emph{relational counterexamples}, such as $f(1) = g(1)$ or $f(g(1),g(2))=f(3)$, our inductive synthesizer compactly represents \emph{tuples} of programs and efficiently searches for their implementations that satisfy all relational counterexamples. In more detail, our relational version space learning algorithm is based on the novel concept of \emph{hierarchical finite tree automata (HFTA)}. Specifically, an HFTA is a hierarchical collection of finite tree automata (FTAs), where each individual FTA represents possible implementations of the different functions to be synthesized. Because relational counterexamples can refer to compositions of functions (e.g., $f({g(1), g(2)})$), the HFTA representation allows us to compose different FTAs according to the hierarchical structure of subterms in the relational counterexamples. Furthermore, our method constructs the HFTA in such a way that tuples of programs that do not satisfy the examples are rejected and therefore excluded from the search space. Thus, the HFTA representation allows us to compactly represent those programs that are consistent with the relational examples. We have implemented the proposed relational synthesis algorithm in a tool called \textsc{Relish}\xspace~\footnote{\textsc{Relish}\xspace stands for RELatIonal SyntHesis.} and evaluate it in the context of two different relational properties. First, we use \textsc{Relish}\xspace to automatically synthesize encoder-decoder pairs that are provably inverses of each other. Second, we use \textsc{Relish}\xspace to automatically generate comparators, which must satisfy three different relational properties, namely, anti-symmetry, transitivity, and totality. Our evaluation shows that \textsc{Relish}\xspace can efficiently solve interesting benchmarks taken from previous literature and online forums. Our evaluation also shows that \textsc{Relish}\xspace significantly outperforms \textsc{EUSolver}\xspace, a general-purpose synthesis tool that won the General Track of the most recent SyGuS competition for syntax-guided synthesis. To summarize, this paper makes the following key contributions: \begin{itemize} \item We introduce the \emph{relational synthesis problem} and take a first step towards solving it. \item We describe a \emph{relational version space} learning algorithm based on the concept of \emph{hierarchical finite tree automata} (HFTA). \item We show how to construct HFTAs from relational examples expressed as ground formulas and describe an algorithm for {finding the desired accepting runs}. \item We experimentally evaluate our approach in two different application domains and demonstrate the advantages of our relational version space learning approach over a state-of-the-art synthesizer based on enumerative search. \end{itemize} \section{Limitations} While we have successfully used the proposed relational synthesis method to synthesize encoder-decoder pairs and comparators, our current approach has some limitations that we plan to address in future work. First, our method only works with simple DSLs without recursion, loops, or let bindings. That is, the programs that can be currently synthesized by \textsc{Relish}\xspace are compositions of built-in functions provided by the DSL. Second, we only allow relational specifications of the form $\forall \vec{x}. \phi(\vec{x})$ where $\phi$ is quantifier-free. Thus, our method does not handle more complex relational specifications with quantifier alternation (e.g., $\forall x. \exists y. f(x,y) = g(y, x)$). \section{Overview} \label{sec:overview} \begin{figure} \small \begin{center} \bgroup \def\arraystretch{1.1 \begin{tabular}{|c|c|} \hline {\bf Property} & {\bf Relational specification} \\ \hline Equivalence & $\forall \vec{x}. \ f_1(\vec{x}) = f_2(\vec{x})$ \\ \hline Commutativity & $\forall x. \ f_1(f_2(x)) = f_2(f_1(x))$ \\ \hline Distributivity & $\forall x, y, z. \ f_2(f_1(x,y), z) = f_1(f_2(x,z), f_2(y,z))$ \\ \hline Associativity & $\forall x,y,z. \ f(f(x,y), z) = f(x, f(y,z))$ \\ \hline Anti-symmetry & $\forall x,y. \ (f(x,y)=\emph{true} \land x \neq y) \Rightarrow f(y, x)=\emph{false}$ \\ \hline \end{tabular} \egroup \end{center} \vspace{-8pt} \caption{Examples of relational specifications for five relational properties.}\label{fig:ex-specs} \vspace{-0.15in} \end{figure} In this section, we define the \emph{relational synthesis} problem and give a few motivating examples. \subsection{Problem Statement} The input to our synthesis algorithm is a \emph{relational specification} defined as follows: \begin{definition}[\textbf{Relational specification}]\label{def:rel-spec} A relational specification with functions $f_1, \ldots, f_n$ is a first-order sentence $\Psi = \forall \vec{x}. ~ \phi(\vec{x})$, where $\phi(\vec{x})$ is a quantifier-free formula with uninterpreted functions $f_1, \ldots, f_n$. \end{definition} Fig.~\ref{fig:ex-specs} shows some familiar relational properties like distributivity and associativity as well as their corresponding specifications. Even though we refer to Definition~\ref{def:rel-spec} as a \emph{relational specification}, observe that it allows us to express combinations of both relational and non-relational properties. For instance, the specification $\forall x. \ (f(x) \geq 0 \land f(x) + g(x) = 0)$ imposes both a non-relational property on $f$ (namely, that all of its outputs must be non-negative) as well as the relational property that the outputs of $f$ and $g$ must always add up to zero. \begin{definition}[\textbf{Relational program synthesis}]\label{def:rel-synth} Given a set of $n$ function symbols $\mathcal{F} = \set{f_1, \ldots, f_n}$, their corresponding domain-specific languages $\set{L_1, \ldots, L_n}$, and a relational specification $\Psi$, the relational program synthesis problem is to find an interpretation $\mathcal{I}$ for $\mathcal{F}$ such that: \begin{enumerate} \item For every function symbol $f_i \in \mathcal{F}$, $\mathcal{I}(f_i)$ is a program in $f_i$'s DSL $L_i$ (i.e., $\mathcal{I}(f_i) \in L_i$). \item The interpretation $\mathcal{I}$ satisfies the relational specification $\Psi$ (i.e., $\mathcal{I}(f_1), \ldots, \mathcal{I}(f_n) \models \Psi$). \end{enumerate} \end{definition} \subsection{Motivating Examples}\label{sec:motivating-ex} We now illustrate the practical relevance of the relational program synthesis problem through several real-world programming scenarios. \begin{example}[\textbf{String encoders and decoders}] \label{ex:codec} Consider a programmer who needs to implement a Base64 encoder \texttt{encode(x)} and its corresponding decoder \texttt{decoder(x)} for any Unicode string \texttt{x}. For example, according to Wikipedia~\footnote{\url{https://en.wikipedia.org/wiki/Base64}}, the encoder should transform the string ``\texttt{Man}'' into ``\texttt{TWFu}'', ``\texttt{Ma}'' into ``\texttt{TWE=}'', and ``\texttt{M}'' to ``\texttt{TQ==}''. In addition, applying the decoder to the encoded string should yield the original string. Implementing this encoder/decoder pair is a relational synthesis problem in which the relational specification is the following: \[ \small \begin{array}{lll} & \texttt{encode(} \text{``} \texttt{Man} \text{''} \texttt{)} = \text{``} \texttt{TWFu} \text{''} \land \texttt{encode(} \text{``} \texttt{Ma} \text{''} \texttt{)} = \text{``} \texttt{TWE=} \text{''} \land \texttt{encode(} \text{``} \texttt{M} \text{''} \texttt{)} = \text{``} \texttt{TQ==} \text{''} & (\emph{input-output examples}) \\ \land & \forall x.~ \texttt{decode(encode($x$))} = x & (\emph{inversion}) \\ \end{array} \] Here, the first part (i.e., the first line) of the specification gives three input-output examples for \texttt{encode}, and the second part states that \texttt{decode} must be the inverse of \texttt{encode}. \textsc{Relish}\xspace can automatically synthesize the correct Base64 encoder and decoder from this specification using a DSL targeted for this domain (see Section~\ref{sec:encoder}). \end{example} \begin{example}[\textbf{Comparators}] \label{ex:comparator} Consider a programmer who needs to implement a comparator for sorting an array of integers according to the number of occurrences of the number \texttt{5}~\footnote{\url{https://stackoverflow.com/questions/19231727/sort-array-based-on-number-of-character-occurrences}}. Specifically, \texttt{compare(x,y)} should return \texttt{-1} (resp. \texttt{1}) if \texttt{x} (resp. \texttt{y}) contains less \texttt{5}'s than \texttt{y} (resp. \texttt{x}), and ties should be broken based on the actual values of the integers. For instance, sorting the array [``\texttt{24}'',``\texttt{15}'',``\texttt{55}'',``\texttt{101}'',``\texttt{555}''] using this comparator should yield array [``\texttt{24}'',``\texttt{101}'',``\texttt{15}'',``\texttt{55}'',``\texttt{555}'']. Furthermore, since the comparator must define a total order, its implementation should satisfy reflexivity, anti-symmetry, transitivity, and totality. The problem of generating a suitable \texttt{compare} method is again a relational synthesis problem and can be defined using the following specification: \[ \small \begin{array}{lll} & \texttt{compare(} \text{``} \texttt{24} \text{''}, \text{``} \texttt{15} \text{''} \texttt{)} = \texttt{-1} \land \texttt{compare(} \text{``} \texttt{101} \text{''}, \text{``} \texttt{24} \text{''} \texttt{)} = \texttt{1} \land \ldots & (\emph{input-output examples}) \\ \land & \forall x. ~ \texttt{compare($x,x$)} = \texttt{0} & (\emph{reflexivity}) \\ \land & \forall x, y. ~ \emph{sgn}(\texttt{compare($x,y$)}) = -\emph{sgn}(\texttt{compare($y, x$)}) & (\emph{anti-symmetry}) \\ \land & \forall x, y, z.~ \texttt{compare($x, y$)} > \texttt{0} \land \texttt{compare($y, z$)} > \texttt{0} \Rightarrow \texttt{compare($x, z$)} > \texttt{0} & (\emph{transitivity}) \\ \land & \forall x, y, z.~ \texttt{compare($x, y$)} = \texttt{0} \Rightarrow \emph{sgn}(\texttt{compare($x, z$)}) = \emph{sgn}(\texttt{compare($y, z$)}) & (\emph{totality}) \\ \end{array} \] A solution to this problem is given by the following implementation of \texttt{compare}: \[ \small \begin{BVerbatim} let a = intCompare (countChar x '5') (countChar y '5') in if a != 0 then a else intCompare (toInt x) (toInt y) \end{BVerbatim} \] where {\tt countChar} function returns the number of occurrences of a character in the input string, and {\tt intCompare} is the standard comparator on integers. \end{example} \begin{example}[\textbf{Equals and hashcode}] A common programming task is to implement {\tt equals} and {\tt hashcode} methods for a given class. These functions are closely related because {\tt equals(x,y)=true} implies that the hash codes of {\tt x} and {\tt y} must be the same. Relational synthesis can be used to simultaneously generate implementations of {\tt equals} and {\tt hashcode}. For example, consider an {\tt ExperimentResults} class that internally maintains an array of numbers where negative integers indicate an anomaly (i.e., failed experiment) and should be ignored when comparing the results of two experiments. For instance, the results \texttt{[23.5,-1,34.7]} and \texttt{[23.5,34.7]} should be equal whereas \texttt{[23.5,34.7]} and \texttt{[34.7,23.5]} should not. The programmer can use a relational synthesizer to generate {\tt equals} and {\tt hashcode} implementations by providing the following specification: \[ \small \begin{array}{lll} & \texttt{equals([23.5,-1,34.7]}, \texttt{[23.5,34.7])} = \texttt{true} & (\emph{input-output examples}) \\ \land & \texttt{equals([23.5,34.7]}, \texttt{[34.7,23.5])} = \texttt{false} \land \ldots \\ \land & \forall x, y. \ \texttt{equals($x, y$)} \Rightarrow (\texttt{hashcode($x$)} = \texttt{hashcode($y$)}) & (\emph{equals-hashcode}) \\ \end{array} \] A possible solution consists of the following pair of implementations of {\tt equals} and {\tt hashcode}: \[\small \begin{BVerbatim} equals(x,y) : (filter (>= 0) x) == (filter (>= 0) y) hashcode(x) : foldl (\u.\v. 31 * (u + v)) 0 (filter (>= 0) x) \end{BVerbatim} \] \end{example} \begin{example}[\textbf{Reducers}] MapReduce is a software framework for writing applications that process large datasets in parallel. Users of this framework need to implement both a \emph{mapper} that maps input key/value pairs to intermediate ones as well as a \emph{reducer} which transforms intermediate values that share a key to a smaller set of values. An important requirement in the MapReduce framework is that the {\tt reduce} function must be associative. Now, consider the task of using the MapReduce framework to sum up sensor measurements obtained from an experiment. In particular, each sensor value is either a real number or \texttt{None}, and the final result should be \texttt{None} if \emph{any} sensor value is \texttt{None}. In order to implement this functionality, the user needs to write a reducer that takes sensor values \texttt{x}, \texttt{y} and returns their sum if neither of them is \texttt{None}, or returns \texttt{None} otherwise. We can express this problem using the following relational specification: \[ \small \begin{array}{lll} & \texttt{reduce(1} , \texttt{2)} = \texttt{3} \land \texttt{reduce(1} , \texttt{None)} = \texttt{None} & (\emph{input-output examples}) \\ \land & \forall x, y, z. ~ \texttt{reduce(reduce($x, y$)$, z$)} = \texttt{reduce(}x, \texttt{reduce(} y, z \texttt{))} & (\emph{associativity}) \\ \end{array} \] The first part of the specification gives input-output examples to illustrate the desired functionality, and the latter part expresses the associativity requirement on {\tt reduce}. A possible solution to this relational synthesis problem is given by the following implementation: \[\small \begin{BVerbatim} reduce(x,y) : if x == None then None else if y == None then None else x + y \end{BVerbatim} \] \end{example} \section{Preliminaries} Since the rest of this paper requires knowledge of finite tree automata and their use in program synthesis, we first briefly review some background material. \subsection{Finite Tree Automata} A tree automaton is a type of state machine that recognizes trees rather than strings. More formally, a (bottom-up) \emph{finite tree automaton} (FTA) is a tuple $\mathcal{A} = (Q, \Sigma, Q_f, \Delta)$, where \begin{itemize} \item $Q$ is a finite set of states. \item $\Sigma$ is an alphabet. \item $Q_f \subseteq Q$ is a set of final states. \item $\Delta \subseteq Q^* \times \Sigma \times Q$ is a set of transitions (or rewrite rules). \end{itemize} Intuitively, a tree automaton $\mathcal{A}$ recognizes a term (i.e., tree) $t$ if we can rewrite $t$ into a final state $q \in Q_f$ using the rewrite rules given by $\Delta$. More formally, suppose we have a tree $t = (V, E, v_r)$ where $V$ is a set of nodes labeled with $\sigma \in \Sigma$, $E \subseteq V \times V$ is a set of edges, and $v_r \in V$ is the root node. A \emph{run} of an FTA $\mathcal{A} = (Q, \Sigma, Q_f, \Delta)$ on $t$ is a mapping $\pi: V \to Q$ compatible with $\Delta$ (i.e., given node $v$ with label $\sigma$ and children $v_1, \ldots, v_n$ in the tree, the run $\pi$ can only map $v, v_1, \ldots, v_n$ to $q, q_1, \ldots, q_n$ if there is a transition $\sigma(q_1, \ldots, q_n) \to q$ in $\Delta$). Run $\pi$ is said to be \emph{accepting} if it maps the root node of $t$ to a final state, and a tree $t$ is \emph{accepted by} an FTA $\mathcal{A}$ if there exists an accepting run of $\mathcal{A}$ on $t$. The \emph{language} $\mathcal{L}(\mathcal{A})$ recognized by $\mathcal{A}$ is the set of all those trees that are accepted by $\mathcal{A}$. \begin{wrapfigure}{R}{2.3in} \centering \includegraphics[scale=0.5]{pics/fta-ex.pdf} \caption{Tree for $1 \land \neg 0$ and its accepting run.} \label{fig:fta-ex} \end{wrapfigure} \begin{example} Consider a finite tree automaton $\mathcal{A} = (Q, \Sigma, Q_f, \Delta)$ with states $Q = \set{q_0, q_1}$, alphabet $\Sigma = \set{0, 1, \neg, \land}$, final states $Q_f = \set{q_1}$, and transitions $\Delta$: \[\small \begin{array}{l l l l} 1 \to q_1 & 0 \to q_0 & \land(q_0, q_0) \to q_0 & \land(q_0, q_1) \to q_0 \\ \neg(q_0) \to q_1 & \neg(q_1) \to q_0 & \land(q_1, q_0) \to q_0 & \land(q_1, q_1) \to q_1 \\ \end{array} \] Intuitively, the states of this FTA correspond to boolean constants (i.e., $q_0, q_1$ represent false and true respectively), and the transitions define the semantics of the boolean connectives $\neg$ and $\land$. Since the final state is $q_1$, the FTA accepts all boolean formulas (containing only $\land$ and $\neg$) that evaluate to $\emph{true}$. For example, $1 \land \neg 0$ is accepted by $\mathcal{A}$, and the corresponding tree and its accepting run are shown in Fig.~\ref{fig:fta-ex}. \end{example} \subsection{Example-based Synthesis using FTAs}\label{sec:fta-construct} Since our relational synthesis algorithm leverages prior work on example-based synthesis using FTAs~\cite{dace,syngar}, we briefly review how FTAs can be used for program synthesis. At a high-level, the idea is to build an FTA that accepts \emph{exactly} the ASTs of those DSL programs that are consistent with the given input-output examples. The states of this FTA correspond to concrete values, and the transitions are constructed using the DSL's semantics. In particular, the FTA states correspond to output values of DSL programs on the input examples, and the final states of the FTA are determined by the given output examples. Once this FTA is constructed, the synthesis task boils down to finding an accepting run of the FTA. A key advantage of using FTAs for synthesis is to enable search space reduction by allowing sharing between programs that have the same input-output behavior. \input{fig-fta-rules} In more detail, suppose we are given a set of input-output examples $\vec{\ex}$ and a context-free grammar $G$ describing the target domain-specific language. We assume that $G$ is of the form $(T, N, P, s_0)$ where $T$ and $N$ are the terminal and non-terminal symbols respectively, $P$ is a set of productions of the form $s \rightarrow \sigma(s_1, \dots, s_n)$, and $s_0 \in N$ is the topmost non-terminal (start symbol) in $G$. We also assume that the program to be synthesized takes arguments $x_1, \ldots, x_n$. Fig. \ref{fig:fta-rules} reviews the construction rules for a single example $e = (e_{\emph{in}}, e_{\emph{out}})$~\cite{syngar}. In particular, the Input rule creates $n$ initial states $q_{x_1}^{e_1}, \ldots, q_{x_n}^{e_n}$ for input example $e_{\emph{in}} = (e_1, \ldots, e_n)$. We then iteratively use the Prod rule to generate new states and transitions using the productions in grammar $G$ and the concrete semantics of the DSL constructs associated with the production. Finally, according to the Output rule, the only final state is $q_{s_0}^c$ where $s_0$ is the start symbol and $c$ is the output example $e_{\emph{out}}$. Observe that we can build an FTA for a \emph{set} of input-output examples by constructing the FTA for each example individually and then taking their intersection using standard techniques~\cite{tata}. \begin{wrapfigure}{R}{0.4\textwidth} \vspace{-12pt} \centering \includegraphics[scale=0.7]{pics/fta-plus.pdf} \caption{FTA for input-output $(1, 3) \to 4$.} \label{fig:fta-plus} \vspace{-10pt} \end{wrapfigure} \vspace{2pt} \noindent {\bf \emph{Remark.}} Since the rules in Fig.~\ref{fig:fta-rules} may generate an infinite set of states and transitions, the FTA is constructed by applying the Prod rule a finite number of times. The number of applications of the Prod rule is determined by a bound on the AST depth of the synthesized program. \begin{example} Consider a very simple DSL specified by the following CFG, where the notation $s \rightarrow s'$ is short-hand for $s \rightarrow \emph{id}(s')$ for a unary identity function $\emph{id}$: \[ \hspace*{-160pt} e \ \ \rightarrow \ \ x_1 ~|~ x_2 ~|~ e + e \] Suppose we want to find a program (of size at most 3) that is consistent with the input-output example $(1, 3) \rightarrow 4$. Using the rules from Fig.~\ref{fig:fta-rules}, we can construct $\mathcal{A} = (\ftaStates, \ftaAlphabet, {\ftaFinal}, \ftaTrans)$ with states $\ftaStates = \set{q^1_{x_1}, q^3_{x_2}, q^1_e, q^2_e, q^3_e, q^4_e, q^6_e}$, alphabet $\ftaAlphabet = \set{x_1, x_2, \emph{id}, +}$, final states ${\ftaFinal} = \set{q^4_e}$, and transitions~$\ftaTrans$: \[\small \begin{array}{llll} x_1 \to q^1_{x_1} & x_2 \to q^3_{x_2} & \emph{id}(q^1_{x_1}) \rightarrow q^1_e & \emph{id}(q^3_{x_2}) \rightarrow q^3_e\\ +(q^1_e, q^1_e) \to q^2_e & +(q^1_e, q^3_e) \to q^4_e & +(q^3_e, q^1_e) \to q^4_e & +(q^3_e, q^3_e) \to q^6_e \\ \end{array} \] For example, the FTA contains the transition $+(q_e^3, q_e^1) \rightarrow q_e^4$ because the grammar contains a production $e \rightarrow e+e$ and we have $3+1 = 4$. Since it is sometimes convenient to visualize FTAs as hypergraphs, Fig.~\ref{fig:fta-plus} shows a hypergraph representation of this FTA, using circles to indicate states, double circles to indicate final states, and labeled (hyper-)edges to represent transitions. The transitions of nullary alphabet symbols (i.e. $x_1 \to q^1_{x_1}, x_2 \to q^3_{x_2}$) are omitted in the hypergraph for brevity. Note that the only two programs that are accepted by this FTA are $x_1 + x_2$ and $x_2 + x_1$. \end{example} \section{Proof of Theorems} \begin{lemma}\label{lem:sound-fta} Suppose FTA $\mathcal{A} = (Q, \Sigma, Q_f, \Delta)$ is built with grammar $G$ and initial states $[Q_1, \ldots, Q_m]$ using rules from Figure~\ref{fig:buildFTA}. Given a tree $t(x_1, \ldots, x_m)$ that is accepted by $\mathcal{A}$, we have \begin{enumerate} \item $t$ conforms to grammar $G$. \item The accepting run of $t$ with $x_i = c_i$, where $q^{c_i}_{x_i} \in Q_i$, is rooted with final state $q^c_s$ and $c = \denot{t(c_1, \ldots, c_m)}$. \end{enumerate} \end{lemma} \begin{proof} To prove by induction on height of $t$, we first strengthen (2) to: the run of $t$ with $x_i = c_i$ is rooted with state $q^c_s$ and $c = \denot{t(c_1, \ldots, c_m)}$. \begin{itemize} \item Base case: height is 1. In this case, $t$ only has one variable $x$, and $q^c_{s}$ is passed as an input. Consider the \textsf{Input} rule in Figure~\ref{fig:buildFTA}, we know $q^c_{x} \in Q$ and $x \to q^c_x \in \Delta$. Given $t$ is accepted by $\mathcal{A}$, $q^c_x$ is a final state, so $x$ conforms to grammar $G$. The run has only one node $q^c_x$, and $c = \denot{t(c)}$. \item Inductive case: Suppose the lemma is correct for term with height no larger than $h$, and the goal is to prove it is correct for term $t$ of height $h+1$. Without loss of generality, assume $t = \sigma(t_1, \ldots, t_n)$ is of height $h+1$, so the height of terms $t_1, \ldots, t_n$ is at most $h$. By inductive hypothesis, $t_i$ conforms to grammar $G$ for $i \in [1, n]$. Furthermore, the run of $t_i$ is rooted with state $q^{a_i}_{s_i}$ where $a_i = \denot{t_i}$ on environment $\set{x_j \mapsto c_j ~|~ j \in [1, m]}$. Based on the \textsf{Prod} rule in Figure~\ref{fig:buildFTA}, we know the transition $\sigma(q^{a_1}_{s_1}, \ldots, q^{a_n}_{s_n}) \to q^c_s$ is only added to $\Delta$ if there is production $\sigma(s_1, \ldots, s_n) \to s$ in grammar $G$ and $c = \denot{\sigma(a_1, \ldots, a_n)}$. Therefore, $t$ conforms to grammar $G$ and the run of $t$ with $x_i = c_i$ is rooted with state $q^c_s$ and $c = \denot{t(c_1, \ldots, c_m)}$. \end{itemize} Since term $t$ is accepted by $\mathcal{A}$, we know $q^c_s$ is a final state, which concludes the proof. \end{proof} \begin{lemma}\label{lem:sound-relaxed-term} Suppose $\mathcal{G},\mathcal{M} \vdash t \leadsto \mathcal{H}$ according to the rules from Figure~\ref{fig:build-hfta}, where $t$ is a term, and let $f_1, \ldots, f_n$ be the function symbols used in $t$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ that is accepted by $\mathcal{H}$, we have: \begin{enumerate} \item $\Upsilon(v_{f_i})$ is a program that conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$ \item The run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^{\denot{t}}_{s}$. \end{enumerate} \end{lemma} \begin{proof} Prove by structural induction on $t$. \begin{itemize} \item Base case: $t = c$. According to the \textsf{Const} rule in Figure~\ref{fig:build-hfta}, the HFTA $\mathcal{H}$ from $t$ contains only a single node $v_c$, so the root node is also $v_c$. The FTA $\Omega(v_c)$ is \[ \mathcal{A} = (\bigset{q^{\denot{c}}_c}, \set{c}, \bigset{q^{\denot{c}}_c}, \bigset{c \to q^{\denot{c}}_c}) \] Given $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$, we know tree $\Upsilon(v_c)$ is accepted by FTA $\Omega(v_c)$ by Definition~\ref{def:hfta-accept}. The run of $\Upsilon(v_c)$ is a single node $q^{\denot{c}}_c$. \item Inductive case: $t = f(t_1, \ldots, t_m)$. Suppose $\mathcal{G}, \mathcal{M} \vdash t_i \leadsto \mathcal{H}_i = (V_i, \Omega_i, v_{r_i}, \Lambda_i)$ for $i \in [1, m]$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$, we denote its $t_i$ part as $\mathcal{T}_i = (V_i, \Upsilon_i, v_{r_i}, E_i)$. On one hand, $V = \cup^{m}_{i=1} V_i \cup \set{v_f}$. $E = \cup^{m}_{i=1} E_i \cup \set{(v_f, v_{r_i}) ~|~ i \in [1, m]}$. $\Upsilon_i(v) = \Upsilon(v)$ for all $v \in V_i$. On the other hand, we have $\Omega_i(v) = \Omega(v)$ for all $v \in V_i$ based on the \textsf{Func} rule in Figure~\ref{fig:build-hfta}. Also, $\Lambda = \cup^{m}_{i=1} \Lambda_i \cup \bigset{q^c_{s_i} \to q^c_{x_i} ~|~ q^c_{s_i} \in Q_{f_i}, i \in [1, m]}$ where $Q_{f_i}$ is the final state set of $\Omega(v_{r_i})$. \hspace{0.1in} According to the acceptance condition of Definition~\ref{def:hfta-accept}, we know $\mathcal{T}_i$ is accepted by $\mathcal{H}_i$ for all $i \in [1, m]$. By inductive hypothesis, we have $\Upsilon_i(v_{f_i})$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$ for all $v_{f_i} \in V_i$. Moreover, we have proved $\Upsilon(v_f)$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f))$ in Lemma~\ref{lem:sound-fta}. Therefore, any function symbol $f_i$ in $t$ conforms to its grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$. Again, by inductive hypothesis, we have the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^{\denot{t_i}}_{s_i}$ for $i \in [1, m]$. Based on Lemma~\ref{lem:sound-fta}, the run of $\Upsilon(v_f)$ on $\Omega(v_f)$ is rooted with final state $q^c_s$ where $c = \denot{t(\denot{t_i}, \ldots, \denot{t_m})} = \denot{t}$. Hence, the lemma is proved for $t = f(t_1, \ldots, t_m)$. \end{itemize} By principle of structural induction, we have proved the lemma. \end{proof} \begin{lemma}\label{lem:sound-relaxed-formula} Suppose $\mathcal{G},\mathcal{M} \vdash \Phi \leadsto \mathcal{H}$ according to the rules from Figure~\ref{fig:build-hfta}, where $\Phi$ is a ground formula, and let $f_1, \ldots, f_n$ be the function symbols used in $\Phi$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ that is accepted by $\mathcal{H}$, we have: \begin{enumerate} \item $\Upsilon(v_{f_i})$ is a program that conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$ \item If the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^\top_{s_0}$, then $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n}) \models \Phi$ \item If the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^\bot_{s_0}$, then $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n}) \not\models \Phi$ \end{enumerate} \end{lemma} \begin{proof} Prove by structural induction on $\Phi$. \begin{itemize} \item Base case: $\Phi = t_1 ~ op ~ t_2$. Suppose $\mathcal{G}, \mathcal{M} \vdash t_1 \leadsto \mathcal{H}_1 = (V_1, \Omega_1, v_{r_1}, \Lambda_1)$ and $\mathcal{G}, \mathcal{M} \vdash t_2 \leadsto \mathcal{H}_2 = (V_2, \Omega_2, v_{r_2}, \Lambda_2)$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$, we denote its $t_1$ part as $\mathcal{T}_1 = (V_1, \Upsilon_1, v_{r_1}, E_1)$ and denote the $t_2$ part as $\mathcal{T}_2 = (V_2, \Upsilon_2, v_{r_2}, E_2)$. On one hand, $V = V_1 \cup V_2 \cup \set{v_{op}}$. $E = E_1 \cup E_2 \cup \set{(v_{op}, v_{r_1}), (v_{op}, v_{r_2})}$. $\Upsilon_i(v) = \Upsilon(v)$ for all $v \in V_i$ where $i = 1, 2$. On the other hand, we have $\Omega_i(v) = \Omega(v)$ for all $v \in V_i$ where $i = 1, 2$ based on the \textsf{Logical} rule in Figure~\ref{fig:build-hfta}. Also, $\Lambda = \Lambda_1 \cup \Lambda_2 \cup \bigset{q^c_{s_i} \to q^c_{x_i} ~|~ q^c_{s_i} \in Q_{f_i}, i = 1, 2}$ where $Q_{f_i}$ is the final state set of $\Omega(v_{r_i})$. \hspace{0.1in} According to the acceptance condition of Definition~\ref{def:hfta-accept}, we know $\mathcal{T}_1$ is accepted by $\mathcal{H}_1$ and $\mathcal{T}_2$ is accepted by $\mathcal{H}_2$. By Lemma~\ref{lem:sound-relaxed-term}, we have $\Upsilon_i(v_f)$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f))$ for all $v_f \in V_i$ where $i = 1, 2$. Therefore, any function symbol $f_i$ in $\Phi$ conforms to its grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$. Also by Lemma~\ref{lem:sound-relaxed-term}, we have the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^{\denot{t_i}}_{s_i}$ for $i = 1, 2$. In addition, consider that \textsf{Logical} rule will add transitions $\bigset{op(q^{c_1}_{x_1}, q^{c_2}_{x_2}) \to q^c_{s_0} ~|~ c = \denot{op}(c_1, c_2)}$ to $\Omega(v_r)$, the run will reach state $q^\top_{s_0}$ if $t_1 ~ op ~ t_2$ evaluates to \emph{true} and reach state $q^\bot_{s_0}$ otherwise. Hence, the lemma is proved for $\Phi = t_1 ~ op ~ t_2$. \item Inductive case: $\Phi = \Phi_1 \odot \Phi_2 ~ (\odot = \land, \lor).$ Suppose $\mathcal{G}, \mathcal{M} \vdash \Phi_1 \leadsto \mathcal{H}_1 = (V_1, \Omega_1, v_{r_1}, \Lambda_1)$ and $\mathcal{G}, \mathcal{M} \vdash \Phi_2 \leadsto \mathcal{H}_2 = (V_2, \Omega_2, v_{r_2}, \Lambda_2)$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$, we denote its $\Phi_1$ part as $\mathcal{T}_1 = (V_1, \Upsilon_1, v_{r_1}, E_1)$ and denote the $\Phi_2$ part as $\mathcal{T}_2 = (V_2, \Upsilon_2, v_{r_2}, E_2)$. On one hand, $V = V_1 \cup V_2 \cup \set{v_{\odot}}$. $E = E_1 \cup E_2 \cup \set{(v_{\odot}, v_{r_1}), (v_{\odot}, v_{r_2})}$. $\Upsilon_i(v) = \Upsilon(v)$ for all $v \in V_i$ where $i = 1, 2$. On the other hand, we have $\Omega_i(v) = \Omega(v)$ for all $v \in V_i$ where $i = 1, 2$ based on the \textsf{Logical} rule in Figure~\ref{fig:build-hfta}. Also, $\Lambda = \Lambda_1 \cup \Lambda_2 \cup \bigset{q^c_{s_i} \to q^c_{x_i} ~|~ q^c_{s_i} \in Q_{f_i}, i = 1, 2}$ where $Q_{f_i}$ is the final state set of $\Omega(v_{r_i})$. \hspace{0.1in} According to the acceptance condition of Definition~\ref{def:hfta-accept}, we know $\mathcal{T}_1$ is accepted by $\mathcal{H}_1$ and $\mathcal{T}_2$ is accepted by $\mathcal{H}_2$. By inductive hypothesis, we have $\Upsilon_i(v_f)$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f))$ for all $v_f \in V_i$ where $i = 1, 2$. Therefore, any function symbol $f_i$ in $\Phi$ conforms to its grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$. Also by inductive hypothesis, we know that if the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^\top_{s_i}$ then $\Upsilon_i(v_{f_i}), v_{f_i} \in V_i $ collectively satisfy $\Phi_i$ for $i = 1, 2$. In addition, the \textsf{Logical} rule is essentially building an FTA that accepts programs with boolean input value $\denot{\Phi_1}$ and $\denot{\Phi_2}$. The run is rooted with $q^\top_{s_0}$ if $\denot{\Phi_1} \odot \denot{\Phi_2}$ evaluates to \emph{true}; otherwise, the run is rooted with $q^\bot_{s_0}$. Hence, the lemma is proved for $\Phi = \Phi_1 \odot \Phi_2$. \item Inductive case: $\Phi = \neg \Phi_1$. Suppose $\mathcal{G}, \mathcal{M} \vdash \Phi_1 \leadsto \mathcal{H}_1 = (V_1, \Omega_1, v_{r_1}, \Lambda_1)$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$, we denote its $\Phi_1$ part as $\mathcal{T}_1 = (V_1, \Upsilon_1, v_{r_1}, E_1)$. On one hand, $V = V_1 \cup \set{v_{\neg}}$. $E = E_1 \cup \set{(v_{\neg}, v_{r_1})}$. $\Upsilon_1(v) = \Upsilon(v)$ for all $v \in V_1$. On the other hand, we have $\Omega_1(v) = \Omega(v)$ for all $v \in V_1$ based on the \textsf{Neg} rule in Figure~\ref{fig:build-hfta}. Also, $\Lambda = \Lambda_1 \cup \bigset{q^c_s \to q^c_{x_1} ~|~ q^c_s \in Q_{f_1}}$ where $Q_{f_1}$ is the final state set of $\Omega(v_{r_1})$. \hspace{0.1in} According to the acceptance condition of Definition~\ref{def:hfta-accept}, we know $\mathcal{T}_1$ is accepted by $\mathcal{H}_1$. By inductive hypothesis, we have $\Upsilon_1(v_f)$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f))$ for all $v_f \in V_1$. Therefore, any function symbol $f_i$ in $\Phi$ conforms to its grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$. Also by inductive hypothesis, we have the run of $\Upsilon_1(v_{r_1})$ on $\Omega_1(v_{r_1})$ is rooted with final state $q^{\denot{t_1}}_{s_1}$. In addition, the \textsf{Neg} rule is building an FTA that accepts programs with boolean input value $\denot{\Phi_1}$. The run is rooted with $q^\top_{s_0}$ if $\neg \denot{\Phi_1}$ evaluates to \emph{true}; otherwise, the run is rooted with $q^\bot_{s_0}$. Hence, the lemma is proved for $\Phi = \neg \Phi_1$. \end{itemize} By principle of structural induction, we have proved the lemma. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:sound-relaxed}] Suppose $\mathcal{G}, \mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$ and $\mathcal{H} = (V, \Omega, v_r, \Lambda)$, we know $\mathcal{G}, \mathcal{M} \vdash \Phi \leadsto \mathcal{H}'$ with $\mathcal{H}' = (V, \Omega', v_r, \Lambda)$ according to the \textrm{Final} rule in Figure~\ref{fig:build-hfta}. The only difference between $\Omega$ and $\Omega'$ is that $\Omega(v_r) = (Q, \Sigma, \set{q^\top_{s_0}}, \Delta)$ but $\Omega'(v_r) = (Q, \Sigma, \set{q^\top_{s_0}, q^\bot_{s_0}}, \Delta)$. By Definition~\ref{def:hfta-accept}, we know $\mathcal{L}(\mathcal{H}) \subseteq \mathcal{L}(\mathcal{H}')$. Thus, if hierarchical tree $\mathcal{T}$ is accepted by $\mathcal{H}$, then $\mathcal{T}$ is also accepted by $\mathcal{H}'$. According to Lemma~\ref{lem:sound-relaxed-formula}, we have $\Upsilon(v_{f_i})$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$. In addition, we have $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n}) \models \Phi$ if the run of $\Upsilon(v_r)$ is rooted with final state $q^\top_{s_0}$. Since $\mathcal{H}$ only has one final state $q^\top_{s_0}$ in $\Omega(v_r)$, we know the accepting run of $\Upsilon(v_r)$ on $\Omega(v_r)$ must be rooted with $q^\top_{s_0}$. Thus, $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n}) \models \Phi$. \end{proof} \begin{lemma}\label{lem:complete-fta} Suppose FTA $\mathcal{A} = (Q, \Sigma, Q_f, \Delta)$ is built with grammar $G$ and initial states $[Q_1, \ldots, Q_m]$ using rules from Figure~\ref{fig:buildFTA}. If there is a program $P$ that conforms to grammar $G$ such that $\denot{P}(a_1, \ldots, a_m) = a$ where $q^{a_i}_{s_i} \in Q_i$ for $i \in [1, m]$, then there exists a tree accepted by $\mathcal{A}$ and the run is rooted with final state $q^a_s$ from initial states $q^{a_1}_{x_1}, \ldots, q^{a_m}_{x_m}$. \end{lemma} \begin{proof} To prove by induction on height of $P$'s AST, we first strengthen the lemma to: the run of tree $t$ is rooted with state $q^a_s$ from initial states $q^{a_1}_{x_1}, \ldots, q^{a_m}_{x_m}$ where $a = \denot{t}(a_1, \ldots, a_m)$. \begin{itemize} \item Base case: height is 1. In this case, $P$ only has one variable $x$, and $q^c_{s}$ is passed as an input. Consider the \textsf{Input} rule in Figure~\ref{fig:buildFTA}, we know $q^c_{x} \in Q$ and $x \to q^c_x \in \Delta$. Then $x$ is accepted by $\mathcal{A}$, and the run is rooted with final state $q^c_{x}$. \item Inductive case: Suppose the lemma is correct for AST with height no larger than $h$, and the goal is to prove it is correct for AST of height $h+1$. Without loss of generality, assume $P = \sigma(P_1, \ldots, P_n)$ is of height $h+1$, so the height of partial programs $P_1, \ldots, P_n$ is at most $h$. By inductive hypothesis, there exists trees $t_1, \ldots, t_n$ whose runs are rooted with state $q^{c_i}_{s_i}$ form initial states $q^{a_1}_{x_1}, \ldots, q^{a_m}_{x_m}$, and $\denot{t_i} = c_i$ for all $i \in [1, n]$. According to the \textsf{Prod} rule in Figure~\ref{fig:buildFTA}, we know the run of $t = \sigma(t_1, \ldots, t_n)$ is rooted with state $\denot{\sigma(\denot{t_1}, \ldots, \denot{t_n})} = \denot{\sigma(c_1, \ldots, c_n)}$. Hence, the inductive case is proved. \end{itemize} Consider $P$ is a program (not partial program), if the run of tree $t$ is rooted with state $q^a_s$, then $s$ must be a starting symbol. By the \textsf{Output} rule in Figure~\ref{fig:buildFTA}, $q^a_s$ is a final state. Therefore, tree $t$ is accepted by $\mathcal{A}$, and we have proved the lemma. \end{proof} \begin{lemma}\label{lem:complete-relaxed-term} Let $t$ be a term where every function symbol $f_1, \ldots, f_n$ occurs exactly once, and suppose we have $\mathcal{G},\mathcal{M} \vdash t \leadsto \mathcal{H}$. If there are implementations $P_i$ of $f_i$ such that $\denot{t} = a$, where $P_i$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$, then there exists a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$ such that $\Upsilon(v_{f_i}) = P_i$ for all $i \in [1,n]$, and the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^a_s$. \end{lemma} \begin{proof} Prove by structural induction on $t$. \begin{itemize} \item Base case: $t = c$. Suppose the HFTA built from \textsf{Const} rule in Figure~\ref{fig:build-hfta} is $\mathcal{H} = (\set{v_c}, \Omega, v_c, \Lambda)$ and $\Omega(v_c)$ is \[ \mathcal{A} = (\bigset{q^{\denot{c}}_c}, \set{c}, \bigset{q^{\denot{c}}_c}, \bigset{c \to q^{\denot{c}}_c}) \] Consider the hierarchical tree $\mathcal{T} = (\set{v_c}, \Upsilon, v_c, \emptyset)$ where $\Upsilon(v_c)$ is a tree that only contains a single node $c$. By Definition~\ref{def:hfta-accept}, $\mathcal{T}$ is accepted by $\mathcal{H}$ and the run of $\Upsilon(v_c)$ on $\Omega(v_c)$ is rooted with final state $q^{\denot{c}}_c$. \item Inductive case: $t = f(t_1, \ldots, t_m)$. Given $\mathcal{G}, \mathcal{M} \vdash t \leadsto \mathcal{H} = (V, \Omega, v_f, \Lambda)$, we denote its $t_i$ part as $\mathcal{H}_i = (V_i, \Omega_i, v_{r_i}, \Lambda_i)$ for $i \in [1, m]$. In addition, we have $V_i(v_i) = V(v_i)$ and $\Omega_i(v_i) = V(v_i)$ for all node $v_i \in V_i$. $\Lambda_i \subset \Lambda$. Note that the $t_i$ part is recognized by inter-FTA transition $q^c_s \to q^c_{x_i} \in \Lambda$ where $q^c_s$ is a state in $\Omega_i(v_{r_i})$ and $q^c_{x_i}$ is a state in $\Omega(v_f)$. According to the \textsf{Func} rule in Figure~\ref{fig:build-hfta}, we have $\mathcal{G}, \mathcal{M} \vdash t_i \leadsto \mathcal{H}_i$ for $i \in [1, m]$. By inductive hypothesis, we know there exists a hierarchical tree $\mathcal{T}_i = (V_i, \Upsilon_i, v_{r_i}, E_i)$ accepted by $\mathcal{H}_i$ and the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^{\denot{t_i}}_{s_i}$ for all $i \in [1, m]$. Moreover, by Lemma~\ref{lem:complete-fta}, we know there exists a tree $T_f$ accepted by $\Omega(v_f)$ and the run is rooted with final state $q^a_s$ from initial states $q^{\denot{t_1}}_{s_1}, \ldots, q^{\denot{t_m}}_{s_m}$, where $a = \denot{f(\denot{t_1}, \ldots, \denot{t_m})}$. Now consider the hierarchical tree \[ \mathcal{T} = (\cup^m_{i=1}V_i \cup \set{v_f}, ~ \cup^m_{i=1}\Upsilon_i \cup \set{v_f \mapsto T_f}, ~ v_f, ~ \cup^m_{i=1} (E_i \cup \set{(v_f, v_{r_i})} )) \] By acceptance condition of HFTA in Defintion~\ref{def:hfta-accept}, $\mathcal{T}$ is accepted by $\mathcal{H}$. Thus, the lemma is proved for $t = f(t_1, \ldots, t_m)$ case. \end{itemize} By principle of structural induction, we have proved the lemma. \end{proof} \begin{lemma}\label{lem:complete-relaxed-formula} Let $\Phi$ be a ground formula where every function symbol $f_1, \ldots, f_n$ occurs exactly once, and suppose we have $\mathcal{G},\mathcal{M} \vdash \Phi \leadsto \mathcal{H}$. If there are implementations $P_i$ of $f_i$ where $P_i$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$, then there exists a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$ such that $\Upsilon(v_{f_i}) = P_i$ for all $i \in [1,n]$, and \begin{enumerate} \item If $P_1, \ldots, P_n \models \Phi$, the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^\top_{s_0}$. \item If $P_1, \ldots, P_n \not\models \Phi$, the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^\bot_{s_0}$. \end{enumerate} \end{lemma} \begin{proof} Prove by structural induction on $\Phi$. \begin{itemize} \item Base case: $\Phi = t_1 ~ op ~ t_2$. Given $\mathcal{G}, \mathcal{M} \vdash t \leadsto \mathcal{H} = (V, \Omega, v_{op}, \Lambda)$, we denote its $t_i$ part as $\mathcal{H}_i = (V_i, \Omega_i, v_{r_i}, \Lambda_i)$ for $i = 1, 2$. In addition, we have $V_i(v_i) = V(v_i)$ and $\Omega_i(v_i) = V(v_i)$ for all node $v_i \in V_i$. $\Lambda_i \subset \Lambda$. Note that the $t_i$ part is recognized by inter-FTA transition $q^c_s \to q^c_{x_i} \in \Lambda$ where $q^c_s$ is a state in $\Omega_i(v_{r_i})$ and $q^c_{x_i}$ is a state in $\Omega(v_{op})$. According to the \textsf{Logical} rule in Figure~\ref{fig:build-hfta}, we have $\mathcal{G}, \mathcal{M} \vdash t_i \leadsto \mathcal{H}_i$ for $i = 1, 2$. By Lemma~\ref{lem:complete-relaxed-term}, we know there exists a hierarchical tree $\mathcal{T}_i = (V_i, \Upsilon_i, v_{r_i}, E_i)$ accepted by $\mathcal{H}_i$ and the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^{\denot{t_i}}_{s_i}$ for $i = 1, 2$. Moreover, based on the \textsf{Logical} rule, we know there exists a tree $T_{op} = op(x_1, x_2)$ accepted by $\Omega(v_{op})$. The run is rooted with $q^\top_{s_0}$ if $\denot{t_1} ~ op ~ \denot{t_2}$ evaluates to \emph{true} (i.e., $P_1, \ldots, P_n \models t_1 ~ op t_2$); the run is rooted with $q^\bot_{s_0}$ otherwise. Now consider the hierarchical tree \[ \mathcal{T} = (V_1 \cup V_2 \cup \set{v_{op}}, ~ \Upsilon_1 \cup \Upsilon_2 \cup \set{v_{op} \mapsto T_{op}}, ~ v_{op}, ~ E_1 \cup E_2 \cup \set{(v_{op}, v_{r_1}), (v_{op}, v_{r_2})} )) \] By acceptance condition of HFTA in Defintion~\ref{def:hfta-accept}, $\mathcal{T}$ is accepted by $\mathcal{H}$. Thus, the lemma is proved for $\Phi = t_1 ~ op ~ t_2$. \item Inductive case: $\Phi = \Phi_1 \odot \Phi_2 ~ (\odot = \land, \lor)$. Given $\mathcal{G}, \mathcal{M} \vdash t \leadsto \mathcal{H} = (V, \Omega, v_{\odot}, \Lambda)$, we denote its $\Phi_i$ part as $\mathcal{H}_i = (V_i, \Omega_i, v_{r_i}, \Lambda_i)$ for $i = 1, 2$. In addition, we have $V_i(v_i) = V(v_i)$ and $\Omega_i(v_i) = V(v_i)$ for all node $v_i \in V_i$. $\Lambda_i \subset \Lambda$. Note that the $\Phi_i$ part is recognized by inter-FTA transition $q^c_s \to q^c_{x_i} \in \Lambda$ where $q^c_s$ is a state in $\Omega_i(v_{r_i})$ and $q^c_{x_i}$ is a state in $\Omega(v_{\odot})$. According to the \textsf{Logical} rule in Figure~\ref{fig:build-hfta}, we have $\mathcal{G}, \mathcal{M} \vdash \Phi_i \leadsto \mathcal{H}_i$ for $i = 1, 2$. By inductive hypothesis, we know there exists a hierarchical tree $\mathcal{T}_i = (V_i, \Upsilon_i, v_{r_i}, E_i)$ accepted by $\mathcal{H}_i$ and the run of $\Upsilon_i(v_{r_i})$ on $\Omega_i(v_{r_i})$ is rooted with final state $q^\top_{s_0}$ if $\Phi_i$ is satisfied; the run is rooted with final state $q^\top_{s_0}$ otherwise. Moreover, based on the \textsf{Logical} rule, we know there exists a tree $T_{\odot} = \odot(x_1, x_2)$ accepted by $\Omega(v_{\odot})$. The run is rooted with $q^\top_{s_0}$ if $\denot{\Phi_1} \odot \denot{\Phi_2}$ evaluates to \emph{true}, but rooted with $q^\bot_{s_0}$ otherwise. Now consider the hierarchical tree \[ \mathcal{T} = (V_1 \cup V_2 \cup \set{v_{\odot}}, ~ \Upsilon_1 \cup \Upsilon_2 \cup \set{v_{\odot} \mapsto T_{\odot}}, ~ v_{\odot}, ~ E_1 \cup E_2 \cup \set{(v_{\odot}, v_{r_1}), (v_{\odot}, v_{r_2})} ) \] By acceptance condition of HFTA in Defintion~\ref{def:hfta-accept}, $\mathcal{T}$ is accepted by $\mathcal{H}$. Thus, the lemma is proved for $\Phi = \Phi_1 \odot \Phi_2$. \item Inductive case: $\Phi = \neg \Phi_1$. Given $\mathcal{G}, \mathcal{M} \vdash t \leadsto \mathcal{H} = (V, \Omega, v_{\neg}, \Lambda)$, we denote its $\Phi_1$ part as $\mathcal{H}_1 = (V_1, \Omega_1, v_{r_1}, \Lambda_1)$. In addition, we have $V_1(v_1) = V(v_1)$, $\Omega_1(v_1) = V(v_1)$, and $\Lambda_1 \subset \Lambda$. Note that the $\Phi_1$ part is recognized by inter-FTA transition $q^c_s \to q^c_{x_1} \in \Lambda$ where $q^c_s$ is a state in $\Omega_1(v_{r_1})$ and $q^c_{x_1}$ is a state in $\Omega(v_{\neg})$. According to the \textsf{Neg} rule in Figure~\ref{fig:build-hfta}, we have $\mathcal{G}, \mathcal{M} \vdash \Phi_1 \leadsto \mathcal{H}_1$. By inductive hypothesis, we know there exists a hierarchical tree $\mathcal{T}_1 = (V_1, \Upsilon_1, v_{r_1}, E_1)$ accepted by $\mathcal{H}_1$ and the run of $\Upsilon_1(v_{r_1})$ on $\Omega_1(v_{r_1})$ is rooted with final state $q^\top_{s_0}$ if $\Phi_1$ is satisfied; the run is rooted with final state $q^\top_{s_0}$ otherwise. Moreover, based on the \textsf{Neg} rule, we know there exists a tree $T_{\neg} = \neg(x_1)$ accepted by $\Omega(v_{\neg})$. The run is rooted with $q^\top_{s_0}$ if $\neg \denot{\Phi_1}$ evaluates to \emph{true}, but rooted with $q^\bot_{s_0}$ otherwise. Now consider the hierarchical tree \[ \mathcal{T} = (V_1 \cup \set{v_{\neg}}, ~ \Upsilon_1 \cup \set{v_{\neg} \mapsto T_{\neg}}, ~ v_{\neg}, ~ E_1 \cup \set{(v_{\neg}, v_{r_1})} ) \] By acceptance condition of HFTA in Defintion~\ref{def:hfta-accept}, $\mathcal{T}$ is accepted by $\mathcal{H}$. Thus, the lemma is proved for $\Phi = \neg \Phi_1$. \end{itemize} By principle of structural induction, we have proved the lemma. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:complete-relaxed}] Suppose $\mathcal{G}, \mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$ and $\mathcal{H} = (V, \Omega, v_r, \Lambda)$, we have $\mathcal{G}, \mathcal{M} \vdash \Phi \leadsto \mathcal{H}'$ with $\mathcal{H}' = (V, \Omega', v_r, \Lambda)$ according to the \textrm{Final} rule in Figure~\ref{fig:build-hfta}. The only difference between $\Omega$ and $\Omega'$ is that $\Omega(v_r) = (Q, \Sigma, \set{q^\top_{s_0}}, \Delta)$ but $\Omega'(v_r) = (Q, \Sigma, \set{q^\top_{s_0}, q^\bot_{s_0}}, \Delta)$. By Lemma~\ref{lem:complete-relaxed-formula}, we know there exists a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}'$ such that $\Upsilon(v_{f_i}) = P_i$ for all $i \in [1, n]$. Furthermore, since $P_1, \ldots, P_n \models \Phi$, we know the run of $\Upsilon(v_r)$ on $\Omega(v_r)$ is rooted with final state $q^\top_{s_0}$. Since $q^\top_{s_0}$ is also a final state in $\Omega'(v_r)$, the run of $\Upsilon(v_r)$ on FTA $\Omega(v_r)$ is the same as the run of $\Upsilon(v_r)$ on FTA $\Omega'(v_r)$. According to the HFTA accepting condition in Definition~\ref{def:hfta-accept}, $\mathcal{T}$ is also accepted by $\mathcal{H}$. \end{proof} \begin{lemma}\label{lem:find-sound} Suppose we have a ground relational specification $\Phi$, its relaxation $\Phi'$ with symbol mapping $\mathcal{M}$, and an HFTA $\mathcal{H}$ built from $\Phi'$, $\mathcal{M}$ and grammars $\mathcal{G}$. If the \textsc{FindProgs} procedure returns programs $\mathcal{P}$ (not \emph{null}) given input $\mathcal{H}$ and $\mathcal{M}$, then $\mathcal{P}$ satisfy $\Phi$, and program $P_i$ for function $f_i$ conforms to the grammar $\mathcal{G}(f_i)$. \end{lemma} \begin{proof} If \textsc{FindProgs} actually returns programs $\mathcal{P}$, then there must be a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$. Otherwise, it will skip the for loop and directly return \emph{null}. For ease of illustration, we assign $\mathcal{H}_0 = \mathcal{H}$ and assign $\mathcal{T}_0 = \mathcal{T} = (V, \Upsilon_0, v_r, E)$. Observe that the \textsc{FindProgs} procedure assigns one program to a function symbol at a time, and terminates when all function symbols are successfully assigned. Without loss of generality, we assume the function symbols are assigned in the order of $f_1, \ldots, f_n$. We denote by $\mathcal{T}_i = (V, \Upsilon_i, v_r, E)$ the hierarchical tree and $\Upsilon_i$ is obtained by substituting \emph{all} $\Upsilon_{i-1}(v_{f'_i})$ in $\mathcal{T}_{i-1}$ with program tree $P_i$, where $f'_i$ is a symbol in $\mathcal{M}(f_i)$. We also denote the HFTA of invoking \textsc{Propagate}($\mathcal{H}_{i-1}, P_i, f_i, \mathcal{M}$) by $\mathcal{H}_i$. To prove programs $P_1, \ldots, P_n$ in $\mathcal{P}$ satisfy $\Phi$ and program $P_i$ for function $f_i$ conforms to grammar $\mathcal{G}(f_i)$, consider the following invariant: For any integer $i \in [0, n]$, \begin{enumerate} \item $\mathcal{L}(\mathcal{H}_i) \subseteq \mathcal{L}(\mathcal{H})$. \item For all $k \leq i$, function symbol $f_k$ in original ground specification $\Phi$ is consistently assigned to program $P_k$, i.e. $P_k = \Upsilon_i(v_{f'_k})$ for all $f'_k \in \mathcal{M}(f_k)$. \end{enumerate} First, let us prove it is indeed an invariant. \begin{itemize} \item Base case: $i = 0$. Given $\mathcal{H}_0 = \mathcal{H}$, we have $\mathcal{L}(\mathcal{H}_0) \subseteq \mathcal{L}(\mathcal{H})$. (2) trivially holds because no program has been assigned for $i = 0$. \item Inductive case: Suppose (1)(2) hold for $i = j$, the goal is to prove they also hold for $i = j+1$. \begin{enumerate} \item Since \textsc{Propagate} will only remove final states and inter-FTA transitions from $\mathcal{H}_j$, we know $\mathcal{L}(\mathcal{H}_{j+1}) \subseteq \mathcal{L}(\mathcal{H}_j)$ by acceptance condition in Definition~\ref{def:hfta-accept}. Since $\mathcal{L}(\mathcal{H}_j) \subseteq \mathcal{L}(\mathcal{H})$ by inductive hypothesis, we have $\mathcal{L}(\mathcal{H}_{j+1}) \subseteq \mathcal{L}(\mathcal{H})$. \item By definition of $\mathcal{T}_{j+1}$, $\Upsilon_{j+1}(v_{f'_{j+1}}) = P_{j+1}$ holds for all $f'_{j+1} \in \mathcal{M}(f_{j+1})$. Also by inductive hypothesis, we know for all $k \leq j$, $P_k = \Upsilon_k(v_{f'_k})$. Thus, for all $k \leq j+1$, we have $P_k = \Upsilon_k(v_{f'_k})$ for all $f'_k \in \mathcal{M}(f_k)$. \end{enumerate} \end{itemize} After assigning programs to all function symbols, we have \begin{enumerate} \item $\mathcal{L}(\mathcal{H}_n) \subseteq \mathcal{L}(\mathcal{H})$. \item For all function symbol $f_i$ in $\Phi$, function occurrences $f'_i$ are consistently assigned to program $P_i$ for all $f'_i \in \mathcal{M}(f_i)$. \end{enumerate} Since the programs $\mathcal{P}$ are accepted by the last HFTA $\mathcal{H}_n$, i.e. $\mathcal{T}_n \in \mathcal{L}(\mathcal{H}_n)$ (otherwise \textsc{Propagate} would return \emph{null} due to the recurrent failure of emptiness check on $\mathcal{H}_n$, we have $\mathcal{T}_n \in \mathcal{L}(\mathcal{H})$. By Theorem~\ref{thm:sound-relaxed}, we have $\Upsilon(v_{f'_i})$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$ and $\Upsilon(v_{f'_1}), \ldots, \Upsilon(v_{f'_n}) \models \Phi'$. Consider the fact that $f_i = \mathcal{M}^{-1}(f'_i) \Leftrightarrow f'_i \in \mathcal{M}(f_i)$ and the only difference between $\Phi$ and $\Phi'$ is captured by $\mathcal{M}$, we have proved $P_1, \ldots, P_n \models \Phi$ and $P_i$ for function $f_i$ conforms to grammar $\mathcal{G}(f_i)$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:sound}] Observe the \textsc{Synthesize} procedure shown in Algorithm~\ref{algo:cegis} can only return a set of programs $\mathcal{P}$ after it invokes the \textsf{Verify} procedure and successfully verifies $\mathcal{P}$ indeed satisfy the relational specification $\Psi$. Thus, the programs $\mathcal{P}$ returned by \textsc{Synthesize} are guaranteed to satisfy $\Psi$. In addition, observe that Algorithm~\ref{algo:cegis} only has two ways to generate candidate programs that may be potentially returned after verification. \begin{enumerate} \item Randomly generate programs $\mathcal{P}_{\emph{random}}$ that conform to context-free grammars $\mathcal{G}$ in the initialization phase. \item Generate programs $\mathcal{P}_{\emph{res}}$ by the \textsc{FindProgs} procedure. According to Lemma~\ref{lem:find-sound}, for program $P_i$ of function $f_i$ in $\mathcal{P}_{\emph{res}}$, program $P_i$ conforms to the grammar $\mathcal{G}(f_i)$. \end{enumerate} Put two cases together, we have program $P_i$ of function symbol $f_i$ conforms to grammar $\mathcal{G}(f_i)$. \end{proof} \begin{lemma}\label{lem:find-complete} Suppose we have a ground relational specification $\Phi$, its relaxation $\Phi'$ with symbol mapping $\mathcal{M}$, and an HFTA $\mathcal{H}$ built from $\Phi'$, $\mathcal{M}$, and grammars $\mathcal{G}$. If there exist a set of programs that (a) satisfy $\Phi$ and (b) can be implemented using the DSLs given by $\mathcal{G}$, then invoking \textsc{FindProgs} on $\mathcal{H}$ and $\mathcal{M}$ will eventually return programs $\mathcal{P}$ satisfying $\Phi$. \end{lemma} \begin{proof} Suppose there exist a set of programs that satisfy $\Phi$ and can be implemented using the DSLs given by $\mathcal{G}$, we denote the program for function $f_i$ by $P_i$. Given that the relationship between $\Phi$ and $\Phi'$ is captured by mapping $\mathcal{M}$, and $f_i = \mathcal{M}^{-1}(f'_i) \Leftrightarrow f'_i \in \mathcal{M}(f_i)$, we know that assigning program $P_i$ for \emph{all} $f'_i$ if $f'_i \in \mathcal{M}(f_i)$ would satisfy $\Phi'$, and $P_i$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f'_i)$. By Theorem~\ref{thm:complete-relaxed}, we know there exists a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$ such that $\Upsilon(v_{f'_i}) = P_i$ for any $i \in [1, n]$. Observe that the \textsc{FindProgs} procedure essentially employs standard backtrack search and may potentially enumerate all possible hierarchical trees accepted by $\mathcal{H}$, it will eventually find $\mathcal{T}$ and thus return the program set $\mathcal{P}$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:complete}] Suppose the ground relational specification obtained by $i$-th loop iteration of Algorithm~\ref{algo:cegis} is denoted by $\Phi_i$, and $\Phi_0 = \top$, let us first prove that $\Psi \Rightarrow \Phi_i$ for any natural number $i$ by induction on $i$. \begin{itemize} \item Base case: $i = 0$. $\Psi \Rightarrow \top$ holds obviously. \item Inductive case: Suppose $\Psi \Rightarrow \Phi_{i-1}$ holds, we would like to prove $\Psi \Rightarrow \Phi_i$. Given $\Phi_i = \Phi_{i-1} \land \emph{GetRelationalCounterexample}(\mathcal{P}, \Psi)$, and consider that $\Psi \Rightarrow \emph{GetRelationalCounterexample}(\mathcal{P}, \Psi)$ because the counterexamples are obtained by instantiating the universal quantifiers in $\Psi$, we have $\Psi \Rightarrow \Phi_i$. \end{itemize} According to the principle of induction, $\Psi \Rightarrow \Phi_i$ holds for any natural number $i$. If there exist programs $\mathcal{P}$ that satisfy $\Psi$ and $\mathcal{P}$ can be implemented using the DSLs given by $\mathcal{G}$, then programs $\mathcal{P}$ also satisfy $\Phi_i$ for any $i$. By Lemma~\ref{lem:find-complete}, we know the \textsc{FindProgs} procedure will always return programs satisfying $\Phi_i$. Since the Algorithm~\ref{algo:cegis} will not terminate until it finds a set of programs that satisfy $\Psi$ or the \textsc{FindProgs} procedure fails to find any programs, and we have proved \textsc{FindProgs} will never return failure, we can conclude that Algorithm~\ref{algo:cegis} will eventually terminate with programs satisfying $\Psi$. \end{proof} \section{Related Work} In this section, we survey prior work that is most closely related to relational program synthesis. \vspace{-0.05in} \paragraph{\textbf{Relational Program Verification.}} There is a large body of work on \emph{verifying} relational properties about programs~\cite{rhl,chl,relsep,product1,mediator,merge}. Existing work on relational verification can be generally categorized into three classes, namely \emph{relational program logics}, \emph{product programs}, and \emph{Constrained Horn Clause (CHC) solving}. The first category includes Benton's Relational Hoare Logic~\cite{rhl} and its variants~\cite{prhl,prhl2,relsep} as well as Cartesian Hoare Logic~\cite{chl,qchl} for verifying $k$-safety properties. In contrast to these approaches that provide a dedicated program logic for reasoning about relational properties, an alternative approach is to build a so-called \emph{product program} that captures the simultaneous execution of multiple programs or different runs of the same program~\cite{product1,barthe2004secure,product-asymmetric}. In the simplest case, this approach sequentially composes different programs (or copies of the same program)~\cite{barthe2004secure}, but more sophisticated product construction methods perform various transformations such as loop fusion and unrolling to make invariant generation easier~\cite{product1, product-asymmetric,program-consolidation,covac}. A common theme underlying all these product construction techniques is to reduce the relational verification problem to a standard safety checking problem. Another alternative approach that has been explored in prior work is to reduce the relational verification problem to solving a (recursive) set of Constrained Horn Clauses (CHC) and apply transformations that make the problem easier to solve~\cite{chc1,chc2}. To the best of our knowledge, this paper is the first one to address the dual \emph{synthesis} variant of the relational verification problem. \vspace{-0.05in} \paragraph{\textbf{Program Synthesis.}} This paper is related to a long line of work on program synthesis dating back to the 1960s~\cite{green}. Generally speaking, program synthesis techniques can be classified into two classes depending on whether they perform \emph{deductive} or \emph{inductive} reasoning. In particular, deductive synthesizers generate correct-by-construction programs by applying refinement and transformation rules~\cite{refinement,manna1986deductive,fiat,leon,synquid}. In contrast, inductive synthesizers learn programs from input-output examples using techniques such as constraint solving~\cite{sketch1,simpl}, enumerative search~\cite{eusolver,lambda2}, version space learning~\cite{prose}, stochastic search~\cite{stratified-synth,stoke}, and statistical models and machine learning~\cite{deepsyn,jsnice}. Similar to most recent work in this area~\cite{sketch1,prose,syngar,gulwani2012synthesis}, our proposed method also uses inductive synthesis. However, a key difference is that we use relational examples in the form of ground formulas rather than the more standard input-output examples. \vspace{-0.05in} \paragraph{\textbf{Counterexample-guided Inductive Synthesis.}} The method proposed in this paper follows the popular counterexample-guided inductive synthesis (CEGIS) methodology~\cite{sketch1,sketch2,sygus}. In the CEGIS paradigm, an inductive synthesizer generates a candidate program $P$ that is consistent with a set of examples, and the verifier checks the correctness of $P$ and provides counterexamples when $P$ does not meet the user-provided specification. Compared to other synthesis algorithms that follow the CEGIS paradigm, the key differences of our method are (a) the use of relational counterexamples and (b) a new inductive synthesis algorithm that utilizes relational specifications. \vspace{-0.05in} \paragraph{\textbf{Version Space Learning.}} As mentioned earlier, the inductive synthesizer used in this work can be viewed as a generalization of \emph{version space learning} to the relational setting. The notion of \emph{version space} was originally introduced in the 1980s as a supervised learning framework~\cite{vs} and has found numerous applications within the field of program synthesis~\cite{vsa,prose,flashfill,dace,fidex,mitra}. Generally speaking, synthesis algorithms based on version space learning construct some sort of data structure that represents all programs that are consistent with the examples. While existing version space learning algorithms only work with input-output examples, the method proposed here works with arbitrary ground formulas representing relational counterexamples and uses hierarchical finite tree automata to compose the version spaces of individual functions. \vspace{-0.05in} \paragraph{\textbf{Program Inversion.}} Prior work has addressed the \emph{program inversion} problem, where the goal is to automatically generate the inverse of a given program~\cite{inv1,inv2,inv3,inversion-swarat,inversion-loris}. Among these, the PINS tool uses inductive synthesis to semi-automate the inversion process based on templates provided by the user~\cite{inversion-swarat}. More recent work describes a fully automated technique, based on symbolic transducers, to generate the inverse of a given program~\cite{inversion-loris}. While program inversion is one of the applications that we consider in this paper, relational synthesis is applicable to many problems beyond program inversion. Furthermore, our approach can be used to synthesize the program and its inverse \emph{simultaneously} rather than requiring the user to provide one of these functions. \section{Relational Program Synthesis using HFTA} \label{sec:synth} In this section, we describe the high-level structure of our relational synthesis algorithm and explain its sub-procedures in the following subsections. Our top-level synthesis algorithm is shown in Algorithm~\ref{algo:cegis}. The procedure {\sc Synthesize} takes as input a tuple of function symbols $\mathcal{F} = (f_1, \ldots, f_n)$, their corresponding context-free grammars $\mathcal{G} = (G_1, \ldots, G_n)$, and a relational specification $\Psi = \forall \vec{x}. \ \phi(\vec{x})$. The output is a mapping $\mathcal{P}$ from each function symbol $f_i$ to a program in grammar $G_i$ such that these programs collectively satisfy the given relational specification $\Psi$, or \emph{null} if no such solution exists. As mentioned in Section~\ref{sec:intro}, our synthesis algorithm is based on the framework of counterexample-guided inductive synthesis (CEGIS)~\cite{cegis}. Specifically, given candidate programs $\mathcal{P}$, the verifier checks whether these programs satisfy the relational specification $\Psi$. If so, the CEGIS loop terminates with $\mathcal{P}$ as a solution (line 6). Otherwise, we obtain a \emph{relational counterexample} in the form of a \emph{ground formula} and use it to strengthen the current \emph{ground relational specification} $\Phi$ (line 7). In particular, a relational counterexample can be obtained by instantiating the quantified variables in the relational specification $\Psi$ with concrete inputs that violate $\Psi$. Then, in the \emph{inductive synthesis phase}, the algorithm finds candidates $\mathcal{P}$ that satisfy the new ground relational specification $\Phi$ (lines 8-11) and repeats the aforementioned process. Initially, the candidate programs $\mathcal{P}$ are chosen at random, and the ground relational specification $\Phi$ is \emph{true}. Since the verification procedure is not a contribution of this paper, the rest of this section focuses on the inductive synthesis phase which proceeds in three steps: \begin{enumerate}[leftmargin=*] \item {\bf \emph{Relaxation:}} Rather than directly taking $\Phi$ as the inductive specification, our algorithm first constructs a relaxation $\Phi'$ of $\Phi$ (line 8) by replacing each occurrence of function $f_i$ in $\Phi$ with a fresh symbol ({e.g., $f_1(f_1(1)) = 1$ is converted to $f_{1,1}(f_{1,2}(1)) = 1$}). This strategy relaxes the requirement that different occurrences of a function symbol must have the same implementation and allows us to construct our relational version space in a \emph{compositional} way. \vspace{1pt} \item {\bf \emph{HFTA construction:}} Given the relaxed specification $\Phi'$, the next step is to construct an HFTA whose language consists of exactly those programs that satisfy $\Phi'$ (line 9). Our HFTA construction follows the recursive decomposition of $\Phi'$ and allows us to compose the individual version spaces of each subterm in a way that the final HFTA is consistent with $\Phi'$. \vspace{1pt} \item {\bf \emph{Enforcing functional consistency:}} Since $\Phi'$ relaxes the original ground relational specification $\Phi$, the programs for different occurrences of the same function symbol in $\Phi$ that are accepted by the HFTA from Step (2) may be different. Thus, in order to guarantee functional consistency, our algorithm looks for programs that are accepted by the HFTA where all occurrences of the same function symbol correspond to the same program (lines 10-11). \end{enumerate} The following subsections discuss each of these three steps in more detail. \input{algo-cegis} \subsection{Specification Relaxation} \label{sec:relax} Given a ground relational specification $\Phi$ and a set of functions $\mathcal{F}$ to be synthesized, the {\sc Relax} procedure generates a relaxed specification $\Phi'$ as well as a mapping $\mathcal{M}$ from each function $f \in \mathcal{F}$ to a \emph{set} of fresh function symbols that represent different occurrences of $f$ in $\Phi$. As mentioned earlier, this relaxation strategy allows us to build a relational version space in a compositional way by composing the individual version spaces for each subterm. In particular, constructing an FTA for a function symbol $f \in \mathcal{F}$ requires knowledge about the possible values of its arguments, which can introduce circular dependencies if we have multiple occurrences of $f$ (e.g., $f(f(1))$). The relaxed specification, however, eliminates such circular dependencies and allows us to build a relational version space in a compositional way. We present our specification relaxation procedure, {\sc Relax}, in the form of inference rules shown in Fig.~\ref{fig:relax}. Specifically, Fig.~\ref{fig:relax} uses judgments of the form $\mathcal{F} \vdash \Phi \hookrightarrow (\Phi', \mathcal{M})$, meaning that ground specification $\Phi$ is transformed into $\Phi'$ and $\mathcal{M}$ gives the mapping between the function symbols used in $\Phi$ and $\Phi'$. In particular, a mapping $f \mapsto \{ f_1, \ldots, f_n\}$ in $\mathcal{M}$ indicates that function symbols $f_1, \ldots, f_n$ in $\Phi'$ all correspond to the same function $f$ in $\Phi$. It is worthwhile to point out that there are no inference rules for variables in Fig.~\ref{fig:relax} because specification $\Phi$ is a ground formula that does not contain any variables. Since the relaxation procedure is pretty straightforward, we do not explain the rules in Fig.~\ref{fig:relax} in detail. At a high level, we recursively transform all subterms and combine the results using a special union operator $\Cup$ defined as follows: \[\small (\mathcal{M}_1 \Cup \mathcal{M}_2)(f) = \left \{ \begin{array}{ll} \emptyset & {\rm if} f \not \in \emph{dom}(\mathcal{M}_1) \text{ and } f \not\in \emph{dom}(\mathcal{M}_2) \\ \mathcal{M}_i(f) & {\rm if} f \in \emph{dom}(\mathcal{M}_i) \text{ and } f \not \in \emph{dom}(\mathcal{M}_j) \\ \mathcal{M}_1(f) \cup \mathcal{M}_2(f) & {\rm otherwise} \\ \end{array} \right . \] In the remainder of this paper, we also use the notation $\mathcal{M}^{-1}$ to denote the inverse of $\mathcal{M}$. That is, $\mathcal{M}^{-1}$ maps each function symbol in the relaxed specification $\Phi'$ to a unique function symbol in the original specification $\Phi$ (i.e., $ f = \mathcal{M}^{-1}(f_i) \ \Leftrightarrow \ f_i \in \mathcal{M}(f) $). \input{fig-relax} \subsection{HFTA Construction} \label{sec:build} We now turn our attention to the relational version space learning algorithm using hierarchical finite tree automata. Given a relaxed relational specification $\Phi$, our algorithm builds an HFTA $\mathcal{H}$ that recognizes exactly those programs that satisfy $\Phi$. Specifically, each node in $\mathcal{H}$ corresponds to a subterm (or subformula) in $\Phi$. For instance, consider the subterm $f(g(1))$ where $f$ and $g$ are functions to be synthesized. In this case, the HFTA contains a node $v_1$ for $f(g(1))$ whose child $v_2$ represents the subterm $g(1)$. The inter-FTA transitions represent data flow from the children subterms to the parent term, and the FTA for each node $v$ is constructed according to the grammar of the top-level constructor of the subterm represented by $v$. Fig.~\ref{fig:build-hfta} describes HFTA construction in more detail using the judgment $\mathcal{G}, \mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$ where $\mathcal{G}$ describes the DSL for each function symbol~\footnote{If there is an interpreted function in the relational specification, we assume that a trivial grammar with a single production for that function is added to $\mathcal{G}$ by default.} and $\mathcal{M}$ is the mapping constructed in Section~\ref{sec:relax}. The meaning of this judgment is that, under CFGs $\mathcal{G}$ and mapping $\mathcal{M}$, HFTA $\mathcal{H}$ represents exactly those programs that are consistent with $\Phi$. In addition to the judgment $\mathcal{G}, \mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$, Fig.~\ref{fig:build-hfta} also uses an auxiliary judgment of the form $\mathcal{G}, \mathcal{M} \vdash \Phi \leadsto \mathcal{H}$ to construct the HFTAs for the subformulas and subterms in the specification $\Phi$. Let us now consider the HFTA construction rules from Fig.~\ref{fig:build-hfta} in more detail. The first rule, called Const, is the base case of our algorithm and builds an HFTA node for a constant $c$. Specifically, we introduce a fresh node $v_c$ and construct a trivial FTA $\mathcal{A}$ that accepts only constant $c$. The next rule, Func, shows how to build an HFTA for a function term $f(t_1, \ldots, t_m)$. In this rule, we first recursively build the HFTAs $\mathcal{H}_i = (V{_i}, \Omega_i, v_{r_i}, \Lambda_i)$ for each subterm $t_i$. Next, we use a procedure called {\sc BuildFTA} to build an FTA $\mathcal{A}$ for the function symbol $f$ that takes $m$ arguments $x_1, \dots, x_m$. The {\sc BuildFTA} procedure is shown in Fig.~\ref{fig:buildFTA} and is a slight adaptation of the example-guided FTA construction technique described in Section~\ref{sec:fta-construct}: {Instead of assigning the input example to each argument $x_i$, our \textsc{BuildFTA} procedure assigns possible values that $x_i$ may take based on the HFTA $\mathcal{H}_i$ for subterm $t_i$. Specifically, let $Q_{f_i}$ denote the final states of the FTA $\mathcal{A}_{f_i}$ associated with the root node of $\mathcal{H}_i$. If there is a final state $q_{s}^c $ in $Q_{f_i}$, this indicates that argument $x_i$ of $f$ may take value $c$; thus our {\sc BuildFTA} procedure adds a state $q_{x_i}^c$ in the FTA $\mathcal{A}$ for $f$ (see the Input rule in Fig.~\ref{fig:buildFTA}). The transitions and new states are built exactly as described in Section~\ref{sec:fta-construct} (see the Prod rule of Fig.~\ref{fig:buildFTA}), and we mark every state $q_{s_0}^c$ as a final state if $s_0$ is the start symbol in $f$'s grammar. To avoid getting an FTA of infinite size, we impose a predefined bound on how many times the Prod rule can be applied when building the FTA. Using this FTA $\mathcal{A}$ for $f$ and the HFTAs $\mathcal{H}_i$ for $f$'s subterms, we now construct a bigger HFTA $\mathcal{H}$ as follows: First, we introduce a new node $v_f$ for function $f$ and annotate it with $\mathcal{A}$. Second, we add appropriate inter-FTA transitions in $\Lambda$ to account for the binding of formal to actual parameters. Specifically, if $q_s^c$ is a final state of $\mathcal{A}_{f_i}$, we add a transition $q_s^c \rightarrow q_{x_i}^c$ to $\Lambda$. Observe that our HFTA is compositional in that there is a separate node for every subterm, and the inter-FTA transitions account for the flow of information from each subterm to its parent. \input{fig-build-hfta} \input{fig-buildFTA} The next rule, called Logical, shows how to build an HFTA for a subterm $\psi_1 \ op \ \psi_2$ where \emph{op} is either a relation constant (e.g., $=, \leq$) or a binary logical connective (e.g., $\land, \leftrightarrow$). As in the previous rule, we first recursively construct the HFTAs $\mathcal{H}_i = (V{_i},\Omega_i, v_{r_i}, \Lambda_i)$ for the subterms $\psi_1$ and $\psi_2$. Next, we introduce a fresh HFTA node $v_{\emph{op}}$ for the subterm ``$\psi_1 \ op \ \psi_2$'' and construct its corresponding FTA $\mathcal{A}$ using \emph{op}'s semantics. The inter-FTA transitions between $\mathcal{A}$ and the FTAs annotating the root nodes of $\mathcal{H}_1$ and $\mathcal{H}_2$ are constructed in the same way as in the previous rule.~\footnote{In fact, the Logical and Neg rules could be viewed as special cases of the Func rule where \emph{op} and $\neg$ have a trivial grammar with a single production. However, we separate these rules for clarity of presentation. } Since the Neg rule for negation is quite similar to the Logical rule, we elide its description in the interest of space. The last rule, Final, in Fig.~\ref{fig:build-hfta} simply assigns the final states of the constructed HFTA. Specifically, given specification $\Phi$, we first use the auxiliary judgment $\leadsto$ to construct the HFTA $\mathcal{H}$ for $\Phi$. However, since we want to accept only those programs that satisfy $\Phi$, we change the final states of the FTA that annotates the root node of $\mathcal{H}$ to be $\{q_{s_0}^\top\}$ rather than $\{q_{s_0}^\top, q_{s_0}^\bot \}$. \begin{figure}[!t] \centering \includegraphics[scale=0.57]{pics/hfta-plus-mult.pdf} \rule{0.75\textwidth}{0.5pt} \vspace{0.05in} \includegraphics[scale=0.57]{pics/htree1.pdf} \rule{0.75\textwidth}{0.5pt} \vspace{0.05in} \includegraphics[scale=0.57]{pics/htree2.pdf} \vspace{-5pt} \caption{HFTA $\mathcal{H} = (V, \Omega, v_=, \Lambda)$ for specification $f_1(2) = g_1(f_2(1))$. FTAs are represented as hypergraphs, where circles correspond to FTA states, double circles indicate final states, and labeled (hyper-)edges correspond to transitions. Inter-FTA transitions $\Lambda$ are represented as dashed lines. Hierarchical trees $\mathcal{T}_1$ and $\mathcal{T}_2$ are both accepted by~$\mathcal{H}$.} \label{fig:hfta-plus-mult} \vspace{-0.05in} \end{figure} \begin{example}\label{ex:hfta-plus-mult}{ Consider the relational specification $\Phi: f(2) = g(f(1))$, where the DSL for $f$ is $e \rightarrow x ~|~ e + 1$ and the DSL for $g$ is $t \rightarrow y ~|~ t \times 2$ (Here, $x,y$ denote the inputs of $f$ and $g$ respectively). We now explain HFTA construction for this specification. \begin{itemize}[leftmargin=*] \item First, the relaxation procedure transforms the specification $\Phi$ to a relaxed version $\Phi': f_1(2) = g_1(f_2(1))$ and produces a mapping $\mathcal{M} = \big\{ f \mapsto \set{f_1, f_2}, ~ g \mapsto \set{g_1} \big\}$ relating the symbols in $\Phi, \Phi'$. \item Fig.~\ref{fig:hfta-plus-mult} shows the HFTA $\mathcal{H}$ constructed for $\Phi'$ where we view $+1$ and $\times 2$ as unary functions for ease of illustration. Specifically, by the \textrm{Const} rule of Fig.~\ref{fig:build-hfta}, we build two nodes $v_1$ and $v_2$ that correspond to two FTAs that only accept constants 1 and 2, respectively. This sets up the initial state set of FTAs corresponding to $f_2$ and $f_1$, which results in two HFTA nodes $v_{f_2}$ and $v_{f_1}$ and their corresponding FTAs $\Omega(v_{f_2})$ and $\Omega(v_{f_1})$ constructed according to the \textrm{Func} rule. (Note that we build the FTAs using only two applications of the Prod rule.) Similarly, we build a node $v_{g_1}$ and FTA $\Omega(v_{g_1})$ by taking the final states of $\Omega(v_{f_2})$ as the initial state set of $\Omega(v_{g_1})$. Finally, the node $v_=$ along with its FTA $\Omega(v_=)$ is built by the \textrm{Logical} rule. \item Fig.~\ref{fig:hfta-plus-mult} also shows two hierarchical trees, $\mathcal{T}_1$ and $\mathcal{T}_2$, that are accepted by the HFTA $\mathcal{H}$ constructed above. However, note that only $\mathcal{T}_1$ corresponds to a valid solution to the original synthesis problem because (1) in $\mathcal{T}_1$ both $f_1$ and $f_2$ refer to the program $f = \lambda x. \ x$, and (2) in $\mathcal{T}_2$ these two occurrences (i.e., $f_1$ and $f_2$) of $f$ correspond to different programs. \end{itemize} } \end{example} \begin{theorem}{\bf (HFTA Soundness)}\label{thm:sound-relaxed} Suppose $\mathcal{G},\mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$ according to the rules from Fig~\ref{fig:build-hfta}, and let $f_1, \ldots, f_n$ be the function symbols used in the relaxed specification $\Phi$. Given a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ that is accepted by $\mathcal{H}$, we have: \begin{enumerate} \item $\Upsilon(v_{f_i})$ is a program that conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$ \item $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n})$ satisfy $\Phi$, i.e. $\Upsilon(v_{f_1}), \ldots, \Upsilon(v_{f_n}) \models \Phi$ \end{enumerate} \end{theorem} \begin{proof} See Appendix A. \end{proof} \begin{theorem}{\bf (HFTA Completeness)}\label{thm:complete-relaxed} Let $\Phi$ be a ground formula where every function symbol $f_1, \ldots, f_n$ occurs exactly once, and suppose we have $\mathcal{G},\mathcal{M} \vdash \Phi \twoheadrightarrow \mathcal{H}$. If there are implementations $P_i$ of $f_i$ such that $P_1, \ldots, P_n \models \Phi$, where $P_i$ conforms to grammar $\mathcal{G}(\mathcal{M}^{-1}(f_i))$, then there exists a hierarchical tree $\mathcal{T} = (V, \Upsilon, v_r, E)$ accepted by $\mathcal{H}$ such that $\Upsilon(v_{f_i}) = P_i$ for all $i \in [1,n]$. \end{theorem} \begin{proof} See Appendix A. \end{proof} \subsection{Enforcing Functional Consistency} \label{sec:find} As stated by Theorems~\ref{thm:sound-relaxed} and~\ref{thm:complete-relaxed}, the HFTA method discussed in the previous subsection gives a sound and complete synthesis procedure with respect to the \emph{relaxed} specification where each function symbol occurs exactly once. However, a hierarchical tree $\mathcal{T}$ that is accepted by the HFTA might still violate functional consistency by assigning different programs to the same function $f$ used in the original specification. In this section, we describe an algorithm for finding hierarchical trees that (a) are accepted by the HFTA and (b) conform to the functional consistency requirement. Algorithm~\ref{algo:find} describes our technique for finding accepting programs that obey functional consistency. Given an HFTA $\mathcal{H}$ and a mapping $\mathcal{M}$ from function symbols in the original specification to those in the relaxed specification, the {\sc FindProgs} procedure finds a mapping $\mathcal{P}_{\text{res}}$ from each function symbol $f_i$ in the domain of $\mathcal{M}$ to a program $p_i$ such that $p_1, \ldots, p_n$ are accepted by $\mathcal{H}$. Since the resulting mapping $\mathcal{P}_{\text{res}}$ maps each function symbol in the \emph{original} specification to a single program, the solution returned by {\sc FindProgs} is guaranteed to be a valid solution for the original relational synthesis problem. We now explain how Algorithm~\ref{algo:find} works in more detail. At a high level, {\sc FindProgs} is a recursive procedure that, in each invocation, updates the current mapping $\mathcal{P}$ by finding a program for a currently unassigned function symbol $f$. In particular, we say that a function symbol $f$ is unassigned if $\mathcal{P}$ maps $f$ to $\bot$. Initially, all function symbols are unassigned, but {\sc FindProgs} iteratively makes assignments to each function symbol in the domain of $\mathcal{P}$. Eventually, if all function symbols have been assigned, this means that $\mathcal{P}$ is a valid solution; thus, Algorithm~\ref{algo:find} returns $\mathcal{P}$ at line 3 if \textsf{ChooseUnassigned} yields \emph{null}. \input{algo-find} Given a function symbol $f$ that is currently unassigned, {\sc FindProgs} first chooses some occurrence $f_i$ of $f$ in the relaxed specification (i.e., $f_i \in \mathcal{M}(f)$) and finds a program $p$ for $f_i$. In particular, given a hierarchical tree $\mathcal{T}$ accepted by $\mathcal{H}$~\footnote{Section~\ref{sec:impl} describes a strategy for lazily enumerating hierarchical trees accepted by a given HFTA.}, we find the program $p$ that corresponds to $f_i$ in $\mathcal{T}$ (line 5). Now, since every occurrence of $f$ must correspond to the same program, we use a procedure called {\sc Propagate} that propagates $p$ to all other occurrences of $f$ (line 6). The procedure {\sc Propagate} is shown in Algorithm~\ref{algo:aux} and returns a modified HFTA $\mathcal{H}'$ that enforces that all occurrences of $f$ are consistent with $p$. (We will discuss the {\sc Propagate} procedure in more detail after finishing the discussion of {\sc FindProgs}). The loop in lines 5--10 of Algorithm~\ref{algo:find} performs backtracking search. In particular, if the language of $\mathcal{H}'$ becomes empty after the call to {\sc Propagate}, this means that $p$ is not a suitable implementation of $f$ --- i.e., given the current mapping $\mathcal{P}$, there is no extension of $\mathcal{P}$ where $p$ is assigned to $f$. Thus, the algorithm backtracks at line 7 by moving on to the next program for $f$. On the other hand, assuming that the call to {\sc Propagate} does not result in a contradiction (i.e., $\mathcal{L}(\mathcal{H}') \neq \emptyset$), we try to find an implementation of the remaining function symbols via the recursive call to {\sc FindProgs} at line 9. If the recursive call does not yield \emph{null}, we have found a set of programs that both satisfy the relational specification and obey functional consistency; thus, we return $\mathcal{P}_{\text{res}}$ at line 10. Otherwise, we again backtrack and look for a different implementation of $f$. \paragraph{\bf \emph{Propagate subroutine.}} We now discuss the {\sc Propagate} subroutine for enforcing that different occurrences of a function have the same implementation. Given an HFTA $\mathcal{H}$ and a function symbol $f$ with candidate implementation $p$, {\sc Propagate} returns a new HFTA $\mathcal{H}'$ such that the FTAs for all occurrences of $f$ only accept programs that have the same input-output behavior as $p$. In more detail, the loop in lines 3--7 of Algorithm~\ref{algo:aux} iterates over all HFTA nodes $v$ that correspond to some occurrence of $f$. Since we want to make sure that the FTA for $v$ only accepts those programs that have the same input-output behavior as $p$, we first compute all final states $Q_f'$ that can be reached by successful runs of the FTA on program $p$ (line 5) and change the final states of $\Omega(v)$ to $Q_f'$ (line 6). Since some of the inter-FTA transitions become spurious as a result of this modification, we also remove inter-FTA transitions $q \rightarrow q'$ where $q$ is no longer a final state of its corresponding FTA (line 7). These modifications ensure that the FTAs for different occurrences of $f$ \emph{only} accept programs that have the same behavior as $p$. \input{algo-aux} \begin{figure}[!t] \centering \includegraphics[scale=0.6]{pics/hfta-propagated2.pdf} \rule{0.75\textwidth}{0.5pt} \vspace{0.05in} \includegraphics[scale=0.6]{pics/hfta-propagated3.pdf} \caption{HFTAs after propagation. $\mathcal{H}_2$ is obtained by propagating $f_1 = f_2 = \lambda x. ~ x+1$ on HFTA $\mathcal{H}$ in Example~\ref{ex:hfta-plus-mult} (shown in Fig.~\ref{fig:hfta-plus-mult}). $\mathcal{H}_3$ is obtained by propagating $f_1 = f_2 = \lambda x. ~ x$ on HFTA $\mathcal{H}$. } \label{fig:hfta-propagated} \vspace{-0.1in} \end{figure} \begin{example} { We now illustrate how {\sc FindProgs} extracts programs from the HFTA in Example~\ref{ex:hfta-plus-mult}. Suppose we first choose occurrence $f_2$ of function $f$ and its implementation $f_2 = \lambda x. ~ x+1$. Since $f_1$ and $f_2$ must have the same implementation, the call to {\sc Propagate} results in the modified HFTA $\mathcal{H}_2$ shown in Fig.~\ref{fig:hfta-propagated}. However, observe that $\mathcal{L}(\mathcal{H}_2)$ is empty because the final state $q^\top_{s_0}$ in FTA $\Omega(v_=)$ is only reachable via state $q^2_{x_1}$, but $q^2_{x_1}$ no longer has incoming transitions. Thus, the algorithm backtracks to the only other implementation choice for $f_2$, namely $\lambda x. x$, and {\sc Propagate} now yields the HFTA $\mathcal{H}_3$ from Fig.~\ref{fig:hfta-propagated}. This time, $\mathcal{L}(\mathcal{H}_3)$ is not empty; hence, the algorithm moves on to function symbol $g$. Since the only hierarchical tree in $\mathcal{L}(\mathcal{H}_3)$ is $\mathcal{T}_1$ from Fig.~\ref{fig:hfta-plus-mult}, we are forced to choose the implementation $g = \lambda y. ~ y \times 2$. Thus, the \textsc{FindProgs} procedure successfully returns $\mathcal{P} = \set{f \mapsto \lambda x. x, ~ g \mapsto \lambda y. ~ y \times 2}$. } \end{example} \subsection{Properties of the Synthesis Algorithm} The following theorems state the soundness and completeness of the synthesis algorithm. \begin{theorem}{\bf (Soundness)}\label{thm:sound} Assuming the soundness of the \textsf{Verify} procedure invoked by Algorithm~\ref{algo:cegis}, then the set of programs $\mathcal{P}$ returned by {\sc Synthesize} are guaranteed to satisfy the relational specification $\Psi$, and program $P_i$ of function $f_i$ conforms to grammar $\mathcal{G}(f_i)$. \end{theorem} \begin{proof} See Appendix A. \end{proof} \begin{theorem}{\bf (Completeness)}\label{thm:complete} If there is a set of programs that (a) satisfy $\Psi$ and (b) can be implemented using the DSLs given by $\mathcal{G}$, then Algorithm~\ref{algo:cegis} will eventually terminate with programs $\mathcal{P}$ satisfying $\Psi$. \end{theorem} \begin{proof} See Appendix A. \end{proof}

d4d40638c11b8bc50f4d32457f1ac9b1f1d535a9

\section{Introduction}\label{introduction} Some time series are stochastic (or random) in the sense that they do not contain underlying information other than noise. The analysis of large databases requires automatic and efficient classifiers to provide the identification of genuine features. This is crucial to reduce the number of misclassifications, to narrow the boundaries between classes, to provide better training sets, as well as to diminish the total processing time \citep[][]{Eyer-2006}. Large volumes of data containing potentially interesting scientific results are left unexplored or have their analysis delayed due to the current limited inventory of tools which are unable to produce clean samples, despite big efforts having been undertaken. In fact, we risk underusing a large part of these data. In order to improve the efficiency of variability indices, we propose discriminating variable sources in correlated and non-correlated data. The correlated data have several measurements close in time, from which accurate correlated indices are computed. On the other hand, the non correlated data are those sources having too few measurements close in time and so they must be analysed using statistical parameters only. The use of correlated and non-correlated indices \citep[see Sect. 4.3 in][]{FerreiraLopes-2016papI}, produces a substantially smaller number of time series which have to be further analysed. However, the resulting selection is still three or more times larger than just the well-defined signals, according to \citet[][]{FerreiraLopes-2016papI,FerreiraLopes-2017papII}. This means that the set of preliminary selection criteria is unable to produce samples comprised only of variable stars, and so, it would be desirable that the following steps of signal searching methods would provide reliable identifications and accurate estimates of periods (frequencies) and amplitudes, even in cases where the preliminary analysis failed to give a confident indicator that the signal was truly variable and not just a noisy time series. Indeed, $\sim75\%$ of parameters used to characterize light curves are derived from the folded light curve using the variability periods \citep[][]{Richards-2011}. This led to a $\sim 11\%$ misclassification rate for non-eclipsing variable stars, for instance, \cite{Dubath-2011}. Nowadays, reliable samples, i.e. samples composed only of variable stars, are increasingly becoming more important than complete samples, i.e samples having all variable stars but also having a large number of misclassified non-variable stars, since visual inspection of all sources is unfeasible. Therefore, an approach that allows us to get unbiased samples having correct periods is mandatory to return quicker scientific results. Therefore, as a continuation of the ``New Insights into Time Series Analysis'' project, the frequency finding methods are reviewed and improved. The periodic signals finding methods can be separated into three main groups if we consider how each component of the figure of merit in the frequency grid is computed: $M_S$ - each epoch provides a single term; $M_P$ - each term is computed using a pair of epochs; $M_B$ - each term is computed by binning the phase diagram. The Lomb Scargle and its generalization (LS - \citealt[][]{Lomb-1976,Scargle-1982} and GLS - \citealt[][]{Zechmeister-2009}) belong to the $M_S$ group. Each epoch is regarded as a single power spectrum term and the periodogram is equivalent to a least squares fit of the folded data at each frequency by a sine wave. Indeed, Fourier methods and their branches are the simplest methods belong to the $M_S$ group. On the other hand, the string length method \citep[STR - ][]{Dworetsky-1983} and the analysis of variance \citep[AOV and AOVMHW - ][]{Schwarzenberg-Czerny-1989, Schwarzenberg-Czerny-1996} belong to the $M_P$ group. However, they follow different approaches since the STR power spectrum is computed using pairs of epochs in the phase diagram while AOV pairs epochs in time. Phase dispersion minimization \citep[PDM and PDM2 - ][]{Stellingwerf-1978, Stellingwerf-2011}, conditional entropy \citep[CE - ][]{Graham-2013}, supersmoother \citep[SS - ][]{Reimann-1994}, and correntropy kernel periodogram \citep[CKP - ][]{Huijse-2012} belong to the $M_B$ group, where the power spectrum is computed by binning the phase diagram. Wavelet analysis also has been used to study time series \citep[e.g.][]{Foster-1996,Bravo-2014}. However it is more suitable to study the evolution of a signal over time and it requires continuous observation. Currently, these are the main frequency finding methods but there are many others \citep[e.g.][]{Huijse-2011,Kato-2012}. The efficiency of the frequency finding methods has been tested in the last few years \citep[e.g.][]{Heck-1985,Swingler-1989,Schwarzenberg-1999, Distefano-2012, Graham-2013}. Usually, the authors analyze the sensitivity as well as the fraction of true periods recovered within a defined accuracy limit. Indeed, research using real data, including for instance irregular sampling, gaps, outliers, and errors, may provide more reliable results. Currently, the most complete of these studies was performed by \citet[][]{Graham-2013}. The authors analysed 11 different methods using light curves of $78$ variable star types. The conditional entropy-based algorithm is the most optimal in terms of completeness and speed according to the authors. However, most frequency finding methods have a selection effect for the identification of weak periodic signals \citet[][]{deJager-1989,Schwarzenberg-Czerny-1999}. Therefore a combination of all methods could potentially increase the recover rate close to $100\%$ according to \citet[][]{Dubath-2011}. However, which method leads to the correct period for a specific light curve within an automated strategy is an open question. Moreover, the main frequencies computed by different methods can be dissimilar and so two questions must be answered to determine the best way to analyse a time series, i.e. how many frequency finding methods should be combined and how to work out which period should be chosen when two or more methods provide different results? The frequency finding methods adopt, as the true frequency ($\mathrm{f_{true}}$), one that defines the main periodic variation displayed by the time series, based on a minimum or maximum of the quantity being tested. However, the main frequency can be a harmonic of $\mathrm{f_{true}}$ or related to an instrumental or spurious variation. It means that only using the period finding method is not enough to set a reliable period and so additional analyses are required. For example, the harmonic fits can be used to set models and, using the $\chi^2$ distribution, establish $\mathrm{f_{true}}$ and its reliability \citep[][]{FerreiraLopes-2015wfcam}. However, what $\chi^2$ threshold, below which a time series may be considered to have a reliable signal, and if the $\chi^2$ alone is enough to do that are open questions. Theoretically, any signal having an amplitude greater than the noise could be detected using a suitable method. The false alarm probability (\textit{FAP}) has been used to determine the typical power spectrum values of the noise and to discard variability due to noise alone. However, sources lacking a periodic signal, such as aperiodic variations and pulses, will also be discarded using this approach. All the period finding methods that depend on the phase diagram are unfit to detect these signals because no frequency will return a smooth phase diagram. Therefore, classifying a time series as noisy requires an investigation of all signal types. On the other hand, correlated noise, seasonal variations, the cadence or the phase coverage can lead to power values above the FAP indicating an applicability limit to using this approach to determine the reliability of selections. Indeed, a large number of Monte-Carlo simulations are usually performed to determine the FAP and hence the running time required should also be taken into account. Then our required list of improvements towards an efficient automation of the variability analysis should include: how to use the current period finding methods to determine the reliability of signals? How to discriminate aperiodic from stochastic variations? Is it possible to study all variation types using a single approach or are different strategies required for different purposes? How to provide a standard cutoff to determine reliable signals? Currently, any frequency finding method is able to compute $\mathrm{f_{true}}$ using a single computation. Therefore, the determination of $\mathrm{f_{true}}$ is performed after computing several times a function that is susceptible to the smoothness of the phase diagrams depending on the method that is used. The phase values are given by; \begin{equation} \mathrm{\phi_{i} = t_{i}\times f_{test} - {\rm \mathrm{\textit{G}}} \left[ t_{i}\times \mathrm{f_{test}} \right]}, \label{eqphi} \end{equation} \noindent where $t_{i}$ is the time, $\mathrm{f_{test}}$ is a test frequency, and the function $\mathrm{\textit{G}}$ returns the integer part of $ t_{i}\times \mathrm{f_{test}}$. This equation provides an interval of values ranging $0 \leq \Phi \leq 1$ where $\mathrm{f_{true}}$ is the frequency which returns the smoothest phase diagram. The minimum ($\mathrm{f_{min}}$) and maximum ($\mathrm{f_{max}}$) frequency as well as the resolution ($\Delta \mathrm{f}$) or number of frequencies ($\mathrm{N_{f}}$) are required as inputs to search periods in all unevenly spaced time series for all frequency finding methods. The $\mathrm{f_{min}}$ is usually defined as $2/T_{tot}$, where $T_{tot}$ is the total timespan of the observations. This definition is commonly used as a requirement to enclose at least two variability cycles in the time series. However, variations having fewer than two cycles can be considered with caution when biases have been identified and removed \citep[e.g.][]{DeMedeiros-2013,FerreiraLopes-2015mgiant}. On the other hand, $\mathrm{f_{max}}$ will be linked with the time interval between the observations $\delta t$. The Nyquist frequency ($\mathrm{f_{max}} = 0.5/\delta t$) must be assumed for evenly-spaced time series since this constitutes an upper limit to the frequency range over which a periodogram can be uniquely calculated. Otherwise, for irregularly sampled cases, the frequency limit becomes dominated by the exposure time \citep{Eyer-1999}. The frequency sampling strategy is crucial to determine $\mathrm{f_{true}}$ using any frequency finding algorithm. A small variation on $\mathrm{f_{true}}$ provides a big variation in the phase diagram mainly when $f_{true}\times\,T_{tot} >> 1$ (see Sect.~\ref{sec_freqsamp}). It means that $\mathrm{f_{true}}$ can be missed if the periodogram is not computed for a sufficiently large number of test frequencies. A reasonable criterion (see previous paragraph) has been used to determine $\mathrm{f_{min}}$ while an empirical criterion has been applied to set $\mathrm{f_{max}}$ and $\Delta \mathrm{f}$. For instance, $\mathrm{f_{min}} = 0$, $\mathrm{f_{max}} = 10d^{-1}$, and $\Delta \mathrm{f} = 0.1/T_{tot}$ were adopted by \citet[][]{Debosscher-2007} and \citet[][]{Richards-2012}. In this case, the authors assumed an empirical cutoff on the maximum frequency below which any reliable frequency could be found: the frequency finding method is able detect all reliable frequencies in a range of $\mathrm{f} \pm \Delta \mathrm{f}/2$. On the other hand, \citet[][]{Schwarzenberg-Czerny-1996} assumes $\mathrm{f_{min}} = 0$, $\mathrm{f_{max}} = 1/2\tau_{med}$, and an optimal grid $\Delta \mathrm{f} = 1/(A\times T_{tot})$, where $\tau_{med}$ is the median difference between successively ordered observation times and $A$ is a factor, typically ranging $10-15$, which takes into account oversampling and binning or the number of harmonics used in the Fourier fit. \citet[][]{Graham-2013} tested $\Delta \mathrm{f}$ values of $0.0001$, $0.001$, $0.01$, and the optimal grid over a frequency range from $\mathrm{f_{min}} = 0$ to $\mathrm{f_{max}} = 20$ for standard frequency finding methods; LS, GLS, AOVs, PDMs, STR, FC, CE, SS, and CKP methods. The data test used by the authors has a number of observations ranging from $105$ to $966$ and a total baseline ranging from $2182$ to $2721$ days. The performance found for $\Delta \mathrm{f} = 0.0001$ and the optimal $\Delta \mathrm{f}$ (the median optimal $\Delta \mathrm{f}$ is $2.5 \times 10^{-5}$) is quite similar for all methods analysed according to the authors. Indeed these results can only be used as a guide for samples that mimic those tested by the authors since the samples tested do not cover all possible intervals of measurements and baselines. Therefore, what is the optimal resolution able to detect all periodic variations and how much finer grain is necessary to get a more accurate period estimation, if $\mathrm{f_{true}}$ is found since an initial value can be found with a coarser grain resolution, are open questions. The majority of frequency finding methods were designed for single time series. Such methods are in accordance with past surveys since they were usually from observations in a single photometric waveband (e.g. VVV - \citealt[][]{Minniti-2010}, LINEAR - \citealt[][]{Sesar-2011}, CoRoT - \citealt[][]{Baglin-2007}, and Kepler - \citealt[][]{Borucki-2010}). However, in the last few years there are multi-wavelength surveys like Gaia \citep[][]{BailerJones-2013}, where the main catalogue is multi-epoch using a wide G filter, but it also contains colour information from simultaneous multi-epoch low resolution spectra. Period finding could be done on G-band and forthcoming surveys like LSST \citep[][]{Ivezic-2008} require multi-wavelength frequency finding methods in order to optimize the period searches. Usually each waveband is analysed separately and posteriorly the results are combined \citep[e.g.][]{Oluseyi-2012,FerreiraLopes-2015wfcam}. However, the combination of different datasets allows us to increase the number of measurements that are extremely important to signal detection. \citet[][]{Suveges-2012} used principal component analysis to extract the best period from analysis of the Welch-Stetson variability index \citep[][]{Welch-1993}. However, the method requires observations taken simultaneously. \citet[][]{VanderPlas-2015} introduce a multi-band periodogram by extending the Lomb-Scargle approach. For that, the authors modeled each waveband as an arbitrary truncated Fourier series using the Tikhonov regularization in order to provide a common model at all wavebands. Such methods and new insights into multi-wavelength frequency finding methods are required to take full advantage of the multi-wavelengths observations. The discussion above provokes questions that must be addressed in the challenge to analyze large amount of photometric data automatically. Some of these questions are addressed in the current paper (III) and the forthcoming papers of this series will address the remaining questions. Sections \ref{sec_freqsamp} and \ref{sec_freqerror} assess the frequency sampling and frequency uncertainties. Sect. \ref{sec_error} establishes an approach to compute period and amplitude variations. In Sect. \ref{sec_results} we show our results on estimating frequency resolution and uncertainties, and discuss them. We address our final remarks in Sect. \ref{sec_conclusion}. \section{Frequency sampling}\label{sec_freqsamp} Consider a periodic signal modeled by function $\mathrm{F}$ having frequency $\mathrm{f_{true}}$ (being a real, positive constant) where $\mathrm{F = \left[F(\,t_{1}\,),F(\,t_{2}\,),\cdots,F(\,t_{n}\,)\right]}$. From which $\mathrm{F(\,t_{i}\,) = F(\,t_{i} + n_{c}/\mathrm{f_{true}}\,)}$ where $n_{c}$ (number of cycles) is a positive integer ranging from zero to $G[T_{tot}\times f_{true}]$. This relationship is also true for phase values, i.e. $\mathrm{\phi_{i}(\,t_{i}\,) = \phi(\,t_{i} + n_{c}/\mathrm{f_{true}}\,) = \phi_{j}(\,t_{j}\,)}$ and therefore $\left| \mathrm{\phi_{j}\,-\,\phi_{i}}\right| = 0$. The phase difference between them for a frequency given by $\mathrm{f = f_{true} + \delta f}$ is written as, \begin{eqnarray} \scriptsize \left| \mathrm{\phi_{j}\,-\,\phi_{i}}\right| &=& \Bigl| \mathrm{t_{j}\times\left(f_{true}+\delta f\right) - G\bigl[ t_{j}\times \left(f_{true}+\delta f\right) \bigr]} \nonumber \\ & & \mathrm{-t_{i}\times\left(f_{true}+\delta f\right) +G\left[ t_{i}\times \left(f_{true}+\delta f\right) \right]} \Bigr| \nonumber \\ \left| \mathrm{\phi_{j}\,-\,\phi_{i}}\right| &=& \Bigl|\mathrm{\left(t_{j}-t_{i}\right)\times\delta f + \Bigl\{ t_{j}\times f_{true} - G \left[ t_{j}\times \left(f_{true}+\delta f\right) \right] \Bigr\} } \nonumber \\ & & \mathrm{-\Bigl\{ t_{i}\times f_{true} - G \left[ t_{i}\times \left(f_{true}+\delta f\right) \right] \Bigr\} \Bigr|}. \label{eq_proof01} \end{eqnarray} Having $\mathrm{\delta f << f}$ implies that $\mathrm{G \left[ t\times \left(f+\delta f\right) \right] = G \left[ t\times f \right]}$. Indeed, this is reasonable since the frequency sampling is usually set as $\mathrm{f_{n} = f_{min}+n\times\delta f}$. For instance, $n=100$ implies that there is a frequency at least hundred times greater than $\mathrm{\delta f}$. Considering this limit, the two last terms (in curly brackets) of Eq. \ref{eq_proof01} are cancelled and so, \begin{equation} \mathrm{\bigr|\phi_{j}\,-\,\phi_{i}\bigr| \simeq \bigr|t_{j}-t_{i}\bigr|\times\delta f} \implies \mathrm{\delta f \simeq \frac{\delta_{\phi_{j,i}}}{\bigr|t_{j}-t_{i}\bigr|}} \label{eq_deltaf} \end{equation} The maximum variation on $\mathrm{\delta_\phi}$ is found for $\left|t_{j}-t_{i}\right| = T_{tot}$, i.e. from the comparison among the measurements at the ends of the time series. Indeed, Eq. \ref{eq_deltaf} was found only assuming that $\mathrm{\delta f << f}$ and hence this expression can be used as an analytical definition of the frequency rate limit, where the number of frequencies is given by: \begin{eqnarray} N_{f} &=& \frac{f_{max}-f_{min} }{ \delta f } = \frac{\left(f_{max}-f_{min}\right)\times T_{tot}}{\delta_\phi } \nonumber \\ N_{f} &\simeq& \frac{f_{max}\times T_{tot}}{\delta_\phi}, \label{eq_nfmax} \end{eqnarray} \noindent where $\mathrm{f_{max} >> f_{min}}$ was assumed when deriving the expression. This expression enables us to determine $N_{f}$ from the phase shift $\delta_\phi$ since $T_{tot}$ is a feature of a time series. On the other hand, $\mathrm{f_{max}}$ can be assumed to be the same for different time series, in the same set of observations, since the upper limit of frequencies, for those time series having evenly spaced data, is given by the Nyquist frequency \citep[e.g.][]{Eyer-1999}. Therefore, the frequency search will be performed with the same resolution in the phase diagram if we assume the same $\delta_\phi$ for different time series. Moreover, it facilitates a strict comparison of frequency finding searches performed by different surveys. Equation \ref{eq_nfmax} was defined only by considering the phase diagram. Therefore, this relation is general and it can be used as an accurate determination of the frequency grid required to find any signal. Indeed, a similar equation has been used to estimate the frequency grid \citep[e.g.][]{Schwarzenberg-Czerny-1996,Debosscher-2007,Richards-2012, VanderPlas-2015,VanderPlas-2018} where the $1/\delta_\phi$ is called the oversampling factor. However, no meaning had been provided for the oversampling factor so far. Values ranging from 5 to 15 have being adopted empirically only to ensure that the frequency grid is sufficient to sample each periodogram peak. The proper meaning of the oversampling factor is defined in Eq. \ref{eq_nfmax} from which a suitable frequency grid for any kind of signal can be determined. \begin{figure} \centering \includegraphics[width=0.49\textwidth,height=0.4\textwidth]{fig_fwhm.eps} \caption{The logarithmic of FWHM as a function of logarithmic of $T_{tot}/P$ for the Generalized Lomb-Scargle (LSG - upper left panel) and for the string length (STR upper right panel) methods using four (ABCD) sinusoidal signal having models of variation and noise, shown in Fig~\ref{fig_discerror} and described in Sect. \ref{sec_error}. The models are set by colors and the solid black line marks the linear fits for LSG and STR methods. The percent relative errors ($\eta$) for both results are shown in the lower panels.} \label{fig_fhwm} \end{figure} \section{Frequency uncertainties}\label{sec_freqerror} Frequency uncertainties were analytically defined from Fourier analysis \citep[e.g.][]{Kovacs-1981,Gregory-2001,Stecchini-2017}, \begin{equation} \sigma_{f} \propto \frac{1}{T_{tot}}\sqrt{\frac{1}{n\times\Sigma}} \label{eq_olderror} \end{equation} \noindent where $n$ is the number of measurements, $\Sigma$ is the signal-to-noise ratio, and $T_{tot}$ is the total baseline of the observations. The uncertainty provided by a well defined periodic signal will be limited by the exposure time and hence Eq. \ref{eq_olderror} is not a suitable definition since it assumes an infinite accuracy. On the other hand, phenomena which result in small variations on the period can be mistaken for an increased uncertainty. Indeed, the uncertainties computed using a time series are given by the sum of intrinsic plus instrumental limitations. The uncertainties related with instrumental limitations can be estimated using a noise model \citep[e.g.][]{FerreiraLopes-2017papII} and by including this we can thus estimate the intrinsic variation. Some inconsistencies are found when the frequency uncertainty ($\sigma_{f}$ - see Eq.~\ref{eq_olderror}) estimation is related with $T_{tot}$, $n$, and $\Sigma$. For instance, a signal having an intrinsic variation in the frequency ($\sigma_{f} \ne 0$), such as light curves of rotational variables, may return a similar estimation of the uncertainty for time series having one hundred or one thousand measurements. On the other hand, for periodic signals a reduction in the dispersion about the model naturally occurs for a longer baseline and its accuracy is limited by instrumental properties instead for large $T_{tot}$ or $\Sigma$. Indeed, the power spectrum of $\mathrm{f_{true}}$ tends to a delta function with increasing $T_{tot}$ while increasing $n$ and $\Sigma$ improves the signal reliability since the probability of a signal being detected increases when the noise is reduced. These properties characterize the signal but they are not directly related with the period variations. The \textit{CoRoT} and \textit{Kepler} databases have in common a large number of measurements and wide coverage time that provide unreliable uncertainty estimations using Eq. \ref{eq_olderror}. Therefore, new approaches have been used to compute frequency uncertainties for semi-periodic signals. For the \textit{Kepler} time series, \citet[][]{Reinhold-2013} compute the frequency uncertainty by fitting a parabola to the peak of the Lomb-Scargle power spectrum (Reinhold, private communication). On the other hand, for the \textit{CoRoT} time series, \citet[][]{DeMedeiros-2013} used a similar equation to that proposed by \citet[][]{Lamm-2004} to estimate the period uncertainty, given by \begin{equation} \delta P = \frac{\delta\nu \times P^2}{2}, \label{eq_eperlamm} \nonumber \end{equation} \noindent where $\delta\nu$ is about $1/T_{tot}$ for non-uniform sampling according to the authors. \citet[][]{FerreiraLopes-2015mgiant, FerreiraLopes-2015cycles} also used the \textit{CoRoT} time series to study non-radial pulsation and stellar activity where the period uncertainties were estimated as the \textit{FWHM} ($\delta P_{(\rm FWHM)}^{(\rm STR)}$) of the String Length power spectrum \citep[][]{Dworetsky-1983}. In particular, the amplitudes and periods vary for light curves of rotational variables that have differential rotation and spot evolution \citep[e.g.][]{Lanza-2014, Reinhold-2015,DasChagas-2016}. The analytical expression given by Eq.~\ref{eq_olderror} or the analysis of the power spectrum are half-way to computing period variations in order to get new clues about physical phenomena that account for such variations. Figure \ref{fig_fhwm} shows $\delta P_{(\rm FWHM)}$ as a function of the number of cycles ($N_{(cycles)} = T_{tot}/P$) for the Generalized Lomb-Scargle \citep[LSG-][]{Zechmeister-2009} and for the string length \citep[STR -][]{Dworetsky-1983} methods using the sinusoidal signal described in Sect. \ref{sec_error}. The best fit models found for the LSG and STR methods are given by, \begin{equation} Log\left(\delta P_{(\rm FWHM)}^{(\rm LSG)}\right) = -1.00 -0.44 \times Log\left(\frac{T_{tot}}{P}\right) \label{eq_eqfwhm_lsg} \end{equation} \noindent and \begin{equation} Log\left(\delta P_{(\rm FWHM)}^{(\rm STR)}\right) = -0.97 -0.54 \times Log\left(\frac{T_{tot}}{P}\right). \label{eq_eqfwhm_str} \end{equation} We create 4 different sinusoidal models which are a single sinusoid (A), sinusoid plus amplitude variation (B), sinusoid plus period variation (C), and a sinusoid plus amplitude and period variations (D), see Sect. \ref{sec_error} and Fig.~\ref{fig_discerror} for more details. However, the results are quite similar for all ABCD models, i.e. the $\delta P_{(\rm FWHM)}$ estimation is mainly defined by the $N_{(cycles)}$ instead of the time series properties. Indeed, it is highlighted for $N_{(cycles)} > 10$ where the percent relative error (i.e. $\eta = 100\times(theoretical-experimental)/\mid theoretical \mid$) is always smaller than $4\%$, see lower panels of Fig.~\ref{fig_fhwm}. Eq. \ref{eq_eqfwhm_lsg} is slightly different from Eq. \ref{eq_eqfwhm_str} (see the solid lines in the two upper panels of Fig. \ref{fig_fhwm}) but the LSG method shows smaller relative errors ($\eta$). The approach using the FWHM of the power spectrum and any frequency finding method does not provide a trustworthy estimation of the period variation (for more detail see Sect. \ref{sec_testreal}). To summarise, the uncertainty computed using the power spectrum gives us a period interval about the variability period that leads to similar phase diagrams. Indeed, the main period and its uncertainty can vary for different period finding methods since the susceptibility to measuring phase diagram variations is not the same \citep[e.g.][]{Eyer-2006,Graham-2013}. Moreover, the main period is assumed to be one that leads to the highest peak of periodogram that, a priori, gives the smoothest phase diagram and also the smallest residuals ($\sigma_r$), i.e. the standard deviation of observed data minus model (or predicted value). However, these assumptions have not been analysed so far, but this empirical criteria has been used all the time. In the sections below these questions are addressed. \begin{figure} \centering \includegraphics[width=0.49\textwidth,height=0.9\textwidth]{fig_propsig.eps} \caption{Sinusoidal light curves, with a fixed period and amplitude (A); with a varying period ($\delta_{\phi} = 0.02$) and constant amplitude (B); varying amplitude ($\delta_{y} = 0.05$) and constant period (C); and both amplitude and period variations (D). The black line shows the model while the red crosses show the measurements. The dark and light grey regions are expanded in the right panels.} \label{fig_discerror} \end{figure} \section{Frequency and Amplitude Variations}\label{sec_error} Eq.~\ref{eq_olderror} is an analytical expression that saves computational time. However, computational methods can be used to perform non-analytical approaches to compute frequency and amplitude variations in order to accurately choose the main variability period as well as give additional information about the phenomena observed. Indeed, any time series can be modeled using Fourier decomposition ($Y(\phi)$ - see Fig. \ref{fig_lctype}). In order to determine the suitable measurements or light curve regions to compute these variations consider the following example: \begin{equation} t^{(o)} = t^{(m)} + \delta_{t} \qquad $and$ \qquad y^{(o)} = \sin\left(2 \times \pi \times t^{(m)} \times \mathrm{f_{true}} \right)+\delta_{y} \label{eq_linmodel} \end{equation} \noindent where $(o)$ and $(m)$ are the mean observed and modeled values, respectively. Indeed, $t^{(o)} = t^{(m)}$ if $\delta_{\phi} = 0$ and $y^{(o)} = y^{(m)}$ if $\delta_{y} = 0$. Four cases are displayed in the Fig. \ref{fig_discerror}; (A) constant period and amplitude, (B) period variation for constant amplitude, (C) amplitude variation for constant period, and (D) period and amplitude variation. It is easier to understand these cases if a linear fit is calculated in the light and dark grey regions in Fig. \ref{fig_discerror}, given by \begin{equation} y^{(o)} = \alpha \times \left( \phi + \delta_{\phi}^{(e)} \right) + \left( \beta + \delta_{y}^{(e)} \right) \label{eq_line} \end{equation} \noindent where $(e)$ means expected value and ($\alpha,\beta$) becomes ($\alpha_{y},\beta_{y}$) or ($\alpha_{\phi},\beta_{\phi}$) to indicate the region used to estimate the amplitude or period variations, respectively. Moreover, $y^{(o)} = y^{(m)}$ if $\delta_{\phi}^{(e)} = 0$ and $\delta_{y}^{(e)} = 0$. The linear fit only takes into account the first order contribution. However, this allows us to determine a simple analytical equation to analyse the contributions of $\delta_{\phi}^{(e)}$ on $\delta_{y}^{(e)}$. For real data, the fitting function, that may have a more complex shape, can be used. The main features derived from Fig. \ref{fig_discerror} are summarized below: \begin{itemize} \item[\textit{i} -- ] The computed amplitude variation ($\delta_{y}^{(c)}$) is defined as the difference between the observed ($o$) and modeled ($m$) amplitude at the same phase, i.e. $[\phi^{(o)},y^{(o)}]$ implied from $y^{(m)} = Y(\phi^{(o)})$, is given by \begin{equation} \delta_{y}^{(c)} = y^{(o)}-y^{(m)} = \alpha_{y} \times \delta_{\phi}^{(e)} + \delta_{y}^{(e)}. \label{eq_uncamp} \end{equation} $\delta_{y}^{(c)}$ can be different from zero if $\delta_{y}^{(e)} = 0$ according to Eq. \ref{eq_uncamp}, i.e. a phase variation $\delta_{\phi}^{(e)}$ can appear as an amplitude variation if $\alpha_{y} \ne 0$. On the other hand, the ideal case will be found when $\delta_{y}^{(c)} = \delta_{y}^{(e)}$ that implies that $\alpha_{y} \times \delta_{\phi}^{(e)} \simeq 0$. For the cases where $\delta_{y}^{(e)} \ne 0$ the ratio of computed to expected values is written as \begin{equation} R^{(\delta_{y})} = \left| \frac{\delta_{y}^{(c)}}{\delta_{y}^{(e)}} \right| = \left|\alpha_{y} \times \frac{\delta_{\phi}^{(e)}}{\delta_{y}^{(e)}} + 1\right| \label{eq_uncamprel} \end{equation} \noindent that indicates whether the computed value is overestimated ($R^{(\delta_{y})} > 1$), equal ($R^{(\delta_{y})} = 1$), or underestimated ($R^{(\delta_{y})} < 1$). Therefore, the estimation of $\delta_{y}^{(e)}$ will be improved if the $\alpha_{y} \times \delta_{\phi}^{(e)} << \delta_{y}^{(e)}$. $\delta_{\phi}^{(e)}$ is a time series property that can not be modified. However those light curve regions having $\alpha_{y} \simeq 0$ provide a better estimation of the amplitude variation. The light-grey region of Fig. \ref{fig_discerror} indicates a suitable region to measure amplitude variation since this contains the smallest $\alpha_{y}$ values. \item[\textit{ii} -- ] The computed phase variation ($\delta_{\phi}^{(c)}$) is given by the difference between the observed and modeled value for the same amplitude, i.e. for each pair of observed measurements $[\phi^{(o)},y^{(o)}]$ implies that $y^{(o)} = Y(\phi^{(m)})$, which can be written as \begin{equation} \delta_{\phi}^{(c)} = \phi^{(o)}-\phi^{(m)} = \frac{\delta_{y}^{(e)}}{\alpha_{\phi}} + \delta_{\phi}^{(e)}. \label{eq_unper} \end{equation} \noindent where $\delta_{\phi}^{(c)}$ will return values different to zero even if $\delta_{\phi}^{(e)}=0$ in the same fashion as the amplitude variation (see item \textit{i}). Indeed, the amplitude and phase variations are coupled equations, i.e. $\delta_{y}^{(c)}$ depends on $\delta_{\phi}^{(e)}$ while $\delta_{\phi}^{(c)}$ depends on $\delta_{y}^{(e)}$ (see Eqs. \ref{eq_uncamp} and \ref{eq_unper}). Moreover, not all observed measurements can be used to compute $\delta_{\phi}$ since those having values bigger or smaller than the maximum and minimum $Y$ values cannot be written as $y^{(o)} = Y(\phi^{(e)})$. Therefore, only the observed measurements having values between the minimum ($Y_{min}$) and maximum ($Y_{max}$) model values can be used to estimate $\delta_{\phi}^{(c)}$, i.e. for all $y^{(o)}$ since $Y_{min} < y^{(o)} < Y_{max}$. The number of measurements used to compute $\delta_{\phi}^{(c)}$ will depend on the signal type (see Fig. \ref{fig_lctype} first panels). However these measurements do not provide a good information of time variation about the model. Using the ratio of computed and expected values is a suitable way to examine agreement between them, given by \begin{equation} R^{(\delta_{\phi})} = \left| \frac{\delta_{\phi}^{(c)}}{\delta_{\phi}^{(e)}} \right| = \left| \alpha_{\phi}^{-1} \times \left( \frac{\delta_{\phi}^{(e)}}{\delta_{y}^{(e)}} \right)^{-1} + 1 \right|. \label{eq_unperrel} \end{equation} The opposite result of $R^{(\delta_{y})}$ is found since the dispersion of $\delta_{\phi}$ values are proportional to the inverse of the angular coefficient and to the inverse of the relation $\delta_{\phi}^{(e)}/\delta_{y}^{(e)}$. This means that the weight of $\delta_{y}^{(e)}$ on $\delta_{\phi}^{(c)}$ and $\delta_{\phi}^{(e)}$ on $\delta_{y}^{(c)}$ will be the same only if $\delta_{\phi}^{(e)}/\delta_{y}^{(e)} = 1$. The regions of the light curve where the highest $\alpha_{\phi}$ values are found will be better to compute $\delta_{\phi}^{(c)}$ since the weight of $\delta_{\phi}^{(e)}/\delta_{y}^{(e)}$ is minimized. For instance, the highest $\alpha_{\phi}$ values for the sinusoidal variation will be found in the dark-grey region of Fig. \ref{fig_discerror}. \item[\textit{iii} -- ] Fig. \ref{fig_discerror} B shows a sinusoidal light curve considering $\delta_{\phi}^{(e)} = 0.02$ and $\delta_{y}^{(e)} = 0$. As expected $\delta_{y}^{(c)} \simeq 0$ is in the flattest region of the light curve. A note of caution, these regions are not well modeled by a straight line and non-linear effects, different from those analysed in items \textit{i} and \textit{ii}, can be found. Therefore, a fit to the whole light curve rather than a linear fit is necessary. The best estimation of the amplitude variation will be found if the region is small enough so that the model and linear fit are in agreement. Indeed, the periodic variation region can be approximately described by a linear model but the estimation of $\delta_{\phi}^{(c)}$ is computed using the time series model (see Sect. \ref{sec_visualdeltaphi}). The size and complexity of regions used to measure the period and amplitude variations are strongly dependent on the time series signal. To summarize, there is a non-zero contribution to the phase variation of the estimation of amplitude variations, if the region cannot be modeled by a horizontal line. On the other hand, $\delta_{\phi}^{(c)}$ is accurately estimated from Eq. \ref{eq_unper} since for this example $\delta_{\phi}^{(c)} = \delta_{\phi}^{(e)}$. \item[\textit{iv} -- ] A sinusoidal light curve considering $\delta_{\phi}^{(e)} = 0$ and $\delta_{y}^{(e)} = 0.05$ is shown in Fig. \ref{fig_discerror} C. From the approach described in item \textit{i} the estimation of $\delta_{y}^{(c)}$ is accurately estimated from Eq. \ref{eq_uncamp} since for this example $\delta_{y}^{(c)} = \delta_{y}^{(e)}$. On the other hand, $\delta_{\phi} \ne 0$ despite $\delta_{\phi}^{(e)} = 0$ for the the dark grey region in Fig. \ref{fig_discerror}. Indeed, $\delta_{\phi}$ will be equal to $\delta_{\phi}^{(e)}$ for the case where $\delta_{y}^{(e)} \ne 0$ only when $\alpha_{\phi} = \infty$, i.e a perpendicular line to the phase axis. Indeed, the phase variation is dominated by the amplitude variation in these cases since $\delta_{y}^{(e)}/\alpha_{\phi} >> \delta_{\phi}^{(e)}$. \item[\textit{v} -- ] Fig. \ref{fig_discerror} D shows the sinusoidal light curve when $\delta_{\phi}^{(e)} = 0.02$ and $\delta_{y}^{(e)} = 0.05$. It exemplifies a real time series where some variation in time and flux is always found. However the ratio of $\delta_{\phi}^{(e)}/\delta_{y}^{(e)}$ will determine the relative weights of each other (see Eqs. \ref{eq_uncamprel} and \ref{eq_unperrel}). For the current example $R^{(\delta_{y})} = \left| 1 + 0.4 \times \alpha_{y} \right|$ and $R^{(\delta_{\phi})} = \left| 1 + 2.5 \times \alpha_{\phi}^{-1} \right|$. Therefore, the best scenario to compute period and amplitude variations is the one where $\alpha_{y} = 0$ and $\alpha_{\phi} = \infty$. However, this is usually not the case, and hence such variations will not be computed precisely. Therefore, Monte-Carlo simulations are performed in Sect. \ref{sec_results} in order to estimate the inaccuracy of $\delta_{y}^{(c)}$ and $\delta_{\phi}^{(c)}$ as proxies for the variation on $\delta_{y}^{(e)}$ and $\delta_{\phi}^{(e)}$. \end{itemize} The discussion above does not take into account any particular light curve shape and hence this argument can be applied to all light curves types. Moreover, multiple regions of the phase diagram can be used to compute the amplitude and period variations. Indeed, these regions must be chosen following the discussion above in order to minimize the amplitude on period variations and vice versa, i.e use the flattest regions to compute the amplitude variation and the regions with the largest gradients (positive or negative) for the period variations. Indeed, time series having saddle regions also are suitable to compute the amplitude variation for the same reason discussed above (see panel A in Fig. \ref{fig_lcreg}). A more detailed description about how to compute the $\delta_{y}^{(c)}$ and $\delta_{\phi}^{(c)}$ is presented in Sect. \ref{sec_computeerror}. \begin{figure*} \centering \includegraphics[width=0.9\textwidth,height=0.9\textwidth]{fig_lctypes.eps} \caption{Ceph (purple - \textit{CoRoT-211626074}), RR (blue - \textit{CoRoT-101370131}), RRblz (yellow - \textit{CoRoT-100689962}), EA (red - \textit{CoRoT-102738809}), and Rot (green - \textit{CoRoT-110843734}) phase diagrams. The dark dots in the first panels indicate the original data while the solid lines set the Fourier models $M(\phi)$. The next panels were build from the models where half of them are set by squares (measurements at $t > t_{0}+1./f$) and another half by crosses (measurements at $t > T_{tot}-1./f$). The frequencies used to build the phase diagram from left to right panels are $\mathrm{f_{true} + \delta f}$ to $\delta_\phi = [0,0.05,0.1,0.2]$, respectively. } \label{fig_lctype} \end{figure*} \subsection{Computing period and amplitude variations}\label{sec_computeerror} Consider a generic light curve modeled by $Y\left(\phi\right)$ for $\left[ \phi_{1},\phi_{2},\cdots,\phi_{N} \right]$ where $\phi_{i}$ are in ascending order. The tangent angles to the model can be written as \begin{equation} \theta_{i} = \arctan \left(\frac{ Y\left( \phi_{i+1} \right) - Y\left( \phi_{i} \right) }{ \phi_{i+1} - \phi_{i}} \right). \label{eq_atan} \end{equation} The angles are better to use than the $\alpha$ values to determine suitable regions to compute the period and amplitude variations because they can be assessed from the model without making any additional computation (see Sect. \ref{sec_error}). The largest $\theta_{i}$ values are associated with the largest $\alpha$ values and the smallest $\theta_{i}$ values are associated with the smallest $\alpha$ values. The regions having smaller or bigger angles will be better to compute $\delta_{y}$ and $\delta_{\phi}$, respectively (see Sect. \ref{sec_error}). After defining the region to compute these variations the following procedures should be performed: \begin{itemize} \item Calculate the amplitude variation ($\sigma_{y}$): consider the region(s) that enclose the majority of measurements having $\mathrm{|\theta| < \theta_y}$. Next for each $[\phi_{i}^{(o)},y_{i}^{(o)}]$ we find its respective $[\phi_{i}^{(m)},y_{i}^{(m)}]$ from which the vector $\delta_{y} = \left[ y_{i}^{(o)}-y_{i}^{(m)} ,\cdots,y_{N}^{(o)}-y_{N`}^{(m)}\right]$ is obtained. Lastly, the amplitude variation is computed as \begin{equation} \sigma_{y} = \gamma \times eMAD\left( \delta_{y_{i}} \right) \label{eq_neweamp} \end{equation} \noindent where $\rm eMAD$ is the even-median absolute deviation of the even-median \citep[][]{FerreiraLopes-2017papII} and $\gamma$ is a correction factor (for more detail see Sect. \ref{sec_sortreg}). The $\rm eMAD$ is a slight modification to the $\rm MAD$ (the median absolute deviation of median). Indeed, $\sigma_{y}$ becomes a robust estimate of the standard deviation to outliers if $\gamma = 1.48$ according to \citet[][]{Hoaglin-1983}. A note of caution, $\delta_{y}$ is computed using $y_{i}^{(m)}$ instead of $y_{i}^{(e)}$ since the first one provides better estimations of expected values if the region cannot be well modeled by a line. Indeed, $y_{i}^{(m)} \simeq y_{i}^{(e)}$ only if $\theta_{i} \approx 0$. \item Calculate the period variation ($\sigma_{P}$): consider the region that encloses the majority of measurements having $\mathrm{|\theta| > \theta_P}$. For each $[\phi_{i}^{(o)},y_{i}^{(o)}]$ we find its respective $[\phi_{i}^{(e)},y_{i}^{(e)}]$, from which the vector $\delta_\phi = \left[ \phi_{i}^{(o)}-\phi_{i}^{(e)} ,\cdots, \phi_{N}^{(o)}-\phi_{N}^{(e)}\right]$ is obtained. Lastly, the period variation is computed as \begin{equation} \sigma_{P} = \gamma \times P \times eMAD\left( \delta_{\phi_i} \right). \label{eq_neweperr} \end{equation} \end{itemize} The current approach estimates the period and amplitude uncertainties taking into account the variations about a model. Eqs. \ref{eq_neweamp} and \ref{eq_neweperr} are computed using only those measurements suitable to reduce the weight of either $\delta_{y}$ or $\delta_{\phi}$. However, the accuracy of $\sigma_{y}$ and $\sigma_{P}$ are extremely dependent on $\theta_y$ and $\theta_P$, respectively. For instance, values of $\theta = [0.1^\circ,1^\circ,5^\circ,10^\circ,70^\circ,80^\circ,89^\circ,89.9^\circ]$ return $\alpha^{(\phi)} = [\sim0.002,\sim0.02,\sim0.09,\sim0.2,\sim2.8,\sim5.7,\sim57,\sim573]$. Indeed, the optimal choice of $\theta$ values is a compromise between the number of measurements enclosed for each limit and the usefulness of these measurements. Moreover, the statistical significance increases with the number of measurements while a higher signal-to-noise reduces the weight of $\delta_{y}$ on $\delta_{\phi}$. Therefore, the number of measurements and signal-to-noise are indirectly implicated in the period and amplitude variations. \section{Results and Discussion}\label{sec_results} Setting correct inputs using either method to select variable stars or to perform frequency finding searches is mandatory to get accurate outputs. The variability indices used to select variable stars candidates were studied deeply in the first two papers of this series, \citet[][]{FerreiraLopes-2016papI, FerreiraLopes-2017papII}. These studies enabled us to provide the optimal constraints on noise models and establish well-defined criteria to settle the best approach to discriminate variable stars from noise as well as to affirm that the selection of a reliable sample is unfeasible using variability indices. Therefore, frequency analysis may also be used to select out untrustworthy variations but all constraints must be properly delimited and understood to avoid mistakes. For instance, the interquartile range can provide an incorrect list of variable star candidates if the time sampling is not taken into account. Therefore, all the relevant points about frequency finding methods were discussed in Sect.~\ref{introduction}. The $\mathrm{f_{min}}$ and $\mathrm{f_{max}}$ are limited by the time series and the maximum reliable frequency, respectively. On the other hand, the sampling frequency was addressed in Sect. \ref{sec_freqsamp} in order to facilitate making a decision about the frequency resolution taking into account the effects on the frequency search. The frequency sampling and a new approach to computing period and amplitude variations are outlined in sections below. \begin{table} \begin{threeparttable} \centering \caption[]{Constraints on the frequency search analysis performed by different surveys. The designation, $\mathrm{f_{min}}$, $\mathrm{f_{max}}$, mean total time span $\mathrm{\overline{T_{tot}}}$, and $\mathrm{N_{f}}$ of each survey are shown. The frequency unit is day$^{-1}(d^{-1})$ and $\mathrm{\overline{T_{tot}}}$ is in days (d). Moreover, $\delta_{\phi}$ Eq.\ref{eq_deltaf} is given in the last column. }\label{tabconstraint} \begin{tabular}{c c c c c c} \hline\hline Survey & $\mathrm{f_{min}}(d^{-1})$ & $\mathrm{f_{max}}(d^{-1})$ & $\mathrm{\overline{T_{tot}}}(d)$ & $\mathrm{N_{f}}$ & $\delta_{\phi}$ \\ \hline CoRoT & $2/T_{tot}$ & $3$ & $\sim 136$ & $2\times 10^{3}$ & $0.20 $ \\ GAIA & $2/T_{tot}$ & $3.9$ & $\sim 1700$ & $\sim3\times 10^{3}$ & $0.19$ \\ Kepler & $\sim 3/T_{tot}$ & $1$ & $\sim 90$ & $1300$ & $0.07$ \\ OGLE & $0$ & $24$ & $\sim 2780$ & $10^{4}$ & $ > 1$ \\ TAROT$^{1}$ & $2/T_{tot}$ & $f_{max}$ & $\sim 900$ & $10^{5}$ & $\sim 0.22$ \\ WFCAM$^{2}$ & $2/T_{tot}$ & $f_{max}$ & $\sim 1058$ & $10^{5}$ & $\sim 0.25$ \\ \hline \end{tabular} \begin{tablenotes} \small \item 1 - The frequency step is taken as described in \citet[][]{Akerlof-1994} and \citet[][]{Larsson-1996}. \item 2 - $f_{max}$ computed according to \citet[][]{Eyer-1999}. \end{tablenotes} \end{threeparttable} \end{table} \subsection{Optimal frequency sampling}\label{sec_resolution} An optimal determination of $\mathrm{f_{max}}$ is critical to reducing running time since it leads to the determination of the resolution and thus the number of frequencies or loops performed by the frequency finding algorithm (see Eq.~\ref{eq_nfmax}). Estimation of $\mathrm{f_{max}}$ using the Nyquist frequency for oversampled data returns an overestimated frequency, i.e. frequencies that are this high are not reliably measured using the available data. Indeed, for unevenly and poorly sampled time series, the Nyquist frequency can be under or over-estimated whatever the estimation of the time interval from the measurements (as a mean or median value). For instance, long and short cadence CoRoT light curves have $\mathrm{f_{max}}$ of about $169d^{-1}$ and $2790d^{-1}$, respectively. These frequencies imply that the search for periodic variations at higher frequencies will not be productive. Therefore, empirical values have been adopted as the frequency limit. $\mathrm{f_{max}} = 10d^{-1}$ has been generally adopted \citep[e.g.][]{Debosscher-2007,Richards-2012,DeMedeiros-2013} but higher values also can be found \citep[e.g.][]{Schwarzenberg-Czerny-1996,Damerdji-2007,FerreiraLopes-2015wfcam}. The parameters used to perform frequency searches for variable star catalogs for some surveys are listed in Table \ref{tabconstraint}; the WFCAM multi-wavelength variable star catalog \citep[WFCAM - ][]{FerreiraLopes-2015wfcam}, the Optical Gravitational Lensing Experiment \citep[OGLE -][]{Soszynski-2009}, the TAROT suspected variable star catalog \citep[TAROT -][]{Damerdji-2007}, GAIA\footnote[1]{https://gaia.esac.esa.int/documentation/GDR1} data release 1 documentation, the semi-sinusoidal variables in the CoRoT mission \citep[SR-CoRoT -][]{DeMedeiros-2013}, rotation periods of 12 000 main-sequence Kepler stars \citep[Kepler -][]{Nielsen-2013}, and the WFCAM multiwavelength Variable Star Catalog \citep[WFCAM -][]{FerreiraLopes-2015wfcam}. The $\mathrm{f_{max}}$ adopted by OGLE was used to estimate $\delta_{\phi}$ for the WFCAM and TAROT catalogs. Indeed, $\mathrm{f_{max}}$ values given by analytical expressions in \citet[][]{Eyer-1999} depend on each time series and such values are usually much higher than those empirically adopted. The frequency sampling defined by Eq. \ref{eq_nfmax} was designed without taking into account any particular criteria and hence this expression may work for any signal type. Indeed, the number of constraints is not reduced, but the frequency sampling given by shifts on the phase $\delta_{\phi}$ instead of shifts in frequency is clearer to read. Moreover, Eq. \ref{eq_deltaf} also enables us to determine how much finer grain resolution is required to get a more accurate frequency estimation if the variability frequency is found since an initial value can be found with a coarser grain resolution. The frequencies not included in the frequency sampling may be detected or not, depending on the response to the frequency finding method for frequencies given by $\mathrm{f \pm \delta f/2}$, for instance. Indeed, the resolution of frequency sampling is critical for a large $T_{tot}$ since we find larger variations in the phase diagram for nearby frequencies. Moreover, as highlighted in previous sections, $\delta_{\phi}$ standardizes the criteria to perform frequency searches for time series having different total time spans. It allows us to compare straightforwardly the frequency analysis performed in different photometric surveys. \begin{figure*} \centering \includegraphics[width=0.9\textwidth,height=0.9\textwidth]{fig_lctypes_reg.eps} \caption{A single period (in contrast to Fig. \ref{fig_lctype} where two periods are shown) of the same models as shown in the Fig. \ref{fig_lctype}. The phase diagrams were built assuming different values of $\delta_{\phi}$ and $\delta_{y}$ that are displayed at the top of each column of panels unless to EA $B*$ and $D*$ panels that was assumed a $\delta_{\phi} = 0.025$. Indeed, the eclipse is missed for larger $\delta_{\phi}$ values. The regions used to compute the amplitude and period uncertainties are indicated by light grey and dark grey shading in all panels (for more details, see Sects. \ref{sec_error} and \ref{sec_testerror}).} \label{fig_lcreg} \end{figure*} \subsection{Visualizing frequency sampling effects}\label{sec_visualdeltaphi} Consider a periodic signal of $1d^{-1}$ with measurements covering a variability cycle from $\mathrm{t = 0}$ to $\mathrm{t = 0 + 1/f}$ and another from $\mathrm{t = T_{tot} - 1/f}$ to $\mathrm{t = T_{tot}}$. Five simulated time series that mimic pulsating stars ($Y_{(Ceph)}$, $Y_{(RR)}$, $Y_{(RRblz)}$), eclipsing binary stars ($Y_{(EA)}$), and rotational variables ($Y_{(Rot)}$) were chosen to illustrate our approach (for more details see Sect. \ref{sec_testreal}). Figure \ref{fig_lctype} shows phase diagrams of simulated light curves where the first column of panels show the results for $\mathrm{f_{true}}$. The grey dots indicate the original light curve while the models are indicated by purple (Ceph), blue (RR), yellow (RRblz), red (EA), and green (Rot) colours. The measurements located at $\mathrm{t = 0}$ to $\mathrm{t = 0 + 1/f}$ are indicated by squares while those at $\mathrm{t = T_{tot} - 1/f}$ to $\mathrm{t = T_{tot}}$ by crosses. The second, third, and fourth columns show phase diagrams using $\mathrm{f_{true} + \delta f}$ (see Eq. \ref{eq_deltaf}) for $\delta_\phi = [0.05,0.1,0.2]$, respectively. The crosses and squares limit the region where all measurements may be arranged considering that phase values computed at the beginning and end of the light curve set the largest variation from the model in the phase diagram as discussed in the Sect.~\ref{sec_freqsamp}. As one can see, the largest distortion of the model is found for binary stars, where the main variation is concentrated in a small part of the phase diagram. These aspects become increasingly important in the presence of noise or poorly sampled time series, when almost all measurements are required to adequately cover all variability phases. On the other hand, a low signal-to-noise is found for small frequency variations about $\mathrm{f_{true}}$ for those models where the variability is observed along the whole phase diagram like Ceph and RR. Indeed, the phase diagram dispersion is larger for those phenomena whose root variability causes period and/or amplitude variations like RRLyrae with the Blazhko effect (RRblz) and rotational variables \citep[e.g.][]{Buchler-2011, FerreiraLopes-2015cycles}. Indeed, non-radial pulsation, exoplanets, and different types of eclipsing and rotational variability enlarge the zoo of phase diagrams that can be produced by astrophysical phenomena \citep[e.g.][]{Prsa-2011, DeMedeiros-2013,FerreiraLopes-2015mgiant, Paz-Chinchon-2015}. To summarize, the phase diagrams of well-defined signals (fixed period and amplitude) only produce slight variations on the true frequency and hence these signals are easily identified compared to those ones with variable period or amplitude where the signal can be completely lost. Of course, the detection of these stars depends on the susceptibility to each frequency finding method. These matters will be addressed in a forthcoming paper of this project. The main conclusion provided by Eq. \ref{eq_nfmax} is a clear limit to the variations in which a smooth phase diagram can be found. \subsubsection{Sorting out $\theta_y$, $\theta_P$, and correction factors}\label{sec_sortreg} The same models described in the Sect. \ref{sec_visualdeltaphi} were used to test our assumptions. Figure~\ref{fig_lcreg} shows the phase diagrams of five typical light curves where the A panels show the model; the B panels show a variation in the period with a constant amplitude; constant period with amplitude variation (C panels), and both amplitude and period variations (D panels). These variations were added to the model using a random uniform distribution, that mimics a non instrumental variation, while an instrumental variation may appear like a normal distribution. Indeed, the real non instrumental variation is more complicated and may include variations with normal, uniform and perhaps more complicated distributions. For instance, the RR and Rot models at the maximum seem to be composed of normal and uniform variations that are not necessarily symmetric about the model, indicating a more complex variation (see Fig. \ref{fig_lctype} first panels). Eqs. \ref{eq_neweamp} and \ref{eq_neweperr} can be considered as a particular case where the noise or variation of amplitude or period is provided by a normal distribution since $1.48 \times MAD$ is approximately the standard deviation value \citep[][]{Hoaglin-1983}. A uniform distribution has a different spread of values. Therefore, a correction factor ($\gamma$) may be considered in order to take account of the distribution type. The percentage of values of $68.27\%$, $95.45\%$ and $99.73\%$ that lie within a band around the mean of a normal distribution is given by $\gamma = 1.48$, $\gamma = 2.96$, and $\gamma = 4.44$, respectively. However, $\gamma \simeq 1.37$, $\gamma \simeq 1.92$, and $\gamma \simeq 2.00$ contain the same fraction of values if an uniform distribution is considered. This factor improves our capability to measure an accurate estimation of the amplitude variation. For our simulation, this factor is not important since the ratio of computed and expected values are analysed (Sect. \ref{sec_testerror}). On the other hand, $\gamma = 1.48$ was used to estimate amplitude variation on real data (Sects. \ref{sec_testreal} and \ref{sec_catalina}). The period and amplitude variations computed are given by the sum of intrinsic and acquired variations. Acquired variations are those which come from the environment or instrument while intrinsic variations come from the source itself. Indeed, low values for the uncertainties are limited by the instrument properties and for constraints related with observability like the sky background, noise from background sources, and blending. For instance, the period and amplitude variations can reveal particularities of phenomena observed if the acquired uncertainties can be deducted from a noise model \citep[e.g.][]{Cross-2009,Aigrain-2009,FerreiraLopes-2017papII}. However the reliability of the period and amplitude variations measured will depend on the ratio $\delta_{\phi}/\delta_{y}$ as well as the regions used to compute them (see Sect. \ref{sec_error} for more detail). \begin{table} \caption[]{Angles ($\theta$) and angular coefficient ($\alpha$) values found for Ceph, RR, RRblz, EA, and Rot models. The angle limits $\theta_y$ and $\theta_P$ with their respective $\alpha_y$ and $\alpha_{\phi}$ values used to set the regions to compute the period and amplitude variations as well as the maximum angle found in each model are displayed below.} \label{tb_angles} \centering \begin{tabular}{l c c c c c} \hline\hline Model & $\mid\theta\mid_{max}$ & $\theta_y$ & $\mid\alpha_y\mid$ & $\theta_P$ & $\mid\alpha_{\phi}\mid$ \\ \hline Ceph & $79.02^o$ & $31.05^o$ & $0.18$ & $76.73^o$ & $4.68$ \\ RR & $85.52^o$ & $28.77^o$ & $0.37$ & $83.67^o$ & $10.99$ \\ RRblz & $81.30^o$ & $18.94^o$ & $0.10$ & $76.25^o$ & $4.74$ \\ EA & $88.39^o$ & $7.87^o$ & $0.04$ & $87.94^o$ & $32.00$ \\ Rot & $74.82^o$ & $34.87^o$ & $0.20$ & $72.90^o$ & $3.48$ \\ \hline \end{tabular} \end{table} \begin{figure} \centering \includegraphics[width=0.5\textwidth,height=0.8\textwidth]{fig_simulation.eps} \caption{The ratio of computed and expected uncertainty values for period ($R^{(\sigma_{P})}$ - blue crosses) and amplitude ($R^{(\sigma_{y})}$ - yellow squares) as function of $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ for EA, RRblz, RR, Ceph, and Rot models. The results where the noise was introduced using an uniform and normal distribution are displayed in the left and right panels, respectively.} \label{fig_simresults} \end{figure} \subsection{Testing frequency uncertainities}\label{sec_testerror} The models described in Sect. \ref{sec_visualdeltaphi} (see A panels of Fig.~\ref{fig_lctype}) were used to perform the simulations. The regions chosen to compute the amplitude and period uncertainties are shown in Fig.~\ref{fig_lcreg}. The measurements in these regions have angles within the defined angle limits which were set to best compute the uncertainties (see Sect. \ref{sec_error} for more details). Indeed, on average the maximum angle values are reduced and the minimum angle values are increased when the noise contribution is increased. Table~\ref{tb_angles} shows the main parameters values found in each model. Next, $10^6$ Monte Carlo simulations were performed setting $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ in the range from $0.1$ to $10$. $\mathrm{\sigma_{P}^{(e)}}$ and $\mathrm{\sigma_{y}^{(e)}}$ were introduced using a uniform distribution or a normal distribution. Finally, the amplitude and period uncertainties were computed according to Eqs. \ref{eq_neweamp} and \ref{eq_neweperr}. The ratio of the computed and expected uncertainty values for period ($R^{(\sigma_{P})}$) and amplitude ($R^{(\sigma_{y})}$) were used to estimate the reliability of computed values. Figure \ref{fig_simresults} shows the main results obtained in the simulations, which are summarized below; \begin{figure} \centering \includegraphics[width=0.5\textwidth,height=0.6\textwidth]{fig_sim_nump.eps} \caption{Percent relative error for $\sigma_P$ and $\sigma_y$ as a function of the number of measurements using an uniform (left panels) and a normal (right panels) distribution. The colours indicate the result for different models in the same way as Figs. \ref{fig_lctype} and \ref{fig_lcreg}.} \label{fig_simnumber} \end{figure} \begin{itemize} \item The results found using uniform and normal distributions are quite similar except for EA models. This happens because the eclipse is ''missed'' more quickly when the uncertainty is introduced by normal distributions than with uniform distributions. Considering the same sigma value for both distributions, a normal distribution of errors provides a larger dispersion of simulations than an uniform distribution. For instance, $\sim 4.44\times eMAD$ is required to enclose $\sim99.7\%$ of observed measurements for a normal distribution while $\sim 2.00\times eMAD$ is required for a uniform distribution (see Sect. \ref{sec_sortreg}). \item $R^{(\sigma_{P})} \simeq R^{(\sigma_{y})} \simeq 1$ is found for $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ ranging from $\sim0.5$ to $\sim2$ for all models as well as for both uniform and normal distributions. Indeed, the EA model has $R^{(\sigma_{P})} \simeq R^{(\sigma_{y})} \simeq 1$ for almost all values of the ratio. $\alpha_{y}$ is smaller than $0.1$ while $\alpha_{\phi}$ is bigger than $10$ for EA model and hence the weight of $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ on the computed uncertainties is reduced (see Table \ref{tb_angles}). \item The greatest difference between computed and expected values (R) are found at extreme ratios, i.e those regions where $\mathrm{\sigma_{P}^{(e)} >> \sigma_{y}^{(e)}}$ or $\mathrm{\sigma_{P}^{(e)} << \sigma_{y}^{(e)}}$. \begin{figure*} \centering \includegraphics[width=0.195\textwidth,height=0.4\textwidth]{fig_211626074.eps} \includegraphics[width=0.195\textwidth,height=0.4\textwidth]{fig_101370131.eps} \includegraphics[width=0.195\textwidth,height=0.4\textwidth]{fig_100689962.eps} \includegraphics[width=0.195\textwidth,height=0.4\textwidth]{fig_102738809.eps} \includegraphics[width=0.195\textwidth,height=0.4\textwidth]{fig_110843734.eps} \caption{Ceph (purple - \textit{CoRoT-211626074}), RR (blue - \textit{CoRoT-101370131}), RRblz (yellow - \textit{CoRoT-100689962}), EA (red - \textit{CoRoT-102738809}), and Rot (green - \textit{CoRoT-110843734}) phase diagrams in normalized flux shown in the top row of panels. The angles found for each models (second row of panels), the $\delta_{y}$ (third row of panels), and $\delta_{\phi}$ values (bottom row of panels) are also shown. Indeed, the last panel only show the results for the region used to compute the period variation. } \label{fig_realdatamodel} \end{figure*} \begin{table*} \begin{threeparttable} \caption[]{Parameters for CoRoT stars used to test the approach proposed in this work. The L indicate the parameters obtained in the literature which the references for are indicated in the last column. Indeed, the values of Ra, Dec, R magnitude, and the exposure time ($T_{exp}$) were obtained from the CoRoT database.}\label{tab_modeldata} \begin{tabular}{l c c c c c c c c c c c c } \hline\hline CoRoT-ID & Var. Type & RA & DEC & R & $\sigma_{(2h)}$ & $P_{(L)}(d)$ & $\delta P(d)$ & $T_{exp}(d)$ & $A_{(L)}(mag)$ & $eA_{(L)}(mag)$ & Ref \\ \hline 211626074 & Ceph & 285.469 & 3.277 & $12.60$ & $1.41\times 10^{-4}$ & $ 5.470600$ & $ \cdot$ & $5.93\times 10^{-3}$ & $ 2.96 \times10^{-1}$ & $1.44 \times10^{-3}$ & $ 1$ \\ 101370131 & RR & 292.060 & 0.101 & $15.28$ & $6.60\times 10^{-4}$ & $6.19332\times10^{-1}$ & $ \cdot$ & $5.93\times 10^{-3}$ & $ \cdot$ & $ \cdot$ & $ 2$ \\ 100689962 & RRblz & 291.000 & 1.697 & $14.65$ & $4.60\times 10^{-4}$ & $3.55997\times10^{-1}$ & $ \cdot$ & $5.93\times 10^{-3}$ & $ \cdot$ & $ \cdot$ & $ 3$ \\ 102738809 & EA & 101.131 & 0.832 & $12.29$ & $1.18\times 10^{-4}$ & $2.035701$ & $ \cdot$ & $3.70\times 10^{-4}$ & $ \cdot$ & $ \cdot$ & $ 4$ \\ 110843734 & Rot & 102.918 & -3.748 & $14.81$ & $5.03\times 10^{-4}$ & $8.186000$ & $ 4.94 \times10^{-2}$ & $5.93\times 10^{-3}$ & $5.66\times10^{-2}$ & $ 1.42\times10^{-2}$ & $ 5$ \\ \hline \end{tabular} \begin{tablenotes} \small \item The last column is regarding to the references that provided the following parameters above; (1) \citet[][]{Poretti-2015}, (2) \citet[][]{Paparo-2009} (3) \citet[][]{Chadid-2010}, (4) \citet[][]{Maciel-2011,Carone-2012} (5) \citet[][]{DeMedeiros-2013}. Moreover, the noise level ($\sigma_{2h}$) were computed using the Eq. 1 described by \citet[][]{Aigrain-2009} where z was computes as the mean value of CoRoT Run analysed by the authors. \end{tablenotes} \end{threeparttable} \end{table*} \item The discussion above can be summarized if the Eqs. \ref{eq_uncamprel} and \ref{eq_unperrel} are extrapolated thus: \begin{equation} R^{(\sigma_{y})} = \frac{\sigma_{y}^{(c)}}{\sigma_{y}^{(e)}} \simeq \left|\alpha_{y}\right| \times \frac{\sigma_{P}^{(e)}}{\sigma_{y}^{(e)}} + 1 \label{eq_uncamprel2} \end{equation} and \begin{equation} R^{(\sigma_{P})} =\frac{\sigma_{P}^{(c)}}{\sigma_{P}^{(e)}} \simeq \left|\alpha_{\phi}^{-1}\right| \times \left(\frac{\sigma_{P}^{(e)}}{\sigma_{y}^{(e)}} \right)^{-1} + 1. \label{eq_unperrel2} \end{equation} $R^{(\sigma_{y})}$ and $R^{(\sigma_{P})}$have opposite behaviour since they vary with $\mathrm{(\sigma_{P}^{(e)}/\sigma_{y}^{(e)}})^{\pm1}$, respectively. $R^{(\sigma_{y})}$ implies a rational function if $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ has values smaller than $1$ while the opposite is found for $R^{(\sigma_{P})}$. However, both functions depend on an angular coefficient ($\left|\alpha_{y}\right|$ or $\left|\alpha_{P}\right|$) that will determine the trend variation. \end{itemize} The simulations are in agreement with the analysis in Sect. \ref{sec_error}. The amplitude and period variations can bias the uncertainty estimations of one another, mainly when $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}} << 1$ or $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}} >> 1$. Indeed, Eqs. \ref{eq_uncamprel2} and \ref{eq_unperrel2} can be used to estimate the reliability of uncertainties if $\mathrm{\sigma_{P}^{(e)}/\sigma_{y}^{(e)}}$ can be estimated somehow. The relative errors of the uncertainties were also analysed as function of the number of measurements (see Fig. \ref{fig_simnumber}). As result, a decrease in the error with the number of measurements is found, as expected. This means that the number of measurements is an implicit parameter in Eqs. \ref{eq_neweamp} and \ref{eq_neweperr} that improve the statistic significance of uncertainties. \subsection{Describing Models and Testing the Approach on Observed Data}\label{sec_testreal} Ceph, RR, RRblz, EA, and Rot models were based on the CoRoT light curves \textit{CoRoT-211626074}, \textit{CoRoT-101370131}, \textit{CoRoT-100689962}, \textit{CoRoT-102738809}, and \textit{CoRoT-110843734}, respectively. The variability types were previously identified by \citet[][]{Debosscher-2007,Poretti-2015,Paparo-2009,Chadid-2010,Maciel-2011,Carone-2012} and \cite[][]{DeMedeiros-2013}. Table \ref{tab_modeldata} shows the main parameters of these sources that were obtained in the literature (L). These light curves were modeled using a harmonic fit with $12$, $12$, $12$, $24$, and $4$ harmonics for $Ceph$, $RR$, $RRblz$, $EA$, and $Rot$ variable stars, respectively. Higher number of harmonics can be used, however this also increases the processing time necessary to model and to perform simulations. The $Y_{(RRblz)}$ and $Y_{(Rot)}$ variable stars present variations in the amplitude and a period-amplitude variation. The $Y_{(RRblz)}$ has a Blazhko effect that is a long-period modulation or a variation in period and amplitude of RR Lyrae stars \citep[e.g.][]{Szabo-2014}. On the other hand, the $Y_{(Rot)}$ displays amplitude variation due the to the magnetic activity cycles and period variation due to differential rotation \citep[e.g.][]{FerreiraLopes-2015cycles,DasChagas-2016}. The exposure time ($T_{exp}$) provided by CoRoT mission and the empirical noise relation ($\sigma_{2h}$) described by \citet[][]{Aigrain-2009} were used to analysis the period and amplitude variation. The tests performed in the sections \ref{sec_error}, \ref{sec_visualdeltaphi}, and \ref{sec_sortreg} used models scaled to unit amplitude. It is useful to test our approaches for different signal types. For instance, the Ceph, RR, RRblz, EA, EB and Rot models have similar angles (see Table \ref{tb_angles}) but a wide difference among them is found when the real data is considered (see Fig. \ref{fig_realdatamodel}) since they have different typical amplitudes. Therefore, the angles found in the real data are not the same as those found for the models tested in the previous sections, as expected. These variations occur because $\tan(\theta) = \delta y/ \delta \phi $, i.e. a bigger $\delta y$ for the same $\delta \phi$ implies a larger angle. Figure \ref{fig_realdatamodel} shows the CoRoT light curves (first row of panels), the angles as a function of phase along the light curve (second row of panels), the observed minus modeled values (third row of panels), and finally the $\delta \phi$ values for the region used do compute the period variation. For example, the $\theta_{max}$ for $Rot$ models is about twelve times bigger than those found when amplitude is scaled to unit amplitude. On the other hand, the $\theta_{max}$ of the Ceph, RR, RRblz, EA, and EB decrease by factors smaller than $0.5$. \begin{figure*} \centering \includegraphics[width=0.95\textwidth,height=0.3\textwidth]{fig_residues.eps} \caption{The normalized standard deviation scaled between 0 and 1 as function of the percent relative error of $P_L$. The results for the residuals and period are shown as black and blue dots, respectively. The red line sets the location of the variability period determined in literature $P_{(L)}$. } \label{fig_normalizedsigma} \end{figure*} \begin{table*} \caption[]{Parameters for CoRoT stars computed from the aproaches proposed in this work.}\label{tab_modeldatanew} \begin{tabular}{l c c c c c c c c c} \hline\hline CoRoT-ID & $P(d)$ & $\delta P_{(FWHM)}^{(LSG)}(d)$ & $\delta P_{(FWHM)}^{(STR)}(d)$ & $\sigma_P(d)$ & $A(mag)$ & $\sigma_A(mag)$ & $\sigma_r(mag)$ & $\sigma_P/\sigma_A$ & $Log(T_{tot}/P)$ \\ \hline 211626074 & $5.47073174$ & $4.55\times 10^{-1}$ & $3.33\times 10^{-1}$ & $5.27\times 10^{-3}$ & $2.99\times 10^{-1}$ & $1.58\times 10^{-3}$ & $1.87\times 10^{-3}$ & $2.86$ & $6.69\times 10^{-1}$ \\ 101370131 & $6.19331408\times 10^{-1}$ & $9.15\times 10^{-4}$ & $7.46\times 10^{-4}$ & $5.80\times 10^{-4}$ & $7.98\times 10^{-1}$ & $7.53\times 10^{-3}$ & $8.14\times 10^{-3}$ & $0.06$ & $2.39$ \\ 100689962 & $3.56090879\times 10^{-1}$ & $3.27\times 10^{-4}$ & $2.93\times 10^{-4}$ & $2.23\times 10^{-2}$ & $4.50\times 10^{-1}$ & $3.82\times 10^{-2}$ & $6.19\times 10^{-2}$ & $0.52$ & $2.60$ \\ 102738809 & $2.03569293$ & $1.17\times 10^{-2}$ & $3.02\times 10^{-3}$ & $7.02\times 10^{-4}$ & $3.30\times 10^{-1}$ & $3.47\times 10^{-3}$ & $3.52\times 10^{-3}$ & $0.25$ & $1.81$ \\ 110843734 & $8.21895695$ & $1.74\times 10^{-1}$ & $1.70\times 10^{-1}$ & $1.71\times 10^{-1}$ & $2.68\times 10^{-2}$ & $6.06\times 10^{-3}$ & $7.76\times 10^{-3}$ & $30.01$ & $1.13$\\ \hline \end{tabular} \end{table*} Figure \ref{fig_realdatamodel} displays, step by step, the procedure that must be used to compute period and amplitude variations: the variability period is computed and the light curve is folded; next, a model is obtained using harmonic fits (see solid lines in the upper panels); from the models the angles are determined (see second row of panels) from where the regions used to compute period and amplitude variations are established; the amplitude variation is given by the standard deviation of the residuals in the region of phase diagram where $\left| \theta \right| < \theta_y$; and the period variation is found by multiplying the variability period by $eMAD$ of $\delta_{\phi}^{(c)}$ (given by Eq. \ref{eq_neweperr}). The periods and amplitudes as well as their uncertainties and variations were computed as described in Sect. \ref{sec_computeerror} (see Table \ref{tab_modeldatanew}). The results were compared with previous ones (see Table \ref{tab_modeldata}) where the main remarks are summarized below; \begin{itemize} \item The period that leads to the smallest $\sigma_{P}$ is not always related with the smallest $\sigma_{r}$ (see Fig. \ref{fig_normalizedsigma}). \item All $\sigma_{A}$ values are bigger than those given by $\sigma_{2h}$. This indicates an underestimation of $\sigma_{2h}$ or that all sources have an intrinsic amplitude variation. A note of caution, the noise values decrease with $Log(T_{tot}/P)$ and hence such a comparison cannot be performed straightforwardly. \item \textit{CoRoT-100689962}, \textit{CoRoT-110843734}, and \textit{CoRoT-102738809} have $\sigma_{P}$ values larger than the exposure time ($T_{exp}$). However, \textit{CoRoT-101370131} has $\sigma_{P}$ ten times smaller than $T_{exp}$. \item The $\sigma_{P}/\sigma_{A}$ values for all sources are smaller than $\sim0.5$ or bigger than $2$ (see Sect. \ref{sec_testerror}). It indicates that all $\sigma_{P}$ and $\sigma_{A}$ values are biased by amplitude or period variation, respectively. Indeed, the intrinsic variation is not known a priori and hence the information provided by the ratio $\sigma_{P}/\sigma_{A}$ will only be accurate if $\mid\alpha_y\mid << 0$ and $\mid\alpha_{\phi}\mid >> 1$ (see Sect. \ref{sec_error}). \item The variability periods determined by us are in agreement with those found in the literature. Indeed, the literature periods are determined as the highest power spectrum peak while those found by us are calculated by minimising $\sigma_{P}$. \item The period uncertainty $\delta P_{(FWHM)}^{(STR)}$ method is always smaller than $\delta P_{(FWHM)}^{(LSG)}$ that indicates that STR is more sensitive to variation in the phase diagram than the LSG method. \item \textit{CoRoT-211626074} - The amplitude ($A_{(L)}$) found in the literature is about $1\%$ smaller than that found by us. However, the authors used the DR2 release while our data come from the DR4 release. Indeed, the amplitudes are in agreement within the error bars. The $\sigma_{A}$ is at least nine times bigger than $\sigma_{2h}$. Moreover, $\sigma_{P}/\sigma_{A} = 2.86$ indicates that the weight of $\sigma_{P}$ in $\sigma_{A}$ is not strong, and vice-versa. It indicates that some of the amplitude variation comes from the sources. This result is supported by the detection of overtone pulsation reported by \citet[][]{Poretti-2015}. Indeed, the determination of amplitude variation reported by us was only settled by determination of $\sigma_{A}$ while the authors use complex analysis. \item \textit{CoRoT-101370131} - The $\sigma_{P}$ is smaller than $T_{exp}$ indicating a non-intrinsic variation related with the period. On the other hand, the amplitude variation $\sigma_{A}$ is about nine times bigger than $\sigma_{2h}$. Moreover, $\sigma_{P}/\sigma_{A} = 0.06$ also indicates that $\sigma_{A}$ is not biased by $\sigma_{P}$. Therefore, an intrinsic variation of the \textit{CoRoT-101370131} can be real if the noise level estimation is reliable. \item \textit{CoRoT-100689962} - The period and amplitude variation is clearly observed in the phase diagram. Moreover, it has the largest $Log(T_{tot}/P)$ and hence the smallest $\delta P_{(FWHM)}$ in agreement with the discussion performed in Sect. \ref{sec_freqerror}. Moreover, $\sigma_{P}/\sigma_{A} = 0.52$ indicates that the period variation is not strongly biased by amplitude variation and vice-versa. Therefore, the $\sigma_{P}$ and $\sigma_{A}$ mean that intrinsic variations come from the source since these variations are $\sim3.8$ times bigger than $T_{exp}$ and $\sim12$ times bigger than $\sigma_{2h}$, respectively. \item \textit{CoRoT-102738809} - The $\sigma_{P}$ is the smallest value among the sources analysed. Indeed, this aspect is caused by the large angles and the shape of the light curve. Moreover, this source has the shortest exposure time (see Table \ref{tab_modeldata}). The $\sigma_{P}$ does not show strong evidence of a period variation since it is smaller than twice $T_{exp}$. On the other hand, $\sigma_{A}$ is three times larger than $\sigma_{2h}$ that indicates a small intrinsic variation related with the amplitude. Indeed, the region used to compute the amplitude variation is related with the eclipse phase where both stars are side by side. Therefore, $\sigma_{A}$ can be related to one or both stars. \item \textit{CoRoT-110843734} - The $\delta P_{(FWHM)}$ is bigger than $\delta P$ and hence the empirical relation given by \citet[][]{Lamm-2004} can provide values smaller than those found for the $\delta P_{(FWHM)}$ estimations. $A_{(L)}$ is about twice that estimated by us. Such a difference can only be achieved by a typing error. On the other hand, the period computed by the authors is in agreement with that found for us. $\theta_{max} \sim 6^{o}$ and hence the period variation is biased by amplitude variation. Indeed, $\sigma_{P}/\sigma_{A} = 30.0$ indicates an unreliable estimation of $\sigma_{P}$ using the phase diagram. Therefore, $\sigma_{P}$ or $\sigma_{A}$ are not useful as indicators of intrinsic variation for rotational variables having small amplitudes. However, the estimation of period and amplitude variation with time instead of phase can reveal important clues about stellar activity cycles \citep[e.g.][]{FerreiraLopes-2015cycles}. \end{itemize} In summary, the period and amplitude variation can provide important information about the intrinsic variation of the source. However, it is trustworthy only if $\theta_{max} >> \theta_{min}$ since the capability to discriminate period and amplitude variation decreases. For a complete characterization of a light curve the period uncertainty as well as period variation must be determined. \begin{figure*} \centering \includegraphics[width=0.48\textwidth,height=0.3\textwidth]{1009030025196_pap3.eps} \includegraphics[width=0.48\textwidth,height=0.3\textwidth]{1149065077771_pap3.eps} \includegraphics[width=0.48\textwidth,height=0.3\textwidth]{1152031059450_pap3.eps} \includegraphics[width=0.48\textwidth,height=0.3\textwidth]{1157050034141_pap3.eps} \caption{ Phase diagram of 4 $EA_{up}$ Catalina stars where we also display the Lomb-Scargle periodogram about the variability period (top right panel for each star) and about half the variability period (bottom right panel for each star). The name and the variability period is shown in each diagram. The solid black line shows the power spectrum considering a number of frequencies obtained from Eq.\ref{eq_nfmax} assuming $\delta_{\phi} = 0.01$ while those values found using $N_{freq} = 10^4$ are marked by blue crosses. Moreover, the maximum power found for both methods of frequency sampling is displayed in the upper right corner. } \label{fig_eacatalina} \end{figure*} \subsection{Testing Frequency Sampling on Observed Data}\label{sec_catalina} The Catalina Real Time Survey found about $\sim47000$ periodic variables stars in Data Release-1 \citep[][]{Drake-2014}. The authors reported a sample of EA variables stars where the period determination was not possible due an insufficient number of observations at the eclipses. These stars were reported as EA variables having unknown-period ($EA_{up}$). The Lomb-Scargle method \citep[][]{Lomb-1976,Scargle-1982} was used to perform a period search but the frequency range and frequency sampling are not given by the authors. Therefore, a mean value of those shown in the Table \ref{tabconstraint} were assumed, i.e. $F_{min} = 2/T_{tot}$, $F_{max} = 10$, and $N_{freq} = 10^4$. These constraints were assumed as those used by the authors to review a small sample of $EA_{up}$ stars. Indeed, EA stars require a high frequency sampling to allow us to determine the variability period otherwise the eclipse region will not be smoothly folded (see $Y_{(EA)}$ panel Fig. \ref{fig_lctype}). Section \ref{sec_freqsamp} discussed the frequency sampling in detail where the $\delta_{\phi}$ required to find the variability periods for EA stars is smaller than $0.05$ in order to be able to fold the eclipse properly. Therefore, four $EA_{up}$ Catalina stars (see Table \ref{table_eacatalina}) were reviewed using the frequency sampling given by $\delta_{\phi} = 0.01$. Indeed, the sample analysed has $T_{tot} \simeq 3000$ days that implies a number of frequencies $\sim 3\times10^6$ (see Eq.\ref{eq_nfmax}). Figure \ref{fig_eacatalina} shows four $EA_{up}$ stars where the variability period was determined. In the right panel of each phase diagram is shown the Lomb-Scargle power spectrum about the variability period using $N_{freq} = 10^4$ (blue crosses) and a number of frequencies obtained from Eq.\ref{eq_nfmax} assuming $\delta_{\phi} = 0.01$ and $F_{max} = 10$. As one can verify the highest peak of the black lines is related to the maximum power of the periodogram. On the other hand, these peaks are not found when the sampling frequency is reduced (blue crosses). Therefore, the variability periods of $EA_{up}$ stars were not identified due to low frequency sampling. The main parameters of the four $EA_{up}$ stars analysed in the present work are presented in the Table \ref{table_eacatalina}. \begin{table*} \caption[]{ Parameters for $EA_{up}$ Catalina stars computed from the approaches proposed in this work.}\label{table_eacatalina} \scriptsize \begin{tabular}{l c c c c c c c c c} \hline\hline Catalina-ID & $P(d)$ & $\delta P_{(FWHM)}^{(LSG)}(d)$ & $\delta P_{(FWHM)}^{(STR)}(d)$ & $\sigma_P$ & V & $A$ & $\sigma_A$ & $\sigma_R$ & $Log(T_{tot}/P)$ \\ \hline CSS\_J053059.3-102647 & $1.16265491$ & $1.32\times 10^{-4}$ & $7.69\times 10^{-5}$ & $2.08\times 10^{-3}$ & $13.350$ & $7.50\times 10^{-1}$ & $2.83\times 10^{-2}$ & $2.17\times 10^{-2}$ & $3.40$ \\ CSS\_J180743.0+502014 & $9.92055466\times 10^{-1}$ & $8.20\times 10^{-5}$ & $6.99\times 10^{-5}$ & $1.22\times 10^{-3}$ & $15.810$ & $1.14$ & $2.71\times 10^{-2}$ & $1.09\times 10^{-1}$ & $3.49$ \\ CSS\_J090355.3+533131 & $3.31637448\times 10^{-1}$ & $1.33\times 10^{-5}$ & $2.49\times 10^{-5}$ & $2.27\times 10^{-3}$ & $15.095$ & $6.32\times 10^{-1}$ & $2.06\times 10^{-2}$ & $3.01\times 10^{-2}$ & $3.94$ \\ CSS\_J164404.4+574227 & $4.92343069\times 10^{-1}$ & $1.14\times 10^{-5}$ & $4.56\times 10^{-5}$ & $1.18\times 10^{-3}$ & $15.088$ & $9.92\times 10^{-1}$ & $1.64\times 10^{-2}$ & $2.91\times 10^{-2}$ & $3.74$ \\ \hline \end{tabular} \end{table*} Indeed, from a methodology viewpoint, the identification of variability periods of $EA_{up}$ stars requires a suitable period finding method and high frequency sampling (see Sect. \ref{sec_freqsamp}) to detect the signal. Moreover, the susceptibility of the period finding methods varies for different signal shapes (see Fig. \ref{fig_lctype}). Therefore, a deeper analysis of all $EA_{up}$ stars will be performed in a forthcoming paper where other methods besides Lomb-Scargle will also be used. As a result, we will define limits on what constitutes an insufficient number of observations. \section{Conclusions}\label{sec_conclusion} Frequency analysis constraints as well as the period and amplitude variations were analysed in the present work. A new approach to compute frequency sampling was introduced. This analysis is fundamental to providing a precise number of frequencies required to perform period finding searches. It also enables us to identify optimal values for searching for particular variable types as well as how much resolution is required to increase the accuracy of the periods found. We consider that this approach is fundamental to efficiently face the challenges of big data science since analytical equations are imposed. The period and amplitude variation of light curves were also reviewed. We show that a complete characterization of a light curve requires separating period uncertainty and period variation, from which important information about the variability nature can be estimated. On the other hand, the noise and amplitude variation also provide new clues about intrinsic variations that come from the source. The analyses performed in this project are very useful since all aspects of the analyses of large photometric surveys are being studied in order to maximize the probability of finding variable stars, reduce the running time, and reduce the number of misclassifications. The current paper is the second step towards unbiased samples, i.e. samples that only enclose reliable variations since this is unfeasible using correlated or non-correlated indices alone. Moreover, in this project we try to standardize the analysis criteria for variable stars in photometric surveys. In spite of this, the dependence of variability indices on the instrumental properties has been reduced and now, we also propose an estimation of frequency sampling that reduces the dependence on the total time span or time sampling of the data. Moreover, an approach to study the amplitude and period variation is presented. We consider that these estimations provide better information about the phenomena observed than previous ones since these estimations are limited by instrument properties or signal features. These must be taken into account for a realistic estimation. This paper concludes our studies about the constraints used to perform frequency searches. A new frequency finding method and new insights to detect aperiodic variations and to determine the false alarm probability will be addressed in a forthcoming paper of this series. \section*{Acknowledgements} C. E. F. L. acknowledges a post-doctoral fellowship from the CNPq. N. J. G. C. acknowledges support from the UK Science and Technology Facilities Council. The authors thank to MCTIC/FINEP (CT-INFRA grant 0112052700) and the Embrace Space Weather Program for the computing facilities at INPE. \bibliographystyle{mnras}

10fa68c9a1f5a4f6c8684b2c3d45dffafb39edbc